brintos

brintos / llvm-project-archived public Read only

0
0
Text · 12.1 MiB · 4c5c56a Raw
19455 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52 3; FIXME: Currently block machineinstr verifier due to SI BUNDLE pass break physical register liveness. Should remove when the issue is fixed up4 5; RUN: llc -mtriple=amdgcn -mcpu=tahiti -verify-machineinstrs=0 < %s | FileCheck -check-prefix=SI %s6; RUN: llc -mtriple=amdgcn -mcpu=tonga -verify-machineinstrs=0 < %s | FileCheck -check-prefix=VI %s7; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs=0 < %s | FileCheck -check-prefix=GFX9 %s8; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs=0 < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s9; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs=0 < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s10 11define <32 x float> @bitcast_v32i32_to_v32f32(<32 x i32> %a, i32 %b) {12; SI-LABEL: bitcast_v32i32_to_v32f32:13; SI:       ; %bb.0:14; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)15; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:416; SI-NEXT:    s_waitcnt vmcnt(0)17; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3118; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s3219; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc20; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]21; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]22; SI-NEXT:    s_cbranch_execz .LBB0_223; SI-NEXT:  ; %bb.1: ; %cmp.true24; SI-NEXT:    s_waitcnt vmcnt(0)25; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v3126; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v3027; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v2928; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v2829; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v2730; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v2631; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v2532; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v2433; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v2334; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v2235; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v2136; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v2037; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v1938; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v1839; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v1740; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v1641; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v1542; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v1443; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v1344; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v1245; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v1146; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v1047; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v948; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v849; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v750; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v651; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v552; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v453; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v354; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v255; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v156; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v057; SI-NEXT:  .LBB0_2: ; %end58; SI-NEXT:    s_or_b64 exec, exec, s[4:5]59; SI-NEXT:    s_waitcnt vmcnt(0)60; SI-NEXT:    s_setpc_b64 s[30:31]61;62; VI-LABEL: bitcast_v32i32_to_v32f32:63; VI:       ; %bb.0:64; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)65; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:466; VI-NEXT:    s_waitcnt vmcnt(0)67; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3168; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s3269; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc70; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]71; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]72; VI-NEXT:    s_cbranch_execz .LBB0_273; VI-NEXT:  ; %bb.1: ; %cmp.true74; VI-NEXT:    s_waitcnt vmcnt(0)75; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v3176; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v3077; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v2978; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v2879; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v2780; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v2681; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v2582; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v2483; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v2384; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v2285; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v2186; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v2087; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v1988; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v1889; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v1790; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v1691; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v1592; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v1493; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v1394; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v1295; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v1196; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v1097; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v998; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v899; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7100; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6101; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5102; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4103; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3104; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2105; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1106; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0107; VI-NEXT:  .LBB0_2: ; %end108; VI-NEXT:    s_or_b64 exec, exec, s[4:5]109; VI-NEXT:    s_waitcnt vmcnt(0)110; VI-NEXT:    s_setpc_b64 s[30:31]111;112; GFX9-LABEL: bitcast_v32i32_to_v32f32:113; GFX9:       ; %bb.0:114; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)115; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4116; GFX9-NEXT:    s_waitcnt vmcnt(0)117; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31118; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32119; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc120; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]121; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]122; GFX9-NEXT:    s_cbranch_execz .LBB0_2123; GFX9-NEXT:  ; %bb.1: ; %cmp.true124; GFX9-NEXT:    s_waitcnt vmcnt(0)125; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31126; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30127; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29128; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28129; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27130; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26131; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25132; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24133; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23134; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22135; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21136; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20137; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19138; GFX9-NEXT:    v_add_u32_e32 v18, 3, v18139; GFX9-NEXT:    v_add_u32_e32 v17, 3, v17140; GFX9-NEXT:    v_add_u32_e32 v16, 3, v16141; GFX9-NEXT:    v_add_u32_e32 v15, 3, v15142; GFX9-NEXT:    v_add_u32_e32 v14, 3, v14143; GFX9-NEXT:    v_add_u32_e32 v13, 3, v13144; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12145; GFX9-NEXT:    v_add_u32_e32 v11, 3, v11146; GFX9-NEXT:    v_add_u32_e32 v10, 3, v10147; GFX9-NEXT:    v_add_u32_e32 v9, 3, v9148; GFX9-NEXT:    v_add_u32_e32 v8, 3, v8149; GFX9-NEXT:    v_add_u32_e32 v7, 3, v7150; GFX9-NEXT:    v_add_u32_e32 v6, 3, v6151; GFX9-NEXT:    v_add_u32_e32 v5, 3, v5152; GFX9-NEXT:    v_add_u32_e32 v4, 3, v4153; GFX9-NEXT:    v_add_u32_e32 v3, 3, v3154; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2155; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1156; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0157; GFX9-NEXT:  .LBB0_2: ; %end158; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]159; GFX9-NEXT:    s_waitcnt vmcnt(0)160; GFX9-NEXT:    s_setpc_b64 s[30:31]161;162; GFX11-LABEL: bitcast_v32i32_to_v32f32:163; GFX11:       ; %bb.0:164; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)165; GFX11-NEXT:    s_clause 0x1166; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4167; GFX11-NEXT:    scratch_load_b32 v31, off, s32168; GFX11-NEXT:    s_mov_b32 s0, exec_lo169; GFX11-NEXT:    s_waitcnt vmcnt(1)170; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32171; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0172; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)173; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0174; GFX11-NEXT:    s_cbranch_execz .LBB0_2175; GFX11-NEXT:  ; %bb.1: ; %cmp.true176; GFX11-NEXT:    s_waitcnt vmcnt(0)177; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v31178; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v30179; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v29180; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v28181; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v27182; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v26183; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v25184; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v24185; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v23186; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v22187; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v21188; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v20189; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v19190; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v18191; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v17192; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v16193; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v15194; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v14195; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v13196; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v12197; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v11198; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v10199; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v9200; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v8201; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v7202; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v6203; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v5204; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v4205; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v3206; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v2207; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v1208; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v0209; GFX11-NEXT:  .LBB0_2: ; %end210; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0211; GFX11-NEXT:    s_waitcnt vmcnt(0)212; GFX11-NEXT:    s_setpc_b64 s[30:31]213  %cmp = icmp eq i32 %b, 0214  br i1 %cmp, label %cmp.true, label %cmp.false215 216cmp.true:217  %a1 = add <32 x i32> %a, splat (i32 3)218  %a2 = bitcast <32 x i32> %a1 to <32 x float>219  br label %end220 221cmp.false:222  %a3 = bitcast <32 x i32> %a to <32 x float>223  br label %end224 225end:226  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]227  ret <32 x float> %phi228}229 230define inreg <32 x float> @bitcast_v32i32_to_v32f32_scalar(<32 x i32> inreg %a, i32 inreg %b) {231; SI-LABEL: bitcast_v32i32_to_v32f32_scalar:232; SI:       ; %bb.0:233; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)234; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v18235; SI-NEXT:    v_mov_b32_e32 v31, v17236; SI-NEXT:    v_mov_b32_e32 v30, v16237; SI-NEXT:    v_mov_b32_e32 v29, v15238; SI-NEXT:    v_mov_b32_e32 v28, v14239; SI-NEXT:    v_mov_b32_e32 v27, v13240; SI-NEXT:    v_mov_b32_e32 v26, v12241; SI-NEXT:    v_mov_b32_e32 v25, v11242; SI-NEXT:    v_mov_b32_e32 v24, v10243; SI-NEXT:    v_mov_b32_e32 v23, v9244; SI-NEXT:    v_mov_b32_e32 v22, v8245; SI-NEXT:    v_mov_b32_e32 v21, v7246; SI-NEXT:    v_mov_b32_e32 v20, v6247; SI-NEXT:    v_mov_b32_e32 v19, v5248; SI-NEXT:    v_mov_b32_e32 v32, v4249; SI-NEXT:    v_mov_b32_e32 v17, v3250; SI-NEXT:    v_mov_b32_e32 v16, v2251; SI-NEXT:    v_mov_b32_e32 v15, v1252; SI-NEXT:    v_mov_b32_e32 v14, v0253; SI-NEXT:    v_mov_b32_e32 v0, s16254; SI-NEXT:    v_mov_b32_e32 v1, s17255; SI-NEXT:    s_and_b64 s[4:5], vcc, exec256; SI-NEXT:    v_mov_b32_e32 v2, s18257; SI-NEXT:    v_mov_b32_e32 v3, s19258; SI-NEXT:    v_mov_b32_e32 v4, s20259; SI-NEXT:    v_mov_b32_e32 v5, s21260; SI-NEXT:    v_mov_b32_e32 v6, s22261; SI-NEXT:    v_mov_b32_e32 v7, s23262; SI-NEXT:    v_mov_b32_e32 v8, s24263; SI-NEXT:    v_mov_b32_e32 v9, s25264; SI-NEXT:    v_mov_b32_e32 v10, s26265; SI-NEXT:    v_mov_b32_e32 v11, s27266; SI-NEXT:    v_mov_b32_e32 v12, s28267; SI-NEXT:    v_mov_b32_e32 v13, s29268; SI-NEXT:    s_cbranch_scc0 .LBB1_4269; SI-NEXT:  ; %bb.1: ; %cmp.false270; SI-NEXT:    s_cbranch_execnz .LBB1_3271; SI-NEXT:  .LBB1_2: ; %cmp.true272; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31273; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30274; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29275; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28276; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27277; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26278; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25279; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24280; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23281; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22282; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21283; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20284; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19285; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32286; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17287; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16288; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15289; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14290; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13291; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12292; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11293; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10294; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9295; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8296; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7297; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6298; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5299; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4300; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3301; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2302; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1303; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0304; SI-NEXT:  .LBB1_3: ; %end305; SI-NEXT:    v_mov_b32_e32 v18, v32306; SI-NEXT:    s_setpc_b64 s[30:31]307; SI-NEXT:  .LBB1_4:308; SI-NEXT:    s_branch .LBB1_2309;310; VI-LABEL: bitcast_v32i32_to_v32f32_scalar:311; VI:       ; %bb.0:312; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)313; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v18314; VI-NEXT:    v_mov_b32_e32 v31, v17315; VI-NEXT:    v_mov_b32_e32 v30, v16316; VI-NEXT:    v_mov_b32_e32 v29, v15317; VI-NEXT:    v_mov_b32_e32 v28, v14318; VI-NEXT:    v_mov_b32_e32 v27, v13319; VI-NEXT:    v_mov_b32_e32 v26, v12320; VI-NEXT:    v_mov_b32_e32 v25, v11321; VI-NEXT:    v_mov_b32_e32 v24, v10322; VI-NEXT:    v_mov_b32_e32 v23, v9323; VI-NEXT:    v_mov_b32_e32 v22, v8324; VI-NEXT:    v_mov_b32_e32 v21, v7325; VI-NEXT:    v_mov_b32_e32 v20, v6326; VI-NEXT:    v_mov_b32_e32 v19, v5327; VI-NEXT:    v_mov_b32_e32 v32, v4328; VI-NEXT:    v_mov_b32_e32 v17, v3329; VI-NEXT:    v_mov_b32_e32 v16, v2330; VI-NEXT:    v_mov_b32_e32 v15, v1331; VI-NEXT:    v_mov_b32_e32 v14, v0332; VI-NEXT:    v_mov_b32_e32 v0, s16333; VI-NEXT:    v_mov_b32_e32 v1, s17334; VI-NEXT:    s_and_b64 s[4:5], vcc, exec335; VI-NEXT:    v_mov_b32_e32 v2, s18336; VI-NEXT:    v_mov_b32_e32 v3, s19337; VI-NEXT:    v_mov_b32_e32 v4, s20338; VI-NEXT:    v_mov_b32_e32 v5, s21339; VI-NEXT:    v_mov_b32_e32 v6, s22340; VI-NEXT:    v_mov_b32_e32 v7, s23341; VI-NEXT:    v_mov_b32_e32 v8, s24342; VI-NEXT:    v_mov_b32_e32 v9, s25343; VI-NEXT:    v_mov_b32_e32 v10, s26344; VI-NEXT:    v_mov_b32_e32 v11, s27345; VI-NEXT:    v_mov_b32_e32 v12, s28346; VI-NEXT:    v_mov_b32_e32 v13, s29347; VI-NEXT:    s_cbranch_scc0 .LBB1_4348; VI-NEXT:  ; %bb.1: ; %cmp.false349; VI-NEXT:    s_cbranch_execnz .LBB1_3350; VI-NEXT:  .LBB1_2: ; %cmp.true351; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31352; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30353; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29354; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28355; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27356; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26357; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25358; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24359; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23360; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22361; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21362; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20363; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19364; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v32365; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v17366; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16367; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v15368; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14369; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v13370; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12371; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v11372; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10373; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v9374; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8375; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7376; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6377; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5378; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4379; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3380; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2381; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1382; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0383; VI-NEXT:  .LBB1_3: ; %end384; VI-NEXT:    v_mov_b32_e32 v18, v32385; VI-NEXT:    s_setpc_b64 s[30:31]386; VI-NEXT:  .LBB1_4:387; VI-NEXT:    s_branch .LBB1_2388;389; GFX9-LABEL: bitcast_v32i32_to_v32f32_scalar:390; GFX9:       ; %bb.0:391; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)392; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v18393; GFX9-NEXT:    v_mov_b32_e32 v31, v17394; GFX9-NEXT:    v_mov_b32_e32 v30, v16395; GFX9-NEXT:    v_mov_b32_e32 v29, v15396; GFX9-NEXT:    v_mov_b32_e32 v28, v14397; GFX9-NEXT:    v_mov_b32_e32 v27, v13398; GFX9-NEXT:    v_mov_b32_e32 v26, v12399; GFX9-NEXT:    v_mov_b32_e32 v25, v11400; GFX9-NEXT:    v_mov_b32_e32 v24, v10401; GFX9-NEXT:    v_mov_b32_e32 v23, v9402; GFX9-NEXT:    v_mov_b32_e32 v22, v8403; GFX9-NEXT:    v_mov_b32_e32 v21, v7404; GFX9-NEXT:    v_mov_b32_e32 v20, v6405; GFX9-NEXT:    v_mov_b32_e32 v19, v5406; GFX9-NEXT:    v_mov_b32_e32 v32, v4407; GFX9-NEXT:    v_mov_b32_e32 v17, v3408; GFX9-NEXT:    v_mov_b32_e32 v16, v2409; GFX9-NEXT:    v_mov_b32_e32 v15, v1410; GFX9-NEXT:    v_mov_b32_e32 v14, v0411; GFX9-NEXT:    v_mov_b32_e32 v0, s16412; GFX9-NEXT:    v_mov_b32_e32 v1, s17413; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec414; GFX9-NEXT:    v_mov_b32_e32 v2, s18415; GFX9-NEXT:    v_mov_b32_e32 v3, s19416; GFX9-NEXT:    v_mov_b32_e32 v4, s20417; GFX9-NEXT:    v_mov_b32_e32 v5, s21418; GFX9-NEXT:    v_mov_b32_e32 v6, s22419; GFX9-NEXT:    v_mov_b32_e32 v7, s23420; GFX9-NEXT:    v_mov_b32_e32 v8, s24421; GFX9-NEXT:    v_mov_b32_e32 v9, s25422; GFX9-NEXT:    v_mov_b32_e32 v10, s26423; GFX9-NEXT:    v_mov_b32_e32 v11, s27424; GFX9-NEXT:    v_mov_b32_e32 v12, s28425; GFX9-NEXT:    v_mov_b32_e32 v13, s29426; GFX9-NEXT:    s_cbranch_scc0 .LBB1_4427; GFX9-NEXT:  ; %bb.1: ; %cmp.false428; GFX9-NEXT:    s_cbranch_execnz .LBB1_3429; GFX9-NEXT:  .LBB1_2: ; %cmp.true430; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31431; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30432; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29433; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28434; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27435; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26436; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25437; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24438; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23439; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22440; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21441; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20442; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19443; GFX9-NEXT:    v_add_u32_e32 v32, 3, v32444; GFX9-NEXT:    v_add_u32_e32 v17, 3, v17445; GFX9-NEXT:    v_add_u32_e32 v16, 3, v16446; GFX9-NEXT:    v_add_u32_e32 v15, 3, v15447; GFX9-NEXT:    v_add_u32_e32 v14, 3, v14448; GFX9-NEXT:    v_add_u32_e32 v13, 3, v13449; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12450; GFX9-NEXT:    v_add_u32_e32 v11, 3, v11451; GFX9-NEXT:    v_add_u32_e32 v10, 3, v10452; GFX9-NEXT:    v_add_u32_e32 v9, 3, v9453; GFX9-NEXT:    v_add_u32_e32 v8, 3, v8454; GFX9-NEXT:    v_add_u32_e32 v7, 3, v7455; GFX9-NEXT:    v_add_u32_e32 v6, 3, v6456; GFX9-NEXT:    v_add_u32_e32 v5, 3, v5457; GFX9-NEXT:    v_add_u32_e32 v4, 3, v4458; GFX9-NEXT:    v_add_u32_e32 v3, 3, v3459; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2460; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1461; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0462; GFX9-NEXT:  .LBB1_3: ; %end463; GFX9-NEXT:    v_mov_b32_e32 v18, v32464; GFX9-NEXT:    s_setpc_b64 s[30:31]465; GFX9-NEXT:  .LBB1_4:466; GFX9-NEXT:    s_branch .LBB1_2467;468; GFX11-LABEL: bitcast_v32i32_to_v32f32_scalar:469; GFX11:       ; %bb.0:470; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)471; GFX11-NEXT:    v_dual_mov_b32 v16, v14 :: v_dual_mov_b32 v31, v13472; GFX11-NEXT:    v_dual_mov_b32 v30, v12 :: v_dual_mov_b32 v29, v11473; GFX11-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v9474; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)475; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v16476; GFX11-NEXT:    v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v25, v7477; GFX11-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v5478; GFX11-NEXT:    v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3479; GFX11-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1480; GFX11-NEXT:    v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1481; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3482; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17483; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19484; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21485; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23486; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25487; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27488; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29489; GFX11-NEXT:    v_mov_b32_e32 v16, s28490; GFX11-NEXT:    s_mov_b32 s0, 0491; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo492; GFX11-NEXT:    s_cbranch_scc0 .LBB1_3493; GFX11-NEXT:  ; %bb.1: ; %Flow494; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0495; GFX11-NEXT:    s_cbranch_vccz .LBB1_4496; GFX11-NEXT:  ; %bb.2: ; %end497; GFX11-NEXT:    s_setpc_b64 s[30:31]498; GFX11-NEXT:  .LBB1_3:499; GFX11-NEXT:  .LBB1_4: ; %cmp.true500; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v31501; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v30502; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v29503; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v28504; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v27505; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v26506; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v25507; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v24508; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v23509; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v22510; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v21511; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v20512; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v19513; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v18514; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v17515; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v16516; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v15517; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v14518; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v13519; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v12520; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v11521; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v10522; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v9523; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v8524; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v7525; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v6526; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v5527; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v4528; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v3529; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v2530; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v1531; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v0532; GFX11-NEXT:    s_setpc_b64 s[30:31]533  %cmp = icmp eq i32 %b, 0534  br i1 %cmp, label %cmp.true, label %cmp.false535 536cmp.true:537  %a1 = add <32 x i32> %a, splat (i32 3)538  %a2 = bitcast <32 x i32> %a1 to <32 x float>539  br label %end540 541cmp.false:542  %a3 = bitcast <32 x i32> %a to <32 x float>543  br label %end544 545end:546  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]547  ret <32 x float> %phi548}549 550define <32 x i32> @bitcast_v32f32_to_v32i32(<32 x float> %a, i32 %b) {551; SI-LABEL: bitcast_v32f32_to_v32i32:552; SI:       ; %bb.0:553; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)554; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4555; SI-NEXT:    s_waitcnt vmcnt(0)556; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31557; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32558; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc559; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]560; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]561; SI-NEXT:    s_cbranch_execz .LBB2_2562; SI-NEXT:  ; %bb.1: ; %cmp.true563; SI-NEXT:    s_waitcnt vmcnt(0)564; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31565; SI-NEXT:    v_add_f32_e32 v30, 1.0, v30566; SI-NEXT:    v_add_f32_e32 v29, 1.0, v29567; SI-NEXT:    v_add_f32_e32 v28, 1.0, v28568; SI-NEXT:    v_add_f32_e32 v27, 1.0, v27569; SI-NEXT:    v_add_f32_e32 v26, 1.0, v26570; SI-NEXT:    v_add_f32_e32 v25, 1.0, v25571; SI-NEXT:    v_add_f32_e32 v24, 1.0, v24572; SI-NEXT:    v_add_f32_e32 v23, 1.0, v23573; SI-NEXT:    v_add_f32_e32 v22, 1.0, v22574; SI-NEXT:    v_add_f32_e32 v21, 1.0, v21575; SI-NEXT:    v_add_f32_e32 v20, 1.0, v20576; SI-NEXT:    v_add_f32_e32 v19, 1.0, v19577; SI-NEXT:    v_add_f32_e32 v18, 1.0, v18578; SI-NEXT:    v_add_f32_e32 v17, 1.0, v17579; SI-NEXT:    v_add_f32_e32 v16, 1.0, v16580; SI-NEXT:    v_add_f32_e32 v15, 1.0, v15581; SI-NEXT:    v_add_f32_e32 v14, 1.0, v14582; SI-NEXT:    v_add_f32_e32 v13, 1.0, v13583; SI-NEXT:    v_add_f32_e32 v12, 1.0, v12584; SI-NEXT:    v_add_f32_e32 v11, 1.0, v11585; SI-NEXT:    v_add_f32_e32 v10, 1.0, v10586; SI-NEXT:    v_add_f32_e32 v9, 1.0, v9587; SI-NEXT:    v_add_f32_e32 v8, 1.0, v8588; SI-NEXT:    v_add_f32_e32 v7, 1.0, v7589; SI-NEXT:    v_add_f32_e32 v6, 1.0, v6590; SI-NEXT:    v_add_f32_e32 v5, 1.0, v5591; SI-NEXT:    v_add_f32_e32 v4, 1.0, v4592; SI-NEXT:    v_add_f32_e32 v3, 1.0, v3593; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2594; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1595; SI-NEXT:    v_add_f32_e32 v0, 1.0, v0596; SI-NEXT:  .LBB2_2: ; %end597; SI-NEXT:    s_or_b64 exec, exec, s[4:5]598; SI-NEXT:    s_waitcnt vmcnt(0)599; SI-NEXT:    s_setpc_b64 s[30:31]600;601; VI-LABEL: bitcast_v32f32_to_v32i32:602; VI:       ; %bb.0:603; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)604; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4605; VI-NEXT:    s_waitcnt vmcnt(0)606; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31607; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32608; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc609; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]610; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]611; VI-NEXT:    s_cbranch_execz .LBB2_2612; VI-NEXT:  ; %bb.1: ; %cmp.true613; VI-NEXT:    s_waitcnt vmcnt(0)614; VI-NEXT:    v_add_f32_e32 v31, 1.0, v31615; VI-NEXT:    v_add_f32_e32 v30, 1.0, v30616; VI-NEXT:    v_add_f32_e32 v29, 1.0, v29617; VI-NEXT:    v_add_f32_e32 v28, 1.0, v28618; VI-NEXT:    v_add_f32_e32 v27, 1.0, v27619; VI-NEXT:    v_add_f32_e32 v26, 1.0, v26620; VI-NEXT:    v_add_f32_e32 v25, 1.0, v25621; VI-NEXT:    v_add_f32_e32 v24, 1.0, v24622; VI-NEXT:    v_add_f32_e32 v23, 1.0, v23623; VI-NEXT:    v_add_f32_e32 v22, 1.0, v22624; VI-NEXT:    v_add_f32_e32 v21, 1.0, v21625; VI-NEXT:    v_add_f32_e32 v20, 1.0, v20626; VI-NEXT:    v_add_f32_e32 v19, 1.0, v19627; VI-NEXT:    v_add_f32_e32 v18, 1.0, v18628; VI-NEXT:    v_add_f32_e32 v17, 1.0, v17629; VI-NEXT:    v_add_f32_e32 v16, 1.0, v16630; VI-NEXT:    v_add_f32_e32 v15, 1.0, v15631; VI-NEXT:    v_add_f32_e32 v14, 1.0, v14632; VI-NEXT:    v_add_f32_e32 v13, 1.0, v13633; VI-NEXT:    v_add_f32_e32 v12, 1.0, v12634; VI-NEXT:    v_add_f32_e32 v11, 1.0, v11635; VI-NEXT:    v_add_f32_e32 v10, 1.0, v10636; VI-NEXT:    v_add_f32_e32 v9, 1.0, v9637; VI-NEXT:    v_add_f32_e32 v8, 1.0, v8638; VI-NEXT:    v_add_f32_e32 v7, 1.0, v7639; VI-NEXT:    v_add_f32_e32 v6, 1.0, v6640; VI-NEXT:    v_add_f32_e32 v5, 1.0, v5641; VI-NEXT:    v_add_f32_e32 v4, 1.0, v4642; VI-NEXT:    v_add_f32_e32 v3, 1.0, v3643; VI-NEXT:    v_add_f32_e32 v2, 1.0, v2644; VI-NEXT:    v_add_f32_e32 v1, 1.0, v1645; VI-NEXT:    v_add_f32_e32 v0, 1.0, v0646; VI-NEXT:  .LBB2_2: ; %end647; VI-NEXT:    s_or_b64 exec, exec, s[4:5]648; VI-NEXT:    s_waitcnt vmcnt(0)649; VI-NEXT:    s_setpc_b64 s[30:31]650;651; GFX9-LABEL: bitcast_v32f32_to_v32i32:652; GFX9:       ; %bb.0:653; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)654; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4655; GFX9-NEXT:    s_waitcnt vmcnt(0)656; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31657; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32658; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc659; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]660; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]661; GFX9-NEXT:    s_cbranch_execz .LBB2_2662; GFX9-NEXT:  ; %bb.1: ; %cmp.true663; GFX9-NEXT:    s_waitcnt vmcnt(0)664; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31665; GFX9-NEXT:    v_add_f32_e32 v30, 1.0, v30666; GFX9-NEXT:    v_add_f32_e32 v29, 1.0, v29667; GFX9-NEXT:    v_add_f32_e32 v28, 1.0, v28668; GFX9-NEXT:    v_add_f32_e32 v27, 1.0, v27669; GFX9-NEXT:    v_add_f32_e32 v26, 1.0, v26670; GFX9-NEXT:    v_add_f32_e32 v25, 1.0, v25671; GFX9-NEXT:    v_add_f32_e32 v24, 1.0, v24672; GFX9-NEXT:    v_add_f32_e32 v23, 1.0, v23673; GFX9-NEXT:    v_add_f32_e32 v22, 1.0, v22674; GFX9-NEXT:    v_add_f32_e32 v21, 1.0, v21675; GFX9-NEXT:    v_add_f32_e32 v20, 1.0, v20676; GFX9-NEXT:    v_add_f32_e32 v19, 1.0, v19677; GFX9-NEXT:    v_add_f32_e32 v18, 1.0, v18678; GFX9-NEXT:    v_add_f32_e32 v17, 1.0, v17679; GFX9-NEXT:    v_add_f32_e32 v16, 1.0, v16680; GFX9-NEXT:    v_add_f32_e32 v15, 1.0, v15681; GFX9-NEXT:    v_add_f32_e32 v14, 1.0, v14682; GFX9-NEXT:    v_add_f32_e32 v13, 1.0, v13683; GFX9-NEXT:    v_add_f32_e32 v12, 1.0, v12684; GFX9-NEXT:    v_add_f32_e32 v11, 1.0, v11685; GFX9-NEXT:    v_add_f32_e32 v10, 1.0, v10686; GFX9-NEXT:    v_add_f32_e32 v9, 1.0, v9687; GFX9-NEXT:    v_add_f32_e32 v8, 1.0, v8688; GFX9-NEXT:    v_add_f32_e32 v7, 1.0, v7689; GFX9-NEXT:    v_add_f32_e32 v6, 1.0, v6690; GFX9-NEXT:    v_add_f32_e32 v5, 1.0, v5691; GFX9-NEXT:    v_add_f32_e32 v4, 1.0, v4692; GFX9-NEXT:    v_add_f32_e32 v3, 1.0, v3693; GFX9-NEXT:    v_add_f32_e32 v2, 1.0, v2694; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1695; GFX9-NEXT:    v_add_f32_e32 v0, 1.0, v0696; GFX9-NEXT:  .LBB2_2: ; %end697; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]698; GFX9-NEXT:    s_waitcnt vmcnt(0)699; GFX9-NEXT:    s_setpc_b64 s[30:31]700;701; GFX11-LABEL: bitcast_v32f32_to_v32i32:702; GFX11:       ; %bb.0:703; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)704; GFX11-NEXT:    s_clause 0x1705; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4706; GFX11-NEXT:    scratch_load_b32 v31, off, s32707; GFX11-NEXT:    s_mov_b32 s0, exec_lo708; GFX11-NEXT:    s_waitcnt vmcnt(1)709; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32710; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0711; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)712; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0713; GFX11-NEXT:    s_cbranch_execz .LBB2_2714; GFX11-NEXT:  ; %bb.1: ; %cmp.true715; GFX11-NEXT:    s_waitcnt vmcnt(0)716; GFX11-NEXT:    v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30717; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28718; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26719; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24720; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22721; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20722; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18723; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16724; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14725; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12726; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10727; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8728; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6729; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4730; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2731; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0732; GFX11-NEXT:  .LBB2_2: ; %end733; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0734; GFX11-NEXT:    s_waitcnt vmcnt(0)735; GFX11-NEXT:    s_setpc_b64 s[30:31]736  %cmp = icmp eq i32 %b, 0737  br i1 %cmp, label %cmp.true, label %cmp.false738 739cmp.true:740  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)741  %a2 = bitcast <32 x float> %a1 to <32 x i32>742  br label %end743 744cmp.false:745  %a3 = bitcast <32 x float> %a to <32 x i32>746  br label %end747 748end:749  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]750  ret <32 x i32> %phi751}752 753define inreg <32 x i32> @bitcast_v32f32_to_v32i32_scalar(<32 x float> inreg %a, i32 inreg %b) {754; SI-LABEL: bitcast_v32f32_to_v32i32_scalar:755; SI:       ; %bb.0:756; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)757; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v18758; SI-NEXT:    v_mov_b32_e32 v31, v17759; SI-NEXT:    v_mov_b32_e32 v30, v16760; SI-NEXT:    v_mov_b32_e32 v29, v15761; SI-NEXT:    v_mov_b32_e32 v28, v14762; SI-NEXT:    v_mov_b32_e32 v27, v13763; SI-NEXT:    v_mov_b32_e32 v26, v12764; SI-NEXT:    v_mov_b32_e32 v25, v11765; SI-NEXT:    v_mov_b32_e32 v24, v10766; SI-NEXT:    v_mov_b32_e32 v23, v9767; SI-NEXT:    v_mov_b32_e32 v22, v8768; SI-NEXT:    v_mov_b32_e32 v21, v7769; SI-NEXT:    v_mov_b32_e32 v20, v6770; SI-NEXT:    v_mov_b32_e32 v19, v5771; SI-NEXT:    v_mov_b32_e32 v32, v4772; SI-NEXT:    v_mov_b32_e32 v17, v3773; SI-NEXT:    v_mov_b32_e32 v16, v2774; SI-NEXT:    v_mov_b32_e32 v15, v1775; SI-NEXT:    v_mov_b32_e32 v14, v0776; SI-NEXT:    v_mov_b32_e32 v0, s16777; SI-NEXT:    v_mov_b32_e32 v1, s17778; SI-NEXT:    s_and_b64 s[4:5], vcc, exec779; SI-NEXT:    v_mov_b32_e32 v2, s18780; SI-NEXT:    v_mov_b32_e32 v3, s19781; SI-NEXT:    v_mov_b32_e32 v4, s20782; SI-NEXT:    v_mov_b32_e32 v5, s21783; SI-NEXT:    v_mov_b32_e32 v6, s22784; SI-NEXT:    v_mov_b32_e32 v7, s23785; SI-NEXT:    v_mov_b32_e32 v8, s24786; SI-NEXT:    v_mov_b32_e32 v9, s25787; SI-NEXT:    v_mov_b32_e32 v10, s26788; SI-NEXT:    v_mov_b32_e32 v11, s27789; SI-NEXT:    v_mov_b32_e32 v12, s28790; SI-NEXT:    v_mov_b32_e32 v13, s29791; SI-NEXT:    s_cbranch_scc0 .LBB3_4792; SI-NEXT:  ; %bb.1: ; %cmp.false793; SI-NEXT:    s_cbranch_execnz .LBB3_3794; SI-NEXT:  .LBB3_2: ; %cmp.true795; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31796; SI-NEXT:    v_add_f32_e32 v30, 1.0, v30797; SI-NEXT:    v_add_f32_e32 v29, 1.0, v29798; SI-NEXT:    v_add_f32_e32 v28, 1.0, v28799; SI-NEXT:    v_add_f32_e32 v27, 1.0, v27800; SI-NEXT:    v_add_f32_e32 v26, 1.0, v26801; SI-NEXT:    v_add_f32_e32 v25, 1.0, v25802; SI-NEXT:    v_add_f32_e32 v24, 1.0, v24803; SI-NEXT:    v_add_f32_e32 v23, 1.0, v23804; SI-NEXT:    v_add_f32_e32 v22, 1.0, v22805; SI-NEXT:    v_add_f32_e32 v21, 1.0, v21806; SI-NEXT:    v_add_f32_e32 v20, 1.0, v20807; SI-NEXT:    v_add_f32_e32 v19, 1.0, v19808; SI-NEXT:    v_add_f32_e32 v32, 1.0, v32809; SI-NEXT:    v_add_f32_e32 v17, 1.0, v17810; SI-NEXT:    v_add_f32_e32 v16, 1.0, v16811; SI-NEXT:    v_add_f32_e32 v15, 1.0, v15812; SI-NEXT:    v_add_f32_e32 v14, 1.0, v14813; SI-NEXT:    v_add_f32_e32 v13, 1.0, v13814; SI-NEXT:    v_add_f32_e32 v12, 1.0, v12815; SI-NEXT:    v_add_f32_e32 v11, 1.0, v11816; SI-NEXT:    v_add_f32_e32 v10, 1.0, v10817; SI-NEXT:    v_add_f32_e32 v9, 1.0, v9818; SI-NEXT:    v_add_f32_e32 v8, 1.0, v8819; SI-NEXT:    v_add_f32_e32 v7, 1.0, v7820; SI-NEXT:    v_add_f32_e32 v6, 1.0, v6821; SI-NEXT:    v_add_f32_e32 v5, 1.0, v5822; SI-NEXT:    v_add_f32_e32 v4, 1.0, v4823; SI-NEXT:    v_add_f32_e32 v3, 1.0, v3824; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2825; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1826; SI-NEXT:    v_add_f32_e32 v0, 1.0, v0827; SI-NEXT:  .LBB3_3: ; %end828; SI-NEXT:    v_mov_b32_e32 v18, v32829; SI-NEXT:    s_setpc_b64 s[30:31]830; SI-NEXT:  .LBB3_4:831; SI-NEXT:    s_branch .LBB3_2832;833; VI-LABEL: bitcast_v32f32_to_v32i32_scalar:834; VI:       ; %bb.0:835; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)836; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v18837; VI-NEXT:    v_mov_b32_e32 v31, v17838; VI-NEXT:    v_mov_b32_e32 v30, v16839; VI-NEXT:    v_mov_b32_e32 v29, v15840; VI-NEXT:    v_mov_b32_e32 v28, v14841; VI-NEXT:    v_mov_b32_e32 v27, v13842; VI-NEXT:    v_mov_b32_e32 v26, v12843; VI-NEXT:    v_mov_b32_e32 v25, v11844; VI-NEXT:    v_mov_b32_e32 v24, v10845; VI-NEXT:    v_mov_b32_e32 v23, v9846; VI-NEXT:    v_mov_b32_e32 v22, v8847; VI-NEXT:    v_mov_b32_e32 v21, v7848; VI-NEXT:    v_mov_b32_e32 v20, v6849; VI-NEXT:    v_mov_b32_e32 v19, v5850; VI-NEXT:    v_mov_b32_e32 v32, v4851; VI-NEXT:    v_mov_b32_e32 v17, v3852; VI-NEXT:    v_mov_b32_e32 v16, v2853; VI-NEXT:    v_mov_b32_e32 v15, v1854; VI-NEXT:    v_mov_b32_e32 v14, v0855; VI-NEXT:    v_mov_b32_e32 v0, s16856; VI-NEXT:    v_mov_b32_e32 v1, s17857; VI-NEXT:    s_and_b64 s[4:5], vcc, exec858; VI-NEXT:    v_mov_b32_e32 v2, s18859; VI-NEXT:    v_mov_b32_e32 v3, s19860; VI-NEXT:    v_mov_b32_e32 v4, s20861; VI-NEXT:    v_mov_b32_e32 v5, s21862; VI-NEXT:    v_mov_b32_e32 v6, s22863; VI-NEXT:    v_mov_b32_e32 v7, s23864; VI-NEXT:    v_mov_b32_e32 v8, s24865; VI-NEXT:    v_mov_b32_e32 v9, s25866; VI-NEXT:    v_mov_b32_e32 v10, s26867; VI-NEXT:    v_mov_b32_e32 v11, s27868; VI-NEXT:    v_mov_b32_e32 v12, s28869; VI-NEXT:    v_mov_b32_e32 v13, s29870; VI-NEXT:    s_cbranch_scc0 .LBB3_4871; VI-NEXT:  ; %bb.1: ; %cmp.false872; VI-NEXT:    s_cbranch_execnz .LBB3_3873; VI-NEXT:  .LBB3_2: ; %cmp.true874; VI-NEXT:    v_add_f32_e32 v31, 1.0, v31875; VI-NEXT:    v_add_f32_e32 v30, 1.0, v30876; VI-NEXT:    v_add_f32_e32 v29, 1.0, v29877; VI-NEXT:    v_add_f32_e32 v28, 1.0, v28878; VI-NEXT:    v_add_f32_e32 v27, 1.0, v27879; VI-NEXT:    v_add_f32_e32 v26, 1.0, v26880; VI-NEXT:    v_add_f32_e32 v25, 1.0, v25881; VI-NEXT:    v_add_f32_e32 v24, 1.0, v24882; VI-NEXT:    v_add_f32_e32 v23, 1.0, v23883; VI-NEXT:    v_add_f32_e32 v22, 1.0, v22884; VI-NEXT:    v_add_f32_e32 v21, 1.0, v21885; VI-NEXT:    v_add_f32_e32 v20, 1.0, v20886; VI-NEXT:    v_add_f32_e32 v19, 1.0, v19887; VI-NEXT:    v_add_f32_e32 v32, 1.0, v32888; VI-NEXT:    v_add_f32_e32 v17, 1.0, v17889; VI-NEXT:    v_add_f32_e32 v16, 1.0, v16890; VI-NEXT:    v_add_f32_e32 v15, 1.0, v15891; VI-NEXT:    v_add_f32_e32 v14, 1.0, v14892; VI-NEXT:    v_add_f32_e32 v13, 1.0, v13893; VI-NEXT:    v_add_f32_e32 v12, 1.0, v12894; VI-NEXT:    v_add_f32_e32 v11, 1.0, v11895; VI-NEXT:    v_add_f32_e32 v10, 1.0, v10896; VI-NEXT:    v_add_f32_e32 v9, 1.0, v9897; VI-NEXT:    v_add_f32_e32 v8, 1.0, v8898; VI-NEXT:    v_add_f32_e32 v7, 1.0, v7899; VI-NEXT:    v_add_f32_e32 v6, 1.0, v6900; VI-NEXT:    v_add_f32_e32 v5, 1.0, v5901; VI-NEXT:    v_add_f32_e32 v4, 1.0, v4902; VI-NEXT:    v_add_f32_e32 v3, 1.0, v3903; VI-NEXT:    v_add_f32_e32 v2, 1.0, v2904; VI-NEXT:    v_add_f32_e32 v1, 1.0, v1905; VI-NEXT:    v_add_f32_e32 v0, 1.0, v0906; VI-NEXT:  .LBB3_3: ; %end907; VI-NEXT:    v_mov_b32_e32 v18, v32908; VI-NEXT:    s_setpc_b64 s[30:31]909; VI-NEXT:  .LBB3_4:910; VI-NEXT:    s_branch .LBB3_2911;912; GFX9-LABEL: bitcast_v32f32_to_v32i32_scalar:913; GFX9:       ; %bb.0:914; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)915; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v18916; GFX9-NEXT:    v_mov_b32_e32 v31, v17917; GFX9-NEXT:    v_mov_b32_e32 v30, v16918; GFX9-NEXT:    v_mov_b32_e32 v29, v15919; GFX9-NEXT:    v_mov_b32_e32 v28, v14920; GFX9-NEXT:    v_mov_b32_e32 v27, v13921; GFX9-NEXT:    v_mov_b32_e32 v26, v12922; GFX9-NEXT:    v_mov_b32_e32 v25, v11923; GFX9-NEXT:    v_mov_b32_e32 v24, v10924; GFX9-NEXT:    v_mov_b32_e32 v23, v9925; GFX9-NEXT:    v_mov_b32_e32 v22, v8926; GFX9-NEXT:    v_mov_b32_e32 v21, v7927; GFX9-NEXT:    v_mov_b32_e32 v20, v6928; GFX9-NEXT:    v_mov_b32_e32 v19, v5929; GFX9-NEXT:    v_mov_b32_e32 v32, v4930; GFX9-NEXT:    v_mov_b32_e32 v17, v3931; GFX9-NEXT:    v_mov_b32_e32 v16, v2932; GFX9-NEXT:    v_mov_b32_e32 v15, v1933; GFX9-NEXT:    v_mov_b32_e32 v14, v0934; GFX9-NEXT:    v_mov_b32_e32 v0, s16935; GFX9-NEXT:    v_mov_b32_e32 v1, s17936; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec937; GFX9-NEXT:    v_mov_b32_e32 v2, s18938; GFX9-NEXT:    v_mov_b32_e32 v3, s19939; GFX9-NEXT:    v_mov_b32_e32 v4, s20940; GFX9-NEXT:    v_mov_b32_e32 v5, s21941; GFX9-NEXT:    v_mov_b32_e32 v6, s22942; GFX9-NEXT:    v_mov_b32_e32 v7, s23943; GFX9-NEXT:    v_mov_b32_e32 v8, s24944; GFX9-NEXT:    v_mov_b32_e32 v9, s25945; GFX9-NEXT:    v_mov_b32_e32 v10, s26946; GFX9-NEXT:    v_mov_b32_e32 v11, s27947; GFX9-NEXT:    v_mov_b32_e32 v12, s28948; GFX9-NEXT:    v_mov_b32_e32 v13, s29949; GFX9-NEXT:    s_cbranch_scc0 .LBB3_4950; GFX9-NEXT:  ; %bb.1: ; %cmp.false951; GFX9-NEXT:    s_cbranch_execnz .LBB3_3952; GFX9-NEXT:  .LBB3_2: ; %cmp.true953; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31954; GFX9-NEXT:    v_add_f32_e32 v30, 1.0, v30955; GFX9-NEXT:    v_add_f32_e32 v29, 1.0, v29956; GFX9-NEXT:    v_add_f32_e32 v28, 1.0, v28957; GFX9-NEXT:    v_add_f32_e32 v27, 1.0, v27958; GFX9-NEXT:    v_add_f32_e32 v26, 1.0, v26959; GFX9-NEXT:    v_add_f32_e32 v25, 1.0, v25960; GFX9-NEXT:    v_add_f32_e32 v24, 1.0, v24961; GFX9-NEXT:    v_add_f32_e32 v23, 1.0, v23962; GFX9-NEXT:    v_add_f32_e32 v22, 1.0, v22963; GFX9-NEXT:    v_add_f32_e32 v21, 1.0, v21964; GFX9-NEXT:    v_add_f32_e32 v20, 1.0, v20965; GFX9-NEXT:    v_add_f32_e32 v19, 1.0, v19966; GFX9-NEXT:    v_add_f32_e32 v32, 1.0, v32967; GFX9-NEXT:    v_add_f32_e32 v17, 1.0, v17968; GFX9-NEXT:    v_add_f32_e32 v16, 1.0, v16969; GFX9-NEXT:    v_add_f32_e32 v15, 1.0, v15970; GFX9-NEXT:    v_add_f32_e32 v14, 1.0, v14971; GFX9-NEXT:    v_add_f32_e32 v13, 1.0, v13972; GFX9-NEXT:    v_add_f32_e32 v12, 1.0, v12973; GFX9-NEXT:    v_add_f32_e32 v11, 1.0, v11974; GFX9-NEXT:    v_add_f32_e32 v10, 1.0, v10975; GFX9-NEXT:    v_add_f32_e32 v9, 1.0, v9976; GFX9-NEXT:    v_add_f32_e32 v8, 1.0, v8977; GFX9-NEXT:    v_add_f32_e32 v7, 1.0, v7978; GFX9-NEXT:    v_add_f32_e32 v6, 1.0, v6979; GFX9-NEXT:    v_add_f32_e32 v5, 1.0, v5980; GFX9-NEXT:    v_add_f32_e32 v4, 1.0, v4981; GFX9-NEXT:    v_add_f32_e32 v3, 1.0, v3982; GFX9-NEXT:    v_add_f32_e32 v2, 1.0, v2983; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1984; GFX9-NEXT:    v_add_f32_e32 v0, 1.0, v0985; GFX9-NEXT:  .LBB3_3: ; %end986; GFX9-NEXT:    v_mov_b32_e32 v18, v32987; GFX9-NEXT:    s_setpc_b64 s[30:31]988; GFX9-NEXT:  .LBB3_4:989; GFX9-NEXT:    s_branch .LBB3_2990;991; GFX11-LABEL: bitcast_v32f32_to_v32i32_scalar:992; GFX11:       ; %bb.0:993; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)994; GFX11-NEXT:    v_dual_mov_b32 v16, v14 :: v_dual_mov_b32 v31, v13995; GFX11-NEXT:    v_dual_mov_b32 v30, v12 :: v_dual_mov_b32 v29, v11996; GFX11-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v9997; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)998; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v16999; GFX11-NEXT:    v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v25, v71000; GFX11-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v51001; GFX11-NEXT:    v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v31002; GFX11-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v11003; GFX11-NEXT:    v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s11004; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s31005; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s171006; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s191007; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s211008; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s231009; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s251010; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s271011; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s291012; GFX11-NEXT:    v_mov_b32_e32 v16, s281013; GFX11-NEXT:    s_mov_b32 s0, 01014; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo1015; GFX11-NEXT:    s_cbranch_scc0 .LBB3_31016; GFX11-NEXT:  ; %bb.1: ; %Flow1017; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s01018; GFX11-NEXT:    s_cbranch_vccz .LBB3_41019; GFX11-NEXT:  ; %bb.2: ; %end1020; GFX11-NEXT:    s_setpc_b64 s[30:31]1021; GFX11-NEXT:  .LBB3_3:1022; GFX11-NEXT:  .LBB3_4: ; %cmp.true1023; GFX11-NEXT:    v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v301024; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v281025; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v261026; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v241027; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v221028; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v201029; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v181030; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v161031; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v141032; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v121033; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v101034; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v81035; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v61036; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v41037; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v21038; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v01039; GFX11-NEXT:    s_setpc_b64 s[30:31]1040  %cmp = icmp eq i32 %b, 01041  br i1 %cmp, label %cmp.true, label %cmp.false1042 1043cmp.true:1044  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)1045  %a2 = bitcast <32 x float> %a1 to <32 x i32>1046  br label %end1047 1048cmp.false:1049  %a3 = bitcast <32 x float> %a to <32 x i32>1050  br label %end1051 1052end:1053  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]1054  ret <32 x i32> %phi1055}1056 1057define <16 x i64> @bitcast_v32i32_to_v16i64(<32 x i32> %a, i32 %b) {1058; SI-LABEL: bitcast_v32i32_to_v16i64:1059; SI:       ; %bb.0:1060; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1061; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:41062; SI-NEXT:    s_waitcnt vmcnt(0)1063; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v311064; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s321065; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc1066; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]1067; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]1068; SI-NEXT:    s_cbranch_execz .LBB4_21069; SI-NEXT:  ; %bb.1: ; %cmp.true1070; SI-NEXT:    s_waitcnt vmcnt(0)1071; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v311072; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v301073; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v291074; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v281075; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v271076; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v261077; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v251078; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v241079; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v231080; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v221081; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v211082; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v201083; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v191084; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v181085; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v171086; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v161087; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v151088; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v141089; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v131090; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v121091; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v111092; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v101093; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v91094; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v81095; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v71096; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v61097; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v51098; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v41099; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v31100; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v21101; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v11102; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v01103; SI-NEXT:  .LBB4_2: ; %end1104; SI-NEXT:    s_or_b64 exec, exec, s[4:5]1105; SI-NEXT:    s_waitcnt vmcnt(0)1106; SI-NEXT:    s_setpc_b64 s[30:31]1107;1108; VI-LABEL: bitcast_v32i32_to_v16i64:1109; VI:       ; %bb.0:1110; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1111; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:41112; VI-NEXT:    s_waitcnt vmcnt(0)1113; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v311114; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s321115; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc1116; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]1117; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]1118; VI-NEXT:    s_cbranch_execz .LBB4_21119; VI-NEXT:  ; %bb.1: ; %cmp.true1120; VI-NEXT:    s_waitcnt vmcnt(0)1121; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v311122; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v301123; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v291124; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v281125; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v271126; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v261127; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v251128; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v241129; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v231130; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v221131; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v211132; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v201133; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v191134; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v181135; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v171136; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v161137; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v151138; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v141139; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v131140; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v121141; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v111142; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v101143; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v91144; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v81145; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v71146; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v61147; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v51148; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v41149; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v31150; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v21151; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v11152; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v01153; VI-NEXT:  .LBB4_2: ; %end1154; VI-NEXT:    s_or_b64 exec, exec, s[4:5]1155; VI-NEXT:    s_waitcnt vmcnt(0)1156; VI-NEXT:    s_setpc_b64 s[30:31]1157;1158; GFX9-LABEL: bitcast_v32i32_to_v16i64:1159; GFX9:       ; %bb.0:1160; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1161; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:41162; GFX9-NEXT:    s_waitcnt vmcnt(0)1163; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v311164; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s321165; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc1166; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]1167; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]1168; GFX9-NEXT:    s_cbranch_execz .LBB4_21169; GFX9-NEXT:  ; %bb.1: ; %cmp.true1170; GFX9-NEXT:    s_waitcnt vmcnt(0)1171; GFX9-NEXT:    v_add_u32_e32 v31, 3, v311172; GFX9-NEXT:    v_add_u32_e32 v30, 3, v301173; GFX9-NEXT:    v_add_u32_e32 v29, 3, v291174; GFX9-NEXT:    v_add_u32_e32 v28, 3, v281175; GFX9-NEXT:    v_add_u32_e32 v27, 3, v271176; GFX9-NEXT:    v_add_u32_e32 v26, 3, v261177; GFX9-NEXT:    v_add_u32_e32 v25, 3, v251178; GFX9-NEXT:    v_add_u32_e32 v24, 3, v241179; GFX9-NEXT:    v_add_u32_e32 v23, 3, v231180; GFX9-NEXT:    v_add_u32_e32 v22, 3, v221181; GFX9-NEXT:    v_add_u32_e32 v21, 3, v211182; GFX9-NEXT:    v_add_u32_e32 v20, 3, v201183; GFX9-NEXT:    v_add_u32_e32 v19, 3, v191184; GFX9-NEXT:    v_add_u32_e32 v18, 3, v181185; GFX9-NEXT:    v_add_u32_e32 v17, 3, v171186; GFX9-NEXT:    v_add_u32_e32 v16, 3, v161187; GFX9-NEXT:    v_add_u32_e32 v15, 3, v151188; GFX9-NEXT:    v_add_u32_e32 v14, 3, v141189; GFX9-NEXT:    v_add_u32_e32 v13, 3, v131190; GFX9-NEXT:    v_add_u32_e32 v12, 3, v121191; GFX9-NEXT:    v_add_u32_e32 v11, 3, v111192; GFX9-NEXT:    v_add_u32_e32 v10, 3, v101193; GFX9-NEXT:    v_add_u32_e32 v9, 3, v91194; GFX9-NEXT:    v_add_u32_e32 v8, 3, v81195; GFX9-NEXT:    v_add_u32_e32 v7, 3, v71196; GFX9-NEXT:    v_add_u32_e32 v6, 3, v61197; GFX9-NEXT:    v_add_u32_e32 v5, 3, v51198; GFX9-NEXT:    v_add_u32_e32 v4, 3, v41199; GFX9-NEXT:    v_add_u32_e32 v3, 3, v31200; GFX9-NEXT:    v_add_u32_e32 v2, 3, v21201; GFX9-NEXT:    v_add_u32_e32 v1, 3, v11202; GFX9-NEXT:    v_add_u32_e32 v0, 3, v01203; GFX9-NEXT:  .LBB4_2: ; %end1204; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]1205; GFX9-NEXT:    s_waitcnt vmcnt(0)1206; GFX9-NEXT:    s_setpc_b64 s[30:31]1207;1208; GFX11-LABEL: bitcast_v32i32_to_v16i64:1209; GFX11:       ; %bb.0:1210; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1211; GFX11-NEXT:    s_clause 0x11212; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:41213; GFX11-NEXT:    scratch_load_b32 v31, off, s321214; GFX11-NEXT:    s_mov_b32 s0, exec_lo1215; GFX11-NEXT:    s_waitcnt vmcnt(1)1216; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v321217; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s01218; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)1219; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s01220; GFX11-NEXT:    s_cbranch_execz .LBB4_21221; GFX11-NEXT:  ; %bb.1: ; %cmp.true1222; GFX11-NEXT:    s_waitcnt vmcnt(0)1223; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v311224; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v301225; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v291226; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v281227; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v271228; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v261229; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v251230; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v241231; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v231232; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v221233; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v211234; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v201235; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v191236; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v181237; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v171238; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v161239; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v151240; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v141241; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v131242; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v121243; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v111244; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v101245; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v91246; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v81247; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v71248; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v61249; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v51250; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v41251; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v31252; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v21253; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v11254; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v01255; GFX11-NEXT:  .LBB4_2: ; %end1256; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s01257; GFX11-NEXT:    s_waitcnt vmcnt(0)1258; GFX11-NEXT:    s_setpc_b64 s[30:31]1259  %cmp = icmp eq i32 %b, 01260  br i1 %cmp, label %cmp.true, label %cmp.false1261 1262cmp.true:1263  %a1 = add <32 x i32> %a, splat (i32 3)1264  %a2 = bitcast <32 x i32> %a1 to <16 x i64>1265  br label %end1266 1267cmp.false:1268  %a3 = bitcast <32 x i32> %a to <16 x i64>1269  br label %end1270 1271end:1272  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]1273  ret <16 x i64> %phi1274}1275 1276define inreg <16 x i64> @bitcast_v32i32_to_v16i64_scalar(<32 x i32> inreg %a, i32 inreg %b) {1277; SI-LABEL: bitcast_v32i32_to_v16i64_scalar:1278; SI:       ; %bb.0:1279; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1280; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v181281; SI-NEXT:    v_mov_b32_e32 v31, v171282; SI-NEXT:    v_mov_b32_e32 v30, v161283; SI-NEXT:    v_mov_b32_e32 v29, v151284; SI-NEXT:    v_mov_b32_e32 v28, v141285; SI-NEXT:    v_mov_b32_e32 v27, v131286; SI-NEXT:    v_mov_b32_e32 v26, v121287; SI-NEXT:    v_mov_b32_e32 v25, v111288; SI-NEXT:    v_mov_b32_e32 v24, v101289; SI-NEXT:    v_mov_b32_e32 v23, v91290; SI-NEXT:    v_mov_b32_e32 v22, v81291; SI-NEXT:    v_mov_b32_e32 v21, v71292; SI-NEXT:    v_mov_b32_e32 v20, v61293; SI-NEXT:    v_mov_b32_e32 v19, v51294; SI-NEXT:    v_mov_b32_e32 v32, v41295; SI-NEXT:    v_mov_b32_e32 v17, v31296; SI-NEXT:    v_mov_b32_e32 v16, v21297; SI-NEXT:    v_mov_b32_e32 v15, v11298; SI-NEXT:    v_mov_b32_e32 v14, v01299; SI-NEXT:    v_mov_b32_e32 v0, s161300; SI-NEXT:    v_mov_b32_e32 v1, s171301; SI-NEXT:    s_and_b64 s[4:5], vcc, exec1302; SI-NEXT:    v_mov_b32_e32 v2, s181303; SI-NEXT:    v_mov_b32_e32 v3, s191304; SI-NEXT:    v_mov_b32_e32 v4, s201305; SI-NEXT:    v_mov_b32_e32 v5, s211306; SI-NEXT:    v_mov_b32_e32 v6, s221307; SI-NEXT:    v_mov_b32_e32 v7, s231308; SI-NEXT:    v_mov_b32_e32 v8, s241309; SI-NEXT:    v_mov_b32_e32 v9, s251310; SI-NEXT:    v_mov_b32_e32 v10, s261311; SI-NEXT:    v_mov_b32_e32 v11, s271312; SI-NEXT:    v_mov_b32_e32 v12, s281313; SI-NEXT:    v_mov_b32_e32 v13, s291314; SI-NEXT:    s_cbranch_scc0 .LBB5_41315; SI-NEXT:  ; %bb.1: ; %cmp.false1316; SI-NEXT:    s_cbranch_execnz .LBB5_31317; SI-NEXT:  .LBB5_2: ; %cmp.true1318; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v311319; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v301320; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v291321; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v281322; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v271323; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v261324; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v251325; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v241326; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v231327; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v221328; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v211329; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v201330; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v191331; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v321332; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v171333; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v161334; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v151335; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v141336; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v131337; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v121338; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v111339; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v101340; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v91341; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v81342; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v71343; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v61344; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v51345; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v41346; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v31347; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v21348; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v11349; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v01350; SI-NEXT:  .LBB5_3: ; %end1351; SI-NEXT:    v_mov_b32_e32 v18, v321352; SI-NEXT:    s_setpc_b64 s[30:31]1353; SI-NEXT:  .LBB5_4:1354; SI-NEXT:    s_branch .LBB5_21355;1356; VI-LABEL: bitcast_v32i32_to_v16i64_scalar:1357; VI:       ; %bb.0:1358; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1359; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v181360; VI-NEXT:    v_mov_b32_e32 v31, v171361; VI-NEXT:    v_mov_b32_e32 v30, v161362; VI-NEXT:    v_mov_b32_e32 v29, v151363; VI-NEXT:    v_mov_b32_e32 v28, v141364; VI-NEXT:    v_mov_b32_e32 v27, v131365; VI-NEXT:    v_mov_b32_e32 v26, v121366; VI-NEXT:    v_mov_b32_e32 v25, v111367; VI-NEXT:    v_mov_b32_e32 v24, v101368; VI-NEXT:    v_mov_b32_e32 v23, v91369; VI-NEXT:    v_mov_b32_e32 v22, v81370; VI-NEXT:    v_mov_b32_e32 v21, v71371; VI-NEXT:    v_mov_b32_e32 v20, v61372; VI-NEXT:    v_mov_b32_e32 v19, v51373; VI-NEXT:    v_mov_b32_e32 v32, v41374; VI-NEXT:    v_mov_b32_e32 v17, v31375; VI-NEXT:    v_mov_b32_e32 v16, v21376; VI-NEXT:    v_mov_b32_e32 v15, v11377; VI-NEXT:    v_mov_b32_e32 v14, v01378; VI-NEXT:    v_mov_b32_e32 v0, s161379; VI-NEXT:    v_mov_b32_e32 v1, s171380; VI-NEXT:    s_and_b64 s[4:5], vcc, exec1381; VI-NEXT:    v_mov_b32_e32 v2, s181382; VI-NEXT:    v_mov_b32_e32 v3, s191383; VI-NEXT:    v_mov_b32_e32 v4, s201384; VI-NEXT:    v_mov_b32_e32 v5, s211385; VI-NEXT:    v_mov_b32_e32 v6, s221386; VI-NEXT:    v_mov_b32_e32 v7, s231387; VI-NEXT:    v_mov_b32_e32 v8, s241388; VI-NEXT:    v_mov_b32_e32 v9, s251389; VI-NEXT:    v_mov_b32_e32 v10, s261390; VI-NEXT:    v_mov_b32_e32 v11, s271391; VI-NEXT:    v_mov_b32_e32 v12, s281392; VI-NEXT:    v_mov_b32_e32 v13, s291393; VI-NEXT:    s_cbranch_scc0 .LBB5_41394; VI-NEXT:  ; %bb.1: ; %cmp.false1395; VI-NEXT:    s_cbranch_execnz .LBB5_31396; VI-NEXT:  .LBB5_2: ; %cmp.true1397; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v311398; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v301399; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v291400; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v281401; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v271402; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v261403; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v251404; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v241405; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v231406; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v221407; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v211408; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v201409; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v191410; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v321411; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v171412; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v161413; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v151414; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v141415; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v131416; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v121417; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v111418; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v101419; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v91420; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v81421; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v71422; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v61423; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v51424; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v41425; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v31426; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v21427; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v11428; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v01429; VI-NEXT:  .LBB5_3: ; %end1430; VI-NEXT:    v_mov_b32_e32 v18, v321431; VI-NEXT:    s_setpc_b64 s[30:31]1432; VI-NEXT:  .LBB5_4:1433; VI-NEXT:    s_branch .LBB5_21434;1435; GFX9-LABEL: bitcast_v32i32_to_v16i64_scalar:1436; GFX9:       ; %bb.0:1437; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1438; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v181439; GFX9-NEXT:    v_mov_b32_e32 v31, v171440; GFX9-NEXT:    v_mov_b32_e32 v30, v161441; GFX9-NEXT:    v_mov_b32_e32 v29, v151442; GFX9-NEXT:    v_mov_b32_e32 v28, v141443; GFX9-NEXT:    v_mov_b32_e32 v27, v131444; GFX9-NEXT:    v_mov_b32_e32 v26, v121445; GFX9-NEXT:    v_mov_b32_e32 v25, v111446; GFX9-NEXT:    v_mov_b32_e32 v24, v101447; GFX9-NEXT:    v_mov_b32_e32 v23, v91448; GFX9-NEXT:    v_mov_b32_e32 v22, v81449; GFX9-NEXT:    v_mov_b32_e32 v21, v71450; GFX9-NEXT:    v_mov_b32_e32 v20, v61451; GFX9-NEXT:    v_mov_b32_e32 v19, v51452; GFX9-NEXT:    v_mov_b32_e32 v32, v41453; GFX9-NEXT:    v_mov_b32_e32 v17, v31454; GFX9-NEXT:    v_mov_b32_e32 v16, v21455; GFX9-NEXT:    v_mov_b32_e32 v15, v11456; GFX9-NEXT:    v_mov_b32_e32 v14, v01457; GFX9-NEXT:    v_mov_b32_e32 v0, s161458; GFX9-NEXT:    v_mov_b32_e32 v1, s171459; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec1460; GFX9-NEXT:    v_mov_b32_e32 v2, s181461; GFX9-NEXT:    v_mov_b32_e32 v3, s191462; GFX9-NEXT:    v_mov_b32_e32 v4, s201463; GFX9-NEXT:    v_mov_b32_e32 v5, s211464; GFX9-NEXT:    v_mov_b32_e32 v6, s221465; GFX9-NEXT:    v_mov_b32_e32 v7, s231466; GFX9-NEXT:    v_mov_b32_e32 v8, s241467; GFX9-NEXT:    v_mov_b32_e32 v9, s251468; GFX9-NEXT:    v_mov_b32_e32 v10, s261469; GFX9-NEXT:    v_mov_b32_e32 v11, s271470; GFX9-NEXT:    v_mov_b32_e32 v12, s281471; GFX9-NEXT:    v_mov_b32_e32 v13, s291472; GFX9-NEXT:    s_cbranch_scc0 .LBB5_41473; GFX9-NEXT:  ; %bb.1: ; %cmp.false1474; GFX9-NEXT:    s_cbranch_execnz .LBB5_31475; GFX9-NEXT:  .LBB5_2: ; %cmp.true1476; GFX9-NEXT:    v_add_u32_e32 v31, 3, v311477; GFX9-NEXT:    v_add_u32_e32 v30, 3, v301478; GFX9-NEXT:    v_add_u32_e32 v29, 3, v291479; GFX9-NEXT:    v_add_u32_e32 v28, 3, v281480; GFX9-NEXT:    v_add_u32_e32 v27, 3, v271481; GFX9-NEXT:    v_add_u32_e32 v26, 3, v261482; GFX9-NEXT:    v_add_u32_e32 v25, 3, v251483; GFX9-NEXT:    v_add_u32_e32 v24, 3, v241484; GFX9-NEXT:    v_add_u32_e32 v23, 3, v231485; GFX9-NEXT:    v_add_u32_e32 v22, 3, v221486; GFX9-NEXT:    v_add_u32_e32 v21, 3, v211487; GFX9-NEXT:    v_add_u32_e32 v20, 3, v201488; GFX9-NEXT:    v_add_u32_e32 v19, 3, v191489; GFX9-NEXT:    v_add_u32_e32 v32, 3, v321490; GFX9-NEXT:    v_add_u32_e32 v17, 3, v171491; GFX9-NEXT:    v_add_u32_e32 v16, 3, v161492; GFX9-NEXT:    v_add_u32_e32 v15, 3, v151493; GFX9-NEXT:    v_add_u32_e32 v14, 3, v141494; GFX9-NEXT:    v_add_u32_e32 v13, 3, v131495; GFX9-NEXT:    v_add_u32_e32 v12, 3, v121496; GFX9-NEXT:    v_add_u32_e32 v11, 3, v111497; GFX9-NEXT:    v_add_u32_e32 v10, 3, v101498; GFX9-NEXT:    v_add_u32_e32 v9, 3, v91499; GFX9-NEXT:    v_add_u32_e32 v8, 3, v81500; GFX9-NEXT:    v_add_u32_e32 v7, 3, v71501; GFX9-NEXT:    v_add_u32_e32 v6, 3, v61502; GFX9-NEXT:    v_add_u32_e32 v5, 3, v51503; GFX9-NEXT:    v_add_u32_e32 v4, 3, v41504; GFX9-NEXT:    v_add_u32_e32 v3, 3, v31505; GFX9-NEXT:    v_add_u32_e32 v2, 3, v21506; GFX9-NEXT:    v_add_u32_e32 v1, 3, v11507; GFX9-NEXT:    v_add_u32_e32 v0, 3, v01508; GFX9-NEXT:  .LBB5_3: ; %end1509; GFX9-NEXT:    v_mov_b32_e32 v18, v321510; GFX9-NEXT:    s_setpc_b64 s[30:31]1511; GFX9-NEXT:  .LBB5_4:1512; GFX9-NEXT:    s_branch .LBB5_21513;1514; GFX11-LABEL: bitcast_v32i32_to_v16i64_scalar:1515; GFX11:       ; %bb.0:1516; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1517; GFX11-NEXT:    v_dual_mov_b32 v16, v14 :: v_dual_mov_b32 v31, v131518; GFX11-NEXT:    v_dual_mov_b32 v30, v12 :: v_dual_mov_b32 v29, v111519; GFX11-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v91520; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)1521; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v161522; GFX11-NEXT:    v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v25, v71523; GFX11-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v51524; GFX11-NEXT:    v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v31525; GFX11-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v11526; GFX11-NEXT:    v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s11527; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s31528; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s171529; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s191530; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s211531; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s231532; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s251533; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s271534; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s291535; GFX11-NEXT:    v_mov_b32_e32 v16, s281536; GFX11-NEXT:    s_mov_b32 s0, 01537; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo1538; GFX11-NEXT:    s_cbranch_scc0 .LBB5_31539; GFX11-NEXT:  ; %bb.1: ; %Flow1540; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s01541; GFX11-NEXT:    s_cbranch_vccz .LBB5_41542; GFX11-NEXT:  ; %bb.2: ; %end1543; GFX11-NEXT:    s_setpc_b64 s[30:31]1544; GFX11-NEXT:  .LBB5_3:1545; GFX11-NEXT:  .LBB5_4: ; %cmp.true1546; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v311547; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v301548; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v291549; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v281550; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v271551; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v261552; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v251553; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v241554; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v231555; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v221556; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v211557; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v201558; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v191559; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v181560; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v171561; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v161562; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v151563; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v141564; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v131565; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v121566; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v111567; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v101568; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v91569; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v81570; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v71571; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v61572; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v51573; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v41574; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v31575; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v21576; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v11577; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v01578; GFX11-NEXT:    s_setpc_b64 s[30:31]1579  %cmp = icmp eq i32 %b, 01580  br i1 %cmp, label %cmp.true, label %cmp.false1581 1582cmp.true:1583  %a1 = add <32 x i32> %a, splat (i32 3)1584  %a2 = bitcast <32 x i32> %a1 to <16 x i64>1585  br label %end1586 1587cmp.false:1588  %a3 = bitcast <32 x i32> %a to <16 x i64>1589  br label %end1590 1591end:1592  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]1593  ret <16 x i64> %phi1594}1595 1596define <32 x i32> @bitcast_v16i64_to_v32i32(<16 x i64> %a, i32 %b) {1597; SI-LABEL: bitcast_v16i64_to_v32i32:1598; SI:       ; %bb.0:1599; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1600; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:41601; SI-NEXT:    s_waitcnt vmcnt(0)1602; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v311603; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s321604; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc1605; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]1606; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]1607; SI-NEXT:    s_cbranch_execz .LBB6_21608; SI-NEXT:  ; %bb.1: ; %cmp.true1609; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v301610; SI-NEXT:    s_waitcnt vmcnt(0)1611; SI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc1612; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v281613; SI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc1614; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v261615; SI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc1616; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v241617; SI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc1618; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v221619; SI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc1620; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v201621; SI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc1622; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v181623; SI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc1624; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v161625; SI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc1626; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v141627; SI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc1628; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v121629; SI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc1630; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v101631; SI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc1632; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v81633; SI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc1634; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v61635; SI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc1636; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v41637; SI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc1638; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v21639; SI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc1640; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v01641; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc1642; SI-NEXT:  .LBB6_2: ; %end1643; SI-NEXT:    s_or_b64 exec, exec, s[4:5]1644; SI-NEXT:    s_waitcnt vmcnt(0)1645; SI-NEXT:    s_setpc_b64 s[30:31]1646;1647; VI-LABEL: bitcast_v16i64_to_v32i32:1648; VI:       ; %bb.0:1649; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1650; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:41651; VI-NEXT:    s_waitcnt vmcnt(0)1652; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v311653; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s321654; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc1655; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]1656; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]1657; VI-NEXT:    s_cbranch_execz .LBB6_21658; VI-NEXT:  ; %bb.1: ; %cmp.true1659; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v301660; VI-NEXT:    s_waitcnt vmcnt(0)1661; VI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc1662; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v281663; VI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc1664; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v261665; VI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc1666; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v241667; VI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc1668; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v221669; VI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc1670; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v201671; VI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc1672; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v181673; VI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc1674; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v161675; VI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc1676; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v141677; VI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc1678; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v121679; VI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc1680; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v101681; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc1682; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v81683; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc1684; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v61685; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc1686; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v41687; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc1688; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v21689; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc1690; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v01691; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc1692; VI-NEXT:  .LBB6_2: ; %end1693; VI-NEXT:    s_or_b64 exec, exec, s[4:5]1694; VI-NEXT:    s_waitcnt vmcnt(0)1695; VI-NEXT:    s_setpc_b64 s[30:31]1696;1697; GFX9-LABEL: bitcast_v16i64_to_v32i32:1698; GFX9:       ; %bb.0:1699; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1700; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:41701; GFX9-NEXT:    s_waitcnt vmcnt(0)1702; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v311703; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s321704; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc1705; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]1706; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]1707; GFX9-NEXT:    s_cbranch_execz .LBB6_21708; GFX9-NEXT:  ; %bb.1: ; %cmp.true1709; GFX9-NEXT:    v_add_co_u32_e32 v30, vcc, 3, v301710; GFX9-NEXT:    s_waitcnt vmcnt(0)1711; GFX9-NEXT:    v_addc_co_u32_e32 v31, vcc, 0, v31, vcc1712; GFX9-NEXT:    v_add_co_u32_e32 v28, vcc, 3, v281713; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, 0, v29, vcc1714; GFX9-NEXT:    v_add_co_u32_e32 v26, vcc, 3, v261715; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, 0, v27, vcc1716; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, 3, v241717; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v25, vcc1718; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, 3, v221719; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, 0, v23, vcc1720; GFX9-NEXT:    v_add_co_u32_e32 v20, vcc, 3, v201721; GFX9-NEXT:    v_addc_co_u32_e32 v21, vcc, 0, v21, vcc1722; GFX9-NEXT:    v_add_co_u32_e32 v18, vcc, 3, v181723; GFX9-NEXT:    v_addc_co_u32_e32 v19, vcc, 0, v19, vcc1724; GFX9-NEXT:    v_add_co_u32_e32 v16, vcc, 3, v161725; GFX9-NEXT:    v_addc_co_u32_e32 v17, vcc, 0, v17, vcc1726; GFX9-NEXT:    v_add_co_u32_e32 v14, vcc, 3, v141727; GFX9-NEXT:    v_addc_co_u32_e32 v15, vcc, 0, v15, vcc1728; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, 3, v121729; GFX9-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v13, vcc1730; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, 3, v101731; GFX9-NEXT:    v_addc_co_u32_e32 v11, vcc, 0, v11, vcc1732; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, 3, v81733; GFX9-NEXT:    v_addc_co_u32_e32 v9, vcc, 0, v9, vcc1734; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, 3, v61735; GFX9-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v7, vcc1736; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, 3, v41737; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v5, vcc1738; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, 3, v21739; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc1740; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 3, v01741; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc1742; GFX9-NEXT:  .LBB6_2: ; %end1743; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]1744; GFX9-NEXT:    s_waitcnt vmcnt(0)1745; GFX9-NEXT:    s_setpc_b64 s[30:31]1746;1747; GFX11-LABEL: bitcast_v16i64_to_v32i32:1748; GFX11:       ; %bb.0:1749; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1750; GFX11-NEXT:    s_clause 0x11751; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:41752; GFX11-NEXT:    scratch_load_b32 v31, off, s321753; GFX11-NEXT:    s_mov_b32 s0, exec_lo1754; GFX11-NEXT:    s_waitcnt vmcnt(1)1755; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v321756; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s01757; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)1758; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s01759; GFX11-NEXT:    s_cbranch_execz .LBB6_21760; GFX11-NEXT:  ; %bb.1: ; %cmp.true1761; GFX11-NEXT:    v_add_co_u32 v30, vcc_lo, v30, 31762; GFX11-NEXT:    s_waitcnt vmcnt(0)1763; GFX11-NEXT:    v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo1764; GFX11-NEXT:    v_add_co_u32 v28, vcc_lo, v28, 31765; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)1766; GFX11-NEXT:    v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo1767; GFX11-NEXT:    v_add_co_u32 v26, vcc_lo, v26, 31768; GFX11-NEXT:    v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo1769; GFX11-NEXT:    v_add_co_u32 v24, vcc_lo, v24, 31770; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)1771; GFX11-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo1772; GFX11-NEXT:    v_add_co_u32 v22, vcc_lo, v22, 31773; GFX11-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo1774; GFX11-NEXT:    v_add_co_u32 v20, vcc_lo, v20, 31775; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)1776; GFX11-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo1777; GFX11-NEXT:    v_add_co_u32 v18, vcc_lo, v18, 31778; GFX11-NEXT:    v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo1779; GFX11-NEXT:    v_add_co_u32 v16, vcc_lo, v16, 31780; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)1781; GFX11-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo1782; GFX11-NEXT:    v_add_co_u32 v14, vcc_lo, v14, 31783; GFX11-NEXT:    v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo1784; GFX11-NEXT:    v_add_co_u32 v12, vcc_lo, v12, 31785; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)1786; GFX11-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo1787; GFX11-NEXT:    v_add_co_u32 v10, vcc_lo, v10, 31788; GFX11-NEXT:    v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo1789; GFX11-NEXT:    v_add_co_u32 v8, vcc_lo, v8, 31790; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)1791; GFX11-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo1792; GFX11-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 31793; GFX11-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo1794; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 31795; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)1796; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo1797; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, 31798; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo1799; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, 31800; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)1801; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo1802; GFX11-NEXT:  .LBB6_2: ; %end1803; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s01804; GFX11-NEXT:    s_waitcnt vmcnt(0)1805; GFX11-NEXT:    s_setpc_b64 s[30:31]1806  %cmp = icmp eq i32 %b, 01807  br i1 %cmp, label %cmp.true, label %cmp.false1808 1809cmp.true:1810  %a1 = add <16 x i64> %a, splat (i64 3)1811  %a2 = bitcast <16 x i64> %a1 to <32 x i32>1812  br label %end1813 1814cmp.false:1815  %a3 = bitcast <16 x i64> %a to <32 x i32>1816  br label %end1817 1818end:1819  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]1820  ret <32 x i32> %phi1821}1822 1823define inreg <32 x i32> @bitcast_v16i64_to_v32i32_scalar(<16 x i64> inreg %a, i32 inreg %b) {1824; SI-LABEL: bitcast_v16i64_to_v32i32_scalar:1825; SI:       ; %bb.0:1826; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1827; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v181828; SI-NEXT:    v_mov_b32_e32 v31, v171829; SI-NEXT:    v_mov_b32_e32 v30, v161830; SI-NEXT:    v_mov_b32_e32 v29, v151831; SI-NEXT:    v_mov_b32_e32 v28, v141832; SI-NEXT:    v_mov_b32_e32 v27, v131833; SI-NEXT:    v_mov_b32_e32 v26, v121834; SI-NEXT:    v_mov_b32_e32 v25, v111835; SI-NEXT:    v_mov_b32_e32 v24, v101836; SI-NEXT:    v_mov_b32_e32 v23, v91837; SI-NEXT:    v_mov_b32_e32 v22, v81838; SI-NEXT:    v_mov_b32_e32 v21, v71839; SI-NEXT:    v_mov_b32_e32 v20, v61840; SI-NEXT:    v_mov_b32_e32 v19, v51841; SI-NEXT:    v_mov_b32_e32 v32, v41842; SI-NEXT:    v_mov_b32_e32 v17, v31843; SI-NEXT:    v_mov_b32_e32 v16, v21844; SI-NEXT:    v_mov_b32_e32 v15, v11845; SI-NEXT:    v_mov_b32_e32 v14, v01846; SI-NEXT:    v_mov_b32_e32 v0, s161847; SI-NEXT:    v_mov_b32_e32 v1, s171848; SI-NEXT:    s_and_b64 s[4:5], vcc, exec1849; SI-NEXT:    v_mov_b32_e32 v2, s181850; SI-NEXT:    v_mov_b32_e32 v3, s191851; SI-NEXT:    v_mov_b32_e32 v4, s201852; SI-NEXT:    v_mov_b32_e32 v5, s211853; SI-NEXT:    v_mov_b32_e32 v6, s221854; SI-NEXT:    v_mov_b32_e32 v7, s231855; SI-NEXT:    v_mov_b32_e32 v8, s241856; SI-NEXT:    v_mov_b32_e32 v9, s251857; SI-NEXT:    v_mov_b32_e32 v10, s261858; SI-NEXT:    v_mov_b32_e32 v11, s271859; SI-NEXT:    v_mov_b32_e32 v12, s281860; SI-NEXT:    v_mov_b32_e32 v13, s291861; SI-NEXT:    s_cbranch_scc0 .LBB7_41862; SI-NEXT:  ; %bb.1: ; %cmp.false1863; SI-NEXT:    s_cbranch_execnz .LBB7_31864; SI-NEXT:  .LBB7_2: ; %cmp.true1865; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v301866; SI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc1867; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v281868; SI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc1869; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v261870; SI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc1871; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v241872; SI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc1873; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v221874; SI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc1875; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v201876; SI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc1877; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v321878; SI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc1879; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v161880; SI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc1881; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v141882; SI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc1883; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v121884; SI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc1885; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v101886; SI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc1887; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v81888; SI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc1889; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v61890; SI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc1891; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v41892; SI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc1893; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v21894; SI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc1895; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v01896; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc1897; SI-NEXT:  .LBB7_3: ; %end1898; SI-NEXT:    v_mov_b32_e32 v18, v321899; SI-NEXT:    s_setpc_b64 s[30:31]1900; SI-NEXT:  .LBB7_4:1901; SI-NEXT:    s_branch .LBB7_21902;1903; VI-LABEL: bitcast_v16i64_to_v32i32_scalar:1904; VI:       ; %bb.0:1905; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1906; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v181907; VI-NEXT:    v_mov_b32_e32 v31, v171908; VI-NEXT:    v_mov_b32_e32 v30, v161909; VI-NEXT:    v_mov_b32_e32 v29, v151910; VI-NEXT:    v_mov_b32_e32 v28, v141911; VI-NEXT:    v_mov_b32_e32 v27, v131912; VI-NEXT:    v_mov_b32_e32 v26, v121913; VI-NEXT:    v_mov_b32_e32 v25, v111914; VI-NEXT:    v_mov_b32_e32 v24, v101915; VI-NEXT:    v_mov_b32_e32 v23, v91916; VI-NEXT:    v_mov_b32_e32 v22, v81917; VI-NEXT:    v_mov_b32_e32 v21, v71918; VI-NEXT:    v_mov_b32_e32 v20, v61919; VI-NEXT:    v_mov_b32_e32 v19, v51920; VI-NEXT:    v_mov_b32_e32 v32, v41921; VI-NEXT:    v_mov_b32_e32 v17, v31922; VI-NEXT:    v_mov_b32_e32 v16, v21923; VI-NEXT:    v_mov_b32_e32 v15, v11924; VI-NEXT:    v_mov_b32_e32 v14, v01925; VI-NEXT:    v_mov_b32_e32 v0, s161926; VI-NEXT:    v_mov_b32_e32 v1, s171927; VI-NEXT:    s_and_b64 s[4:5], vcc, exec1928; VI-NEXT:    v_mov_b32_e32 v2, s181929; VI-NEXT:    v_mov_b32_e32 v3, s191930; VI-NEXT:    v_mov_b32_e32 v4, s201931; VI-NEXT:    v_mov_b32_e32 v5, s211932; VI-NEXT:    v_mov_b32_e32 v6, s221933; VI-NEXT:    v_mov_b32_e32 v7, s231934; VI-NEXT:    v_mov_b32_e32 v8, s241935; VI-NEXT:    v_mov_b32_e32 v9, s251936; VI-NEXT:    v_mov_b32_e32 v10, s261937; VI-NEXT:    v_mov_b32_e32 v11, s271938; VI-NEXT:    v_mov_b32_e32 v12, s281939; VI-NEXT:    v_mov_b32_e32 v13, s291940; VI-NEXT:    s_cbranch_scc0 .LBB7_41941; VI-NEXT:  ; %bb.1: ; %cmp.false1942; VI-NEXT:    s_cbranch_execnz .LBB7_31943; VI-NEXT:  .LBB7_2: ; %cmp.true1944; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v301945; VI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc1946; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v281947; VI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc1948; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v261949; VI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc1950; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v241951; VI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc1952; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v221953; VI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc1954; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v201955; VI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc1956; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v321957; VI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc1958; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v161959; VI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc1960; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v141961; VI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc1962; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v121963; VI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc1964; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v101965; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc1966; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v81967; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc1968; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v61969; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc1970; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v41971; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc1972; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v21973; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc1974; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v01975; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc1976; VI-NEXT:  .LBB7_3: ; %end1977; VI-NEXT:    v_mov_b32_e32 v18, v321978; VI-NEXT:    s_setpc_b64 s[30:31]1979; VI-NEXT:  .LBB7_4:1980; VI-NEXT:    s_branch .LBB7_21981;1982; GFX9-LABEL: bitcast_v16i64_to_v32i32_scalar:1983; GFX9:       ; %bb.0:1984; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1985; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v181986; GFX9-NEXT:    v_mov_b32_e32 v31, v171987; GFX9-NEXT:    v_mov_b32_e32 v30, v161988; GFX9-NEXT:    v_mov_b32_e32 v29, v151989; GFX9-NEXT:    v_mov_b32_e32 v28, v141990; GFX9-NEXT:    v_mov_b32_e32 v27, v131991; GFX9-NEXT:    v_mov_b32_e32 v26, v121992; GFX9-NEXT:    v_mov_b32_e32 v25, v111993; GFX9-NEXT:    v_mov_b32_e32 v24, v101994; GFX9-NEXT:    v_mov_b32_e32 v23, v91995; GFX9-NEXT:    v_mov_b32_e32 v22, v81996; GFX9-NEXT:    v_mov_b32_e32 v21, v71997; GFX9-NEXT:    v_mov_b32_e32 v20, v61998; GFX9-NEXT:    v_mov_b32_e32 v19, v51999; GFX9-NEXT:    v_mov_b32_e32 v32, v42000; GFX9-NEXT:    v_mov_b32_e32 v17, v32001; GFX9-NEXT:    v_mov_b32_e32 v16, v22002; GFX9-NEXT:    v_mov_b32_e32 v15, v12003; GFX9-NEXT:    v_mov_b32_e32 v14, v02004; GFX9-NEXT:    v_mov_b32_e32 v0, s162005; GFX9-NEXT:    v_mov_b32_e32 v1, s172006; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec2007; GFX9-NEXT:    v_mov_b32_e32 v2, s182008; GFX9-NEXT:    v_mov_b32_e32 v3, s192009; GFX9-NEXT:    v_mov_b32_e32 v4, s202010; GFX9-NEXT:    v_mov_b32_e32 v5, s212011; GFX9-NEXT:    v_mov_b32_e32 v6, s222012; GFX9-NEXT:    v_mov_b32_e32 v7, s232013; GFX9-NEXT:    v_mov_b32_e32 v8, s242014; GFX9-NEXT:    v_mov_b32_e32 v9, s252015; GFX9-NEXT:    v_mov_b32_e32 v10, s262016; GFX9-NEXT:    v_mov_b32_e32 v11, s272017; GFX9-NEXT:    v_mov_b32_e32 v12, s282018; GFX9-NEXT:    v_mov_b32_e32 v13, s292019; GFX9-NEXT:    s_cbranch_scc0 .LBB7_42020; GFX9-NEXT:  ; %bb.1: ; %cmp.false2021; GFX9-NEXT:    s_cbranch_execnz .LBB7_32022; GFX9-NEXT:  .LBB7_2: ; %cmp.true2023; GFX9-NEXT:    v_add_co_u32_e32 v30, vcc, 3, v302024; GFX9-NEXT:    v_addc_co_u32_e32 v31, vcc, 0, v31, vcc2025; GFX9-NEXT:    v_add_co_u32_e32 v28, vcc, 3, v282026; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, 0, v29, vcc2027; GFX9-NEXT:    v_add_co_u32_e32 v26, vcc, 3, v262028; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, 0, v27, vcc2029; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, 3, v242030; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v25, vcc2031; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, 3, v222032; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, 0, v23, vcc2033; GFX9-NEXT:    v_add_co_u32_e32 v20, vcc, 3, v202034; GFX9-NEXT:    v_addc_co_u32_e32 v21, vcc, 0, v21, vcc2035; GFX9-NEXT:    v_add_co_u32_e32 v32, vcc, 3, v322036; GFX9-NEXT:    v_addc_co_u32_e32 v19, vcc, 0, v19, vcc2037; GFX9-NEXT:    v_add_co_u32_e32 v16, vcc, 3, v162038; GFX9-NEXT:    v_addc_co_u32_e32 v17, vcc, 0, v17, vcc2039; GFX9-NEXT:    v_add_co_u32_e32 v14, vcc, 3, v142040; GFX9-NEXT:    v_addc_co_u32_e32 v15, vcc, 0, v15, vcc2041; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, 3, v122042; GFX9-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v13, vcc2043; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, 3, v102044; GFX9-NEXT:    v_addc_co_u32_e32 v11, vcc, 0, v11, vcc2045; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, 3, v82046; GFX9-NEXT:    v_addc_co_u32_e32 v9, vcc, 0, v9, vcc2047; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, 3, v62048; GFX9-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v7, vcc2049; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, 3, v42050; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v5, vcc2051; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, 3, v22052; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc2053; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 3, v02054; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc2055; GFX9-NEXT:  .LBB7_3: ; %end2056; GFX9-NEXT:    v_mov_b32_e32 v18, v322057; GFX9-NEXT:    s_setpc_b64 s[30:31]2058; GFX9-NEXT:  .LBB7_4:2059; GFX9-NEXT:    s_branch .LBB7_22060;2061; GFX11-LABEL: bitcast_v16i64_to_v32i32_scalar:2062; GFX11:       ; %bb.0:2063; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2064; GFX11-NEXT:    v_dual_mov_b32 v16, v14 :: v_dual_mov_b32 v31, v132065; GFX11-NEXT:    v_dual_mov_b32 v30, v12 :: v_dual_mov_b32 v29, v112066; GFX11-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v92067; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)2068; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v162069; GFX11-NEXT:    v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v25, v72070; GFX11-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v52071; GFX11-NEXT:    v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v32072; GFX11-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v12073; GFX11-NEXT:    v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s12074; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s32075; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s172076; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s192077; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s212078; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s232079; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s252080; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s272081; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s292082; GFX11-NEXT:    v_mov_b32_e32 v16, s282083; GFX11-NEXT:    s_mov_b32 s0, 02084; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo2085; GFX11-NEXT:    s_cbranch_scc0 .LBB7_32086; GFX11-NEXT:  ; %bb.1: ; %Flow2087; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s02088; GFX11-NEXT:    s_cbranch_vccz .LBB7_42089; GFX11-NEXT:  ; %bb.2: ; %end2090; GFX11-NEXT:    s_setpc_b64 s[30:31]2091; GFX11-NEXT:  .LBB7_3:2092; GFX11-NEXT:  .LBB7_4: ; %cmp.true2093; GFX11-NEXT:    v_add_co_u32 v30, vcc_lo, v30, 32094; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)2095; GFX11-NEXT:    v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo2096; GFX11-NEXT:    v_add_co_u32 v28, vcc_lo, v28, 32097; GFX11-NEXT:    v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo2098; GFX11-NEXT:    v_add_co_u32 v26, vcc_lo, v26, 32099; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)2100; GFX11-NEXT:    v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo2101; GFX11-NEXT:    v_add_co_u32 v24, vcc_lo, v24, 32102; GFX11-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo2103; GFX11-NEXT:    v_add_co_u32 v22, vcc_lo, v22, 32104; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)2105; GFX11-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo2106; GFX11-NEXT:    v_add_co_u32 v20, vcc_lo, v20, 32107; GFX11-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo2108; GFX11-NEXT:    v_add_co_u32 v18, vcc_lo, v18, 32109; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)2110; GFX11-NEXT:    v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo2111; GFX11-NEXT:    v_add_co_u32 v16, vcc_lo, v16, 32112; GFX11-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo2113; GFX11-NEXT:    v_add_co_u32 v14, vcc_lo, v14, 32114; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)2115; GFX11-NEXT:    v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo2116; GFX11-NEXT:    v_add_co_u32 v12, vcc_lo, v12, 32117; GFX11-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo2118; GFX11-NEXT:    v_add_co_u32 v10, vcc_lo, v10, 32119; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)2120; GFX11-NEXT:    v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo2121; GFX11-NEXT:    v_add_co_u32 v8, vcc_lo, v8, 32122; GFX11-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo2123; GFX11-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 32124; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)2125; GFX11-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo2126; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 32127; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo2128; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, 32129; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)2130; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo2131; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, 32132; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo2133; GFX11-NEXT:    s_setpc_b64 s[30:31]2134  %cmp = icmp eq i32 %b, 02135  br i1 %cmp, label %cmp.true, label %cmp.false2136 2137cmp.true:2138  %a1 = add <16 x i64> %a, splat (i64 3)2139  %a2 = bitcast <16 x i64> %a1 to <32 x i32>2140  br label %end2141 2142cmp.false:2143  %a3 = bitcast <16 x i64> %a to <32 x i32>2144  br label %end2145 2146end:2147  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]2148  ret <32 x i32> %phi2149}2150 2151define <16 x double> @bitcast_v32i32_to_v16f64(<32 x i32> %a, i32 %b) {2152; SI-LABEL: bitcast_v32i32_to_v16f64:2153; SI:       ; %bb.0:2154; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2155; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:42156; SI-NEXT:    s_waitcnt vmcnt(0)2157; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v312158; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s322159; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc2160; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]2161; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]2162; SI-NEXT:    s_cbranch_execz .LBB8_22163; SI-NEXT:  ; %bb.1: ; %cmp.true2164; SI-NEXT:    s_waitcnt vmcnt(0)2165; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v312166; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v302167; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v292168; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v282169; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v272170; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v262171; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v252172; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v242173; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v232174; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v222175; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v212176; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v202177; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v192178; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v182179; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v172180; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v162181; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v152182; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v142183; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v132184; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v122185; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v112186; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v102187; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v92188; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v82189; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v72190; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v62191; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v52192; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v42193; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v32194; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v22195; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v12196; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v02197; SI-NEXT:  .LBB8_2: ; %end2198; SI-NEXT:    s_or_b64 exec, exec, s[4:5]2199; SI-NEXT:    s_waitcnt vmcnt(0)2200; SI-NEXT:    s_setpc_b64 s[30:31]2201;2202; VI-LABEL: bitcast_v32i32_to_v16f64:2203; VI:       ; %bb.0:2204; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2205; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:42206; VI-NEXT:    s_waitcnt vmcnt(0)2207; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v312208; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s322209; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc2210; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]2211; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]2212; VI-NEXT:    s_cbranch_execz .LBB8_22213; VI-NEXT:  ; %bb.1: ; %cmp.true2214; VI-NEXT:    s_waitcnt vmcnt(0)2215; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v312216; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v302217; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v292218; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v282219; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v272220; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v262221; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v252222; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v242223; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v232224; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v222225; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v212226; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v202227; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v192228; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v182229; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v172230; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v162231; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v152232; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v142233; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v132234; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v122235; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v112236; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v102237; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v92238; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v82239; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v72240; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v62241; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v52242; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v42243; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v32244; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v22245; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v12246; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v02247; VI-NEXT:  .LBB8_2: ; %end2248; VI-NEXT:    s_or_b64 exec, exec, s[4:5]2249; VI-NEXT:    s_waitcnt vmcnt(0)2250; VI-NEXT:    s_setpc_b64 s[30:31]2251;2252; GFX9-LABEL: bitcast_v32i32_to_v16f64:2253; GFX9:       ; %bb.0:2254; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2255; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:42256; GFX9-NEXT:    s_waitcnt vmcnt(0)2257; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v312258; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s322259; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc2260; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]2261; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]2262; GFX9-NEXT:    s_cbranch_execz .LBB8_22263; GFX9-NEXT:  ; %bb.1: ; %cmp.true2264; GFX9-NEXT:    s_waitcnt vmcnt(0)2265; GFX9-NEXT:    v_add_u32_e32 v31, 3, v312266; GFX9-NEXT:    v_add_u32_e32 v30, 3, v302267; GFX9-NEXT:    v_add_u32_e32 v29, 3, v292268; GFX9-NEXT:    v_add_u32_e32 v28, 3, v282269; GFX9-NEXT:    v_add_u32_e32 v27, 3, v272270; GFX9-NEXT:    v_add_u32_e32 v26, 3, v262271; GFX9-NEXT:    v_add_u32_e32 v25, 3, v252272; GFX9-NEXT:    v_add_u32_e32 v24, 3, v242273; GFX9-NEXT:    v_add_u32_e32 v23, 3, v232274; GFX9-NEXT:    v_add_u32_e32 v22, 3, v222275; GFX9-NEXT:    v_add_u32_e32 v21, 3, v212276; GFX9-NEXT:    v_add_u32_e32 v20, 3, v202277; GFX9-NEXT:    v_add_u32_e32 v19, 3, v192278; GFX9-NEXT:    v_add_u32_e32 v18, 3, v182279; GFX9-NEXT:    v_add_u32_e32 v17, 3, v172280; GFX9-NEXT:    v_add_u32_e32 v16, 3, v162281; GFX9-NEXT:    v_add_u32_e32 v15, 3, v152282; GFX9-NEXT:    v_add_u32_e32 v14, 3, v142283; GFX9-NEXT:    v_add_u32_e32 v13, 3, v132284; GFX9-NEXT:    v_add_u32_e32 v12, 3, v122285; GFX9-NEXT:    v_add_u32_e32 v11, 3, v112286; GFX9-NEXT:    v_add_u32_e32 v10, 3, v102287; GFX9-NEXT:    v_add_u32_e32 v9, 3, v92288; GFX9-NEXT:    v_add_u32_e32 v8, 3, v82289; GFX9-NEXT:    v_add_u32_e32 v7, 3, v72290; GFX9-NEXT:    v_add_u32_e32 v6, 3, v62291; GFX9-NEXT:    v_add_u32_e32 v5, 3, v52292; GFX9-NEXT:    v_add_u32_e32 v4, 3, v42293; GFX9-NEXT:    v_add_u32_e32 v3, 3, v32294; GFX9-NEXT:    v_add_u32_e32 v2, 3, v22295; GFX9-NEXT:    v_add_u32_e32 v1, 3, v12296; GFX9-NEXT:    v_add_u32_e32 v0, 3, v02297; GFX9-NEXT:  .LBB8_2: ; %end2298; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]2299; GFX9-NEXT:    s_waitcnt vmcnt(0)2300; GFX9-NEXT:    s_setpc_b64 s[30:31]2301;2302; GFX11-LABEL: bitcast_v32i32_to_v16f64:2303; GFX11:       ; %bb.0:2304; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2305; GFX11-NEXT:    s_clause 0x12306; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:42307; GFX11-NEXT:    scratch_load_b32 v31, off, s322308; GFX11-NEXT:    s_mov_b32 s0, exec_lo2309; GFX11-NEXT:    s_waitcnt vmcnt(1)2310; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v322311; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s02312; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)2313; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s02314; GFX11-NEXT:    s_cbranch_execz .LBB8_22315; GFX11-NEXT:  ; %bb.1: ; %cmp.true2316; GFX11-NEXT:    s_waitcnt vmcnt(0)2317; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v312318; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v302319; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v292320; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v282321; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v272322; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v262323; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v252324; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v242325; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v232326; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v222327; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v212328; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v202329; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v192330; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v182331; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v172332; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v162333; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v152334; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v142335; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v132336; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v122337; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v112338; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v102339; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v92340; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v82341; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v72342; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v62343; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v52344; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v42345; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v32346; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v22347; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v12348; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v02349; GFX11-NEXT:  .LBB8_2: ; %end2350; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s02351; GFX11-NEXT:    s_waitcnt vmcnt(0)2352; GFX11-NEXT:    s_setpc_b64 s[30:31]2353  %cmp = icmp eq i32 %b, 02354  br i1 %cmp, label %cmp.true, label %cmp.false2355 2356cmp.true:2357  %a1 = add <32 x i32> %a, splat (i32 3)2358  %a2 = bitcast <32 x i32> %a1 to <16 x double>2359  br label %end2360 2361cmp.false:2362  %a3 = bitcast <32 x i32> %a to <16 x double>2363  br label %end2364 2365end:2366  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]2367  ret <16 x double> %phi2368}2369 2370define inreg <16 x double> @bitcast_v32i32_to_v16f64_scalar(<32 x i32> inreg %a, i32 inreg %b) {2371; SI-LABEL: bitcast_v32i32_to_v16f64_scalar:2372; SI:       ; %bb.0:2373; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2374; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v182375; SI-NEXT:    v_mov_b32_e32 v31, v172376; SI-NEXT:    v_mov_b32_e32 v30, v162377; SI-NEXT:    v_mov_b32_e32 v29, v152378; SI-NEXT:    v_mov_b32_e32 v28, v142379; SI-NEXT:    v_mov_b32_e32 v27, v132380; SI-NEXT:    v_mov_b32_e32 v26, v122381; SI-NEXT:    v_mov_b32_e32 v25, v112382; SI-NEXT:    v_mov_b32_e32 v24, v102383; SI-NEXT:    v_mov_b32_e32 v23, v92384; SI-NEXT:    v_mov_b32_e32 v22, v82385; SI-NEXT:    v_mov_b32_e32 v21, v72386; SI-NEXT:    v_mov_b32_e32 v20, v62387; SI-NEXT:    v_mov_b32_e32 v19, v52388; SI-NEXT:    v_mov_b32_e32 v32, v42389; SI-NEXT:    v_mov_b32_e32 v17, v32390; SI-NEXT:    v_mov_b32_e32 v16, v22391; SI-NEXT:    v_mov_b32_e32 v15, v12392; SI-NEXT:    v_mov_b32_e32 v14, v02393; SI-NEXT:    v_mov_b32_e32 v0, s162394; SI-NEXT:    v_mov_b32_e32 v1, s172395; SI-NEXT:    s_and_b64 s[4:5], vcc, exec2396; SI-NEXT:    v_mov_b32_e32 v2, s182397; SI-NEXT:    v_mov_b32_e32 v3, s192398; SI-NEXT:    v_mov_b32_e32 v4, s202399; SI-NEXT:    v_mov_b32_e32 v5, s212400; SI-NEXT:    v_mov_b32_e32 v6, s222401; SI-NEXT:    v_mov_b32_e32 v7, s232402; SI-NEXT:    v_mov_b32_e32 v8, s242403; SI-NEXT:    v_mov_b32_e32 v9, s252404; SI-NEXT:    v_mov_b32_e32 v10, s262405; SI-NEXT:    v_mov_b32_e32 v11, s272406; SI-NEXT:    v_mov_b32_e32 v12, s282407; SI-NEXT:    v_mov_b32_e32 v13, s292408; SI-NEXT:    s_cbranch_scc0 .LBB9_42409; SI-NEXT:  ; %bb.1: ; %cmp.false2410; SI-NEXT:    s_cbranch_execnz .LBB9_32411; SI-NEXT:  .LBB9_2: ; %cmp.true2412; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v312413; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v302414; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v292415; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v282416; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v272417; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v262418; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v252419; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v242420; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v232421; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v222422; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v212423; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v202424; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v192425; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v322426; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v172427; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v162428; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v152429; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v142430; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v132431; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v122432; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v112433; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v102434; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v92435; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v82436; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v72437; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v62438; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v52439; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v42440; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v32441; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v22442; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v12443; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v02444; SI-NEXT:  .LBB9_3: ; %end2445; SI-NEXT:    v_mov_b32_e32 v18, v322446; SI-NEXT:    s_setpc_b64 s[30:31]2447; SI-NEXT:  .LBB9_4:2448; SI-NEXT:    s_branch .LBB9_22449;2450; VI-LABEL: bitcast_v32i32_to_v16f64_scalar:2451; VI:       ; %bb.0:2452; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2453; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v182454; VI-NEXT:    v_mov_b32_e32 v31, v172455; VI-NEXT:    v_mov_b32_e32 v30, v162456; VI-NEXT:    v_mov_b32_e32 v29, v152457; VI-NEXT:    v_mov_b32_e32 v28, v142458; VI-NEXT:    v_mov_b32_e32 v27, v132459; VI-NEXT:    v_mov_b32_e32 v26, v122460; VI-NEXT:    v_mov_b32_e32 v25, v112461; VI-NEXT:    v_mov_b32_e32 v24, v102462; VI-NEXT:    v_mov_b32_e32 v23, v92463; VI-NEXT:    v_mov_b32_e32 v22, v82464; VI-NEXT:    v_mov_b32_e32 v21, v72465; VI-NEXT:    v_mov_b32_e32 v20, v62466; VI-NEXT:    v_mov_b32_e32 v19, v52467; VI-NEXT:    v_mov_b32_e32 v32, v42468; VI-NEXT:    v_mov_b32_e32 v17, v32469; VI-NEXT:    v_mov_b32_e32 v16, v22470; VI-NEXT:    v_mov_b32_e32 v15, v12471; VI-NEXT:    v_mov_b32_e32 v14, v02472; VI-NEXT:    v_mov_b32_e32 v0, s162473; VI-NEXT:    v_mov_b32_e32 v1, s172474; VI-NEXT:    s_and_b64 s[4:5], vcc, exec2475; VI-NEXT:    v_mov_b32_e32 v2, s182476; VI-NEXT:    v_mov_b32_e32 v3, s192477; VI-NEXT:    v_mov_b32_e32 v4, s202478; VI-NEXT:    v_mov_b32_e32 v5, s212479; VI-NEXT:    v_mov_b32_e32 v6, s222480; VI-NEXT:    v_mov_b32_e32 v7, s232481; VI-NEXT:    v_mov_b32_e32 v8, s242482; VI-NEXT:    v_mov_b32_e32 v9, s252483; VI-NEXT:    v_mov_b32_e32 v10, s262484; VI-NEXT:    v_mov_b32_e32 v11, s272485; VI-NEXT:    v_mov_b32_e32 v12, s282486; VI-NEXT:    v_mov_b32_e32 v13, s292487; VI-NEXT:    s_cbranch_scc0 .LBB9_42488; VI-NEXT:  ; %bb.1: ; %cmp.false2489; VI-NEXT:    s_cbranch_execnz .LBB9_32490; VI-NEXT:  .LBB9_2: ; %cmp.true2491; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v312492; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v302493; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v292494; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v282495; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v272496; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v262497; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v252498; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v242499; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v232500; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v222501; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v212502; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v202503; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v192504; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v322505; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v172506; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v162507; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v152508; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v142509; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v132510; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v122511; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v112512; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v102513; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v92514; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v82515; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v72516; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v62517; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v52518; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v42519; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v32520; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v22521; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v12522; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v02523; VI-NEXT:  .LBB9_3: ; %end2524; VI-NEXT:    v_mov_b32_e32 v18, v322525; VI-NEXT:    s_setpc_b64 s[30:31]2526; VI-NEXT:  .LBB9_4:2527; VI-NEXT:    s_branch .LBB9_22528;2529; GFX9-LABEL: bitcast_v32i32_to_v16f64_scalar:2530; GFX9:       ; %bb.0:2531; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2532; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v182533; GFX9-NEXT:    v_mov_b32_e32 v31, v172534; GFX9-NEXT:    v_mov_b32_e32 v30, v162535; GFX9-NEXT:    v_mov_b32_e32 v29, v152536; GFX9-NEXT:    v_mov_b32_e32 v28, v142537; GFX9-NEXT:    v_mov_b32_e32 v27, v132538; GFX9-NEXT:    v_mov_b32_e32 v26, v122539; GFX9-NEXT:    v_mov_b32_e32 v25, v112540; GFX9-NEXT:    v_mov_b32_e32 v24, v102541; GFX9-NEXT:    v_mov_b32_e32 v23, v92542; GFX9-NEXT:    v_mov_b32_e32 v22, v82543; GFX9-NEXT:    v_mov_b32_e32 v21, v72544; GFX9-NEXT:    v_mov_b32_e32 v20, v62545; GFX9-NEXT:    v_mov_b32_e32 v19, v52546; GFX9-NEXT:    v_mov_b32_e32 v32, v42547; GFX9-NEXT:    v_mov_b32_e32 v17, v32548; GFX9-NEXT:    v_mov_b32_e32 v16, v22549; GFX9-NEXT:    v_mov_b32_e32 v15, v12550; GFX9-NEXT:    v_mov_b32_e32 v14, v02551; GFX9-NEXT:    v_mov_b32_e32 v0, s162552; GFX9-NEXT:    v_mov_b32_e32 v1, s172553; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec2554; GFX9-NEXT:    v_mov_b32_e32 v2, s182555; GFX9-NEXT:    v_mov_b32_e32 v3, s192556; GFX9-NEXT:    v_mov_b32_e32 v4, s202557; GFX9-NEXT:    v_mov_b32_e32 v5, s212558; GFX9-NEXT:    v_mov_b32_e32 v6, s222559; GFX9-NEXT:    v_mov_b32_e32 v7, s232560; GFX9-NEXT:    v_mov_b32_e32 v8, s242561; GFX9-NEXT:    v_mov_b32_e32 v9, s252562; GFX9-NEXT:    v_mov_b32_e32 v10, s262563; GFX9-NEXT:    v_mov_b32_e32 v11, s272564; GFX9-NEXT:    v_mov_b32_e32 v12, s282565; GFX9-NEXT:    v_mov_b32_e32 v13, s292566; GFX9-NEXT:    s_cbranch_scc0 .LBB9_42567; GFX9-NEXT:  ; %bb.1: ; %cmp.false2568; GFX9-NEXT:    s_cbranch_execnz .LBB9_32569; GFX9-NEXT:  .LBB9_2: ; %cmp.true2570; GFX9-NEXT:    v_add_u32_e32 v31, 3, v312571; GFX9-NEXT:    v_add_u32_e32 v30, 3, v302572; GFX9-NEXT:    v_add_u32_e32 v29, 3, v292573; GFX9-NEXT:    v_add_u32_e32 v28, 3, v282574; GFX9-NEXT:    v_add_u32_e32 v27, 3, v272575; GFX9-NEXT:    v_add_u32_e32 v26, 3, v262576; GFX9-NEXT:    v_add_u32_e32 v25, 3, v252577; GFX9-NEXT:    v_add_u32_e32 v24, 3, v242578; GFX9-NEXT:    v_add_u32_e32 v23, 3, v232579; GFX9-NEXT:    v_add_u32_e32 v22, 3, v222580; GFX9-NEXT:    v_add_u32_e32 v21, 3, v212581; GFX9-NEXT:    v_add_u32_e32 v20, 3, v202582; GFX9-NEXT:    v_add_u32_e32 v19, 3, v192583; GFX9-NEXT:    v_add_u32_e32 v32, 3, v322584; GFX9-NEXT:    v_add_u32_e32 v17, 3, v172585; GFX9-NEXT:    v_add_u32_e32 v16, 3, v162586; GFX9-NEXT:    v_add_u32_e32 v15, 3, v152587; GFX9-NEXT:    v_add_u32_e32 v14, 3, v142588; GFX9-NEXT:    v_add_u32_e32 v13, 3, v132589; GFX9-NEXT:    v_add_u32_e32 v12, 3, v122590; GFX9-NEXT:    v_add_u32_e32 v11, 3, v112591; GFX9-NEXT:    v_add_u32_e32 v10, 3, v102592; GFX9-NEXT:    v_add_u32_e32 v9, 3, v92593; GFX9-NEXT:    v_add_u32_e32 v8, 3, v82594; GFX9-NEXT:    v_add_u32_e32 v7, 3, v72595; GFX9-NEXT:    v_add_u32_e32 v6, 3, v62596; GFX9-NEXT:    v_add_u32_e32 v5, 3, v52597; GFX9-NEXT:    v_add_u32_e32 v4, 3, v42598; GFX9-NEXT:    v_add_u32_e32 v3, 3, v32599; GFX9-NEXT:    v_add_u32_e32 v2, 3, v22600; GFX9-NEXT:    v_add_u32_e32 v1, 3, v12601; GFX9-NEXT:    v_add_u32_e32 v0, 3, v02602; GFX9-NEXT:  .LBB9_3: ; %end2603; GFX9-NEXT:    v_mov_b32_e32 v18, v322604; GFX9-NEXT:    s_setpc_b64 s[30:31]2605; GFX9-NEXT:  .LBB9_4:2606; GFX9-NEXT:    s_branch .LBB9_22607;2608; GFX11-LABEL: bitcast_v32i32_to_v16f64_scalar:2609; GFX11:       ; %bb.0:2610; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2611; GFX11-NEXT:    v_dual_mov_b32 v16, v14 :: v_dual_mov_b32 v31, v132612; GFX11-NEXT:    v_dual_mov_b32 v30, v12 :: v_dual_mov_b32 v29, v112613; GFX11-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v92614; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)2615; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v162616; GFX11-NEXT:    v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v25, v72617; GFX11-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v52618; GFX11-NEXT:    v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v32619; GFX11-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v12620; GFX11-NEXT:    v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s12621; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s32622; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s172623; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s192624; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s212625; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s232626; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s252627; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s272628; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s292629; GFX11-NEXT:    v_mov_b32_e32 v16, s282630; GFX11-NEXT:    s_mov_b32 s0, 02631; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo2632; GFX11-NEXT:    s_cbranch_scc0 .LBB9_32633; GFX11-NEXT:  ; %bb.1: ; %Flow2634; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s02635; GFX11-NEXT:    s_cbranch_vccz .LBB9_42636; GFX11-NEXT:  ; %bb.2: ; %end2637; GFX11-NEXT:    s_setpc_b64 s[30:31]2638; GFX11-NEXT:  .LBB9_3:2639; GFX11-NEXT:  .LBB9_4: ; %cmp.true2640; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v312641; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v302642; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v292643; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v282644; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v272645; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v262646; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v252647; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v242648; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v232649; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v222650; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v212651; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v202652; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v192653; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v182654; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v172655; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v162656; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v152657; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v142658; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v132659; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v122660; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v112661; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v102662; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v92663; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v82664; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v72665; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v62666; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v52667; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v42668; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v32669; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v22670; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v12671; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v02672; GFX11-NEXT:    s_setpc_b64 s[30:31]2673  %cmp = icmp eq i32 %b, 02674  br i1 %cmp, label %cmp.true, label %cmp.false2675 2676cmp.true:2677  %a1 = add <32 x i32> %a, splat (i32 3)2678  %a2 = bitcast <32 x i32> %a1 to <16 x double>2679  br label %end2680 2681cmp.false:2682  %a3 = bitcast <32 x i32> %a to <16 x double>2683  br label %end2684 2685end:2686  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]2687  ret <16 x double> %phi2688}2689 2690define <32 x i32> @bitcast_v16f64_to_v32i32(<16 x double> %a, i32 %b) {2691; SI-LABEL: bitcast_v16f64_to_v32i32:2692; SI:       ; %bb.0:2693; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2694; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:42695; SI-NEXT:    s_waitcnt vmcnt(0)2696; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v312697; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s322698; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc2699; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]2700; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]2701; SI-NEXT:    s_cbranch_execz .LBB10_22702; SI-NEXT:  ; %bb.1: ; %cmp.true2703; SI-NEXT:    s_waitcnt vmcnt(0)2704; SI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.02705; SI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.02706; SI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.02707; SI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.02708; SI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.02709; SI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.02710; SI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.02711; SI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.02712; SI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.02713; SI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.02714; SI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.02715; SI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.02716; SI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.02717; SI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.02718; SI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.02719; SI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.02720; SI-NEXT:  .LBB10_2: ; %end2721; SI-NEXT:    s_or_b64 exec, exec, s[4:5]2722; SI-NEXT:    s_waitcnt vmcnt(0)2723; SI-NEXT:    s_setpc_b64 s[30:31]2724;2725; VI-LABEL: bitcast_v16f64_to_v32i32:2726; VI:       ; %bb.0:2727; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2728; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:42729; VI-NEXT:    s_waitcnt vmcnt(0)2730; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v312731; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s322732; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc2733; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]2734; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]2735; VI-NEXT:    s_cbranch_execz .LBB10_22736; VI-NEXT:  ; %bb.1: ; %cmp.true2737; VI-NEXT:    s_waitcnt vmcnt(0)2738; VI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.02739; VI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.02740; VI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.02741; VI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.02742; VI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.02743; VI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.02744; VI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.02745; VI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.02746; VI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.02747; VI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.02748; VI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.02749; VI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.02750; VI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.02751; VI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.02752; VI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.02753; VI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.02754; VI-NEXT:  .LBB10_2: ; %end2755; VI-NEXT:    s_or_b64 exec, exec, s[4:5]2756; VI-NEXT:    s_waitcnt vmcnt(0)2757; VI-NEXT:    s_setpc_b64 s[30:31]2758;2759; GFX9-LABEL: bitcast_v16f64_to_v32i32:2760; GFX9:       ; %bb.0:2761; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2762; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:42763; GFX9-NEXT:    s_waitcnt vmcnt(0)2764; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v312765; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s322766; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc2767; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]2768; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]2769; GFX9-NEXT:    s_cbranch_execz .LBB10_22770; GFX9-NEXT:  ; %bb.1: ; %cmp.true2771; GFX9-NEXT:    s_waitcnt vmcnt(0)2772; GFX9-NEXT:    v_add_f64 v[30:31], v[30:31], 1.02773; GFX9-NEXT:    v_add_f64 v[28:29], v[28:29], 1.02774; GFX9-NEXT:    v_add_f64 v[26:27], v[26:27], 1.02775; GFX9-NEXT:    v_add_f64 v[24:25], v[24:25], 1.02776; GFX9-NEXT:    v_add_f64 v[22:23], v[22:23], 1.02777; GFX9-NEXT:    v_add_f64 v[20:21], v[20:21], 1.02778; GFX9-NEXT:    v_add_f64 v[18:19], v[18:19], 1.02779; GFX9-NEXT:    v_add_f64 v[16:17], v[16:17], 1.02780; GFX9-NEXT:    v_add_f64 v[14:15], v[14:15], 1.02781; GFX9-NEXT:    v_add_f64 v[12:13], v[12:13], 1.02782; GFX9-NEXT:    v_add_f64 v[10:11], v[10:11], 1.02783; GFX9-NEXT:    v_add_f64 v[8:9], v[8:9], 1.02784; GFX9-NEXT:    v_add_f64 v[6:7], v[6:7], 1.02785; GFX9-NEXT:    v_add_f64 v[4:5], v[4:5], 1.02786; GFX9-NEXT:    v_add_f64 v[2:3], v[2:3], 1.02787; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], 1.02788; GFX9-NEXT:  .LBB10_2: ; %end2789; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]2790; GFX9-NEXT:    s_waitcnt vmcnt(0)2791; GFX9-NEXT:    s_setpc_b64 s[30:31]2792;2793; GFX11-LABEL: bitcast_v16f64_to_v32i32:2794; GFX11:       ; %bb.0:2795; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2796; GFX11-NEXT:    s_clause 0x12797; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:42798; GFX11-NEXT:    scratch_load_b32 v31, off, s322799; GFX11-NEXT:    s_mov_b32 s0, exec_lo2800; GFX11-NEXT:    s_waitcnt vmcnt(1)2801; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v322802; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s02803; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)2804; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s02805; GFX11-NEXT:    s_cbranch_execz .LBB10_22806; GFX11-NEXT:  ; %bb.1: ; %cmp.true2807; GFX11-NEXT:    s_waitcnt vmcnt(0)2808; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.02809; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.02810; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.02811; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.02812; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.02813; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.02814; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.02815; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.02816; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.02817; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.02818; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.02819; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.02820; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.02821; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.02822; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.02823; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.02824; GFX11-NEXT:  .LBB10_2: ; %end2825; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s02826; GFX11-NEXT:    s_waitcnt vmcnt(0)2827; GFX11-NEXT:    s_setpc_b64 s[30:31]2828  %cmp = icmp eq i32 %b, 02829  br i1 %cmp, label %cmp.true, label %cmp.false2830 2831cmp.true:2832  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)2833  %a2 = bitcast <16 x double> %a1 to <32 x i32>2834  br label %end2835 2836cmp.false:2837  %a3 = bitcast <16 x double> %a to <32 x i32>2838  br label %end2839 2840end:2841  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]2842  ret <32 x i32> %phi2843}2844 2845define inreg <32 x i32> @bitcast_v16f64_to_v32i32_scalar(<16 x double> inreg %a, i32 inreg %b) {2846; SI-LABEL: bitcast_v16f64_to_v32i32_scalar:2847; SI:       ; %bb.0:2848; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2849; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v182850; SI-NEXT:    v_mov_b32_e32 v31, v172851; SI-NEXT:    v_mov_b32_e32 v30, v162852; SI-NEXT:    v_mov_b32_e32 v29, v152853; SI-NEXT:    v_mov_b32_e32 v28, v142854; SI-NEXT:    v_mov_b32_e32 v27, v132855; SI-NEXT:    v_mov_b32_e32 v26, v122856; SI-NEXT:    v_mov_b32_e32 v25, v112857; SI-NEXT:    v_mov_b32_e32 v24, v102858; SI-NEXT:    v_mov_b32_e32 v23, v92859; SI-NEXT:    v_mov_b32_e32 v22, v82860; SI-NEXT:    v_mov_b32_e32 v21, v72861; SI-NEXT:    v_mov_b32_e32 v20, v62862; SI-NEXT:    v_mov_b32_e32 v33, v52863; SI-NEXT:    v_mov_b32_e32 v32, v42864; SI-NEXT:    v_mov_b32_e32 v17, v32865; SI-NEXT:    v_mov_b32_e32 v16, v22866; SI-NEXT:    v_mov_b32_e32 v15, v12867; SI-NEXT:    v_mov_b32_e32 v14, v02868; SI-NEXT:    v_mov_b32_e32 v0, s162869; SI-NEXT:    v_mov_b32_e32 v1, s172870; SI-NEXT:    v_mov_b32_e32 v2, s182871; SI-NEXT:    v_mov_b32_e32 v3, s192872; SI-NEXT:    v_mov_b32_e32 v4, s202873; SI-NEXT:    v_mov_b32_e32 v5, s212874; SI-NEXT:    v_mov_b32_e32 v6, s222875; SI-NEXT:    v_mov_b32_e32 v7, s232876; SI-NEXT:    v_mov_b32_e32 v8, s242877; SI-NEXT:    v_mov_b32_e32 v9, s252878; SI-NEXT:    v_mov_b32_e32 v10, s262879; SI-NEXT:    v_mov_b32_e32 v11, s272880; SI-NEXT:    s_and_b64 s[4:5], vcc, exec2881; SI-NEXT:    v_mov_b32_e32 v12, s282882; SI-NEXT:    v_mov_b32_e32 v13, s292883; SI-NEXT:    s_cbranch_scc0 .LBB11_42884; SI-NEXT:  ; %bb.1: ; %cmp.false2885; SI-NEXT:    s_cbranch_execnz .LBB11_32886; SI-NEXT:  .LBB11_2: ; %cmp.true2887; SI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.02888; SI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.02889; SI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.02890; SI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.02891; SI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.02892; SI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.02893; SI-NEXT:    v_add_f64 v[32:33], v[32:33], 1.02894; SI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.02895; SI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.02896; SI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.02897; SI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.02898; SI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.02899; SI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.02900; SI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.02901; SI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.02902; SI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.02903; SI-NEXT:  .LBB11_3: ; %end2904; SI-NEXT:    v_mov_b32_e32 v18, v322905; SI-NEXT:    v_mov_b32_e32 v19, v332906; SI-NEXT:    s_setpc_b64 s[30:31]2907; SI-NEXT:  .LBB11_4:2908; SI-NEXT:    s_branch .LBB11_22909;2910; VI-LABEL: bitcast_v16f64_to_v32i32_scalar:2911; VI:       ; %bb.0:2912; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2913; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v182914; VI-NEXT:    v_mov_b32_e32 v31, v172915; VI-NEXT:    v_mov_b32_e32 v30, v162916; VI-NEXT:    v_mov_b32_e32 v29, v152917; VI-NEXT:    v_mov_b32_e32 v28, v142918; VI-NEXT:    v_mov_b32_e32 v27, v132919; VI-NEXT:    v_mov_b32_e32 v26, v122920; VI-NEXT:    v_mov_b32_e32 v25, v112921; VI-NEXT:    v_mov_b32_e32 v24, v102922; VI-NEXT:    v_mov_b32_e32 v23, v92923; VI-NEXT:    v_mov_b32_e32 v22, v82924; VI-NEXT:    v_mov_b32_e32 v21, v72925; VI-NEXT:    v_mov_b32_e32 v20, v62926; VI-NEXT:    v_mov_b32_e32 v33, v52927; VI-NEXT:    v_mov_b32_e32 v32, v42928; VI-NEXT:    v_mov_b32_e32 v17, v32929; VI-NEXT:    v_mov_b32_e32 v16, v22930; VI-NEXT:    v_mov_b32_e32 v15, v12931; VI-NEXT:    v_mov_b32_e32 v14, v02932; VI-NEXT:    v_mov_b32_e32 v0, s162933; VI-NEXT:    v_mov_b32_e32 v1, s172934; VI-NEXT:    v_mov_b32_e32 v2, s182935; VI-NEXT:    v_mov_b32_e32 v3, s192936; VI-NEXT:    v_mov_b32_e32 v4, s202937; VI-NEXT:    v_mov_b32_e32 v5, s212938; VI-NEXT:    v_mov_b32_e32 v6, s222939; VI-NEXT:    v_mov_b32_e32 v7, s232940; VI-NEXT:    v_mov_b32_e32 v8, s242941; VI-NEXT:    v_mov_b32_e32 v9, s252942; VI-NEXT:    v_mov_b32_e32 v10, s262943; VI-NEXT:    v_mov_b32_e32 v11, s272944; VI-NEXT:    s_and_b64 s[4:5], vcc, exec2945; VI-NEXT:    v_mov_b32_e32 v12, s282946; VI-NEXT:    v_mov_b32_e32 v13, s292947; VI-NEXT:    s_cbranch_scc0 .LBB11_42948; VI-NEXT:  ; %bb.1: ; %cmp.false2949; VI-NEXT:    s_cbranch_execnz .LBB11_32950; VI-NEXT:  .LBB11_2: ; %cmp.true2951; VI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.02952; VI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.02953; VI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.02954; VI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.02955; VI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.02956; VI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.02957; VI-NEXT:    v_add_f64 v[32:33], v[32:33], 1.02958; VI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.02959; VI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.02960; VI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.02961; VI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.02962; VI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.02963; VI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.02964; VI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.02965; VI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.02966; VI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.02967; VI-NEXT:  .LBB11_3: ; %end2968; VI-NEXT:    v_mov_b32_e32 v18, v322969; VI-NEXT:    v_mov_b32_e32 v19, v332970; VI-NEXT:    s_setpc_b64 s[30:31]2971; VI-NEXT:  .LBB11_4:2972; VI-NEXT:    s_branch .LBB11_22973;2974; GFX9-LABEL: bitcast_v16f64_to_v32i32_scalar:2975; GFX9:       ; %bb.0:2976; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2977; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v182978; GFX9-NEXT:    v_mov_b32_e32 v31, v172979; GFX9-NEXT:    v_mov_b32_e32 v30, v162980; GFX9-NEXT:    v_mov_b32_e32 v29, v152981; GFX9-NEXT:    v_mov_b32_e32 v28, v142982; GFX9-NEXT:    v_mov_b32_e32 v27, v132983; GFX9-NEXT:    v_mov_b32_e32 v26, v122984; GFX9-NEXT:    v_mov_b32_e32 v25, v112985; GFX9-NEXT:    v_mov_b32_e32 v24, v102986; GFX9-NEXT:    v_mov_b32_e32 v23, v92987; GFX9-NEXT:    v_mov_b32_e32 v22, v82988; GFX9-NEXT:    v_mov_b32_e32 v21, v72989; GFX9-NEXT:    v_mov_b32_e32 v20, v62990; GFX9-NEXT:    v_mov_b32_e32 v33, v52991; GFX9-NEXT:    v_mov_b32_e32 v32, v42992; GFX9-NEXT:    v_mov_b32_e32 v17, v32993; GFX9-NEXT:    v_mov_b32_e32 v16, v22994; GFX9-NEXT:    v_mov_b32_e32 v15, v12995; GFX9-NEXT:    v_mov_b32_e32 v14, v02996; GFX9-NEXT:    v_mov_b32_e32 v0, s162997; GFX9-NEXT:    v_mov_b32_e32 v1, s172998; GFX9-NEXT:    v_mov_b32_e32 v2, s182999; GFX9-NEXT:    v_mov_b32_e32 v3, s193000; GFX9-NEXT:    v_mov_b32_e32 v4, s203001; GFX9-NEXT:    v_mov_b32_e32 v5, s213002; GFX9-NEXT:    v_mov_b32_e32 v6, s223003; GFX9-NEXT:    v_mov_b32_e32 v7, s233004; GFX9-NEXT:    v_mov_b32_e32 v8, s243005; GFX9-NEXT:    v_mov_b32_e32 v9, s253006; GFX9-NEXT:    v_mov_b32_e32 v10, s263007; GFX9-NEXT:    v_mov_b32_e32 v11, s273008; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec3009; GFX9-NEXT:    v_mov_b32_e32 v12, s283010; GFX9-NEXT:    v_mov_b32_e32 v13, s293011; GFX9-NEXT:    s_cbranch_scc0 .LBB11_43012; GFX9-NEXT:  ; %bb.1: ; %cmp.false3013; GFX9-NEXT:    s_cbranch_execnz .LBB11_33014; GFX9-NEXT:  .LBB11_2: ; %cmp.true3015; GFX9-NEXT:    v_add_f64 v[30:31], v[30:31], 1.03016; GFX9-NEXT:    v_add_f64 v[28:29], v[28:29], 1.03017; GFX9-NEXT:    v_add_f64 v[26:27], v[26:27], 1.03018; GFX9-NEXT:    v_add_f64 v[24:25], v[24:25], 1.03019; GFX9-NEXT:    v_add_f64 v[22:23], v[22:23], 1.03020; GFX9-NEXT:    v_add_f64 v[20:21], v[20:21], 1.03021; GFX9-NEXT:    v_add_f64 v[32:33], v[32:33], 1.03022; GFX9-NEXT:    v_add_f64 v[16:17], v[16:17], 1.03023; GFX9-NEXT:    v_add_f64 v[14:15], v[14:15], 1.03024; GFX9-NEXT:    v_add_f64 v[12:13], v[12:13], 1.03025; GFX9-NEXT:    v_add_f64 v[10:11], v[10:11], 1.03026; GFX9-NEXT:    v_add_f64 v[8:9], v[8:9], 1.03027; GFX9-NEXT:    v_add_f64 v[6:7], v[6:7], 1.03028; GFX9-NEXT:    v_add_f64 v[4:5], v[4:5], 1.03029; GFX9-NEXT:    v_add_f64 v[2:3], v[2:3], 1.03030; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], 1.03031; GFX9-NEXT:  .LBB11_3: ; %end3032; GFX9-NEXT:    v_mov_b32_e32 v18, v323033; GFX9-NEXT:    v_mov_b32_e32 v19, v333034; GFX9-NEXT:    s_setpc_b64 s[30:31]3035; GFX9-NEXT:  .LBB11_4:3036; GFX9-NEXT:    s_branch .LBB11_23037;3038; GFX11-LABEL: bitcast_v16f64_to_v32i32_scalar:3039; GFX11:       ; %bb.0:3040; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)3041; GFX11-NEXT:    v_dual_mov_b32 v16, v14 :: v_dual_mov_b32 v31, v133042; GFX11-NEXT:    v_dual_mov_b32 v30, v12 :: v_dual_mov_b32 v29, v113043; GFX11-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v93044; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)3045; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v163046; GFX11-NEXT:    v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v25, v73047; GFX11-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v53048; GFX11-NEXT:    v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v33049; GFX11-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v13050; GFX11-NEXT:    v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s13051; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s33052; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s173053; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s193054; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s213055; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s233056; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s253057; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s273058; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s293059; GFX11-NEXT:    v_mov_b32_e32 v16, s283060; GFX11-NEXT:    s_mov_b32 s0, 03061; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo3062; GFX11-NEXT:    s_cbranch_scc0 .LBB11_33063; GFX11-NEXT:  ; %bb.1: ; %Flow3064; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s03065; GFX11-NEXT:    s_cbranch_vccz .LBB11_43066; GFX11-NEXT:  ; %bb.2: ; %end3067; GFX11-NEXT:    s_setpc_b64 s[30:31]3068; GFX11-NEXT:  .LBB11_3:3069; GFX11-NEXT:  .LBB11_4: ; %cmp.true3070; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.03071; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.03072; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.03073; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.03074; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.03075; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.03076; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.03077; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.03078; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.03079; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.03080; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.03081; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.03082; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.03083; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.03084; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.03085; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.03086; GFX11-NEXT:    s_setpc_b64 s[30:31]3087  %cmp = icmp eq i32 %b, 03088  br i1 %cmp, label %cmp.true, label %cmp.false3089 3090cmp.true:3091  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)3092  %a2 = bitcast <16 x double> %a1 to <32 x i32>3093  br label %end3094 3095cmp.false:3096  %a3 = bitcast <16 x double> %a to <32 x i32>3097  br label %end3098 3099end:3100  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]3101  ret <32 x i32> %phi3102}3103 3104define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {3105; SI-LABEL: bitcast_v32i32_to_v128i8:3106; SI:       ; %bb.0:3107; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)3108; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:43109; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:83110; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s323111; SI-NEXT:    ; implicit-def: $vgpr363112; SI-NEXT:    ; kill: killed $vgpr363113; SI-NEXT:    ; implicit-def: $vgpr363114; SI-NEXT:    ; kill: killed $vgpr363115; SI-NEXT:    ; implicit-def: $vgpr363116; SI-NEXT:    ; kill: killed $vgpr363117; SI-NEXT:    ; implicit-def: $vgpr363118; SI-NEXT:    ; kill: killed $vgpr363119; SI-NEXT:    ; implicit-def: $vgpr363120; SI-NEXT:    ; kill: killed $vgpr363121; SI-NEXT:    ; implicit-def: $vgpr363122; SI-NEXT:    ; kill: killed $vgpr363123; SI-NEXT:    ; implicit-def: $vgpr363124; SI-NEXT:    ; kill: killed $vgpr363125; SI-NEXT:    ; implicit-def: $vgpr363126; SI-NEXT:    ; kill: killed $vgpr363127; SI-NEXT:    ; implicit-def: $vgpr363128; SI-NEXT:    ; kill: killed $vgpr363129; SI-NEXT:    ; implicit-def: $vgpr363130; SI-NEXT:    ; kill: killed $vgpr363131; SI-NEXT:    ; implicit-def: $vgpr363132; SI-NEXT:    ; kill: killed $vgpr363133; SI-NEXT:    ; implicit-def: $vgpr363134; SI-NEXT:    ; kill: killed $vgpr363135; SI-NEXT:    ; implicit-def: $vgpr363136; SI-NEXT:    ; kill: killed $vgpr363137; SI-NEXT:    ; implicit-def: $vgpr363138; SI-NEXT:    ; kill: killed $vgpr363139; SI-NEXT:    ; implicit-def: $vgpr363140; SI-NEXT:    ; kill: killed $vgpr363141; SI-NEXT:    ; implicit-def: $vgpr363142; SI-NEXT:    ; kill: killed $vgpr363143; SI-NEXT:    ; implicit-def: $vgpr363144; SI-NEXT:    ; kill: killed $vgpr363145; SI-NEXT:    ; implicit-def: $vgpr363146; SI-NEXT:    ; kill: killed $vgpr363147; SI-NEXT:    ; implicit-def: $vgpr363148; SI-NEXT:    ; kill: killed $vgpr363149; SI-NEXT:    ; implicit-def: $vgpr363150; SI-NEXT:    ; kill: killed $vgpr363151; SI-NEXT:    ; implicit-def: $vgpr363152; SI-NEXT:    ; kill: killed $vgpr363153; SI-NEXT:    ; implicit-def: $vgpr363154; SI-NEXT:    ; kill: killed $vgpr363155; SI-NEXT:    ; implicit-def: $vgpr363156; SI-NEXT:    ; kill: killed $vgpr363157; SI-NEXT:    ; implicit-def: $vgpr363158; SI-NEXT:    ; kill: killed $vgpr363159; SI-NEXT:    ; implicit-def: $vgpr363160; SI-NEXT:    ; kill: killed $vgpr363161; SI-NEXT:    ; implicit-def: $vgpr363162; SI-NEXT:    ; kill: killed $vgpr363163; SI-NEXT:    ; implicit-def: $vgpr363164; SI-NEXT:    ; kill: killed $vgpr363165; SI-NEXT:    ; implicit-def: $vgpr363166; SI-NEXT:    ; kill: killed $vgpr363167; SI-NEXT:    ; implicit-def: $vgpr363168; SI-NEXT:    ; kill: killed $vgpr363169; SI-NEXT:    ; implicit-def: $vgpr363170; SI-NEXT:    ; kill: killed $vgpr363171; SI-NEXT:    ; implicit-def: $vgpr363172; SI-NEXT:    ; kill: killed $vgpr363173; SI-NEXT:    ; implicit-def: $vgpr363174; SI-NEXT:    ; kill: killed $vgpr363175; SI-NEXT:    ; implicit-def: $vgpr363176; SI-NEXT:    ; kill: killed $vgpr363177; SI-NEXT:    ; implicit-def: $vgpr363178; SI-NEXT:    ; kill: killed $vgpr363179; SI-NEXT:    ; implicit-def: $vgpr363180; SI-NEXT:    ; kill: killed $vgpr363181; SI-NEXT:    ; implicit-def: $vgpr363182; SI-NEXT:    ; kill: killed $vgpr363183; SI-NEXT:    ; implicit-def: $vgpr363184; SI-NEXT:    ; kill: killed $vgpr363185; SI-NEXT:    ; implicit-def: $vgpr363186; SI-NEXT:    ; kill: killed $vgpr363187; SI-NEXT:    ; implicit-def: $vgpr363188; SI-NEXT:    ; kill: killed $vgpr363189; SI-NEXT:    ; implicit-def: $vgpr363190; SI-NEXT:    ; kill: killed $vgpr363191; SI-NEXT:    ; implicit-def: $vgpr363192; SI-NEXT:    ; kill: killed $vgpr363193; SI-NEXT:    ; implicit-def: $vgpr363194; SI-NEXT:    ; kill: killed $vgpr363195; SI-NEXT:    ; implicit-def: $vgpr363196; SI-NEXT:    ; kill: killed $vgpr363197; SI-NEXT:    ; implicit-def: $vgpr363198; SI-NEXT:    ; kill: killed $vgpr363199; SI-NEXT:    ; implicit-def: $vgpr363200; SI-NEXT:    ; kill: killed $vgpr363201; SI-NEXT:    ; implicit-def: $vgpr363202; SI-NEXT:    ; kill: killed $vgpr363203; SI-NEXT:    ; implicit-def: $vgpr363204; SI-NEXT:    ; kill: killed $vgpr363205; SI-NEXT:    ; implicit-def: $vgpr363206; SI-NEXT:    ; kill: killed $vgpr363207; SI-NEXT:    ; implicit-def: $vgpr363208; SI-NEXT:    ; kill: killed $vgpr363209; SI-NEXT:    ; implicit-def: $vgpr363210; SI-NEXT:    ; kill: killed $vgpr363211; SI-NEXT:    ; implicit-def: $vgpr363212; SI-NEXT:    ; kill: killed $vgpr363213; SI-NEXT:    ; implicit-def: $vgpr363214; SI-NEXT:    ; kill: killed $vgpr363215; SI-NEXT:    ; implicit-def: $vgpr363216; SI-NEXT:    ; kill: killed $vgpr363217; SI-NEXT:    ; implicit-def: $vgpr363218; SI-NEXT:    ; kill: killed $vgpr363219; SI-NEXT:    ; implicit-def: $vgpr363220; SI-NEXT:    ; kill: killed $vgpr363221; SI-NEXT:    ; implicit-def: $vgpr363222; SI-NEXT:    ; kill: killed $vgpr363223; SI-NEXT:    ; implicit-def: $vgpr363224; SI-NEXT:    ; kill: killed $vgpr363225; SI-NEXT:    ; implicit-def: $vgpr363226; SI-NEXT:    ; kill: killed $vgpr363227; SI-NEXT:    ; implicit-def: $vgpr363228; SI-NEXT:    ; kill: killed $vgpr363229; SI-NEXT:    ; implicit-def: $vgpr363230; SI-NEXT:    ; kill: killed $vgpr363231; SI-NEXT:    ; implicit-def: $vgpr363232; SI-NEXT:    ; kill: killed $vgpr363233; SI-NEXT:    ; implicit-def: $vgpr363234; SI-NEXT:    ; kill: killed $vgpr363235; SI-NEXT:    ; implicit-def: $vgpr363236; SI-NEXT:    ; kill: killed $vgpr363237; SI-NEXT:    ; implicit-def: $vgpr363238; SI-NEXT:    ; kill: killed $vgpr363239; SI-NEXT:    ; implicit-def: $vgpr363240; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill3241; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill3242; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill3243; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill3244; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill3245; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill3246; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill3247; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill3248; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill3249; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill3250; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill3251; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill3252; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill3253; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill3254; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill3255; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill3256; SI-NEXT:    ; implicit-def: $vgpr453257; SI-NEXT:    ; implicit-def: $vgpr433258; SI-NEXT:    ; implicit-def: $vgpr413259; SI-NEXT:    ; implicit-def: $vgpr353260; SI-NEXT:    ; implicit-def: $vgpr343261; SI-NEXT:    ; implicit-def: $vgpr553262; SI-NEXT:    ; implicit-def: $vgpr533263; SI-NEXT:    ; implicit-def: $vgpr513264; SI-NEXT:    ; implicit-def: $vgpr633265; SI-NEXT:    ; implicit-def: $vgpr623266; SI-NEXT:    ; implicit-def: $vgpr613267; SI-NEXT:    ; implicit-def: $vgpr483268; SI-NEXT:    ; implicit-def: $vgpr383269; SI-NEXT:    ; implicit-def: $vgpr373270; SI-NEXT:    ; implicit-def: $vgpr603271; SI-NEXT:    ; implicit-def: $vgpr593272; SI-NEXT:    ; implicit-def: $vgpr583273; SI-NEXT:    ; implicit-def: $vgpr573274; SI-NEXT:    ; implicit-def: $vgpr563275; SI-NEXT:    ; implicit-def: $vgpr473276; SI-NEXT:    ; implicit-def: $vgpr463277; SI-NEXT:    ; implicit-def: $vgpr443278; SI-NEXT:    ; implicit-def: $vgpr423279; SI-NEXT:    ; implicit-def: $vgpr403280; SI-NEXT:    ; implicit-def: $vgpr543281; SI-NEXT:    ; implicit-def: $vgpr523282; SI-NEXT:    ; implicit-def: $vgpr503283; SI-NEXT:    ; implicit-def: $vgpr493284; SI-NEXT:    ; implicit-def: $vgpr393285; SI-NEXT:    ; kill: killed $vgpr363286; SI-NEXT:    ; implicit-def: $vgpr363287; SI-NEXT:    s_waitcnt vmcnt(14)3288; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v333289; SI-NEXT:    ; implicit-def: $vgpr333290; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc3291; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]3292; SI-NEXT:    s_cbranch_execz .LBB12_23293; SI-NEXT:  ; %bb.1: ; %cmp.false3294; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 243295; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill3296; SI-NEXT:    s_waitcnt expcnt(0)3297; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 163298; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill3299; SI-NEXT:    s_waitcnt expcnt(0)3300; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 83301; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill3302; SI-NEXT:    s_waitcnt expcnt(0)3303; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 243304; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill3305; SI-NEXT:    s_waitcnt expcnt(0)3306; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 163307; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill3308; SI-NEXT:    s_waitcnt expcnt(0)3309; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 83310; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill3311; SI-NEXT:    s_waitcnt expcnt(0)3312; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 243313; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill3314; SI-NEXT:    s_waitcnt expcnt(0)3315; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 163316; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill3317; SI-NEXT:    s_waitcnt expcnt(0)3318; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 83319; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill3320; SI-NEXT:    s_waitcnt expcnt(0)3321; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 243322; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill3323; SI-NEXT:    s_waitcnt expcnt(0)3324; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 163325; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill3326; SI-NEXT:    s_waitcnt expcnt(0)3327; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 83328; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill3329; SI-NEXT:    s_waitcnt expcnt(0)3330; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 243331; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill3332; SI-NEXT:    s_waitcnt expcnt(0)3333; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 163334; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill3335; SI-NEXT:    s_waitcnt expcnt(0)3336; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 83337; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill3338; SI-NEXT:    s_waitcnt expcnt(0)3339; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 243340; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill3341; SI-NEXT:    s_waitcnt expcnt(0)3342; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 163343; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill3344; SI-NEXT:    s_waitcnt expcnt(0)3345; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 83346; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill3347; SI-NEXT:    s_waitcnt expcnt(0)3348; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 243349; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill3350; SI-NEXT:    s_waitcnt expcnt(0)3351; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 163352; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill3353; SI-NEXT:    s_waitcnt expcnt(0)3354; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 83355; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill3356; SI-NEXT:    s_waitcnt expcnt(0)3357; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 243358; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill3359; SI-NEXT:    s_waitcnt expcnt(0)3360; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 163361; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill3362; SI-NEXT:    s_waitcnt expcnt(0)3363; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 83364; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill3365; SI-NEXT:    s_waitcnt expcnt(0)3366; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 243367; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill3368; SI-NEXT:    s_waitcnt expcnt(0)3369; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 163370; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill3371; SI-NEXT:    s_waitcnt expcnt(0)3372; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 83373; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill3374; SI-NEXT:    s_waitcnt expcnt(0)3375; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 243376; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill3377; SI-NEXT:    s_waitcnt expcnt(0)3378; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 163379; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill3380; SI-NEXT:    s_waitcnt expcnt(0)3381; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 83382; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill3383; SI-NEXT:    s_waitcnt expcnt(0)3384; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 243385; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill3386; SI-NEXT:    s_waitcnt expcnt(0)3387; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 163388; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill3389; SI-NEXT:    s_waitcnt expcnt(0)3390; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 83391; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill3392; SI-NEXT:    s_waitcnt expcnt(0)3393; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 243394; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill3395; SI-NEXT:    s_waitcnt expcnt(0)3396; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 163397; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill3398; SI-NEXT:    s_waitcnt expcnt(0)3399; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 83400; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill3401; SI-NEXT:    s_waitcnt expcnt(0)3402; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 243403; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill3404; SI-NEXT:    s_waitcnt expcnt(0)3405; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 163406; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill3407; SI-NEXT:    s_waitcnt expcnt(0)3408; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 83409; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill3410; SI-NEXT:    s_waitcnt expcnt(0)3411; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v313412; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill3413; SI-NEXT:    s_waitcnt expcnt(0)3414; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v313415; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill3416; SI-NEXT:    s_waitcnt expcnt(0)3417; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v303418; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill3419; SI-NEXT:    s_waitcnt expcnt(0)3420; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v303421; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill3422; SI-NEXT:    s_waitcnt expcnt(0)3423; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v303424; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill3425; SI-NEXT:    s_waitcnt expcnt(0)3426; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v283427; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill3428; SI-NEXT:    s_waitcnt expcnt(0)3429; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v283430; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill3431; SI-NEXT:    s_waitcnt expcnt(0)3432; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v283433; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill3434; SI-NEXT:    s_waitcnt expcnt(0)3435; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v263436; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill3437; SI-NEXT:    s_waitcnt expcnt(0)3438; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v263439; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill3440; SI-NEXT:    s_waitcnt expcnt(0)3441; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v263442; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill3443; SI-NEXT:    s_waitcnt expcnt(0)3444; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v243445; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill3446; SI-NEXT:    s_waitcnt expcnt(0)3447; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v243448; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill3449; SI-NEXT:    s_waitcnt expcnt(0)3450; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v243451; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill3452; SI-NEXT:    s_waitcnt expcnt(0)3453; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v223454; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill3455; SI-NEXT:    s_waitcnt expcnt(0)3456; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v223457; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill3458; SI-NEXT:    s_waitcnt expcnt(0)3459; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v223460; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill3461; SI-NEXT:    s_waitcnt expcnt(0)3462; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v203463; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill3464; SI-NEXT:    s_waitcnt expcnt(0)3465; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v203466; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill3467; SI-NEXT:    s_waitcnt expcnt(0)3468; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v203469; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill3470; SI-NEXT:    s_waitcnt expcnt(0)3471; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v183472; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill3473; SI-NEXT:    s_waitcnt expcnt(0)3474; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v183475; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill3476; SI-NEXT:    s_waitcnt expcnt(0)3477; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v183478; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill3479; SI-NEXT:    s_waitcnt expcnt(0)3480; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v163481; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill3482; SI-NEXT:    s_waitcnt expcnt(0)3483; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v163484; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill3485; SI-NEXT:    s_waitcnt expcnt(0)3486; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v163487; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 243488; SI-NEXT:    v_alignbit_b32 v38, v6, v5, 163489; SI-NEXT:    v_alignbit_b32 v48, v6, v5, 83490; SI-NEXT:    v_alignbit_b32 v51, v4, v3, 243491; SI-NEXT:    v_alignbit_b32 v53, v4, v3, 163492; SI-NEXT:    v_alignbit_b32 v55, v4, v3, 83493; SI-NEXT:    v_alignbit_b32 v41, v2, v1, 243494; SI-NEXT:    v_alignbit_b32 v43, v2, v1, 163495; SI-NEXT:    v_alignbit_b32 v45, v2, v1, 83496; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v313497; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill3498; SI-NEXT:    v_lshrrev_b32_e32 v39, 24, v143499; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v143500; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v143501; SI-NEXT:    v_lshrrev_b32_e32 v52, 24, v123502; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v123503; SI-NEXT:    v_lshrrev_b32_e32 v40, 8, v123504; SI-NEXT:    v_lshrrev_b32_e32 v42, 24, v103505; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v103506; SI-NEXT:    v_lshrrev_b32_e32 v46, 8, v103507; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v83508; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v83509; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v83510; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v63511; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v63512; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v63513; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v43514; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v43515; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v43516; SI-NEXT:    s_waitcnt expcnt(0)3517; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v23518; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v23519; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v23520; SI-NEXT:  .LBB12_2: ; %Flow3521; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]3522; SI-NEXT:    s_cbranch_execz .LBB12_43523; SI-NEXT:  ; %bb.3: ; %cmp.true3524; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v313525; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v323526; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 243527; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill3528; SI-NEXT:    s_waitcnt expcnt(0)3529; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 163530; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v303531; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v293532; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill3533; SI-NEXT:    s_waitcnt expcnt(0)3534; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 83535; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill3536; SI-NEXT:    s_waitcnt expcnt(0)3537; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 243538; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill3539; SI-NEXT:    s_waitcnt expcnt(0)3540; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 163541; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v283542; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v273543; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill3544; SI-NEXT:    s_waitcnt expcnt(0)3545; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 83546; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill3547; SI-NEXT:    s_waitcnt expcnt(0)3548; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 243549; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill3550; SI-NEXT:    s_waitcnt expcnt(0)3551; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 163552; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v263553; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v253554; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill3555; SI-NEXT:    s_waitcnt expcnt(0)3556; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 83557; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill3558; SI-NEXT:    s_waitcnt expcnt(0)3559; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 243560; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill3561; SI-NEXT:    s_waitcnt expcnt(0)3562; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 163563; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v243564; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v233565; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill3566; SI-NEXT:    s_waitcnt expcnt(0)3567; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 83568; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill3569; SI-NEXT:    s_waitcnt expcnt(0)3570; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 243571; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill3572; SI-NEXT:    s_waitcnt expcnt(0)3573; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 163574; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v223575; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v213576; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill3577; SI-NEXT:    s_waitcnt expcnt(0)3578; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 83579; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill3580; SI-NEXT:    s_waitcnt expcnt(0)3581; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 243582; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill3583; SI-NEXT:    s_waitcnt expcnt(0)3584; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 163585; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v203586; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v193587; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill3588; SI-NEXT:    s_waitcnt expcnt(0)3589; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 83590; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill3591; SI-NEXT:    s_waitcnt expcnt(0)3592; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 243593; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill3594; SI-NEXT:    s_waitcnt expcnt(0)3595; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 163596; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v183597; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v173598; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill3599; SI-NEXT:    s_waitcnt expcnt(0)3600; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 83601; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill3602; SI-NEXT:    s_waitcnt expcnt(0)3603; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 243604; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill3605; SI-NEXT:    s_waitcnt expcnt(0)3606; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 163607; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v163608; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v153609; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill3610; SI-NEXT:    s_waitcnt expcnt(0)3611; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 83612; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill3613; SI-NEXT:    s_waitcnt expcnt(0)3614; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 243615; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill3616; SI-NEXT:    s_waitcnt expcnt(0)3617; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 163618; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v143619; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v133620; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill3621; SI-NEXT:    s_waitcnt expcnt(0)3622; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 83623; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill3624; SI-NEXT:    s_waitcnt expcnt(0)3625; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 243626; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill3627; SI-NEXT:    s_waitcnt expcnt(0)3628; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 163629; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v123630; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v113631; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill3632; SI-NEXT:    s_waitcnt expcnt(0)3633; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 83634; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill3635; SI-NEXT:    s_waitcnt expcnt(0)3636; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 243637; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill3638; SI-NEXT:    s_waitcnt expcnt(0)3639; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 163640; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v103641; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v93642; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill3643; SI-NEXT:    s_waitcnt expcnt(0)3644; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 83645; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill3646; SI-NEXT:    s_waitcnt expcnt(0)3647; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 243648; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill3649; SI-NEXT:    s_waitcnt expcnt(0)3650; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 163651; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v83652; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v73653; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill3654; SI-NEXT:    s_waitcnt expcnt(0)3655; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 83656; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill3657; SI-NEXT:    s_waitcnt expcnt(0)3658; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 243659; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill3660; SI-NEXT:    s_waitcnt expcnt(0)3661; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 163662; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill3663; SI-NEXT:    s_waitcnt expcnt(0)3664; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 83665; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill3666; SI-NEXT:    s_waitcnt expcnt(0)3667; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v313668; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill3669; SI-NEXT:    s_waitcnt expcnt(0)3670; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v313671; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill3672; SI-NEXT:    s_waitcnt expcnt(0)3673; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v303674; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill3675; SI-NEXT:    s_waitcnt expcnt(0)3676; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v303677; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill3678; SI-NEXT:    s_waitcnt expcnt(0)3679; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v303680; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill3681; SI-NEXT:    s_waitcnt expcnt(0)3682; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v283683; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill3684; SI-NEXT:    s_waitcnt expcnt(0)3685; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v283686; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill3687; SI-NEXT:    s_waitcnt expcnt(0)3688; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v283689; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill3690; SI-NEXT:    s_waitcnt expcnt(0)3691; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v263692; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill3693; SI-NEXT:    s_waitcnt expcnt(0)3694; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v263695; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill3696; SI-NEXT:    s_waitcnt expcnt(0)3697; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v263698; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill3699; SI-NEXT:    s_waitcnt expcnt(0)3700; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v243701; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill3702; SI-NEXT:    s_waitcnt expcnt(0)3703; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v243704; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill3705; SI-NEXT:    s_waitcnt expcnt(0)3706; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v243707; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill3708; SI-NEXT:    s_waitcnt expcnt(0)3709; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v223710; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill3711; SI-NEXT:    s_waitcnt expcnt(0)3712; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v223713; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill3714; SI-NEXT:    s_waitcnt expcnt(0)3715; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v223716; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill3717; SI-NEXT:    s_waitcnt expcnt(0)3718; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v203719; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill3720; SI-NEXT:    s_waitcnt expcnt(0)3721; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v203722; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill3723; SI-NEXT:    s_waitcnt expcnt(0)3724; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v203725; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill3726; SI-NEXT:    s_waitcnt expcnt(0)3727; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v183728; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill3729; SI-NEXT:    s_waitcnt expcnt(0)3730; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v183731; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill3732; SI-NEXT:    s_waitcnt expcnt(0)3733; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v183734; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill3735; SI-NEXT:    s_waitcnt expcnt(0)3736; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v163737; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill3738; SI-NEXT:    s_waitcnt expcnt(0)3739; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v163740; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v23741; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v13742; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v43743; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v33744; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v63745; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v53746; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill3747; SI-NEXT:    s_waitcnt expcnt(0)3748; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v163749; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 243750; SI-NEXT:    v_alignbit_b32 v38, v6, v5, 163751; SI-NEXT:    v_alignbit_b32 v48, v6, v5, 83752; SI-NEXT:    v_alignbit_b32 v51, v4, v3, 243753; SI-NEXT:    v_alignbit_b32 v53, v4, v3, 163754; SI-NEXT:    v_alignbit_b32 v55, v4, v3, 83755; SI-NEXT:    v_alignbit_b32 v41, v2, v1, 243756; SI-NEXT:    v_alignbit_b32 v43, v2, v1, 163757; SI-NEXT:    v_alignbit_b32 v45, v2, v1, 83758; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v313759; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill3760; SI-NEXT:    v_lshrrev_b32_e32 v39, 24, v143761; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v143762; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v143763; SI-NEXT:    v_lshrrev_b32_e32 v52, 24, v123764; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v123765; SI-NEXT:    v_lshrrev_b32_e32 v40, 8, v123766; SI-NEXT:    v_lshrrev_b32_e32 v42, 24, v103767; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v103768; SI-NEXT:    v_lshrrev_b32_e32 v46, 8, v103769; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v83770; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v83771; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v83772; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v63773; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v63774; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v63775; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v43776; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v43777; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v43778; SI-NEXT:    s_waitcnt expcnt(0)3779; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v23780; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v23781; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v23782; SI-NEXT:  .LBB12_4: ; %end3783; SI-NEXT:    s_or_b64 exec, exec, s[4:5]3784; SI-NEXT:    v_and_b32_e32 v1, 0xff, v13785; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v453786; SI-NEXT:    v_and_b32_e32 v43, 0xff, v433787; SI-NEXT:    v_or_b32_e32 v1, v1, v453788; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v433789; SI-NEXT:    v_lshlrev_b32_e32 v41, 24, v413790; SI-NEXT:    v_or_b32_e32 v41, v41, v433791; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13792; SI-NEXT:    v_or_b32_e32 v1, v1, v413793; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen3794; SI-NEXT:    s_waitcnt expcnt(0)3795; SI-NEXT:    v_and_b32_e32 v1, 0xff, v23796; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v353797; SI-NEXT:    v_or_b32_e32 v1, v1, v23798; SI-NEXT:    v_and_b32_e32 v2, 0xff, v343799; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23800; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v333801; SI-NEXT:    v_or_b32_e32 v2, v33, v23802; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13803; SI-NEXT:    v_or_b32_e32 v1, v1, v23804; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v03805; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3806; SI-NEXT:    s_waitcnt expcnt(0)3807; SI-NEXT:    v_and_b32_e32 v1, 0xff, v33808; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v553809; SI-NEXT:    v_or_b32_e32 v1, v1, v23810; SI-NEXT:    v_and_b32_e32 v2, 0xff, v533811; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23812; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v513813; SI-NEXT:    v_or_b32_e32 v2, v3, v23814; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13815; SI-NEXT:    v_or_b32_e32 v1, v1, v23816; SI-NEXT:    v_add_i32_e32 v2, vcc, 8, v03817; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3818; SI-NEXT:    s_waitcnt expcnt(0)3819; SI-NEXT:    v_and_b32_e32 v1, 0xff, v43820; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v633821; SI-NEXT:    v_or_b32_e32 v1, v1, v23822; SI-NEXT:    v_and_b32_e32 v2, 0xff, v623823; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23824; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v613825; SI-NEXT:    v_or_b32_e32 v2, v3, v23826; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13827; SI-NEXT:    v_or_b32_e32 v1, v1, v23828; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v03829; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3830; SI-NEXT:    s_waitcnt expcnt(0)3831; SI-NEXT:    v_and_b32_e32 v1, 0xff, v53832; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v483833; SI-NEXT:    v_or_b32_e32 v1, v1, v23834; SI-NEXT:    v_and_b32_e32 v2, 0xff, v383835; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23836; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v373837; SI-NEXT:    v_or_b32_e32 v2, v3, v23838; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13839; SI-NEXT:    v_or_b32_e32 v1, v1, v23840; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v03841; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3842; SI-NEXT:    s_waitcnt expcnt(0)3843; SI-NEXT:    v_and_b32_e32 v1, 0xff, v63844; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v603845; SI-NEXT:    v_or_b32_e32 v1, v1, v23846; SI-NEXT:    v_and_b32_e32 v2, 0xff, v593847; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23848; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v583849; SI-NEXT:    v_or_b32_e32 v2, v3, v23850; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13851; SI-NEXT:    v_or_b32_e32 v1, v1, v23852; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v03853; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3854; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload3855; SI-NEXT:    s_waitcnt expcnt(0)3856; SI-NEXT:    v_and_b32_e32 v1, 0xff, v73857; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload3858; SI-NEXT:    s_waitcnt vmcnt(1)3859; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v23860; SI-NEXT:    v_or_b32_e32 v1, v1, v23861; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload3862; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13863; SI-NEXT:    s_waitcnt vmcnt(1)3864; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v33865; SI-NEXT:    s_waitcnt vmcnt(0)3866; SI-NEXT:    v_and_b32_e32 v2, 0xff, v23867; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23868; SI-NEXT:    v_or_b32_e32 v2, v3, v23869; SI-NEXT:    v_or_b32_e32 v1, v1, v23870; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v03871; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3872; SI-NEXT:    s_waitcnt expcnt(0)3873; SI-NEXT:    v_and_b32_e32 v1, 0xff, v83874; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v573875; SI-NEXT:    v_or_b32_e32 v1, v1, v23876; SI-NEXT:    v_and_b32_e32 v2, 0xff, v563877; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23878; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v473879; SI-NEXT:    v_or_b32_e32 v2, v3, v23880; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13881; SI-NEXT:    v_or_b32_e32 v1, v1, v23882; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v03883; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3884; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload3885; SI-NEXT:    s_waitcnt expcnt(0)3886; SI-NEXT:    v_and_b32_e32 v1, 0xff, v93887; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload3888; SI-NEXT:    s_waitcnt vmcnt(1)3889; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v23890; SI-NEXT:    v_or_b32_e32 v1, v1, v23891; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload3892; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13893; SI-NEXT:    s_waitcnt vmcnt(1)3894; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v33895; SI-NEXT:    s_waitcnt vmcnt(0)3896; SI-NEXT:    v_and_b32_e32 v2, 0xff, v23897; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23898; SI-NEXT:    v_or_b32_e32 v2, v3, v23899; SI-NEXT:    v_or_b32_e32 v1, v1, v23900; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v03901; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3902; SI-NEXT:    s_waitcnt expcnt(0)3903; SI-NEXT:    v_and_b32_e32 v1, 0xff, v103904; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v463905; SI-NEXT:    v_or_b32_e32 v1, v1, v23906; SI-NEXT:    v_and_b32_e32 v2, 0xff, v443907; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23908; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v423909; SI-NEXT:    v_or_b32_e32 v2, v3, v23910; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13911; SI-NEXT:    v_or_b32_e32 v1, v1, v23912; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v03913; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3914; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload3915; SI-NEXT:    s_waitcnt expcnt(0)3916; SI-NEXT:    v_and_b32_e32 v1, 0xff, v113917; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload3918; SI-NEXT:    s_waitcnt vmcnt(1)3919; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v23920; SI-NEXT:    v_or_b32_e32 v1, v1, v23921; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload3922; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13923; SI-NEXT:    s_waitcnt vmcnt(1)3924; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v33925; SI-NEXT:    s_waitcnt vmcnt(0)3926; SI-NEXT:    v_and_b32_e32 v2, 0xff, v23927; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23928; SI-NEXT:    v_or_b32_e32 v2, v3, v23929; SI-NEXT:    v_or_b32_e32 v1, v1, v23930; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v03931; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3932; SI-NEXT:    s_waitcnt expcnt(0)3933; SI-NEXT:    v_and_b32_e32 v1, 0xff, v123934; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v403935; SI-NEXT:    v_or_b32_e32 v1, v1, v23936; SI-NEXT:    v_and_b32_e32 v2, 0xff, v543937; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23938; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v523939; SI-NEXT:    v_or_b32_e32 v2, v3, v23940; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13941; SI-NEXT:    v_or_b32_e32 v1, v1, v23942; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v03943; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3944; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload3945; SI-NEXT:    s_waitcnt expcnt(0)3946; SI-NEXT:    v_and_b32_e32 v1, 0xff, v133947; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload3948; SI-NEXT:    s_waitcnt vmcnt(1)3949; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v23950; SI-NEXT:    v_or_b32_e32 v1, v1, v23951; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload3952; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13953; SI-NEXT:    s_waitcnt vmcnt(1)3954; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v33955; SI-NEXT:    s_waitcnt vmcnt(0)3956; SI-NEXT:    v_and_b32_e32 v2, 0xff, v23957; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23958; SI-NEXT:    v_or_b32_e32 v2, v3, v23959; SI-NEXT:    v_or_b32_e32 v1, v1, v23960; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v03961; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3962; SI-NEXT:    s_waitcnt expcnt(0)3963; SI-NEXT:    v_and_b32_e32 v1, 0xff, v143964; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v503965; SI-NEXT:    v_or_b32_e32 v1, v1, v23966; SI-NEXT:    v_and_b32_e32 v2, 0xff, v493967; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23968; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v393969; SI-NEXT:    v_or_b32_e32 v2, v3, v23970; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13971; SI-NEXT:    v_or_b32_e32 v1, v1, v23972; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v03973; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3974; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload3975; SI-NEXT:    s_waitcnt expcnt(0)3976; SI-NEXT:    v_and_b32_e32 v1, 0xff, v153977; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload3978; SI-NEXT:    s_waitcnt vmcnt(1)3979; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v23980; SI-NEXT:    v_or_b32_e32 v1, v1, v23981; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload3982; SI-NEXT:    s_waitcnt vmcnt(1)3983; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v33984; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13985; SI-NEXT:    s_waitcnt vmcnt(0)3986; SI-NEXT:    v_and_b32_e32 v2, 0xff, v23987; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23988; SI-NEXT:    v_or_b32_e32 v2, v3, v23989; SI-NEXT:    v_or_b32_e32 v1, v1, v23990; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v03991; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen3992; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload3993; SI-NEXT:    s_waitcnt expcnt(0)3994; SI-NEXT:    v_and_b32_e32 v1, 0xff, v163995; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload3996; SI-NEXT:    s_waitcnt vmcnt(1)3997; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v23998; SI-NEXT:    v_or_b32_e32 v1, v1, v23999; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload4000; SI-NEXT:    s_waitcnt vmcnt(1)4001; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34002; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14003; SI-NEXT:    s_waitcnt vmcnt(0)4004; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24005; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24006; SI-NEXT:    v_or_b32_e32 v2, v3, v24007; SI-NEXT:    v_or_b32_e32 v1, v1, v24008; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v04009; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4010; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload4011; SI-NEXT:    s_waitcnt expcnt(0)4012; SI-NEXT:    v_and_b32_e32 v1, 0xff, v174013; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload4014; SI-NEXT:    s_waitcnt vmcnt(1)4015; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24016; SI-NEXT:    v_or_b32_e32 v1, v1, v24017; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload4018; SI-NEXT:    s_waitcnt vmcnt(1)4019; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34020; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14021; SI-NEXT:    s_waitcnt vmcnt(0)4022; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24023; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24024; SI-NEXT:    v_or_b32_e32 v2, v3, v24025; SI-NEXT:    v_or_b32_e32 v1, v1, v24026; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v04027; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4028; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload4029; SI-NEXT:    s_waitcnt expcnt(0)4030; SI-NEXT:    v_and_b32_e32 v1, 0xff, v184031; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload4032; SI-NEXT:    s_waitcnt vmcnt(1)4033; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24034; SI-NEXT:    v_or_b32_e32 v1, v1, v24035; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload4036; SI-NEXT:    s_waitcnt vmcnt(1)4037; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34038; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14039; SI-NEXT:    s_waitcnt vmcnt(0)4040; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24041; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24042; SI-NEXT:    v_or_b32_e32 v2, v3, v24043; SI-NEXT:    v_or_b32_e32 v1, v1, v24044; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v04045; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4046; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload4047; SI-NEXT:    s_waitcnt expcnt(0)4048; SI-NEXT:    v_and_b32_e32 v1, 0xff, v194049; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload4050; SI-NEXT:    s_waitcnt vmcnt(1)4051; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24052; SI-NEXT:    v_or_b32_e32 v1, v1, v24053; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload4054; SI-NEXT:    s_waitcnt vmcnt(1)4055; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34056; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14057; SI-NEXT:    s_waitcnt vmcnt(0)4058; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24059; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24060; SI-NEXT:    v_or_b32_e32 v2, v3, v24061; SI-NEXT:    v_or_b32_e32 v1, v1, v24062; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v04063; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4064; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload4065; SI-NEXT:    s_waitcnt expcnt(0)4066; SI-NEXT:    v_and_b32_e32 v1, 0xff, v204067; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload4068; SI-NEXT:    s_waitcnt vmcnt(1)4069; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24070; SI-NEXT:    v_or_b32_e32 v1, v1, v24071; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload4072; SI-NEXT:    s_waitcnt vmcnt(1)4073; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34074; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14075; SI-NEXT:    s_waitcnt vmcnt(0)4076; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24077; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24078; SI-NEXT:    v_or_b32_e32 v2, v3, v24079; SI-NEXT:    v_or_b32_e32 v1, v1, v24080; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v04081; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4082; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload4083; SI-NEXT:    s_waitcnt expcnt(0)4084; SI-NEXT:    v_and_b32_e32 v1, 0xff, v214085; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload4086; SI-NEXT:    s_waitcnt vmcnt(1)4087; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24088; SI-NEXT:    v_or_b32_e32 v1, v1, v24089; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload4090; SI-NEXT:    s_waitcnt vmcnt(1)4091; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34092; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14093; SI-NEXT:    s_waitcnt vmcnt(0)4094; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24095; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24096; SI-NEXT:    v_or_b32_e32 v2, v3, v24097; SI-NEXT:    v_or_b32_e32 v1, v1, v24098; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v04099; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4100; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload4101; SI-NEXT:    s_waitcnt expcnt(0)4102; SI-NEXT:    v_and_b32_e32 v1, 0xff, v224103; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload4104; SI-NEXT:    s_waitcnt vmcnt(1)4105; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24106; SI-NEXT:    v_or_b32_e32 v1, v1, v24107; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload4108; SI-NEXT:    s_waitcnt vmcnt(1)4109; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34110; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14111; SI-NEXT:    s_waitcnt vmcnt(0)4112; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24113; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24114; SI-NEXT:    v_or_b32_e32 v2, v3, v24115; SI-NEXT:    v_or_b32_e32 v1, v1, v24116; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v04117; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4118; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload4119; SI-NEXT:    s_waitcnt expcnt(0)4120; SI-NEXT:    v_and_b32_e32 v1, 0xff, v234121; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload4122; SI-NEXT:    s_waitcnt vmcnt(1)4123; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24124; SI-NEXT:    v_or_b32_e32 v1, v1, v24125; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload4126; SI-NEXT:    s_waitcnt vmcnt(1)4127; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34128; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14129; SI-NEXT:    s_waitcnt vmcnt(0)4130; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24131; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24132; SI-NEXT:    v_or_b32_e32 v2, v3, v24133; SI-NEXT:    v_or_b32_e32 v1, v1, v24134; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v04135; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4136; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload4137; SI-NEXT:    s_waitcnt expcnt(0)4138; SI-NEXT:    v_and_b32_e32 v1, 0xff, v244139; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload4140; SI-NEXT:    s_waitcnt vmcnt(1)4141; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24142; SI-NEXT:    v_or_b32_e32 v1, v1, v24143; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload4144; SI-NEXT:    s_waitcnt vmcnt(1)4145; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34146; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14147; SI-NEXT:    s_waitcnt vmcnt(0)4148; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24149; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24150; SI-NEXT:    v_or_b32_e32 v2, v3, v24151; SI-NEXT:    v_or_b32_e32 v1, v1, v24152; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v04153; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4154; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload4155; SI-NEXT:    s_waitcnt expcnt(0)4156; SI-NEXT:    v_and_b32_e32 v1, 0xff, v254157; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload4158; SI-NEXT:    s_waitcnt vmcnt(1)4159; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24160; SI-NEXT:    v_or_b32_e32 v1, v1, v24161; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload4162; SI-NEXT:    s_waitcnt vmcnt(1)4163; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34164; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14165; SI-NEXT:    s_waitcnt vmcnt(0)4166; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24167; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24168; SI-NEXT:    v_or_b32_e32 v2, v3, v24169; SI-NEXT:    v_or_b32_e32 v1, v1, v24170; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v04171; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4172; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload4173; SI-NEXT:    s_waitcnt expcnt(0)4174; SI-NEXT:    v_and_b32_e32 v1, 0xff, v264175; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload4176; SI-NEXT:    s_waitcnt vmcnt(1)4177; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24178; SI-NEXT:    v_or_b32_e32 v1, v1, v24179; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload4180; SI-NEXT:    s_waitcnt vmcnt(1)4181; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34182; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14183; SI-NEXT:    s_waitcnt vmcnt(0)4184; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24185; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24186; SI-NEXT:    v_or_b32_e32 v2, v3, v24187; SI-NEXT:    v_or_b32_e32 v1, v1, v24188; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v04189; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4190; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload4191; SI-NEXT:    s_waitcnt expcnt(0)4192; SI-NEXT:    v_and_b32_e32 v1, 0xff, v274193; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload4194; SI-NEXT:    s_waitcnt vmcnt(1)4195; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24196; SI-NEXT:    v_or_b32_e32 v1, v1, v24197; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload4198; SI-NEXT:    s_waitcnt vmcnt(1)4199; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34200; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14201; SI-NEXT:    s_waitcnt vmcnt(0)4202; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24203; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24204; SI-NEXT:    v_or_b32_e32 v2, v3, v24205; SI-NEXT:    v_or_b32_e32 v1, v1, v24206; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v04207; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4208; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload4209; SI-NEXT:    s_waitcnt expcnt(0)4210; SI-NEXT:    v_and_b32_e32 v1, 0xff, v284211; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload4212; SI-NEXT:    s_waitcnt vmcnt(1)4213; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24214; SI-NEXT:    v_or_b32_e32 v1, v1, v24215; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload4216; SI-NEXT:    s_waitcnt vmcnt(1)4217; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34218; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14219; SI-NEXT:    s_waitcnt vmcnt(0)4220; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24221; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24222; SI-NEXT:    v_or_b32_e32 v2, v3, v24223; SI-NEXT:    v_or_b32_e32 v1, v1, v24224; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v04225; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4226; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload4227; SI-NEXT:    s_waitcnt expcnt(0)4228; SI-NEXT:    v_and_b32_e32 v1, 0xff, v294229; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload4230; SI-NEXT:    s_waitcnt vmcnt(1)4231; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24232; SI-NEXT:    v_or_b32_e32 v1, v1, v24233; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload4234; SI-NEXT:    s_waitcnt vmcnt(1)4235; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34236; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14237; SI-NEXT:    s_waitcnt vmcnt(0)4238; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24239; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24240; SI-NEXT:    v_or_b32_e32 v2, v3, v24241; SI-NEXT:    v_or_b32_e32 v1, v1, v24242; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v04243; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4244; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload4245; SI-NEXT:    s_waitcnt expcnt(0)4246; SI-NEXT:    v_and_b32_e32 v1, 0xff, v304247; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload4248; SI-NEXT:    s_waitcnt vmcnt(1)4249; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24250; SI-NEXT:    v_or_b32_e32 v1, v1, v24251; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload4252; SI-NEXT:    s_waitcnt vmcnt(1)4253; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34254; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14255; SI-NEXT:    s_waitcnt vmcnt(0)4256; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24257; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24258; SI-NEXT:    v_or_b32_e32 v2, v3, v24259; SI-NEXT:    v_or_b32_e32 v1, v1, v24260; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v04261; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4262; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload4263; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload4264; SI-NEXT:    s_waitcnt expcnt(0)4265; SI-NEXT:    v_and_b32_e32 v1, 0xff, v324266; SI-NEXT:    s_waitcnt vmcnt(1)4267; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24268; SI-NEXT:    v_or_b32_e32 v1, v1, v24269; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload4270; SI-NEXT:    s_waitcnt vmcnt(1)4271; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34272; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14273; SI-NEXT:    s_waitcnt vmcnt(0)4274; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24275; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24276; SI-NEXT:    v_or_b32_e32 v2, v3, v24277; SI-NEXT:    v_or_b32_e32 v1, v1, v24278; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v04279; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4280; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload4281; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload4282; SI-NEXT:    s_waitcnt expcnt(0)4283; SI-NEXT:    v_and_b32_e32 v1, 0xff, v314284; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v04285; SI-NEXT:    s_waitcnt vmcnt(1)4286; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24287; SI-NEXT:    v_or_b32_e32 v1, v1, v24288; SI-NEXT:    v_and_b32_e32 v2, 0xff, v364289; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24290; SI-NEXT:    s_waitcnt vmcnt(0)4291; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v34292; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14293; SI-NEXT:    v_or_b32_e32 v2, v3, v24294; SI-NEXT:    v_or_b32_e32 v1, v1, v24295; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen4296; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload4297; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload4298; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload4299; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload4300; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload4301; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload4302; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload4303; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload4304; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload4305; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload4306; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload4307; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload4308; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload4309; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload4310; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload4311; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload4312; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)4313; SI-NEXT:    s_setpc_b64 s[30:31]4314;4315; VI-LABEL: bitcast_v32i32_to_v128i8:4316; VI:       ; %bb.0:4317; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)4318; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:44319; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:84320; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s324321; VI-NEXT:    ; implicit-def: $vgpr394322; VI-NEXT:    ; kill: killed $vgpr394323; VI-NEXT:    ; implicit-def: $vgpr394324; VI-NEXT:    ; kill: killed $vgpr394325; VI-NEXT:    ; implicit-def: $vgpr394326; VI-NEXT:    ; kill: killed $vgpr394327; VI-NEXT:    ; implicit-def: $vgpr394328; VI-NEXT:    ; kill: killed $vgpr394329; VI-NEXT:    ; implicit-def: $vgpr394330; VI-NEXT:    ; kill: killed $vgpr394331; VI-NEXT:    ; implicit-def: $vgpr394332; VI-NEXT:    ; kill: killed $vgpr394333; VI-NEXT:    ; implicit-def: $vgpr394334; VI-NEXT:    ; kill: killed $vgpr394335; VI-NEXT:    ; implicit-def: $vgpr394336; VI-NEXT:    ; kill: killed $vgpr394337; VI-NEXT:    ; implicit-def: $vgpr394338; VI-NEXT:    ; kill: killed $vgpr394339; VI-NEXT:    ; implicit-def: $vgpr394340; VI-NEXT:    ; kill: killed $vgpr394341; VI-NEXT:    ; implicit-def: $vgpr394342; VI-NEXT:    ; kill: killed $vgpr394343; VI-NEXT:    ; implicit-def: $vgpr394344; VI-NEXT:    ; kill: killed $vgpr394345; VI-NEXT:    ; implicit-def: $vgpr394346; VI-NEXT:    ; kill: killed $vgpr394347; VI-NEXT:    ; implicit-def: $vgpr394348; VI-NEXT:    ; kill: killed $vgpr394349; VI-NEXT:    ; implicit-def: $vgpr394350; VI-NEXT:    ; kill: killed $vgpr394351; VI-NEXT:    ; implicit-def: $vgpr394352; VI-NEXT:    ; kill: killed $vgpr394353; VI-NEXT:    ; implicit-def: $vgpr394354; VI-NEXT:    ; kill: killed $vgpr394355; VI-NEXT:    ; implicit-def: $vgpr394356; VI-NEXT:    ; kill: killed $vgpr394357; VI-NEXT:    ; implicit-def: $vgpr394358; VI-NEXT:    ; kill: killed $vgpr394359; VI-NEXT:    ; implicit-def: $vgpr394360; VI-NEXT:    ; kill: killed $vgpr394361; VI-NEXT:    ; implicit-def: $vgpr394362; VI-NEXT:    ; kill: killed $vgpr394363; VI-NEXT:    ; implicit-def: $vgpr394364; VI-NEXT:    ; kill: killed $vgpr394365; VI-NEXT:    ; implicit-def: $vgpr394366; VI-NEXT:    ; kill: killed $vgpr394367; VI-NEXT:    ; implicit-def: $vgpr394368; VI-NEXT:    ; kill: killed $vgpr394369; VI-NEXT:    ; implicit-def: $vgpr394370; VI-NEXT:    ; kill: killed $vgpr394371; VI-NEXT:    ; implicit-def: $vgpr394372; VI-NEXT:    ; kill: killed $vgpr394373; VI-NEXT:    ; implicit-def: $vgpr394374; VI-NEXT:    ; kill: killed $vgpr394375; VI-NEXT:    ; implicit-def: $vgpr394376; VI-NEXT:    ; kill: killed $vgpr394377; VI-NEXT:    ; implicit-def: $vgpr394378; VI-NEXT:    ; kill: killed $vgpr394379; VI-NEXT:    ; implicit-def: $vgpr394380; VI-NEXT:    ; kill: killed $vgpr394381; VI-NEXT:    ; implicit-def: $vgpr394382; VI-NEXT:    ; kill: killed $vgpr394383; VI-NEXT:    ; implicit-def: $vgpr394384; VI-NEXT:    ; kill: killed $vgpr394385; VI-NEXT:    ; implicit-def: $vgpr394386; VI-NEXT:    ; kill: killed $vgpr394387; VI-NEXT:    ; implicit-def: $vgpr394388; VI-NEXT:    ; kill: killed $vgpr394389; VI-NEXT:    ; implicit-def: $vgpr394390; VI-NEXT:    ; kill: killed $vgpr394391; VI-NEXT:    ; implicit-def: $vgpr394392; VI-NEXT:    ; kill: killed $vgpr394393; VI-NEXT:    ; implicit-def: $vgpr394394; VI-NEXT:    ; kill: killed $vgpr394395; VI-NEXT:    ; implicit-def: $vgpr394396; VI-NEXT:    ; kill: killed $vgpr394397; VI-NEXT:    ; implicit-def: $vgpr394398; VI-NEXT:    ; kill: killed $vgpr394399; VI-NEXT:    ; implicit-def: $vgpr394400; VI-NEXT:    ; kill: killed $vgpr394401; VI-NEXT:    ; implicit-def: $vgpr394402; VI-NEXT:    ; kill: killed $vgpr394403; VI-NEXT:    ; implicit-def: $vgpr394404; VI-NEXT:    ; kill: killed $vgpr394405; VI-NEXT:    ; implicit-def: $vgpr394406; VI-NEXT:    ; kill: killed $vgpr394407; VI-NEXT:    ; implicit-def: $vgpr394408; VI-NEXT:    ; kill: killed $vgpr394409; VI-NEXT:    ; implicit-def: $vgpr394410; VI-NEXT:    ; kill: killed $vgpr394411; VI-NEXT:    ; implicit-def: $vgpr394412; VI-NEXT:    ; kill: killed $vgpr394413; VI-NEXT:    ; implicit-def: $vgpr394414; VI-NEXT:    ; kill: killed $vgpr394415; VI-NEXT:    ; implicit-def: $vgpr394416; VI-NEXT:    ; kill: killed $vgpr394417; VI-NEXT:    ; implicit-def: $vgpr394418; VI-NEXT:    ; kill: killed $vgpr394419; VI-NEXT:    ; implicit-def: $vgpr394420; VI-NEXT:    ; kill: killed $vgpr394421; VI-NEXT:    ; implicit-def: $vgpr394422; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill4423; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill4424; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill4425; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill4426; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill4427; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill4428; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill4429; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill4430; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill4431; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill4432; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill4433; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill4434; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill4435; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill4436; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill4437; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill4438; VI-NEXT:    ; implicit-def: $vgpr594439; VI-NEXT:    ; kill: killed $vgpr394440; VI-NEXT:    ; implicit-def: $vgpr394441; VI-NEXT:    ; implicit-def: $vgpr584442; VI-NEXT:    ; kill: killed $vgpr394443; VI-NEXT:    ; implicit-def: $vgpr394444; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill4445; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill4446; VI-NEXT:    ; implicit-def: $vgpr584447; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill4448; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill4449; VI-NEXT:    ; implicit-def: $vgpr584450; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill4451; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill4452; VI-NEXT:    ; implicit-def: $vgpr584453; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill4454; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill4455; VI-NEXT:    ; implicit-def: $vgpr584456; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill4457; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill4458; VI-NEXT:    ; implicit-def: $vgpr584459; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill4460; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill4461; VI-NEXT:    ; implicit-def: $vgpr584462; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill4463; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill4464; VI-NEXT:    ; implicit-def: $vgpr584465; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill4466; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill4467; VI-NEXT:    ; implicit-def: $vgpr584468; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill4469; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill4470; VI-NEXT:    ; implicit-def: $vgpr584471; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill4472; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill4473; VI-NEXT:    ; implicit-def: $vgpr584474; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill4475; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill4476; VI-NEXT:    ; implicit-def: $vgpr584477; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill4478; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill4479; VI-NEXT:    ; implicit-def: $vgpr484480; VI-NEXT:    ; implicit-def: $vgpr534481; VI-NEXT:    ; implicit-def: $vgpr574482; VI-NEXT:    ; implicit-def: $vgpr564483; VI-NEXT:    ; implicit-def: $vgpr344484; VI-NEXT:    ; implicit-def: $vgpr384485; VI-NEXT:    ; implicit-def: $vgpr524486; VI-NEXT:    ; implicit-def: $vgpr474487; VI-NEXT:    ; implicit-def: $vgpr464488; VI-NEXT:    ; implicit-def: $vgpr454489; VI-NEXT:    ; implicit-def: $vgpr444490; VI-NEXT:    ; implicit-def: $vgpr514491; VI-NEXT:    ; implicit-def: $vgpr374492; VI-NEXT:    ; implicit-def: $vgpr434493; VI-NEXT:    ; implicit-def: $vgpr504494; VI-NEXT:    ; implicit-def: $vgpr634495; VI-NEXT:    ; implicit-def: $vgpr364496; VI-NEXT:    ; implicit-def: $vgpr624497; VI-NEXT:    ; implicit-def: $vgpr614498; VI-NEXT:    ; implicit-def: $vgpr494499; VI-NEXT:    ; implicit-def: $vgpr604500; VI-NEXT:    ; implicit-def: $vgpr354501; VI-NEXT:    ; kill: killed $vgpr394502; VI-NEXT:    ; implicit-def: $vgpr424503; VI-NEXT:    ; implicit-def: $vgpr554504; VI-NEXT:    ; implicit-def: $vgpr414505; VI-NEXT:    ; implicit-def: $vgpr404506; VI-NEXT:    ; implicit-def: $vgpr394507; VI-NEXT:    ; implicit-def: $vgpr544508; VI-NEXT:    ; implicit-def: $vgpr584509; VI-NEXT:    s_waitcnt vmcnt(14)4510; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v334511; VI-NEXT:    ; implicit-def: $vgpr334512; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc4513; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]4514; VI-NEXT:    s_cbranch_execz .LBB12_24515; VI-NEXT:  ; %bb.1: ; %cmp.false4516; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v324517; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill4518; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v314519; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill4520; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v314521; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill4522; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v304523; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill4524; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v304525; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill4526; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v304527; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill4528; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v294529; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill4530; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v294531; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill4532; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v284533; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill4534; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v284535; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill4536; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v284537; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill4538; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v274539; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill4540; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v264541; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill4542; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v264543; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[31:32]4544; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill4545; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v264546; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill4547; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v254548; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill4549; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill4550; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill4551; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v254552; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[29:30]4553; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill4554; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v244555; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill4556; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v244557; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill4558; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill4559; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill4560; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v244561; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[27:28]4562; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill4563; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v234564; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill4565; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v234566; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill4567; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill4568; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill4569; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v224570; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[25:26]4571; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill4572; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v224573; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill4574; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v224575; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill4576; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill4577; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill4578; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v214579; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]4580; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill4581; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v214582; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill4583; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v204584; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill4585; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill4586; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill4587; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v204588; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[21:22]4589; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill4590; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v204591; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill4592; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v194593; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill4594; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill4595; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill4596; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v194597; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[19:20]4598; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill4599; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v184600; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill4601; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v184602; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill4603; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill4604; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill4605; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v184606; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]4607; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill4608; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v174609; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill4610; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v174611; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill4612; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill4613; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill4614; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v164615; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[15:16]4616; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill4617; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v164618; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill4619; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v164620; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill4621; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill4622; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill4623; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v154624; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[13:14]4625; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill4626; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v154627; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill4628; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v144629; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill4630; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill4631; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill4632; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v144633; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[11:12]4634; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill4635; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v144636; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill4637; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v134638; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill4639; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill4640; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill4641; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v134642; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[9:10]4643; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill4644; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v124645; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill4646; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v124647; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill4648; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill4649; VI-NEXT:    v_lshrrev_b32_e32 v39, 24, v324650; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill4651; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v124652; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]4653; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill4654; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v114655; VI-NEXT:    v_mov_b32_e32 v55, v394656; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]4657; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill4658; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v114659; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]4660; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill4661; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v104662; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]4663; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v274664; VI-NEXT:    v_lshrrev_b32_e32 v59, 24, v104665; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill4666; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v104667; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v94668; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v94669; VI-NEXT:    v_lshrrev_b32_e32 v61, 24, v84670; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v84671; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v84672; VI-NEXT:    v_lshrrev_b32_e32 v63, 16, v74673; VI-NEXT:    v_lshrrev_b32_e32 v50, 8, v74674; VI-NEXT:    v_lshrrev_b32_e32 v43, 24, v64675; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v64676; VI-NEXT:    v_lshrrev_b32_e32 v37, 8, v64677; VI-NEXT:    v_lshrrev_b32_e32 v51, 16, v54678; VI-NEXT:    v_lshrrev_b32_e32 v44, 8, v54679; VI-NEXT:    v_lshrrev_b32_e32 v45, 24, v44680; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v44681; VI-NEXT:    v_lshrrev_b32_e32 v47, 8, v44682; VI-NEXT:    v_lshrrev_b32_e32 v52, 16, v34683; VI-NEXT:    v_lshrrev_b32_e32 v38, 8, v34684; VI-NEXT:    v_lshrrev_b32_e32 v34, 24, v24685; VI-NEXT:    v_lshrrev_b32_e32 v56, 16, v24686; VI-NEXT:    v_lshrrev_b32_e32 v57, 8, v24687; VI-NEXT:    v_lshrrev_b32_e32 v53, 16, v14688; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v14689; VI-NEXT:    v_lshrrev_b32_e32 v42, 16, v324690; VI-NEXT:  .LBB12_2: ; %Flow4691; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]4692; VI-NEXT:    s_cbranch_execz .LBB12_44693; VI-NEXT:  ; %bb.3: ; %cmp.true4694; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v324695; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v314696; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]4697; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v304698; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v294699; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill4700; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill4701; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]4702; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v284703; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v274704; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill4705; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill4706; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]4707; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v264708; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v254709; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill4710; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill4711; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]4712; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v244713; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v234714; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill4715; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill4716; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]4717; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v224718; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v214719; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill4720; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill4721; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]4722; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v204723; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v194724; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill4725; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill4726; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]4727; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v184728; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v174729; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill4730; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill4731; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]4732; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v164733; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v154734; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill4735; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill4736; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]4737; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v144738; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v134739; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill4740; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill4741; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]4742; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v124743; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v114744; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill4745; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill4746; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]4747; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v104748; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v94749; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill4750; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill4751; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]4752; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill4753; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill4754; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v324755; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill4756; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v314757; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill4758; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v314759; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill4760; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v304761; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill4762; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v304763; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill4764; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v304765; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill4766; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v294767; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill4768; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v294769; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill4770; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v284771; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill4772; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v284773; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill4774; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v284775; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill4776; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v274777; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill4778; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v264779; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill4780; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v264781; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill4782; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v264783; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill4784; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v254785; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill4786; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v254787; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill4788; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v244789; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill4790; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v244791; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill4792; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v244793; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill4794; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v234795; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill4796; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v234797; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill4798; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v224799; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill4800; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v224801; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill4802; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v224803; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill4804; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v214805; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill4806; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v214807; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill4808; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v204809; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill4810; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v204811; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill4812; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v204813; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill4814; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v194815; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill4816; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v194817; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill4818; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v184819; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill4820; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v184821; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill4822; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v184823; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill4824; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v174825; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill4826; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v174827; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill4828; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v164829; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill4830; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v164831; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill4832; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v164833; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill4834; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v154835; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill4836; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v154837; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill4838; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v144839; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill4840; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v144841; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill4842; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v144843; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill4844; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v134845; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill4846; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v134847; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill4848; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v124849; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill4850; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v124851; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v64852; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v54853; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill4854; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v124855; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v44856; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v34857; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]4858; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill4859; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v114860; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v24861; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v14862; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v84863; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v74864; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]4865; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill4866; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v114867; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]4868; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]4869; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill4870; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v104871; VI-NEXT:    v_lshrrev_b32_e32 v55, 24, v324872; VI-NEXT:    v_lshrrev_b32_e32 v42, 16, v324873; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v274874; VI-NEXT:    v_lshrrev_b32_e32 v59, 24, v104875; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill4876; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v104877; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v94878; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v94879; VI-NEXT:    v_lshrrev_b32_e32 v61, 24, v84880; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v84881; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v84882; VI-NEXT:    v_lshrrev_b32_e32 v63, 16, v74883; VI-NEXT:    v_lshrrev_b32_e32 v50, 8, v74884; VI-NEXT:    v_lshrrev_b32_e32 v43, 24, v64885; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v64886; VI-NEXT:    v_lshrrev_b32_e32 v37, 8, v64887; VI-NEXT:    v_lshrrev_b32_e32 v51, 16, v54888; VI-NEXT:    v_lshrrev_b32_e32 v44, 8, v54889; VI-NEXT:    v_lshrrev_b32_e32 v45, 24, v44890; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v44891; VI-NEXT:    v_lshrrev_b32_e32 v47, 8, v44892; VI-NEXT:    v_lshrrev_b32_e32 v52, 16, v34893; VI-NEXT:    v_lshrrev_b32_e32 v38, 8, v34894; VI-NEXT:    v_lshrrev_b32_e32 v34, 24, v24895; VI-NEXT:    v_lshrrev_b32_e32 v56, 16, v24896; VI-NEXT:    v_lshrrev_b32_e32 v57, 8, v24897; VI-NEXT:    v_lshrrev_b32_e32 v53, 16, v14898; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v14899; VI-NEXT:  .LBB12_4: ; %end4900; VI-NEXT:    s_or_b64 exec, exec, s[4:5]4901; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v484902; VI-NEXT:    v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4903; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v574904; VI-NEXT:    v_or_b32_sdwa v2, v2, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4905; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v414906; VI-NEXT:    v_or_b32_sdwa v48, v53, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4907; VI-NEXT:    v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4908; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen4909; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v344910; VI-NEXT:    v_or_b32_sdwa v1, v56, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4911; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4912; VI-NEXT:    v_add_u32_e32 v2, vcc, 4, v04913; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4914; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v404915; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v384916; VI-NEXT:    v_or_b32_sdwa v1, v52, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4917; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4918; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4919; VI-NEXT:    v_add_u32_e32 v2, vcc, 8, v04920; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4921; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v474922; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v454923; VI-NEXT:    v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4924; VI-NEXT:    v_or_b32_sdwa v2, v46, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4925; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4926; VI-NEXT:    v_add_u32_e32 v2, vcc, 12, v04927; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4928; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v394929; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v444930; VI-NEXT:    v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4931; VI-NEXT:    v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4932; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4933; VI-NEXT:    v_add_u32_e32 v2, vcc, 16, v04934; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4935; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v374936; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v434937; VI-NEXT:    v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4938; VI-NEXT:    v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4939; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4940; VI-NEXT:    v_add_u32_e32 v2, vcc, 20, v04941; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4942; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v544943; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v504944; VI-NEXT:    v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4945; VI-NEXT:    v_or_b32_sdwa v2, v7, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4946; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4947; VI-NEXT:    v_add_u32_e32 v2, vcc, 24, v04948; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4949; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v364950; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v614951; VI-NEXT:    v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4952; VI-NEXT:    v_or_b32_sdwa v2, v62, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4953; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4954; VI-NEXT:    v_add_u32_e32 v2, vcc, 28, v04955; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4956; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload4957; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload4958; VI-NEXT:    s_waitcnt vmcnt(0)4959; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v494960; VI-NEXT:    v_or_b32_sdwa v2, v9, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4961; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v14962; VI-NEXT:    v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4963; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4964; VI-NEXT:    v_add_u32_e32 v2, vcc, 32, v04965; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4966; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload4967; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v354968; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v594969; VI-NEXT:    v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4970; VI-NEXT:    s_waitcnt vmcnt(0)4971; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4972; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4973; VI-NEXT:    v_add_u32_e32 v2, vcc, 36, v04974; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4975; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload4976; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload4977; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload4978; VI-NEXT:    s_waitcnt vmcnt(2)4979; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v14980; VI-NEXT:    s_waitcnt vmcnt(0)4981; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4982; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload4983; VI-NEXT:    s_waitcnt vmcnt(0)4984; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v24985; VI-NEXT:    v_or_b32_sdwa v2, v11, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4986; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD4987; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v04988; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen4989; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload4990; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload4991; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload4992; VI-NEXT:    s_waitcnt vmcnt(2)4993; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v14994; VI-NEXT:    s_waitcnt vmcnt(1)4995; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v24996; VI-NEXT:    v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4997; VI-NEXT:    s_waitcnt vmcnt(0)4998; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD4999; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5000; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v05001; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5002; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload5003; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload5004; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload5005; VI-NEXT:    s_waitcnt vmcnt(2)5006; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15007; VI-NEXT:    s_waitcnt vmcnt(0)5008; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5009; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload5010; VI-NEXT:    s_waitcnt vmcnt(0)5011; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25012; VI-NEXT:    v_or_b32_sdwa v2, v13, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5013; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5014; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v05015; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5016; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload5017; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload5018; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload5019; VI-NEXT:    s_waitcnt vmcnt(2)5020; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15021; VI-NEXT:    s_waitcnt vmcnt(1)5022; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25023; VI-NEXT:    v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5024; VI-NEXT:    s_waitcnt vmcnt(0)5025; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5026; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5027; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v05028; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5029; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload5030; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload5031; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload5032; VI-NEXT:    s_waitcnt vmcnt(2)5033; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15034; VI-NEXT:    s_waitcnt vmcnt(0)5035; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5036; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload5037; VI-NEXT:    s_waitcnt vmcnt(0)5038; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25039; VI-NEXT:    v_or_b32_sdwa v2, v15, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5040; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5041; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v05042; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5043; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload5044; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload5045; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload5046; VI-NEXT:    s_waitcnt vmcnt(2)5047; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15048; VI-NEXT:    s_waitcnt vmcnt(1)5049; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25050; VI-NEXT:    v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5051; VI-NEXT:    s_waitcnt vmcnt(0)5052; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5053; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5054; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v05055; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5056; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload5057; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload5058; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload5059; VI-NEXT:    s_waitcnt vmcnt(2)5060; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15061; VI-NEXT:    s_waitcnt vmcnt(0)5062; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5063; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload5064; VI-NEXT:    s_waitcnt vmcnt(0)5065; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25066; VI-NEXT:    v_or_b32_sdwa v2, v17, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5067; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5068; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v05069; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5070; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload5071; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload5072; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload5073; VI-NEXT:    s_waitcnt vmcnt(2)5074; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15075; VI-NEXT:    s_waitcnt vmcnt(1)5076; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25077; VI-NEXT:    v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5078; VI-NEXT:    s_waitcnt vmcnt(0)5079; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5080; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5081; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v05082; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5083; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload5084; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload5085; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload5086; VI-NEXT:    s_waitcnt vmcnt(2)5087; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15088; VI-NEXT:    s_waitcnt vmcnt(0)5089; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5090; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload5091; VI-NEXT:    s_waitcnt vmcnt(0)5092; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25093; VI-NEXT:    v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5094; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5095; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v05096; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5097; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload5098; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload5099; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload5100; VI-NEXT:    s_waitcnt vmcnt(2)5101; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15102; VI-NEXT:    s_waitcnt vmcnt(1)5103; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25104; VI-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5105; VI-NEXT:    s_waitcnt vmcnt(0)5106; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5107; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5108; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v05109; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5110; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload5111; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload5112; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload5113; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload5114; VI-NEXT:    s_waitcnt vmcnt(2)5115; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25116; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15117; VI-NEXT:    v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5118; VI-NEXT:    s_waitcnt vmcnt(0)5119; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5120; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5121; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v05122; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5123; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload5124; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload5125; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload5126; VI-NEXT:    s_waitcnt vmcnt(2)5127; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15128; VI-NEXT:    s_waitcnt vmcnt(1)5129; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25130; VI-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5131; VI-NEXT:    s_waitcnt vmcnt(0)5132; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5133; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5134; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v05135; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5136; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload5137; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload5138; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload5139; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload5140; VI-NEXT:    s_waitcnt vmcnt(3)5141; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15142; VI-NEXT:    s_waitcnt vmcnt(2)5143; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25144; VI-NEXT:    v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5145; VI-NEXT:    s_waitcnt vmcnt(0)5146; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5147; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5148; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v05149; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5150; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload5151; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload5152; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload5153; VI-NEXT:    s_waitcnt vmcnt(2)5154; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15155; VI-NEXT:    s_waitcnt vmcnt(1)5156; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25157; VI-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5158; VI-NEXT:    s_waitcnt vmcnt(0)5159; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5160; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5161; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v05162; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5163; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload5164; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload5165; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload5166; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload5167; VI-NEXT:    s_waitcnt vmcnt(3)5168; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15169; VI-NEXT:    s_waitcnt vmcnt(2)5170; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25171; VI-NEXT:    v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5172; VI-NEXT:    s_waitcnt vmcnt(0)5173; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5174; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5175; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v05176; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5177; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload5178; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload5179; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload5180; VI-NEXT:    s_waitcnt vmcnt(2)5181; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15182; VI-NEXT:    s_waitcnt vmcnt(1)5183; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25184; VI-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5185; VI-NEXT:    s_waitcnt vmcnt(0)5186; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5187; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5188; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v05189; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5190; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload5191; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload5192; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload5193; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v585194; VI-NEXT:    v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5195; VI-NEXT:    s_waitcnt vmcnt(2)5196; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25197; VI-NEXT:    s_waitcnt vmcnt(0)5198; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5199; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5200; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v05201; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5202; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload5203; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload5204; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload5205; VI-NEXT:    s_waitcnt vmcnt(2)5206; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15207; VI-NEXT:    s_waitcnt vmcnt(1)5208; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25209; VI-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5210; VI-NEXT:    s_waitcnt vmcnt(0)5211; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5212; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5213; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v05214; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5215; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload5216; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload5217; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload5218; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload5219; VI-NEXT:    s_waitcnt vmcnt(3)5220; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15221; VI-NEXT:    s_waitcnt vmcnt(2)5222; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25223; VI-NEXT:    v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5224; VI-NEXT:    s_waitcnt vmcnt(0)5225; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5226; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5227; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v05228; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5229; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload5230; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload5231; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload5232; VI-NEXT:    s_waitcnt vmcnt(2)5233; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15234; VI-NEXT:    s_waitcnt vmcnt(1)5235; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25236; VI-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5237; VI-NEXT:    s_waitcnt vmcnt(0)5238; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5239; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5240; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v05241; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5242; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload5243; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload5244; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload5245; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload5246; VI-NEXT:    s_waitcnt vmcnt(3)5247; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15248; VI-NEXT:    s_waitcnt vmcnt(2)5249; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v25250; VI-NEXT:    v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5251; VI-NEXT:    s_waitcnt vmcnt(0)5252; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5253; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5254; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v05255; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen5256; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload5257; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v555258; VI-NEXT:    v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5259; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v05260; VI-NEXT:    s_waitcnt vmcnt(0)5261; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v15262; VI-NEXT:    v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5263; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5264; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen5265; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload5266; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload5267; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload5268; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload5269; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload5270; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload5271; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload5272; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload5273; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload5274; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload5275; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload5276; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload5277; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload5278; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload5279; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload5280; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload5281; VI-NEXT:    s_waitcnt vmcnt(0)5282; VI-NEXT:    s_setpc_b64 s[30:31]5283;5284; GFX9-LABEL: bitcast_v32i32_to_v128i8:5285; GFX9:       ; %bb.0:5286; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)5287; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:45288; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:85289; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s325290; GFX9-NEXT:    s_nop 05291; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill5292; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill5293; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill5294; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill5295; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill5296; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill5297; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill5298; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill5299; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill5300; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill5301; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill5302; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill5303; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill5304; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill5305; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill5306; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill5307; GFX9-NEXT:    ; implicit-def: $vgpr405308; GFX9-NEXT:    ; kill: killed $vgpr405309; GFX9-NEXT:    ; implicit-def: $vgpr405310; GFX9-NEXT:    ; kill: killed $vgpr405311; GFX9-NEXT:    ; implicit-def: $vgpr405312; GFX9-NEXT:    ; kill: killed $vgpr405313; GFX9-NEXT:    ; implicit-def: $vgpr405314; GFX9-NEXT:    ; kill: killed $vgpr405315; GFX9-NEXT:    ; implicit-def: $vgpr405316; GFX9-NEXT:    ; kill: killed $vgpr405317; GFX9-NEXT:    ; implicit-def: $vgpr405318; GFX9-NEXT:    ; kill: killed $vgpr405319; GFX9-NEXT:    ; implicit-def: $vgpr405320; GFX9-NEXT:    ; kill: killed $vgpr405321; GFX9-NEXT:    ; implicit-def: $vgpr405322; GFX9-NEXT:    ; kill: killed $vgpr405323; GFX9-NEXT:    ; implicit-def: $vgpr405324; GFX9-NEXT:    ; kill: killed $vgpr405325; GFX9-NEXT:    ; implicit-def: $vgpr405326; GFX9-NEXT:    ; kill: killed $vgpr405327; GFX9-NEXT:    ; implicit-def: $vgpr405328; GFX9-NEXT:    ; kill: killed $vgpr405329; GFX9-NEXT:    ; implicit-def: $vgpr405330; GFX9-NEXT:    ; kill: killed $vgpr405331; GFX9-NEXT:    ; implicit-def: $vgpr405332; GFX9-NEXT:    ; kill: killed $vgpr405333; GFX9-NEXT:    ; implicit-def: $vgpr405334; GFX9-NEXT:    ; kill: killed $vgpr405335; GFX9-NEXT:    ; implicit-def: $vgpr405336; GFX9-NEXT:    ; kill: killed $vgpr405337; GFX9-NEXT:    ; implicit-def: $vgpr405338; GFX9-NEXT:    ; kill: killed $vgpr405339; GFX9-NEXT:    ; implicit-def: $vgpr405340; GFX9-NEXT:    ; kill: killed $vgpr405341; GFX9-NEXT:    ; implicit-def: $vgpr405342; GFX9-NEXT:    ; kill: killed $vgpr405343; GFX9-NEXT:    ; implicit-def: $vgpr405344; GFX9-NEXT:    ; kill: killed $vgpr405345; GFX9-NEXT:    ; implicit-def: $vgpr405346; GFX9-NEXT:    ; kill: killed $vgpr405347; GFX9-NEXT:    ; implicit-def: $vgpr405348; GFX9-NEXT:    ; kill: killed $vgpr405349; GFX9-NEXT:    ; implicit-def: $vgpr405350; GFX9-NEXT:    ; kill: killed $vgpr405351; GFX9-NEXT:    ; implicit-def: $vgpr405352; GFX9-NEXT:    ; kill: killed $vgpr405353; GFX9-NEXT:    ; implicit-def: $vgpr405354; GFX9-NEXT:    ; kill: killed $vgpr405355; GFX9-NEXT:    ; implicit-def: $vgpr405356; GFX9-NEXT:    ; kill: killed $vgpr405357; GFX9-NEXT:    ; implicit-def: $vgpr405358; GFX9-NEXT:    ; kill: killed $vgpr405359; GFX9-NEXT:    ; implicit-def: $vgpr405360; GFX9-NEXT:    ; kill: killed $vgpr405361; GFX9-NEXT:    ; implicit-def: $vgpr405362; GFX9-NEXT:    ; kill: killed $vgpr405363; GFX9-NEXT:    ; implicit-def: $vgpr405364; GFX9-NEXT:    ; kill: killed $vgpr405365; GFX9-NEXT:    ; implicit-def: $vgpr405366; GFX9-NEXT:    ; kill: killed $vgpr405367; GFX9-NEXT:    ; implicit-def: $vgpr405368; GFX9-NEXT:    ; kill: killed $vgpr405369; GFX9-NEXT:    ; implicit-def: $vgpr405370; GFX9-NEXT:    ; kill: killed $vgpr405371; GFX9-NEXT:    ; implicit-def: $vgpr405372; GFX9-NEXT:    ; kill: killed $vgpr405373; GFX9-NEXT:    ; implicit-def: $vgpr405374; GFX9-NEXT:    ; kill: killed $vgpr405375; GFX9-NEXT:    ; implicit-def: $vgpr405376; GFX9-NEXT:    ; kill: killed $vgpr405377; GFX9-NEXT:    ; implicit-def: $vgpr405378; GFX9-NEXT:    ; kill: killed $vgpr405379; GFX9-NEXT:    ; implicit-def: $vgpr405380; GFX9-NEXT:    ; kill: killed $vgpr405381; GFX9-NEXT:    ; implicit-def: $vgpr405382; GFX9-NEXT:    ; kill: killed $vgpr405383; GFX9-NEXT:    ; implicit-def: $vgpr405384; GFX9-NEXT:    ; kill: killed $vgpr405385; GFX9-NEXT:    ; implicit-def: $vgpr405386; GFX9-NEXT:    ; kill: killed $vgpr405387; GFX9-NEXT:    ; implicit-def: $vgpr405388; GFX9-NEXT:    ; kill: killed $vgpr405389; GFX9-NEXT:    ; implicit-def: $vgpr405390; GFX9-NEXT:    ; kill: killed $vgpr405391; GFX9-NEXT:    ; implicit-def: $vgpr405392; GFX9-NEXT:    ; implicit-def: $vgpr365393; GFX9-NEXT:    ; kill: killed $vgpr405394; GFX9-NEXT:    ; implicit-def: $vgpr405395; GFX9-NEXT:    ; kill: killed $vgpr365396; GFX9-NEXT:    ; implicit-def: $vgpr365397; GFX9-NEXT:    ; implicit-def: $vgpr505398; GFX9-NEXT:    ; kill: killed $vgpr405399; GFX9-NEXT:    ; implicit-def: $vgpr405400; GFX9-NEXT:    ; implicit-def: $vgpr435401; GFX9-NEXT:    ; kill: killed $vgpr365402; GFX9-NEXT:    ; implicit-def: $vgpr365403; GFX9-NEXT:    ; kill: killed $vgpr505404; GFX9-NEXT:    ; implicit-def: $vgpr505405; GFX9-NEXT:    ; kill: killed $vgpr405406; GFX9-NEXT:    ; implicit-def: $vgpr405407; GFX9-NEXT:    ; implicit-def: $vgpr425408; GFX9-NEXT:    ; implicit-def: $vgpr565409; GFX9-NEXT:    ; implicit-def: $vgpr445410; GFX9-NEXT:    ; implicit-def: $vgpr385411; GFX9-NEXT:    ; implicit-def: $vgpr535412; GFX9-NEXT:    ; implicit-def: $vgpr395413; GFX9-NEXT:    ; implicit-def: $vgpr635414; GFX9-NEXT:    ; implicit-def: $vgpr475415; GFX9-NEXT:    ; implicit-def: $vgpr515416; GFX9-NEXT:    ; implicit-def: $vgpr625417; GFX9-NEXT:    ; implicit-def: $vgpr465418; GFX9-NEXT:    ; implicit-def: $vgpr615419; GFX9-NEXT:    ; implicit-def: $vgpr455420; GFX9-NEXT:    ; implicit-def: $vgpr495421; GFX9-NEXT:    ; implicit-def: $vgpr355422; GFX9-NEXT:    ; implicit-def: $vgpr555423; GFX9-NEXT:    ; implicit-def: $vgpr585424; GFX9-NEXT:    ; implicit-def: $vgpr545425; GFX9-NEXT:    ; implicit-def: $vgpr485426; GFX9-NEXT:    ; implicit-def: $vgpr345427; GFX9-NEXT:    ; implicit-def: $vgpr375428; GFX9-NEXT:    ; implicit-def: $vgpr525429; GFX9-NEXT:    ; kill: killed $vgpr365430; GFX9-NEXT:    ; implicit-def: $vgpr365431; GFX9-NEXT:    ; kill: killed $vgpr505432; GFX9-NEXT:    ; implicit-def: $vgpr505433; GFX9-NEXT:    ; implicit-def: $vgpr605434; GFX9-NEXT:    ; implicit-def: $vgpr595435; GFX9-NEXT:    ; implicit-def: $vgpr575436; GFX9-NEXT:    ; kill: killed $vgpr405437; GFX9-NEXT:    ; implicit-def: $vgpr415438; GFX9-NEXT:    ; implicit-def: $vgpr405439; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill5440; GFX9-NEXT:    s_nop 05441; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill5442; GFX9-NEXT:    ; implicit-def: $vgpr425443; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill5444; GFX9-NEXT:    s_nop 05445; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill5446; GFX9-NEXT:    ; implicit-def: $vgpr425447; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill5448; GFX9-NEXT:    s_nop 05449; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill5450; GFX9-NEXT:    ; implicit-def: $vgpr425451; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill5452; GFX9-NEXT:    s_nop 05453; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill5454; GFX9-NEXT:    ; implicit-def: $vgpr425455; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill5456; GFX9-NEXT:    s_nop 05457; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill5458; GFX9-NEXT:    ; implicit-def: $vgpr425459; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill5460; GFX9-NEXT:    s_nop 05461; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill5462; GFX9-NEXT:    ; implicit-def: $vgpr425463; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill5464; GFX9-NEXT:    s_nop 05465; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill5466; GFX9-NEXT:    ; implicit-def: $vgpr425467; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill5468; GFX9-NEXT:    s_nop 05469; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill5470; GFX9-NEXT:    ; implicit-def: $vgpr425471; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill5472; GFX9-NEXT:    s_nop 05473; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill5474; GFX9-NEXT:    ; implicit-def: $vgpr425475; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill5476; GFX9-NEXT:    s_nop 05477; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill5478; GFX9-NEXT:    ; implicit-def: $vgpr425479; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill5480; GFX9-NEXT:    s_nop 05481; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill5482; GFX9-NEXT:    ; implicit-def: $vgpr425483; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill5484; GFX9-NEXT:    s_nop 05485; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill5486; GFX9-NEXT:    ; implicit-def: $vgpr425487; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill5488; GFX9-NEXT:    s_nop 05489; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill5490; GFX9-NEXT:    ; implicit-def: $vgpr425491; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill5492; GFX9-NEXT:    s_nop 05493; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill5494; GFX9-NEXT:    s_waitcnt vmcnt(45)5495; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v335496; GFX9-NEXT:    ; implicit-def: $vgpr335497; GFX9-NEXT:    ; kill: killed $vgpr335498; GFX9-NEXT:    ; implicit-def: $vgpr335499; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc5500; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]5501; GFX9-NEXT:    s_cbranch_execz .LBB12_25502; GFX9-NEXT:  ; %bb.1: ; %cmp.false5503; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v325504; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill5505; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v325506; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill5507; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v325508; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill5509; GFX9-NEXT:    s_waitcnt vmcnt(47)5510; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v315511; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill5512; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v315513; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill5514; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v305515; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill5516; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v305517; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill5518; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v305519; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill5520; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v295521; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[31:32]5522; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill5523; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v295524; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill5525; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v285526; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill5527; GFX9-NEXT:    s_nop 05528; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill5529; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill5530; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v285531; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[29:30]5532; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill5533; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v285534; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill5535; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v275536; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill5537; GFX9-NEXT:    s_nop 05538; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill5539; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill5540; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v275541; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[27:28]5542; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill5543; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v265544; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill5545; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v265546; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill5547; GFX9-NEXT:    s_nop 05548; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill5549; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill5550; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v265551; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[25:26]5552; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill5553; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v255554; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill5555; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v255556; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill5557; GFX9-NEXT:    s_nop 05558; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill5559; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill5560; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v245561; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[23:24]5562; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill5563; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v245564; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill5565; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v245566; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill5567; GFX9-NEXT:    s_nop 05568; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill5569; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill5570; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v235571; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[21:22]5572; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill5573; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v235574; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill5575; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v225576; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill5577; GFX9-NEXT:    s_nop 05578; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill5579; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill5580; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v225581; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[19:20]5582; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill5583; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v225584; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill5585; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v215586; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill5587; GFX9-NEXT:    s_nop 05588; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill5589; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill5590; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v215591; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[17:18]5592; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill5593; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v205594; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill5595; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v205596; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill5597; GFX9-NEXT:    s_nop 05598; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill5599; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill5600; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v205601; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[15:16]5602; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill5603; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v195604; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill5605; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v195606; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill5607; GFX9-NEXT:    s_nop 05608; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill5609; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill5610; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v185611; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[13:14]5612; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill5613; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v185614; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill5615; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v185616; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill5617; GFX9-NEXT:    s_nop 05618; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill5619; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill5620; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v175621; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[11:12]5622; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill5623; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v175624; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill5625; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v165626; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill5627; GFX9-NEXT:    s_nop 05628; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill5629; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill5630; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v165631; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[9:10]5632; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill5633; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v155634; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill5635; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v145636; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill5637; GFX9-NEXT:    s_nop 05638; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill5639; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill5640; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v145641; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[7:8]5642; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill5643; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v135644; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill5645; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v125646; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill5647; GFX9-NEXT:    s_nop 05648; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill5649; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill5650; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v125651; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[5:6]5652; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill5653; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v115654; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill5655; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v105656; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill5657; GFX9-NEXT:    s_nop 05658; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill5659; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill5660; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v105661; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v95662; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]5663; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v165664; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v155665; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v145666; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 8, v135667; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v125668; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v115669; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill5670; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v105671; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill5672; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 8, v95673; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v85674; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v85675; GFX9-NEXT:    v_lshrrev_b32_e32 v54, 8, v85676; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v75677; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 8, v75678; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 24, v65679; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v65680; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v65681; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v55682; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v55683; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 24, v45684; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 16, v45685; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v45686; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v35687; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 8, v35688; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v25689; GFX9-NEXT:    v_lshrrev_b32_e32 v53, 16, v25690; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 8, v25691; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 16, v15692; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v15693; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]5694; GFX9-NEXT:  .LBB12_2: ; %Flow5695; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]5696; GFX9-NEXT:    s_cbranch_execz .LBB12_45697; GFX9-NEXT:  ; %bb.3: ; %cmp.true5698; GFX9-NEXT:    v_add_u32_e32 v32, 3, v325699; GFX9-NEXT:    s_waitcnt vmcnt(44)5700; GFX9-NEXT:    v_add_u32_e32 v31, 3, v315701; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]5702; GFX9-NEXT:    v_add_u32_e32 v30, 3, v305703; GFX9-NEXT:    v_add_u32_e32 v29, 3, v295704; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill5705; GFX9-NEXT:    s_nop 05706; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill5707; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]5708; GFX9-NEXT:    v_add_u32_e32 v28, 3, v285709; GFX9-NEXT:    v_add_u32_e32 v27, 3, v275710; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill5711; GFX9-NEXT:    s_nop 05712; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill5713; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]5714; GFX9-NEXT:    v_add_u32_e32 v26, 3, v265715; GFX9-NEXT:    v_add_u32_e32 v25, 3, v255716; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill5717; GFX9-NEXT:    s_nop 05718; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill5719; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]5720; GFX9-NEXT:    v_add_u32_e32 v24, 3, v245721; GFX9-NEXT:    v_add_u32_e32 v23, 3, v235722; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill5723; GFX9-NEXT:    s_nop 05724; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill5725; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]5726; GFX9-NEXT:    v_add_u32_e32 v22, 3, v225727; GFX9-NEXT:    v_add_u32_e32 v21, 3, v215728; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill5729; GFX9-NEXT:    s_nop 05730; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill5731; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]5732; GFX9-NEXT:    v_add_u32_e32 v20, 3, v205733; GFX9-NEXT:    v_add_u32_e32 v19, 3, v195734; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill5735; GFX9-NEXT:    s_nop 05736; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill5737; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]5738; GFX9-NEXT:    v_add_u32_e32 v18, 3, v185739; GFX9-NEXT:    v_add_u32_e32 v17, 3, v175740; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill5741; GFX9-NEXT:    s_nop 05742; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill5743; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]5744; GFX9-NEXT:    v_add_u32_e32 v16, 3, v165745; GFX9-NEXT:    v_add_u32_e32 v15, 3, v155746; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill5747; GFX9-NEXT:    s_nop 05748; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill5749; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]5750; GFX9-NEXT:    v_add_u32_e32 v14, 3, v145751; GFX9-NEXT:    v_add_u32_e32 v13, 3, v135752; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill5753; GFX9-NEXT:    s_nop 05754; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill5755; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]5756; GFX9-NEXT:    v_add_u32_e32 v12, 3, v125757; GFX9-NEXT:    v_add_u32_e32 v11, 3, v115758; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill5759; GFX9-NEXT:    s_nop 05760; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill5761; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]5762; GFX9-NEXT:    v_add_u32_e32 v10, 3, v105763; GFX9-NEXT:    v_add_u32_e32 v9, 3, v95764; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill5765; GFX9-NEXT:    s_nop 05766; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill5767; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]5768; GFX9-NEXT:    v_add_u32_e32 v8, 3, v85769; GFX9-NEXT:    v_add_u32_e32 v7, 3, v75770; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill5771; GFX9-NEXT:    s_nop 05772; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill5773; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[7:8]5774; GFX9-NEXT:    v_add_u32_e32 v6, 3, v65775; GFX9-NEXT:    v_add_u32_e32 v5, 3, v55776; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill5777; GFX9-NEXT:    s_nop 05778; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill5779; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[5:6]5780; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill5781; GFX9-NEXT:    s_nop 05782; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill5783; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v325784; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill5785; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v325786; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill5787; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v325788; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill5789; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v315790; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill5791; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v315792; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill5793; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v305794; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill5795; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v305796; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill5797; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v305798; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill5799; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v295800; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill5801; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v295802; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill5803; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v285804; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill5805; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v285806; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill5807; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v285808; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill5809; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v275810; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill5811; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v275812; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill5813; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v265814; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill5815; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v265816; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill5817; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v265818; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill5819; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v255820; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill5821; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v255822; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill5823; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v245824; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill5825; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v245826; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill5827; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v245828; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill5829; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v235830; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill5831; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v235832; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill5833; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v225834; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill5835; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v225836; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill5837; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v225838; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill5839; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v215840; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill5841; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v215842; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill5843; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v205844; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill5845; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v205846; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill5847; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v205848; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill5849; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v195850; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill5851; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v195852; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill5853; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v185854; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill5855; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v185856; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill5857; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v185858; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill5859; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v175860; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill5861; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v175862; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill5863; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v165864; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill5865; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v165866; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill5867; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v155868; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill5869; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v145870; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill5871; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v145872; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill5873; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v135874; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill5875; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v125876; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill5877; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v125878; GFX9-NEXT:    v_add_u32_e32 v4, 3, v45879; GFX9-NEXT:    v_add_u32_e32 v3, 3, v35880; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill5881; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v115882; GFX9-NEXT:    v_add_u32_e32 v2, 3, v25883; GFX9-NEXT:    v_add_u32_e32 v1, 3, v15884; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]5885; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill5886; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v105887; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]5888; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill5889; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v105890; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v95891; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v165892; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v155893; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v145894; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 8, v135895; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v125896; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v115897; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill5898; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v105899; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill5900; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 8, v95901; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v85902; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v85903; GFX9-NEXT:    v_lshrrev_b32_e32 v54, 8, v85904; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v75905; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 8, v75906; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 24, v65907; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v65908; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v65909; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v55910; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v55911; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 24, v45912; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 16, v45913; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v45914; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v35915; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 8, v35916; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v25917; GFX9-NEXT:    v_lshrrev_b32_e32 v53, 16, v25918; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 8, v25919; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 16, v15920; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v15921; GFX9-NEXT:  .LBB12_4: ; %end5922; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]5923; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v335924; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5925; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v525926; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5927; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v365928; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5929; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v505930; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5931; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v605932; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5933; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v595934; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5935; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v575936; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5937; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload5938; GFX9-NEXT:    v_lshlrev_b16_e32 v42, 8, v565939; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5940; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v385941; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5942; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v395943; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5944; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v475945; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5946; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v465947; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5948; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v455949; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5950; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v555951; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5952; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v545953; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5954; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v375955; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5956; GFX9-NEXT:    s_waitcnt vmcnt(0)5957; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v335958; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5959; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload5960; GFX9-NEXT:    s_waitcnt vmcnt(0)5961; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v335962; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5963; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v415964; GFX9-NEXT:    v_or_b32_sdwa v33, v44, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5965; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5966; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen5967; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v435968; GFX9-NEXT:    v_or_b32_sdwa v1, v53, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5969; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5970; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:45971; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v405972; GFX9-NEXT:    v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5973; GFX9-NEXT:    v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5974; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:85975; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v625976; GFX9-NEXT:    v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5977; GFX9-NEXT:    v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5978; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:125979; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload5980; GFX9-NEXT:    s_nop 05981; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload5982; GFX9-NEXT:    s_waitcnt vmcnt(1)5983; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v15984; GFX9-NEXT:    v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5985; GFX9-NEXT:    v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5986; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:165987; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v355988; GFX9-NEXT:    v_or_b32_sdwa v1, v49, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5989; GFX9-NEXT:    v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5990; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:205991; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload5992; GFX9-NEXT:    s_nop 05993; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload5994; GFX9-NEXT:    s_waitcnt vmcnt(1)5995; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v15996; GFX9-NEXT:    v_or_b32_sdwa v1, v58, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD5997; GFX9-NEXT:    v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD5998; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:245999; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v346000; GFX9-NEXT:    v_or_b32_sdwa v1, v48, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6001; GFX9-NEXT:    v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6002; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:286003; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload6004; GFX9-NEXT:    s_nop 06005; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload6006; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload6007; GFX9-NEXT:    s_waitcnt vmcnt(2)6008; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16009; GFX9-NEXT:    s_waitcnt vmcnt(0)6010; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6011; GFX9-NEXT:    v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6012; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:326013; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload6014; GFX9-NEXT:    s_nop 06015; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload6016; GFX9-NEXT:    s_waitcnt vmcnt(1)6017; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16018; GFX9-NEXT:    s_waitcnt vmcnt(0)6019; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6020; GFX9-NEXT:    v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6021; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:366022; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload6023; GFX9-NEXT:    s_nop 06024; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload6025; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload6026; GFX9-NEXT:    s_waitcnt vmcnt(2)6027; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16028; GFX9-NEXT:    s_waitcnt vmcnt(0)6029; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6030; GFX9-NEXT:    v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6031; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:406032; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload6033; GFX9-NEXT:    s_nop 06034; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload6035; GFX9-NEXT:    s_waitcnt vmcnt(1)6036; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16037; GFX9-NEXT:    s_waitcnt vmcnt(0)6038; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6039; GFX9-NEXT:    v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6040; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:446041; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload6042; GFX9-NEXT:    s_nop 06043; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload6044; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload6045; GFX9-NEXT:    s_waitcnt vmcnt(2)6046; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16047; GFX9-NEXT:    s_waitcnt vmcnt(0)6048; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6049; GFX9-NEXT:    v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6050; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:486051; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload6052; GFX9-NEXT:    s_nop 06053; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload6054; GFX9-NEXT:    s_waitcnt vmcnt(1)6055; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16056; GFX9-NEXT:    s_waitcnt vmcnt(0)6057; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6058; GFX9-NEXT:    v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6059; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:526060; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload6061; GFX9-NEXT:    s_nop 06062; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload6063; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload6064; GFX9-NEXT:    s_waitcnt vmcnt(2)6065; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16066; GFX9-NEXT:    s_waitcnt vmcnt(0)6067; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6068; GFX9-NEXT:    v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6069; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:566070; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload6071; GFX9-NEXT:    s_nop 06072; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload6073; GFX9-NEXT:    s_waitcnt vmcnt(1)6074; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16075; GFX9-NEXT:    s_waitcnt vmcnt(0)6076; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6077; GFX9-NEXT:    v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6078; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:606079; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload6080; GFX9-NEXT:    s_nop 06081; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload6082; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload6083; GFX9-NEXT:    s_waitcnt vmcnt(2)6084; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16085; GFX9-NEXT:    s_waitcnt vmcnt(0)6086; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6087; GFX9-NEXT:    v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6088; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:646089; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload6090; GFX9-NEXT:    s_nop 06091; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload6092; GFX9-NEXT:    s_waitcnt vmcnt(1)6093; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16094; GFX9-NEXT:    s_waitcnt vmcnt(0)6095; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6096; GFX9-NEXT:    v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6097; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:686098; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload6099; GFX9-NEXT:    s_nop 06100; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload6101; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload6102; GFX9-NEXT:    s_waitcnt vmcnt(2)6103; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16104; GFX9-NEXT:    s_waitcnt vmcnt(0)6105; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6106; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload6107; GFX9-NEXT:    s_waitcnt vmcnt(0)6108; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26109; GFX9-NEXT:    v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6110; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6111; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:726112; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload6113; GFX9-NEXT:    s_nop 06114; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload6115; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload6116; GFX9-NEXT:    s_waitcnt vmcnt(2)6117; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16118; GFX9-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6119; GFX9-NEXT:    s_waitcnt vmcnt(1)6120; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26121; GFX9-NEXT:    s_waitcnt vmcnt(0)6122; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6123; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6124; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:766125; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload6126; GFX9-NEXT:    s_nop 06127; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload6128; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload6129; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload6130; GFX9-NEXT:    s_waitcnt vmcnt(2)6131; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26132; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16133; GFX9-NEXT:    v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6134; GFX9-NEXT:    s_waitcnt vmcnt(0)6135; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6136; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6137; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:806138; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload6139; GFX9-NEXT:    s_nop 06140; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload6141; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload6142; GFX9-NEXT:    s_waitcnt vmcnt(2)6143; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16144; GFX9-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6145; GFX9-NEXT:    s_waitcnt vmcnt(1)6146; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26147; GFX9-NEXT:    s_waitcnt vmcnt(0)6148; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6149; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6150; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:846151; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload6152; GFX9-NEXT:    s_nop 06153; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload6154; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload6155; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload6156; GFX9-NEXT:    s_waitcnt vmcnt(2)6157; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26158; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16159; GFX9-NEXT:    v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6160; GFX9-NEXT:    s_waitcnt vmcnt(0)6161; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6162; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6163; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:886164; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload6165; GFX9-NEXT:    s_nop 06166; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload6167; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload6168; GFX9-NEXT:    s_waitcnt vmcnt(2)6169; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16170; GFX9-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6171; GFX9-NEXT:    s_waitcnt vmcnt(1)6172; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26173; GFX9-NEXT:    s_waitcnt vmcnt(0)6174; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6175; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6176; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:926177; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload6178; GFX9-NEXT:    s_nop 06179; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload6180; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload6181; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload6182; GFX9-NEXT:    s_waitcnt vmcnt(2)6183; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26184; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16185; GFX9-NEXT:    v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6186; GFX9-NEXT:    s_waitcnt vmcnt(0)6187; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6188; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6189; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:966190; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload6191; GFX9-NEXT:    s_nop 06192; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload6193; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload6194; GFX9-NEXT:    s_waitcnt vmcnt(2)6195; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16196; GFX9-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6197; GFX9-NEXT:    s_waitcnt vmcnt(1)6198; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26199; GFX9-NEXT:    s_waitcnt vmcnt(0)6200; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6201; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6202; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:1006203; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload6204; GFX9-NEXT:    s_nop 06205; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload6206; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload6207; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload6208; GFX9-NEXT:    s_waitcnt vmcnt(2)6209; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26210; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16211; GFX9-NEXT:    v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6212; GFX9-NEXT:    s_waitcnt vmcnt(0)6213; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6214; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6215; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:1046216; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload6217; GFX9-NEXT:    s_nop 06218; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload6219; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload6220; GFX9-NEXT:    s_waitcnt vmcnt(2)6221; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16222; GFX9-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6223; GFX9-NEXT:    s_waitcnt vmcnt(1)6224; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26225; GFX9-NEXT:    s_waitcnt vmcnt(0)6226; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6227; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6228; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:1086229; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload6230; GFX9-NEXT:    s_nop 06231; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload6232; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload6233; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload6234; GFX9-NEXT:    s_waitcnt vmcnt(2)6235; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26236; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16237; GFX9-NEXT:    v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6238; GFX9-NEXT:    s_waitcnt vmcnt(0)6239; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6240; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6241; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:1126242; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload6243; GFX9-NEXT:    s_nop 06244; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload6245; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload6246; GFX9-NEXT:    s_waitcnt vmcnt(2)6247; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16248; GFX9-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6249; GFX9-NEXT:    s_waitcnt vmcnt(1)6250; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26251; GFX9-NEXT:    s_waitcnt vmcnt(0)6252; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6253; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6254; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:1166255; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload6256; GFX9-NEXT:    s_nop 06257; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload6258; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload6259; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload6260; GFX9-NEXT:    s_waitcnt vmcnt(2)6261; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26262; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16263; GFX9-NEXT:    v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6264; GFX9-NEXT:    s_waitcnt vmcnt(0)6265; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6266; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6267; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:1206268; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload6269; GFX9-NEXT:    s_nop 06270; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload6271; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload6272; GFX9-NEXT:    s_waitcnt vmcnt(2)6273; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v16274; GFX9-NEXT:    v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6275; GFX9-NEXT:    s_waitcnt vmcnt(1)6276; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v26277; GFX9-NEXT:    s_waitcnt vmcnt(0)6278; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD6279; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD6280; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:1246281; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload6282; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload6283; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload6284; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload6285; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload6286; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload6287; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload6288; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload6289; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload6290; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload6291; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload6292; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload6293; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload6294; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload6295; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload6296; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload6297; GFX9-NEXT:    s_waitcnt vmcnt(0)6298; GFX9-NEXT:    s_setpc_b64 s[30:31]6299;6300; GFX11-TRUE16-LABEL: bitcast_v32i32_to_v128i8:6301; GFX11-TRUE16:       ; %bb.0:6302; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)6303; GFX11-TRUE16-NEXT:    s_clause 0x26304; GFX11-TRUE16-NEXT:    scratch_load_b32 v33, off, s32 offset:86305; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:46306; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s326307; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo166308; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo166309; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo166310; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo166311; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo166312; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo166313; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo166314; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo166315; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo166316; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo166317; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo166318; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo166319; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo166320; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo166321; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo166322; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo166323; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo166324; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo166325; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo166326; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo166327; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo166328; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo166329; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo166330; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo166331; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo166332; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo166333; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo166334; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo166335; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo166336; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo166337; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo166338; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo166339; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo166340; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo166341; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo166342; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo166343; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo166344; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo166345; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo166346; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo166347; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo166348; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo166349; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo166350; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo166351; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo166352; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo166353; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo166354; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo166355; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo166356; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo166357; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo166358; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo166359; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo166360; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo166361; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo166362; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo166363; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo166364; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo166365; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo166366; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo166367; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo166368; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo166369; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo166370; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)6371; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v336372; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo166373; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo6374; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)6375; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s06376; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB12_26377; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false6378; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)6379; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]6380; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]6381; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]6382; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]6383; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]6384; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]6385; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]6386; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]6387; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]6388; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]6389; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]6390; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]6391; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]6392; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]6393; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v326394; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v326395; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v316396; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v306397; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v306398; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v296399; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v286400; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v286401; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v276402; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v266403; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v266404; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v256405; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v246406; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v246407; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v236408; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v226409; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v226410; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v216411; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v206412; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v206413; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v196414; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v186415; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v186416; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v176417; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v166418; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v166419; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v156420; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v146421; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v146422; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v136423; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v126424; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v126425; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v116426; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v106427; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v106428; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v96429; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v86430; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v86431; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v76432; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v66433; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v66434; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v56435; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v46436; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v46437; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v36438; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v26439; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v26440; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]6441; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]6442; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v16443; GFX11-TRUE16-NEXT:  .LBB12_2: ; %Flow6444; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s06445; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB12_46446; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true6447; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v206448; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v196449; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v186450; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v176451; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)6452; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v326453; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)6454; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v316455; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v166456; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v156457; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v306458; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v296459; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v146460; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v136461; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v286462; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v276463; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v66464; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v56465; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v126466; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v116467; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v266468; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v256469; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v46470; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v36471; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v106472; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v96473; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v246474; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v236475; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]6476; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v26477; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v16478; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v86479; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v76480; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v226481; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v216482; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]6483; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]6484; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]6485; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]6486; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]6487; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]6488; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]6489; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]6490; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]6491; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]6492; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]6493; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]6494; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]6495; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]6496; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]6497; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v326498; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v326499; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v316500; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v306501; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v306502; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v296503; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v286504; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v286505; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v276506; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v266507; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v266508; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v256509; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v246510; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v246511; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v236512; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v226513; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v226514; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v216515; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v206516; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v206517; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v196518; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v186519; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v186520; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v176521; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v166522; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v166523; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v156524; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v146525; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v146526; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v136527; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v126528; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v126529; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v116530; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v106531; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v106532; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v96533; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v86534; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v86535; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v76536; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v66537; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v66538; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v56539; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v46540; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v46541; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v36542; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v26543; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v26544; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v16545; GFX11-TRUE16-NEXT:  .LBB12_4: ; %end6546; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s06547; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.l6548; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)6549; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.h, 8, v39.l6550; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h6551; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.h, 8, v66.l6552; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l6553; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v2.h6554; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.l, v33.h6555; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.h, 8, v162.l6556; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v1.h, v34.h6557; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.h, 8, v161.l6558; GFX11-TRUE16-NEXT:    v_and_b16 v3.l, 0xff, v3.l6559; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v35.h, 8, v160.l6560; GFX11-TRUE16-NEXT:    v_and_b16 v3.h, 0xff, v3.h6561; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v36.h, 8, v65.l6562; GFX11-TRUE16-NEXT:    v_or_b16 v2.l, v2.l, v33.h6563; GFX11-TRUE16-NEXT:    v_or_b16 v2.h, v2.h, v34.h6564; GFX11-TRUE16-NEXT:    v_or_b16 v3.l, v3.l, v35.h6565; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v4.l6566; GFX11-TRUE16-NEXT:    v_or_b16 v3.h, v3.h, v36.h6567; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.h, 8, v151.l6568; GFX11-TRUE16-NEXT:    v_and_b16 v4.h, 0xff, v4.h6569; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.h, 8, v150.l6570; GFX11-TRUE16-NEXT:    v_and_b16 v5.l, 0xff, v5.l6571; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v35.h, 8, v149.l6572; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v5.h6573; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v36.h, 8, v64.l6574; GFX11-TRUE16-NEXT:    v_and_b16 v6.l, 0xff, v6.l6575; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v37.h, 8, v148.l6576; GFX11-TRUE16-NEXT:    v_or_b16 v4.l, v4.l, v33.h6577; GFX11-TRUE16-NEXT:    v_or_b16 v4.h, v4.h, v34.h6578; GFX11-TRUE16-NEXT:    v_or_b16 v5.l, v5.l, v35.h6579; GFX11-TRUE16-NEXT:    v_or_b16 v5.h, v5.h, v36.h6580; GFX11-TRUE16-NEXT:    v_or_b16 v6.l, v6.l, v37.h6581; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v6.h6582; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.h, 8, v147.l6583; GFX11-TRUE16-NEXT:    v_and_b16 v7.l, 0xff, v7.l6584; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.h, 8, v146.l6585; GFX11-TRUE16-NEXT:    v_and_b16 v7.h, 0xff, v7.h6586; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v35.h, 8, v54.l6587; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v8.l6588; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v36.h, 8, v145.l6589; GFX11-TRUE16-NEXT:    v_and_b16 v8.h, 0xff, v8.h6590; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v37.h, 8, v144.l6591; GFX11-TRUE16-NEXT:    v_or_b16 v6.h, v6.h, v33.h6592; GFX11-TRUE16-NEXT:    v_or_b16 v7.l, v7.l, v34.h6593; GFX11-TRUE16-NEXT:    v_or_b16 v7.h, v7.h, v35.h6594; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.l, v36.h6595; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v8.h, v37.h6596; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v9.l6597; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.h, 8, v135.l6598; GFX11-TRUE16-NEXT:    v_and_b16 v9.h, 0xff, v9.h6599; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.h, 8, v53.l6600; GFX11-TRUE16-NEXT:    s_clause 0x16601; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off6602; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:166603; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v10.h6604; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v133.l6605; GFX11-TRUE16-NEXT:    v_and_b16 v3.h, 0xff, v11.l6606; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v4.l, 8, v132.l6607; GFX11-TRUE16-NEXT:    v_and_b16 v4.h, 0xff, v11.h6608; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v5.l, 8, v52.l6609; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v12.l6610; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.l, 8, v131.l6611; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v12.h6612; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.l, 8, v130.l6613; GFX11-TRUE16-NEXT:    v_and_b16 v10.l, 0xff, v10.l6614; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v35.h, 8, v134.l6615; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v9.l, v33.h6616; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v9.h, v34.h6617; GFX11-TRUE16-NEXT:    v_or_b16 v2.h, v2.h, v3.l6618; GFX11-TRUE16-NEXT:    v_or_b16 v3.l, v3.h, v4.l6619; GFX11-TRUE16-NEXT:    v_or_b16 v3.h, v4.h, v5.l6620; GFX11-TRUE16-NEXT:    v_or_b16 v4.l, v5.h, v6.l6621; GFX11-TRUE16-NEXT:    v_or_b16 v4.h, v6.h, v7.l6622; GFX11-TRUE16-NEXT:    v_and_b16 v5.l, 0xff, v13.l6623; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v5.h, 8, v129.l6624; GFX11-TRUE16-NEXT:    v_and_b16 v6.l, 0xff, v13.h6625; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.h, 8, v51.l6626; GFX11-TRUE16-NEXT:    v_and_b16 v7.l, 0xff, v14.l6627; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.h, 8, v128.l6628; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v14.h6629; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.h, 8, v119.l6630; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v15.l6631; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.h, 8, v118.l6632; GFX11-TRUE16-NEXT:    v_or_b16 v2.l, v10.l, v35.h6633; GFX11-TRUE16-NEXT:    v_or_b16 v5.l, v5.l, v5.h6634; GFX11-TRUE16-NEXT:    v_or_b16 v5.h, v6.l, v6.h6635; GFX11-TRUE16-NEXT:    v_or_b16 v6.l, v7.l, v7.h6636; GFX11-TRUE16-NEXT:    v_or_b16 v6.h, v8.l, v8.h6637; GFX11-TRUE16-NEXT:    v_or_b16 v7.l, v9.l, v9.h6638; GFX11-TRUE16-NEXT:    v_and_b16 v7.h, 0xff, v15.h6639; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.l, 8, v50.l6640; GFX11-TRUE16-NEXT:    v_and_b16 v8.h, 0xff, v16.l6641; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.l, 8, v117.l6642; GFX11-TRUE16-NEXT:    v_and_b16 v9.h, 0xff, v16.h6643; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v10.l, 8, v116.l6644; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v17.l6645; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.l, 8, v115.l6646; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v17.h6647; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.l, 8, v49.l6648; GFX11-TRUE16-NEXT:    v_or_b16 v7.h, v7.h, v8.l6649; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.h, v9.l6650; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v9.h, v10.l6651; GFX11-TRUE16-NEXT:    v_or_b16 v9.l, v10.h, v11.l6652; GFX11-TRUE16-NEXT:    v_or_b16 v9.h, v11.h, v12.l6653; GFX11-TRUE16-NEXT:    v_and_b16 v10.l, 0xff, v18.l6654; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v10.h, 8, v114.l6655; GFX11-TRUE16-NEXT:    v_and_b16 v11.l, 0xff, v18.h6656; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.h, 8, v113.l6657; GFX11-TRUE16-NEXT:    v_and_b16 v12.l, 0xff, v19.l6658; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.h, 8, v112.l6659; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v19.h6660; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.h, 8, v48.l6661; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v20.l6662; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.h, 8, v103.l6663; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v10.l, v10.h6664; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v11.l, v11.h6665; GFX11-TRUE16-NEXT:    v_or_b16 v11.l, v12.l, v12.h6666; GFX11-TRUE16-NEXT:    v_or_b16 v11.h, v13.l, v13.h6667; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v14.l, v14.h6668; GFX11-TRUE16-NEXT:    v_and_b16 v12.h, 0xff, v20.h6669; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.l, 8, v102.l6670; GFX11-TRUE16-NEXT:    v_and_b16 v13.h, 0xff, v21.l6671; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.l, 8, v101.l6672; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v21.h6673; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v15.l, 8, v38.l6674; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v22.l6675; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.l, 8, v100.l6676; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v22.h6677; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.l, 8, v99.l6678; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v12.h, v13.l6679; GFX11-TRUE16-NEXT:    v_or_b16 v13.l, v13.h, v14.l6680; GFX11-TRUE16-NEXT:    v_or_b16 v13.h, v14.h, v15.l6681; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v15.h, v16.l6682; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v16.h, v17.l6683; GFX11-TRUE16-NEXT:    v_and_b16 v15.l, 0xff, v23.l6684; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v15.h, 8, v98.l6685; GFX11-TRUE16-NEXT:    v_and_b16 v16.l, 0xff, v23.h6686; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.h, 8, v37.l6687; GFX11-TRUE16-NEXT:    v_and_b16 v17.l, 0xff, v24.l6688; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.h, 8, v97.l6689; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v24.h6690; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.h, 8, v96.l6691; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v25.l6692; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.h, 8, v87.l6693; GFX11-TRUE16-NEXT:    v_or_b16 v15.l, v15.l, v15.h6694; GFX11-TRUE16-NEXT:    v_or_b16 v15.h, v16.l, v16.h6695; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v17.l, v17.h6696; GFX11-TRUE16-NEXT:    v_or_b16 v16.h, v18.l, v18.h6697; GFX11-TRUE16-NEXT:    v_or_b16 v17.l, v19.l, v19.h6698; GFX11-TRUE16-NEXT:    v_and_b16 v17.h, 0xff, v25.h6699; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.l, 8, v36.l6700; GFX11-TRUE16-NEXT:    v_and_b16 v18.h, 0xff, v26.l6701; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.l, 8, v86.l6702; GFX11-TRUE16-NEXT:    v_and_b16 v19.h, 0xff, v26.h6703; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v20.l, 8, v85.l6704; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v27.l6705; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.l, 8, v84.l6706; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v27.h6707; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.l, 8, v35.l6708; GFX11-TRUE16-NEXT:    v_or_b16 v17.h, v17.h, v18.l6709; GFX11-TRUE16-NEXT:    v_or_b16 v18.l, v18.h, v19.l6710; GFX11-TRUE16-NEXT:    v_or_b16 v18.h, v19.h, v20.l6711; GFX11-TRUE16-NEXT:    v_or_b16 v19.l, v20.h, v21.l6712; GFX11-TRUE16-NEXT:    v_or_b16 v19.h, v21.h, v22.l6713; GFX11-TRUE16-NEXT:    v_and_b16 v20.l, 0xff, v28.l6714; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v20.h, 8, v83.l6715; GFX11-TRUE16-NEXT:    v_and_b16 v21.l, 0xff, v28.h6716; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.h, 8, v82.l6717; GFX11-TRUE16-NEXT:    v_and_b16 v22.l, 0xff, v29.l6718; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.h, 8, v81.l6719; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v29.h6720; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.h, 8, v34.l6721; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v30.l6722; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.h, 8, v80.l6723; GFX11-TRUE16-NEXT:    v_or_b16 v20.l, v20.l, v20.h6724; GFX11-TRUE16-NEXT:    v_or_b16 v20.h, v21.l, v21.h6725; GFX11-TRUE16-NEXT:    v_or_b16 v21.l, v22.l, v22.h6726; GFX11-TRUE16-NEXT:    v_or_b16 v21.h, v23.l, v23.h6727; GFX11-TRUE16-NEXT:    v_or_b16 v22.l, v24.l, v24.h6728; GFX11-TRUE16-NEXT:    v_and_b16 v22.h, 0xff, v30.h6729; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.l, 8, v71.l6730; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)6731; GFX11-TRUE16-NEXT:    v_and_b16 v23.h, 0xff, v31.l6732; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.l, 8, v70.l6733; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v31.h6734; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v25.l, 8, v33.l6735; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v32.l6736; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v26.l, 8, v69.l6737; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v32.h6738; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.l, 8, v68.l6739; GFX11-TRUE16-NEXT:    v_or_b16 v22.h, v22.h, v23.l6740; GFX11-TRUE16-NEXT:    v_or_b16 v23.l, v23.h, v24.l6741; GFX11-TRUE16-NEXT:    v_or_b16 v23.h, v24.h, v25.l6742; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v25.h, v26.l6743; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v26.h, v27.l6744; GFX11-TRUE16-NEXT:    s_clause 0x56745; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:326746; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:486747; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:646748; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:806749; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:966750; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:1126751; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]6752;6753; GFX11-FAKE16-LABEL: bitcast_v32i32_to_v128i8:6754; GFX11-FAKE16:       ; %bb.0:6755; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)6756; GFX11-FAKE16-NEXT:    s_clause 0x26757; GFX11-FAKE16-NEXT:    scratch_load_b32 v33, off, s32 offset:86758; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:46759; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s326760; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Spill6761; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:886762; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:846763; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:806764; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:766765; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:726766; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:686767; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:646768; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:606769; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:566770; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:526771; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:486772; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:446773; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:406774; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:366775; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:326776; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:286777; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:246778; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:206779; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:166780; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:126781; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr756782; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr746783; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr666784; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr736785; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr726786; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr636787; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr626788; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr616789; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr656790; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr606791; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr596792; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr586793; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr576794; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr566795; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr646796; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr476797; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr466798; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr456799; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr446800; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr436801; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr546802; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr426803; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr416804; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr406805; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1836806; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1826807; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr536808; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1816809; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1806810; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1796811; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1786812; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1776813; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr526814; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1766815; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1676816; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1666817; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1656818; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1646819; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr516820; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1636821; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1626822; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1616823; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1606824; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1516825; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr506826; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1506827; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1496828; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1486829; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1476830; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1466831; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr496832; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1456833; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1446834; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1356835; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1346836; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1336837; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr486838; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1326839; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1316840; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1306841; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1296842; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1286843; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr386844; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1196845; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1186846; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1176847; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1166848; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1156849; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr376850; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1146851; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1136852; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1126853; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1036854; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1026855; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr366856; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1016857; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr1006858; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr996859; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr986860; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr976861; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr356862; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr966863; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr876864; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr866865; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr856866; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr846867; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr836868; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr826869; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr816870; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr806871; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr716872; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr706873; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr696874; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr686875; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr346876; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)6877; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v336878; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr336879; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo6880; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)6881; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s06882; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB12_26883; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false6884; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]6885; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)6886; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]6887; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]6888; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]6889; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]6890; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]6891; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]6892; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]6893; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]6894; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]6895; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]6896; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]6897; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]6898; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v326899; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v326900; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v326901; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v316902; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v316903; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v306904; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v306905; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v306906; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v296907; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v296908; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v286909; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v286910; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v286911; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v276912; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v276913; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v266914; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v266915; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v266916; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v256917; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v256918; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v246919; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v246920; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v246921; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v236922; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v236923; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v226924; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v226925; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v226926; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v216927; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v216928; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v206929; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v206930; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v206931; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v196932; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v196933; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v186934; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v186935; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v186936; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v176937; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v176938; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v166939; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v166940; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v166941; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v156942; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v156943; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v146944; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v146945; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v146946; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v136947; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v136948; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v126949; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v126950; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v126951; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v116952; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v116953; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v106954; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v106955; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v106956; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v96957; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v96958; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v86959; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v86960; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v86961; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v76962; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v76963; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v66964; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v66965; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v66966; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v56967; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v56968; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v46969; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v46970; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v46971; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v36972; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v36973; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v26974; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v26975; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v26976; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v16977; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v16978; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]6979; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]6980; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]6981; GFX11-FAKE16-NEXT:  .LBB12_2: ; %Flow6982; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s06983; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB12_46984; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true6985; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v206986; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v196987; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v186988; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v176989; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)6990; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 3, v326991; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)6992; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v316993; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 3, v166994; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 3, v156995; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v306996; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v296997; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 3, v146998; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 3, v136999; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v287000; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v277001; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v67002; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v57003; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 3, v127004; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 3, v117005; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v267006; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v257007; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v47008; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v37009; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 3, v107010; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 3, v97011; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v247012; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v237013; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]7014; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v27015; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v17016; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 3, v87017; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 3, v77018; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v227019; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v217020; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]7021; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]7022; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]7023; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]7024; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]7025; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]7026; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]7027; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]7028; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]7029; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]7030; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]7031; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]7032; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]7033; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]7034; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]7035; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v327036; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v327037; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v327038; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v317039; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v317040; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v307041; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v307042; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v307043; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v297044; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v297045; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v287046; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v287047; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v287048; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v277049; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v277050; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v267051; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v267052; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v267053; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v257054; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v257055; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v247056; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v247057; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v247058; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v237059; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v237060; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v227061; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v227062; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v227063; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v217064; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v217065; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v207066; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v207067; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v207068; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v197069; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v197070; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v187071; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v187072; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v187073; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v177074; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v177075; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v167076; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v167077; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v167078; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v157079; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v157080; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v147081; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v147082; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v147083; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v137084; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v137085; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v127086; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v127087; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v127088; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v117089; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v117090; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v107091; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v107092; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v107093; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v97094; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v97095; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v87096; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v87097; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v87098; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v77099; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v77100; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v67101; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v67102; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v67103; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v57104; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v57105; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v47106; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v47107; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v47108; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v37109; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v37110; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v27111; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v27112; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v27113; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v17114; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v17115; GFX11-FAKE16-NEXT:  .LBB12_4: ; %end7116; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s07117; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)7118; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v55, 8, v757119; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v17120; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v39, 8, v667121; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v27122; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v66, 8, v637123; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v65, 8, v657124; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v557125; GFX11-FAKE16-NEXT:    v_and_b32_e32 v55, 0xff, v747126; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v37127; GFX11-FAKE16-NEXT:    v_and_b32_e32 v67, 0xff, v617128; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v47129; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v17130; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, v55, v397131; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v55, 8, v737132; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, v67, v657133; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v67, 8, v587134; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v57135; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v397136; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v557137; GFX11-FAKE16-NEXT:    v_and_b32_e32 v55, 0xff, v727138; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v64, 8, v647139; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v67140; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v397141; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v27142; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, v55, v667143; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v66, 8, v627144; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v77145; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v54, 8, v547146; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v87147; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v557148; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v667149; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v55, 8, v607150; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xff, v597151; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v53, 8, v537152; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v397153; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v37154; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v657155; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v557156; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, v66, v677157; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v65, 8, v577158; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xff, v567159; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v397160; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v47161; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v557162; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v657163; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, v66, v647164; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v64, 8, v477165; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xff, v467166; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v66, 8, v457167; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v397168; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v57169; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v557170; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v647171; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, v65, v667172; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v64, 8, v447173; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xff, v437174; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v397175; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v67176; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v557177; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v647178; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, v65, v547179; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v55, 8, v427180; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xff, v417181; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v65, 8, v407182; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v397183; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v547184; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v557185; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v77186; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, v64, v657187; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xff, v1827188; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v97189; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v55, 8, v1837190; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v107191; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v65, 8, v1817192; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v87193; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v547194; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v64, v537195; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v557196; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v657197; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v397198; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v547199; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v537200; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xff, v1807201; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v54, 8, v1797202; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v117203; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v55, 8, v1787204; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xff, v1777205; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v52, 8, v527206; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v127207; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v65, 8, v1767208; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xff, v1677209; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v67, 8, v1667210; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v53, v547211; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v557212; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v64, v527213; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v657214; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, v66, v677215; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v97216; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v107217; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v537218; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v117219; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v527220; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v127221; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v547222; GFX11-FAKE16-NEXT:    s_clause 0x17223; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off7224; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:167225; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v9, v397226; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v10, v537227; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v11, v527228; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v12, v547229; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v137230; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v1657231; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v1647232; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v8, 8, v517233; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v147234; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v10, 8, v1637235; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v1627236; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v1617237; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v157238; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v14, 8, v1607239; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v67240; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v87241; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v107242; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v11, v127243; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v13, v147244; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v1517245; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v507246; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v167247; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v13, 8, v1507248; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v1497249; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v15, 8, v1487250; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v177251; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v1477252; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xff, v1467253; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v49, 8, v497254; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v117255; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v137256; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v14, v157257; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v16, v177258; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v39, v497259; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v57260; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v67261; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v77262; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v87263; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v97264; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v107265; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v117266; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v127267; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v137268; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v147269; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v67270; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v87271; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v107272; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v11, v127273; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v13, v147274; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v187275; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v1457276; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v1447277; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v13, 8, v1357278; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v197279; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v15, 8, v1347280; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v1337281; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v487282; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v207283; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v19, 8, v1327284; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v117285; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v137286; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v14, v157287; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v16, v177288; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v18, v197289; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v1317290; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v1307291; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v217292; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v18, 8, v1297293; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v1287294; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v20, 8, v387295; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v227296; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v1197297; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v1187298; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v39, 8, v1177299; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v167300; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v187301; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v207302; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v21, v227303; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v38, v397304; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v107305; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v117306; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v127307; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v137308; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v147309; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v157310; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v167311; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v177312; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v187313; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v197314; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v117315; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v137316; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v14, v157317; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v16, v177318; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v18, v197319; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v237320; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v1167321; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v1157322; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v18, 8, v377323; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v247324; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v20, 8, v1147325; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v1137326; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v1127327; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v257328; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v24, 8, v1037329; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v167330; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v187331; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v207332; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v21, v227333; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v23, v247334; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v1027335; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v367336; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v267337; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v23, 8, v1017338; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v1007339; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v25, 8, v997340; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v277341; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v987342; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v977343; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v35, 8, v357344; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v217345; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v237346; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v24, v257347; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v26, v277348; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v36, v357349; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v157350; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v167351; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v177352; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v187353; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v197354; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v207355; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v217356; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v227357; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v237358; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v247359; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v167360; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v187361; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v207362; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v21, v227363; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v23, v247364; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v287365; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v967366; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v877367; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v23, 8, v867368; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v297369; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v25, 8, v857370; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v847371; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v347372; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v307373; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v29, 8, v837374; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v217375; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v237376; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v24, v257377; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v26, v277378; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v28, v297379; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v827380; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v817381; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)7382; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v317383; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v28, 8, v807384; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v717385; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v30, 8, v337386; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v327387; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v32, 8, v707388; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v697389; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v687390; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v25, v267391; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v287392; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v307393; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v31, v327394; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v33, v347395; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v207396; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v217397; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v227398; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v237399; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v247400; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v257401; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v267402; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v277403; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v287404; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v297405; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v217406; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v237407; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v24, v257408; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v26, v277409; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v28, v297410; GFX11-FAKE16-NEXT:    s_clause 0x57411; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:327412; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:487413; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:647414; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:807415; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:967416; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:1127417; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Reload7418; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:127419; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:167420; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:207421; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:247422; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:287423; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:327424; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:367425; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:407426; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:447427; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:487428; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:527429; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:567430; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:607431; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:647432; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:687433; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:727434; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:767435; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:807436; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:847437; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:887438; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)7439; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]7440  %cmp = icmp eq i32 %b, 07441  br i1 %cmp, label %cmp.true, label %cmp.false7442 7443cmp.true:7444  %a1 = add <32 x i32> %a, splat (i32 3)7445  %a2 = bitcast <32 x i32> %a1 to <128 x i8>7446  br label %end7447 7448cmp.false:7449  %a3 = bitcast <32 x i32> %a to <128 x i8>7450  br label %end7451 7452end:7453  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]7454  ret <128 x i8> %phi7455}7456 7457define inreg <128 x i8> @bitcast_v32i32_to_v128i8_scalar(<32 x i32> inreg %a, i32 inreg %b) {7458; SI-LABEL: bitcast_v32i32_to_v128i8_scalar:7459; SI:       ; %bb.0:7460; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)7461; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -17462; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 ; 4-byte Folded Spill7463; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill7464; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill7465; SI-NEXT:    s_mov_b64 exec, s[4:5]7466; SI-NEXT:    s_waitcnt expcnt(2)7467; SI-NEXT:    v_writelane_b32 v21, s30, 07468; SI-NEXT:    v_writelane_b32 v21, s31, 17469; SI-NEXT:    v_writelane_b32 v21, s34, 27470; SI-NEXT:    v_writelane_b32 v21, s35, 37471; SI-NEXT:    v_writelane_b32 v21, s36, 47472; SI-NEXT:    v_writelane_b32 v21, s37, 57473; SI-NEXT:    v_writelane_b32 v21, s38, 67474; SI-NEXT:    v_writelane_b32 v21, s39, 77475; SI-NEXT:    v_writelane_b32 v21, s48, 87476; SI-NEXT:    v_writelane_b32 v21, s49, 97477; SI-NEXT:    v_writelane_b32 v21, s50, 107478; SI-NEXT:    v_writelane_b32 v21, s51, 117479; SI-NEXT:    v_writelane_b32 v21, s52, 127480; SI-NEXT:    v_writelane_b32 v21, s53, 137481; SI-NEXT:    v_writelane_b32 v21, s54, 147482; SI-NEXT:    v_writelane_b32 v21, s55, 157483; SI-NEXT:    v_writelane_b32 v21, s64, 167484; SI-NEXT:    v_writelane_b32 v21, s65, 177485; SI-NEXT:    v_writelane_b32 v21, s66, 187486; SI-NEXT:    v_writelane_b32 v21, s67, 197487; SI-NEXT:    v_writelane_b32 v21, s68, 207488; SI-NEXT:    v_mov_b32_e32 v20, s167489; SI-NEXT:    v_writelane_b32 v21, s69, 217490; SI-NEXT:    v_readfirstlane_b32 s56, v207491; SI-NEXT:    v_mov_b32_e32 v20, s177492; SI-NEXT:    v_writelane_b32 v21, s70, 227493; SI-NEXT:    v_readfirstlane_b32 s57, v207494; SI-NEXT:    v_mov_b32_e32 v20, s187495; SI-NEXT:    v_writelane_b32 v21, s71, 237496; SI-NEXT:    v_readfirstlane_b32 s46, v207497; SI-NEXT:    v_mov_b32_e32 v20, s197498; SI-NEXT:    v_writelane_b32 v21, s80, 247499; SI-NEXT:    v_readfirstlane_b32 s47, v207500; SI-NEXT:    v_mov_b32_e32 v20, s207501; SI-NEXT:    v_writelane_b32 v21, s81, 257502; SI-NEXT:    v_readfirstlane_b32 s44, v207503; SI-NEXT:    v_mov_b32_e32 v20, s217504; SI-NEXT:    v_writelane_b32 v21, s82, 267505; SI-NEXT:    v_readfirstlane_b32 s45, v207506; SI-NEXT:    v_mov_b32_e32 v20, s227507; SI-NEXT:    v_writelane_b32 v21, s83, 277508; SI-NEXT:    v_readfirstlane_b32 s42, v207509; SI-NEXT:    v_mov_b32_e32 v20, s237510; SI-NEXT:    v_writelane_b32 v21, s84, 287511; SI-NEXT:    v_readfirstlane_b32 s43, v207512; SI-NEXT:    v_mov_b32_e32 v20, s247513; SI-NEXT:    v_writelane_b32 v21, s85, 297514; SI-NEXT:    v_readfirstlane_b32 s40, v207515; SI-NEXT:    v_mov_b32_e32 v20, s257516; SI-NEXT:    v_writelane_b32 v21, s86, 307517; SI-NEXT:    v_readfirstlane_b32 s41, v207518; SI-NEXT:    v_mov_b32_e32 v20, s267519; SI-NEXT:    v_writelane_b32 v21, s87, 317520; SI-NEXT:    v_readfirstlane_b32 s24, v207521; SI-NEXT:    v_mov_b32_e32 v20, s277522; SI-NEXT:    v_writelane_b32 v21, s96, 327523; SI-NEXT:    v_readfirstlane_b32 s25, v207524; SI-NEXT:    v_mov_b32_e32 v20, s287525; SI-NEXT:    v_writelane_b32 v21, s97, 337526; SI-NEXT:    v_readfirstlane_b32 s22, v207527; SI-NEXT:    v_mov_b32_e32 v20, s297528; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v197529; SI-NEXT:    v_writelane_b32 v21, s98, 347530; SI-NEXT:    v_readfirstlane_b32 s23, v207531; SI-NEXT:    v_readfirstlane_b32 s20, v17532; SI-NEXT:    v_readfirstlane_b32 s21, v27533; SI-NEXT:    v_readfirstlane_b32 s18, v37534; SI-NEXT:    v_readfirstlane_b32 s19, v47535; SI-NEXT:    v_readfirstlane_b32 s16, v57536; SI-NEXT:    v_readfirstlane_b32 s17, v67537; SI-NEXT:    v_readfirstlane_b32 s14, v77538; SI-NEXT:    v_readfirstlane_b32 s15, v87539; SI-NEXT:    v_readfirstlane_b32 s12, v97540; SI-NEXT:    v_readfirstlane_b32 s13, v107541; SI-NEXT:    v_readfirstlane_b32 s10, v117542; SI-NEXT:    v_readfirstlane_b32 s11, v127543; SI-NEXT:    v_readfirstlane_b32 s8, v137544; SI-NEXT:    v_readfirstlane_b32 s9, v147545; SI-NEXT:    v_readfirstlane_b32 s6, v157546; SI-NEXT:    v_readfirstlane_b32 s7, v167547; SI-NEXT:    v_readfirstlane_b32 s4, v177548; SI-NEXT:    s_and_b64 s[26:27], vcc, exec7549; SI-NEXT:    v_readfirstlane_b32 s5, v187550; SI-NEXT:    v_writelane_b32 v21, s99, 357551; SI-NEXT:    ; implicit-def: $vgpr23 : SGPR spill to VGPR lane7552; SI-NEXT:    ; implicit-def: $vgpr22 : SGPR spill to VGPR lane7553; SI-NEXT:    s_cbranch_scc0 .LBB13_47554; SI-NEXT:  ; %bb.1: ; %cmp.false7555; SI-NEXT:    s_lshr_b32 s26, s5, 247556; SI-NEXT:    s_waitcnt expcnt(0)7557; SI-NEXT:    v_writelane_b32 v23, s26, 427558; SI-NEXT:    s_lshr_b32 s26, s5, 167559; SI-NEXT:    v_writelane_b32 v23, s26, 437560; SI-NEXT:    s_lshr_b32 s26, s5, 87561; SI-NEXT:    v_writelane_b32 v23, s26, 447562; SI-NEXT:    s_lshr_b32 s26, s7, 247563; SI-NEXT:    v_writelane_b32 v23, s26, 457564; SI-NEXT:    s_lshr_b32 s26, s7, 167565; SI-NEXT:    v_writelane_b32 v23, s26, 467566; SI-NEXT:    s_lshr_b32 s26, s7, 87567; SI-NEXT:    v_writelane_b32 v23, s26, 477568; SI-NEXT:    s_lshr_b32 s26, s9, 167569; SI-NEXT:    v_writelane_b32 v23, s26, 487570; SI-NEXT:    s_lshr_b32 s26, s9, 87571; SI-NEXT:    v_writelane_b32 v23, s26, 497572; SI-NEXT:    s_lshr_b32 s26, s11, 247573; SI-NEXT:    v_writelane_b32 v23, s26, 507574; SI-NEXT:    s_lshr_b32 s26, s11, 167575; SI-NEXT:    v_writelane_b32 v23, s26, 517576; SI-NEXT:    s_lshr_b32 s26, s11, 87577; SI-NEXT:    v_writelane_b32 v23, s26, 527578; SI-NEXT:    s_lshr_b32 s26, s13, 247579; SI-NEXT:    v_writelane_b32 v23, s26, 537580; SI-NEXT:    s_lshr_b32 s26, s13, 167581; SI-NEXT:    v_writelane_b32 v23, s26, 547582; SI-NEXT:    s_lshr_b32 s26, s13, 87583; SI-NEXT:    v_writelane_b32 v23, s26, 557584; SI-NEXT:    s_lshr_b32 s26, s15, 247585; SI-NEXT:    v_writelane_b32 v23, s26, 567586; SI-NEXT:    s_lshr_b32 s26, s15, 167587; SI-NEXT:    v_writelane_b32 v23, s26, 577588; SI-NEXT:    s_lshr_b32 s26, s15, 87589; SI-NEXT:    v_writelane_b32 v23, s26, 587590; SI-NEXT:    s_lshr_b32 s26, s17, 247591; SI-NEXT:    v_writelane_b32 v23, s26, 597592; SI-NEXT:    s_lshr_b32 s26, s17, 167593; SI-NEXT:    v_writelane_b32 v23, s26, 607594; SI-NEXT:    s_lshr_b32 s26, s17, 87595; SI-NEXT:    v_writelane_b32 v23, s26, 617596; SI-NEXT:    s_lshr_b32 s26, s19, 247597; SI-NEXT:    v_writelane_b32 v23, s26, 627598; SI-NEXT:    s_lshr_b32 s26, s19, 167599; SI-NEXT:    v_writelane_b32 v23, s26, 637600; SI-NEXT:    s_lshr_b32 s26, s19, 87601; SI-NEXT:    v_writelane_b32 v22, s26, 07602; SI-NEXT:    s_lshr_b32 s26, s21, 247603; SI-NEXT:    v_writelane_b32 v22, s26, 17604; SI-NEXT:    s_lshr_b32 s26, s21, 167605; SI-NEXT:    v_writelane_b32 v22, s26, 27606; SI-NEXT:    s_lshr_b32 s26, s21, 87607; SI-NEXT:    v_writelane_b32 v22, s26, 37608; SI-NEXT:    s_lshr_b32 s26, s23, 247609; SI-NEXT:    v_writelane_b32 v22, s26, 47610; SI-NEXT:    s_lshr_b32 s26, s23, 167611; SI-NEXT:    v_writelane_b32 v22, s26, 57612; SI-NEXT:    s_lshr_b32 s26, s23, 87613; SI-NEXT:    v_writelane_b32 v22, s26, 67614; SI-NEXT:    s_lshr_b32 s26, s25, 247615; SI-NEXT:    v_writelane_b32 v22, s26, 77616; SI-NEXT:    s_lshr_b32 s26, s25, 167617; SI-NEXT:    v_writelane_b32 v22, s26, 87618; SI-NEXT:    s_lshr_b32 s26, s25, 87619; SI-NEXT:    v_writelane_b32 v22, s26, 97620; SI-NEXT:    s_lshr_b32 s26, s41, 247621; SI-NEXT:    v_writelane_b32 v22, s26, 107622; SI-NEXT:    s_lshr_b32 s26, s41, 167623; SI-NEXT:    v_writelane_b32 v22, s26, 117624; SI-NEXT:    s_lshr_b32 s26, s41, 87625; SI-NEXT:    v_writelane_b32 v22, s26, 127626; SI-NEXT:    s_lshr_b32 s26, s43, 247627; SI-NEXT:    v_writelane_b32 v22, s26, 137628; SI-NEXT:    s_lshr_b32 s26, s43, 167629; SI-NEXT:    v_writelane_b32 v22, s26, 147630; SI-NEXT:    s_lshr_b32 s26, s43, 87631; SI-NEXT:    v_writelane_b32 v22, s26, 157632; SI-NEXT:    s_lshr_b32 s26, s45, 247633; SI-NEXT:    v_writelane_b32 v22, s26, 167634; SI-NEXT:    s_lshr_b32 s26, s45, 167635; SI-NEXT:    v_writelane_b32 v22, s26, 177636; SI-NEXT:    s_lshr_b64 s[26:27], s[4:5], 247637; SI-NEXT:    v_writelane_b32 v23, s26, 407638; SI-NEXT:    v_writelane_b32 v23, s27, 417639; SI-NEXT:    s_lshr_b64 s[26:27], s[4:5], 167640; SI-NEXT:    v_writelane_b32 v23, s26, 387641; SI-NEXT:    v_writelane_b32 v23, s27, 397642; SI-NEXT:    s_lshr_b64 s[26:27], s[4:5], 87643; SI-NEXT:    v_writelane_b32 v23, s26, 367644; SI-NEXT:    v_writelane_b32 v23, s27, 377645; SI-NEXT:    s_lshr_b64 s[26:27], s[6:7], 247646; SI-NEXT:    v_writelane_b32 v23, s26, 347647; SI-NEXT:    v_writelane_b32 v23, s27, 357648; SI-NEXT:    s_lshr_b64 s[26:27], s[6:7], 167649; SI-NEXT:    v_writelane_b32 v23, s26, 327650; SI-NEXT:    v_writelane_b32 v23, s27, 337651; SI-NEXT:    s_lshr_b64 s[26:27], s[6:7], 87652; SI-NEXT:    v_writelane_b32 v23, s26, 307653; SI-NEXT:    v_writelane_b32 v23, s27, 317654; SI-NEXT:    s_lshr_b64 s[26:27], s[8:9], 247655; SI-NEXT:    v_writelane_b32 v23, s26, 287656; SI-NEXT:    v_writelane_b32 v23, s27, 297657; SI-NEXT:    s_lshr_b64 s[26:27], s[8:9], 167658; SI-NEXT:    v_writelane_b32 v23, s26, 267659; SI-NEXT:    v_writelane_b32 v23, s27, 277660; SI-NEXT:    s_lshr_b64 s[26:27], s[8:9], 87661; SI-NEXT:    v_writelane_b32 v23, s26, 247662; SI-NEXT:    v_writelane_b32 v23, s27, 257663; SI-NEXT:    s_lshr_b64 s[26:27], s[10:11], 247664; SI-NEXT:    v_writelane_b32 v23, s26, 227665; SI-NEXT:    v_writelane_b32 v23, s27, 237666; SI-NEXT:    s_lshr_b64 s[26:27], s[10:11], 167667; SI-NEXT:    v_writelane_b32 v23, s26, 207668; SI-NEXT:    v_writelane_b32 v23, s27, 217669; SI-NEXT:    s_lshr_b64 s[26:27], s[10:11], 87670; SI-NEXT:    v_writelane_b32 v23, s26, 187671; SI-NEXT:    v_writelane_b32 v23, s27, 197672; SI-NEXT:    s_lshr_b64 s[26:27], s[12:13], 247673; SI-NEXT:    v_writelane_b32 v23, s26, 167674; SI-NEXT:    v_writelane_b32 v23, s27, 177675; SI-NEXT:    s_lshr_b64 s[26:27], s[12:13], 167676; SI-NEXT:    v_writelane_b32 v23, s26, 147677; SI-NEXT:    v_writelane_b32 v23, s27, 157678; SI-NEXT:    s_lshr_b64 s[26:27], s[12:13], 87679; SI-NEXT:    v_writelane_b32 v23, s26, 127680; SI-NEXT:    v_writelane_b32 v23, s27, 137681; SI-NEXT:    s_lshr_b64 s[26:27], s[14:15], 247682; SI-NEXT:    v_writelane_b32 v23, s26, 107683; SI-NEXT:    v_writelane_b32 v23, s27, 117684; SI-NEXT:    s_lshr_b64 s[26:27], s[14:15], 167685; SI-NEXT:    v_writelane_b32 v23, s26, 87686; SI-NEXT:    v_writelane_b32 v23, s27, 97687; SI-NEXT:    s_lshr_b64 s[26:27], s[14:15], 87688; SI-NEXT:    v_writelane_b32 v23, s26, 67689; SI-NEXT:    v_writelane_b32 v23, s27, 77690; SI-NEXT:    s_lshr_b64 s[26:27], s[16:17], 247691; SI-NEXT:    v_writelane_b32 v23, s26, 47692; SI-NEXT:    v_writelane_b32 v23, s27, 57693; SI-NEXT:    s_lshr_b64 s[26:27], s[16:17], 167694; SI-NEXT:    v_writelane_b32 v23, s26, 27695; SI-NEXT:    v_writelane_b32 v23, s27, 37696; SI-NEXT:    s_lshr_b64 s[26:27], s[16:17], 87697; SI-NEXT:    v_writelane_b32 v23, s26, 07698; SI-NEXT:    s_lshr_b32 s50, s9, 247699; SI-NEXT:    s_lshr_b32 s51, s45, 87700; SI-NEXT:    s_lshr_b32 s48, s47, 247701; SI-NEXT:    s_lshr_b32 s52, s47, 167702; SI-NEXT:    s_lshr_b32 s53, s47, 87703; SI-NEXT:    s_lshr_b32 s54, s57, 247704; SI-NEXT:    s_lshr_b32 s55, s57, 167705; SI-NEXT:    s_lshr_b32 s49, s57, 87706; SI-NEXT:    v_writelane_b32 v23, s27, 17707; SI-NEXT:    s_lshr_b64 s[64:65], s[18:19], 247708; SI-NEXT:    s_lshr_b64 s[66:67], s[18:19], 167709; SI-NEXT:    s_lshr_b64 s[68:69], s[18:19], 87710; SI-NEXT:    s_lshr_b64 s[70:71], s[20:21], 247711; SI-NEXT:    s_lshr_b64 s[80:81], s[20:21], 167712; SI-NEXT:    s_lshr_b64 s[82:83], s[20:21], 87713; SI-NEXT:    s_lshr_b64 s[84:85], s[22:23], 247714; SI-NEXT:    s_lshr_b64 s[86:87], s[22:23], 167715; SI-NEXT:    s_lshr_b64 s[96:97], s[22:23], 87716; SI-NEXT:    s_lshr_b64 s[98:99], s[24:25], 247717; SI-NEXT:    s_lshr_b64 s[26:27], s[24:25], 167718; SI-NEXT:    s_lshr_b64 s[28:29], s[24:25], 87719; SI-NEXT:    s_lshr_b64 s[58:59], s[40:41], 247720; SI-NEXT:    s_lshr_b64 s[60:61], s[40:41], 167721; SI-NEXT:    s_lshr_b64 s[62:63], s[40:41], 87722; SI-NEXT:    s_lshr_b64 s[72:73], s[42:43], 247723; SI-NEXT:    s_lshr_b64 s[74:75], s[42:43], 167724; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 87725; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 247726; SI-NEXT:    s_lshr_b64 s[88:89], s[44:45], 167727; SI-NEXT:    s_lshr_b64 s[90:91], s[44:45], 87728; SI-NEXT:    s_lshr_b64 s[92:93], s[46:47], 247729; SI-NEXT:    s_lshr_b64 s[94:95], s[46:47], 167730; SI-NEXT:    s_lshr_b64 s[30:31], s[46:47], 87731; SI-NEXT:    s_lshr_b64 s[34:35], s[56:57], 247732; SI-NEXT:    s_lshr_b64 s[36:37], s[56:57], 167733; SI-NEXT:    s_lshr_b64 s[38:39], s[56:57], 87734; SI-NEXT:    s_cbranch_execnz .LBB13_37735; SI-NEXT:  .LBB13_2: ; %cmp.true7736; SI-NEXT:    s_add_i32 s5, s5, 37737; SI-NEXT:    s_add_i32 s4, s4, 37738; SI-NEXT:    s_lshr_b64 s[26:27], s[4:5], 247739; SI-NEXT:    v_writelane_b32 v23, s26, 407740; SI-NEXT:    v_writelane_b32 v23, s27, 417741; SI-NEXT:    s_lshr_b64 s[26:27], s[4:5], 167742; SI-NEXT:    v_writelane_b32 v23, s26, 387743; SI-NEXT:    v_writelane_b32 v23, s27, 397744; SI-NEXT:    s_lshr_b32 s26, s5, 247745; SI-NEXT:    v_writelane_b32 v23, s26, 427746; SI-NEXT:    s_lshr_b32 s26, s5, 167747; SI-NEXT:    s_add_i32 s7, s7, 37748; SI-NEXT:    v_writelane_b32 v23, s26, 437749; SI-NEXT:    s_lshr_b32 s26, s5, 87750; SI-NEXT:    v_writelane_b32 v23, s26, 447751; SI-NEXT:    s_lshr_b32 s26, s7, 247752; SI-NEXT:    v_writelane_b32 v23, s26, 457753; SI-NEXT:    s_lshr_b32 s26, s7, 167754; SI-NEXT:    s_add_i32 s9, s9, 37755; SI-NEXT:    v_writelane_b32 v23, s26, 467756; SI-NEXT:    s_lshr_b32 s26, s7, 87757; SI-NEXT:    v_writelane_b32 v23, s26, 477758; SI-NEXT:    s_lshr_b32 s26, s9, 167759; SI-NEXT:    s_add_i32 s11, s11, 37760; SI-NEXT:    v_writelane_b32 v23, s26, 487761; SI-NEXT:    s_lshr_b32 s26, s9, 87762; SI-NEXT:    v_writelane_b32 v23, s26, 497763; SI-NEXT:    s_lshr_b32 s26, s11, 247764; SI-NEXT:    v_writelane_b32 v23, s26, 507765; SI-NEXT:    s_lshr_b32 s26, s11, 167766; SI-NEXT:    s_add_i32 s13, s13, 37767; SI-NEXT:    v_writelane_b32 v23, s26, 517768; SI-NEXT:    s_lshr_b32 s26, s11, 87769; SI-NEXT:    v_writelane_b32 v23, s26, 527770; SI-NEXT:    s_lshr_b32 s26, s13, 247771; SI-NEXT:    v_writelane_b32 v23, s26, 537772; SI-NEXT:    s_lshr_b32 s26, s13, 167773; SI-NEXT:    s_add_i32 s15, s15, 37774; SI-NEXT:    v_writelane_b32 v23, s26, 547775; SI-NEXT:    s_lshr_b32 s26, s13, 87776; SI-NEXT:    v_writelane_b32 v23, s26, 557777; SI-NEXT:    s_lshr_b32 s26, s15, 247778; SI-NEXT:    v_writelane_b32 v23, s26, 567779; SI-NEXT:    s_lshr_b32 s26, s15, 167780; SI-NEXT:    s_add_i32 s17, s17, 37781; SI-NEXT:    v_writelane_b32 v23, s26, 577782; SI-NEXT:    s_lshr_b32 s26, s15, 87783; SI-NEXT:    v_writelane_b32 v23, s26, 587784; SI-NEXT:    s_lshr_b32 s26, s17, 247785; SI-NEXT:    v_writelane_b32 v23, s26, 597786; SI-NEXT:    s_lshr_b32 s26, s17, 167787; SI-NEXT:    s_add_i32 s19, s19, 37788; SI-NEXT:    v_writelane_b32 v23, s26, 607789; SI-NEXT:    s_lshr_b32 s26, s17, 87790; SI-NEXT:    v_writelane_b32 v23, s26, 617791; SI-NEXT:    s_lshr_b32 s26, s19, 247792; SI-NEXT:    v_writelane_b32 v23, s26, 627793; SI-NEXT:    s_lshr_b32 s26, s19, 167794; SI-NEXT:    s_add_i32 s21, s21, 37795; SI-NEXT:    v_writelane_b32 v23, s26, 637796; SI-NEXT:    s_lshr_b32 s26, s19, 87797; SI-NEXT:    v_writelane_b32 v22, s26, 07798; SI-NEXT:    s_lshr_b32 s26, s21, 247799; SI-NEXT:    v_writelane_b32 v22, s26, 17800; SI-NEXT:    s_lshr_b32 s26, s21, 167801; SI-NEXT:    s_add_i32 s23, s23, 37802; SI-NEXT:    v_writelane_b32 v22, s26, 27803; SI-NEXT:    s_lshr_b32 s26, s21, 87804; SI-NEXT:    v_writelane_b32 v22, s26, 37805; SI-NEXT:    s_lshr_b32 s26, s23, 247806; SI-NEXT:    v_writelane_b32 v22, s26, 47807; SI-NEXT:    s_lshr_b32 s26, s23, 167808; SI-NEXT:    s_add_i32 s25, s25, 37809; SI-NEXT:    v_writelane_b32 v22, s26, 57810; SI-NEXT:    s_lshr_b32 s26, s23, 87811; SI-NEXT:    v_writelane_b32 v22, s26, 67812; SI-NEXT:    s_lshr_b32 s26, s25, 247813; SI-NEXT:    v_writelane_b32 v22, s26, 77814; SI-NEXT:    s_lshr_b32 s26, s25, 167815; SI-NEXT:    s_add_i32 s41, s41, 37816; SI-NEXT:    v_writelane_b32 v22, s26, 87817; SI-NEXT:    s_lshr_b32 s26, s25, 87818; SI-NEXT:    v_writelane_b32 v22, s26, 97819; SI-NEXT:    s_lshr_b32 s26, s41, 247820; SI-NEXT:    v_writelane_b32 v22, s26, 107821; SI-NEXT:    s_lshr_b32 s26, s41, 167822; SI-NEXT:    s_add_i32 s43, s43, 37823; SI-NEXT:    v_writelane_b32 v22, s26, 117824; SI-NEXT:    s_lshr_b32 s26, s41, 87825; SI-NEXT:    v_writelane_b32 v22, s26, 127826; SI-NEXT:    s_lshr_b32 s26, s43, 247827; SI-NEXT:    v_writelane_b32 v22, s26, 137828; SI-NEXT:    s_lshr_b32 s26, s43, 167829; SI-NEXT:    s_add_i32 s45, s45, 37830; SI-NEXT:    v_writelane_b32 v22, s26, 147831; SI-NEXT:    s_lshr_b32 s26, s43, 87832; SI-NEXT:    v_writelane_b32 v22, s26, 157833; SI-NEXT:    s_lshr_b32 s26, s45, 247834; SI-NEXT:    v_writelane_b32 v22, s26, 167835; SI-NEXT:    s_lshr_b32 s26, s45, 167836; SI-NEXT:    v_writelane_b32 v22, s26, 177837; SI-NEXT:    s_lshr_b64 s[26:27], s[4:5], 87838; SI-NEXT:    s_add_i32 s6, s6, 37839; SI-NEXT:    v_writelane_b32 v23, s26, 367840; SI-NEXT:    v_writelane_b32 v23, s27, 377841; SI-NEXT:    s_lshr_b64 s[26:27], s[6:7], 247842; SI-NEXT:    v_writelane_b32 v23, s26, 347843; SI-NEXT:    v_writelane_b32 v23, s27, 357844; SI-NEXT:    s_lshr_b64 s[26:27], s[6:7], 167845; SI-NEXT:    v_writelane_b32 v23, s26, 327846; SI-NEXT:    v_writelane_b32 v23, s27, 337847; SI-NEXT:    s_lshr_b64 s[26:27], s[6:7], 87848; SI-NEXT:    s_add_i32 s8, s8, 37849; SI-NEXT:    v_writelane_b32 v23, s26, 307850; SI-NEXT:    v_writelane_b32 v23, s27, 317851; SI-NEXT:    s_lshr_b64 s[26:27], s[8:9], 247852; SI-NEXT:    v_writelane_b32 v23, s26, 287853; SI-NEXT:    v_writelane_b32 v23, s27, 297854; SI-NEXT:    s_lshr_b64 s[26:27], s[8:9], 167855; SI-NEXT:    v_writelane_b32 v23, s26, 267856; SI-NEXT:    v_writelane_b32 v23, s27, 277857; SI-NEXT:    s_lshr_b64 s[26:27], s[8:9], 87858; SI-NEXT:    s_add_i32 s10, s10, 37859; SI-NEXT:    v_writelane_b32 v23, s26, 247860; SI-NEXT:    v_writelane_b32 v23, s27, 257861; SI-NEXT:    s_lshr_b64 s[26:27], s[10:11], 247862; SI-NEXT:    v_writelane_b32 v23, s26, 227863; SI-NEXT:    v_writelane_b32 v23, s27, 237864; SI-NEXT:    s_lshr_b64 s[26:27], s[10:11], 167865; SI-NEXT:    v_writelane_b32 v23, s26, 207866; SI-NEXT:    v_writelane_b32 v23, s27, 217867; SI-NEXT:    s_lshr_b64 s[26:27], s[10:11], 87868; SI-NEXT:    s_add_i32 s12, s12, 37869; SI-NEXT:    v_writelane_b32 v23, s26, 187870; SI-NEXT:    v_writelane_b32 v23, s27, 197871; SI-NEXT:    s_lshr_b64 s[26:27], s[12:13], 247872; SI-NEXT:    v_writelane_b32 v23, s26, 167873; SI-NEXT:    v_writelane_b32 v23, s27, 177874; SI-NEXT:    s_lshr_b64 s[26:27], s[12:13], 167875; SI-NEXT:    v_writelane_b32 v23, s26, 147876; SI-NEXT:    v_writelane_b32 v23, s27, 157877; SI-NEXT:    s_lshr_b64 s[26:27], s[12:13], 87878; SI-NEXT:    s_add_i32 s14, s14, 37879; SI-NEXT:    v_writelane_b32 v23, s26, 127880; SI-NEXT:    v_writelane_b32 v23, s27, 137881; SI-NEXT:    s_lshr_b64 s[26:27], s[14:15], 247882; SI-NEXT:    v_writelane_b32 v23, s26, 107883; SI-NEXT:    v_writelane_b32 v23, s27, 117884; SI-NEXT:    s_lshr_b64 s[26:27], s[14:15], 167885; SI-NEXT:    v_writelane_b32 v23, s26, 87886; SI-NEXT:    v_writelane_b32 v23, s27, 97887; SI-NEXT:    s_lshr_b64 s[26:27], s[14:15], 87888; SI-NEXT:    s_add_i32 s16, s16, 37889; SI-NEXT:    v_writelane_b32 v23, s26, 67890; SI-NEXT:    v_writelane_b32 v23, s27, 77891; SI-NEXT:    s_lshr_b64 s[26:27], s[16:17], 247892; SI-NEXT:    v_writelane_b32 v23, s26, 47893; SI-NEXT:    v_writelane_b32 v23, s27, 57894; SI-NEXT:    s_lshr_b64 s[26:27], s[16:17], 167895; SI-NEXT:    v_writelane_b32 v23, s26, 27896; SI-NEXT:    v_writelane_b32 v23, s27, 37897; SI-NEXT:    s_lshr_b64 s[26:27], s[16:17], 87898; SI-NEXT:    s_add_i32 s57, s57, 37899; SI-NEXT:    s_add_i32 s56, s56, 37900; SI-NEXT:    s_add_i32 s47, s47, 37901; SI-NEXT:    s_add_i32 s46, s46, 37902; SI-NEXT:    s_add_i32 s44, s44, 37903; SI-NEXT:    s_add_i32 s42, s42, 37904; SI-NEXT:    s_add_i32 s40, s40, 37905; SI-NEXT:    s_add_i32 s24, s24, 37906; SI-NEXT:    s_add_i32 s22, s22, 37907; SI-NEXT:    s_add_i32 s20, s20, 37908; SI-NEXT:    s_add_i32 s18, s18, 37909; SI-NEXT:    v_writelane_b32 v23, s26, 07910; SI-NEXT:    s_lshr_b32 s50, s9, 247911; SI-NEXT:    s_lshr_b32 s51, s45, 87912; SI-NEXT:    s_lshr_b32 s48, s47, 247913; SI-NEXT:    s_lshr_b32 s52, s47, 167914; SI-NEXT:    s_lshr_b32 s53, s47, 87915; SI-NEXT:    s_lshr_b32 s54, s57, 247916; SI-NEXT:    s_lshr_b32 s55, s57, 167917; SI-NEXT:    s_lshr_b32 s49, s57, 87918; SI-NEXT:    v_writelane_b32 v23, s27, 17919; SI-NEXT:    s_lshr_b64 s[64:65], s[18:19], 247920; SI-NEXT:    s_lshr_b64 s[66:67], s[18:19], 167921; SI-NEXT:    s_lshr_b64 s[68:69], s[18:19], 87922; SI-NEXT:    s_lshr_b64 s[70:71], s[20:21], 247923; SI-NEXT:    s_lshr_b64 s[80:81], s[20:21], 167924; SI-NEXT:    s_lshr_b64 s[82:83], s[20:21], 87925; SI-NEXT:    s_lshr_b64 s[84:85], s[22:23], 247926; SI-NEXT:    s_lshr_b64 s[86:87], s[22:23], 167927; SI-NEXT:    s_lshr_b64 s[96:97], s[22:23], 87928; SI-NEXT:    s_lshr_b64 s[98:99], s[24:25], 247929; SI-NEXT:    s_lshr_b64 s[26:27], s[24:25], 167930; SI-NEXT:    s_lshr_b64 s[28:29], s[24:25], 87931; SI-NEXT:    s_lshr_b64 s[58:59], s[40:41], 247932; SI-NEXT:    s_lshr_b64 s[60:61], s[40:41], 167933; SI-NEXT:    s_lshr_b64 s[62:63], s[40:41], 87934; SI-NEXT:    s_lshr_b64 s[72:73], s[42:43], 247935; SI-NEXT:    s_lshr_b64 s[74:75], s[42:43], 167936; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 87937; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 247938; SI-NEXT:    s_lshr_b64 s[88:89], s[44:45], 167939; SI-NEXT:    s_lshr_b64 s[90:91], s[44:45], 87940; SI-NEXT:    s_lshr_b64 s[92:93], s[46:47], 247941; SI-NEXT:    s_lshr_b64 s[94:95], s[46:47], 167942; SI-NEXT:    s_lshr_b64 s[30:31], s[46:47], 87943; SI-NEXT:    s_lshr_b64 s[34:35], s[56:57], 247944; SI-NEXT:    s_lshr_b64 s[36:37], s[56:57], 167945; SI-NEXT:    s_lshr_b64 s[38:39], s[56:57], 87946; SI-NEXT:  .LBB13_3: ; %end7947; SI-NEXT:    s_lshl_b32 s27, s38, 87948; SI-NEXT:    s_and_b32 s29, s56, 0xff7949; SI-NEXT:    s_or_b32 s27, s29, s277950; SI-NEXT:    s_and_b32 s29, s36, 0xff7951; SI-NEXT:    s_lshl_b32 s56, s34, 247952; SI-NEXT:    s_lshl_b32 s29, s29, 167953; SI-NEXT:    s_or_b32 s29, s56, s297954; SI-NEXT:    s_and_b32 s27, s27, 0xffff7955; SI-NEXT:    s_or_b32 s27, s27, s297956; SI-NEXT:    v_mov_b32_e32 v1, s277957; SI-NEXT:    s_and_b32 s27, s57, 0xff7958; SI-NEXT:    s_lshl_b32 s29, s49, 87959; SI-NEXT:    s_or_b32 s27, s27, s297960; SI-NEXT:    s_and_b32 s29, s55, 0xff7961; SI-NEXT:    s_lshl_b32 s29, s29, 167962; SI-NEXT:    s_lshl_b32 s56, s54, 247963; SI-NEXT:    s_or_b32 s29, s56, s297964; SI-NEXT:    s_and_b32 s27, s27, 0xffff7965; SI-NEXT:    s_or_b32 s27, s27, s297966; SI-NEXT:    v_mov_b32_e32 v2, s277967; SI-NEXT:    s_lshl_b32 s27, s30, 87968; SI-NEXT:    s_and_b32 s29, s46, 0xff7969; SI-NEXT:    s_or_b32 s27, s29, s277970; SI-NEXT:    s_and_b32 s29, s94, 0xff7971; SI-NEXT:    s_lshl_b32 s46, s92, 247972; SI-NEXT:    s_lshl_b32 s29, s29, 167973; SI-NEXT:    s_or_b32 s29, s46, s297974; SI-NEXT:    s_and_b32 s27, s27, 0xffff7975; SI-NEXT:    s_or_b32 s27, s27, s297976; SI-NEXT:    v_mov_b32_e32 v3, s277977; SI-NEXT:    s_and_b32 s27, s47, 0xff7978; SI-NEXT:    s_lshl_b32 s29, s53, 87979; SI-NEXT:    s_or_b32 s27, s27, s297980; SI-NEXT:    s_and_b32 s29, s52, 0xff7981; SI-NEXT:    s_lshl_b32 s29, s29, 167982; SI-NEXT:    s_lshl_b32 s46, s48, 247983; SI-NEXT:    s_or_b32 s29, s46, s297984; SI-NEXT:    s_and_b32 s27, s27, 0xffff7985; SI-NEXT:    s_or_b32 s27, s27, s297986; SI-NEXT:    v_mov_b32_e32 v4, s277987; SI-NEXT:    s_lshl_b32 s27, s90, 87988; SI-NEXT:    s_and_b32 s29, s44, 0xff7989; SI-NEXT:    s_or_b32 s27, s29, s277990; SI-NEXT:    s_and_b32 s29, s88, 0xff7991; SI-NEXT:    s_lshl_b32 s44, s78, 247992; SI-NEXT:    s_lshl_b32 s29, s29, 167993; SI-NEXT:    s_or_b32 s29, s44, s297994; SI-NEXT:    s_and_b32 s27, s27, 0xffff7995; SI-NEXT:    s_or_b32 s27, s27, s297996; SI-NEXT:    v_mov_b32_e32 v5, s277997; SI-NEXT:    s_and_b32 s27, s45, 0xff7998; SI-NEXT:    s_lshl_b32 s29, s51, 87999; SI-NEXT:    s_or_b32 s27, s27, s298000; SI-NEXT:    v_readlane_b32 s29, v22, 178001; SI-NEXT:    s_and_b32 s29, s29, 0xff8002; SI-NEXT:    v_readlane_b32 s44, v22, 168003; SI-NEXT:    s_lshl_b32 s29, s29, 168004; SI-NEXT:    s_lshl_b32 s44, s44, 248005; SI-NEXT:    s_or_b32 s29, s44, s298006; SI-NEXT:    s_and_b32 s27, s27, 0xffff8007; SI-NEXT:    s_or_b32 s27, s27, s298008; SI-NEXT:    v_mov_b32_e32 v6, s278009; SI-NEXT:    s_lshl_b32 s27, s76, 88010; SI-NEXT:    s_and_b32 s29, s42, 0xff8011; SI-NEXT:    s_or_b32 s27, s29, s278012; SI-NEXT:    s_and_b32 s29, s74, 0xff8013; SI-NEXT:    s_lshl_b32 s42, s72, 248014; SI-NEXT:    s_lshl_b32 s29, s29, 168015; SI-NEXT:    s_or_b32 s29, s42, s298016; SI-NEXT:    s_and_b32 s27, s27, 0xffff8017; SI-NEXT:    s_or_b32 s27, s27, s298018; SI-NEXT:    v_readlane_b32 s29, v22, 158019; SI-NEXT:    v_mov_b32_e32 v7, s278020; SI-NEXT:    s_and_b32 s27, s43, 0xff8021; SI-NEXT:    s_lshl_b32 s29, s29, 88022; SI-NEXT:    s_or_b32 s27, s27, s298023; SI-NEXT:    v_readlane_b32 s29, v22, 148024; SI-NEXT:    s_and_b32 s29, s29, 0xff8025; SI-NEXT:    v_readlane_b32 s42, v22, 138026; SI-NEXT:    s_lshl_b32 s29, s29, 168027; SI-NEXT:    s_lshl_b32 s42, s42, 248028; SI-NEXT:    s_or_b32 s29, s42, s298029; SI-NEXT:    s_and_b32 s27, s27, 0xffff8030; SI-NEXT:    s_or_b32 s27, s27, s298031; SI-NEXT:    v_mov_b32_e32 v8, s278032; SI-NEXT:    s_lshl_b32 s27, s62, 88033; SI-NEXT:    s_and_b32 s29, s40, 0xff8034; SI-NEXT:    s_or_b32 s27, s29, s278035; SI-NEXT:    s_and_b32 s29, s60, 0xff8036; SI-NEXT:    s_lshl_b32 s40, s58, 248037; SI-NEXT:    s_lshl_b32 s29, s29, 168038; SI-NEXT:    s_or_b32 s29, s40, s298039; SI-NEXT:    s_and_b32 s27, s27, 0xffff8040; SI-NEXT:    s_or_b32 s27, s27, s298041; SI-NEXT:    v_readlane_b32 s29, v22, 128042; SI-NEXT:    v_mov_b32_e32 v9, s278043; SI-NEXT:    s_and_b32 s27, s41, 0xff8044; SI-NEXT:    s_lshl_b32 s29, s29, 88045; SI-NEXT:    s_or_b32 s27, s27, s298046; SI-NEXT:    v_readlane_b32 s29, v22, 118047; SI-NEXT:    s_and_b32 s29, s29, 0xff8048; SI-NEXT:    v_readlane_b32 s40, v22, 108049; SI-NEXT:    s_lshl_b32 s29, s29, 168050; SI-NEXT:    s_lshl_b32 s40, s40, 248051; SI-NEXT:    s_or_b32 s29, s40, s298052; SI-NEXT:    s_and_b32 s27, s27, 0xffff8053; SI-NEXT:    s_or_b32 s27, s27, s298054; SI-NEXT:    v_mov_b32_e32 v10, s278055; SI-NEXT:    s_lshl_b32 s27, s28, 88056; SI-NEXT:    s_and_b32 s24, s24, 0xff8057; SI-NEXT:    s_and_b32 s26, s26, 0xff8058; SI-NEXT:    s_or_b32 s24, s24, s278059; SI-NEXT:    s_lshl_b32 s27, s98, 248060; SI-NEXT:    s_lshl_b32 s26, s26, 168061; SI-NEXT:    s_or_b32 s26, s27, s268062; SI-NEXT:    s_and_b32 s24, s24, 0xffff8063; SI-NEXT:    s_or_b32 s24, s24, s268064; SI-NEXT:    v_mov_b32_e32 v11, s248065; SI-NEXT:    s_and_b32 s24, s25, 0xff8066; SI-NEXT:    v_readlane_b32 s25, v22, 98067; SI-NEXT:    s_lshl_b32 s25, s25, 88068; SI-NEXT:    s_or_b32 s24, s24, s258069; SI-NEXT:    v_readlane_b32 s25, v22, 88070; SI-NEXT:    s_and_b32 s25, s25, 0xff8071; SI-NEXT:    v_readlane_b32 s26, v22, 78072; SI-NEXT:    s_lshl_b32 s25, s25, 168073; SI-NEXT:    s_lshl_b32 s26, s26, 248074; SI-NEXT:    s_or_b32 s25, s26, s258075; SI-NEXT:    s_and_b32 s24, s24, 0xffff8076; SI-NEXT:    s_or_b32 s24, s24, s258077; SI-NEXT:    v_mov_b32_e32 v12, s248078; SI-NEXT:    s_lshl_b32 s24, s96, 88079; SI-NEXT:    s_and_b32 s22, s22, 0xff8080; SI-NEXT:    s_or_b32 s22, s22, s248081; SI-NEXT:    s_and_b32 s24, s86, 0xff8082; SI-NEXT:    s_lshl_b32 s25, s84, 248083; SI-NEXT:    s_lshl_b32 s24, s24, 168084; SI-NEXT:    s_or_b32 s24, s25, s248085; SI-NEXT:    s_and_b32 s22, s22, 0xffff8086; SI-NEXT:    s_or_b32 s22, s22, s248087; SI-NEXT:    v_mov_b32_e32 v13, s228088; SI-NEXT:    s_and_b32 s22, s23, 0xff8089; SI-NEXT:    v_readlane_b32 s23, v22, 68090; SI-NEXT:    s_lshl_b32 s23, s23, 88091; SI-NEXT:    s_or_b32 s22, s22, s238092; SI-NEXT:    v_readlane_b32 s23, v22, 58093; SI-NEXT:    s_and_b32 s23, s23, 0xff8094; SI-NEXT:    v_readlane_b32 s24, v22, 48095; SI-NEXT:    s_lshl_b32 s23, s23, 168096; SI-NEXT:    s_lshl_b32 s24, s24, 248097; SI-NEXT:    s_or_b32 s23, s24, s238098; SI-NEXT:    s_and_b32 s22, s22, 0xffff8099; SI-NEXT:    s_or_b32 s22, s22, s238100; SI-NEXT:    v_mov_b32_e32 v14, s228101; SI-NEXT:    s_lshl_b32 s22, s82, 88102; SI-NEXT:    s_and_b32 s20, s20, 0xff8103; SI-NEXT:    s_or_b32 s20, s20, s228104; SI-NEXT:    s_and_b32 s22, s80, 0xff8105; SI-NEXT:    s_lshl_b32 s23, s70, 248106; SI-NEXT:    s_lshl_b32 s22, s22, 168107; SI-NEXT:    s_or_b32 s22, s23, s228108; SI-NEXT:    s_and_b32 s20, s20, 0xffff8109; SI-NEXT:    s_or_b32 s20, s20, s228110; SI-NEXT:    v_mov_b32_e32 v15, s208111; SI-NEXT:    s_and_b32 s20, s21, 0xff8112; SI-NEXT:    v_readlane_b32 s21, v22, 38113; SI-NEXT:    s_lshl_b32 s21, s21, 88114; SI-NEXT:    s_or_b32 s20, s20, s218115; SI-NEXT:    v_readlane_b32 s21, v22, 28116; SI-NEXT:    s_and_b32 s21, s21, 0xff8117; SI-NEXT:    v_readlane_b32 s22, v22, 18118; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen8119; SI-NEXT:    s_waitcnt expcnt(0)8120; SI-NEXT:    v_add_i32_e32 v1, vcc, 4, v08121; SI-NEXT:    s_lshl_b32 s21, s21, 168122; SI-NEXT:    s_lshl_b32 s22, s22, 248123; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8124; SI-NEXT:    v_add_i32_e32 v1, vcc, 8, v08125; SI-NEXT:    s_or_b32 s21, s22, s218126; SI-NEXT:    s_and_b32 s20, s20, 0xffff8127; SI-NEXT:    buffer_store_dword v3, v1, s[0:3], 0 offen8128; SI-NEXT:    v_add_i32_e32 v1, vcc, 12, v08129; SI-NEXT:    s_or_b32 s20, s20, s218130; SI-NEXT:    buffer_store_dword v4, v1, s[0:3], 0 offen8131; SI-NEXT:    v_add_i32_e32 v1, vcc, 16, v08132; SI-NEXT:    v_mov_b32_e32 v16, s208133; SI-NEXT:    s_lshl_b32 s20, s68, 88134; SI-NEXT:    s_and_b32 s18, s18, 0xff8135; SI-NEXT:    buffer_store_dword v5, v1, s[0:3], 0 offen8136; SI-NEXT:    v_add_i32_e32 v1, vcc, 20, v08137; SI-NEXT:    s_or_b32 s18, s18, s208138; SI-NEXT:    s_and_b32 s20, s66, 0xff8139; SI-NEXT:    buffer_store_dword v6, v1, s[0:3], 0 offen8140; SI-NEXT:    v_add_i32_e32 v1, vcc, 24, v08141; SI-NEXT:    s_lshl_b32 s21, s64, 248142; SI-NEXT:    s_lshl_b32 s20, s20, 168143; SI-NEXT:    buffer_store_dword v7, v1, s[0:3], 0 offen8144; SI-NEXT:    v_add_i32_e32 v1, vcc, 28, v08145; SI-NEXT:    s_or_b32 s20, s21, s208146; SI-NEXT:    s_and_b32 s18, s18, 0xffff8147; SI-NEXT:    buffer_store_dword v8, v1, s[0:3], 0 offen8148; SI-NEXT:    v_add_i32_e32 v1, vcc, 32, v08149; SI-NEXT:    s_or_b32 s18, s18, s208150; SI-NEXT:    buffer_store_dword v9, v1, s[0:3], 0 offen8151; SI-NEXT:    v_add_i32_e32 v1, vcc, 36, v08152; SI-NEXT:    buffer_store_dword v10, v1, s[0:3], 0 offen8153; SI-NEXT:    v_add_i32_e32 v1, vcc, 40, v08154; SI-NEXT:    v_mov_b32_e32 v2, s188155; SI-NEXT:    s_and_b32 s18, s19, 0xff8156; SI-NEXT:    v_readlane_b32 s19, v22, 08157; SI-NEXT:    buffer_store_dword v11, v1, s[0:3], 0 offen8158; SI-NEXT:    v_add_i32_e32 v1, vcc, 44, v08159; SI-NEXT:    s_lshl_b32 s19, s19, 88160; SI-NEXT:    buffer_store_dword v12, v1, s[0:3], 0 offen8161; SI-NEXT:    v_add_i32_e32 v1, vcc, 48, v08162; SI-NEXT:    s_or_b32 s18, s18, s198163; SI-NEXT:    v_readlane_b32 s19, v23, 638164; SI-NEXT:    buffer_store_dword v13, v1, s[0:3], 0 offen8165; SI-NEXT:    v_add_i32_e32 v1, vcc, 52, v08166; SI-NEXT:    s_and_b32 s19, s19, 0xff8167; SI-NEXT:    v_readlane_b32 s20, v23, 628168; SI-NEXT:    buffer_store_dword v14, v1, s[0:3], 0 offen8169; SI-NEXT:    v_add_i32_e32 v1, vcc, 56, v08170; SI-NEXT:    s_lshl_b32 s19, s19, 168171; SI-NEXT:    s_lshl_b32 s20, s20, 248172; SI-NEXT:    buffer_store_dword v15, v1, s[0:3], 0 offen8173; SI-NEXT:    v_add_i32_e32 v1, vcc, 60, v08174; SI-NEXT:    s_and_b32 s18, s18, 0xffff8175; SI-NEXT:    s_or_b32 s19, s20, s198176; SI-NEXT:    buffer_store_dword v16, v1, s[0:3], 0 offen8177; SI-NEXT:    v_add_i32_e32 v1, vcc, 64, v08178; SI-NEXT:    s_or_b32 s18, s18, s198179; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8180; SI-NEXT:    s_waitcnt expcnt(0)8181; SI-NEXT:    v_mov_b32_e32 v2, s188182; SI-NEXT:    v_readlane_b32 s18, v23, 08183; SI-NEXT:    s_and_b32 s16, s16, 0xff8184; SI-NEXT:    v_readlane_b32 s19, v23, 18185; SI-NEXT:    s_lshl_b32 s18, s18, 88186; SI-NEXT:    s_or_b32 s16, s16, s188187; SI-NEXT:    v_readlane_b32 s18, v23, 28188; SI-NEXT:    v_readlane_b32 s19, v23, 38189; SI-NEXT:    s_and_b32 s18, s18, 0xff8190; SI-NEXT:    v_readlane_b32 s20, v23, 48191; SI-NEXT:    s_lshl_b32 s18, s18, 168192; SI-NEXT:    s_lshl_b32 s19, s20, 248193; SI-NEXT:    s_and_b32 s16, s16, 0xffff8194; SI-NEXT:    s_or_b32 s18, s19, s188195; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x44, v08196; SI-NEXT:    s_or_b32 s16, s16, s188197; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8198; SI-NEXT:    s_waitcnt expcnt(0)8199; SI-NEXT:    v_mov_b32_e32 v2, s168200; SI-NEXT:    s_and_b32 s16, s17, 0xff8201; SI-NEXT:    v_readlane_b32 s17, v23, 618202; SI-NEXT:    s_lshl_b32 s17, s17, 88203; SI-NEXT:    s_or_b32 s16, s16, s178204; SI-NEXT:    v_readlane_b32 s17, v23, 608205; SI-NEXT:    s_and_b32 s17, s17, 0xff8206; SI-NEXT:    v_readlane_b32 s18, v23, 598207; SI-NEXT:    s_lshl_b32 s17, s17, 168208; SI-NEXT:    s_lshl_b32 s18, s18, 248209; SI-NEXT:    s_and_b32 s16, s16, 0xffff8210; SI-NEXT:    s_or_b32 s17, s18, s178211; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x48, v08212; SI-NEXT:    s_or_b32 s16, s16, s178213; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8214; SI-NEXT:    s_waitcnt expcnt(0)8215; SI-NEXT:    v_mov_b32_e32 v2, s168216; SI-NEXT:    v_readlane_b32 s16, v23, 68217; SI-NEXT:    s_and_b32 s14, s14, 0xff8218; SI-NEXT:    v_readlane_b32 s17, v23, 78219; SI-NEXT:    s_lshl_b32 s16, s16, 88220; SI-NEXT:    s_or_b32 s14, s14, s168221; SI-NEXT:    v_readlane_b32 s16, v23, 88222; SI-NEXT:    v_readlane_b32 s17, v23, 98223; SI-NEXT:    s_and_b32 s16, s16, 0xff8224; SI-NEXT:    v_readlane_b32 s18, v23, 108225; SI-NEXT:    s_lshl_b32 s16, s16, 168226; SI-NEXT:    s_lshl_b32 s17, s18, 248227; SI-NEXT:    s_and_b32 s14, s14, 0xffff8228; SI-NEXT:    s_or_b32 s16, s17, s168229; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x4c, v08230; SI-NEXT:    s_or_b32 s14, s14, s168231; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8232; SI-NEXT:    s_waitcnt expcnt(0)8233; SI-NEXT:    v_mov_b32_e32 v2, s148234; SI-NEXT:    s_and_b32 s14, s15, 0xff8235; SI-NEXT:    v_readlane_b32 s15, v23, 588236; SI-NEXT:    s_lshl_b32 s15, s15, 88237; SI-NEXT:    s_or_b32 s14, s14, s158238; SI-NEXT:    v_readlane_b32 s15, v23, 578239; SI-NEXT:    s_and_b32 s15, s15, 0xff8240; SI-NEXT:    v_readlane_b32 s16, v23, 568241; SI-NEXT:    s_lshl_b32 s15, s15, 168242; SI-NEXT:    s_lshl_b32 s16, s16, 248243; SI-NEXT:    s_and_b32 s14, s14, 0xffff8244; SI-NEXT:    s_or_b32 s15, s16, s158245; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x50, v08246; SI-NEXT:    s_or_b32 s14, s14, s158247; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8248; SI-NEXT:    s_waitcnt expcnt(0)8249; SI-NEXT:    v_mov_b32_e32 v2, s148250; SI-NEXT:    v_readlane_b32 s14, v23, 128251; SI-NEXT:    s_and_b32 s12, s12, 0xff8252; SI-NEXT:    v_readlane_b32 s15, v23, 138253; SI-NEXT:    s_lshl_b32 s14, s14, 88254; SI-NEXT:    s_or_b32 s12, s12, s148255; SI-NEXT:    v_readlane_b32 s14, v23, 148256; SI-NEXT:    v_readlane_b32 s15, v23, 158257; SI-NEXT:    s_and_b32 s14, s14, 0xff8258; SI-NEXT:    v_readlane_b32 s16, v23, 168259; SI-NEXT:    s_lshl_b32 s14, s14, 168260; SI-NEXT:    s_lshl_b32 s15, s16, 248261; SI-NEXT:    s_and_b32 s12, s12, 0xffff8262; SI-NEXT:    s_or_b32 s14, s15, s148263; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x54, v08264; SI-NEXT:    s_or_b32 s12, s12, s148265; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8266; SI-NEXT:    s_waitcnt expcnt(0)8267; SI-NEXT:    v_mov_b32_e32 v2, s128268; SI-NEXT:    s_and_b32 s12, s13, 0xff8269; SI-NEXT:    v_readlane_b32 s13, v23, 558270; SI-NEXT:    s_lshl_b32 s13, s13, 88271; SI-NEXT:    s_or_b32 s12, s12, s138272; SI-NEXT:    v_readlane_b32 s13, v23, 548273; SI-NEXT:    s_and_b32 s13, s13, 0xff8274; SI-NEXT:    v_readlane_b32 s14, v23, 538275; SI-NEXT:    s_lshl_b32 s13, s13, 168276; SI-NEXT:    s_lshl_b32 s14, s14, 248277; SI-NEXT:    s_and_b32 s12, s12, 0xffff8278; SI-NEXT:    s_or_b32 s13, s14, s138279; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x58, v08280; SI-NEXT:    s_or_b32 s12, s12, s138281; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8282; SI-NEXT:    s_waitcnt expcnt(0)8283; SI-NEXT:    v_mov_b32_e32 v2, s128284; SI-NEXT:    v_readlane_b32 s12, v23, 188285; SI-NEXT:    s_and_b32 s10, s10, 0xff8286; SI-NEXT:    v_readlane_b32 s13, v23, 198287; SI-NEXT:    s_lshl_b32 s12, s12, 88288; SI-NEXT:    s_or_b32 s10, s10, s128289; SI-NEXT:    v_readlane_b32 s12, v23, 208290; SI-NEXT:    v_readlane_b32 s13, v23, 218291; SI-NEXT:    s_and_b32 s12, s12, 0xff8292; SI-NEXT:    v_readlane_b32 s14, v23, 228293; SI-NEXT:    s_lshl_b32 s12, s12, 168294; SI-NEXT:    s_lshl_b32 s13, s14, 248295; SI-NEXT:    s_and_b32 s10, s10, 0xffff8296; SI-NEXT:    s_or_b32 s12, s13, s128297; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x5c, v08298; SI-NEXT:    s_or_b32 s10, s10, s128299; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8300; SI-NEXT:    s_waitcnt expcnt(0)8301; SI-NEXT:    v_mov_b32_e32 v2, s108302; SI-NEXT:    s_and_b32 s10, s11, 0xff8303; SI-NEXT:    v_readlane_b32 s11, v23, 528304; SI-NEXT:    s_lshl_b32 s11, s11, 88305; SI-NEXT:    s_or_b32 s10, s10, s118306; SI-NEXT:    v_readlane_b32 s11, v23, 518307; SI-NEXT:    s_and_b32 s11, s11, 0xff8308; SI-NEXT:    v_readlane_b32 s12, v23, 508309; SI-NEXT:    s_lshl_b32 s11, s11, 168310; SI-NEXT:    s_lshl_b32 s12, s12, 248311; SI-NEXT:    s_and_b32 s10, s10, 0xffff8312; SI-NEXT:    s_or_b32 s11, s12, s118313; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x60, v08314; SI-NEXT:    s_or_b32 s10, s10, s118315; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8316; SI-NEXT:    s_waitcnt expcnt(0)8317; SI-NEXT:    v_mov_b32_e32 v2, s108318; SI-NEXT:    v_readlane_b32 s10, v23, 248319; SI-NEXT:    s_and_b32 s8, s8, 0xff8320; SI-NEXT:    v_readlane_b32 s11, v23, 258321; SI-NEXT:    s_lshl_b32 s10, s10, 88322; SI-NEXT:    s_or_b32 s8, s8, s108323; SI-NEXT:    v_readlane_b32 s10, v23, 268324; SI-NEXT:    v_readlane_b32 s11, v23, 278325; SI-NEXT:    s_and_b32 s10, s10, 0xff8326; SI-NEXT:    v_readlane_b32 s12, v23, 288327; SI-NEXT:    s_lshl_b32 s10, s10, 168328; SI-NEXT:    s_lshl_b32 s11, s12, 248329; SI-NEXT:    s_and_b32 s8, s8, 0xffff8330; SI-NEXT:    s_or_b32 s10, s11, s108331; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x64, v08332; SI-NEXT:    s_or_b32 s8, s8, s108333; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8334; SI-NEXT:    s_waitcnt expcnt(0)8335; SI-NEXT:    v_mov_b32_e32 v2, s88336; SI-NEXT:    s_and_b32 s8, s9, 0xff8337; SI-NEXT:    v_readlane_b32 s9, v23, 498338; SI-NEXT:    s_lshl_b32 s9, s9, 88339; SI-NEXT:    s_or_b32 s8, s8, s98340; SI-NEXT:    v_readlane_b32 s9, v23, 488341; SI-NEXT:    s_and_b32 s9, s9, 0xff8342; SI-NEXT:    s_lshl_b32 s9, s9, 168343; SI-NEXT:    s_lshl_b32 s10, s50, 248344; SI-NEXT:    s_and_b32 s8, s8, 0xffff8345; SI-NEXT:    s_or_b32 s9, s10, s98346; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x68, v08347; SI-NEXT:    s_or_b32 s8, s8, s98348; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8349; SI-NEXT:    s_waitcnt expcnt(0)8350; SI-NEXT:    v_mov_b32_e32 v2, s88351; SI-NEXT:    v_readlane_b32 s8, v23, 308352; SI-NEXT:    s_and_b32 s6, s6, 0xff8353; SI-NEXT:    v_readlane_b32 s9, v23, 318354; SI-NEXT:    s_lshl_b32 s8, s8, 88355; SI-NEXT:    s_or_b32 s6, s6, s88356; SI-NEXT:    v_readlane_b32 s8, v23, 328357; SI-NEXT:    v_readlane_b32 s9, v23, 338358; SI-NEXT:    s_and_b32 s8, s8, 0xff8359; SI-NEXT:    v_readlane_b32 s10, v23, 348360; SI-NEXT:    s_lshl_b32 s8, s8, 168361; SI-NEXT:    s_lshl_b32 s9, s10, 248362; SI-NEXT:    s_and_b32 s6, s6, 0xffff8363; SI-NEXT:    s_or_b32 s8, s9, s88364; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x6c, v08365; SI-NEXT:    s_or_b32 s6, s6, s88366; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8367; SI-NEXT:    s_waitcnt expcnt(0)8368; SI-NEXT:    v_mov_b32_e32 v2, s68369; SI-NEXT:    s_and_b32 s6, s7, 0xff8370; SI-NEXT:    v_readlane_b32 s7, v23, 478371; SI-NEXT:    s_lshl_b32 s7, s7, 88372; SI-NEXT:    s_or_b32 s6, s6, s78373; SI-NEXT:    v_readlane_b32 s7, v23, 468374; SI-NEXT:    s_and_b32 s7, s7, 0xff8375; SI-NEXT:    v_readlane_b32 s8, v23, 458376; SI-NEXT:    s_lshl_b32 s7, s7, 168377; SI-NEXT:    s_lshl_b32 s8, s8, 248378; SI-NEXT:    s_and_b32 s6, s6, 0xffff8379; SI-NEXT:    s_or_b32 s7, s8, s78380; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x70, v08381; SI-NEXT:    s_or_b32 s6, s6, s78382; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8383; SI-NEXT:    s_waitcnt expcnt(0)8384; SI-NEXT:    v_mov_b32_e32 v2, s68385; SI-NEXT:    v_readlane_b32 s6, v23, 368386; SI-NEXT:    s_and_b32 s4, s4, 0xff8387; SI-NEXT:    v_readlane_b32 s7, v23, 378388; SI-NEXT:    s_lshl_b32 s6, s6, 88389; SI-NEXT:    s_or_b32 s4, s4, s68390; SI-NEXT:    v_readlane_b32 s6, v23, 388391; SI-NEXT:    v_readlane_b32 s7, v23, 398392; SI-NEXT:    s_and_b32 s6, s6, 0xff8393; SI-NEXT:    v_readlane_b32 s8, v23, 408394; SI-NEXT:    s_lshl_b32 s6, s6, 168395; SI-NEXT:    s_lshl_b32 s7, s8, 248396; SI-NEXT:    s_and_b32 s4, s4, 0xffff8397; SI-NEXT:    s_or_b32 s6, s7, s68398; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x74, v08399; SI-NEXT:    s_or_b32 s4, s4, s68400; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8401; SI-NEXT:    s_waitcnt expcnt(0)8402; SI-NEXT:    v_mov_b32_e32 v2, s48403; SI-NEXT:    s_and_b32 s4, s5, 0xff8404; SI-NEXT:    v_readlane_b32 s5, v23, 448405; SI-NEXT:    s_lshl_b32 s5, s5, 88406; SI-NEXT:    s_or_b32 s4, s4, s58407; SI-NEXT:    v_readlane_b32 s5, v23, 438408; SI-NEXT:    s_and_b32 s5, s5, 0xff8409; SI-NEXT:    v_readlane_b32 s6, v23, 428410; SI-NEXT:    s_lshl_b32 s5, s5, 168411; SI-NEXT:    s_lshl_b32 s6, s6, 248412; SI-NEXT:    s_and_b32 s4, s4, 0xffff8413; SI-NEXT:    s_or_b32 s5, s6, s58414; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x78, v08415; SI-NEXT:    s_or_b32 s4, s4, s58416; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen8417; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v08418; SI-NEXT:    v_mov_b32_e32 v1, s48419; SI-NEXT:    v_readlane_b32 s21, v23, 58420; SI-NEXT:    v_readlane_b32 s19, v23, 118421; SI-NEXT:    v_readlane_b32 s17, v23, 178422; SI-NEXT:    v_readlane_b32 s15, v23, 238423; SI-NEXT:    v_readlane_b32 s13, v23, 298424; SI-NEXT:    v_readlane_b32 s11, v23, 358425; SI-NEXT:    v_readlane_b32 s9, v23, 418426; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen8427; SI-NEXT:    v_readlane_b32 s99, v21, 358428; SI-NEXT:    v_readlane_b32 s98, v21, 348429; SI-NEXT:    v_readlane_b32 s97, v21, 338430; SI-NEXT:    v_readlane_b32 s96, v21, 328431; SI-NEXT:    v_readlane_b32 s87, v21, 318432; SI-NEXT:    v_readlane_b32 s86, v21, 308433; SI-NEXT:    v_readlane_b32 s85, v21, 298434; SI-NEXT:    v_readlane_b32 s84, v21, 288435; SI-NEXT:    v_readlane_b32 s83, v21, 278436; SI-NEXT:    v_readlane_b32 s82, v21, 268437; SI-NEXT:    v_readlane_b32 s81, v21, 258438; SI-NEXT:    v_readlane_b32 s80, v21, 248439; SI-NEXT:    v_readlane_b32 s71, v21, 238440; SI-NEXT:    v_readlane_b32 s70, v21, 228441; SI-NEXT:    v_readlane_b32 s69, v21, 218442; SI-NEXT:    v_readlane_b32 s68, v21, 208443; SI-NEXT:    v_readlane_b32 s67, v21, 198444; SI-NEXT:    v_readlane_b32 s66, v21, 188445; SI-NEXT:    v_readlane_b32 s65, v21, 178446; SI-NEXT:    v_readlane_b32 s64, v21, 168447; SI-NEXT:    v_readlane_b32 s55, v21, 158448; SI-NEXT:    v_readlane_b32 s54, v21, 148449; SI-NEXT:    v_readlane_b32 s53, v21, 138450; SI-NEXT:    v_readlane_b32 s52, v21, 128451; SI-NEXT:    v_readlane_b32 s51, v21, 118452; SI-NEXT:    v_readlane_b32 s50, v21, 108453; SI-NEXT:    v_readlane_b32 s49, v21, 98454; SI-NEXT:    v_readlane_b32 s48, v21, 88455; SI-NEXT:    v_readlane_b32 s39, v21, 78456; SI-NEXT:    v_readlane_b32 s38, v21, 68457; SI-NEXT:    v_readlane_b32 s37, v21, 58458; SI-NEXT:    v_readlane_b32 s36, v21, 48459; SI-NEXT:    v_readlane_b32 s35, v21, 38460; SI-NEXT:    v_readlane_b32 s34, v21, 28461; SI-NEXT:    v_readlane_b32 s31, v21, 18462; SI-NEXT:    v_readlane_b32 s30, v21, 08463; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -18464; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 ; 4-byte Folded Reload8465; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload8466; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload8467; SI-NEXT:    s_mov_b64 exec, s[4:5]8468; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)8469; SI-NEXT:    s_setpc_b64 s[30:31]8470; SI-NEXT:  .LBB13_4:8471; SI-NEXT:    ; implicit-def: $sgpr518472; SI-NEXT:    ; implicit-def: $sgpr508473; SI-NEXT:    ; implicit-def: $sgpr268474; SI-NEXT:    ; kill: killed $sgpr268475; SI-NEXT:    s_waitcnt expcnt(0)8476; SI-NEXT:    v_writelane_b32 v23, s50, 08477; SI-NEXT:    v_writelane_b32 v23, s51, 18478; SI-NEXT:    ; implicit-def: $sgpr508479; SI-NEXT:    ; implicit-def: $sgpr268480; SI-NEXT:    ; kill: killed $sgpr268481; SI-NEXT:    ; implicit-def: $sgpr268482; SI-NEXT:    v_writelane_b32 v23, s50, 28483; SI-NEXT:    v_writelane_b32 v23, s51, 38484; SI-NEXT:    ; implicit-def: $sgpr508485; SI-NEXT:    ; kill: killed $sgpr268486; SI-NEXT:    ; implicit-def: $sgpr268487; SI-NEXT:    ; kill: killed $sgpr268488; SI-NEXT:    ; implicit-def: $sgpr268489; SI-NEXT:    v_writelane_b32 v23, s50, 48490; SI-NEXT:    v_writelane_b32 v23, s51, 58491; SI-NEXT:    ; implicit-def: $sgpr508492; SI-NEXT:    ; kill: killed $sgpr268493; SI-NEXT:    ; implicit-def: $sgpr268494; SI-NEXT:    ; kill: killed $sgpr268495; SI-NEXT:    ; implicit-def: $sgpr268496; SI-NEXT:    v_writelane_b32 v23, s50, 68497; SI-NEXT:    v_writelane_b32 v23, s51, 78498; SI-NEXT:    ; implicit-def: $sgpr508499; SI-NEXT:    ; kill: killed $sgpr268500; SI-NEXT:    ; implicit-def: $sgpr268501; SI-NEXT:    ; kill: killed $sgpr268502; SI-NEXT:    ; implicit-def: $sgpr268503; SI-NEXT:    v_writelane_b32 v23, s50, 88504; SI-NEXT:    v_writelane_b32 v23, s51, 98505; SI-NEXT:    ; implicit-def: $sgpr508506; SI-NEXT:    ; kill: killed $sgpr268507; SI-NEXT:    ; implicit-def: $sgpr268508; SI-NEXT:    ; kill: killed $sgpr268509; SI-NEXT:    ; implicit-def: $sgpr268510; SI-NEXT:    v_writelane_b32 v23, s50, 108511; SI-NEXT:    v_writelane_b32 v23, s51, 118512; SI-NEXT:    ; implicit-def: $sgpr508513; SI-NEXT:    ; kill: killed $sgpr268514; SI-NEXT:    ; implicit-def: $sgpr268515; SI-NEXT:    ; kill: killed $sgpr268516; SI-NEXT:    ; implicit-def: $sgpr268517; SI-NEXT:    v_writelane_b32 v23, s50, 128518; SI-NEXT:    v_writelane_b32 v23, s51, 138519; SI-NEXT:    ; implicit-def: $sgpr508520; SI-NEXT:    ; kill: killed $sgpr268521; SI-NEXT:    ; implicit-def: $sgpr268522; SI-NEXT:    ; kill: killed $sgpr268523; SI-NEXT:    ; implicit-def: $sgpr268524; SI-NEXT:    v_writelane_b32 v23, s50, 148525; SI-NEXT:    v_writelane_b32 v23, s51, 158526; SI-NEXT:    ; implicit-def: $sgpr508527; SI-NEXT:    ; kill: killed $sgpr268528; SI-NEXT:    ; implicit-def: $sgpr268529; SI-NEXT:    ; kill: killed $sgpr268530; SI-NEXT:    ; implicit-def: $sgpr268531; SI-NEXT:    v_writelane_b32 v23, s50, 168532; SI-NEXT:    v_writelane_b32 v23, s51, 178533; SI-NEXT:    ; implicit-def: $sgpr508534; SI-NEXT:    ; kill: killed $sgpr268535; SI-NEXT:    ; implicit-def: $sgpr268536; SI-NEXT:    ; kill: killed $sgpr268537; SI-NEXT:    ; implicit-def: $sgpr268538; SI-NEXT:    v_writelane_b32 v23, s50, 188539; SI-NEXT:    v_writelane_b32 v23, s51, 198540; SI-NEXT:    ; implicit-def: $sgpr508541; SI-NEXT:    ; kill: killed $sgpr268542; SI-NEXT:    ; implicit-def: $sgpr268543; SI-NEXT:    ; kill: killed $sgpr268544; SI-NEXT:    ; implicit-def: $sgpr268545; SI-NEXT:    v_writelane_b32 v23, s50, 208546; SI-NEXT:    v_writelane_b32 v23, s51, 218547; SI-NEXT:    ; implicit-def: $sgpr508548; SI-NEXT:    ; kill: killed $sgpr268549; SI-NEXT:    ; implicit-def: $sgpr268550; SI-NEXT:    ; kill: killed $sgpr268551; SI-NEXT:    ; implicit-def: $sgpr268552; SI-NEXT:    v_writelane_b32 v23, s50, 228553; SI-NEXT:    v_writelane_b32 v23, s51, 238554; SI-NEXT:    ; implicit-def: $sgpr508555; SI-NEXT:    ; kill: killed $sgpr268556; SI-NEXT:    ; implicit-def: $sgpr268557; SI-NEXT:    ; kill: killed $sgpr268558; SI-NEXT:    ; implicit-def: $sgpr268559; SI-NEXT:    v_writelane_b32 v23, s50, 248560; SI-NEXT:    v_writelane_b32 v23, s51, 258561; SI-NEXT:    ; implicit-def: $sgpr508562; SI-NEXT:    ; kill: killed $sgpr268563; SI-NEXT:    ; implicit-def: $sgpr268564; SI-NEXT:    ; kill: killed $sgpr268565; SI-NEXT:    ; implicit-def: $sgpr268566; SI-NEXT:    v_writelane_b32 v23, s50, 268567; SI-NEXT:    v_writelane_b32 v23, s51, 278568; SI-NEXT:    ; implicit-def: $sgpr508569; SI-NEXT:    ; kill: killed $sgpr268570; SI-NEXT:    ; implicit-def: $sgpr268571; SI-NEXT:    ; kill: killed $sgpr268572; SI-NEXT:    ; implicit-def: $sgpr268573; SI-NEXT:    v_writelane_b32 v23, s50, 288574; SI-NEXT:    v_writelane_b32 v23, s51, 298575; SI-NEXT:    ; implicit-def: $sgpr508576; SI-NEXT:    ; kill: killed $sgpr268577; SI-NEXT:    ; implicit-def: $sgpr268578; SI-NEXT:    ; kill: killed $sgpr268579; SI-NEXT:    ; implicit-def: $sgpr268580; SI-NEXT:    v_writelane_b32 v23, s50, 308581; SI-NEXT:    v_writelane_b32 v23, s51, 318582; SI-NEXT:    ; implicit-def: $sgpr508583; SI-NEXT:    ; kill: killed $sgpr268584; SI-NEXT:    ; implicit-def: $sgpr268585; SI-NEXT:    ; kill: killed $sgpr268586; SI-NEXT:    ; implicit-def: $sgpr268587; SI-NEXT:    v_writelane_b32 v23, s50, 328588; SI-NEXT:    v_writelane_b32 v23, s51, 338589; SI-NEXT:    ; implicit-def: $sgpr508590; SI-NEXT:    ; kill: killed $sgpr268591; SI-NEXT:    ; implicit-def: $sgpr268592; SI-NEXT:    ; kill: killed $sgpr268593; SI-NEXT:    ; implicit-def: $sgpr268594; SI-NEXT:    v_writelane_b32 v23, s50, 348595; SI-NEXT:    v_writelane_b32 v23, s51, 358596; SI-NEXT:    ; implicit-def: $sgpr508597; SI-NEXT:    ; kill: killed $sgpr268598; SI-NEXT:    ; implicit-def: $sgpr268599; SI-NEXT:    ; kill: killed $sgpr268600; SI-NEXT:    ; implicit-def: $sgpr268601; SI-NEXT:    v_writelane_b32 v23, s50, 368602; SI-NEXT:    v_writelane_b32 v23, s51, 378603; SI-NEXT:    ; implicit-def: $sgpr508604; SI-NEXT:    ; kill: killed $sgpr268605; SI-NEXT:    ; implicit-def: $sgpr268606; SI-NEXT:    ; kill: killed $sgpr268607; SI-NEXT:    ; implicit-def: $sgpr268608; SI-NEXT:    v_writelane_b32 v23, s50, 388609; SI-NEXT:    v_writelane_b32 v23, s51, 398610; SI-NEXT:    ; implicit-def: $sgpr508611; SI-NEXT:    ; kill: killed $sgpr268612; SI-NEXT:    ; implicit-def: $sgpr268613; SI-NEXT:    ; implicit-def: $sgpr498614; SI-NEXT:    ; implicit-def: $sgpr558615; SI-NEXT:    ; implicit-def: $sgpr548616; SI-NEXT:    ; implicit-def: $sgpr538617; SI-NEXT:    ; implicit-def: $sgpr528618; SI-NEXT:    ; implicit-def: $sgpr488619; SI-NEXT:    ; kill: killed $sgpr268620; SI-NEXT:    ; implicit-def: $sgpr388621; SI-NEXT:    ; implicit-def: $sgpr368622; SI-NEXT:    ; implicit-def: $sgpr348623; SI-NEXT:    ; implicit-def: $sgpr308624; SI-NEXT:    ; implicit-def: $sgpr948625; SI-NEXT:    ; implicit-def: $sgpr928626; SI-NEXT:    ; implicit-def: $sgpr908627; SI-NEXT:    ; implicit-def: $sgpr888628; SI-NEXT:    ; implicit-def: $sgpr788629; SI-NEXT:    ; implicit-def: $sgpr768630; SI-NEXT:    ; implicit-def: $sgpr748631; SI-NEXT:    ; implicit-def: $sgpr728632; SI-NEXT:    ; implicit-def: $sgpr628633; SI-NEXT:    ; implicit-def: $sgpr608634; SI-NEXT:    ; implicit-def: $sgpr588635; SI-NEXT:    ; implicit-def: $sgpr288636; SI-NEXT:    ; implicit-def: $sgpr268637; SI-NEXT:    v_writelane_b32 v23, s50, 408638; SI-NEXT:    ; implicit-def: $sgpr988639; SI-NEXT:    ; implicit-def: $sgpr968640; SI-NEXT:    ; implicit-def: $sgpr868641; SI-NEXT:    ; implicit-def: $sgpr848642; SI-NEXT:    ; implicit-def: $sgpr828643; SI-NEXT:    ; implicit-def: $sgpr808644; SI-NEXT:    ; implicit-def: $sgpr708645; SI-NEXT:    ; implicit-def: $sgpr688646; SI-NEXT:    ; implicit-def: $sgpr668647; SI-NEXT:    ; implicit-def: $sgpr648648; SI-NEXT:    v_writelane_b32 v23, s51, 418649; SI-NEXT:    ; implicit-def: $sgpr508650; SI-NEXT:    s_branch .LBB13_28651;8652; VI-LABEL: bitcast_v32i32_to_v128i8_scalar:8653; VI:       ; %bb.0:8654; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)8655; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -18656; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 ; 4-byte Folded Spill8657; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill8658; VI-NEXT:    s_mov_b64 exec, s[4:5]8659; VI-NEXT:    v_writelane_b32 v21, s30, 08660; VI-NEXT:    v_writelane_b32 v21, s31, 18661; VI-NEXT:    v_writelane_b32 v21, s34, 28662; VI-NEXT:    v_writelane_b32 v21, s35, 38663; VI-NEXT:    v_writelane_b32 v21, s36, 48664; VI-NEXT:    v_writelane_b32 v21, s37, 58665; VI-NEXT:    v_writelane_b32 v21, s38, 68666; VI-NEXT:    v_writelane_b32 v21, s39, 78667; VI-NEXT:    v_writelane_b32 v21, s48, 88668; VI-NEXT:    v_writelane_b32 v21, s49, 98669; VI-NEXT:    v_writelane_b32 v21, s50, 108670; VI-NEXT:    v_writelane_b32 v21, s51, 118671; VI-NEXT:    v_writelane_b32 v21, s52, 128672; VI-NEXT:    v_writelane_b32 v21, s53, 138673; VI-NEXT:    v_writelane_b32 v21, s54, 148674; VI-NEXT:    v_writelane_b32 v21, s55, 158675; VI-NEXT:    v_writelane_b32 v21, s64, 168676; VI-NEXT:    v_mov_b32_e32 v20, s168677; VI-NEXT:    v_writelane_b32 v21, s65, 178678; VI-NEXT:    v_readfirstlane_b32 s56, v208679; VI-NEXT:    v_mov_b32_e32 v20, s178680; VI-NEXT:    v_writelane_b32 v21, s66, 188681; VI-NEXT:    v_readfirstlane_b32 s57, v208682; VI-NEXT:    v_mov_b32_e32 v20, s188683; VI-NEXT:    v_writelane_b32 v21, s67, 198684; VI-NEXT:    v_readfirstlane_b32 s46, v208685; VI-NEXT:    v_mov_b32_e32 v20, s198686; VI-NEXT:    v_writelane_b32 v21, s68, 208687; VI-NEXT:    v_readfirstlane_b32 s47, v208688; VI-NEXT:    v_mov_b32_e32 v20, s208689; VI-NEXT:    v_writelane_b32 v21, s69, 218690; VI-NEXT:    v_readfirstlane_b32 s44, v208691; VI-NEXT:    v_mov_b32_e32 v20, s218692; VI-NEXT:    v_writelane_b32 v21, s70, 228693; VI-NEXT:    v_readfirstlane_b32 s45, v208694; VI-NEXT:    v_mov_b32_e32 v20, s228695; VI-NEXT:    v_writelane_b32 v21, s71, 238696; VI-NEXT:    v_readfirstlane_b32 s42, v208697; VI-NEXT:    v_mov_b32_e32 v20, s238698; VI-NEXT:    v_writelane_b32 v21, s80, 248699; VI-NEXT:    v_readfirstlane_b32 s43, v208700; VI-NEXT:    v_mov_b32_e32 v20, s248701; VI-NEXT:    v_writelane_b32 v21, s81, 258702; VI-NEXT:    v_readfirstlane_b32 s40, v208703; VI-NEXT:    v_mov_b32_e32 v20, s258704; VI-NEXT:    v_writelane_b32 v21, s82, 268705; VI-NEXT:    v_readfirstlane_b32 s41, v208706; VI-NEXT:    v_mov_b32_e32 v20, s268707; VI-NEXT:    v_writelane_b32 v21, s83, 278708; VI-NEXT:    v_readfirstlane_b32 s24, v208709; VI-NEXT:    v_mov_b32_e32 v20, s278710; VI-NEXT:    v_writelane_b32 v21, s84, 288711; VI-NEXT:    v_readfirstlane_b32 s25, v208712; VI-NEXT:    v_mov_b32_e32 v20, s288713; VI-NEXT:    v_writelane_b32 v21, s85, 298714; VI-NEXT:    v_readfirstlane_b32 s22, v208715; VI-NEXT:    v_mov_b32_e32 v20, s298716; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v198717; VI-NEXT:    v_writelane_b32 v21, s86, 308718; VI-NEXT:    v_readfirstlane_b32 s23, v208719; VI-NEXT:    v_readfirstlane_b32 s20, v18720; VI-NEXT:    v_readfirstlane_b32 s21, v28721; VI-NEXT:    v_readfirstlane_b32 s18, v38722; VI-NEXT:    v_readfirstlane_b32 s19, v48723; VI-NEXT:    v_readfirstlane_b32 s16, v58724; VI-NEXT:    v_readfirstlane_b32 s17, v68725; VI-NEXT:    v_readfirstlane_b32 s14, v78726; VI-NEXT:    v_readfirstlane_b32 s15, v88727; VI-NEXT:    v_readfirstlane_b32 s12, v98728; VI-NEXT:    v_readfirstlane_b32 s13, v108729; VI-NEXT:    v_readfirstlane_b32 s10, v118730; VI-NEXT:    v_readfirstlane_b32 s11, v128731; VI-NEXT:    v_readfirstlane_b32 s8, v138732; VI-NEXT:    v_readfirstlane_b32 s9, v148733; VI-NEXT:    v_readfirstlane_b32 s6, v158734; VI-NEXT:    v_readfirstlane_b32 s7, v168735; VI-NEXT:    v_readfirstlane_b32 s4, v178736; VI-NEXT:    s_and_b64 s[26:27], vcc, exec8737; VI-NEXT:    v_readfirstlane_b32 s5, v188738; VI-NEXT:    v_writelane_b32 v21, s87, 318739; VI-NEXT:    ; implicit-def: $vgpr22 : SGPR spill to VGPR lane8740; VI-NEXT:    s_cbranch_scc0 .LBB13_48741; VI-NEXT:  ; %bb.1: ; %cmp.false8742; VI-NEXT:    s_lshr_b32 s26, s5, 248743; VI-NEXT:    v_writelane_b32 v22, s26, 88744; VI-NEXT:    s_lshr_b32 s26, s5, 168745; VI-NEXT:    v_writelane_b32 v22, s26, 98746; VI-NEXT:    s_lshr_b32 s26, s5, 88747; VI-NEXT:    v_writelane_b32 v22, s26, 108748; VI-NEXT:    s_lshr_b32 s26, s4, 168749; VI-NEXT:    v_writelane_b32 v22, s26, 118750; VI-NEXT:    s_lshr_b32 s26, s4, 88751; VI-NEXT:    v_writelane_b32 v22, s26, 128752; VI-NEXT:    s_lshr_b32 s26, s7, 248753; VI-NEXT:    v_writelane_b32 v22, s26, 138754; VI-NEXT:    s_lshr_b32 s26, s7, 168755; VI-NEXT:    v_writelane_b32 v22, s26, 148756; VI-NEXT:    s_lshr_b32 s26, s7, 88757; VI-NEXT:    v_writelane_b32 v22, s26, 158758; VI-NEXT:    s_lshr_b32 s26, s6, 168759; VI-NEXT:    v_writelane_b32 v22, s26, 168760; VI-NEXT:    s_lshr_b32 s26, s6, 88761; VI-NEXT:    v_writelane_b32 v22, s26, 178762; VI-NEXT:    s_lshr_b32 s26, s9, 248763; VI-NEXT:    v_writelane_b32 v22, s26, 188764; VI-NEXT:    s_lshr_b32 s26, s9, 168765; VI-NEXT:    v_writelane_b32 v22, s26, 198766; VI-NEXT:    s_lshr_b32 s26, s9, 88767; VI-NEXT:    v_writelane_b32 v22, s26, 208768; VI-NEXT:    s_lshr_b32 s26, s8, 168769; VI-NEXT:    v_writelane_b32 v22, s26, 218770; VI-NEXT:    s_lshr_b32 s26, s8, 88771; VI-NEXT:    v_writelane_b32 v22, s26, 228772; VI-NEXT:    s_lshr_b32 s26, s11, 248773; VI-NEXT:    v_writelane_b32 v22, s26, 238774; VI-NEXT:    s_lshr_b32 s26, s11, 168775; VI-NEXT:    v_writelane_b32 v22, s26, 248776; VI-NEXT:    s_lshr_b32 s26, s11, 88777; VI-NEXT:    v_writelane_b32 v22, s26, 258778; VI-NEXT:    s_lshr_b32 s26, s10, 168779; VI-NEXT:    v_writelane_b32 v22, s26, 268780; VI-NEXT:    s_lshr_b32 s26, s10, 88781; VI-NEXT:    v_writelane_b32 v22, s26, 278782; VI-NEXT:    s_lshr_b32 s26, s13, 248783; VI-NEXT:    v_writelane_b32 v22, s26, 288784; VI-NEXT:    s_lshr_b32 s26, s13, 168785; VI-NEXT:    v_writelane_b32 v22, s26, 298786; VI-NEXT:    s_lshr_b32 s26, s13, 88787; VI-NEXT:    v_writelane_b32 v22, s26, 308788; VI-NEXT:    s_lshr_b32 s26, s12, 168789; VI-NEXT:    v_writelane_b32 v22, s26, 318790; VI-NEXT:    s_lshr_b32 s26, s12, 88791; VI-NEXT:    v_writelane_b32 v22, s26, 328792; VI-NEXT:    s_lshr_b32 s26, s15, 248793; VI-NEXT:    v_writelane_b32 v22, s26, 338794; VI-NEXT:    s_lshr_b32 s26, s15, 168795; VI-NEXT:    v_writelane_b32 v22, s26, 348796; VI-NEXT:    s_lshr_b32 s26, s15, 88797; VI-NEXT:    v_writelane_b32 v22, s26, 358798; VI-NEXT:    s_lshr_b32 s26, s14, 168799; VI-NEXT:    v_writelane_b32 v22, s26, 368800; VI-NEXT:    s_lshr_b32 s26, s14, 88801; VI-NEXT:    v_writelane_b32 v22, s26, 378802; VI-NEXT:    s_lshr_b32 s26, s17, 248803; VI-NEXT:    v_writelane_b32 v22, s26, 388804; VI-NEXT:    s_lshr_b32 s26, s17, 168805; VI-NEXT:    v_writelane_b32 v22, s26, 398806; VI-NEXT:    s_lshr_b32 s26, s17, 88807; VI-NEXT:    v_writelane_b32 v22, s26, 408808; VI-NEXT:    s_lshr_b32 s26, s16, 168809; VI-NEXT:    v_writelane_b32 v22, s26, 418810; VI-NEXT:    s_lshr_b32 s26, s16, 88811; VI-NEXT:    v_writelane_b32 v22, s26, 428812; VI-NEXT:    s_lshr_b32 s26, s19, 248813; VI-NEXT:    v_writelane_b32 v22, s26, 438814; VI-NEXT:    s_lshr_b32 s26, s19, 168815; VI-NEXT:    v_writelane_b32 v22, s26, 448816; VI-NEXT:    s_lshr_b32 s26, s19, 88817; VI-NEXT:    v_writelane_b32 v22, s26, 458818; VI-NEXT:    s_lshr_b32 s26, s18, 168819; VI-NEXT:    v_writelane_b32 v22, s26, 468820; VI-NEXT:    s_lshr_b32 s26, s18, 88821; VI-NEXT:    v_writelane_b32 v22, s26, 478822; VI-NEXT:    s_lshr_b32 s26, s21, 248823; VI-NEXT:    v_writelane_b32 v22, s26, 488824; VI-NEXT:    s_lshr_b32 s26, s21, 168825; VI-NEXT:    v_writelane_b32 v22, s26, 498826; VI-NEXT:    s_lshr_b32 s26, s21, 88827; VI-NEXT:    v_writelane_b32 v22, s26, 508828; VI-NEXT:    s_lshr_b32 s26, s20, 168829; VI-NEXT:    v_writelane_b32 v22, s26, 518830; VI-NEXT:    s_lshr_b32 s26, s20, 88831; VI-NEXT:    v_writelane_b32 v22, s26, 528832; VI-NEXT:    s_lshr_b32 s26, s23, 248833; VI-NEXT:    v_writelane_b32 v22, s26, 538834; VI-NEXT:    s_lshr_b32 s26, s23, 168835; VI-NEXT:    v_writelane_b32 v22, s26, 548836; VI-NEXT:    s_lshr_b32 s26, s23, 88837; VI-NEXT:    v_writelane_b32 v22, s26, 558838; VI-NEXT:    s_lshr_b32 s26, s22, 168839; VI-NEXT:    v_writelane_b32 v22, s26, 568840; VI-NEXT:    s_lshr_b32 s26, s22, 88841; VI-NEXT:    v_writelane_b32 v22, s26, 578842; VI-NEXT:    s_lshr_b32 s26, s25, 248843; VI-NEXT:    v_writelane_b32 v22, s26, 588844; VI-NEXT:    s_lshr_b32 s26, s25, 168845; VI-NEXT:    v_writelane_b32 v22, s26, 598846; VI-NEXT:    s_lshr_b64 s[60:61], s[4:5], 248847; VI-NEXT:    v_writelane_b32 v22, s60, 68848; VI-NEXT:    v_writelane_b32 v22, s61, 78849; VI-NEXT:    s_lshr_b64 s[60:61], s[6:7], 248850; VI-NEXT:    v_writelane_b32 v22, s60, 48851; VI-NEXT:    v_writelane_b32 v22, s61, 58852; VI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 248853; VI-NEXT:    v_writelane_b32 v22, s60, 28854; VI-NEXT:    v_writelane_b32 v22, s61, 38855; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 248856; VI-NEXT:    v_writelane_b32 v22, s60, 08857; VI-NEXT:    s_lshr_b32 s66, s25, 88858; VI-NEXT:    s_lshr_b32 s67, s24, 168859; VI-NEXT:    s_lshr_b32 s68, s24, 88860; VI-NEXT:    s_lshr_b32 s69, s41, 248861; VI-NEXT:    s_lshr_b32 s70, s41, 168862; VI-NEXT:    s_lshr_b32 s71, s41, 88863; VI-NEXT:    s_lshr_b32 s80, s40, 168864; VI-NEXT:    s_lshr_b32 s81, s40, 88865; VI-NEXT:    s_lshr_b32 s82, s43, 248866; VI-NEXT:    s_lshr_b32 s83, s43, 168867; VI-NEXT:    s_lshr_b32 s84, s43, 88868; VI-NEXT:    s_lshr_b32 s85, s42, 168869; VI-NEXT:    s_lshr_b32 s86, s42, 88870; VI-NEXT:    s_lshr_b32 s87, s45, 248871; VI-NEXT:    s_lshr_b32 s50, s45, 168872; VI-NEXT:    s_lshr_b32 s26, s45, 88873; VI-NEXT:    s_lshr_b32 s27, s44, 168874; VI-NEXT:    s_lshr_b32 s28, s44, 88875; VI-NEXT:    s_lshr_b32 s29, s47, 248876; VI-NEXT:    s_lshr_b32 s51, s47, 168877; VI-NEXT:    s_lshr_b32 s52, s47, 88878; VI-NEXT:    s_lshr_b32 s53, s46, 168879; VI-NEXT:    s_lshr_b32 s54, s46, 88880; VI-NEXT:    s_lshr_b32 s58, s57, 248881; VI-NEXT:    s_lshr_b32 s59, s57, 168882; VI-NEXT:    s_lshr_b32 s55, s57, 88883; VI-NEXT:    s_lshr_b32 s64, s56, 168884; VI-NEXT:    s_lshr_b32 s65, s56, 88885; VI-NEXT:    v_writelane_b32 v22, s61, 18886; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 248887; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 248888; VI-NEXT:    s_lshr_b64 s[74:75], s[16:17], 248889; VI-NEXT:    s_lshr_b64 s[76:77], s[18:19], 248890; VI-NEXT:    s_lshr_b64 s[78:79], s[20:21], 248891; VI-NEXT:    s_lshr_b64 s[88:89], s[22:23], 248892; VI-NEXT:    s_lshr_b64 s[90:91], s[24:25], 248893; VI-NEXT:    s_lshr_b64 s[30:31], s[40:41], 248894; VI-NEXT:    s_lshr_b64 s[34:35], s[42:43], 248895; VI-NEXT:    s_lshr_b64 s[36:37], s[44:45], 248896; VI-NEXT:    s_lshr_b64 s[38:39], s[46:47], 248897; VI-NEXT:    s_lshr_b64 s[48:49], s[56:57], 248898; VI-NEXT:    s_cbranch_execnz .LBB13_38899; VI-NEXT:  .LBB13_2: ; %cmp.true8900; VI-NEXT:    s_add_i32 s5, s5, 38901; VI-NEXT:    s_lshr_b32 s26, s5, 248902; VI-NEXT:    v_writelane_b32 v22, s26, 88903; VI-NEXT:    s_lshr_b32 s26, s5, 168904; VI-NEXT:    s_add_i32 s4, s4, 38905; VI-NEXT:    v_writelane_b32 v22, s26, 98906; VI-NEXT:    s_lshr_b32 s26, s5, 88907; VI-NEXT:    v_writelane_b32 v22, s26, 108908; VI-NEXT:    s_lshr_b32 s26, s4, 168909; VI-NEXT:    s_add_i32 s7, s7, 38910; VI-NEXT:    v_writelane_b32 v22, s26, 118911; VI-NEXT:    s_lshr_b32 s26, s4, 88912; VI-NEXT:    v_writelane_b32 v22, s26, 128913; VI-NEXT:    s_lshr_b32 s26, s7, 248914; VI-NEXT:    v_writelane_b32 v22, s26, 138915; VI-NEXT:    s_lshr_b32 s26, s7, 168916; VI-NEXT:    s_add_i32 s6, s6, 38917; VI-NEXT:    v_writelane_b32 v22, s26, 148918; VI-NEXT:    s_lshr_b32 s26, s7, 88919; VI-NEXT:    v_writelane_b32 v22, s26, 158920; VI-NEXT:    s_lshr_b32 s26, s6, 168921; VI-NEXT:    s_add_i32 s9, s9, 38922; VI-NEXT:    v_writelane_b32 v22, s26, 168923; VI-NEXT:    s_lshr_b32 s26, s6, 88924; VI-NEXT:    v_writelane_b32 v22, s26, 178925; VI-NEXT:    s_lshr_b32 s26, s9, 248926; VI-NEXT:    v_writelane_b32 v22, s26, 188927; VI-NEXT:    s_lshr_b32 s26, s9, 168928; VI-NEXT:    s_add_i32 s8, s8, 38929; VI-NEXT:    v_writelane_b32 v22, s26, 198930; VI-NEXT:    s_lshr_b32 s26, s9, 88931; VI-NEXT:    v_writelane_b32 v22, s26, 208932; VI-NEXT:    s_lshr_b32 s26, s8, 168933; VI-NEXT:    s_add_i32 s11, s11, 38934; VI-NEXT:    v_writelane_b32 v22, s26, 218935; VI-NEXT:    s_lshr_b32 s26, s8, 88936; VI-NEXT:    v_writelane_b32 v22, s26, 228937; VI-NEXT:    s_lshr_b32 s26, s11, 248938; VI-NEXT:    v_writelane_b32 v22, s26, 238939; VI-NEXT:    s_lshr_b32 s26, s11, 168940; VI-NEXT:    s_add_i32 s10, s10, 38941; VI-NEXT:    v_writelane_b32 v22, s26, 248942; VI-NEXT:    s_lshr_b32 s26, s11, 88943; VI-NEXT:    v_writelane_b32 v22, s26, 258944; VI-NEXT:    s_lshr_b32 s26, s10, 168945; VI-NEXT:    s_add_i32 s13, s13, 38946; VI-NEXT:    v_writelane_b32 v22, s26, 268947; VI-NEXT:    s_lshr_b32 s26, s10, 88948; VI-NEXT:    v_writelane_b32 v22, s26, 278949; VI-NEXT:    s_lshr_b32 s26, s13, 248950; VI-NEXT:    v_writelane_b32 v22, s26, 288951; VI-NEXT:    s_lshr_b32 s26, s13, 168952; VI-NEXT:    s_add_i32 s12, s12, 38953; VI-NEXT:    v_writelane_b32 v22, s26, 298954; VI-NEXT:    s_lshr_b32 s26, s13, 88955; VI-NEXT:    v_writelane_b32 v22, s26, 308956; VI-NEXT:    s_lshr_b32 s26, s12, 168957; VI-NEXT:    s_add_i32 s15, s15, 38958; VI-NEXT:    v_writelane_b32 v22, s26, 318959; VI-NEXT:    s_lshr_b32 s26, s12, 88960; VI-NEXT:    v_writelane_b32 v22, s26, 328961; VI-NEXT:    s_lshr_b32 s26, s15, 248962; VI-NEXT:    v_writelane_b32 v22, s26, 338963; VI-NEXT:    s_lshr_b32 s26, s15, 168964; VI-NEXT:    s_add_i32 s14, s14, 38965; VI-NEXT:    v_writelane_b32 v22, s26, 348966; VI-NEXT:    s_lshr_b32 s26, s15, 88967; VI-NEXT:    v_writelane_b32 v22, s26, 358968; VI-NEXT:    s_lshr_b32 s26, s14, 168969; VI-NEXT:    s_add_i32 s17, s17, 38970; VI-NEXT:    v_writelane_b32 v22, s26, 368971; VI-NEXT:    s_lshr_b32 s26, s14, 88972; VI-NEXT:    v_writelane_b32 v22, s26, 378973; VI-NEXT:    s_lshr_b32 s26, s17, 248974; VI-NEXT:    v_writelane_b32 v22, s26, 388975; VI-NEXT:    s_lshr_b32 s26, s17, 168976; VI-NEXT:    s_add_i32 s16, s16, 38977; VI-NEXT:    v_writelane_b32 v22, s26, 398978; VI-NEXT:    s_lshr_b32 s26, s17, 88979; VI-NEXT:    v_writelane_b32 v22, s26, 408980; VI-NEXT:    s_lshr_b32 s26, s16, 168981; VI-NEXT:    s_add_i32 s19, s19, 38982; VI-NEXT:    v_writelane_b32 v22, s26, 418983; VI-NEXT:    s_lshr_b32 s26, s16, 88984; VI-NEXT:    v_writelane_b32 v22, s26, 428985; VI-NEXT:    s_lshr_b32 s26, s19, 248986; VI-NEXT:    v_writelane_b32 v22, s26, 438987; VI-NEXT:    s_lshr_b32 s26, s19, 168988; VI-NEXT:    s_add_i32 s18, s18, 38989; VI-NEXT:    v_writelane_b32 v22, s26, 448990; VI-NEXT:    s_lshr_b32 s26, s19, 88991; VI-NEXT:    v_writelane_b32 v22, s26, 458992; VI-NEXT:    s_lshr_b32 s26, s18, 168993; VI-NEXT:    s_add_i32 s21, s21, 38994; VI-NEXT:    v_writelane_b32 v22, s26, 468995; VI-NEXT:    s_lshr_b32 s26, s18, 88996; VI-NEXT:    v_writelane_b32 v22, s26, 478997; VI-NEXT:    s_lshr_b32 s26, s21, 248998; VI-NEXT:    v_writelane_b32 v22, s26, 488999; VI-NEXT:    s_lshr_b32 s26, s21, 169000; VI-NEXT:    s_add_i32 s20, s20, 39001; VI-NEXT:    v_writelane_b32 v22, s26, 499002; VI-NEXT:    s_lshr_b32 s26, s21, 89003; VI-NEXT:    v_writelane_b32 v22, s26, 509004; VI-NEXT:    s_lshr_b32 s26, s20, 169005; VI-NEXT:    s_add_i32 s23, s23, 39006; VI-NEXT:    v_writelane_b32 v22, s26, 519007; VI-NEXT:    s_lshr_b32 s26, s20, 89008; VI-NEXT:    v_writelane_b32 v22, s26, 529009; VI-NEXT:    s_lshr_b32 s26, s23, 249010; VI-NEXT:    v_writelane_b32 v22, s26, 539011; VI-NEXT:    s_lshr_b32 s26, s23, 169012; VI-NEXT:    s_add_i32 s22, s22, 39013; VI-NEXT:    v_writelane_b32 v22, s26, 549014; VI-NEXT:    s_lshr_b32 s26, s23, 89015; VI-NEXT:    v_writelane_b32 v22, s26, 559016; VI-NEXT:    s_lshr_b32 s26, s22, 169017; VI-NEXT:    s_add_i32 s25, s25, 39018; VI-NEXT:    v_writelane_b32 v22, s26, 569019; VI-NEXT:    s_lshr_b32 s26, s22, 89020; VI-NEXT:    v_writelane_b32 v22, s26, 579021; VI-NEXT:    s_lshr_b32 s26, s25, 249022; VI-NEXT:    v_writelane_b32 v22, s26, 589023; VI-NEXT:    s_lshr_b32 s26, s25, 169024; VI-NEXT:    v_writelane_b32 v22, s26, 599025; VI-NEXT:    s_lshr_b64 s[60:61], s[4:5], 249026; VI-NEXT:    v_writelane_b32 v22, s60, 69027; VI-NEXT:    v_writelane_b32 v22, s61, 79028; VI-NEXT:    s_lshr_b64 s[60:61], s[6:7], 249029; VI-NEXT:    v_writelane_b32 v22, s60, 49030; VI-NEXT:    v_writelane_b32 v22, s61, 59031; VI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 249032; VI-NEXT:    v_writelane_b32 v22, s60, 29033; VI-NEXT:    s_add_i32 s57, s57, 39034; VI-NEXT:    s_add_i32 s56, s56, 39035; VI-NEXT:    s_add_i32 s47, s47, 39036; VI-NEXT:    s_add_i32 s46, s46, 39037; VI-NEXT:    s_add_i32 s45, s45, 39038; VI-NEXT:    s_add_i32 s44, s44, 39039; VI-NEXT:    s_add_i32 s43, s43, 39040; VI-NEXT:    s_add_i32 s42, s42, 39041; VI-NEXT:    s_add_i32 s41, s41, 39042; VI-NEXT:    s_add_i32 s40, s40, 39043; VI-NEXT:    s_add_i32 s24, s24, 39044; VI-NEXT:    v_writelane_b32 v22, s61, 39045; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 249046; VI-NEXT:    s_lshr_b32 s66, s25, 89047; VI-NEXT:    s_lshr_b32 s67, s24, 169048; VI-NEXT:    s_lshr_b32 s68, s24, 89049; VI-NEXT:    s_lshr_b32 s69, s41, 249050; VI-NEXT:    s_lshr_b32 s70, s41, 169051; VI-NEXT:    s_lshr_b32 s71, s41, 89052; VI-NEXT:    s_lshr_b32 s80, s40, 169053; VI-NEXT:    s_lshr_b32 s81, s40, 89054; VI-NEXT:    s_lshr_b32 s82, s43, 249055; VI-NEXT:    s_lshr_b32 s83, s43, 169056; VI-NEXT:    s_lshr_b32 s84, s43, 89057; VI-NEXT:    s_lshr_b32 s85, s42, 169058; VI-NEXT:    s_lshr_b32 s86, s42, 89059; VI-NEXT:    s_lshr_b32 s87, s45, 249060; VI-NEXT:    s_lshr_b32 s50, s45, 169061; VI-NEXT:    s_lshr_b32 s26, s45, 89062; VI-NEXT:    s_lshr_b32 s27, s44, 169063; VI-NEXT:    s_lshr_b32 s28, s44, 89064; VI-NEXT:    s_lshr_b32 s29, s47, 249065; VI-NEXT:    s_lshr_b32 s51, s47, 169066; VI-NEXT:    s_lshr_b32 s52, s47, 89067; VI-NEXT:    s_lshr_b32 s53, s46, 169068; VI-NEXT:    s_lshr_b32 s54, s46, 89069; VI-NEXT:    s_lshr_b32 s58, s57, 249070; VI-NEXT:    s_lshr_b32 s59, s57, 169071; VI-NEXT:    s_lshr_b32 s55, s57, 89072; VI-NEXT:    s_lshr_b32 s64, s56, 169073; VI-NEXT:    s_lshr_b32 s65, s56, 89074; VI-NEXT:    v_writelane_b32 v22, s60, 09075; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 249076; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 249077; VI-NEXT:    s_lshr_b64 s[74:75], s[16:17], 249078; VI-NEXT:    s_lshr_b64 s[76:77], s[18:19], 249079; VI-NEXT:    s_lshr_b64 s[78:79], s[20:21], 249080; VI-NEXT:    s_lshr_b64 s[88:89], s[22:23], 249081; VI-NEXT:    s_lshr_b64 s[90:91], s[24:25], 249082; VI-NEXT:    s_lshr_b64 s[30:31], s[40:41], 249083; VI-NEXT:    s_lshr_b64 s[34:35], s[42:43], 249084; VI-NEXT:    s_lshr_b64 s[36:37], s[44:45], 249085; VI-NEXT:    s_lshr_b64 s[38:39], s[46:47], 249086; VI-NEXT:    s_lshr_b64 s[48:49], s[56:57], 249087; VI-NEXT:    v_writelane_b32 v22, s61, 19088; VI-NEXT:  .LBB13_3: ; %end9089; VI-NEXT:    s_lshl_b32 s61, s65, 89090; VI-NEXT:    s_and_b32 s56, s56, 0xff9091; VI-NEXT:    s_or_b32 s56, s56, s619092; VI-NEXT:    s_lshl_b32 s61, s48, 89093; VI-NEXT:    s_and_b32 s63, s64, 0xff9094; VI-NEXT:    s_or_b32 s61, s63, s619095; VI-NEXT:    s_and_b32 s56, s56, 0xffff9096; VI-NEXT:    s_lshl_b32 s61, s61, 169097; VI-NEXT:    s_or_b32 s56, s56, s619098; VI-NEXT:    v_mov_b32_e32 v1, s569099; VI-NEXT:    s_and_b32 s56, s57, 0xff9100; VI-NEXT:    s_lshl_b32 s57, s55, 89101; VI-NEXT:    s_or_b32 s56, s56, s579102; VI-NEXT:    s_and_b32 s57, s59, 0xff9103; VI-NEXT:    s_lshl_b32 s58, s58, 89104; VI-NEXT:    s_or_b32 s57, s57, s589105; VI-NEXT:    s_and_b32 s56, s56, 0xffff9106; VI-NEXT:    s_lshl_b32 s57, s57, 169107; VI-NEXT:    s_or_b32 s56, s56, s579108; VI-NEXT:    v_mov_b32_e32 v2, s569109; VI-NEXT:    s_lshl_b32 s56, s54, 89110; VI-NEXT:    s_and_b32 s46, s46, 0xff9111; VI-NEXT:    s_or_b32 s46, s46, s569112; VI-NEXT:    s_lshl_b32 s56, s38, 89113; VI-NEXT:    s_and_b32 s57, s53, 0xff9114; VI-NEXT:    s_or_b32 s56, s57, s569115; VI-NEXT:    s_and_b32 s46, s46, 0xffff9116; VI-NEXT:    s_lshl_b32 s56, s56, 169117; VI-NEXT:    s_or_b32 s46, s46, s569118; VI-NEXT:    v_mov_b32_e32 v3, s469119; VI-NEXT:    s_and_b32 s46, s47, 0xff9120; VI-NEXT:    s_lshl_b32 s47, s52, 89121; VI-NEXT:    s_or_b32 s46, s46, s479122; VI-NEXT:    s_and_b32 s47, s51, 0xff9123; VI-NEXT:    s_lshl_b32 s29, s29, 89124; VI-NEXT:    s_or_b32 s29, s47, s299125; VI-NEXT:    s_and_b32 s46, s46, 0xffff9126; VI-NEXT:    s_lshl_b32 s29, s29, 169127; VI-NEXT:    s_or_b32 s29, s46, s299128; VI-NEXT:    v_mov_b32_e32 v4, s299129; VI-NEXT:    s_lshl_b32 s28, s28, 89130; VI-NEXT:    s_and_b32 s29, s44, 0xff9131; VI-NEXT:    s_or_b32 s28, s29, s289132; VI-NEXT:    s_lshl_b32 s29, s36, 89133; VI-NEXT:    s_and_b32 s27, s27, 0xff9134; VI-NEXT:    s_or_b32 s27, s27, s299135; VI-NEXT:    s_and_b32 s28, s28, 0xffff9136; VI-NEXT:    s_lshl_b32 s27, s27, 169137; VI-NEXT:    s_or_b32 s27, s28, s279138; VI-NEXT:    v_mov_b32_e32 v5, s279139; VI-NEXT:    s_and_b32 s27, s45, 0xff9140; VI-NEXT:    s_lshl_b32 s26, s26, 89141; VI-NEXT:    s_or_b32 s26, s27, s269142; VI-NEXT:    s_and_b32 s27, s50, 0xff9143; VI-NEXT:    s_lshl_b32 s28, s87, 89144; VI-NEXT:    s_or_b32 s27, s27, s289145; VI-NEXT:    s_and_b32 s26, s26, 0xffff9146; VI-NEXT:    s_lshl_b32 s27, s27, 169147; VI-NEXT:    s_or_b32 s26, s26, s279148; VI-NEXT:    v_mov_b32_e32 v6, s269149; VI-NEXT:    s_lshl_b32 s26, s86, 89150; VI-NEXT:    s_and_b32 s27, s42, 0xff9151; VI-NEXT:    s_or_b32 s26, s27, s269152; VI-NEXT:    s_lshl_b32 s27, s34, 89153; VI-NEXT:    s_and_b32 s28, s85, 0xff9154; VI-NEXT:    s_or_b32 s27, s28, s279155; VI-NEXT:    s_and_b32 s26, s26, 0xffff9156; VI-NEXT:    s_lshl_b32 s27, s27, 169157; VI-NEXT:    s_or_b32 s26, s26, s279158; VI-NEXT:    v_mov_b32_e32 v7, s269159; VI-NEXT:    s_and_b32 s26, s43, 0xff9160; VI-NEXT:    s_lshl_b32 s27, s84, 89161; VI-NEXT:    s_or_b32 s26, s26, s279162; VI-NEXT:    s_and_b32 s27, s83, 0xff9163; VI-NEXT:    s_lshl_b32 s28, s82, 89164; VI-NEXT:    s_or_b32 s27, s27, s289165; VI-NEXT:    s_and_b32 s26, s26, 0xffff9166; VI-NEXT:    s_lshl_b32 s27, s27, 169167; VI-NEXT:    s_or_b32 s26, s26, s279168; VI-NEXT:    v_mov_b32_e32 v8, s269169; VI-NEXT:    s_lshl_b32 s26, s81, 89170; VI-NEXT:    s_and_b32 s27, s40, 0xff9171; VI-NEXT:    s_or_b32 s26, s27, s269172; VI-NEXT:    s_lshl_b32 s27, s30, 89173; VI-NEXT:    s_and_b32 s28, s80, 0xff9174; VI-NEXT:    s_or_b32 s27, s28, s279175; VI-NEXT:    s_and_b32 s26, s26, 0xffff9176; VI-NEXT:    s_lshl_b32 s27, s27, 169177; VI-NEXT:    s_or_b32 s26, s26, s279178; VI-NEXT:    v_mov_b32_e32 v9, s269179; VI-NEXT:    s_and_b32 s26, s41, 0xff9180; VI-NEXT:    s_lshl_b32 s27, s71, 89181; VI-NEXT:    s_or_b32 s26, s26, s279182; VI-NEXT:    s_and_b32 s27, s70, 0xff9183; VI-NEXT:    s_lshl_b32 s28, s69, 89184; VI-NEXT:    s_or_b32 s27, s27, s289185; VI-NEXT:    s_and_b32 s26, s26, 0xffff9186; VI-NEXT:    s_lshl_b32 s27, s27, 169187; VI-NEXT:    s_or_b32 s26, s26, s279188; VI-NEXT:    v_mov_b32_e32 v10, s269189; VI-NEXT:    s_lshl_b32 s26, s68, 89190; VI-NEXT:    s_and_b32 s24, s24, 0xff9191; VI-NEXT:    s_or_b32 s24, s24, s269192; VI-NEXT:    s_lshl_b32 s26, s90, 89193; VI-NEXT:    s_and_b32 s27, s67, 0xff9194; VI-NEXT:    s_or_b32 s26, s27, s269195; VI-NEXT:    s_and_b32 s24, s24, 0xffff9196; VI-NEXT:    s_lshl_b32 s26, s26, 169197; VI-NEXT:    s_or_b32 s24, s24, s269198; VI-NEXT:    v_mov_b32_e32 v11, s249199; VI-NEXT:    s_and_b32 s24, s25, 0xff9200; VI-NEXT:    s_lshl_b32 s25, s66, 89201; VI-NEXT:    s_or_b32 s24, s24, s259202; VI-NEXT:    v_readlane_b32 s25, v22, 599203; VI-NEXT:    v_readlane_b32 s26, v22, 589204; VI-NEXT:    s_and_b32 s25, s25, 0xff9205; VI-NEXT:    s_lshl_b32 s26, s26, 89206; VI-NEXT:    s_or_b32 s25, s25, s269207; VI-NEXT:    s_and_b32 s24, s24, 0xffff9208; VI-NEXT:    s_lshl_b32 s25, s25, 169209; VI-NEXT:    s_or_b32 s24, s24, s259210; VI-NEXT:    v_mov_b32_e32 v12, s249211; VI-NEXT:    v_readlane_b32 s24, v22, 579212; VI-NEXT:    s_lshl_b32 s24, s24, 89213; VI-NEXT:    s_and_b32 s22, s22, 0xff9214; VI-NEXT:    v_readlane_b32 s25, v22, 569215; VI-NEXT:    s_or_b32 s22, s22, s249216; VI-NEXT:    s_lshl_b32 s24, s88, 89217; VI-NEXT:    s_and_b32 s25, s25, 0xff9218; VI-NEXT:    s_or_b32 s24, s25, s249219; VI-NEXT:    s_and_b32 s22, s22, 0xffff9220; VI-NEXT:    s_lshl_b32 s24, s24, 169221; VI-NEXT:    s_or_b32 s22, s22, s249222; VI-NEXT:    v_mov_b32_e32 v13, s229223; VI-NEXT:    s_and_b32 s22, s23, 0xff9224; VI-NEXT:    v_readlane_b32 s23, v22, 559225; VI-NEXT:    s_lshl_b32 s23, s23, 89226; VI-NEXT:    s_or_b32 s22, s22, s239227; VI-NEXT:    v_readlane_b32 s23, v22, 549228; VI-NEXT:    v_readlane_b32 s24, v22, 539229; VI-NEXT:    s_and_b32 s23, s23, 0xff9230; VI-NEXT:    s_lshl_b32 s24, s24, 89231; VI-NEXT:    s_or_b32 s23, s23, s249232; VI-NEXT:    s_and_b32 s22, s22, 0xffff9233; VI-NEXT:    s_lshl_b32 s23, s23, 169234; VI-NEXT:    s_or_b32 s22, s22, s239235; VI-NEXT:    v_mov_b32_e32 v14, s229236; VI-NEXT:    v_readlane_b32 s22, v22, 529237; VI-NEXT:    s_lshl_b32 s22, s22, 89238; VI-NEXT:    s_and_b32 s20, s20, 0xff9239; VI-NEXT:    v_readlane_b32 s23, v22, 519240; VI-NEXT:    s_or_b32 s20, s20, s229241; VI-NEXT:    s_lshl_b32 s22, s78, 89242; VI-NEXT:    s_and_b32 s23, s23, 0xff9243; VI-NEXT:    s_or_b32 s22, s23, s229244; VI-NEXT:    s_and_b32 s20, s20, 0xffff9245; VI-NEXT:    s_lshl_b32 s22, s22, 169246; VI-NEXT:    s_or_b32 s20, s20, s229247; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen9248; VI-NEXT:    v_add_u32_e32 v1, vcc, 4, v09249; VI-NEXT:    v_mov_b32_e32 v15, s209250; VI-NEXT:    s_and_b32 s20, s21, 0xff9251; VI-NEXT:    v_readlane_b32 s21, v22, 509252; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9253; VI-NEXT:    v_add_u32_e32 v1, vcc, 8, v09254; VI-NEXT:    s_lshl_b32 s21, s21, 89255; VI-NEXT:    buffer_store_dword v3, v1, s[0:3], 0 offen9256; VI-NEXT:    v_add_u32_e32 v1, vcc, 12, v09257; VI-NEXT:    s_or_b32 s20, s20, s219258; VI-NEXT:    buffer_store_dword v4, v1, s[0:3], 0 offen9259; VI-NEXT:    v_add_u32_e32 v1, vcc, 16, v09260; VI-NEXT:    v_readlane_b32 s21, v22, 499261; VI-NEXT:    v_readlane_b32 s22, v22, 489262; VI-NEXT:    buffer_store_dword v5, v1, s[0:3], 0 offen9263; VI-NEXT:    v_add_u32_e32 v1, vcc, 20, v09264; VI-NEXT:    s_and_b32 s21, s21, 0xff9265; VI-NEXT:    s_lshl_b32 s22, s22, 89266; VI-NEXT:    buffer_store_dword v6, v1, s[0:3], 0 offen9267; VI-NEXT:    v_add_u32_e32 v1, vcc, 24, v09268; VI-NEXT:    s_or_b32 s21, s21, s229269; VI-NEXT:    buffer_store_dword v7, v1, s[0:3], 0 offen9270; VI-NEXT:    v_add_u32_e32 v1, vcc, 28, v09271; VI-NEXT:    s_and_b32 s20, s20, 0xffff9272; VI-NEXT:    s_lshl_b32 s21, s21, 169273; VI-NEXT:    buffer_store_dword v8, v1, s[0:3], 0 offen9274; VI-NEXT:    v_add_u32_e32 v1, vcc, 32, v09275; VI-NEXT:    s_or_b32 s20, s20, s219276; VI-NEXT:    buffer_store_dword v9, v1, s[0:3], 0 offen9277; VI-NEXT:    v_add_u32_e32 v1, vcc, 36, v09278; VI-NEXT:    v_mov_b32_e32 v2, s209279; VI-NEXT:    v_readlane_b32 s20, v22, 479280; VI-NEXT:    buffer_store_dword v10, v1, s[0:3], 0 offen9281; VI-NEXT:    v_add_u32_e32 v1, vcc, 40, v09282; VI-NEXT:    s_and_b32 s18, s18, 0xff9283; VI-NEXT:    s_lshl_b32 s20, s20, 89284; VI-NEXT:    buffer_store_dword v11, v1, s[0:3], 0 offen9285; VI-NEXT:    v_add_u32_e32 v1, vcc, 44, v09286; VI-NEXT:    s_or_b32 s18, s18, s209287; VI-NEXT:    v_readlane_b32 s20, v22, 469288; VI-NEXT:    buffer_store_dword v12, v1, s[0:3], 0 offen9289; VI-NEXT:    v_add_u32_e32 v1, vcc, 48, v09290; VI-NEXT:    s_and_b32 s20, s20, 0xff9291; VI-NEXT:    s_lshl_b32 s21, s76, 89292; VI-NEXT:    buffer_store_dword v13, v1, s[0:3], 0 offen9293; VI-NEXT:    v_add_u32_e32 v1, vcc, 52, v09294; VI-NEXT:    s_or_b32 s20, s20, s219295; VI-NEXT:    buffer_store_dword v14, v1, s[0:3], 0 offen9296; VI-NEXT:    v_add_u32_e32 v1, vcc, 56, v09297; VI-NEXT:    s_and_b32 s18, s18, 0xffff9298; VI-NEXT:    s_lshl_b32 s20, s20, 169299; VI-NEXT:    buffer_store_dword v15, v1, s[0:3], 0 offen9300; VI-NEXT:    v_add_u32_e32 v1, vcc, 60, v09301; VI-NEXT:    s_or_b32 s18, s18, s209302; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9303; VI-NEXT:    v_mov_b32_e32 v2, s189304; VI-NEXT:    s_and_b32 s18, s19, 0xff9305; VI-NEXT:    v_readlane_b32 s19, v22, 459306; VI-NEXT:    s_lshl_b32 s19, s19, 89307; VI-NEXT:    s_or_b32 s18, s18, s199308; VI-NEXT:    v_readlane_b32 s19, v22, 449309; VI-NEXT:    v_readlane_b32 s20, v22, 439310; VI-NEXT:    s_and_b32 s19, s19, 0xff9311; VI-NEXT:    s_lshl_b32 s20, s20, 89312; VI-NEXT:    s_or_b32 s19, s19, s209313; VI-NEXT:    s_and_b32 s18, s18, 0xffff9314; VI-NEXT:    s_lshl_b32 s19, s19, 169315; VI-NEXT:    v_add_u32_e32 v1, vcc, 64, v09316; VI-NEXT:    s_or_b32 s18, s18, s199317; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9318; VI-NEXT:    v_mov_b32_e32 v2, s189319; VI-NEXT:    v_readlane_b32 s18, v22, 429320; VI-NEXT:    s_and_b32 s16, s16, 0xff9321; VI-NEXT:    s_lshl_b32 s18, s18, 89322; VI-NEXT:    s_or_b32 s16, s16, s189323; VI-NEXT:    v_readlane_b32 s18, v22, 419324; VI-NEXT:    s_and_b32 s18, s18, 0xff9325; VI-NEXT:    s_lshl_b32 s19, s74, 89326; VI-NEXT:    s_or_b32 s18, s18, s199327; VI-NEXT:    s_and_b32 s16, s16, 0xffff9328; VI-NEXT:    s_lshl_b32 s18, s18, 169329; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x44, v09330; VI-NEXT:    s_or_b32 s16, s16, s189331; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9332; VI-NEXT:    v_mov_b32_e32 v2, s169333; VI-NEXT:    s_and_b32 s16, s17, 0xff9334; VI-NEXT:    v_readlane_b32 s17, v22, 409335; VI-NEXT:    s_lshl_b32 s17, s17, 89336; VI-NEXT:    s_or_b32 s16, s16, s179337; VI-NEXT:    v_readlane_b32 s17, v22, 399338; VI-NEXT:    v_readlane_b32 s18, v22, 389339; VI-NEXT:    s_and_b32 s17, s17, 0xff9340; VI-NEXT:    s_lshl_b32 s18, s18, 89341; VI-NEXT:    s_or_b32 s17, s17, s189342; VI-NEXT:    s_and_b32 s16, s16, 0xffff9343; VI-NEXT:    s_lshl_b32 s17, s17, 169344; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x48, v09345; VI-NEXT:    s_or_b32 s16, s16, s179346; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9347; VI-NEXT:    v_mov_b32_e32 v2, s169348; VI-NEXT:    v_readlane_b32 s16, v22, 379349; VI-NEXT:    s_and_b32 s14, s14, 0xff9350; VI-NEXT:    s_lshl_b32 s16, s16, 89351; VI-NEXT:    s_or_b32 s14, s14, s169352; VI-NEXT:    v_readlane_b32 s16, v22, 369353; VI-NEXT:    s_and_b32 s16, s16, 0xff9354; VI-NEXT:    s_lshl_b32 s17, s72, 89355; VI-NEXT:    s_or_b32 s16, s16, s179356; VI-NEXT:    s_and_b32 s14, s14, 0xffff9357; VI-NEXT:    s_lshl_b32 s16, s16, 169358; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x4c, v09359; VI-NEXT:    s_or_b32 s14, s14, s169360; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9361; VI-NEXT:    v_mov_b32_e32 v2, s149362; VI-NEXT:    s_and_b32 s14, s15, 0xff9363; VI-NEXT:    v_readlane_b32 s15, v22, 359364; VI-NEXT:    s_lshl_b32 s15, s15, 89365; VI-NEXT:    s_or_b32 s14, s14, s159366; VI-NEXT:    v_readlane_b32 s15, v22, 349367; VI-NEXT:    v_readlane_b32 s16, v22, 339368; VI-NEXT:    s_and_b32 s15, s15, 0xff9369; VI-NEXT:    s_lshl_b32 s16, s16, 89370; VI-NEXT:    s_or_b32 s15, s15, s169371; VI-NEXT:    s_and_b32 s14, s14, 0xffff9372; VI-NEXT:    s_lshl_b32 s15, s15, 169373; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x50, v09374; VI-NEXT:    s_or_b32 s14, s14, s159375; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9376; VI-NEXT:    v_mov_b32_e32 v2, s149377; VI-NEXT:    v_readlane_b32 s14, v22, 329378; VI-NEXT:    s_and_b32 s12, s12, 0xff9379; VI-NEXT:    s_lshl_b32 s14, s14, 89380; VI-NEXT:    s_or_b32 s12, s12, s149381; VI-NEXT:    v_readlane_b32 s14, v22, 319382; VI-NEXT:    s_and_b32 s14, s14, 0xff9383; VI-NEXT:    s_lshl_b32 s15, s62, 89384; VI-NEXT:    s_or_b32 s14, s14, s159385; VI-NEXT:    s_and_b32 s12, s12, 0xffff9386; VI-NEXT:    s_lshl_b32 s14, s14, 169387; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x54, v09388; VI-NEXT:    s_or_b32 s12, s12, s149389; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9390; VI-NEXT:    v_mov_b32_e32 v2, s129391; VI-NEXT:    s_and_b32 s12, s13, 0xff9392; VI-NEXT:    v_readlane_b32 s13, v22, 309393; VI-NEXT:    s_lshl_b32 s13, s13, 89394; VI-NEXT:    s_or_b32 s12, s12, s139395; VI-NEXT:    v_readlane_b32 s13, v22, 299396; VI-NEXT:    v_readlane_b32 s14, v22, 289397; VI-NEXT:    s_and_b32 s13, s13, 0xff9398; VI-NEXT:    s_lshl_b32 s14, s14, 89399; VI-NEXT:    s_or_b32 s13, s13, s149400; VI-NEXT:    s_and_b32 s12, s12, 0xffff9401; VI-NEXT:    s_lshl_b32 s13, s13, 169402; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x58, v09403; VI-NEXT:    s_or_b32 s12, s12, s139404; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9405; VI-NEXT:    v_mov_b32_e32 v2, s129406; VI-NEXT:    v_readlane_b32 s12, v22, 279407; VI-NEXT:    s_and_b32 s10, s10, 0xff9408; VI-NEXT:    s_lshl_b32 s12, s12, 89409; VI-NEXT:    s_or_b32 s10, s10, s129410; VI-NEXT:    v_readlane_b32 s12, v22, 269411; VI-NEXT:    v_readlane_b32 s14, v22, 09412; VI-NEXT:    s_and_b32 s12, s12, 0xff9413; VI-NEXT:    s_lshl_b32 s13, s14, 89414; VI-NEXT:    s_or_b32 s12, s12, s139415; VI-NEXT:    s_and_b32 s10, s10, 0xffff9416; VI-NEXT:    s_lshl_b32 s12, s12, 169417; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x5c, v09418; VI-NEXT:    s_or_b32 s10, s10, s129419; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9420; VI-NEXT:    v_mov_b32_e32 v2, s109421; VI-NEXT:    s_and_b32 s10, s11, 0xff9422; VI-NEXT:    v_readlane_b32 s11, v22, 259423; VI-NEXT:    s_lshl_b32 s11, s11, 89424; VI-NEXT:    s_or_b32 s10, s10, s119425; VI-NEXT:    v_readlane_b32 s11, v22, 249426; VI-NEXT:    v_readlane_b32 s12, v22, 239427; VI-NEXT:    s_and_b32 s11, s11, 0xff9428; VI-NEXT:    s_lshl_b32 s12, s12, 89429; VI-NEXT:    s_or_b32 s11, s11, s129430; VI-NEXT:    s_and_b32 s10, s10, 0xffff9431; VI-NEXT:    s_lshl_b32 s11, s11, 169432; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x60, v09433; VI-NEXT:    s_or_b32 s10, s10, s119434; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9435; VI-NEXT:    v_mov_b32_e32 v2, s109436; VI-NEXT:    v_readlane_b32 s10, v22, 229437; VI-NEXT:    s_and_b32 s8, s8, 0xff9438; VI-NEXT:    s_lshl_b32 s10, s10, 89439; VI-NEXT:    s_or_b32 s8, s8, s109440; VI-NEXT:    v_readlane_b32 s10, v22, 219441; VI-NEXT:    v_readlane_b32 s12, v22, 29442; VI-NEXT:    s_and_b32 s10, s10, 0xff9443; VI-NEXT:    s_lshl_b32 s11, s12, 89444; VI-NEXT:    s_or_b32 s10, s10, s119445; VI-NEXT:    s_and_b32 s8, s8, 0xffff9446; VI-NEXT:    s_lshl_b32 s10, s10, 169447; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x64, v09448; VI-NEXT:    s_or_b32 s8, s8, s109449; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9450; VI-NEXT:    v_mov_b32_e32 v2, s89451; VI-NEXT:    s_and_b32 s8, s9, 0xff9452; VI-NEXT:    v_readlane_b32 s9, v22, 209453; VI-NEXT:    s_lshl_b32 s9, s9, 89454; VI-NEXT:    s_or_b32 s8, s8, s99455; VI-NEXT:    v_readlane_b32 s9, v22, 199456; VI-NEXT:    v_readlane_b32 s10, v22, 189457; VI-NEXT:    s_and_b32 s9, s9, 0xff9458; VI-NEXT:    s_lshl_b32 s10, s10, 89459; VI-NEXT:    s_or_b32 s9, s9, s109460; VI-NEXT:    s_and_b32 s8, s8, 0xffff9461; VI-NEXT:    s_lshl_b32 s9, s9, 169462; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x68, v09463; VI-NEXT:    s_or_b32 s8, s8, s99464; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9465; VI-NEXT:    v_mov_b32_e32 v2, s89466; VI-NEXT:    v_readlane_b32 s8, v22, 179467; VI-NEXT:    s_and_b32 s6, s6, 0xff9468; VI-NEXT:    s_lshl_b32 s8, s8, 89469; VI-NEXT:    s_or_b32 s6, s6, s89470; VI-NEXT:    v_readlane_b32 s8, v22, 169471; VI-NEXT:    v_readlane_b32 s10, v22, 49472; VI-NEXT:    s_and_b32 s8, s8, 0xff9473; VI-NEXT:    s_lshl_b32 s9, s10, 89474; VI-NEXT:    s_or_b32 s8, s8, s99475; VI-NEXT:    s_and_b32 s6, s6, 0xffff9476; VI-NEXT:    s_lshl_b32 s8, s8, 169477; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x6c, v09478; VI-NEXT:    s_or_b32 s6, s6, s89479; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9480; VI-NEXT:    v_mov_b32_e32 v2, s69481; VI-NEXT:    s_and_b32 s6, s7, 0xff9482; VI-NEXT:    v_readlane_b32 s7, v22, 159483; VI-NEXT:    s_lshl_b32 s7, s7, 89484; VI-NEXT:    s_or_b32 s6, s6, s79485; VI-NEXT:    v_readlane_b32 s7, v22, 149486; VI-NEXT:    v_readlane_b32 s8, v22, 139487; VI-NEXT:    s_and_b32 s7, s7, 0xff9488; VI-NEXT:    s_lshl_b32 s8, s8, 89489; VI-NEXT:    s_or_b32 s7, s7, s89490; VI-NEXT:    s_and_b32 s6, s6, 0xffff9491; VI-NEXT:    s_lshl_b32 s7, s7, 169492; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x70, v09493; VI-NEXT:    s_or_b32 s6, s6, s79494; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9495; VI-NEXT:    v_mov_b32_e32 v2, s69496; VI-NEXT:    v_readlane_b32 s6, v22, 129497; VI-NEXT:    s_and_b32 s4, s4, 0xff9498; VI-NEXT:    s_lshl_b32 s6, s6, 89499; VI-NEXT:    s_or_b32 s4, s4, s69500; VI-NEXT:    v_readlane_b32 s6, v22, 119501; VI-NEXT:    v_readlane_b32 s8, v22, 69502; VI-NEXT:    s_and_b32 s6, s6, 0xff9503; VI-NEXT:    s_lshl_b32 s7, s8, 89504; VI-NEXT:    s_or_b32 s6, s6, s79505; VI-NEXT:    s_and_b32 s4, s4, 0xffff9506; VI-NEXT:    s_lshl_b32 s6, s6, 169507; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x74, v09508; VI-NEXT:    s_or_b32 s4, s4, s69509; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9510; VI-NEXT:    v_mov_b32_e32 v2, s49511; VI-NEXT:    s_and_b32 s4, s5, 0xff9512; VI-NEXT:    v_readlane_b32 s5, v22, 109513; VI-NEXT:    s_lshl_b32 s5, s5, 89514; VI-NEXT:    s_or_b32 s4, s4, s59515; VI-NEXT:    v_readlane_b32 s5, v22, 99516; VI-NEXT:    v_readlane_b32 s6, v22, 89517; VI-NEXT:    s_and_b32 s5, s5, 0xff9518; VI-NEXT:    s_lshl_b32 s6, s6, 89519; VI-NEXT:    s_or_b32 s5, s5, s69520; VI-NEXT:    s_and_b32 s4, s4, 0xffff9521; VI-NEXT:    s_lshl_b32 s5, s5, 169522; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x78, v09523; VI-NEXT:    s_or_b32 s4, s4, s59524; VI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen9525; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v09526; VI-NEXT:    v_mov_b32_e32 v1, s49527; VI-NEXT:    v_readlane_b32 s15, v22, 19528; VI-NEXT:    v_readlane_b32 s13, v22, 39529; VI-NEXT:    v_readlane_b32 s11, v22, 59530; VI-NEXT:    v_readlane_b32 s9, v22, 79531; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen9532; VI-NEXT:    v_readlane_b32 s87, v21, 319533; VI-NEXT:    v_readlane_b32 s86, v21, 309534; VI-NEXT:    v_readlane_b32 s85, v21, 299535; VI-NEXT:    v_readlane_b32 s84, v21, 289536; VI-NEXT:    v_readlane_b32 s83, v21, 279537; VI-NEXT:    v_readlane_b32 s82, v21, 269538; VI-NEXT:    v_readlane_b32 s81, v21, 259539; VI-NEXT:    v_readlane_b32 s80, v21, 249540; VI-NEXT:    v_readlane_b32 s71, v21, 239541; VI-NEXT:    v_readlane_b32 s70, v21, 229542; VI-NEXT:    v_readlane_b32 s69, v21, 219543; VI-NEXT:    v_readlane_b32 s68, v21, 209544; VI-NEXT:    v_readlane_b32 s67, v21, 199545; VI-NEXT:    v_readlane_b32 s66, v21, 189546; VI-NEXT:    v_readlane_b32 s65, v21, 179547; VI-NEXT:    v_readlane_b32 s64, v21, 169548; VI-NEXT:    v_readlane_b32 s55, v21, 159549; VI-NEXT:    v_readlane_b32 s54, v21, 149550; VI-NEXT:    v_readlane_b32 s53, v21, 139551; VI-NEXT:    v_readlane_b32 s52, v21, 129552; VI-NEXT:    v_readlane_b32 s51, v21, 119553; VI-NEXT:    v_readlane_b32 s50, v21, 109554; VI-NEXT:    v_readlane_b32 s49, v21, 99555; VI-NEXT:    v_readlane_b32 s48, v21, 89556; VI-NEXT:    v_readlane_b32 s39, v21, 79557; VI-NEXT:    v_readlane_b32 s38, v21, 69558; VI-NEXT:    v_readlane_b32 s37, v21, 59559; VI-NEXT:    v_readlane_b32 s36, v21, 49560; VI-NEXT:    v_readlane_b32 s35, v21, 39561; VI-NEXT:    v_readlane_b32 s34, v21, 29562; VI-NEXT:    v_readlane_b32 s31, v21, 19563; VI-NEXT:    v_readlane_b32 s30, v21, 09564; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -19565; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 ; 4-byte Folded Reload9566; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload9567; VI-NEXT:    s_mov_b64 exec, s[4:5]9568; VI-NEXT:    s_waitcnt vmcnt(0)9569; VI-NEXT:    s_setpc_b64 s[30:31]9570; VI-NEXT:  .LBB13_4:9571; VI-NEXT:    ; implicit-def: $sgpr609572; VI-NEXT:    ; kill: killed $sgpr609573; VI-NEXT:    ; implicit-def: $sgpr609574; VI-NEXT:    ; kill: killed $sgpr609575; VI-NEXT:    ; implicit-def: $sgpr659576; VI-NEXT:    ; implicit-def: $sgpr649577; VI-NEXT:    ; implicit-def: $sgpr559578; VI-NEXT:    ; implicit-def: $sgpr599579; VI-NEXT:    ; implicit-def: $sgpr589580; VI-NEXT:    ; implicit-def: $sgpr549581; VI-NEXT:    ; implicit-def: $sgpr539582; VI-NEXT:    ; implicit-def: $sgpr529583; VI-NEXT:    ; implicit-def: $sgpr519584; VI-NEXT:    ; implicit-def: $sgpr299585; VI-NEXT:    ; implicit-def: $sgpr289586; VI-NEXT:    ; implicit-def: $sgpr279587; VI-NEXT:    ; implicit-def: $sgpr269588; VI-NEXT:    ; implicit-def: $sgpr509589; VI-NEXT:    ; implicit-def: $sgpr879590; VI-NEXT:    ; implicit-def: $sgpr869591; VI-NEXT:    ; implicit-def: $sgpr859592; VI-NEXT:    ; implicit-def: $sgpr849593; VI-NEXT:    ; implicit-def: $sgpr839594; VI-NEXT:    ; implicit-def: $sgpr829595; VI-NEXT:    ; implicit-def: $sgpr819596; VI-NEXT:    ; implicit-def: $sgpr809597; VI-NEXT:    ; implicit-def: $sgpr719598; VI-NEXT:    ; implicit-def: $sgpr709599; VI-NEXT:    ; implicit-def: $sgpr699600; VI-NEXT:    ; implicit-def: $sgpr689601; VI-NEXT:    ; implicit-def: $sgpr679602; VI-NEXT:    ; implicit-def: $sgpr669603; VI-NEXT:    ; implicit-def: $sgpr489604; VI-NEXT:    ; implicit-def: $sgpr389605; VI-NEXT:    ; implicit-def: $sgpr369606; VI-NEXT:    ; implicit-def: $sgpr349607; VI-NEXT:    ; implicit-def: $sgpr309608; VI-NEXT:    ; implicit-def: $sgpr909609; VI-NEXT:    ; implicit-def: $sgpr889610; VI-NEXT:    ; implicit-def: $sgpr789611; VI-NEXT:    ; implicit-def: $sgpr769612; VI-NEXT:    ; implicit-def: $sgpr749613; VI-NEXT:    ; implicit-def: $sgpr729614; VI-NEXT:    ; implicit-def: $sgpr629615; VI-NEXT:    ; implicit-def: $sgpr609616; VI-NEXT:    ; kill: killed $sgpr609617; VI-NEXT:    ; implicit-def: $sgpr609618; VI-NEXT:    ; kill: killed $sgpr609619; VI-NEXT:    ; implicit-def: $sgpr609620; VI-NEXT:    ; kill: killed $sgpr609621; VI-NEXT:    ; implicit-def: $sgpr609622; VI-NEXT:    ; kill: killed $sgpr609623; VI-NEXT:    ; implicit-def: $sgpr609624; VI-NEXT:    ; kill: killed $sgpr609625; VI-NEXT:    ; implicit-def: $sgpr609626; VI-NEXT:    ; kill: killed $sgpr609627; VI-NEXT:    ; implicit-def: $sgpr609628; VI-NEXT:    ; kill: killed $sgpr609629; VI-NEXT:    ; implicit-def: $sgpr609630; VI-NEXT:    ; kill: killed $sgpr609631; VI-NEXT:    ; implicit-def: $sgpr609632; VI-NEXT:    ; kill: killed $sgpr609633; VI-NEXT:    ; implicit-def: $sgpr609634; VI-NEXT:    ; kill: killed $sgpr609635; VI-NEXT:    ; implicit-def: $sgpr609636; VI-NEXT:    ; kill: killed $sgpr609637; VI-NEXT:    ; implicit-def: $sgpr609638; VI-NEXT:    ; kill: killed $sgpr609639; VI-NEXT:    ; implicit-def: $sgpr609640; VI-NEXT:    ; kill: killed $sgpr609641; VI-NEXT:    ; implicit-def: $sgpr609642; VI-NEXT:    ; kill: killed $sgpr609643; VI-NEXT:    ; implicit-def: $sgpr609644; VI-NEXT:    ; kill: killed $sgpr609645; VI-NEXT:    ; implicit-def: $sgpr609646; VI-NEXT:    ; kill: killed $sgpr609647; VI-NEXT:    ; implicit-def: $sgpr609648; VI-NEXT:    ; kill: killed $sgpr609649; VI-NEXT:    ; implicit-def: $sgpr609650; VI-NEXT:    ; kill: killed $sgpr609651; VI-NEXT:    ; implicit-def: $sgpr609652; VI-NEXT:    ; kill: killed $sgpr609653; VI-NEXT:    ; implicit-def: $sgpr609654; VI-NEXT:    ; kill: killed $sgpr609655; VI-NEXT:    ; implicit-def: $sgpr609656; VI-NEXT:    ; kill: killed $sgpr609657; VI-NEXT:    ; implicit-def: $sgpr609658; VI-NEXT:    ; kill: killed $sgpr609659; VI-NEXT:    ; implicit-def: $sgpr609660; VI-NEXT:    ; kill: killed $sgpr609661; VI-NEXT:    ; implicit-def: $sgpr609662; VI-NEXT:    ; kill: killed $sgpr609663; VI-NEXT:    ; implicit-def: $sgpr609664; VI-NEXT:    ; kill: killed $sgpr609665; VI-NEXT:    ; implicit-def: $sgpr609666; VI-NEXT:    ; kill: killed $sgpr609667; VI-NEXT:    ; implicit-def: $sgpr609668; VI-NEXT:    ; kill: killed $sgpr609669; VI-NEXT:    ; implicit-def: $sgpr609670; VI-NEXT:    ; kill: killed $sgpr609671; VI-NEXT:    ; implicit-def: $sgpr609672; VI-NEXT:    ; kill: killed $sgpr609673; VI-NEXT:    ; implicit-def: $sgpr609674; VI-NEXT:    ; kill: killed $sgpr609675; VI-NEXT:    ; implicit-def: $sgpr609676; VI-NEXT:    ; kill: killed $sgpr609677; VI-NEXT:    ; implicit-def: $sgpr609678; VI-NEXT:    ; kill: killed $sgpr609679; VI-NEXT:    ; implicit-def: $sgpr609680; VI-NEXT:    ; kill: killed $sgpr609681; VI-NEXT:    ; implicit-def: $sgpr609682; VI-NEXT:    ; kill: killed $sgpr609683; VI-NEXT:    ; implicit-def: $sgpr609684; VI-NEXT:    ; kill: killed $sgpr609685; VI-NEXT:    ; implicit-def: $sgpr609686; VI-NEXT:    ; kill: killed $sgpr609687; VI-NEXT:    ; implicit-def: $sgpr609688; VI-NEXT:    ; kill: killed $sgpr609689; VI-NEXT:    ; implicit-def: $sgpr609690; VI-NEXT:    ; kill: killed $sgpr609691; VI-NEXT:    ; implicit-def: $sgpr609692; VI-NEXT:    ; kill: killed $sgpr609693; VI-NEXT:    ; implicit-def: $sgpr609694; VI-NEXT:    ; kill: killed $sgpr609695; VI-NEXT:    ; implicit-def: $sgpr609696; VI-NEXT:    ; kill: killed $sgpr609697; VI-NEXT:    ; implicit-def: $sgpr609698; VI-NEXT:    ; kill: killed $sgpr609699; VI-NEXT:    ; implicit-def: $sgpr609700; VI-NEXT:    ; kill: killed $sgpr609701; VI-NEXT:    ; implicit-def: $sgpr609702; VI-NEXT:    ; kill: killed $sgpr609703; VI-NEXT:    ; implicit-def: $sgpr609704; VI-NEXT:    ; kill: killed $sgpr609705; VI-NEXT:    ; implicit-def: $sgpr609706; VI-NEXT:    ; kill: killed $sgpr609707; VI-NEXT:    ; implicit-def: $sgpr609708; VI-NEXT:    ; kill: killed $sgpr609709; VI-NEXT:    ; implicit-def: $sgpr609710; VI-NEXT:    ; kill: killed $sgpr609711; VI-NEXT:    ; implicit-def: $sgpr609712; VI-NEXT:    ; kill: killed $sgpr609713; VI-NEXT:    ; implicit-def: $sgpr609714; VI-NEXT:    ; kill: killed $sgpr609715; VI-NEXT:    ; implicit-def: $sgpr609716; VI-NEXT:    v_writelane_b32 v22, s60, 09717; VI-NEXT:    v_writelane_b32 v22, s61, 19718; VI-NEXT:    ; implicit-def: $sgpr609719; VI-NEXT:    v_writelane_b32 v22, s60, 29720; VI-NEXT:    v_writelane_b32 v22, s61, 39721; VI-NEXT:    ; implicit-def: $sgpr609722; VI-NEXT:    v_writelane_b32 v22, s60, 49723; VI-NEXT:    v_writelane_b32 v22, s61, 59724; VI-NEXT:    ; implicit-def: $sgpr609725; VI-NEXT:    v_writelane_b32 v22, s60, 69726; VI-NEXT:    v_writelane_b32 v22, s61, 79727; VI-NEXT:    s_branch .LBB13_29728;9729; GFX9-LABEL: bitcast_v32i32_to_v128i8_scalar:9730; GFX9:       ; %bb.0:9731; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)9732; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -19733; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 ; 4-byte Folded Spill9734; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill9735; GFX9-NEXT:    s_mov_b64 exec, s[4:5]9736; GFX9-NEXT:    v_writelane_b32 v21, s30, 09737; GFX9-NEXT:    v_writelane_b32 v21, s31, 19738; GFX9-NEXT:    v_writelane_b32 v21, s34, 29739; GFX9-NEXT:    v_writelane_b32 v21, s35, 39740; GFX9-NEXT:    v_writelane_b32 v21, s36, 49741; GFX9-NEXT:    v_writelane_b32 v21, s37, 59742; GFX9-NEXT:    v_writelane_b32 v21, s38, 69743; GFX9-NEXT:    v_writelane_b32 v21, s39, 79744; GFX9-NEXT:    v_writelane_b32 v21, s48, 89745; GFX9-NEXT:    v_writelane_b32 v21, s49, 99746; GFX9-NEXT:    v_writelane_b32 v21, s50, 109747; GFX9-NEXT:    v_writelane_b32 v21, s51, 119748; GFX9-NEXT:    v_writelane_b32 v21, s52, 129749; GFX9-NEXT:    v_writelane_b32 v21, s53, 139750; GFX9-NEXT:    v_writelane_b32 v21, s54, 149751; GFX9-NEXT:    v_writelane_b32 v21, s55, 159752; GFX9-NEXT:    v_writelane_b32 v21, s64, 169753; GFX9-NEXT:    v_writelane_b32 v21, s65, 179754; GFX9-NEXT:    v_writelane_b32 v21, s66, 189755; GFX9-NEXT:    v_writelane_b32 v21, s67, 199756; GFX9-NEXT:    v_writelane_b32 v21, s68, 209757; GFX9-NEXT:    v_mov_b32_e32 v20, s169758; GFX9-NEXT:    v_writelane_b32 v21, s69, 219759; GFX9-NEXT:    v_readfirstlane_b32 s56, v209760; GFX9-NEXT:    v_mov_b32_e32 v20, s179761; GFX9-NEXT:    v_writelane_b32 v21, s70, 229762; GFX9-NEXT:    v_readfirstlane_b32 s57, v209763; GFX9-NEXT:    v_mov_b32_e32 v20, s189764; GFX9-NEXT:    v_writelane_b32 v21, s71, 239765; GFX9-NEXT:    v_readfirstlane_b32 s46, v209766; GFX9-NEXT:    v_mov_b32_e32 v20, s199767; GFX9-NEXT:    v_writelane_b32 v21, s80, 249768; GFX9-NEXT:    v_readfirstlane_b32 s47, v209769; GFX9-NEXT:    v_mov_b32_e32 v20, s209770; GFX9-NEXT:    v_writelane_b32 v21, s81, 259771; GFX9-NEXT:    v_readfirstlane_b32 s44, v209772; GFX9-NEXT:    v_mov_b32_e32 v20, s219773; GFX9-NEXT:    v_writelane_b32 v21, s82, 269774; GFX9-NEXT:    v_readfirstlane_b32 s45, v209775; GFX9-NEXT:    v_mov_b32_e32 v20, s229776; GFX9-NEXT:    v_writelane_b32 v21, s83, 279777; GFX9-NEXT:    v_readfirstlane_b32 s42, v209778; GFX9-NEXT:    v_mov_b32_e32 v20, s239779; GFX9-NEXT:    v_writelane_b32 v21, s84, 289780; GFX9-NEXT:    v_readfirstlane_b32 s43, v209781; GFX9-NEXT:    v_mov_b32_e32 v20, s249782; GFX9-NEXT:    v_writelane_b32 v21, s85, 299783; GFX9-NEXT:    v_readfirstlane_b32 s40, v209784; GFX9-NEXT:    v_mov_b32_e32 v20, s259785; GFX9-NEXT:    v_writelane_b32 v21, s86, 309786; GFX9-NEXT:    v_readfirstlane_b32 s41, v209787; GFX9-NEXT:    v_mov_b32_e32 v20, s269788; GFX9-NEXT:    v_writelane_b32 v21, s87, 319789; GFX9-NEXT:    v_readfirstlane_b32 s24, v209790; GFX9-NEXT:    v_mov_b32_e32 v20, s279791; GFX9-NEXT:    v_writelane_b32 v21, s96, 329792; GFX9-NEXT:    v_readfirstlane_b32 s25, v209793; GFX9-NEXT:    v_mov_b32_e32 v20, s289794; GFX9-NEXT:    v_writelane_b32 v21, s97, 339795; GFX9-NEXT:    v_readfirstlane_b32 s22, v209796; GFX9-NEXT:    v_mov_b32_e32 v20, s299797; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v199798; GFX9-NEXT:    v_writelane_b32 v21, s98, 349799; GFX9-NEXT:    v_readfirstlane_b32 s23, v209800; GFX9-NEXT:    v_readfirstlane_b32 s20, v19801; GFX9-NEXT:    v_readfirstlane_b32 s21, v29802; GFX9-NEXT:    v_readfirstlane_b32 s18, v39803; GFX9-NEXT:    v_readfirstlane_b32 s19, v49804; GFX9-NEXT:    v_readfirstlane_b32 s16, v59805; GFX9-NEXT:    v_readfirstlane_b32 s17, v69806; GFX9-NEXT:    v_readfirstlane_b32 s14, v79807; GFX9-NEXT:    v_readfirstlane_b32 s15, v89808; GFX9-NEXT:    v_readfirstlane_b32 s12, v99809; GFX9-NEXT:    v_readfirstlane_b32 s13, v109810; GFX9-NEXT:    v_readfirstlane_b32 s10, v119811; GFX9-NEXT:    v_readfirstlane_b32 s11, v129812; GFX9-NEXT:    v_readfirstlane_b32 s8, v139813; GFX9-NEXT:    v_readfirstlane_b32 s9, v149814; GFX9-NEXT:    v_readfirstlane_b32 s6, v159815; GFX9-NEXT:    v_readfirstlane_b32 s7, v169816; GFX9-NEXT:    v_readfirstlane_b32 s4, v179817; GFX9-NEXT:    s_and_b64 s[26:27], vcc, exec9818; GFX9-NEXT:    v_readfirstlane_b32 s5, v189819; GFX9-NEXT:    v_writelane_b32 v21, s99, 359820; GFX9-NEXT:    ; implicit-def: $vgpr22 : SGPR spill to VGPR lane9821; GFX9-NEXT:    s_cbranch_scc0 .LBB13_49822; GFX9-NEXT:  ; %bb.1: ; %cmp.false9823; GFX9-NEXT:    s_lshr_b32 s26, s5, 249824; GFX9-NEXT:    v_writelane_b32 v22, s26, 29825; GFX9-NEXT:    s_lshr_b32 s26, s5, 169826; GFX9-NEXT:    v_writelane_b32 v22, s26, 39827; GFX9-NEXT:    s_lshr_b32 s26, s5, 89828; GFX9-NEXT:    v_writelane_b32 v22, s26, 49829; GFX9-NEXT:    s_lshr_b32 s26, s4, 169830; GFX9-NEXT:    v_writelane_b32 v22, s26, 59831; GFX9-NEXT:    s_lshr_b32 s26, s4, 89832; GFX9-NEXT:    v_writelane_b32 v22, s26, 69833; GFX9-NEXT:    s_lshr_b32 s26, s7, 249834; GFX9-NEXT:    v_writelane_b32 v22, s26, 79835; GFX9-NEXT:    s_lshr_b32 s26, s7, 169836; GFX9-NEXT:    v_writelane_b32 v22, s26, 89837; GFX9-NEXT:    s_lshr_b32 s26, s7, 89838; GFX9-NEXT:    v_writelane_b32 v22, s26, 99839; GFX9-NEXT:    s_lshr_b32 s26, s6, 169840; GFX9-NEXT:    v_writelane_b32 v22, s26, 109841; GFX9-NEXT:    s_lshr_b32 s26, s6, 89842; GFX9-NEXT:    v_writelane_b32 v22, s26, 119843; GFX9-NEXT:    s_lshr_b32 s26, s9, 249844; GFX9-NEXT:    v_writelane_b32 v22, s26, 129845; GFX9-NEXT:    s_lshr_b32 s26, s9, 169846; GFX9-NEXT:    v_writelane_b32 v22, s26, 139847; GFX9-NEXT:    s_lshr_b32 s26, s9, 89848; GFX9-NEXT:    v_writelane_b32 v22, s26, 149849; GFX9-NEXT:    s_lshr_b32 s26, s8, 169850; GFX9-NEXT:    v_writelane_b32 v22, s26, 159851; GFX9-NEXT:    s_lshr_b32 s26, s8, 89852; GFX9-NEXT:    v_writelane_b32 v22, s26, 169853; GFX9-NEXT:    s_lshr_b32 s26, s11, 249854; GFX9-NEXT:    v_writelane_b32 v22, s26, 179855; GFX9-NEXT:    s_lshr_b32 s26, s11, 169856; GFX9-NEXT:    v_writelane_b32 v22, s26, 189857; GFX9-NEXT:    s_lshr_b32 s26, s11, 89858; GFX9-NEXT:    v_writelane_b32 v22, s26, 199859; GFX9-NEXT:    s_lshr_b32 s26, s10, 169860; GFX9-NEXT:    v_writelane_b32 v22, s26, 209861; GFX9-NEXT:    s_lshr_b32 s26, s10, 89862; GFX9-NEXT:    v_writelane_b32 v22, s26, 219863; GFX9-NEXT:    s_lshr_b32 s26, s13, 249864; GFX9-NEXT:    v_writelane_b32 v22, s26, 229865; GFX9-NEXT:    s_lshr_b32 s26, s13, 169866; GFX9-NEXT:    v_writelane_b32 v22, s26, 239867; GFX9-NEXT:    s_lshr_b32 s26, s13, 89868; GFX9-NEXT:    v_writelane_b32 v22, s26, 249869; GFX9-NEXT:    s_lshr_b32 s26, s12, 169870; GFX9-NEXT:    v_writelane_b32 v22, s26, 259871; GFX9-NEXT:    s_lshr_b32 s26, s12, 89872; GFX9-NEXT:    v_writelane_b32 v22, s26, 269873; GFX9-NEXT:    s_lshr_b32 s26, s15, 249874; GFX9-NEXT:    v_writelane_b32 v22, s26, 279875; GFX9-NEXT:    s_lshr_b32 s26, s15, 169876; GFX9-NEXT:    v_writelane_b32 v22, s26, 289877; GFX9-NEXT:    s_lshr_b32 s26, s15, 89878; GFX9-NEXT:    v_writelane_b32 v22, s26, 299879; GFX9-NEXT:    s_lshr_b32 s26, s14, 169880; GFX9-NEXT:    v_writelane_b32 v22, s26, 309881; GFX9-NEXT:    s_lshr_b32 s26, s14, 89882; GFX9-NEXT:    v_writelane_b32 v22, s26, 319883; GFX9-NEXT:    s_lshr_b32 s26, s17, 249884; GFX9-NEXT:    v_writelane_b32 v22, s26, 329885; GFX9-NEXT:    s_lshr_b32 s26, s17, 169886; GFX9-NEXT:    v_writelane_b32 v22, s26, 339887; GFX9-NEXT:    s_lshr_b32 s26, s17, 89888; GFX9-NEXT:    v_writelane_b32 v22, s26, 349889; GFX9-NEXT:    s_lshr_b32 s26, s16, 169890; GFX9-NEXT:    v_writelane_b32 v22, s26, 359891; GFX9-NEXT:    s_lshr_b32 s26, s16, 89892; GFX9-NEXT:    v_writelane_b32 v22, s26, 369893; GFX9-NEXT:    s_lshr_b32 s26, s19, 249894; GFX9-NEXT:    v_writelane_b32 v22, s26, 379895; GFX9-NEXT:    s_lshr_b32 s26, s19, 169896; GFX9-NEXT:    v_writelane_b32 v22, s26, 389897; GFX9-NEXT:    s_lshr_b32 s26, s19, 89898; GFX9-NEXT:    v_writelane_b32 v22, s26, 399899; GFX9-NEXT:    s_lshr_b32 s26, s18, 169900; GFX9-NEXT:    v_writelane_b32 v22, s26, 409901; GFX9-NEXT:    s_lshr_b32 s26, s18, 89902; GFX9-NEXT:    v_writelane_b32 v22, s26, 419903; GFX9-NEXT:    s_lshr_b32 s26, s21, 249904; GFX9-NEXT:    v_writelane_b32 v22, s26, 429905; GFX9-NEXT:    s_lshr_b32 s26, s21, 169906; GFX9-NEXT:    v_writelane_b32 v22, s26, 439907; GFX9-NEXT:    s_lshr_b32 s26, s21, 89908; GFX9-NEXT:    v_writelane_b32 v22, s26, 449909; GFX9-NEXT:    s_lshr_b32 s26, s20, 169910; GFX9-NEXT:    v_writelane_b32 v22, s26, 459911; GFX9-NEXT:    s_lshr_b32 s26, s20, 89912; GFX9-NEXT:    v_writelane_b32 v22, s26, 469913; GFX9-NEXT:    s_lshr_b32 s26, s23, 249914; GFX9-NEXT:    v_writelane_b32 v22, s26, 479915; GFX9-NEXT:    s_lshr_b32 s26, s23, 169916; GFX9-NEXT:    v_writelane_b32 v22, s26, 489917; GFX9-NEXT:    s_lshr_b32 s26, s23, 89918; GFX9-NEXT:    v_writelane_b32 v22, s26, 499919; GFX9-NEXT:    s_lshr_b32 s26, s22, 169920; GFX9-NEXT:    v_writelane_b32 v22, s26, 509921; GFX9-NEXT:    s_lshr_b64 s[28:29], s[4:5], 249922; GFX9-NEXT:    v_writelane_b32 v22, s28, 09923; GFX9-NEXT:    s_lshr_b32 s82, s22, 89924; GFX9-NEXT:    s_lshr_b32 s83, s25, 249925; GFX9-NEXT:    s_lshr_b32 s81, s25, 169926; GFX9-NEXT:    s_lshr_b32 s84, s25, 89927; GFX9-NEXT:    s_lshr_b32 s85, s24, 169928; GFX9-NEXT:    s_lshr_b32 s86, s24, 89929; GFX9-NEXT:    s_lshr_b32 s87, s41, 249930; GFX9-NEXT:    s_lshr_b32 s96, s41, 169931; GFX9-NEXT:    s_lshr_b32 s97, s41, 89932; GFX9-NEXT:    s_lshr_b32 s98, s40, 169933; GFX9-NEXT:    s_lshr_b32 s99, s40, 89934; GFX9-NEXT:    s_lshr_b32 s38, s43, 249935; GFX9-NEXT:    s_lshr_b32 s39, s43, 169936; GFX9-NEXT:    s_lshr_b32 s48, s43, 89937; GFX9-NEXT:    s_lshr_b32 s49, s42, 169938; GFX9-NEXT:    s_lshr_b32 s50, s42, 89939; GFX9-NEXT:    s_lshr_b32 s51, s45, 249940; GFX9-NEXT:    s_lshr_b32 s52, s45, 169941; GFX9-NEXT:    s_lshr_b32 s53, s45, 89942; GFX9-NEXT:    s_lshr_b32 s54, s44, 169943; GFX9-NEXT:    s_lshr_b32 s55, s44, 89944; GFX9-NEXT:    s_lshr_b32 s64, s47, 249945; GFX9-NEXT:    s_lshr_b32 s65, s47, 169946; GFX9-NEXT:    s_lshr_b32 s66, s47, 89947; GFX9-NEXT:    s_lshr_b32 s67, s46, 169948; GFX9-NEXT:    s_lshr_b32 s68, s46, 89949; GFX9-NEXT:    s_lshr_b32 s69, s57, 249950; GFX9-NEXT:    s_lshr_b32 s70, s57, 169951; GFX9-NEXT:    s_lshr_b32 s71, s57, 89952; GFX9-NEXT:    s_lshr_b32 s80, s56, 169953; GFX9-NEXT:    s_lshr_b32 s26, s56, 89954; GFX9-NEXT:    v_writelane_b32 v22, s29, 19955; GFX9-NEXT:    s_lshr_b64 s[28:29], s[6:7], 249956; GFX9-NEXT:    s_lshr_b64 s[58:59], s[8:9], 249957; GFX9-NEXT:    s_lshr_b64 s[60:61], s[10:11], 249958; GFX9-NEXT:    s_lshr_b64 s[62:63], s[12:13], 249959; GFX9-NEXT:    s_lshr_b64 s[72:73], s[14:15], 249960; GFX9-NEXT:    s_lshr_b64 s[74:75], s[16:17], 249961; GFX9-NEXT:    s_lshr_b64 s[76:77], s[18:19], 249962; GFX9-NEXT:    s_lshr_b64 s[78:79], s[20:21], 249963; GFX9-NEXT:    s_lshr_b64 s[88:89], s[22:23], 249964; GFX9-NEXT:    s_lshr_b64 s[90:91], s[24:25], 249965; GFX9-NEXT:    s_lshr_b64 s[92:93], s[40:41], 249966; GFX9-NEXT:    s_lshr_b64 s[94:95], s[42:43], 249967; GFX9-NEXT:    s_lshr_b64 s[30:31], s[44:45], 249968; GFX9-NEXT:    s_lshr_b64 s[34:35], s[46:47], 249969; GFX9-NEXT:    s_lshr_b64 s[36:37], s[56:57], 249970; GFX9-NEXT:    s_cbranch_execnz .LBB13_39971; GFX9-NEXT:  .LBB13_2: ; %cmp.true9972; GFX9-NEXT:    s_add_i32 s5, s5, 39973; GFX9-NEXT:    s_lshr_b32 s26, s5, 249974; GFX9-NEXT:    v_writelane_b32 v22, s26, 29975; GFX9-NEXT:    s_lshr_b32 s26, s5, 169976; GFX9-NEXT:    s_add_i32 s4, s4, 39977; GFX9-NEXT:    v_writelane_b32 v22, s26, 39978; GFX9-NEXT:    s_lshr_b32 s26, s5, 89979; GFX9-NEXT:    v_writelane_b32 v22, s26, 49980; GFX9-NEXT:    s_lshr_b32 s26, s4, 169981; GFX9-NEXT:    s_add_i32 s7, s7, 39982; GFX9-NEXT:    v_writelane_b32 v22, s26, 59983; GFX9-NEXT:    s_lshr_b32 s26, s4, 89984; GFX9-NEXT:    v_writelane_b32 v22, s26, 69985; GFX9-NEXT:    s_lshr_b32 s26, s7, 249986; GFX9-NEXT:    v_writelane_b32 v22, s26, 79987; GFX9-NEXT:    s_lshr_b32 s26, s7, 169988; GFX9-NEXT:    s_add_i32 s6, s6, 39989; GFX9-NEXT:    v_writelane_b32 v22, s26, 89990; GFX9-NEXT:    s_lshr_b32 s26, s7, 89991; GFX9-NEXT:    v_writelane_b32 v22, s26, 99992; GFX9-NEXT:    s_lshr_b32 s26, s6, 169993; GFX9-NEXT:    s_add_i32 s9, s9, 39994; GFX9-NEXT:    v_writelane_b32 v22, s26, 109995; GFX9-NEXT:    s_lshr_b32 s26, s6, 89996; GFX9-NEXT:    v_writelane_b32 v22, s26, 119997; GFX9-NEXT:    s_lshr_b32 s26, s9, 249998; GFX9-NEXT:    v_writelane_b32 v22, s26, 129999; GFX9-NEXT:    s_lshr_b32 s26, s9, 1610000; GFX9-NEXT:    s_add_i32 s8, s8, 310001; GFX9-NEXT:    v_writelane_b32 v22, s26, 1310002; GFX9-NEXT:    s_lshr_b32 s26, s9, 810003; GFX9-NEXT:    v_writelane_b32 v22, s26, 1410004; GFX9-NEXT:    s_lshr_b32 s26, s8, 1610005; GFX9-NEXT:    s_add_i32 s11, s11, 310006; GFX9-NEXT:    v_writelane_b32 v22, s26, 1510007; GFX9-NEXT:    s_lshr_b32 s26, s8, 810008; GFX9-NEXT:    v_writelane_b32 v22, s26, 1610009; GFX9-NEXT:    s_lshr_b32 s26, s11, 2410010; GFX9-NEXT:    v_writelane_b32 v22, s26, 1710011; GFX9-NEXT:    s_lshr_b32 s26, s11, 1610012; GFX9-NEXT:    s_add_i32 s10, s10, 310013; GFX9-NEXT:    v_writelane_b32 v22, s26, 1810014; GFX9-NEXT:    s_lshr_b32 s26, s11, 810015; GFX9-NEXT:    v_writelane_b32 v22, s26, 1910016; GFX9-NEXT:    s_lshr_b32 s26, s10, 1610017; GFX9-NEXT:    s_add_i32 s13, s13, 310018; GFX9-NEXT:    v_writelane_b32 v22, s26, 2010019; GFX9-NEXT:    s_lshr_b32 s26, s10, 810020; GFX9-NEXT:    v_writelane_b32 v22, s26, 2110021; GFX9-NEXT:    s_lshr_b32 s26, s13, 2410022; GFX9-NEXT:    v_writelane_b32 v22, s26, 2210023; GFX9-NEXT:    s_lshr_b32 s26, s13, 1610024; GFX9-NEXT:    s_add_i32 s12, s12, 310025; GFX9-NEXT:    v_writelane_b32 v22, s26, 2310026; GFX9-NEXT:    s_lshr_b32 s26, s13, 810027; GFX9-NEXT:    v_writelane_b32 v22, s26, 2410028; GFX9-NEXT:    s_lshr_b32 s26, s12, 1610029; GFX9-NEXT:    s_add_i32 s15, s15, 310030; GFX9-NEXT:    v_writelane_b32 v22, s26, 2510031; GFX9-NEXT:    s_lshr_b32 s26, s12, 810032; GFX9-NEXT:    v_writelane_b32 v22, s26, 2610033; GFX9-NEXT:    s_lshr_b32 s26, s15, 2410034; GFX9-NEXT:    v_writelane_b32 v22, s26, 2710035; GFX9-NEXT:    s_lshr_b32 s26, s15, 1610036; GFX9-NEXT:    s_add_i32 s14, s14, 310037; GFX9-NEXT:    v_writelane_b32 v22, s26, 2810038; GFX9-NEXT:    s_lshr_b32 s26, s15, 810039; GFX9-NEXT:    v_writelane_b32 v22, s26, 2910040; GFX9-NEXT:    s_lshr_b32 s26, s14, 1610041; GFX9-NEXT:    s_add_i32 s17, s17, 310042; GFX9-NEXT:    v_writelane_b32 v22, s26, 3010043; GFX9-NEXT:    s_lshr_b32 s26, s14, 810044; GFX9-NEXT:    v_writelane_b32 v22, s26, 3110045; GFX9-NEXT:    s_lshr_b32 s26, s17, 2410046; GFX9-NEXT:    v_writelane_b32 v22, s26, 3210047; GFX9-NEXT:    s_lshr_b32 s26, s17, 1610048; GFX9-NEXT:    s_add_i32 s16, s16, 310049; GFX9-NEXT:    v_writelane_b32 v22, s26, 3310050; GFX9-NEXT:    s_lshr_b32 s26, s17, 810051; GFX9-NEXT:    v_writelane_b32 v22, s26, 3410052; GFX9-NEXT:    s_lshr_b32 s26, s16, 1610053; GFX9-NEXT:    s_add_i32 s19, s19, 310054; GFX9-NEXT:    v_writelane_b32 v22, s26, 3510055; GFX9-NEXT:    s_lshr_b32 s26, s16, 810056; GFX9-NEXT:    v_writelane_b32 v22, s26, 3610057; GFX9-NEXT:    s_lshr_b32 s26, s19, 2410058; GFX9-NEXT:    v_writelane_b32 v22, s26, 3710059; GFX9-NEXT:    s_lshr_b32 s26, s19, 1610060; GFX9-NEXT:    s_add_i32 s18, s18, 310061; GFX9-NEXT:    v_writelane_b32 v22, s26, 3810062; GFX9-NEXT:    s_lshr_b32 s26, s19, 810063; GFX9-NEXT:    v_writelane_b32 v22, s26, 3910064; GFX9-NEXT:    s_lshr_b32 s26, s18, 1610065; GFX9-NEXT:    s_add_i32 s21, s21, 310066; GFX9-NEXT:    v_writelane_b32 v22, s26, 4010067; GFX9-NEXT:    s_lshr_b32 s26, s18, 810068; GFX9-NEXT:    v_writelane_b32 v22, s26, 4110069; GFX9-NEXT:    s_lshr_b32 s26, s21, 2410070; GFX9-NEXT:    v_writelane_b32 v22, s26, 4210071; GFX9-NEXT:    s_lshr_b32 s26, s21, 1610072; GFX9-NEXT:    s_add_i32 s20, s20, 310073; GFX9-NEXT:    v_writelane_b32 v22, s26, 4310074; GFX9-NEXT:    s_lshr_b32 s26, s21, 810075; GFX9-NEXT:    v_writelane_b32 v22, s26, 4410076; GFX9-NEXT:    s_lshr_b32 s26, s20, 1610077; GFX9-NEXT:    s_add_i32 s23, s23, 310078; GFX9-NEXT:    v_writelane_b32 v22, s26, 4510079; GFX9-NEXT:    s_lshr_b32 s26, s20, 810080; GFX9-NEXT:    v_writelane_b32 v22, s26, 4610081; GFX9-NEXT:    s_lshr_b32 s26, s23, 2410082; GFX9-NEXT:    v_writelane_b32 v22, s26, 4710083; GFX9-NEXT:    s_lshr_b32 s26, s23, 1610084; GFX9-NEXT:    s_add_i32 s22, s22, 310085; GFX9-NEXT:    v_writelane_b32 v22, s26, 4810086; GFX9-NEXT:    s_lshr_b32 s26, s23, 810087; GFX9-NEXT:    v_writelane_b32 v22, s26, 4910088; GFX9-NEXT:    s_lshr_b32 s26, s22, 1610089; GFX9-NEXT:    v_writelane_b32 v22, s26, 5010090; GFX9-NEXT:    s_lshr_b64 s[28:29], s[4:5], 2410091; GFX9-NEXT:    s_add_i32 s57, s57, 310092; GFX9-NEXT:    s_add_i32 s56, s56, 310093; GFX9-NEXT:    s_add_i32 s47, s47, 310094; GFX9-NEXT:    s_add_i32 s46, s46, 310095; GFX9-NEXT:    s_add_i32 s45, s45, 310096; GFX9-NEXT:    s_add_i32 s44, s44, 310097; GFX9-NEXT:    s_add_i32 s43, s43, 310098; GFX9-NEXT:    s_add_i32 s42, s42, 310099; GFX9-NEXT:    s_add_i32 s41, s41, 310100; GFX9-NEXT:    s_add_i32 s40, s40, 310101; GFX9-NEXT:    s_add_i32 s25, s25, 310102; GFX9-NEXT:    s_add_i32 s24, s24, 310103; GFX9-NEXT:    v_writelane_b32 v22, s28, 010104; GFX9-NEXT:    s_lshr_b32 s82, s22, 810105; GFX9-NEXT:    s_lshr_b32 s83, s25, 2410106; GFX9-NEXT:    s_lshr_b32 s81, s25, 1610107; GFX9-NEXT:    s_lshr_b32 s84, s25, 810108; GFX9-NEXT:    s_lshr_b32 s85, s24, 1610109; GFX9-NEXT:    s_lshr_b32 s86, s24, 810110; GFX9-NEXT:    s_lshr_b32 s87, s41, 2410111; GFX9-NEXT:    s_lshr_b32 s96, s41, 1610112; GFX9-NEXT:    s_lshr_b32 s97, s41, 810113; GFX9-NEXT:    s_lshr_b32 s98, s40, 1610114; GFX9-NEXT:    s_lshr_b32 s99, s40, 810115; GFX9-NEXT:    s_lshr_b32 s38, s43, 2410116; GFX9-NEXT:    s_lshr_b32 s39, s43, 1610117; GFX9-NEXT:    s_lshr_b32 s48, s43, 810118; GFX9-NEXT:    s_lshr_b32 s49, s42, 1610119; GFX9-NEXT:    s_lshr_b32 s50, s42, 810120; GFX9-NEXT:    s_lshr_b32 s51, s45, 2410121; GFX9-NEXT:    s_lshr_b32 s52, s45, 1610122; GFX9-NEXT:    s_lshr_b32 s53, s45, 810123; GFX9-NEXT:    s_lshr_b32 s54, s44, 1610124; GFX9-NEXT:    s_lshr_b32 s55, s44, 810125; GFX9-NEXT:    s_lshr_b32 s64, s47, 2410126; GFX9-NEXT:    s_lshr_b32 s65, s47, 1610127; GFX9-NEXT:    s_lshr_b32 s66, s47, 810128; GFX9-NEXT:    s_lshr_b32 s67, s46, 1610129; GFX9-NEXT:    s_lshr_b32 s68, s46, 810130; GFX9-NEXT:    s_lshr_b32 s69, s57, 2410131; GFX9-NEXT:    s_lshr_b32 s70, s57, 1610132; GFX9-NEXT:    s_lshr_b32 s71, s57, 810133; GFX9-NEXT:    s_lshr_b32 s80, s56, 1610134; GFX9-NEXT:    s_lshr_b32 s26, s56, 810135; GFX9-NEXT:    v_writelane_b32 v22, s29, 110136; GFX9-NEXT:    s_lshr_b64 s[28:29], s[6:7], 2410137; GFX9-NEXT:    s_lshr_b64 s[58:59], s[8:9], 2410138; GFX9-NEXT:    s_lshr_b64 s[60:61], s[10:11], 2410139; GFX9-NEXT:    s_lshr_b64 s[62:63], s[12:13], 2410140; GFX9-NEXT:    s_lshr_b64 s[72:73], s[14:15], 2410141; GFX9-NEXT:    s_lshr_b64 s[74:75], s[16:17], 2410142; GFX9-NEXT:    s_lshr_b64 s[76:77], s[18:19], 2410143; GFX9-NEXT:    s_lshr_b64 s[78:79], s[20:21], 2410144; GFX9-NEXT:    s_lshr_b64 s[88:89], s[22:23], 2410145; GFX9-NEXT:    s_lshr_b64 s[90:91], s[24:25], 2410146; GFX9-NEXT:    s_lshr_b64 s[92:93], s[40:41], 2410147; GFX9-NEXT:    s_lshr_b64 s[94:95], s[42:43], 2410148; GFX9-NEXT:    s_lshr_b64 s[30:31], s[44:45], 2410149; GFX9-NEXT:    s_lshr_b64 s[34:35], s[46:47], 2410150; GFX9-NEXT:    s_lshr_b64 s[36:37], s[56:57], 2410151; GFX9-NEXT:  .LBB13_3: ; %end10152; GFX9-NEXT:    s_lshl_b32 s26, s26, 810153; GFX9-NEXT:    s_and_b32 s27, s56, 0xff10154; GFX9-NEXT:    s_or_b32 s26, s27, s2610155; GFX9-NEXT:    s_lshl_b32 s27, s36, 810156; GFX9-NEXT:    s_and_b32 s29, s80, 0xff10157; GFX9-NEXT:    s_or_b32 s27, s29, s2710158; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10159; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610160; GFX9-NEXT:    s_or_b32 s26, s26, s2710161; GFX9-NEXT:    v_mov_b32_e32 v1, s2610162; GFX9-NEXT:    s_and_b32 s26, s57, 0xff10163; GFX9-NEXT:    s_lshl_b32 s27, s71, 810164; GFX9-NEXT:    s_or_b32 s26, s26, s2710165; GFX9-NEXT:    s_and_b32 s27, s70, 0xff10166; GFX9-NEXT:    s_lshl_b32 s29, s69, 810167; GFX9-NEXT:    s_or_b32 s27, s27, s2910168; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10169; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610170; GFX9-NEXT:    s_or_b32 s26, s26, s2710171; GFX9-NEXT:    v_mov_b32_e32 v2, s2610172; GFX9-NEXT:    s_lshl_b32 s26, s68, 810173; GFX9-NEXT:    s_and_b32 s27, s46, 0xff10174; GFX9-NEXT:    s_or_b32 s26, s27, s2610175; GFX9-NEXT:    s_lshl_b32 s27, s34, 810176; GFX9-NEXT:    s_and_b32 s29, s67, 0xff10177; GFX9-NEXT:    s_or_b32 s27, s29, s2710178; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10179; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610180; GFX9-NEXT:    s_or_b32 s26, s26, s2710181; GFX9-NEXT:    v_mov_b32_e32 v3, s2610182; GFX9-NEXT:    s_and_b32 s26, s47, 0xff10183; GFX9-NEXT:    s_lshl_b32 s27, s66, 810184; GFX9-NEXT:    s_or_b32 s26, s26, s2710185; GFX9-NEXT:    s_and_b32 s27, s65, 0xff10186; GFX9-NEXT:    s_lshl_b32 s29, s64, 810187; GFX9-NEXT:    s_or_b32 s27, s27, s2910188; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10189; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610190; GFX9-NEXT:    s_or_b32 s26, s26, s2710191; GFX9-NEXT:    v_mov_b32_e32 v4, s2610192; GFX9-NEXT:    s_lshl_b32 s26, s55, 810193; GFX9-NEXT:    s_and_b32 s27, s44, 0xff10194; GFX9-NEXT:    s_or_b32 s26, s27, s2610195; GFX9-NEXT:    s_lshl_b32 s27, s30, 810196; GFX9-NEXT:    s_and_b32 s29, s54, 0xff10197; GFX9-NEXT:    s_or_b32 s27, s29, s2710198; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10199; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610200; GFX9-NEXT:    s_or_b32 s26, s26, s2710201; GFX9-NEXT:    v_mov_b32_e32 v5, s2610202; GFX9-NEXT:    s_and_b32 s26, s45, 0xff10203; GFX9-NEXT:    s_lshl_b32 s27, s53, 810204; GFX9-NEXT:    s_or_b32 s26, s26, s2710205; GFX9-NEXT:    s_and_b32 s27, s52, 0xff10206; GFX9-NEXT:    s_lshl_b32 s29, s51, 810207; GFX9-NEXT:    s_or_b32 s27, s27, s2910208; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10209; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610210; GFX9-NEXT:    s_or_b32 s26, s26, s2710211; GFX9-NEXT:    v_mov_b32_e32 v6, s2610212; GFX9-NEXT:    s_lshl_b32 s26, s50, 810213; GFX9-NEXT:    s_and_b32 s27, s42, 0xff10214; GFX9-NEXT:    s_or_b32 s26, s27, s2610215; GFX9-NEXT:    s_lshl_b32 s27, s94, 810216; GFX9-NEXT:    s_and_b32 s29, s49, 0xff10217; GFX9-NEXT:    s_or_b32 s27, s29, s2710218; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10219; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610220; GFX9-NEXT:    s_or_b32 s26, s26, s2710221; GFX9-NEXT:    v_mov_b32_e32 v7, s2610222; GFX9-NEXT:    s_and_b32 s26, s43, 0xff10223; GFX9-NEXT:    s_lshl_b32 s27, s48, 810224; GFX9-NEXT:    s_or_b32 s26, s26, s2710225; GFX9-NEXT:    s_and_b32 s27, s39, 0xff10226; GFX9-NEXT:    s_lshl_b32 s29, s38, 810227; GFX9-NEXT:    s_or_b32 s27, s27, s2910228; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10229; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610230; GFX9-NEXT:    s_or_b32 s26, s26, s2710231; GFX9-NEXT:    v_mov_b32_e32 v8, s2610232; GFX9-NEXT:    s_lshl_b32 s26, s99, 810233; GFX9-NEXT:    s_and_b32 s27, s40, 0xff10234; GFX9-NEXT:    s_or_b32 s26, s27, s2610235; GFX9-NEXT:    s_lshl_b32 s27, s92, 810236; GFX9-NEXT:    s_and_b32 s29, s98, 0xff10237; GFX9-NEXT:    s_or_b32 s27, s29, s2710238; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10239; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610240; GFX9-NEXT:    s_or_b32 s26, s26, s2710241; GFX9-NEXT:    v_mov_b32_e32 v9, s2610242; GFX9-NEXT:    s_and_b32 s26, s41, 0xff10243; GFX9-NEXT:    s_lshl_b32 s27, s97, 810244; GFX9-NEXT:    s_or_b32 s26, s26, s2710245; GFX9-NEXT:    s_and_b32 s27, s96, 0xff10246; GFX9-NEXT:    s_lshl_b32 s29, s87, 810247; GFX9-NEXT:    s_or_b32 s27, s27, s2910248; GFX9-NEXT:    s_and_b32 s26, s26, 0xffff10249; GFX9-NEXT:    s_lshl_b32 s27, s27, 1610250; GFX9-NEXT:    s_or_b32 s26, s26, s2710251; GFX9-NEXT:    v_mov_b32_e32 v10, s2610252; GFX9-NEXT:    s_lshl_b32 s26, s86, 810253; GFX9-NEXT:    s_and_b32 s24, s24, 0xff10254; GFX9-NEXT:    s_or_b32 s24, s24, s2610255; GFX9-NEXT:    s_lshl_b32 s26, s90, 810256; GFX9-NEXT:    s_and_b32 s27, s85, 0xff10257; GFX9-NEXT:    s_or_b32 s26, s27, s2610258; GFX9-NEXT:    s_and_b32 s24, s24, 0xffff10259; GFX9-NEXT:    s_lshl_b32 s26, s26, 1610260; GFX9-NEXT:    s_or_b32 s24, s24, s2610261; GFX9-NEXT:    v_mov_b32_e32 v11, s2410262; GFX9-NEXT:    s_and_b32 s24, s25, 0xff10263; GFX9-NEXT:    s_lshl_b32 s25, s84, 810264; GFX9-NEXT:    s_or_b32 s24, s24, s2510265; GFX9-NEXT:    s_and_b32 s25, s81, 0xff10266; GFX9-NEXT:    s_lshl_b32 s26, s83, 810267; GFX9-NEXT:    s_or_b32 s25, s25, s2610268; GFX9-NEXT:    s_and_b32 s24, s24, 0xffff10269; GFX9-NEXT:    s_lshl_b32 s25, s25, 1610270; GFX9-NEXT:    s_or_b32 s24, s24, s2510271; GFX9-NEXT:    v_mov_b32_e32 v12, s2410272; GFX9-NEXT:    s_lshl_b32 s24, s82, 810273; GFX9-NEXT:    s_and_b32 s22, s22, 0xff10274; GFX9-NEXT:    v_readlane_b32 s25, v22, 5010275; GFX9-NEXT:    s_or_b32 s22, s22, s2410276; GFX9-NEXT:    s_lshl_b32 s24, s88, 810277; GFX9-NEXT:    s_and_b32 s25, s25, 0xff10278; GFX9-NEXT:    s_or_b32 s24, s25, s2410279; GFX9-NEXT:    s_and_b32 s22, s22, 0xffff10280; GFX9-NEXT:    s_lshl_b32 s24, s24, 1610281; GFX9-NEXT:    s_or_b32 s22, s22, s2410282; GFX9-NEXT:    v_mov_b32_e32 v13, s2210283; GFX9-NEXT:    s_and_b32 s22, s23, 0xff10284; GFX9-NEXT:    v_readlane_b32 s23, v22, 4910285; GFX9-NEXT:    s_lshl_b32 s23, s23, 810286; GFX9-NEXT:    s_or_b32 s22, s22, s2310287; GFX9-NEXT:    v_readlane_b32 s23, v22, 4810288; GFX9-NEXT:    v_readlane_b32 s24, v22, 4710289; GFX9-NEXT:    s_and_b32 s23, s23, 0xff10290; GFX9-NEXT:    s_lshl_b32 s24, s24, 810291; GFX9-NEXT:    s_or_b32 s23, s23, s2410292; GFX9-NEXT:    s_and_b32 s22, s22, 0xffff10293; GFX9-NEXT:    s_lshl_b32 s23, s23, 1610294; GFX9-NEXT:    s_or_b32 s22, s22, s2310295; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen10296; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:410297; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:810298; GFX9-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:1210299; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:1610300; GFX9-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:2010301; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:2410302; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:2810303; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:3210304; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:3610305; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:4010306; GFX9-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:4410307; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:4810308; GFX9-NEXT:    v_mov_b32_e32 v1, s2210309; GFX9-NEXT:    v_readlane_b32 s22, v22, 4610310; GFX9-NEXT:    s_and_b32 s20, s20, 0xff10311; GFX9-NEXT:    s_lshl_b32 s22, s22, 810312; GFX9-NEXT:    s_or_b32 s20, s20, s2210313; GFX9-NEXT:    v_readlane_b32 s22, v22, 4510314; GFX9-NEXT:    s_and_b32 s22, s22, 0xff10315; GFX9-NEXT:    s_lshl_b32 s23, s78, 810316; GFX9-NEXT:    s_or_b32 s22, s22, s2310317; GFX9-NEXT:    s_and_b32 s20, s20, 0xffff10318; GFX9-NEXT:    s_lshl_b32 s22, s22, 1610319; GFX9-NEXT:    s_or_b32 s20, s20, s2210320; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:5210321; GFX9-NEXT:    v_mov_b32_e32 v1, s2010322; GFX9-NEXT:    s_and_b32 s20, s21, 0xff10323; GFX9-NEXT:    v_readlane_b32 s21, v22, 4410324; GFX9-NEXT:    s_lshl_b32 s21, s21, 810325; GFX9-NEXT:    s_or_b32 s20, s20, s2110326; GFX9-NEXT:    v_readlane_b32 s21, v22, 4310327; GFX9-NEXT:    v_readlane_b32 s22, v22, 4210328; GFX9-NEXT:    s_and_b32 s21, s21, 0xff10329; GFX9-NEXT:    s_lshl_b32 s22, s22, 810330; GFX9-NEXT:    s_or_b32 s21, s21, s2210331; GFX9-NEXT:    s_and_b32 s20, s20, 0xffff10332; GFX9-NEXT:    s_lshl_b32 s21, s21, 1610333; GFX9-NEXT:    s_or_b32 s20, s20, s2110334; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:5610335; GFX9-NEXT:    v_mov_b32_e32 v1, s2010336; GFX9-NEXT:    v_readlane_b32 s20, v22, 4110337; GFX9-NEXT:    s_and_b32 s18, s18, 0xff10338; GFX9-NEXT:    s_lshl_b32 s20, s20, 810339; GFX9-NEXT:    s_or_b32 s18, s18, s2010340; GFX9-NEXT:    v_readlane_b32 s20, v22, 4010341; GFX9-NEXT:    s_and_b32 s20, s20, 0xff10342; GFX9-NEXT:    s_lshl_b32 s21, s76, 810343; GFX9-NEXT:    s_or_b32 s20, s20, s2110344; GFX9-NEXT:    s_and_b32 s18, s18, 0xffff10345; GFX9-NEXT:    s_lshl_b32 s20, s20, 1610346; GFX9-NEXT:    s_or_b32 s18, s18, s2010347; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:6010348; GFX9-NEXT:    v_mov_b32_e32 v1, s1810349; GFX9-NEXT:    s_and_b32 s18, s19, 0xff10350; GFX9-NEXT:    v_readlane_b32 s19, v22, 3910351; GFX9-NEXT:    s_lshl_b32 s19, s19, 810352; GFX9-NEXT:    s_or_b32 s18, s18, s1910353; GFX9-NEXT:    v_readlane_b32 s19, v22, 3810354; GFX9-NEXT:    v_readlane_b32 s20, v22, 3710355; GFX9-NEXT:    s_and_b32 s19, s19, 0xff10356; GFX9-NEXT:    s_lshl_b32 s20, s20, 810357; GFX9-NEXT:    s_or_b32 s19, s19, s2010358; GFX9-NEXT:    s_and_b32 s18, s18, 0xffff10359; GFX9-NEXT:    s_lshl_b32 s19, s19, 1610360; GFX9-NEXT:    s_or_b32 s18, s18, s1910361; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:6410362; GFX9-NEXT:    v_mov_b32_e32 v1, s1810363; GFX9-NEXT:    v_readlane_b32 s18, v22, 3610364; GFX9-NEXT:    s_and_b32 s16, s16, 0xff10365; GFX9-NEXT:    s_lshl_b32 s18, s18, 810366; GFX9-NEXT:    s_or_b32 s16, s16, s1810367; GFX9-NEXT:    v_readlane_b32 s18, v22, 3510368; GFX9-NEXT:    s_and_b32 s18, s18, 0xff10369; GFX9-NEXT:    s_lshl_b32 s19, s74, 810370; GFX9-NEXT:    s_or_b32 s18, s18, s1910371; GFX9-NEXT:    s_and_b32 s16, s16, 0xffff10372; GFX9-NEXT:    s_lshl_b32 s18, s18, 1610373; GFX9-NEXT:    s_or_b32 s16, s16, s1810374; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:6810375; GFX9-NEXT:    v_mov_b32_e32 v1, s1610376; GFX9-NEXT:    s_and_b32 s16, s17, 0xff10377; GFX9-NEXT:    v_readlane_b32 s17, v22, 3410378; GFX9-NEXT:    s_lshl_b32 s17, s17, 810379; GFX9-NEXT:    s_or_b32 s16, s16, s1710380; GFX9-NEXT:    v_readlane_b32 s17, v22, 3310381; GFX9-NEXT:    v_readlane_b32 s18, v22, 3210382; GFX9-NEXT:    s_and_b32 s17, s17, 0xff10383; GFX9-NEXT:    s_lshl_b32 s18, s18, 810384; GFX9-NEXT:    s_or_b32 s17, s17, s1810385; GFX9-NEXT:    s_and_b32 s16, s16, 0xffff10386; GFX9-NEXT:    s_lshl_b32 s17, s17, 1610387; GFX9-NEXT:    s_or_b32 s16, s16, s1710388; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:7210389; GFX9-NEXT:    v_mov_b32_e32 v1, s1610390; GFX9-NEXT:    v_readlane_b32 s16, v22, 3110391; GFX9-NEXT:    s_and_b32 s14, s14, 0xff10392; GFX9-NEXT:    s_lshl_b32 s16, s16, 810393; GFX9-NEXT:    s_or_b32 s14, s14, s1610394; GFX9-NEXT:    v_readlane_b32 s16, v22, 3010395; GFX9-NEXT:    s_and_b32 s16, s16, 0xff10396; GFX9-NEXT:    s_lshl_b32 s17, s72, 810397; GFX9-NEXT:    s_or_b32 s16, s16, s1710398; GFX9-NEXT:    s_and_b32 s14, s14, 0xffff10399; GFX9-NEXT:    s_lshl_b32 s16, s16, 1610400; GFX9-NEXT:    s_or_b32 s14, s14, s1610401; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:7610402; GFX9-NEXT:    v_mov_b32_e32 v1, s1410403; GFX9-NEXT:    s_and_b32 s14, s15, 0xff10404; GFX9-NEXT:    v_readlane_b32 s15, v22, 2910405; GFX9-NEXT:    s_lshl_b32 s15, s15, 810406; GFX9-NEXT:    s_or_b32 s14, s14, s1510407; GFX9-NEXT:    v_readlane_b32 s15, v22, 2810408; GFX9-NEXT:    v_readlane_b32 s16, v22, 2710409; GFX9-NEXT:    s_and_b32 s15, s15, 0xff10410; GFX9-NEXT:    s_lshl_b32 s16, s16, 810411; GFX9-NEXT:    s_or_b32 s15, s15, s1610412; GFX9-NEXT:    s_and_b32 s14, s14, 0xffff10413; GFX9-NEXT:    s_lshl_b32 s15, s15, 1610414; GFX9-NEXT:    s_or_b32 s14, s14, s1510415; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:8010416; GFX9-NEXT:    v_mov_b32_e32 v1, s1410417; GFX9-NEXT:    v_readlane_b32 s14, v22, 2610418; GFX9-NEXT:    s_and_b32 s12, s12, 0xff10419; GFX9-NEXT:    s_lshl_b32 s14, s14, 810420; GFX9-NEXT:    s_or_b32 s12, s12, s1410421; GFX9-NEXT:    v_readlane_b32 s14, v22, 2510422; GFX9-NEXT:    s_and_b32 s14, s14, 0xff10423; GFX9-NEXT:    s_lshl_b32 s15, s62, 810424; GFX9-NEXT:    s_or_b32 s14, s14, s1510425; GFX9-NEXT:    s_and_b32 s12, s12, 0xffff10426; GFX9-NEXT:    s_lshl_b32 s14, s14, 1610427; GFX9-NEXT:    s_or_b32 s12, s12, s1410428; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:8410429; GFX9-NEXT:    v_mov_b32_e32 v1, s1210430; GFX9-NEXT:    s_and_b32 s12, s13, 0xff10431; GFX9-NEXT:    v_readlane_b32 s13, v22, 2410432; GFX9-NEXT:    s_lshl_b32 s13, s13, 810433; GFX9-NEXT:    s_or_b32 s12, s12, s1310434; GFX9-NEXT:    v_readlane_b32 s13, v22, 2310435; GFX9-NEXT:    v_readlane_b32 s14, v22, 2210436; GFX9-NEXT:    s_and_b32 s13, s13, 0xff10437; GFX9-NEXT:    s_lshl_b32 s14, s14, 810438; GFX9-NEXT:    s_or_b32 s13, s13, s1410439; GFX9-NEXT:    s_and_b32 s12, s12, 0xffff10440; GFX9-NEXT:    s_lshl_b32 s13, s13, 1610441; GFX9-NEXT:    s_or_b32 s12, s12, s1310442; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:8810443; GFX9-NEXT:    v_mov_b32_e32 v1, s1210444; GFX9-NEXT:    v_readlane_b32 s12, v22, 2110445; GFX9-NEXT:    s_and_b32 s10, s10, 0xff10446; GFX9-NEXT:    s_lshl_b32 s12, s12, 810447; GFX9-NEXT:    s_or_b32 s10, s10, s1210448; GFX9-NEXT:    v_readlane_b32 s12, v22, 2010449; GFX9-NEXT:    s_and_b32 s12, s12, 0xff10450; GFX9-NEXT:    s_lshl_b32 s13, s60, 810451; GFX9-NEXT:    s_or_b32 s12, s12, s1310452; GFX9-NEXT:    s_and_b32 s10, s10, 0xffff10453; GFX9-NEXT:    s_lshl_b32 s12, s12, 1610454; GFX9-NEXT:    s_or_b32 s10, s10, s1210455; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:9210456; GFX9-NEXT:    v_mov_b32_e32 v1, s1010457; GFX9-NEXT:    s_and_b32 s10, s11, 0xff10458; GFX9-NEXT:    v_readlane_b32 s11, v22, 1910459; GFX9-NEXT:    s_lshl_b32 s11, s11, 810460; GFX9-NEXT:    s_or_b32 s10, s10, s1110461; GFX9-NEXT:    v_readlane_b32 s11, v22, 1810462; GFX9-NEXT:    v_readlane_b32 s12, v22, 1710463; GFX9-NEXT:    s_and_b32 s11, s11, 0xff10464; GFX9-NEXT:    s_lshl_b32 s12, s12, 810465; GFX9-NEXT:    s_or_b32 s11, s11, s1210466; GFX9-NEXT:    s_and_b32 s10, s10, 0xffff10467; GFX9-NEXT:    s_lshl_b32 s11, s11, 1610468; GFX9-NEXT:    s_or_b32 s10, s10, s1110469; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:9610470; GFX9-NEXT:    v_mov_b32_e32 v1, s1010471; GFX9-NEXT:    v_readlane_b32 s10, v22, 1610472; GFX9-NEXT:    s_and_b32 s8, s8, 0xff10473; GFX9-NEXT:    s_lshl_b32 s10, s10, 810474; GFX9-NEXT:    s_or_b32 s8, s8, s1010475; GFX9-NEXT:    v_readlane_b32 s10, v22, 1510476; GFX9-NEXT:    s_and_b32 s10, s10, 0xff10477; GFX9-NEXT:    s_lshl_b32 s11, s58, 810478; GFX9-NEXT:    s_or_b32 s10, s10, s1110479; GFX9-NEXT:    s_and_b32 s8, s8, 0xffff10480; GFX9-NEXT:    s_lshl_b32 s10, s10, 1610481; GFX9-NEXT:    s_or_b32 s8, s8, s1010482; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:10010483; GFX9-NEXT:    v_mov_b32_e32 v1, s810484; GFX9-NEXT:    s_and_b32 s8, s9, 0xff10485; GFX9-NEXT:    v_readlane_b32 s9, v22, 1410486; GFX9-NEXT:    s_lshl_b32 s9, s9, 810487; GFX9-NEXT:    s_or_b32 s8, s8, s910488; GFX9-NEXT:    v_readlane_b32 s9, v22, 1310489; GFX9-NEXT:    v_readlane_b32 s10, v22, 1210490; GFX9-NEXT:    s_and_b32 s9, s9, 0xff10491; GFX9-NEXT:    s_lshl_b32 s10, s10, 810492; GFX9-NEXT:    s_or_b32 s9, s9, s1010493; GFX9-NEXT:    s_and_b32 s8, s8, 0xffff10494; GFX9-NEXT:    s_lshl_b32 s9, s9, 1610495; GFX9-NEXT:    s_or_b32 s8, s8, s910496; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:10410497; GFX9-NEXT:    v_mov_b32_e32 v1, s810498; GFX9-NEXT:    v_readlane_b32 s8, v22, 1110499; GFX9-NEXT:    s_and_b32 s6, s6, 0xff10500; GFX9-NEXT:    s_lshl_b32 s8, s8, 810501; GFX9-NEXT:    s_or_b32 s6, s6, s810502; GFX9-NEXT:    v_readlane_b32 s8, v22, 1010503; GFX9-NEXT:    s_and_b32 s8, s8, 0xff10504; GFX9-NEXT:    s_lshl_b32 s9, s28, 810505; GFX9-NEXT:    s_or_b32 s8, s8, s910506; GFX9-NEXT:    s_and_b32 s6, s6, 0xffff10507; GFX9-NEXT:    s_lshl_b32 s8, s8, 1610508; GFX9-NEXT:    s_or_b32 s6, s6, s810509; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:10810510; GFX9-NEXT:    v_mov_b32_e32 v1, s610511; GFX9-NEXT:    s_and_b32 s6, s7, 0xff10512; GFX9-NEXT:    v_readlane_b32 s7, v22, 910513; GFX9-NEXT:    s_lshl_b32 s7, s7, 810514; GFX9-NEXT:    s_or_b32 s6, s6, s710515; GFX9-NEXT:    v_readlane_b32 s7, v22, 810516; GFX9-NEXT:    v_readlane_b32 s8, v22, 710517; GFX9-NEXT:    s_and_b32 s7, s7, 0xff10518; GFX9-NEXT:    s_lshl_b32 s8, s8, 810519; GFX9-NEXT:    s_or_b32 s7, s7, s810520; GFX9-NEXT:    s_and_b32 s6, s6, 0xffff10521; GFX9-NEXT:    s_lshl_b32 s7, s7, 1610522; GFX9-NEXT:    s_or_b32 s6, s6, s710523; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:11210524; GFX9-NEXT:    v_mov_b32_e32 v1, s610525; GFX9-NEXT:    v_readlane_b32 s6, v22, 610526; GFX9-NEXT:    s_and_b32 s4, s4, 0xff10527; GFX9-NEXT:    s_lshl_b32 s6, s6, 810528; GFX9-NEXT:    s_or_b32 s4, s4, s610529; GFX9-NEXT:    v_readlane_b32 s6, v22, 510530; GFX9-NEXT:    v_readlane_b32 s8, v22, 010531; GFX9-NEXT:    s_and_b32 s6, s6, 0xff10532; GFX9-NEXT:    s_lshl_b32 s7, s8, 810533; GFX9-NEXT:    s_or_b32 s6, s6, s710534; GFX9-NEXT:    s_and_b32 s4, s4, 0xffff10535; GFX9-NEXT:    s_lshl_b32 s6, s6, 1610536; GFX9-NEXT:    s_or_b32 s4, s4, s610537; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:11610538; GFX9-NEXT:    v_mov_b32_e32 v1, s410539; GFX9-NEXT:    s_and_b32 s4, s5, 0xff10540; GFX9-NEXT:    v_readlane_b32 s5, v22, 410541; GFX9-NEXT:    s_lshl_b32 s5, s5, 810542; GFX9-NEXT:    s_or_b32 s4, s4, s510543; GFX9-NEXT:    v_readlane_b32 s5, v22, 310544; GFX9-NEXT:    v_readlane_b32 s6, v22, 210545; GFX9-NEXT:    s_and_b32 s5, s5, 0xff10546; GFX9-NEXT:    s_lshl_b32 s6, s6, 810547; GFX9-NEXT:    s_or_b32 s5, s5, s610548; GFX9-NEXT:    s_and_b32 s4, s4, 0xffff10549; GFX9-NEXT:    s_lshl_b32 s5, s5, 1610550; GFX9-NEXT:    s_or_b32 s4, s4, s510551; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:12010552; GFX9-NEXT:    v_mov_b32_e32 v1, s410553; GFX9-NEXT:    v_readlane_b32 s9, v22, 110554; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:12410555; GFX9-NEXT:    v_readlane_b32 s99, v21, 3510556; GFX9-NEXT:    v_readlane_b32 s98, v21, 3410557; GFX9-NEXT:    v_readlane_b32 s97, v21, 3310558; GFX9-NEXT:    v_readlane_b32 s96, v21, 3210559; GFX9-NEXT:    v_readlane_b32 s87, v21, 3110560; GFX9-NEXT:    v_readlane_b32 s86, v21, 3010561; GFX9-NEXT:    v_readlane_b32 s85, v21, 2910562; GFX9-NEXT:    v_readlane_b32 s84, v21, 2810563; GFX9-NEXT:    v_readlane_b32 s83, v21, 2710564; GFX9-NEXT:    v_readlane_b32 s82, v21, 2610565; GFX9-NEXT:    v_readlane_b32 s81, v21, 2510566; GFX9-NEXT:    v_readlane_b32 s80, v21, 2410567; GFX9-NEXT:    v_readlane_b32 s71, v21, 2310568; GFX9-NEXT:    v_readlane_b32 s70, v21, 2210569; GFX9-NEXT:    v_readlane_b32 s69, v21, 2110570; GFX9-NEXT:    v_readlane_b32 s68, v21, 2010571; GFX9-NEXT:    v_readlane_b32 s67, v21, 1910572; GFX9-NEXT:    v_readlane_b32 s66, v21, 1810573; GFX9-NEXT:    v_readlane_b32 s65, v21, 1710574; GFX9-NEXT:    v_readlane_b32 s64, v21, 1610575; GFX9-NEXT:    v_readlane_b32 s55, v21, 1510576; GFX9-NEXT:    v_readlane_b32 s54, v21, 1410577; GFX9-NEXT:    v_readlane_b32 s53, v21, 1310578; GFX9-NEXT:    v_readlane_b32 s52, v21, 1210579; GFX9-NEXT:    v_readlane_b32 s51, v21, 1110580; GFX9-NEXT:    v_readlane_b32 s50, v21, 1010581; GFX9-NEXT:    v_readlane_b32 s49, v21, 910582; GFX9-NEXT:    v_readlane_b32 s48, v21, 810583; GFX9-NEXT:    v_readlane_b32 s39, v21, 710584; GFX9-NEXT:    v_readlane_b32 s38, v21, 610585; GFX9-NEXT:    v_readlane_b32 s37, v21, 510586; GFX9-NEXT:    v_readlane_b32 s36, v21, 410587; GFX9-NEXT:    v_readlane_b32 s35, v21, 310588; GFX9-NEXT:    v_readlane_b32 s34, v21, 210589; GFX9-NEXT:    v_readlane_b32 s31, v21, 110590; GFX9-NEXT:    v_readlane_b32 s30, v21, 010591; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -110592; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 ; 4-byte Folded Reload10593; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload10594; GFX9-NEXT:    s_mov_b64 exec, s[4:5]10595; GFX9-NEXT:    s_waitcnt vmcnt(0)10596; GFX9-NEXT:    s_setpc_b64 s[30:31]10597; GFX9-NEXT:  .LBB13_4:10598; GFX9-NEXT:    ; implicit-def: $sgpr2710599; GFX9-NEXT:    ; kill: killed $sgpr2710600; GFX9-NEXT:    ; implicit-def: $sgpr8310601; GFX9-NEXT:    ; implicit-def: $sgpr8210602; GFX9-NEXT:    ; implicit-def: $sgpr2710603; GFX9-NEXT:    ; kill: killed $sgpr2710604; GFX9-NEXT:    v_writelane_b32 v22, s82, 010605; GFX9-NEXT:    ; implicit-def: $sgpr2710606; GFX9-NEXT:    ; kill: killed $sgpr2710607; GFX9-NEXT:    ; implicit-def: $sgpr2610608; GFX9-NEXT:    ; implicit-def: $sgpr8010609; GFX9-NEXT:    ; implicit-def: $sgpr7110610; GFX9-NEXT:    ; implicit-def: $sgpr7010611; GFX9-NEXT:    ; implicit-def: $sgpr6910612; GFX9-NEXT:    ; implicit-def: $sgpr6810613; GFX9-NEXT:    ; implicit-def: $sgpr6710614; GFX9-NEXT:    ; implicit-def: $sgpr6610615; GFX9-NEXT:    ; implicit-def: $sgpr6510616; GFX9-NEXT:    ; implicit-def: $sgpr6410617; GFX9-NEXT:    ; implicit-def: $sgpr5510618; GFX9-NEXT:    ; implicit-def: $sgpr5410619; GFX9-NEXT:    ; implicit-def: $sgpr5310620; GFX9-NEXT:    ; implicit-def: $sgpr5210621; GFX9-NEXT:    ; implicit-def: $sgpr5110622; GFX9-NEXT:    ; implicit-def: $sgpr5010623; GFX9-NEXT:    ; implicit-def: $sgpr4910624; GFX9-NEXT:    ; implicit-def: $sgpr4810625; GFX9-NEXT:    ; implicit-def: $sgpr3910626; GFX9-NEXT:    ; implicit-def: $sgpr3810627; GFX9-NEXT:    ; implicit-def: $sgpr9910628; GFX9-NEXT:    ; implicit-def: $sgpr9810629; GFX9-NEXT:    ; implicit-def: $sgpr9710630; GFX9-NEXT:    ; implicit-def: $sgpr9610631; GFX9-NEXT:    ; implicit-def: $sgpr8710632; GFX9-NEXT:    ; implicit-def: $sgpr8610633; GFX9-NEXT:    ; implicit-def: $sgpr8510634; GFX9-NEXT:    ; implicit-def: $sgpr8410635; GFX9-NEXT:    ; implicit-def: $sgpr8110636; GFX9-NEXT:    ; implicit-def: $sgpr3610637; GFX9-NEXT:    ; implicit-def: $sgpr3410638; GFX9-NEXT:    ; implicit-def: $sgpr3010639; GFX9-NEXT:    ; implicit-def: $sgpr9410640; GFX9-NEXT:    ; implicit-def: $sgpr9210641; GFX9-NEXT:    ; implicit-def: $sgpr9010642; GFX9-NEXT:    ; implicit-def: $sgpr8810643; GFX9-NEXT:    ; implicit-def: $sgpr7810644; GFX9-NEXT:    ; implicit-def: $sgpr7610645; GFX9-NEXT:    ; implicit-def: $sgpr7410646; GFX9-NEXT:    ; implicit-def: $sgpr7210647; GFX9-NEXT:    ; implicit-def: $sgpr6210648; GFX9-NEXT:    ; implicit-def: $sgpr6010649; GFX9-NEXT:    ; implicit-def: $sgpr5810650; GFX9-NEXT:    ; implicit-def: $sgpr2810651; GFX9-NEXT:    v_writelane_b32 v22, s83, 110652; GFX9-NEXT:    ; implicit-def: $sgpr2710653; GFX9-NEXT:    ; kill: killed $sgpr2710654; GFX9-NEXT:    ; implicit-def: $sgpr8210655; GFX9-NEXT:    ; implicit-def: $sgpr2710656; GFX9-NEXT:    ; kill: killed $sgpr2710657; GFX9-NEXT:    ; implicit-def: $sgpr2710658; GFX9-NEXT:    ; kill: killed $sgpr2710659; GFX9-NEXT:    ; implicit-def: $sgpr2710660; GFX9-NEXT:    ; kill: killed $sgpr2710661; GFX9-NEXT:    ; implicit-def: $sgpr2710662; GFX9-NEXT:    ; kill: killed $sgpr2710663; GFX9-NEXT:    ; implicit-def: $sgpr2710664; GFX9-NEXT:    ; kill: killed $sgpr2710665; GFX9-NEXT:    ; implicit-def: $sgpr2710666; GFX9-NEXT:    ; kill: killed $sgpr2710667; GFX9-NEXT:    ; implicit-def: $sgpr2710668; GFX9-NEXT:    ; kill: killed $sgpr2710669; GFX9-NEXT:    ; implicit-def: $sgpr2710670; GFX9-NEXT:    ; kill: killed $sgpr2710671; GFX9-NEXT:    ; implicit-def: $sgpr2710672; GFX9-NEXT:    ; kill: killed $sgpr2710673; GFX9-NEXT:    ; implicit-def: $sgpr2710674; GFX9-NEXT:    ; kill: killed $sgpr2710675; GFX9-NEXT:    ; implicit-def: $sgpr2710676; GFX9-NEXT:    ; kill: killed $sgpr2710677; GFX9-NEXT:    ; implicit-def: $sgpr2710678; GFX9-NEXT:    ; kill: killed $sgpr2710679; GFX9-NEXT:    ; implicit-def: $sgpr2710680; GFX9-NEXT:    ; kill: killed $sgpr2710681; GFX9-NEXT:    ; implicit-def: $sgpr2710682; GFX9-NEXT:    ; kill: killed $sgpr2710683; GFX9-NEXT:    ; implicit-def: $sgpr2710684; GFX9-NEXT:    ; kill: killed $sgpr2710685; GFX9-NEXT:    ; implicit-def: $sgpr2710686; GFX9-NEXT:    ; kill: killed $sgpr2710687; GFX9-NEXT:    ; implicit-def: $sgpr2710688; GFX9-NEXT:    ; kill: killed $sgpr2710689; GFX9-NEXT:    ; implicit-def: $sgpr2710690; GFX9-NEXT:    ; kill: killed $sgpr2710691; GFX9-NEXT:    ; implicit-def: $sgpr2710692; GFX9-NEXT:    ; kill: killed $sgpr2710693; GFX9-NEXT:    ; implicit-def: $sgpr2710694; GFX9-NEXT:    ; kill: killed $sgpr2710695; GFX9-NEXT:    ; implicit-def: $sgpr2710696; GFX9-NEXT:    ; kill: killed $sgpr2710697; GFX9-NEXT:    ; implicit-def: $sgpr2710698; GFX9-NEXT:    ; kill: killed $sgpr2710699; GFX9-NEXT:    ; implicit-def: $sgpr2710700; GFX9-NEXT:    ; kill: killed $sgpr2710701; GFX9-NEXT:    ; implicit-def: $sgpr2710702; GFX9-NEXT:    ; kill: killed $sgpr2710703; GFX9-NEXT:    ; implicit-def: $sgpr2710704; GFX9-NEXT:    ; kill: killed $sgpr2710705; GFX9-NEXT:    ; implicit-def: $sgpr2710706; GFX9-NEXT:    ; kill: killed $sgpr2710707; GFX9-NEXT:    ; implicit-def: $sgpr2710708; GFX9-NEXT:    ; kill: killed $sgpr2710709; GFX9-NEXT:    ; implicit-def: $sgpr2710710; GFX9-NEXT:    ; kill: killed $sgpr2710711; GFX9-NEXT:    ; implicit-def: $sgpr2710712; GFX9-NEXT:    ; kill: killed $sgpr2710713; GFX9-NEXT:    ; implicit-def: $sgpr2710714; GFX9-NEXT:    ; kill: killed $sgpr2710715; GFX9-NEXT:    ; implicit-def: $sgpr2710716; GFX9-NEXT:    ; kill: killed $sgpr2710717; GFX9-NEXT:    ; implicit-def: $sgpr2710718; GFX9-NEXT:    ; kill: killed $sgpr2710719; GFX9-NEXT:    ; implicit-def: $sgpr2710720; GFX9-NEXT:    ; kill: killed $sgpr2710721; GFX9-NEXT:    ; implicit-def: $sgpr2710722; GFX9-NEXT:    ; kill: killed $sgpr2710723; GFX9-NEXT:    ; implicit-def: $sgpr2710724; GFX9-NEXT:    ; kill: killed $sgpr2710725; GFX9-NEXT:    ; implicit-def: $sgpr2710726; GFX9-NEXT:    ; kill: killed $sgpr2710727; GFX9-NEXT:    ; implicit-def: $sgpr2710728; GFX9-NEXT:    ; kill: killed $sgpr2710729; GFX9-NEXT:    ; implicit-def: $sgpr2710730; GFX9-NEXT:    ; kill: killed $sgpr2710731; GFX9-NEXT:    ; implicit-def: $sgpr2710732; GFX9-NEXT:    ; kill: killed $sgpr2710733; GFX9-NEXT:    ; implicit-def: $sgpr2710734; GFX9-NEXT:    ; kill: killed $sgpr2710735; GFX9-NEXT:    ; implicit-def: $sgpr2710736; GFX9-NEXT:    ; kill: killed $sgpr2710737; GFX9-NEXT:    ; implicit-def: $sgpr2710738; GFX9-NEXT:    ; kill: killed $sgpr2710739; GFX9-NEXT:    ; implicit-def: $sgpr2710740; GFX9-NEXT:    ; kill: killed $sgpr2710741; GFX9-NEXT:    ; implicit-def: $sgpr2710742; GFX9-NEXT:    ; kill: killed $sgpr2710743; GFX9-NEXT:    ; implicit-def: $sgpr2710744; GFX9-NEXT:    ; kill: killed $sgpr2710745; GFX9-NEXT:    s_branch .LBB13_210746;10747; GFX11-LABEL: bitcast_v32i32_to_v128i8_scalar:10748; GFX11:       ; %bb.0:10749; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)10750; GFX11-NEXT:    s_xor_saveexec_b32 s4, -110751; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Spill10752; GFX11-NEXT:    scratch_store_b32 off, v34, s3210753; GFX11-NEXT:    scratch_store_b32 off, v35, s32 offset:410754; GFX11-NEXT:    scratch_store_b32 off, v36, s32 offset:810755; GFX11-NEXT:    scratch_store_b32 off, v37, s32 offset:1210756; GFX11-NEXT:    s_mov_b32 exec_lo, s410757; GFX11-NEXT:    v_writelane_b32 v34, s30, 010758; GFX11-NEXT:    v_writelane_b32 v35, s96, 010759; GFX11-NEXT:    v_dual_mov_b32 v16, s0 :: v_dual_mov_b32 v17, s110760; GFX11-NEXT:    v_dual_mov_b32 v18, s2 :: v_dual_mov_b32 v19, s310761; GFX11-NEXT:    v_writelane_b32 v34, s31, 110762; GFX11-NEXT:    v_writelane_b32 v35, s97, 110763; GFX11-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v21, s1710764; GFX11-NEXT:    v_dual_mov_b32 v22, s18 :: v_dual_mov_b32 v23, s1910765; GFX11-NEXT:    v_writelane_b32 v34, s34, 210766; GFX11-NEXT:    v_writelane_b32 v35, s98, 210767; GFX11-NEXT:    v_dual_mov_b32 v24, s20 :: v_dual_mov_b32 v25, s2110768; GFX11-NEXT:    v_dual_mov_b32 v26, s22 :: v_dual_mov_b32 v27, s2310769; GFX11-NEXT:    v_writelane_b32 v34, s35, 310770; GFX11-NEXT:    v_writelane_b32 v35, s99, 310771; GFX11-NEXT:    v_dual_mov_b32 v28, s24 :: v_dual_mov_b32 v29, s2510772; GFX11-NEXT:    v_dual_mov_b32 v30, s26 :: v_dual_mov_b32 v31, s2710773; GFX11-NEXT:    v_writelane_b32 v34, s36, 410774; GFX11-NEXT:    v_writelane_b32 v35, s100, 410775; GFX11-NEXT:    v_dual_mov_b32 v32, s28 :: v_dual_mov_b32 v33, s2910776; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v1510777; GFX11-NEXT:    v_writelane_b32 v34, s37, 510778; GFX11-NEXT:    v_writelane_b32 v35, s101, 510779; GFX11-NEXT:    v_readfirstlane_b32 s40, v1610780; GFX11-NEXT:    v_readfirstlane_b32 s41, v1710781; GFX11-NEXT:    v_readfirstlane_b32 s28, v1810782; GFX11-NEXT:    v_writelane_b32 v34, s38, 610783; GFX11-NEXT:    v_writelane_b32 v35, s102, 610784; GFX11-NEXT:    v_readfirstlane_b32 s29, v1910785; GFX11-NEXT:    v_readfirstlane_b32 s26, v2010786; GFX11-NEXT:    v_readfirstlane_b32 s27, v2110787; GFX11-NEXT:    v_writelane_b32 v34, s39, 710788; GFX11-NEXT:    v_writelane_b32 v35, s103, 710789; GFX11-NEXT:    v_readfirstlane_b32 s24, v2210790; GFX11-NEXT:    v_readfirstlane_b32 s25, v2310791; GFX11-NEXT:    v_readfirstlane_b32 s22, v2410792; GFX11-NEXT:    v_writelane_b32 v34, s48, 810793; GFX11-NEXT:    v_readfirstlane_b32 s23, v2510794; GFX11-NEXT:    v_readfirstlane_b32 s20, v2610795; GFX11-NEXT:    v_readfirstlane_b32 s21, v2710796; GFX11-NEXT:    v_readfirstlane_b32 s18, v2810797; GFX11-NEXT:    v_writelane_b32 v34, s49, 910798; GFX11-NEXT:    v_readfirstlane_b32 s19, v2910799; GFX11-NEXT:    v_readfirstlane_b32 s16, v3010800; GFX11-NEXT:    v_readfirstlane_b32 s17, v3110801; GFX11-NEXT:    v_readfirstlane_b32 s14, v3210802; GFX11-NEXT:    v_writelane_b32 v34, s50, 1010803; GFX11-NEXT:    v_readfirstlane_b32 s15, v3310804; GFX11-NEXT:    v_readfirstlane_b32 s12, v110805; GFX11-NEXT:    v_readfirstlane_b32 s13, v210806; GFX11-NEXT:    v_readfirstlane_b32 s10, v310807; GFX11-NEXT:    v_writelane_b32 v34, s51, 1110808; GFX11-NEXT:    v_readfirstlane_b32 s11, v410809; GFX11-NEXT:    v_readfirstlane_b32 s8, v510810; GFX11-NEXT:    v_readfirstlane_b32 s9, v610811; GFX11-NEXT:    v_readfirstlane_b32 s6, v710812; GFX11-NEXT:    v_writelane_b32 v34, s52, 1210813; GFX11-NEXT:    v_readfirstlane_b32 s7, v810814; GFX11-NEXT:    v_readfirstlane_b32 s4, v910815; GFX11-NEXT:    v_readfirstlane_b32 s5, v1010816; GFX11-NEXT:    v_readfirstlane_b32 s2, v1110817; GFX11-NEXT:    v_writelane_b32 v34, s53, 1310818; GFX11-NEXT:    v_readfirstlane_b32 s3, v1210819; GFX11-NEXT:    v_readfirstlane_b32 s0, v1310820; GFX11-NEXT:    v_readfirstlane_b32 s1, v1410821; GFX11-NEXT:    s_mov_b32 s101, 010822; GFX11-NEXT:    v_writelane_b32 v34, s54, 1410823; GFX11-NEXT:    s_and_b32 s42, vcc_lo, exec_lo10824; GFX11-NEXT:    v_writelane_b32 v35, s104, 810825; GFX11-NEXT:    ; implicit-def: $vgpr37 : SGPR spill to VGPR lane10826; GFX11-NEXT:    ; implicit-def: $vgpr36 : SGPR spill to VGPR lane10827; GFX11-NEXT:    v_writelane_b32 v34, s55, 1510828; GFX11-NEXT:    v_writelane_b32 v34, s64, 1610829; GFX11-NEXT:    v_writelane_b32 v34, s65, 1710830; GFX11-NEXT:    v_writelane_b32 v34, s66, 1810831; GFX11-NEXT:    v_writelane_b32 v34, s67, 1910832; GFX11-NEXT:    v_writelane_b32 v34, s68, 2010833; GFX11-NEXT:    v_writelane_b32 v34, s69, 2110834; GFX11-NEXT:    v_writelane_b32 v34, s70, 2210835; GFX11-NEXT:    v_writelane_b32 v34, s71, 2310836; GFX11-NEXT:    v_writelane_b32 v34, s80, 2410837; GFX11-NEXT:    v_writelane_b32 v34, s81, 2510838; GFX11-NEXT:    v_writelane_b32 v34, s82, 2610839; GFX11-NEXT:    v_writelane_b32 v34, s83, 2710840; GFX11-NEXT:    v_writelane_b32 v34, s84, 2810841; GFX11-NEXT:    v_writelane_b32 v34, s85, 2910842; GFX11-NEXT:    v_writelane_b32 v34, s86, 3010843; GFX11-NEXT:    v_writelane_b32 v34, s87, 3110844; GFX11-NEXT:    s_cbranch_scc0 .LBB13_210845; GFX11-NEXT:  ; %bb.1: ; %cmp.false10846; GFX11-NEXT:    s_lshr_b32 s43, s19, 810847; GFX11-NEXT:    s_lshr_b64 s[62:63], s[0:1], 2410848; GFX11-NEXT:    v_writelane_b32 v37, s43, 1610849; GFX11-NEXT:    s_lshr_b32 s43, s18, 1610850; GFX11-NEXT:    s_lshr_b32 s104, s1, 2410851; GFX11-NEXT:    s_lshr_b32 s102, s1, 1610852; GFX11-NEXT:    s_lshr_b32 s103, s1, 810853; GFX11-NEXT:    v_writelane_b32 v37, s43, 1710854; GFX11-NEXT:    s_lshr_b32 s43, s18, 810855; GFX11-NEXT:    s_lshr_b32 s57, s0, 1610856; GFX11-NEXT:    s_lshr_b32 s47, s0, 810857; GFX11-NEXT:    s_lshr_b32 s46, s3, 2410858; GFX11-NEXT:    v_writelane_b32 v37, s43, 1810859; GFX11-NEXT:    s_lshr_b32 s43, s21, 2410860; GFX11-NEXT:    s_lshr_b32 vcc_hi, s3, 1610861; GFX11-NEXT:    s_lshr_b32 s34, s3, 810862; GFX11-NEXT:    s_lshr_b32 s69, s2, 1610863; GFX11-NEXT:    v_writelane_b32 v37, s43, 1910864; GFX11-NEXT:    s_lshr_b32 s43, s21, 1610865; GFX11-NEXT:    s_lshr_b32 s56, s2, 810866; GFX11-NEXT:    s_lshr_b32 s35, s5, 2410867; GFX11-NEXT:    s_lshr_b32 s36, s5, 1610868; GFX11-NEXT:    v_writelane_b32 v37, s43, 2010869; GFX11-NEXT:    s_lshr_b32 s43, s21, 810870; GFX11-NEXT:    s_lshr_b32 s37, s5, 810871; GFX11-NEXT:    s_lshr_b32 s38, s4, 1610872; GFX11-NEXT:    s_lshr_b32 s39, s4, 810873; GFX11-NEXT:    v_writelane_b32 v37, s43, 2110874; GFX11-NEXT:    s_lshr_b32 s43, s20, 1610875; GFX11-NEXT:    s_lshr_b32 s48, s7, 2410876; GFX11-NEXT:    s_lshr_b32 s49, s7, 1610877; GFX11-NEXT:    s_lshr_b32 s50, s7, 810878; GFX11-NEXT:    v_writelane_b32 v37, s43, 2210879; GFX11-NEXT:    s_lshr_b32 s43, s20, 810880; GFX11-NEXT:    s_lshr_b32 s51, s6, 1610881; GFX11-NEXT:    s_lshr_b32 s52, s6, 810882; GFX11-NEXT:    s_lshr_b32 s53, s9, 2410883; GFX11-NEXT:    v_writelane_b32 v37, s43, 2310884; GFX11-NEXT:    s_lshr_b32 s43, s23, 2410885; GFX11-NEXT:    s_lshr_b32 s54, s9, 1610886; GFX11-NEXT:    s_lshr_b32 s55, s9, 810887; GFX11-NEXT:    s_lshr_b32 s64, s8, 1610888; GFX11-NEXT:    v_writelane_b32 v37, s43, 2410889; GFX11-NEXT:    s_lshr_b32 s43, s23, 1610890; GFX11-NEXT:    s_lshr_b32 s65, s8, 810891; GFX11-NEXT:    s_lshr_b32 s66, s11, 2410892; GFX11-NEXT:    s_lshr_b32 s67, s11, 1610893; GFX11-NEXT:    v_writelane_b32 v37, s43, 2510894; GFX11-NEXT:    s_lshr_b32 s43, s23, 810895; GFX11-NEXT:    s_lshr_b32 s68, s11, 810896; GFX11-NEXT:    s_lshr_b32 s59, s10, 1610897; GFX11-NEXT:    s_lshr_b32 s58, s10, 810898; GFX11-NEXT:    v_writelane_b32 v37, s43, 2610899; GFX11-NEXT:    s_lshr_b32 s43, s22, 1610900; GFX11-NEXT:    s_lshr_b32 s70, s13, 2410901; GFX11-NEXT:    s_lshr_b32 s71, s13, 1610902; GFX11-NEXT:    s_lshr_b32 s60, s13, 810903; GFX11-NEXT:    v_writelane_b32 v37, s43, 2710904; GFX11-NEXT:    s_lshr_b32 s43, s22, 810905; GFX11-NEXT:    s_lshr_b32 s80, s12, 1610906; GFX11-NEXT:    s_lshr_b32 s61, s12, 810907; GFX11-NEXT:    s_lshr_b32 s81, s15, 2410908; GFX11-NEXT:    v_writelane_b32 v37, s43, 2810909; GFX11-NEXT:    s_lshr_b32 s43, s25, 2410910; GFX11-NEXT:    s_lshr_b32 s82, s15, 1610911; GFX11-NEXT:    s_lshr_b32 s83, s15, 810912; GFX11-NEXT:    s_lshr_b32 s84, s14, 1610913; GFX11-NEXT:    v_writelane_b32 v37, s43, 2910914; GFX11-NEXT:    s_lshr_b32 s43, s25, 1610915; GFX11-NEXT:    s_lshr_b32 s85, s14, 810916; GFX11-NEXT:    s_lshr_b32 s86, s17, 2410917; GFX11-NEXT:    s_lshr_b32 s72, s17, 1610918; GFX11-NEXT:    v_writelane_b32 v37, s43, 3010919; GFX11-NEXT:    s_lshr_b32 s43, s25, 810920; GFX11-NEXT:    s_lshr_b32 s87, s17, 810921; GFX11-NEXT:    s_lshr_b32 s73, s16, 1610922; GFX11-NEXT:    s_lshr_b32 s96, s16, 810923; GFX11-NEXT:    v_writelane_b32 v37, s43, 3110924; GFX11-NEXT:    s_lshr_b32 s43, s24, 1610925; GFX11-NEXT:    s_lshr_b32 s97, s19, 2410926; GFX11-NEXT:    v_writelane_b32 v36, s43, 010927; GFX11-NEXT:    s_lshr_b32 s43, s24, 810928; GFX11-NEXT:    v_writelane_b32 v37, s62, 1410929; GFX11-NEXT:    s_lshr_b32 s42, s19, 1610930; GFX11-NEXT:    s_lshr_b32 s74, s28, 1610931; GFX11-NEXT:    v_writelane_b32 v36, s43, 110932; GFX11-NEXT:    s_lshr_b32 s43, s27, 2410933; GFX11-NEXT:    v_writelane_b32 v37, s63, 1510934; GFX11-NEXT:    s_lshr_b64 s[62:63], s[2:3], 2410935; GFX11-NEXT:    s_lshr_b32 s98, s41, 2410936; GFX11-NEXT:    v_writelane_b32 v36, s43, 210937; GFX11-NEXT:    s_lshr_b32 s43, s27, 1610938; GFX11-NEXT:    v_writelane_b32 v37, s62, 1210939; GFX11-NEXT:    s_lshr_b32 s99, s41, 1610940; GFX11-NEXT:    s_lshr_b32 s100, s41, 810941; GFX11-NEXT:    v_writelane_b32 v36, s43, 310942; GFX11-NEXT:    s_lshr_b32 s43, s27, 810943; GFX11-NEXT:    v_writelane_b32 v37, s63, 1310944; GFX11-NEXT:    s_lshr_b64 s[62:63], s[4:5], 2410945; GFX11-NEXT:    s_lshr_b32 s44, s40, 1610946; GFX11-NEXT:    v_writelane_b32 v36, s43, 410947; GFX11-NEXT:    s_lshr_b32 s43, s26, 1610948; GFX11-NEXT:    v_writelane_b32 v37, s62, 1010949; GFX11-NEXT:    s_lshr_b32 s45, s40, 810950; GFX11-NEXT:    s_lshr_b64 s[76:77], s[16:17], 2410951; GFX11-NEXT:    v_writelane_b32 v36, s43, 510952; GFX11-NEXT:    s_lshr_b32 s43, s26, 810953; GFX11-NEXT:    v_writelane_b32 v37, s63, 1110954; GFX11-NEXT:    s_lshr_b64 s[62:63], s[6:7], 2410955; GFX11-NEXT:    s_lshr_b64 s[88:89], s[18:19], 2410956; GFX11-NEXT:    v_writelane_b32 v36, s43, 610957; GFX11-NEXT:    s_lshr_b32 s43, s29, 2410958; GFX11-NEXT:    v_writelane_b32 v37, s62, 810959; GFX11-NEXT:    s_lshr_b64 s[78:79], s[22:23], 2410960; GFX11-NEXT:    s_lshr_b64 s[90:91], s[24:25], 2410961; GFX11-NEXT:    v_writelane_b32 v36, s43, 710962; GFX11-NEXT:    s_lshr_b32 s43, s29, 1610963; GFX11-NEXT:    v_writelane_b32 v37, s63, 910964; GFX11-NEXT:    s_lshr_b64 s[62:63], s[8:9], 2410965; GFX11-NEXT:    s_lshr_b64 s[92:93], s[26:27], 2410966; GFX11-NEXT:    v_writelane_b32 v36, s43, 810967; GFX11-NEXT:    s_lshr_b32 s43, s29, 810968; GFX11-NEXT:    v_writelane_b32 v37, s62, 610969; GFX11-NEXT:    s_lshr_b64 s[94:95], s[28:29], 2410970; GFX11-NEXT:    s_lshr_b64 s[30:31], s[40:41], 2410971; GFX11-NEXT:    v_writelane_b32 v36, s43, 910972; GFX11-NEXT:    s_lshr_b32 s43, s28, 810973; GFX11-NEXT:    v_writelane_b32 v37, s63, 710974; GFX11-NEXT:    s_lshr_b64 s[62:63], s[10:11], 2410975; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)10976; GFX11-NEXT:    v_writelane_b32 v37, s62, 410977; GFX11-NEXT:    v_writelane_b32 v37, s63, 510978; GFX11-NEXT:    s_lshr_b64 s[62:63], s[12:13], 2410979; GFX11-NEXT:    v_writelane_b32 v37, s62, 210980; GFX11-NEXT:    v_writelane_b32 v37, s63, 310981; GFX11-NEXT:    s_lshr_b64 s[62:63], s[14:15], 2410982; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)10983; GFX11-NEXT:    v_writelane_b32 v37, s62, 010984; GFX11-NEXT:    v_writelane_b32 v37, s63, 110985; GFX11-NEXT:    s_lshr_b64 s[62:63], s[20:21], 2410986; GFX11-NEXT:    s_branch .LBB13_310987; GFX11-NEXT:  .LBB13_2:10988; GFX11-NEXT:    ; implicit-def: $vcc_hi10989; GFX11-NEXT:    ; implicit-def: $vcc_lo10990; GFX11-NEXT:    ; implicit-def: $sgpr4210991; GFX11-NEXT:    ; kill: killed $sgpr4210992; GFX11-NEXT:    s_mov_b32 s101, -110993; GFX11-NEXT:    v_writelane_b32 v37, vcc_lo, 010994; GFX11-NEXT:    ; implicit-def: $sgpr4210995; GFX11-NEXT:    ; kill: killed $sgpr4210996; GFX11-NEXT:    ; implicit-def: $sgpr4210997; GFX11-NEXT:    ; kill: killed $sgpr4210998; GFX11-NEXT:    ; implicit-def: $sgpr4210999; GFX11-NEXT:    ; kill: killed $sgpr4211000; GFX11-NEXT:    ; implicit-def: $sgpr4211001; GFX11-NEXT:    ; kill: killed $sgpr4211002; GFX11-NEXT:    ; implicit-def: $sgpr4211003; GFX11-NEXT:    v_writelane_b32 v37, vcc_hi, 111004; GFX11-NEXT:    ; implicit-def: $vcc_lo11005; GFX11-NEXT:    ; kill: killed $sgpr4211006; GFX11-NEXT:    ; implicit-def: $sgpr4211007; GFX11-NEXT:    ; kill: killed $sgpr4211008; GFX11-NEXT:    ; implicit-def: $sgpr4211009; GFX11-NEXT:    ; kill: killed $sgpr4211010; GFX11-NEXT:    ; implicit-def: $sgpr4211011; GFX11-NEXT:    ; kill: killed $sgpr4211012; GFX11-NEXT:    ; implicit-def: $sgpr4211013; GFX11-NEXT:    ; kill: killed $sgpr4211014; GFX11-NEXT:    ; implicit-def: $sgpr4211015; GFX11-NEXT:    v_writelane_b32 v37, vcc_lo, 211016; GFX11-NEXT:    ; kill: killed $sgpr4211017; GFX11-NEXT:    ; implicit-def: $sgpr4211018; GFX11-NEXT:    ; kill: killed $sgpr4211019; GFX11-NEXT:    ; implicit-def: $sgpr4211020; GFX11-NEXT:    ; kill: killed $sgpr4211021; GFX11-NEXT:    ; implicit-def: $sgpr4211022; GFX11-NEXT:    ; kill: killed $sgpr4211023; GFX11-NEXT:    ; implicit-def: $sgpr4211024; GFX11-NEXT:    ; kill: killed $sgpr4211025; GFX11-NEXT:    ; implicit-def: $sgpr4211026; GFX11-NEXT:    v_writelane_b32 v37, vcc_hi, 311027; GFX11-NEXT:    ; implicit-def: $vcc_lo11028; GFX11-NEXT:    ; kill: killed $sgpr4211029; GFX11-NEXT:    ; implicit-def: $sgpr4211030; GFX11-NEXT:    ; kill: killed $sgpr4211031; GFX11-NEXT:    ; implicit-def: $sgpr4211032; GFX11-NEXT:    ; kill: killed $sgpr4211033; GFX11-NEXT:    ; implicit-def: $sgpr4211034; GFX11-NEXT:    ; kill: killed $sgpr4211035; GFX11-NEXT:    ; implicit-def: $sgpr4211036; GFX11-NEXT:    ; kill: killed $sgpr4211037; GFX11-NEXT:    ; implicit-def: $sgpr4211038; GFX11-NEXT:    v_writelane_b32 v37, vcc_lo, 411039; GFX11-NEXT:    ; kill: killed $sgpr4211040; GFX11-NEXT:    ; implicit-def: $sgpr4211041; GFX11-NEXT:    ; kill: killed $sgpr4211042; GFX11-NEXT:    ; implicit-def: $sgpr4211043; GFX11-NEXT:    ; kill: killed $sgpr4211044; GFX11-NEXT:    ; implicit-def: $sgpr4211045; GFX11-NEXT:    ; kill: killed $sgpr4211046; GFX11-NEXT:    ; implicit-def: $sgpr4211047; GFX11-NEXT:    ; kill: killed $sgpr4211048; GFX11-NEXT:    ; implicit-def: $sgpr4211049; GFX11-NEXT:    v_writelane_b32 v37, vcc_hi, 511050; GFX11-NEXT:    ; implicit-def: $vcc_lo11051; GFX11-NEXT:    ; implicit-def: $sgpr4511052; GFX11-NEXT:    ; implicit-def: $sgpr4411053; GFX11-NEXT:    ; implicit-def: $sgpr3011054; GFX11-NEXT:    ; implicit-def: $sgpr10011055; GFX11-NEXT:    ; implicit-def: $sgpr9911056; GFX11-NEXT:    ; implicit-def: $sgpr9811057; GFX11-NEXT:    ; implicit-def: $sgpr4311058; GFX11-NEXT:    ; implicit-def: $sgpr7411059; GFX11-NEXT:    ; implicit-def: $sgpr9411060; GFX11-NEXT:    ; implicit-def: $sgpr9211061; GFX11-NEXT:    ; implicit-def: $sgpr9011062; GFX11-NEXT:    ; implicit-def: $sgpr7811063; GFX11-NEXT:    ; implicit-def: $sgpr6211064; GFX11-NEXT:    ; kill: killed $sgpr4211065; GFX11-NEXT:    ; implicit-def: $sgpr4211066; GFX11-NEXT:    ; implicit-def: $sgpr9711067; GFX11-NEXT:    ; implicit-def: $sgpr9611068; GFX11-NEXT:    ; implicit-def: $sgpr7311069; GFX11-NEXT:    ; implicit-def: $sgpr8711070; GFX11-NEXT:    ; implicit-def: $sgpr7211071; GFX11-NEXT:    ; implicit-def: $sgpr8611072; GFX11-NEXT:    ; implicit-def: $sgpr8511073; GFX11-NEXT:    ; implicit-def: $sgpr8411074; GFX11-NEXT:    ; implicit-def: $sgpr8311075; GFX11-NEXT:    ; implicit-def: $sgpr8211076; GFX11-NEXT:    ; implicit-def: $sgpr8111077; GFX11-NEXT:    ; implicit-def: $sgpr6111078; GFX11-NEXT:    ; implicit-def: $sgpr8011079; GFX11-NEXT:    ; implicit-def: $sgpr6011080; GFX11-NEXT:    ; implicit-def: $sgpr7111081; GFX11-NEXT:    ; implicit-def: $sgpr7011082; GFX11-NEXT:    ; implicit-def: $sgpr5811083; GFX11-NEXT:    ; implicit-def: $sgpr5911084; GFX11-NEXT:    ; implicit-def: $sgpr6811085; GFX11-NEXT:    ; implicit-def: $sgpr6711086; GFX11-NEXT:    ; implicit-def: $sgpr6611087; GFX11-NEXT:    ; implicit-def: $sgpr6511088; GFX11-NEXT:    ; implicit-def: $sgpr6411089; GFX11-NEXT:    ; implicit-def: $sgpr5511090; GFX11-NEXT:    ; implicit-def: $sgpr5411091; GFX11-NEXT:    ; implicit-def: $sgpr5311092; GFX11-NEXT:    ; implicit-def: $sgpr5211093; GFX11-NEXT:    ; implicit-def: $sgpr5111094; GFX11-NEXT:    ; implicit-def: $sgpr5011095; GFX11-NEXT:    ; implicit-def: $sgpr4911096; GFX11-NEXT:    ; implicit-def: $sgpr4811097; GFX11-NEXT:    ; implicit-def: $sgpr3911098; GFX11-NEXT:    ; implicit-def: $sgpr3811099; GFX11-NEXT:    ; implicit-def: $sgpr3711100; GFX11-NEXT:    ; implicit-def: $sgpr3611101; GFX11-NEXT:    ; implicit-def: $sgpr3511102; GFX11-NEXT:    ; implicit-def: $sgpr5611103; GFX11-NEXT:    ; implicit-def: $sgpr6911104; GFX11-NEXT:    ; implicit-def: $sgpr3411105; GFX11-NEXT:    ; implicit-def: $sgpr4611106; GFX11-NEXT:    ; implicit-def: $sgpr4711107; GFX11-NEXT:    ; implicit-def: $sgpr5711108; GFX11-NEXT:    ; implicit-def: $sgpr10311109; GFX11-NEXT:    ; implicit-def: $sgpr10211110; GFX11-NEXT:    ; implicit-def: $sgpr10411111; GFX11-NEXT:    ; implicit-def: $sgpr8811112; GFX11-NEXT:    ; implicit-def: $sgpr7611113; GFX11-NEXT:    v_writelane_b32 v37, vcc_lo, 611114; GFX11-NEXT:    v_writelane_b32 v37, vcc_hi, 711115; GFX11-NEXT:    ; implicit-def: $vcc_lo11116; GFX11-NEXT:    v_writelane_b32 v37, vcc_lo, 811117; GFX11-NEXT:    v_writelane_b32 v37, vcc_hi, 911118; GFX11-NEXT:    ; implicit-def: $vcc_lo11119; GFX11-NEXT:    v_writelane_b32 v37, vcc_lo, 1011120; GFX11-NEXT:    v_writelane_b32 v37, vcc_hi, 1111121; GFX11-NEXT:    ; implicit-def: $vcc_lo11122; GFX11-NEXT:    v_writelane_b32 v37, vcc_lo, 1211123; GFX11-NEXT:    v_writelane_b32 v37, vcc_hi, 1311124; GFX11-NEXT:    ; implicit-def: $vcc_lo11125; GFX11-NEXT:    v_writelane_b32 v37, vcc_lo, 1411126; GFX11-NEXT:    v_writelane_b32 v37, vcc_hi, 1511127; GFX11-NEXT:  .LBB13_3: ; %Flow11128; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s10111129; GFX11-NEXT:    s_mov_b32 s101, s10411130; GFX11-NEXT:    s_mov_b32 s104, s5711131; GFX11-NEXT:    s_mov_b32 s57, s6911132; GFX11-NEXT:    s_mov_b32 s69, s4211133; GFX11-NEXT:    s_cbranch_vccnz .LBB13_511134; GFX11-NEXT:  ; %bb.4: ; %cmp.true11135; GFX11-NEXT:    s_add_i32 s19, s19, 311136; GFX11-NEXT:    s_add_i32 s18, s18, 311137; GFX11-NEXT:    s_lshr_b32 s42, s19, 811138; GFX11-NEXT:    s_add_i32 s21, s21, 311139; GFX11-NEXT:    v_writelane_b32 v37, s42, 1611140; GFX11-NEXT:    s_lshr_b32 s42, s18, 1611141; GFX11-NEXT:    s_add_i32 s20, s20, 311142; GFX11-NEXT:    s_add_i32 s23, s23, 311143; GFX11-NEXT:    s_add_i32 s22, s22, 311144; GFX11-NEXT:    v_writelane_b32 v37, s42, 1711145; GFX11-NEXT:    s_lshr_b32 s42, s18, 811146; GFX11-NEXT:    s_add_i32 s25, s25, 311147; GFX11-NEXT:    s_add_i32 s1, s1, 311148; GFX11-NEXT:    s_add_i32 s0, s0, 311149; GFX11-NEXT:    v_writelane_b32 v37, s42, 1811150; GFX11-NEXT:    s_lshr_b32 s42, s21, 2411151; GFX11-NEXT:    s_lshr_b64 s[62:63], s[0:1], 2411152; GFX11-NEXT:    s_add_i32 s3, s3, 311153; GFX11-NEXT:    s_add_i32 s2, s2, 311154; GFX11-NEXT:    v_writelane_b32 v37, s42, 1911155; GFX11-NEXT:    s_lshr_b32 s42, s21, 1611156; GFX11-NEXT:    s_add_i32 s5, s5, 311157; GFX11-NEXT:    s_add_i32 s4, s4, 311158; GFX11-NEXT:    s_add_i32 s7, s7, 311159; GFX11-NEXT:    v_writelane_b32 v37, s42, 2011160; GFX11-NEXT:    s_lshr_b32 s42, s21, 811161; GFX11-NEXT:    s_add_i32 s6, s6, 311162; GFX11-NEXT:    s_add_i32 s24, s24, 311163; GFX11-NEXT:    s_add_i32 s9, s9, 311164; GFX11-NEXT:    v_writelane_b32 v37, s42, 2111165; GFX11-NEXT:    s_lshr_b32 s42, s20, 1611166; GFX11-NEXT:    s_add_i32 s8, s8, 311167; GFX11-NEXT:    s_add_i32 s27, s27, 311168; GFX11-NEXT:    s_add_i32 s11, s11, 311169; GFX11-NEXT:    v_writelane_b32 v37, s42, 2211170; GFX11-NEXT:    s_lshr_b32 s42, s20, 811171; GFX11-NEXT:    s_add_i32 s10, s10, 311172; GFX11-NEXT:    s_add_i32 s26, s26, 311173; GFX11-NEXT:    s_add_i32 s13, s13, 311174; GFX11-NEXT:    v_writelane_b32 v37, s42, 2311175; GFX11-NEXT:    s_lshr_b32 s42, s23, 2411176; GFX11-NEXT:    s_add_i32 s12, s12, 311177; GFX11-NEXT:    s_add_i32 s29, s29, 311178; GFX11-NEXT:    s_add_i32 s15, s15, 311179; GFX11-NEXT:    v_writelane_b32 v37, s42, 2411180; GFX11-NEXT:    s_lshr_b32 s42, s23, 1611181; GFX11-NEXT:    s_add_i32 s14, s14, 311182; GFX11-NEXT:    s_add_i32 s41, s41, 311183; GFX11-NEXT:    s_add_i32 s40, s40, 311184; GFX11-NEXT:    v_writelane_b32 v37, s42, 2511185; GFX11-NEXT:    s_lshr_b32 s42, s23, 811186; GFX11-NEXT:    s_add_i32 s28, s28, 311187; GFX11-NEXT:    s_add_i32 s17, s17, 311188; GFX11-NEXT:    s_add_i32 s16, s16, 311189; GFX11-NEXT:    v_writelane_b32 v37, s42, 2611190; GFX11-NEXT:    s_lshr_b32 s42, s22, 1611191; GFX11-NEXT:    s_lshr_b32 s101, s1, 2411192; GFX11-NEXT:    s_lshr_b32 s102, s1, 1611193; GFX11-NEXT:    s_lshr_b32 s103, s1, 811194; GFX11-NEXT:    v_writelane_b32 v37, s42, 2711195; GFX11-NEXT:    s_lshr_b32 s42, s22, 811196; GFX11-NEXT:    s_lshr_b32 s104, s0, 1611197; GFX11-NEXT:    s_lshr_b32 s47, s0, 811198; GFX11-NEXT:    s_lshr_b32 s46, s3, 2411199; GFX11-NEXT:    v_writelane_b32 v37, s42, 2811200; GFX11-NEXT:    s_lshr_b32 s42, s25, 2411201; GFX11-NEXT:    s_lshr_b32 vcc_hi, s3, 1611202; GFX11-NEXT:    s_lshr_b32 s34, s3, 811203; GFX11-NEXT:    s_lshr_b32 s57, s2, 1611204; GFX11-NEXT:    v_writelane_b32 v37, s42, 2911205; GFX11-NEXT:    s_lshr_b32 s42, s25, 1611206; GFX11-NEXT:    s_lshr_b32 s56, s2, 811207; GFX11-NEXT:    s_lshr_b32 s35, s5, 2411208; GFX11-NEXT:    s_lshr_b32 s36, s5, 1611209; GFX11-NEXT:    v_writelane_b32 v37, s42, 3011210; GFX11-NEXT:    s_lshr_b32 s42, s25, 811211; GFX11-NEXT:    s_lshr_b32 s37, s5, 811212; GFX11-NEXT:    s_lshr_b32 s38, s4, 1611213; GFX11-NEXT:    s_lshr_b32 s39, s4, 811214; GFX11-NEXT:    v_writelane_b32 v37, s42, 3111215; GFX11-NEXT:    s_lshr_b32 s42, s24, 1611216; GFX11-NEXT:    s_lshr_b32 s48, s7, 2411217; GFX11-NEXT:    v_writelane_b32 v36, s42, 011218; GFX11-NEXT:    s_lshr_b32 s42, s24, 811219; GFX11-NEXT:    v_writelane_b32 v37, s62, 1411220; GFX11-NEXT:    s_lshr_b32 s49, s7, 1611221; GFX11-NEXT:    s_lshr_b32 s50, s7, 811222; GFX11-NEXT:    v_writelane_b32 v36, s42, 111223; GFX11-NEXT:    s_lshr_b32 s42, s27, 2411224; GFX11-NEXT:    v_writelane_b32 v37, s63, 1511225; GFX11-NEXT:    s_lshr_b64 s[62:63], s[2:3], 2411226; GFX11-NEXT:    s_lshr_b32 s51, s6, 1611227; GFX11-NEXT:    v_writelane_b32 v36, s42, 211228; GFX11-NEXT:    s_lshr_b32 s42, s27, 1611229; GFX11-NEXT:    v_writelane_b32 v37, s62, 1211230; GFX11-NEXT:    s_lshr_b32 s52, s6, 811231; GFX11-NEXT:    s_lshr_b32 s53, s9, 2411232; GFX11-NEXT:    v_writelane_b32 v36, s42, 311233; GFX11-NEXT:    s_lshr_b32 s42, s27, 811234; GFX11-NEXT:    v_writelane_b32 v37, s63, 1311235; GFX11-NEXT:    s_lshr_b64 s[62:63], s[4:5], 2411236; GFX11-NEXT:    s_lshr_b32 s54, s9, 1611237; GFX11-NEXT:    v_writelane_b32 v36, s42, 411238; GFX11-NEXT:    s_lshr_b32 s42, s26, 1611239; GFX11-NEXT:    v_writelane_b32 v37, s62, 1011240; GFX11-NEXT:    s_lshr_b32 s55, s9, 811241; GFX11-NEXT:    s_lshr_b32 s64, s8, 1611242; GFX11-NEXT:    v_writelane_b32 v36, s42, 511243; GFX11-NEXT:    s_lshr_b32 s42, s26, 811244; GFX11-NEXT:    v_writelane_b32 v37, s63, 1111245; GFX11-NEXT:    s_lshr_b64 s[62:63], s[6:7], 2411246; GFX11-NEXT:    s_lshr_b32 s65, s8, 811247; GFX11-NEXT:    v_writelane_b32 v36, s42, 611248; GFX11-NEXT:    s_lshr_b32 s42, s29, 2411249; GFX11-NEXT:    v_writelane_b32 v37, s62, 811250; GFX11-NEXT:    s_lshr_b32 s66, s11, 2411251; GFX11-NEXT:    s_lshr_b32 s67, s11, 1611252; GFX11-NEXT:    v_writelane_b32 v36, s42, 711253; GFX11-NEXT:    s_lshr_b32 s42, s29, 1611254; GFX11-NEXT:    v_writelane_b32 v37, s63, 911255; GFX11-NEXT:    s_lshr_b64 s[62:63], s[8:9], 2411256; GFX11-NEXT:    s_lshr_b32 s68, s11, 811257; GFX11-NEXT:    v_writelane_b32 v36, s42, 811258; GFX11-NEXT:    s_lshr_b32 s59, s10, 1611259; GFX11-NEXT:    v_writelane_b32 v37, s62, 611260; GFX11-NEXT:    s_lshr_b32 s58, s10, 811261; GFX11-NEXT:    s_lshr_b32 s70, s13, 2411262; GFX11-NEXT:    s_lshr_b32 s71, s13, 1611263; GFX11-NEXT:    s_lshr_b32 s60, s13, 811264; GFX11-NEXT:    v_writelane_b32 v37, s63, 711265; GFX11-NEXT:    s_lshr_b64 s[62:63], s[10:11], 2411266; GFX11-NEXT:    s_lshr_b32 s80, s12, 1611267; GFX11-NEXT:    s_lshr_b32 s61, s12, 811268; GFX11-NEXT:    s_lshr_b32 s81, s15, 2411269; GFX11-NEXT:    v_writelane_b32 v37, s62, 411270; GFX11-NEXT:    s_lshr_b32 s82, s15, 1611271; GFX11-NEXT:    s_lshr_b32 s83, s15, 811272; GFX11-NEXT:    s_lshr_b32 s84, s14, 1611273; GFX11-NEXT:    s_lshr_b32 s85, s14, 811274; GFX11-NEXT:    v_writelane_b32 v37, s63, 511275; GFX11-NEXT:    s_lshr_b64 s[62:63], s[12:13], 2411276; GFX11-NEXT:    s_lshr_b32 s86, s17, 2411277; GFX11-NEXT:    s_lshr_b32 s72, s17, 1611278; GFX11-NEXT:    s_lshr_b32 s87, s17, 811279; GFX11-NEXT:    v_writelane_b32 v37, s62, 211280; GFX11-NEXT:    s_lshr_b32 s73, s16, 1611281; GFX11-NEXT:    s_lshr_b32 s96, s16, 811282; GFX11-NEXT:    s_lshr_b32 s97, s19, 2411283; GFX11-NEXT:    s_lshr_b32 s69, s19, 1611284; GFX11-NEXT:    v_writelane_b32 v37, s63, 311285; GFX11-NEXT:    s_lshr_b64 s[62:63], s[14:15], 2411286; GFX11-NEXT:    s_lshr_b32 s42, s29, 811287; GFX11-NEXT:    s_lshr_b32 s74, s28, 1611288; GFX11-NEXT:    s_lshr_b32 s43, s28, 811289; GFX11-NEXT:    v_writelane_b32 v37, s62, 011290; GFX11-NEXT:    s_lshr_b32 s98, s41, 2411291; GFX11-NEXT:    s_lshr_b32 s99, s41, 1611292; GFX11-NEXT:    s_lshr_b32 s100, s41, 811293; GFX11-NEXT:    s_lshr_b32 s44, s40, 1611294; GFX11-NEXT:    s_lshr_b32 s45, s40, 811295; GFX11-NEXT:    v_writelane_b32 v37, s63, 111296; GFX11-NEXT:    s_lshr_b64 s[76:77], s[16:17], 2411297; GFX11-NEXT:    s_lshr_b64 s[88:89], s[18:19], 2411298; GFX11-NEXT:    s_lshr_b64 s[62:63], s[20:21], 2411299; GFX11-NEXT:    s_lshr_b64 s[78:79], s[22:23], 2411300; GFX11-NEXT:    s_lshr_b64 s[90:91], s[24:25], 2411301; GFX11-NEXT:    s_lshr_b64 s[92:93], s[26:27], 2411302; GFX11-NEXT:    s_lshr_b64 s[94:95], s[28:29], 2411303; GFX11-NEXT:    s_lshr_b64 s[30:31], s[40:41], 2411304; GFX11-NEXT:    v_writelane_b32 v36, s42, 911305; GFX11-NEXT:  .LBB13_5: ; %end11306; GFX11-NEXT:    s_lshl_b32 s43, s43, 811307; GFX11-NEXT:    s_and_b32 s28, s28, 0xff11308; GFX11-NEXT:    s_and_b32 s42, s74, 0xff11309; GFX11-NEXT:    s_or_b32 s28, s28, s4311310; GFX11-NEXT:    s_lshl_b32 s43, s94, 811311; GFX11-NEXT:    s_and_b32 s28, s28, 0xffff11312; GFX11-NEXT:    s_or_b32 s42, s42, s4311313; GFX11-NEXT:    s_and_b32 s29, s29, 0xff11314; GFX11-NEXT:    s_lshl_b32 s42, s42, 1611315; GFX11-NEXT:    v_readlane_b32 s43, v36, 711316; GFX11-NEXT:    s_or_b32 s28, s28, s4211317; GFX11-NEXT:    v_readlane_b32 s42, v36, 911318; GFX11-NEXT:    s_and_b32 s26, s26, 0xff11319; GFX11-NEXT:    s_and_b32 s27, s27, 0xff11320; GFX11-NEXT:    s_lshl_b32 s43, s43, 811321; GFX11-NEXT:    s_and_b32 s24, s24, 0xff11322; GFX11-NEXT:    s_lshl_b32 s42, s42, 811323; GFX11-NEXT:    s_and_b32 s25, s25, 0xff11324; GFX11-NEXT:    s_or_b32 s29, s29, s4211325; GFX11-NEXT:    v_readlane_b32 s42, v36, 811326; GFX11-NEXT:    s_and_b32 s29, s29, 0xffff11327; GFX11-NEXT:    s_and_b32 s22, s22, 0xff11328; GFX11-NEXT:    s_and_b32 s23, s23, 0xff11329; GFX11-NEXT:    s_and_b32 s20, s20, 0xff11330; GFX11-NEXT:    s_and_b32 s42, s42, 0xff11331; GFX11-NEXT:    s_and_b32 s21, s21, 0xff11332; GFX11-NEXT:    s_or_b32 s42, s42, s4311333; GFX11-NEXT:    s_lshl_b32 s45, s45, 811334; GFX11-NEXT:    s_lshl_b32 s42, s42, 1611335; GFX11-NEXT:    s_and_b32 s40, s40, 0xff11336; GFX11-NEXT:    s_or_b32 s29, s29, s4211337; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)11338; GFX11-NEXT:    v_dual_mov_b32 v3, s28 :: v_dual_mov_b32 v4, s2911339; GFX11-NEXT:    v_readlane_b32 s28, v36, 611340; GFX11-NEXT:    v_readlane_b32 s29, v36, 511341; GFX11-NEXT:    s_or_b32 s40, s40, s4511342; GFX11-NEXT:    s_lshl_b32 s45, s30, 811343; GFX11-NEXT:    s_and_b32 s44, s44, 0xff11344; GFX11-NEXT:    s_lshl_b32 s28, s28, 811345; GFX11-NEXT:    s_and_b32 s29, s29, 0xff11346; GFX11-NEXT:    s_or_b32 s26, s26, s2811347; GFX11-NEXT:    s_lshl_b32 s28, s92, 811348; GFX11-NEXT:    s_and_b32 s26, s26, 0xffff11349; GFX11-NEXT:    s_or_b32 s28, s29, s2811350; GFX11-NEXT:    v_readlane_b32 s29, v36, 211351; GFX11-NEXT:    s_lshl_b32 s28, s28, 1611352; GFX11-NEXT:    s_or_b32 s44, s44, s4511353; GFX11-NEXT:    s_or_b32 s26, s26, s2811354; GFX11-NEXT:    v_readlane_b32 s28, v36, 411355; GFX11-NEXT:    s_lshl_b32 s29, s29, 811356; GFX11-NEXT:    s_and_b32 s18, s18, 0xff11357; GFX11-NEXT:    s_and_b32 s40, s40, 0xffff11358; GFX11-NEXT:    s_lshl_b32 s44, s44, 1611359; GFX11-NEXT:    s_lshl_b32 s28, s28, 811360; GFX11-NEXT:    s_or_b32 s40, s40, s4411361; GFX11-NEXT:    s_or_b32 s27, s27, s2811362; GFX11-NEXT:    v_readlane_b32 s28, v36, 311363; GFX11-NEXT:    s_and_b32 s27, s27, 0xffff11364; GFX11-NEXT:    s_and_b32 s41, s41, 0xff11365; GFX11-NEXT:    s_lshl_b32 s44, s100, 811366; GFX11-NEXT:    s_lshl_b32 s45, s98, 811367; GFX11-NEXT:    s_and_b32 s28, s28, 0xff11368; GFX11-NEXT:    s_or_b32 s41, s41, s4411369; GFX11-NEXT:    s_or_b32 s28, s28, s2911370; GFX11-NEXT:    v_readlane_b32 s29, v36, 011371; GFX11-NEXT:    s_lshl_b32 s28, s28, 1611372; GFX11-NEXT:    s_and_b32 s44, s99, 0xff11373; GFX11-NEXT:    s_or_b32 s27, s27, s2811374; GFX11-NEXT:    v_readlane_b32 s28, v36, 111375; GFX11-NEXT:    s_and_b32 s29, s29, 0xff11376; GFX11-NEXT:    v_dual_mov_b32 v5, s26 :: v_dual_mov_b32 v6, s2711377; GFX11-NEXT:    v_readlane_b32 s26, v37, 1911378; GFX11-NEXT:    s_lshl_b32 s28, s28, 811379; GFX11-NEXT:    s_or_b32 s44, s44, s4511380; GFX11-NEXT:    s_or_b32 s24, s24, s2811381; GFX11-NEXT:    s_lshl_b32 s28, s90, 811382; GFX11-NEXT:    s_and_b32 s24, s24, 0xffff11383; GFX11-NEXT:    s_or_b32 s28, s29, s2811384; GFX11-NEXT:    v_readlane_b32 s29, v37, 2911385; GFX11-NEXT:    s_lshl_b32 s28, s28, 1611386; GFX11-NEXT:    s_lshl_b32 s26, s26, 811387; GFX11-NEXT:    s_or_b32 s24, s24, s2811388; GFX11-NEXT:    v_readlane_b32 s28, v37, 3111389; GFX11-NEXT:    s_lshl_b32 s29, s29, 811390; GFX11-NEXT:    s_and_b32 s19, s19, 0xff11391; GFX11-NEXT:    s_and_b32 s41, s41, 0xffff11392; GFX11-NEXT:    s_lshl_b32 s44, s44, 1611393; GFX11-NEXT:    s_lshl_b32 s28, s28, 811394; GFX11-NEXT:    s_or_b32 s41, s41, s4411395; GFX11-NEXT:    s_or_b32 s25, s25, s2811396; GFX11-NEXT:    v_readlane_b32 s28, v37, 3011397; GFX11-NEXT:    s_and_b32 s25, s25, 0xffff11398; GFX11-NEXT:    v_dual_mov_b32 v1, s40 :: v_dual_mov_b32 v2, s4111399; GFX11-NEXT:    s_and_b32 s16, s16, 0xff11400; GFX11-NEXT:    s_and_b32 s28, s28, 0xff11401; GFX11-NEXT:    s_and_b32 s17, s17, 0xff11402; GFX11-NEXT:    s_or_b32 s28, s28, s2911403; GFX11-NEXT:    s_and_b32 s14, s14, 0xff11404; GFX11-NEXT:    s_lshl_b32 s28, s28, 1611405; GFX11-NEXT:    s_and_b32 s15, s15, 0xff11406; GFX11-NEXT:    s_or_b32 s25, s25, s2811407; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)11408; GFX11-NEXT:    v_dual_mov_b32 v7, s24 :: v_dual_mov_b32 v8, s2511409; GFX11-NEXT:    v_readlane_b32 s24, v37, 2811410; GFX11-NEXT:    v_readlane_b32 s25, v37, 2711411; GFX11-NEXT:    s_clause 0x111412; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off11413; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:1611414; GFX11-NEXT:    s_and_b32 s12, s12, 0xff11415; GFX11-NEXT:    s_lshl_b32 s24, s24, 811416; GFX11-NEXT:    s_and_b32 s25, s25, 0xff11417; GFX11-NEXT:    s_or_b32 s22, s22, s2411418; GFX11-NEXT:    s_lshl_b32 s24, s78, 811419; GFX11-NEXT:    s_and_b32 s22, s22, 0xffff11420; GFX11-NEXT:    s_or_b32 s24, s25, s2411421; GFX11-NEXT:    v_readlane_b32 s25, v37, 2411422; GFX11-NEXT:    s_lshl_b32 s24, s24, 1611423; GFX11-NEXT:    s_and_b32 s13, s13, 0xff11424; GFX11-NEXT:    s_or_b32 s22, s22, s2411425; GFX11-NEXT:    v_readlane_b32 s24, v37, 2611426; GFX11-NEXT:    s_lshl_b32 s25, s25, 811427; GFX11-NEXT:    s_and_b32 s10, s10, 0xff11428; GFX11-NEXT:    s_and_b32 s11, s11, 0xff11429; GFX11-NEXT:    s_and_b32 s8, s8, 0xff11430; GFX11-NEXT:    s_lshl_b32 s24, s24, 811431; GFX11-NEXT:    s_and_b32 s9, s9, 0xff11432; GFX11-NEXT:    s_or_b32 s23, s23, s2411433; GFX11-NEXT:    v_readlane_b32 s24, v37, 2511434; GFX11-NEXT:    s_and_b32 s23, s23, 0xffff11435; GFX11-NEXT:    s_and_b32 s6, s6, 0xff11436; GFX11-NEXT:    s_and_b32 s7, s7, 0xff11437; GFX11-NEXT:    s_and_b32 s4, s4, 0xff11438; GFX11-NEXT:    s_and_b32 s24, s24, 0xff11439; GFX11-NEXT:    s_and_b32 s5, s5, 0xff11440; GFX11-NEXT:    s_or_b32 s24, s24, s2511441; GFX11-NEXT:    v_readlane_b32 s25, v37, 2211442; GFX11-NEXT:    s_lshl_b32 s24, s24, 1611443; GFX11-NEXT:    s_and_b32 s2, s2, 0xff11444; GFX11-NEXT:    s_or_b32 s23, s23, s2411445; GFX11-NEXT:    v_readlane_b32 s24, v37, 2311446; GFX11-NEXT:    s_and_b32 s25, s25, 0xff11447; GFX11-NEXT:    v_dual_mov_b32 v9, s22 :: v_dual_mov_b32 v10, s2311448; GFX11-NEXT:    s_lshl_b32 s22, s88, 811449; GFX11-NEXT:    s_lshl_b32 s24, s24, 811450; GFX11-NEXT:    s_lshl_b32 s23, s97, 811451; GFX11-NEXT:    s_or_b32 s20, s20, s2411452; GFX11-NEXT:    s_lshl_b32 s24, s62, 811453; GFX11-NEXT:    s_and_b32 s20, s20, 0xffff11454; GFX11-NEXT:    s_or_b32 s24, s25, s2411455; GFX11-NEXT:    v_readlane_b32 s25, v37, 2111456; GFX11-NEXT:    s_lshl_b32 s24, s24, 1611457; GFX11-NEXT:    s_and_b32 s3, s3, 0xff11458; GFX11-NEXT:    s_or_b32 s20, s20, s2411459; GFX11-NEXT:    s_and_b32 s0, s0, 0xff11460; GFX11-NEXT:    s_lshl_b32 s25, s25, 811461; GFX11-NEXT:    s_and_b32 s1, s1, 0xff11462; GFX11-NEXT:    s_or_b32 s21, s21, s2511463; GFX11-NEXT:    v_readlane_b32 s25, v37, 2011464; GFX11-NEXT:    s_and_b32 s21, s21, 0xffff11465; GFX11-NEXT:    v_readlane_b32 s100, v35, 411466; GFX11-NEXT:    v_readlane_b32 s99, v35, 311467; GFX11-NEXT:    v_readlane_b32 s98, v35, 211468; GFX11-NEXT:    s_and_b32 s25, s25, 0xff11469; GFX11-NEXT:    v_readlane_b32 s97, v35, 111470; GFX11-NEXT:    s_or_b32 s25, s25, s2611471; GFX11-NEXT:    v_readlane_b32 s31, v34, 111472; GFX11-NEXT:    s_lshl_b32 s24, s25, 1611473; GFX11-NEXT:    v_readlane_b32 s30, v34, 011474; GFX11-NEXT:    s_or_b32 s21, s21, s2411475; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)11476; GFX11-NEXT:    v_dual_mov_b32 v11, s20 :: v_dual_mov_b32 v12, s2111477; GFX11-NEXT:    v_readlane_b32 s20, v37, 1811478; GFX11-NEXT:    v_readlane_b32 s21, v37, 1711479; GFX11-NEXT:    s_lshl_b32 s20, s20, 811480; GFX11-NEXT:    s_and_b32 s21, s21, 0xff11481; GFX11-NEXT:    s_or_b32 s18, s18, s2011482; GFX11-NEXT:    s_or_b32 s20, s21, s2211483; GFX11-NEXT:    v_readlane_b32 s21, v37, 1611484; GFX11-NEXT:    s_and_b32 s22, s69, 0xff11485; GFX11-NEXT:    s_and_b32 s18, s18, 0xffff11486; GFX11-NEXT:    s_lshl_b32 s20, s20, 1611487; GFX11-NEXT:    v_readlane_b32 s69, v34, 2111488; GFX11-NEXT:    s_lshl_b32 s21, s21, 811489; GFX11-NEXT:    s_or_b32 s18, s18, s2011490; GFX11-NEXT:    s_or_b32 s19, s19, s2111491; GFX11-NEXT:    s_or_b32 s21, s22, s2311492; GFX11-NEXT:    s_and_b32 s19, s19, 0xffff11493; GFX11-NEXT:    s_lshl_b32 s21, s21, 1611494; GFX11-NEXT:    s_lshl_b32 s20, s96, 811495; GFX11-NEXT:    s_or_b32 s19, s19, s2111496; GFX11-NEXT:    s_and_b32 s21, s73, 0xff11497; GFX11-NEXT:    s_lshl_b32 s22, s76, 811498; GFX11-NEXT:    s_or_b32 s16, s16, s2011499; GFX11-NEXT:    s_or_b32 s20, s21, s2211500; GFX11-NEXT:    s_lshl_b32 s21, s87, 811501; GFX11-NEXT:    s_and_b32 s22, s72, 0xff11502; GFX11-NEXT:    s_lshl_b32 s23, s86, 811503; GFX11-NEXT:    s_or_b32 s17, s17, s2111504; GFX11-NEXT:    s_or_b32 s21, s22, s2311505; GFX11-NEXT:    v_dual_mov_b32 v1, s18 :: v_dual_mov_b32 v2, s1911506; GFX11-NEXT:    v_readlane_b32 s18, v37, 011507; GFX11-NEXT:    s_and_b32 s16, s16, 0xffff11508; GFX11-NEXT:    s_lshl_b32 s20, s20, 1611509; GFX11-NEXT:    s_and_b32 s17, s17, 0xffff11510; GFX11-NEXT:    s_lshl_b32 s21, s21, 1611511; GFX11-NEXT:    s_or_b32 s16, s16, s2011512; GFX11-NEXT:    s_or_b32 s17, s17, s2111513; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)11514; GFX11-NEXT:    v_dual_mov_b32 v3, s16 :: v_dual_mov_b32 v4, s1711515; GFX11-NEXT:    s_lshl_b32 s16, s85, 811516; GFX11-NEXT:    s_and_b32 s17, s84, 0xff11517; GFX11-NEXT:    s_lshl_b32 s18, s18, 811518; GFX11-NEXT:    v_readlane_b32 s19, v37, 111519; GFX11-NEXT:    s_or_b32 s14, s14, s1611520; GFX11-NEXT:    s_or_b32 s16, s17, s1811521; GFX11-NEXT:    s_lshl_b32 s17, s83, 811522; GFX11-NEXT:    s_and_b32 s18, s82, 0xff11523; GFX11-NEXT:    s_lshl_b32 s19, s81, 811524; GFX11-NEXT:    s_or_b32 s15, s15, s1711525; GFX11-NEXT:    s_or_b32 s17, s18, s1911526; GFX11-NEXT:    v_readlane_b32 s18, v37, 211527; GFX11-NEXT:    s_and_b32 s14, s14, 0xffff11528; GFX11-NEXT:    s_lshl_b32 s16, s16, 1611529; GFX11-NEXT:    s_and_b32 s15, s15, 0xffff11530; GFX11-NEXT:    s_lshl_b32 s17, s17, 1611531; GFX11-NEXT:    s_or_b32 s14, s14, s1611532; GFX11-NEXT:    s_or_b32 s15, s15, s1711533; GFX11-NEXT:    s_lshl_b32 s16, s61, 811534; GFX11-NEXT:    s_and_b32 s17, s80, 0xff11535; GFX11-NEXT:    s_lshl_b32 s18, s18, 811536; GFX11-NEXT:    v_readlane_b32 s19, v37, 311537; GFX11-NEXT:    s_or_b32 s12, s12, s1611538; GFX11-NEXT:    s_or_b32 s16, s17, s1811539; GFX11-NEXT:    s_lshl_b32 s17, s60, 811540; GFX11-NEXT:    s_and_b32 s18, s71, 0xff11541; GFX11-NEXT:    s_lshl_b32 s19, s70, 811542; GFX11-NEXT:    s_or_b32 s13, s13, s1711543; GFX11-NEXT:    s_or_b32 s17, s18, s1911544; GFX11-NEXT:    v_dual_mov_b32 v5, s14 :: v_dual_mov_b32 v6, s1511545; GFX11-NEXT:    v_readlane_b32 s14, v37, 411546; GFX11-NEXT:    s_and_b32 s12, s12, 0xffff11547; GFX11-NEXT:    s_lshl_b32 s16, s16, 1611548; GFX11-NEXT:    s_and_b32 s13, s13, 0xffff11549; GFX11-NEXT:    s_lshl_b32 s17, s17, 1611550; GFX11-NEXT:    s_or_b32 s12, s12, s1611551; GFX11-NEXT:    s_or_b32 s13, s13, s1711552; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)11553; GFX11-NEXT:    v_dual_mov_b32 v7, s12 :: v_dual_mov_b32 v8, s1311554; GFX11-NEXT:    s_lshl_b32 s12, s58, 811555; GFX11-NEXT:    s_and_b32 s13, s59, 0xff11556; GFX11-NEXT:    s_lshl_b32 s14, s14, 811557; GFX11-NEXT:    v_readlane_b32 s15, v37, 511558; GFX11-NEXT:    s_or_b32 s10, s10, s1211559; GFX11-NEXT:    s_or_b32 s12, s13, s1411560; GFX11-NEXT:    s_lshl_b32 s13, s68, 811561; GFX11-NEXT:    s_and_b32 s14, s67, 0xff11562; GFX11-NEXT:    s_lshl_b32 s15, s66, 811563; GFX11-NEXT:    s_or_b32 s11, s11, s1311564; GFX11-NEXT:    s_or_b32 s13, s14, s1511565; GFX11-NEXT:    v_readlane_b32 s14, v37, 611566; GFX11-NEXT:    s_and_b32 s10, s10, 0xffff11567; GFX11-NEXT:    s_lshl_b32 s12, s12, 1611568; GFX11-NEXT:    s_and_b32 s11, s11, 0xffff11569; GFX11-NEXT:    s_lshl_b32 s13, s13, 1611570; GFX11-NEXT:    s_or_b32 s10, s10, s1211571; GFX11-NEXT:    s_or_b32 s11, s11, s1311572; GFX11-NEXT:    s_lshl_b32 s12, s65, 811573; GFX11-NEXT:    s_and_b32 s13, s64, 0xff11574; GFX11-NEXT:    s_lshl_b32 s14, s14, 811575; GFX11-NEXT:    v_readlane_b32 s15, v37, 711576; GFX11-NEXT:    s_or_b32 s8, s8, s1211577; GFX11-NEXT:    s_or_b32 s12, s13, s1411578; GFX11-NEXT:    s_lshl_b32 s13, s55, 811579; GFX11-NEXT:    s_and_b32 s14, s54, 0xff11580; GFX11-NEXT:    s_lshl_b32 s15, s53, 811581; GFX11-NEXT:    s_or_b32 s9, s9, s1311582; GFX11-NEXT:    s_or_b32 s13, s14, s1511583; GFX11-NEXT:    s_clause 0x111584; GFX11-NEXT:    scratch_store_b128 v0, v[9:12], off offset:3211585; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:4811586; GFX11-NEXT:    v_dual_mov_b32 v9, s10 :: v_dual_mov_b32 v10, s1111587; GFX11-NEXT:    v_readlane_b32 s10, v37, 811588; GFX11-NEXT:    s_and_b32 s8, s8, 0xffff11589; GFX11-NEXT:    s_lshl_b32 s12, s12, 1611590; GFX11-NEXT:    s_and_b32 s9, s9, 0xffff11591; GFX11-NEXT:    s_lshl_b32 s13, s13, 1611592; GFX11-NEXT:    s_or_b32 s8, s8, s1211593; GFX11-NEXT:    s_or_b32 s9, s9, s1311594; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)11595; GFX11-NEXT:    v_dual_mov_b32 v11, s8 :: v_dual_mov_b32 v12, s911596; GFX11-NEXT:    s_lshl_b32 s8, s52, 811597; GFX11-NEXT:    s_and_b32 s9, s51, 0xff11598; GFX11-NEXT:    s_lshl_b32 s10, s10, 811599; GFX11-NEXT:    v_readlane_b32 s11, v37, 911600; GFX11-NEXT:    s_or_b32 s6, s6, s811601; GFX11-NEXT:    s_or_b32 s8, s9, s1011602; GFX11-NEXT:    s_lshl_b32 s9, s50, 811603; GFX11-NEXT:    s_and_b32 s10, s49, 0xff11604; GFX11-NEXT:    s_lshl_b32 s11, s48, 811605; GFX11-NEXT:    s_or_b32 s7, s7, s911606; GFX11-NEXT:    s_or_b32 s9, s10, s1111607; GFX11-NEXT:    v_readlane_b32 s10, v37, 1011608; GFX11-NEXT:    s_and_b32 s6, s6, 0xffff11609; GFX11-NEXT:    s_lshl_b32 s8, s8, 1611610; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff11611; GFX11-NEXT:    s_lshl_b32 s9, s9, 1611612; GFX11-NEXT:    s_or_b32 s6, s6, s811613; GFX11-NEXT:    s_or_b32 s7, s7, s911614; GFX11-NEXT:    s_lshl_b32 s8, s39, 811615; GFX11-NEXT:    s_and_b32 s9, s38, 0xff11616; GFX11-NEXT:    s_lshl_b32 s10, s10, 811617; GFX11-NEXT:    v_readlane_b32 s11, v37, 1111618; GFX11-NEXT:    s_or_b32 s4, s4, s811619; GFX11-NEXT:    s_or_b32 s8, s9, s1011620; GFX11-NEXT:    s_lshl_b32 s9, s37, 811621; GFX11-NEXT:    s_and_b32 s10, s36, 0xff11622; GFX11-NEXT:    s_lshl_b32 s11, s35, 811623; GFX11-NEXT:    s_or_b32 s5, s5, s911624; GFX11-NEXT:    s_or_b32 s9, s10, s1111625; GFX11-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s711626; GFX11-NEXT:    v_readlane_b32 s6, v37, 1211627; GFX11-NEXT:    s_and_b32 s4, s4, 0xffff11628; GFX11-NEXT:    s_lshl_b32 s8, s8, 1611629; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff11630; GFX11-NEXT:    s_lshl_b32 s9, s9, 1611631; GFX11-NEXT:    s_or_b32 s4, s4, s811632; GFX11-NEXT:    s_or_b32 s5, s5, s911633; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)11634; GFX11-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v4, s511635; GFX11-NEXT:    s_lshl_b32 s4, s56, 811636; GFX11-NEXT:    s_and_b32 s5, s57, 0xff11637; GFX11-NEXT:    s_lshl_b32 s6, s6, 811638; GFX11-NEXT:    v_readlane_b32 s7, v37, 1311639; GFX11-NEXT:    s_or_b32 s2, s2, s411640; GFX11-NEXT:    s_or_b32 s4, s5, s611641; GFX11-NEXT:    s_lshl_b32 s5, s34, 811642; GFX11-NEXT:    s_and_b32 s6, vcc_hi, 0xff11643; GFX11-NEXT:    s_lshl_b32 s7, s46, 811644; GFX11-NEXT:    s_or_b32 s3, s3, s511645; GFX11-NEXT:    s_or_b32 s5, s6, s711646; GFX11-NEXT:    v_readlane_b32 s6, v37, 1411647; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff11648; GFX11-NEXT:    s_lshl_b32 s4, s4, 1611649; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff11650; GFX11-NEXT:    s_lshl_b32 s5, s5, 1611651; GFX11-NEXT:    s_or_b32 s2, s2, s411652; GFX11-NEXT:    s_or_b32 s3, s3, s511653; GFX11-NEXT:    s_lshl_b32 s4, s47, 811654; GFX11-NEXT:    s_and_b32 s5, s104, 0xff11655; GFX11-NEXT:    s_lshl_b32 s6, s6, 811656; GFX11-NEXT:    v_readlane_b32 s7, v37, 1511657; GFX11-NEXT:    s_or_b32 s0, s0, s411658; GFX11-NEXT:    s_or_b32 s4, s5, s611659; GFX11-NEXT:    s_lshl_b32 s5, s103, 811660; GFX11-NEXT:    s_and_b32 s6, s102, 0xff11661; GFX11-NEXT:    s_lshl_b32 s7, s101, 811662; GFX11-NEXT:    s_or_b32 s1, s1, s511663; GFX11-NEXT:    s_or_b32 s5, s6, s711664; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff11665; GFX11-NEXT:    s_lshl_b32 s4, s4, 1611666; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff11667; GFX11-NEXT:    s_lshl_b32 s5, s5, 1611668; GFX11-NEXT:    s_or_b32 s0, s0, s411669; GFX11-NEXT:    s_or_b32 s1, s1, s511670; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:6411671; GFX11-NEXT:    v_dual_mov_b32 v5, s2 :: v_dual_mov_b32 v6, s311672; GFX11-NEXT:    v_dual_mov_b32 v7, s0 :: v_dual_mov_b32 v8, s111673; GFX11-NEXT:    s_clause 0x211674; GFX11-NEXT:    scratch_store_b128 v0, v[9:12], off offset:8011675; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:9611676; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:11211677; GFX11-NEXT:    v_readlane_b32 s104, v35, 811678; GFX11-NEXT:    v_readlane_b32 s103, v35, 711679; GFX11-NEXT:    v_readlane_b32 s102, v35, 611680; GFX11-NEXT:    v_readlane_b32 s101, v35, 511681; GFX11-NEXT:    v_readlane_b32 s96, v35, 011682; GFX11-NEXT:    v_readlane_b32 s87, v34, 3111683; GFX11-NEXT:    v_readlane_b32 s86, v34, 3011684; GFX11-NEXT:    v_readlane_b32 s85, v34, 2911685; GFX11-NEXT:    v_readlane_b32 s84, v34, 2811686; GFX11-NEXT:    v_readlane_b32 s83, v34, 2711687; GFX11-NEXT:    v_readlane_b32 s82, v34, 2611688; GFX11-NEXT:    v_readlane_b32 s81, v34, 2511689; GFX11-NEXT:    v_readlane_b32 s80, v34, 2411690; GFX11-NEXT:    v_readlane_b32 s71, v34, 2311691; GFX11-NEXT:    v_readlane_b32 s70, v34, 2211692; GFX11-NEXT:    v_readlane_b32 s68, v34, 2011693; GFX11-NEXT:    v_readlane_b32 s67, v34, 1911694; GFX11-NEXT:    v_readlane_b32 s66, v34, 1811695; GFX11-NEXT:    v_readlane_b32 s65, v34, 1711696; GFX11-NEXT:    v_readlane_b32 s64, v34, 1611697; GFX11-NEXT:    v_readlane_b32 s55, v34, 1511698; GFX11-NEXT:    v_readlane_b32 s54, v34, 1411699; GFX11-NEXT:    v_readlane_b32 s53, v34, 1311700; GFX11-NEXT:    v_readlane_b32 s52, v34, 1211701; GFX11-NEXT:    v_readlane_b32 s51, v34, 1111702; GFX11-NEXT:    v_readlane_b32 s50, v34, 1011703; GFX11-NEXT:    v_readlane_b32 s49, v34, 911704; GFX11-NEXT:    v_readlane_b32 s48, v34, 811705; GFX11-NEXT:    v_readlane_b32 s39, v34, 711706; GFX11-NEXT:    v_readlane_b32 s38, v34, 611707; GFX11-NEXT:    v_readlane_b32 s37, v34, 511708; GFX11-NEXT:    v_readlane_b32 s36, v34, 411709; GFX11-NEXT:    v_readlane_b32 s35, v34, 311710; GFX11-NEXT:    v_readlane_b32 s34, v34, 211711; GFX11-NEXT:    s_xor_saveexec_b32 s0, -111712; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Reload11713; GFX11-NEXT:    scratch_load_b32 v34, off, s3211714; GFX11-NEXT:    scratch_load_b32 v35, off, s32 offset:411715; GFX11-NEXT:    scratch_load_b32 v36, off, s32 offset:811716; GFX11-NEXT:    scratch_load_b32 v37, off, s32 offset:1211717; GFX11-NEXT:    s_mov_b32 exec_lo, s011718; GFX11-NEXT:    s_waitcnt vmcnt(0)11719; GFX11-NEXT:    s_setpc_b64 s[30:31]11720  %cmp = icmp eq i32 %b, 011721  br i1 %cmp, label %cmp.true, label %cmp.false11722 11723cmp.true:11724  %a1 = add <32 x i32> %a, splat (i32 3)11725  %a2 = bitcast <32 x i32> %a1 to <128 x i8>11726  br label %end11727 11728cmp.false:11729  %a3 = bitcast <32 x i32> %a to <128 x i8>11730  br label %end11731 11732end:11733  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]11734  ret <128 x i8> %phi11735}11736 11737define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {11738; SI-LABEL: bitcast_v128i8_to_v32i32:11739; SI:       ; %bb.0:11740; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)11741; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill11742; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill11743; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill11744; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill11745; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill11746; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill11747; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill11748; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill11749; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill11750; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill11751; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill11752; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill11753; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill11754; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill11755; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill11756; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill11757; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill11758; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill11759; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill11760; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill11761; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill11762; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill11763; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill11764; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill11765; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill11766; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill11767; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill11768; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill11769; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill11770; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill11771; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill11772; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill11773; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:38811774; SI-NEXT:    s_waitcnt expcnt(0)11775; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s3211776; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:811777; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:1611778; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:2411779; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:3211780; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:4011781; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:4811782; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:5611783; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:6411784; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:7211785; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:8011786; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:8811787; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:9611788; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:10411789; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:11211790; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:12011791; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:12811792; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:13611793; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:14411794; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:15211795; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:16011796; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:16811797; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:17611798; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:18411799; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v111800; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v311801; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v511802; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v711803; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v911804; SI-NEXT:    v_lshlrev_b32_e32 v34, 24, v1111805; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v1311806; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v1511807; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v1711808; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v1911809; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v2111810; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v2311811; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v2511812; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v2711813; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v2911814; SI-NEXT:    s_waitcnt vmcnt(14)11815; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3011816; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v011817; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v211818; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v411819; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill11820; SI-NEXT:    s_waitcnt expcnt(0)11821; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v611822; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill11823; SI-NEXT:    s_waitcnt expcnt(0)11824; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v811825; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill11826; SI-NEXT:    s_waitcnt expcnt(0)11827; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1011828; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill11829; SI-NEXT:    s_waitcnt expcnt(0)11830; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v1211831; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill11832; SI-NEXT:    s_waitcnt expcnt(0)11833; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1411834; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill11835; SI-NEXT:    s_waitcnt expcnt(0)11836; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v1611837; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill11838; SI-NEXT:    s_waitcnt expcnt(0)11839; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1811840; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill11841; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)11842; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2011843; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill11844; SI-NEXT:    s_waitcnt expcnt(0)11845; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2211846; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill11847; SI-NEXT:    s_waitcnt expcnt(0)11848; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2411849; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill11850; SI-NEXT:    s_waitcnt expcnt(0)11851; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2611852; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill11853; SI-NEXT:    s_waitcnt expcnt(0)11854; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2811855; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill11856; SI-NEXT:    s_waitcnt expcnt(0)11857; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:11611858; SI-NEXT:    s_waitcnt vmcnt(0)11859; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill11860; SI-NEXT:    s_waitcnt expcnt(0)11861; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3111862; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill11863; SI-NEXT:    s_waitcnt expcnt(0)11864; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:12411865; SI-NEXT:    s_waitcnt vmcnt(0)11866; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill11867; SI-NEXT:    s_waitcnt expcnt(0)11868; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:13211869; SI-NEXT:    s_waitcnt vmcnt(0)11870; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill11871; SI-NEXT:    s_waitcnt expcnt(0)11872; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:14011873; SI-NEXT:    s_waitcnt vmcnt(0)11874; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill11875; SI-NEXT:    s_waitcnt expcnt(0)11876; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:14811877; SI-NEXT:    s_waitcnt vmcnt(0)11878; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill11879; SI-NEXT:    s_waitcnt expcnt(0)11880; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3311881; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill11882; SI-NEXT:    s_waitcnt expcnt(0)11883; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3511884; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill11885; SI-NEXT:    s_waitcnt expcnt(0)11886; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3611887; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill11888; SI-NEXT:    s_waitcnt expcnt(0)11889; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3911890; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill11891; SI-NEXT:    s_waitcnt expcnt(0)11892; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:15611893; SI-NEXT:    s_waitcnt vmcnt(0)11894; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill11895; SI-NEXT:    s_waitcnt expcnt(0)11896; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16411897; SI-NEXT:    s_waitcnt vmcnt(0)11898; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill11899; SI-NEXT:    s_waitcnt expcnt(0)11900; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:17211901; SI-NEXT:    s_waitcnt vmcnt(0)11902; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill11903; SI-NEXT:    s_waitcnt expcnt(0)11904; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:18011905; SI-NEXT:    s_waitcnt vmcnt(0)11906; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill11907; SI-NEXT:    s_waitcnt expcnt(0)11908; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4911909; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill11910; SI-NEXT:    s_waitcnt expcnt(0)11911; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5011912; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill11913; SI-NEXT:    s_waitcnt expcnt(0)11914; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v5111915; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill11916; SI-NEXT:    s_waitcnt expcnt(0)11917; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5211918; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill11919; SI-NEXT:    s_waitcnt expcnt(0)11920; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:19211921; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:20011922; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:20811923; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:21611924; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:18811925; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:4411926; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:3611927; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:2811928; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:2011929; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:1211930; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:411931; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:10811932; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:10011933; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:9211934; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:8411935; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:7611936; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:6811937; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:6011938; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:5211939; SI-NEXT:    s_waitcnt vmcnt(14)11940; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v011941; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill11942; SI-NEXT:    s_waitcnt expcnt(0)11943; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v111944; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill11945; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill11946; SI-NEXT:    s_waitcnt expcnt(0)11947; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:19611948; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v211949; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill11950; SI-NEXT:    s_waitcnt expcnt(0)11951; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v311952; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill11953; SI-NEXT:    s_waitcnt vmcnt(2)11954; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill11955; SI-NEXT:    s_waitcnt expcnt(0)11956; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:20411957; SI-NEXT:    s_waitcnt vmcnt(0)11958; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill11959; SI-NEXT:    s_waitcnt expcnt(0)11960; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:21211961; SI-NEXT:    s_waitcnt vmcnt(0)11962; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill11963; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:22411964; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:23211965; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24011966; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:24811967; SI-NEXT:    s_waitcnt expcnt(0)11968; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:22011969; SI-NEXT:    s_waitcnt vmcnt(4)11970; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v011971; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill11972; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)11973; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v111974; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill11975; SI-NEXT:    s_waitcnt vmcnt(2)11976; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill11977; SI-NEXT:    s_waitcnt expcnt(0)11978; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:22811979; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v211980; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill11981; SI-NEXT:    s_waitcnt expcnt(0)11982; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v311983; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill11984; SI-NEXT:    s_waitcnt vmcnt(2)11985; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill11986; SI-NEXT:    s_waitcnt expcnt(0)11987; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:23611988; SI-NEXT:    s_waitcnt vmcnt(0)11989; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill11990; SI-NEXT:    s_waitcnt expcnt(0)11991; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:24411992; SI-NEXT:    s_waitcnt vmcnt(0)11993; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill11994; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:25611995; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:26411996; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:27211997; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:28011998; SI-NEXT:    s_waitcnt expcnt(0)11999; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:25212000; SI-NEXT:    s_waitcnt vmcnt(4)12001; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v012002; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill12003; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)12004; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v112005; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill12006; SI-NEXT:    s_waitcnt vmcnt(2)12007; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill12008; SI-NEXT:    s_waitcnt expcnt(0)12009; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:26012010; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v212011; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill12012; SI-NEXT:    s_waitcnt expcnt(0)12013; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v312014; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill12015; SI-NEXT:    s_waitcnt vmcnt(2)12016; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill12017; SI-NEXT:    s_waitcnt expcnt(0)12018; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:26812019; SI-NEXT:    s_waitcnt vmcnt(0)12020; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill12021; SI-NEXT:    s_waitcnt expcnt(0)12022; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:27612023; SI-NEXT:    s_waitcnt vmcnt(0)12024; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill12025; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28812026; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:29612027; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:30412028; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:31212029; SI-NEXT:    s_waitcnt expcnt(0)12030; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:28412031; SI-NEXT:    s_waitcnt vmcnt(4)12032; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v012033; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill12034; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)12035; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v112036; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill12037; SI-NEXT:    s_waitcnt vmcnt(2)12038; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill12039; SI-NEXT:    s_waitcnt expcnt(0)12040; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:29212041; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v212042; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill12043; SI-NEXT:    s_waitcnt expcnt(0)12044; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v312045; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill12046; SI-NEXT:    s_waitcnt vmcnt(2)12047; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill12048; SI-NEXT:    s_waitcnt expcnt(0)12049; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:30012050; SI-NEXT:    s_waitcnt vmcnt(0)12051; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill12052; SI-NEXT:    s_waitcnt expcnt(0)12053; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:30812054; SI-NEXT:    s_waitcnt vmcnt(0)12055; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill12056; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32012057; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:32812058; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:33612059; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:34412060; SI-NEXT:    s_waitcnt expcnt(0)12061; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:31612062; SI-NEXT:    s_waitcnt vmcnt(4)12063; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v012064; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill12065; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)12066; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v112067; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill12068; SI-NEXT:    s_waitcnt vmcnt(2)12069; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill12070; SI-NEXT:    s_waitcnt expcnt(0)12071; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32412072; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v212073; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill12074; SI-NEXT:    s_waitcnt expcnt(0)12075; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v312076; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill12077; SI-NEXT:    s_waitcnt vmcnt(2)12078; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill12079; SI-NEXT:    s_waitcnt expcnt(0)12080; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:33212081; SI-NEXT:    s_waitcnt vmcnt(0)12082; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill12083; SI-NEXT:    s_waitcnt expcnt(0)12084; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:34012085; SI-NEXT:    s_waitcnt vmcnt(0)12086; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill12087; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:35212088; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:36012089; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36812090; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:37612091; SI-NEXT:    s_waitcnt expcnt(0)12092; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:34812093; SI-NEXT:    s_waitcnt vmcnt(4)12094; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v012095; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill12096; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)12097; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v112098; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill12099; SI-NEXT:    s_waitcnt vmcnt(2)12100; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill12101; SI-NEXT:    s_waitcnt expcnt(0)12102; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:35612103; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v212104; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill12105; SI-NEXT:    s_waitcnt expcnt(0)12106; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:38412107; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:38012108; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v312109; SI-NEXT:    s_waitcnt vmcnt(3)12110; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill12111; SI-NEXT:    s_waitcnt expcnt(0)12112; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:36412113; SI-NEXT:    s_waitcnt vmcnt(3)12114; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v012115; SI-NEXT:    s_waitcnt vmcnt(2)12116; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill12117; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill12118; SI-NEXT:    s_waitcnt vmcnt(2)12119; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill12120; SI-NEXT:    s_waitcnt expcnt(0)12121; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:37212122; SI-NEXT:    s_waitcnt vmcnt(0)12123; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill12124; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr3112125; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc12126; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]12127; SI-NEXT:    s_cbranch_execz .LBB14_212128; SI-NEXT:  ; %bb.1: ; %cmp.false12129; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload12130; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload12131; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload12132; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload12133; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload12134; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload12135; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload12136; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload12137; SI-NEXT:    v_and_b32_e32 v9, 0xff, v4912138; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v912139; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload12140; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload12141; SI-NEXT:    s_waitcnt expcnt(0)12142; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload12143; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload12144; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload12145; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload12146; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload12147; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload12148; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload12149; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload12150; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload12151; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload12152; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload12153; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload12154; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload12155; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload12156; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload12157; SI-NEXT:    ; implicit-def: $vgpr4912158; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload12159; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload12160; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload12161; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload12162; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload12163; SI-NEXT:    s_waitcnt vmcnt(14)12164; SI-NEXT:    v_and_b32_e32 v0, 0xff, v012165; SI-NEXT:    v_and_b32_e32 v2, 0xff, v212166; SI-NEXT:    v_and_b32_e32 v3, 0xff, v312167; SI-NEXT:    v_or_b32_e32 v0, v0, v5412168; SI-NEXT:    v_or_b32_e32 v2, v2, v4112169; SI-NEXT:    v_and_b32_e32 v1, 0xff, v112170; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v112171; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v312172; SI-NEXT:    v_or_b32_e32 v1, v42, v112173; SI-NEXT:    v_or_b32_e32 v3, v40, v312174; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v012175; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v212176; SI-NEXT:    v_or_b32_e32 v0, v0, v112177; SI-NEXT:    v_or_b32_e32 v1, v2, v312178; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload12179; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload12180; SI-NEXT:    v_or_b32_e32 v9, v10, v912181; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload12182; SI-NEXT:    v_and_b32_e32 v4, 0xff, v412183; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v412184; SI-NEXT:    v_or_b32_e32 v4, v32, v412185; SI-NEXT:    v_and_b32_e32 v5, 0xff, v512186; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v512187; SI-NEXT:    v_or_b32_e32 v5, v60, v512188; SI-NEXT:    s_waitcnt vmcnt(14)12189; SI-NEXT:    v_and_b32_e32 v6, 0xff, v612190; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v612191; SI-NEXT:    v_or_b32_e32 v6, v58, v612192; SI-NEXT:    s_waitcnt vmcnt(10)12193; SI-NEXT:    v_and_b32_e32 v7, 0xff, v712194; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v712195; SI-NEXT:    v_or_b32_e32 v7, v37, v712196; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload12197; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload12198; SI-NEXT:    ; implicit-def: $vgpr5412199; SI-NEXT:    ; implicit-def: $vgpr4212200; SI-NEXT:    ; implicit-def: $vgpr4112201; SI-NEXT:    ; implicit-def: $vgpr4012202; SI-NEXT:    ; implicit-def: $vgpr6012203; SI-NEXT:    ; implicit-def: $vgpr5812204; SI-NEXT:    ; implicit-def: $vgpr3712205; SI-NEXT:    s_waitcnt vmcnt(4)12206; SI-NEXT:    v_and_b32_e32 v2, 0xff, v212207; SI-NEXT:    s_waitcnt vmcnt(3)12208; SI-NEXT:    v_and_b32_e32 v3, 0xff, v312209; SI-NEXT:    v_or_b32_e32 v2, v2, v5312210; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v312211; SI-NEXT:    v_or_b32_e32 v3, v34, v312212; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v212213; SI-NEXT:    v_or_b32_e32 v2, v2, v312214; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload12215; SI-NEXT:    ; implicit-def: $vgpr5312216; SI-NEXT:    ; implicit-def: $vgpr3412217; SI-NEXT:    s_waitcnt vmcnt(1)12218; SI-NEXT:    v_and_b32_e32 v8, 0xff, v812219; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v812220; SI-NEXT:    v_or_b32_e32 v8, v55, v812221; SI-NEXT:    ; implicit-def: $vgpr5512222; SI-NEXT:    s_waitcnt vmcnt(0)12223; SI-NEXT:    v_and_b32_e32 v3, 0xff, v312224; SI-NEXT:    v_or_b32_e32 v3, v3, v4812225; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v312226; SI-NEXT:    v_or_b32_e32 v3, v3, v412227; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload12228; SI-NEXT:    ; implicit-def: $vgpr4812229; SI-NEXT:    s_waitcnt vmcnt(0)12230; SI-NEXT:    v_and_b32_e32 v4, 0xff, v412231; SI-NEXT:    v_or_b32_e32 v4, v4, v6212232; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v412233; SI-NEXT:    v_or_b32_e32 v4, v4, v512234; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload12235; SI-NEXT:    ; implicit-def: $vgpr6212236; SI-NEXT:    s_waitcnt vmcnt(0)12237; SI-NEXT:    v_and_b32_e32 v5, 0xff, v512238; SI-NEXT:    v_or_b32_e32 v5, v5, v3812239; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v512240; SI-NEXT:    v_or_b32_e32 v5, v5, v612241; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload12242; SI-NEXT:    ; implicit-def: $vgpr3812243; SI-NEXT:    s_waitcnt vmcnt(0)12244; SI-NEXT:    v_and_b32_e32 v6, 0xff, v612245; SI-NEXT:    v_or_b32_e32 v6, v6, v5612246; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v612247; SI-NEXT:    v_or_b32_e32 v6, v6, v712248; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload12249; SI-NEXT:    ; implicit-def: $vgpr5612250; SI-NEXT:    s_waitcnt vmcnt(0)12251; SI-NEXT:    v_and_b32_e32 v7, 0xff, v712252; SI-NEXT:    v_or_b32_e32 v7, v7, v4612253; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v712254; SI-NEXT:    v_or_b32_e32 v7, v7, v812255; SI-NEXT:    v_and_b32_e32 v8, 0xff, v3512256; SI-NEXT:    v_or_b32_e32 v8, v8, v4412257; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v812258; SI-NEXT:    v_or_b32_e32 v8, v8, v912259; SI-NEXT:    v_and_b32_e32 v9, 0xff, v3312260; SI-NEXT:    v_or_b32_e32 v9, v9, v1012261; SI-NEXT:    v_and_b32_e32 v10, 0xff, v5212262; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v1012263; SI-NEXT:    v_or_b32_e32 v10, v11, v1012264; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload12265; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v912266; SI-NEXT:    v_or_b32_e32 v9, v9, v1012267; SI-NEXT:    v_and_b32_e32 v10, 0xff, v6312268; SI-NEXT:    ; implicit-def: $vgpr3512269; SI-NEXT:    ; implicit-def: $vgpr5212270; SI-NEXT:    ; implicit-def: $vgpr6312271; SI-NEXT:    ; implicit-def: $vgpr4612272; SI-NEXT:    ; implicit-def: $vgpr4412273; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload12274; SI-NEXT:    s_waitcnt vmcnt(1)12275; SI-NEXT:    v_or_b32_e32 v10, v10, v1112276; SI-NEXT:    v_and_b32_e32 v11, 0xff, v3912277; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v1112278; SI-NEXT:    v_or_b32_e32 v11, v12, v1112279; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload12280; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v1012281; SI-NEXT:    v_or_b32_e32 v10, v10, v1112282; SI-NEXT:    v_and_b32_e32 v11, 0xff, v6112283; SI-NEXT:    ; implicit-def: $vgpr3912284; SI-NEXT:    ; implicit-def: $vgpr6112285; SI-NEXT:    s_waitcnt vmcnt(0)12286; SI-NEXT:    v_or_b32_e32 v11, v11, v1212287; SI-NEXT:    v_and_b32_e32 v12, 0xff, v5912288; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v1212289; SI-NEXT:    v_or_b32_e32 v12, v13, v1212290; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload12291; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v1112292; SI-NEXT:    v_or_b32_e32 v11, v11, v1212293; SI-NEXT:    v_and_b32_e32 v12, 0xff, v5112294; SI-NEXT:    ; implicit-def: $vgpr5912295; SI-NEXT:    ; implicit-def: $vgpr5112296; SI-NEXT:    s_waitcnt vmcnt(0)12297; SI-NEXT:    v_or_b32_e32 v12, v12, v1312298; SI-NEXT:    v_and_b32_e32 v13, 0xff, v5712299; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v1312300; SI-NEXT:    v_or_b32_e32 v13, v14, v1312301; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload12302; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v1212303; SI-NEXT:    v_or_b32_e32 v12, v12, v1312304; SI-NEXT:    v_and_b32_e32 v13, 0xff, v4712305; SI-NEXT:    ; implicit-def: $vgpr5712306; SI-NEXT:    ; implicit-def: $vgpr4712307; SI-NEXT:    s_waitcnt vmcnt(0)12308; SI-NEXT:    v_or_b32_e32 v13, v13, v1412309; SI-NEXT:    v_and_b32_e32 v14, 0xff, v5012310; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v1412311; SI-NEXT:    v_or_b32_e32 v14, v15, v1412312; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload12313; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v1312314; SI-NEXT:    v_or_b32_e32 v13, v13, v1412315; SI-NEXT:    v_and_b32_e32 v14, 0xff, v4512316; SI-NEXT:    ; implicit-def: $vgpr5012317; SI-NEXT:    ; implicit-def: $vgpr4512318; SI-NEXT:    s_waitcnt vmcnt(0)12319; SI-NEXT:    v_or_b32_e32 v14, v14, v1512320; SI-NEXT:    v_and_b32_e32 v15, 0xff, v3612321; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v1512322; SI-NEXT:    v_or_b32_e32 v15, v16, v1512323; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v1412324; SI-NEXT:    v_or_b32_e32 v14, v14, v1512325; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload12326; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload12327; SI-NEXT:    ; implicit-def: $vgpr3612328; SI-NEXT:    s_waitcnt vmcnt(1)12329; SI-NEXT:    v_and_b32_e32 v15, 0xff, v1512330; SI-NEXT:    s_waitcnt vmcnt(0)12331; SI-NEXT:    v_or_b32_e32 v15, v15, v1612332; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload12333; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v1512334; SI-NEXT:    s_waitcnt vmcnt(0)12335; SI-NEXT:    v_and_b32_e32 v16, 0xff, v1612336; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v1612337; SI-NEXT:    v_or_b32_e32 v16, v17, v1612338; SI-NEXT:    v_or_b32_e32 v15, v15, v1612339; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload12340; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload12341; SI-NEXT:    s_waitcnt vmcnt(1)12342; SI-NEXT:    v_and_b32_e32 v16, 0xff, v1612343; SI-NEXT:    s_waitcnt vmcnt(0)12344; SI-NEXT:    v_or_b32_e32 v16, v16, v1712345; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload12346; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v1612347; SI-NEXT:    s_waitcnt vmcnt(0)12348; SI-NEXT:    v_and_b32_e32 v17, 0xff, v1712349; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v1712350; SI-NEXT:    v_or_b32_e32 v17, v18, v1712351; SI-NEXT:    v_or_b32_e32 v16, v16, v1712352; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload12353; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload12354; SI-NEXT:    s_waitcnt vmcnt(1)12355; SI-NEXT:    v_and_b32_e32 v17, 0xff, v1712356; SI-NEXT:    s_waitcnt vmcnt(0)12357; SI-NEXT:    v_or_b32_e32 v17, v17, v1812358; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload12359; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v1712360; SI-NEXT:    s_waitcnt vmcnt(0)12361; SI-NEXT:    v_and_b32_e32 v18, 0xff, v1812362; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v1812363; SI-NEXT:    v_or_b32_e32 v18, v19, v1812364; SI-NEXT:    v_or_b32_e32 v17, v17, v1812365; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload12366; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload12367; SI-NEXT:    s_waitcnt vmcnt(1)12368; SI-NEXT:    v_and_b32_e32 v18, 0xff, v1812369; SI-NEXT:    s_waitcnt vmcnt(0)12370; SI-NEXT:    v_or_b32_e32 v18, v18, v1912371; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload12372; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v1812373; SI-NEXT:    s_waitcnt vmcnt(0)12374; SI-NEXT:    v_and_b32_e32 v19, 0xff, v1912375; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v1912376; SI-NEXT:    v_or_b32_e32 v19, v20, v1912377; SI-NEXT:    v_or_b32_e32 v18, v18, v1912378; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload12379; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload12380; SI-NEXT:    s_waitcnt vmcnt(1)12381; SI-NEXT:    v_and_b32_e32 v19, 0xff, v1912382; SI-NEXT:    s_waitcnt vmcnt(0)12383; SI-NEXT:    v_or_b32_e32 v19, v19, v2012384; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload12385; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v1912386; SI-NEXT:    s_waitcnt vmcnt(0)12387; SI-NEXT:    v_and_b32_e32 v20, 0xff, v2012388; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v2012389; SI-NEXT:    v_or_b32_e32 v20, v21, v2012390; SI-NEXT:    v_or_b32_e32 v19, v19, v2012391; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload12392; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload12393; SI-NEXT:    s_waitcnt vmcnt(1)12394; SI-NEXT:    v_and_b32_e32 v20, 0xff, v2012395; SI-NEXT:    s_waitcnt vmcnt(0)12396; SI-NEXT:    v_or_b32_e32 v20, v20, v2112397; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload12398; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v2012399; SI-NEXT:    s_waitcnt vmcnt(0)12400; SI-NEXT:    v_and_b32_e32 v21, 0xff, v2112401; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v2112402; SI-NEXT:    v_or_b32_e32 v21, v22, v2112403; SI-NEXT:    v_or_b32_e32 v20, v20, v2112404; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload12405; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload12406; SI-NEXT:    s_waitcnt vmcnt(1)12407; SI-NEXT:    v_and_b32_e32 v21, 0xff, v2112408; SI-NEXT:    s_waitcnt vmcnt(0)12409; SI-NEXT:    v_or_b32_e32 v21, v21, v2212410; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload12411; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v2112412; SI-NEXT:    s_waitcnt vmcnt(0)12413; SI-NEXT:    v_and_b32_e32 v22, 0xff, v2212414; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v2212415; SI-NEXT:    v_or_b32_e32 v22, v23, v2212416; SI-NEXT:    v_or_b32_e32 v21, v21, v2212417; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload12418; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload12419; SI-NEXT:    s_waitcnt vmcnt(1)12420; SI-NEXT:    v_and_b32_e32 v22, 0xff, v2212421; SI-NEXT:    s_waitcnt vmcnt(0)12422; SI-NEXT:    v_or_b32_e32 v22, v22, v2312423; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload12424; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v2212425; SI-NEXT:    s_waitcnt vmcnt(0)12426; SI-NEXT:    v_and_b32_e32 v23, 0xff, v2312427; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v2312428; SI-NEXT:    v_or_b32_e32 v23, v24, v2312429; SI-NEXT:    v_or_b32_e32 v22, v22, v2312430; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload12431; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload12432; SI-NEXT:    s_waitcnt vmcnt(1)12433; SI-NEXT:    v_and_b32_e32 v23, 0xff, v2312434; SI-NEXT:    s_waitcnt vmcnt(0)12435; SI-NEXT:    v_or_b32_e32 v23, v23, v2412436; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload12437; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v2312438; SI-NEXT:    s_waitcnt vmcnt(0)12439; SI-NEXT:    v_and_b32_e32 v24, 0xff, v2412440; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v2412441; SI-NEXT:    v_or_b32_e32 v24, v25, v2412442; SI-NEXT:    v_or_b32_e32 v23, v23, v2412443; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload12444; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload12445; SI-NEXT:    s_waitcnt vmcnt(1)12446; SI-NEXT:    v_and_b32_e32 v24, 0xff, v2412447; SI-NEXT:    s_waitcnt vmcnt(0)12448; SI-NEXT:    v_or_b32_e32 v24, v24, v2512449; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload12450; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v2412451; SI-NEXT:    s_waitcnt vmcnt(0)12452; SI-NEXT:    v_and_b32_e32 v25, 0xff, v2512453; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v2512454; SI-NEXT:    v_or_b32_e32 v25, v26, v2512455; SI-NEXT:    v_or_b32_e32 v24, v24, v2512456; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload12457; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload12458; SI-NEXT:    s_waitcnt vmcnt(1)12459; SI-NEXT:    v_and_b32_e32 v25, 0xff, v2512460; SI-NEXT:    s_waitcnt vmcnt(0)12461; SI-NEXT:    v_or_b32_e32 v25, v25, v2612462; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload12463; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v2512464; SI-NEXT:    s_waitcnt vmcnt(0)12465; SI-NEXT:    v_and_b32_e32 v26, 0xff, v2612466; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v2612467; SI-NEXT:    v_or_b32_e32 v26, v27, v2612468; SI-NEXT:    v_or_b32_e32 v25, v25, v2612469; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload12470; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload12471; SI-NEXT:    s_waitcnt vmcnt(1)12472; SI-NEXT:    v_and_b32_e32 v26, 0xff, v2612473; SI-NEXT:    s_waitcnt vmcnt(0)12474; SI-NEXT:    v_or_b32_e32 v26, v26, v2712475; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload12476; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v2612477; SI-NEXT:    s_waitcnt vmcnt(0)12478; SI-NEXT:    v_and_b32_e32 v27, 0xff, v2712479; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v2712480; SI-NEXT:    v_or_b32_e32 v27, v28, v2712481; SI-NEXT:    v_or_b32_e32 v26, v26, v2712482; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload12483; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload12484; SI-NEXT:    s_waitcnt vmcnt(1)12485; SI-NEXT:    v_and_b32_e32 v27, 0xff, v2712486; SI-NEXT:    s_waitcnt vmcnt(0)12487; SI-NEXT:    v_or_b32_e32 v27, v27, v2812488; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload12489; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v2712490; SI-NEXT:    s_waitcnt vmcnt(0)12491; SI-NEXT:    v_and_b32_e32 v28, 0xff, v2812492; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v2812493; SI-NEXT:    v_or_b32_e32 v28, v29, v2812494; SI-NEXT:    v_or_b32_e32 v27, v27, v2812495; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload12496; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload12497; SI-NEXT:    s_waitcnt vmcnt(1)12498; SI-NEXT:    v_and_b32_e32 v28, 0xff, v2812499; SI-NEXT:    s_waitcnt vmcnt(0)12500; SI-NEXT:    v_or_b32_e32 v28, v28, v2912501; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload12502; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v2812503; SI-NEXT:    s_waitcnt vmcnt(0)12504; SI-NEXT:    v_and_b32_e32 v29, 0xff, v2912505; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v2912506; SI-NEXT:    v_or_b32_e32 v29, v30, v2912507; SI-NEXT:    v_or_b32_e32 v28, v28, v2912508; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload12509; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload12510; SI-NEXT:    s_waitcnt vmcnt(1)12511; SI-NEXT:    v_and_b32_e32 v29, 0xff, v2912512; SI-NEXT:    s_waitcnt vmcnt(0)12513; SI-NEXT:    v_or_b32_e32 v29, v29, v3012514; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload12515; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v2912516; SI-NEXT:    s_waitcnt vmcnt(0)12517; SI-NEXT:    v_and_b32_e32 v30, 0xff, v3012518; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v3012519; SI-NEXT:    v_or_b32_e32 v30, v31, v3012520; SI-NEXT:    v_or_b32_e32 v29, v29, v3012521; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload12522; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload12523; SI-NEXT:    s_waitcnt vmcnt(1)12524; SI-NEXT:    v_and_b32_e32 v30, 0xff, v3012525; SI-NEXT:    s_waitcnt vmcnt(0)12526; SI-NEXT:    v_or_b32_e32 v30, v30, v3112527; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload12528; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v3012529; SI-NEXT:    s_waitcnt vmcnt(0)12530; SI-NEXT:    v_and_b32_e32 v31, 0xff, v3112531; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v3112532; SI-NEXT:    v_or_b32_e32 v31, v32, v3112533; SI-NEXT:    v_or_b32_e32 v30, v30, v3112534; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload12535; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload12536; SI-NEXT:    s_waitcnt vmcnt(1)12537; SI-NEXT:    v_and_b32_e32 v31, 0xff, v3112538; SI-NEXT:    s_waitcnt vmcnt(0)12539; SI-NEXT:    v_and_b32_e32 v32, 0xff, v3212540; SI-NEXT:    v_or_b32_e32 v31, v31, v4312541; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v3212542; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v3112543; SI-NEXT:    v_or_b32_e32 v32, v33, v3212544; SI-NEXT:    v_or_b32_e32 v31, v31, v3212545; SI-NEXT:    ; implicit-def: $vgpr3212546; SI-NEXT:    ; kill: killed $vgpr3212547; SI-NEXT:    ; implicit-def: $vgpr3212548; SI-NEXT:    ; kill: killed $vgpr3212549; SI-NEXT:    ; implicit-def: $vgpr3212550; SI-NEXT:    ; kill: killed $vgpr3212551; SI-NEXT:    ; implicit-def: $vgpr3212552; SI-NEXT:    ; kill: killed $vgpr3212553; SI-NEXT:    ; implicit-def: $vgpr3212554; SI-NEXT:    ; implicit-def: $vgpr4312555; SI-NEXT:    ; kill: killed $vgpr3212556; SI-NEXT:    ; implicit-def: $vgpr3212557; SI-NEXT:    ; kill: killed $vgpr4312558; SI-NEXT:    ; implicit-def: $vgpr4312559; SI-NEXT:    ; kill: killed $vgpr3212560; SI-NEXT:    ; implicit-def: $vgpr3212561; SI-NEXT:    ; kill: killed $vgpr4312562; SI-NEXT:    ; implicit-def: $vgpr4312563; SI-NEXT:    ; kill: killed $vgpr3212564; SI-NEXT:    ; implicit-def: $vgpr3212565; SI-NEXT:    ; kill: killed $vgpr4312566; SI-NEXT:    ; implicit-def: $vgpr4312567; SI-NEXT:    ; kill: killed $vgpr3212568; SI-NEXT:    ; implicit-def: $vgpr3212569; SI-NEXT:    ; kill: killed $vgpr4312570; SI-NEXT:    ; implicit-def: $vgpr4312571; SI-NEXT:    ; kill: killed $vgpr3212572; SI-NEXT:    ; implicit-def: $vgpr3212573; SI-NEXT:    ; kill: killed $vgpr4312574; SI-NEXT:    ; implicit-def: $vgpr4312575; SI-NEXT:    ; kill: killed $vgpr3212576; SI-NEXT:    ; implicit-def: $vgpr3212577; SI-NEXT:    ; kill: killed $vgpr4312578; SI-NEXT:    ; implicit-def: $vgpr4312579; SI-NEXT:    ; kill: killed $vgpr3212580; SI-NEXT:    ; implicit-def: $vgpr3212581; SI-NEXT:    ; kill: killed $vgpr4312582; SI-NEXT:    ; implicit-def: $vgpr4312583; SI-NEXT:    ; kill: killed $vgpr3212584; SI-NEXT:    ; implicit-def: $vgpr3212585; SI-NEXT:    ; kill: killed $vgpr4312586; SI-NEXT:    ; implicit-def: $vgpr4312587; SI-NEXT:    ; kill: killed $vgpr3212588; SI-NEXT:    ; implicit-def: $vgpr3212589; SI-NEXT:    ; kill: killed $vgpr4312590; SI-NEXT:    ; implicit-def: $vgpr4312591; SI-NEXT:    ; kill: killed $vgpr3212592; SI-NEXT:    ; implicit-def: $vgpr3212593; SI-NEXT:    ; kill: killed $vgpr4312594; SI-NEXT:    ; implicit-def: $vgpr4312595; SI-NEXT:    ; kill: killed $vgpr3212596; SI-NEXT:    ; implicit-def: $vgpr3212597; SI-NEXT:    ; kill: killed $vgpr4312598; SI-NEXT:    ; implicit-def: $vgpr4312599; SI-NEXT:    ; kill: killed $vgpr3212600; SI-NEXT:    ; implicit-def: $vgpr3212601; SI-NEXT:    ; kill: killed $vgpr4312602; SI-NEXT:    ; implicit-def: $vgpr4312603; SI-NEXT:    ; kill: killed $vgpr3212604; SI-NEXT:    ; implicit-def: $vgpr3212605; SI-NEXT:    ; kill: killed $vgpr4312606; SI-NEXT:    ; implicit-def: $vgpr4312607; SI-NEXT:    ; kill: killed $vgpr3212608; SI-NEXT:    ; implicit-def: $vgpr3212609; SI-NEXT:    ; kill: killed $vgpr4312610; SI-NEXT:    ; implicit-def: $vgpr4312611; SI-NEXT:    ; kill: killed $vgpr3212612; SI-NEXT:    ; implicit-def: $vgpr3212613; SI-NEXT:    ; kill: killed $vgpr4312614; SI-NEXT:    ; implicit-def: $vgpr4312615; SI-NEXT:    ; kill: killed $vgpr3212616; SI-NEXT:    ; implicit-def: $vgpr3212617; SI-NEXT:    ; kill: killed $vgpr4312618; SI-NEXT:    ; implicit-def: $vgpr4312619; SI-NEXT:    ; kill: killed $vgpr3212620; SI-NEXT:    ; implicit-def: $vgpr3212621; SI-NEXT:    ; kill: killed $vgpr4312622; SI-NEXT:    ; implicit-def: $vgpr4312623; SI-NEXT:    ; kill: killed $vgpr3212624; SI-NEXT:    ; implicit-def: $vgpr3212625; SI-NEXT:    ; kill: killed $vgpr4312626; SI-NEXT:    ; implicit-def: $vgpr4312627; SI-NEXT:    ; kill: killed $vgpr3212628; SI-NEXT:    ; implicit-def: $vgpr3212629; SI-NEXT:    ; kill: killed $vgpr4312630; SI-NEXT:    ; implicit-def: $vgpr4312631; SI-NEXT:    ; kill: killed $vgpr3212632; SI-NEXT:    ; implicit-def: $vgpr3212633; SI-NEXT:    ; kill: killed $vgpr4312634; SI-NEXT:    ; implicit-def: $vgpr4312635; SI-NEXT:    ; kill: killed $vgpr3212636; SI-NEXT:    ; implicit-def: $vgpr3212637; SI-NEXT:    ; kill: killed $vgpr4312638; SI-NEXT:    ; implicit-def: $vgpr4312639; SI-NEXT:    ; kill: killed $vgpr3212640; SI-NEXT:    ; implicit-def: $vgpr3212641; SI-NEXT:    ; kill: killed $vgpr4312642; SI-NEXT:    ; implicit-def: $vgpr4312643; SI-NEXT:    ; kill: killed $vgpr3212644; SI-NEXT:    ; implicit-def: $vgpr3212645; SI-NEXT:    ; kill: killed $vgpr4312646; SI-NEXT:    ; implicit-def: $vgpr4312647; SI-NEXT:    ; kill: killed $vgpr3212648; SI-NEXT:    ; implicit-def: $vgpr3212649; SI-NEXT:    ; kill: killed $vgpr4312650; SI-NEXT:    ; implicit-def: $vgpr4312651; SI-NEXT:    ; kill: killed $vgpr3212652; SI-NEXT:    ; implicit-def: $vgpr3212653; SI-NEXT:    ; kill: killed $vgpr4312654; SI-NEXT:    ; implicit-def: $vgpr4312655; SI-NEXT:    ; kill: killed $vgpr3212656; SI-NEXT:    ; implicit-def: $vgpr3212657; SI-NEXT:    ; kill: killed $vgpr4312658; SI-NEXT:    ; implicit-def: $vgpr4312659; SI-NEXT:    ; kill: killed $vgpr3212660; SI-NEXT:    ; implicit-def: $vgpr3212661; SI-NEXT:    ; kill: killed $vgpr4312662; SI-NEXT:    ; implicit-def: $vgpr4312663; SI-NEXT:    ; kill: killed $vgpr3212664; SI-NEXT:    ; implicit-def: $vgpr3212665; SI-NEXT:    ; kill: killed $vgpr4312666; SI-NEXT:    ; implicit-def: $vgpr4312667; SI-NEXT:    ; kill: killed $vgpr3212668; SI-NEXT:    ; implicit-def: $vgpr3212669; SI-NEXT:    ; kill: killed $vgpr4312670; SI-NEXT:    ; implicit-def: $vgpr4312671; SI-NEXT:    ; kill: killed $vgpr3212672; SI-NEXT:    ; implicit-def: $vgpr3212673; SI-NEXT:    ; kill: killed $vgpr4312674; SI-NEXT:    ; implicit-def: $vgpr4312675; SI-NEXT:    ; kill: killed $vgpr3212676; SI-NEXT:    ; implicit-def: $vgpr3212677; SI-NEXT:    ; kill: killed $vgpr4312678; SI-NEXT:    ; implicit-def: $vgpr4312679; SI-NEXT:    ; kill: killed $vgpr3212680; SI-NEXT:    ; implicit-def: $vgpr3212681; SI-NEXT:    ; kill: killed $vgpr4312682; SI-NEXT:    ; implicit-def: $vgpr4312683; SI-NEXT:    ; kill: killed $vgpr3212684; SI-NEXT:    ; implicit-def: $vgpr3212685; SI-NEXT:    ; kill: killed $vgpr4312686; SI-NEXT:    ; implicit-def: $vgpr4312687; SI-NEXT:    ; kill: killed $vgpr3212688; SI-NEXT:    ; implicit-def: $vgpr3212689; SI-NEXT:    ; kill: killed $vgpr4312690; SI-NEXT:    ; implicit-def: $vgpr4312691; SI-NEXT:    ; kill: killed $vgpr3212692; SI-NEXT:    ; implicit-def: $vgpr3212693; SI-NEXT:    ; kill: killed $vgpr4312694; SI-NEXT:    ; implicit-def: $vgpr4312695; SI-NEXT:    ; kill: killed $vgpr3212696; SI-NEXT:    ; implicit-def: $vgpr3212697; SI-NEXT:    ; kill: killed $vgpr4312698; SI-NEXT:    ; implicit-def: $vgpr4312699; SI-NEXT:    ; kill: killed $vgpr3212700; SI-NEXT:    ; implicit-def: $vgpr3212701; SI-NEXT:    ; kill: killed $vgpr4312702; SI-NEXT:    ; implicit-def: $vgpr4312703; SI-NEXT:    ; kill: killed $vgpr3212704; SI-NEXT:    ; implicit-def: $vgpr3212705; SI-NEXT:    ; kill: killed $vgpr4312706; SI-NEXT:    ; implicit-def: $vgpr4312707; SI-NEXT:    ; kill: killed $vgpr3212708; SI-NEXT:    ; implicit-def: $vgpr3212709; SI-NEXT:    ; kill: killed $vgpr4312710; SI-NEXT:    ; implicit-def: $vgpr4312711; SI-NEXT:    ; kill: killed $vgpr3212712; SI-NEXT:    ; implicit-def: $vgpr3212713; SI-NEXT:    ; kill: killed $vgpr4312714; SI-NEXT:    ; implicit-def: $vgpr4312715; SI-NEXT:    ; kill: killed $vgpr3212716; SI-NEXT:    ; implicit-def: $vgpr3212717; SI-NEXT:    ; kill: killed $vgpr4312718; SI-NEXT:    ; implicit-def: $vgpr4312719; SI-NEXT:    ; kill: killed $vgpr3212720; SI-NEXT:    ; implicit-def: $vgpr3212721; SI-NEXT:    ; kill: killed $vgpr4312722; SI-NEXT:    ; implicit-def: $vgpr4312723; SI-NEXT:    ; kill: killed $vgpr3212724; SI-NEXT:    ; implicit-def: $vgpr3212725; SI-NEXT:    ; kill: killed $vgpr4312726; SI-NEXT:    ; implicit-def: $vgpr4312727; SI-NEXT:    ; kill: killed $vgpr3212728; SI-NEXT:    ; implicit-def: $vgpr3212729; SI-NEXT:    ; kill: killed $vgpr4312730; SI-NEXT:    ; implicit-def: $vgpr4312731; SI-NEXT:    ; kill: killed $vgpr3212732; SI-NEXT:    ; implicit-def: $vgpr3212733; SI-NEXT:    ; kill: killed $vgpr4312734; SI-NEXT:    ; implicit-def: $vgpr4312735; SI-NEXT:    ; implicit-def: $vgpr3312736; SI-NEXT:    ; kill: killed $vgpr3212737; SI-NEXT:    ; implicit-def: $vgpr3212738; SI-NEXT:    ; kill: killed $vgpr4312739; SI-NEXT:    ; implicit-def: $vgpr4312740; SI-NEXT:  .LBB14_2: ; %Flow12741; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]12742; SI-NEXT:    s_cbranch_execz .LBB14_412743; SI-NEXT:  ; %bb.3: ; %cmp.true12744; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload12745; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload12746; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload12747; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload12748; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload12749; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload12750; SI-NEXT:    s_movk_i32 s6, 0x30012751; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload12752; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload12753; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload12754; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload12755; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload12756; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload12757; SI-NEXT:    s_waitcnt expcnt(0)12758; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload12759; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v4912760; SI-NEXT:    v_and_b32_e32 v9, 0xff, v912761; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v912762; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload12763; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload12764; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload12765; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload12766; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload12767; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload12768; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload12769; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload12770; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload12771; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload12772; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload12773; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload12774; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload12775; SI-NEXT:    s_mov_b32 s7, 0x300000012776; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload12777; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload12778; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload12779; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload12780; SI-NEXT:    s_waitcnt vmcnt(14)12781; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v012782; SI-NEXT:    v_and_b32_e32 v0, 0xff, v012783; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v112784; SI-NEXT:    v_or_b32_e32 v0, v54, v012785; SI-NEXT:    v_and_b32_e32 v1, 0xff, v112786; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v112787; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v012788; SI-NEXT:    v_or_b32_e32 v1, v42, v112789; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v012790; SI-NEXT:    v_or_b32_e32 v0, v1, v012791; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload12792; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v212793; SI-NEXT:    v_and_b32_e32 v2, 0xff, v212794; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v212795; SI-NEXT:    v_or_b32_e32 v2, v40, v212796; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v312797; SI-NEXT:    v_and_b32_e32 v3, 0xff, v312798; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v312799; SI-NEXT:    v_or_b32_e32 v3, v34, v312800; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v412801; SI-NEXT:    v_and_b32_e32 v4, 0xff, v412802; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v412803; SI-NEXT:    v_or_b32_e32 v4, v32, v412804; SI-NEXT:    v_or_b32_e32 v9, v10, v912805; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload12806; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v512807; SI-NEXT:    v_and_b32_e32 v5, 0xff, v512808; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v512809; SI-NEXT:    v_or_b32_e32 v5, v60, v512810; SI-NEXT:    s_waitcnt vmcnt(12)12811; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v612812; SI-NEXT:    v_and_b32_e32 v6, 0xff, v612813; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v612814; SI-NEXT:    v_or_b32_e32 v6, v58, v612815; SI-NEXT:    s_waitcnt vmcnt(8)12816; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v712817; SI-NEXT:    v_and_b32_e32 v7, 0xff, v712818; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v712819; SI-NEXT:    v_or_b32_e32 v7, v37, v712820; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload12821; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload12822; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v012823; SI-NEXT:    s_waitcnt vmcnt(3)12824; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v112825; SI-NEXT:    v_and_b32_e32 v1, 0xff, v112826; SI-NEXT:    v_or_b32_e32 v1, v41, v112827; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v112828; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v112829; SI-NEXT:    v_or_b32_e32 v1, v2, v112830; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload12831; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v112832; SI-NEXT:    s_waitcnt vmcnt(1)12833; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v812834; SI-NEXT:    v_and_b32_e32 v8, 0xff, v812835; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v812836; SI-NEXT:    v_or_b32_e32 v8, v55, v812837; SI-NEXT:    s_waitcnt vmcnt(0)12838; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v212839; SI-NEXT:    v_and_b32_e32 v2, 0xff, v212840; SI-NEXT:    v_or_b32_e32 v2, v53, v212841; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v212842; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v212843; SI-NEXT:    v_or_b32_e32 v2, v3, v212844; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload12845; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v212846; SI-NEXT:    s_waitcnt vmcnt(0)12847; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v312848; SI-NEXT:    v_and_b32_e32 v3, 0xff, v312849; SI-NEXT:    v_or_b32_e32 v3, v48, v312850; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v312851; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v312852; SI-NEXT:    v_or_b32_e32 v3, v4, v312853; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload12854; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v312855; SI-NEXT:    s_waitcnt vmcnt(0)12856; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v412857; SI-NEXT:    v_and_b32_e32 v4, 0xff, v412858; SI-NEXT:    v_or_b32_e32 v4, v62, v412859; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v412860; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v412861; SI-NEXT:    v_or_b32_e32 v4, v5, v412862; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload12863; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v412864; SI-NEXT:    s_waitcnt vmcnt(0)12865; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v512866; SI-NEXT:    v_and_b32_e32 v5, 0xff, v512867; SI-NEXT:    v_or_b32_e32 v5, v38, v512868; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v512869; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v512870; SI-NEXT:    v_or_b32_e32 v5, v6, v512871; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload12872; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v512873; SI-NEXT:    s_waitcnt vmcnt(0)12874; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v612875; SI-NEXT:    v_and_b32_e32 v6, 0xff, v612876; SI-NEXT:    v_or_b32_e32 v6, v56, v612877; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v612878; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v612879; SI-NEXT:    v_or_b32_e32 v6, v7, v612880; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload12881; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v612882; SI-NEXT:    s_waitcnt vmcnt(0)12883; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v712884; SI-NEXT:    v_and_b32_e32 v7, 0xff, v712885; SI-NEXT:    v_or_b32_e32 v7, v46, v712886; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v712887; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v712888; SI-NEXT:    v_or_b32_e32 v7, v8, v712889; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v3512890; SI-NEXT:    v_and_b32_e32 v8, 0xff, v812891; SI-NEXT:    v_or_b32_e32 v8, v44, v812892; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v812893; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v812894; SI-NEXT:    v_or_b32_e32 v8, v9, v812895; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v3312896; SI-NEXT:    v_and_b32_e32 v9, 0xff, v912897; SI-NEXT:    v_or_b32_e32 v9, v10, v912898; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v5212899; SI-NEXT:    v_and_b32_e32 v10, 0xff, v1012900; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v1012901; SI-NEXT:    v_or_b32_e32 v10, v11, v1012902; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload12903; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v912904; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v912905; SI-NEXT:    v_or_b32_e32 v9, v10, v912906; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v6312907; SI-NEXT:    v_and_b32_e32 v10, 0xff, v1012908; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v712909; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v812910; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v912911; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload12912; SI-NEXT:    s_waitcnt vmcnt(1)12913; SI-NEXT:    v_or_b32_e32 v10, v11, v1012914; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v3912915; SI-NEXT:    v_and_b32_e32 v11, 0xff, v1112916; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v1112917; SI-NEXT:    v_or_b32_e32 v11, v12, v1112918; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload12919; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v1012920; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v1012921; SI-NEXT:    v_or_b32_e32 v10, v11, v1012922; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v6112923; SI-NEXT:    v_and_b32_e32 v11, 0xff, v1112924; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v1012925; SI-NEXT:    s_waitcnt vmcnt(0)12926; SI-NEXT:    v_or_b32_e32 v11, v12, v1112927; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v5912928; SI-NEXT:    v_and_b32_e32 v12, 0xff, v1212929; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v1212930; SI-NEXT:    v_or_b32_e32 v12, v13, v1212931; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload12932; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v1112933; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v1112934; SI-NEXT:    v_or_b32_e32 v11, v12, v1112935; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v5112936; SI-NEXT:    v_and_b32_e32 v12, 0xff, v1212937; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v1112938; SI-NEXT:    s_waitcnt vmcnt(0)12939; SI-NEXT:    v_or_b32_e32 v12, v13, v1212940; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v5712941; SI-NEXT:    v_and_b32_e32 v13, 0xff, v1312942; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v1312943; SI-NEXT:    v_or_b32_e32 v13, v14, v1312944; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload12945; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v1212946; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v1212947; SI-NEXT:    v_or_b32_e32 v12, v13, v1212948; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v4712949; SI-NEXT:    v_and_b32_e32 v13, 0xff, v1312950; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v1212951; SI-NEXT:    s_waitcnt vmcnt(0)12952; SI-NEXT:    v_or_b32_e32 v13, v14, v1312953; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v5012954; SI-NEXT:    v_and_b32_e32 v14, 0xff, v1412955; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v1412956; SI-NEXT:    v_or_b32_e32 v14, v15, v1412957; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload12958; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v1312959; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v1312960; SI-NEXT:    v_or_b32_e32 v13, v14, v1312961; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v4512962; SI-NEXT:    v_and_b32_e32 v14, 0xff, v1412963; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v1312964; SI-NEXT:    s_waitcnt vmcnt(0)12965; SI-NEXT:    v_or_b32_e32 v14, v15, v1412966; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v3612967; SI-NEXT:    v_and_b32_e32 v15, 0xff, v1512968; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v1512969; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v1412970; SI-NEXT:    v_or_b32_e32 v15, v16, v1512971; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v1412972; SI-NEXT:    v_or_b32_e32 v14, v15, v1412973; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload12974; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload12975; SI-NEXT:    v_add_i32_e32 v14, vcc, s7, v1412976; SI-NEXT:    s_waitcnt vmcnt(1)12977; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v1512978; SI-NEXT:    v_and_b32_e32 v15, 0xff, v1512979; SI-NEXT:    s_waitcnt vmcnt(0)12980; SI-NEXT:    v_or_b32_e32 v15, v16, v1512981; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload12982; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v1512983; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v1512984; SI-NEXT:    s_waitcnt vmcnt(0)12985; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v1612986; SI-NEXT:    v_and_b32_e32 v16, 0xff, v1612987; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v1612988; SI-NEXT:    v_or_b32_e32 v16, v17, v1612989; SI-NEXT:    v_or_b32_e32 v15, v16, v1512990; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload12991; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload12992; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v1512993; SI-NEXT:    s_waitcnt vmcnt(1)12994; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v1612995; SI-NEXT:    v_and_b32_e32 v16, 0xff, v1612996; SI-NEXT:    s_waitcnt vmcnt(0)12997; SI-NEXT:    v_or_b32_e32 v16, v17, v1612998; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload12999; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v1613000; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v1613001; SI-NEXT:    s_waitcnt vmcnt(0)13002; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v1713003; SI-NEXT:    v_and_b32_e32 v17, 0xff, v1713004; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v1713005; SI-NEXT:    v_or_b32_e32 v17, v18, v1713006; SI-NEXT:    v_or_b32_e32 v16, v17, v1613007; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload13008; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload13009; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v1613010; SI-NEXT:    s_waitcnt vmcnt(1)13011; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v1713012; SI-NEXT:    v_and_b32_e32 v17, 0xff, v1713013; SI-NEXT:    s_waitcnt vmcnt(0)13014; SI-NEXT:    v_or_b32_e32 v17, v18, v1713015; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload13016; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v1713017; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v1713018; SI-NEXT:    s_waitcnt vmcnt(0)13019; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v1813020; SI-NEXT:    v_and_b32_e32 v18, 0xff, v1813021; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v1813022; SI-NEXT:    v_or_b32_e32 v18, v19, v1813023; SI-NEXT:    v_or_b32_e32 v17, v18, v1713024; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload13025; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload13026; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v1713027; SI-NEXT:    s_waitcnt vmcnt(1)13028; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v1813029; SI-NEXT:    v_and_b32_e32 v18, 0xff, v1813030; SI-NEXT:    s_waitcnt vmcnt(0)13031; SI-NEXT:    v_or_b32_e32 v18, v19, v1813032; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload13033; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v1813034; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v1813035; SI-NEXT:    s_waitcnt vmcnt(0)13036; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v1913037; SI-NEXT:    v_and_b32_e32 v19, 0xff, v1913038; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v1913039; SI-NEXT:    v_or_b32_e32 v19, v20, v1913040; SI-NEXT:    v_or_b32_e32 v18, v19, v1813041; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload13042; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload13043; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v1813044; SI-NEXT:    s_waitcnt vmcnt(1)13045; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v1913046; SI-NEXT:    v_and_b32_e32 v19, 0xff, v1913047; SI-NEXT:    s_waitcnt vmcnt(0)13048; SI-NEXT:    v_or_b32_e32 v19, v20, v1913049; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload13050; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v1913051; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v1913052; SI-NEXT:    s_waitcnt vmcnt(0)13053; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v2013054; SI-NEXT:    v_and_b32_e32 v20, 0xff, v2013055; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v2013056; SI-NEXT:    v_or_b32_e32 v20, v21, v2013057; SI-NEXT:    v_or_b32_e32 v19, v20, v1913058; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload13059; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload13060; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v1913061; SI-NEXT:    s_waitcnt vmcnt(1)13062; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v2013063; SI-NEXT:    v_and_b32_e32 v20, 0xff, v2013064; SI-NEXT:    s_waitcnt vmcnt(0)13065; SI-NEXT:    v_or_b32_e32 v20, v21, v2013066; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload13067; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v2013068; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v2013069; SI-NEXT:    s_waitcnt vmcnt(0)13070; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v2113071; SI-NEXT:    v_and_b32_e32 v21, 0xff, v2113072; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v2113073; SI-NEXT:    v_or_b32_e32 v21, v22, v2113074; SI-NEXT:    v_or_b32_e32 v20, v21, v2013075; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload13076; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload13077; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v2013078; SI-NEXT:    s_waitcnt vmcnt(1)13079; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v2113080; SI-NEXT:    v_and_b32_e32 v21, 0xff, v2113081; SI-NEXT:    s_waitcnt vmcnt(0)13082; SI-NEXT:    v_or_b32_e32 v21, v22, v2113083; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload13084; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v2113085; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v2113086; SI-NEXT:    s_waitcnt vmcnt(0)13087; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v2213088; SI-NEXT:    v_and_b32_e32 v22, 0xff, v2213089; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v2213090; SI-NEXT:    v_or_b32_e32 v22, v23, v2213091; SI-NEXT:    v_or_b32_e32 v21, v22, v2113092; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload13093; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload13094; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v2113095; SI-NEXT:    s_waitcnt vmcnt(1)13096; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v2213097; SI-NEXT:    v_and_b32_e32 v22, 0xff, v2213098; SI-NEXT:    s_waitcnt vmcnt(0)13099; SI-NEXT:    v_or_b32_e32 v22, v23, v2213100; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload13101; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v2213102; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v2213103; SI-NEXT:    s_waitcnt vmcnt(0)13104; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v2313105; SI-NEXT:    v_and_b32_e32 v23, 0xff, v2313106; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v2313107; SI-NEXT:    v_or_b32_e32 v23, v24, v2313108; SI-NEXT:    v_or_b32_e32 v22, v23, v2213109; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload13110; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload13111; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v2213112; SI-NEXT:    s_waitcnt vmcnt(1)13113; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v2313114; SI-NEXT:    v_and_b32_e32 v23, 0xff, v2313115; SI-NEXT:    s_waitcnt vmcnt(0)13116; SI-NEXT:    v_or_b32_e32 v23, v24, v2313117; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload13118; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v2313119; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v2313120; SI-NEXT:    s_waitcnt vmcnt(0)13121; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v2413122; SI-NEXT:    v_and_b32_e32 v24, 0xff, v2413123; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v2413124; SI-NEXT:    v_or_b32_e32 v24, v25, v2413125; SI-NEXT:    v_or_b32_e32 v23, v24, v2313126; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload13127; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload13128; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v2313129; SI-NEXT:    s_waitcnt vmcnt(1)13130; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v2413131; SI-NEXT:    v_and_b32_e32 v24, 0xff, v2413132; SI-NEXT:    s_waitcnt vmcnt(0)13133; SI-NEXT:    v_or_b32_e32 v24, v25, v2413134; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload13135; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v2413136; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v2413137; SI-NEXT:    s_waitcnt vmcnt(0)13138; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v2513139; SI-NEXT:    v_and_b32_e32 v25, 0xff, v2513140; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v2513141; SI-NEXT:    v_or_b32_e32 v25, v26, v2513142; SI-NEXT:    v_or_b32_e32 v24, v25, v2413143; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload13144; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload13145; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v2413146; SI-NEXT:    s_waitcnt vmcnt(1)13147; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v2513148; SI-NEXT:    v_and_b32_e32 v25, 0xff, v2513149; SI-NEXT:    s_waitcnt vmcnt(0)13150; SI-NEXT:    v_or_b32_e32 v25, v26, v2513151; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload13152; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v2513153; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v2513154; SI-NEXT:    s_waitcnt vmcnt(0)13155; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v2613156; SI-NEXT:    v_and_b32_e32 v26, 0xff, v2613157; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v2613158; SI-NEXT:    v_or_b32_e32 v26, v27, v2613159; SI-NEXT:    v_or_b32_e32 v25, v26, v2513160; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload13161; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload13162; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v2513163; SI-NEXT:    s_waitcnt vmcnt(1)13164; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v2613165; SI-NEXT:    v_and_b32_e32 v26, 0xff, v2613166; SI-NEXT:    s_waitcnt vmcnt(0)13167; SI-NEXT:    v_or_b32_e32 v26, v27, v2613168; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload13169; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v2613170; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v2613171; SI-NEXT:    s_waitcnt vmcnt(0)13172; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v2713173; SI-NEXT:    v_and_b32_e32 v27, 0xff, v2713174; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v2713175; SI-NEXT:    v_or_b32_e32 v27, v28, v2713176; SI-NEXT:    v_or_b32_e32 v26, v27, v2613177; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload13178; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload13179; SI-NEXT:    v_add_i32_e32 v26, vcc, s7, v2613180; SI-NEXT:    s_waitcnt vmcnt(1)13181; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v2713182; SI-NEXT:    v_and_b32_e32 v27, 0xff, v2713183; SI-NEXT:    s_waitcnt vmcnt(0)13184; SI-NEXT:    v_or_b32_e32 v27, v28, v2713185; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload13186; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v2713187; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v2713188; SI-NEXT:    s_waitcnt vmcnt(0)13189; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v2813190; SI-NEXT:    v_and_b32_e32 v28, 0xff, v2813191; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v2813192; SI-NEXT:    v_or_b32_e32 v28, v29, v2813193; SI-NEXT:    v_or_b32_e32 v27, v28, v2713194; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload13195; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload13196; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v2713197; SI-NEXT:    s_waitcnt vmcnt(1)13198; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v2813199; SI-NEXT:    v_and_b32_e32 v28, 0xff, v2813200; SI-NEXT:    s_waitcnt vmcnt(0)13201; SI-NEXT:    v_or_b32_e32 v28, v29, v2813202; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload13203; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v2813204; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v2813205; SI-NEXT:    s_waitcnt vmcnt(0)13206; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v2913207; SI-NEXT:    v_and_b32_e32 v29, 0xff, v2913208; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v2913209; SI-NEXT:    v_or_b32_e32 v29, v30, v2913210; SI-NEXT:    v_or_b32_e32 v28, v29, v2813211; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload13212; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload13213; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v2813214; SI-NEXT:    s_waitcnt vmcnt(1)13215; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v2913216; SI-NEXT:    v_and_b32_e32 v29, 0xff, v2913217; SI-NEXT:    s_waitcnt vmcnt(0)13218; SI-NEXT:    v_or_b32_e32 v29, v30, v2913219; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload13220; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v2913221; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v2913222; SI-NEXT:    s_waitcnt vmcnt(0)13223; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v3013224; SI-NEXT:    v_and_b32_e32 v30, 0xff, v3013225; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v3013226; SI-NEXT:    v_or_b32_e32 v30, v31, v3013227; SI-NEXT:    v_or_b32_e32 v29, v30, v2913228; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload13229; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload13230; SI-NEXT:    v_add_i32_e32 v29, vcc, s7, v2913231; SI-NEXT:    s_waitcnt vmcnt(1)13232; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v3013233; SI-NEXT:    v_and_b32_e32 v30, 0xff, v3013234; SI-NEXT:    s_waitcnt vmcnt(0)13235; SI-NEXT:    v_or_b32_e32 v30, v31, v3013236; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload13237; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v3013238; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v3013239; SI-NEXT:    s_waitcnt vmcnt(0)13240; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v3113241; SI-NEXT:    v_and_b32_e32 v31, 0xff, v3113242; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v3113243; SI-NEXT:    v_or_b32_e32 v31, v32, v3113244; SI-NEXT:    v_or_b32_e32 v30, v31, v3013245; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload13246; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload13247; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v3013248; SI-NEXT:    s_waitcnt vmcnt(1)13249; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v3113250; SI-NEXT:    v_and_b32_e32 v31, 0xff, v3113251; SI-NEXT:    s_waitcnt vmcnt(0)13252; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v3213253; SI-NEXT:    v_or_b32_e32 v31, v43, v3113254; SI-NEXT:    v_and_b32_e32 v32, 0xff, v3213255; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x300, v3113256; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v3213257; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v3113258; SI-NEXT:    v_or_b32_e32 v32, v33, v3213259; SI-NEXT:    v_or_b32_e32 v31, v32, v3113260; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v3113261; SI-NEXT:  .LBB14_4: ; %end13262; SI-NEXT:    s_or_b64 exec, exec, s[4:5]13263; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload13264; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload13265; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload13266; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload13267; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload13268; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload13269; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload13270; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload13271; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload13272; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload13273; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload13274; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload13275; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload13276; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload13277; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload13278; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload13279; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)13280; SI-NEXT:    s_setpc_b64 s[30:31]13281;13282; VI-LABEL: bitcast_v128i8_to_v32i32:13283; VI:       ; %bb.0:13284; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)13285; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill13286; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill13287; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill13288; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill13289; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill13290; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill13291; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill13292; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill13293; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill13294; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill13295; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill13296; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill13297; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill13298; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill13299; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill13300; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill13301; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill13302; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill13303; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill13304; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill13305; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill13306; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill13307; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill13308; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill13309; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill13310; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill13311; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill13312; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill13313; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill13314; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill13315; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill13316; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill13317; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:38813318; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s3213319; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:813320; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:1613321; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:2413322; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:3213323; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:4013324; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:4813325; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:5613326; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:6413327; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:7213328; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:8013329; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:8813330; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:9613331; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:10413332; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:11213333; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:12013334; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:12813335; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:13613336; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:14413337; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:15213338; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:16013339; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:16813340; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:17613341; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:18413342; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v113343; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v313344; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v513345; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v713346; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v913347; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v1113348; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v1313349; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v1513350; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v1713351; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v1913352; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v2113353; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v2313354; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v2513355; VI-NEXT:    v_lshlrev_b16_e32 v36, 8, v2713356; VI-NEXT:    v_lshlrev_b16_e32 v35, 8, v2913357; VI-NEXT:    s_waitcnt vmcnt(14)13358; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3013359; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v013360; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v213361; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v413362; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill13363; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v613364; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill13365; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v813366; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill13367; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1013368; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill13369; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1213370; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill13371; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1413372; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill13373; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1613374; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill13375; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1813376; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill13377; VI-NEXT:    s_waitcnt vmcnt(14)13378; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2013379; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill13380; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2213381; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill13382; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2413383; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill13384; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2613385; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill13386; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2813387; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill13388; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:11613389; VI-NEXT:    s_waitcnt vmcnt(0)13390; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill13391; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3113392; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill13393; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:12413394; VI-NEXT:    s_waitcnt vmcnt(0)13395; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill13396; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:13213397; VI-NEXT:    s_waitcnt vmcnt(0)13398; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill13399; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:14013400; VI-NEXT:    s_waitcnt vmcnt(0)13401; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill13402; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:14813403; VI-NEXT:    s_waitcnt vmcnt(0)13404; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill13405; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3213406; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill13407; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v4013408; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill13409; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v4213410; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill13411; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v4413412; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill13413; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:15613414; VI-NEXT:    s_waitcnt vmcnt(0)13415; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill13416; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16413417; VI-NEXT:    s_waitcnt vmcnt(0)13418; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill13419; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:17213420; VI-NEXT:    s_waitcnt vmcnt(0)13421; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill13422; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:18013423; VI-NEXT:    s_waitcnt vmcnt(0)13424; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill13425; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v4513426; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill13427; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v4613428; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill13429; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v4713430; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill13431; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v5613432; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill13433; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:19213434; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:20013435; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:20813436; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:21613437; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:18813438; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:4413439; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:3613440; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:2813441; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:2013442; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:1213443; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:413444; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:10813445; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:10013446; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:9213447; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:8413448; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:7613449; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:6813450; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:6013451; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:5213452; VI-NEXT:    s_waitcnt vmcnt(14)13453; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v013454; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill13455; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v113456; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill13457; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill13458; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:19613459; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v213460; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill13461; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v313462; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill13463; VI-NEXT:    s_waitcnt vmcnt(2)13464; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill13465; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:20413466; VI-NEXT:    s_waitcnt vmcnt(0)13467; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill13468; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:21213469; VI-NEXT:    s_waitcnt vmcnt(0)13470; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill13471; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:22413472; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:23213473; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:24013474; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:24813475; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:22013476; VI-NEXT:    s_waitcnt vmcnt(4)13477; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v013478; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill13479; VI-NEXT:    s_waitcnt vmcnt(4)13480; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v113481; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill13482; VI-NEXT:    s_waitcnt vmcnt(2)13483; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill13484; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:22813485; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v213486; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill13487; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v313488; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill13489; VI-NEXT:    s_waitcnt vmcnt(2)13490; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill13491; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:23613492; VI-NEXT:    s_waitcnt vmcnt(0)13493; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill13494; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:24413495; VI-NEXT:    s_waitcnt vmcnt(0)13496; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill13497; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:25613498; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:26413499; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:27213500; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:28013501; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:25213502; VI-NEXT:    s_waitcnt vmcnt(4)13503; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v013504; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill13505; VI-NEXT:    s_waitcnt vmcnt(4)13506; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v113507; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill13508; VI-NEXT:    s_waitcnt vmcnt(2)13509; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill13510; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:26013511; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v213512; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill13513; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v313514; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill13515; VI-NEXT:    s_waitcnt vmcnt(2)13516; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill13517; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:26813518; VI-NEXT:    s_waitcnt vmcnt(0)13519; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill13520; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:27613521; VI-NEXT:    s_waitcnt vmcnt(0)13522; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill13523; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:28813524; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:29613525; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:30413526; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:31213527; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:28413528; VI-NEXT:    s_waitcnt vmcnt(4)13529; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v013530; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill13531; VI-NEXT:    s_waitcnt vmcnt(4)13532; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v113533; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill13534; VI-NEXT:    s_waitcnt vmcnt(2)13535; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill13536; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:29213537; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v213538; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill13539; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v313540; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill13541; VI-NEXT:    s_waitcnt vmcnt(2)13542; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill13543; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:30013544; VI-NEXT:    s_waitcnt vmcnt(0)13545; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill13546; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:30813547; VI-NEXT:    s_waitcnt vmcnt(0)13548; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill13549; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:32013550; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:32813551; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:33613552; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:34413553; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:31613554; VI-NEXT:    s_waitcnt vmcnt(4)13555; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v013556; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill13557; VI-NEXT:    s_waitcnt vmcnt(4)13558; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v113559; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill13560; VI-NEXT:    s_waitcnt vmcnt(2)13561; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill13562; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32413563; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v213564; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill13565; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v313566; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill13567; VI-NEXT:    s_waitcnt vmcnt(2)13568; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill13569; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:33213570; VI-NEXT:    s_waitcnt vmcnt(0)13571; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill13572; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:34013573; VI-NEXT:    s_waitcnt vmcnt(0)13574; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill13575; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:35213576; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:36013577; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:36813578; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:37613579; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:34813580; VI-NEXT:    s_waitcnt vmcnt(4)13581; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v013582; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill13583; VI-NEXT:    s_waitcnt vmcnt(4)13584; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v113585; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill13586; VI-NEXT:    s_waitcnt vmcnt(2)13587; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill13588; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:35613589; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v213590; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill13591; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v313592; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill13593; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:38413594; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:38013595; VI-NEXT:    s_waitcnt vmcnt(4)13596; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill13597; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:36413598; VI-NEXT:    s_waitcnt vmcnt(3)13599; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v013600; VI-NEXT:    s_waitcnt vmcnt(2)13601; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill13602; VI-NEXT:    s_waitcnt vmcnt(1)13603; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill13604; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:37213605; VI-NEXT:    s_waitcnt vmcnt(0)13606; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill13607; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr3113608; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc13609; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]13610; VI-NEXT:    s_cbranch_execz .LBB14_213611; VI-NEXT:  ; %bb.1: ; %cmp.false13612; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload13613; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload13614; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload13615; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload13616; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload13617; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload13618; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload13619; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload13620; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload13621; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload13622; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload13623; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload13624; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload13625; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload13626; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload13627; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload13628; VI-NEXT:    s_waitcnt vmcnt(14)13629; VI-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13630; VI-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13631; VI-NEXT:    s_waitcnt vmcnt(13)13632; VI-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13633; VI-NEXT:    s_waitcnt vmcnt(12)13634; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13635; VI-NEXT:    s_waitcnt vmcnt(11)13636; VI-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13637; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13638; VI-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13639; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload13640; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload13641; VI-NEXT:    s_waitcnt vmcnt(10)13642; VI-NEXT:    v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13643; VI-NEXT:    s_waitcnt vmcnt(8)13644; VI-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13645; VI-NEXT:    s_waitcnt vmcnt(6)13646; VI-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13647; VI-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13648; VI-NEXT:    s_waitcnt vmcnt(4)13649; VI-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13650; VI-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13651; VI-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13652; VI-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13653; VI-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13654; VI-NEXT:    ; implicit-def: $vgpr6113655; VI-NEXT:    ; implicit-def: $vgpr6013656; VI-NEXT:    ; implicit-def: $vgpr5613657; VI-NEXT:    ; implicit-def: $vgpr5713658; VI-NEXT:    ; implicit-def: $vgpr4613659; VI-NEXT:    ; implicit-def: $vgpr4413660; VI-NEXT:    ; implicit-def: $vgpr4313661; VI-NEXT:    ; implicit-def: $vgpr4113662; VI-NEXT:    ; implicit-def: $vgpr5513663; VI-NEXT:    ; implicit-def: $vgpr5413664; VI-NEXT:    ; implicit-def: $vgpr5013665; VI-NEXT:    ; implicit-def: $vgpr4813666; VI-NEXT:    ; implicit-def: $vgpr3813667; VI-NEXT:    ; implicit-def: $vgpr3613668; VI-NEXT:    s_waitcnt vmcnt(2)13669; VI-NEXT:    v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13670; VI-NEXT:    v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13671; VI-NEXT:    ; implicit-def: $vgpr4013672; VI-NEXT:    ; implicit-def: $vgpr3413673; VI-NEXT:    s_waitcnt vmcnt(1)13674; VI-NEXT:    v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13675; VI-NEXT:    s_waitcnt vmcnt(0)13676; VI-NEXT:    v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13677; VI-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13678; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload13679; VI-NEXT:    ; implicit-def: $vgpr5313680; VI-NEXT:    ; implicit-def: $vgpr5213681; VI-NEXT:    s_waitcnt vmcnt(0)13682; VI-NEXT:    v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13683; VI-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13684; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload13685; VI-NEXT:    ; implicit-def: $vgpr5113686; VI-NEXT:    s_waitcnt vmcnt(0)13687; VI-NEXT:    v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13688; VI-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13689; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload13690; VI-NEXT:    ; implicit-def: $vgpr4913691; VI-NEXT:    s_waitcnt vmcnt(0)13692; VI-NEXT:    v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13693; VI-NEXT:    v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13694; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload13695; VI-NEXT:    ; implicit-def: $vgpr3913696; VI-NEXT:    s_waitcnt vmcnt(0)13697; VI-NEXT:    v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13698; VI-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13699; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload13700; VI-NEXT:    ; implicit-def: $vgpr3713701; VI-NEXT:    s_waitcnt vmcnt(0)13702; VI-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13703; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13704; VI-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13705; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13706; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload13707; VI-NEXT:    ; implicit-def: $vgpr3313708; VI-NEXT:    ; kill: killed $vgpr3313709; VI-NEXT:    ; implicit-def: $vgpr3313710; VI-NEXT:    ; kill: killed $vgpr3313711; VI-NEXT:    ; implicit-def: $vgpr3313712; VI-NEXT:    ; kill: killed $vgpr3313713; VI-NEXT:    ; implicit-def: $vgpr3313714; VI-NEXT:    ; kill: killed $vgpr3313715; VI-NEXT:    ; implicit-def: $vgpr3313716; VI-NEXT:    ; kill: killed $vgpr3313717; VI-NEXT:    ; implicit-def: $vgpr3313718; VI-NEXT:    ; kill: killed $vgpr3313719; VI-NEXT:    ; implicit-def: $vgpr3313720; VI-NEXT:    ; kill: killed $vgpr3313721; VI-NEXT:    ; implicit-def: $vgpr3313722; VI-NEXT:    ; kill: killed $vgpr3313723; VI-NEXT:    ; implicit-def: $vgpr3313724; VI-NEXT:    ; kill: killed $vgpr3313725; VI-NEXT:    ; implicit-def: $vgpr3313726; VI-NEXT:    ; kill: killed $vgpr3313727; VI-NEXT:    ; implicit-def: $vgpr3313728; VI-NEXT:    ; kill: killed $vgpr3313729; VI-NEXT:    ; implicit-def: $vgpr3313730; VI-NEXT:    ; kill: killed $vgpr3313731; VI-NEXT:    ; implicit-def: $vgpr3313732; VI-NEXT:    ; kill: killed $vgpr3313733; VI-NEXT:    ; implicit-def: $vgpr3313734; VI-NEXT:    ; kill: killed $vgpr3313735; VI-NEXT:    ; implicit-def: $vgpr3313736; VI-NEXT:    ; kill: killed $vgpr3313737; VI-NEXT:    ; implicit-def: $vgpr3313738; VI-NEXT:    ; kill: killed $vgpr3313739; VI-NEXT:    ; implicit-def: $vgpr3313740; VI-NEXT:    ; kill: killed $vgpr3313741; VI-NEXT:    ; implicit-def: $vgpr3313742; VI-NEXT:    ; kill: killed $vgpr3313743; VI-NEXT:    ; implicit-def: $vgpr3313744; VI-NEXT:    ; kill: killed $vgpr3313745; VI-NEXT:    ; implicit-def: $vgpr3313746; VI-NEXT:    ; kill: killed $vgpr3313747; VI-NEXT:    ; implicit-def: $vgpr3313748; VI-NEXT:    ; kill: killed $vgpr3313749; VI-NEXT:    ; implicit-def: $vgpr3313750; VI-NEXT:    ; kill: killed $vgpr3313751; VI-NEXT:    ; implicit-def: $vgpr3313752; VI-NEXT:    ; kill: killed $vgpr3313753; VI-NEXT:    ; implicit-def: $vgpr3313754; VI-NEXT:    ; kill: killed $vgpr3313755; VI-NEXT:    ; implicit-def: $vgpr3313756; VI-NEXT:    ; kill: killed $vgpr3313757; VI-NEXT:    ; implicit-def: $vgpr3313758; VI-NEXT:    ; kill: killed $vgpr3313759; VI-NEXT:    ; implicit-def: $vgpr3313760; VI-NEXT:    ; kill: killed $vgpr3313761; VI-NEXT:    ; implicit-def: $vgpr3313762; VI-NEXT:    ; kill: killed $vgpr3313763; VI-NEXT:    ; implicit-def: $vgpr3313764; VI-NEXT:    ; kill: killed $vgpr3313765; VI-NEXT:    ; implicit-def: $vgpr3313766; VI-NEXT:    ; kill: killed $vgpr3313767; VI-NEXT:    ; implicit-def: $vgpr3313768; VI-NEXT:    ; kill: killed $vgpr3313769; VI-NEXT:    ; implicit-def: $vgpr3313770; VI-NEXT:    ; kill: killed $vgpr3313771; VI-NEXT:    ; implicit-def: $vgpr3313772; VI-NEXT:    ; kill: killed $vgpr3313773; VI-NEXT:    ; implicit-def: $vgpr3313774; VI-NEXT:    ; implicit-def: $vgpr6213775; VI-NEXT:    ; kill: killed $vgpr3313776; VI-NEXT:    ; implicit-def: $vgpr3513777; VI-NEXT:    ; implicit-def: $vgpr3313778; VI-NEXT:    s_waitcnt vmcnt(0)13779; VI-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13780; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13781; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload13782; VI-NEXT:    s_waitcnt vmcnt(0)13783; VI-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13784; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13785; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload13786; VI-NEXT:    ; implicit-def: $vgpr5813787; VI-NEXT:    s_waitcnt vmcnt(0)13788; VI-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13789; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13790; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload13791; VI-NEXT:    ; implicit-def: $vgpr5913792; VI-NEXT:    s_waitcnt vmcnt(0)13793; VI-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13794; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13795; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload13796; VI-NEXT:    ; implicit-def: $vgpr4713797; VI-NEXT:    s_waitcnt vmcnt(0)13798; VI-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13799; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13800; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload13801; VI-NEXT:    ; implicit-def: $vgpr4513802; VI-NEXT:    s_waitcnt vmcnt(0)13803; VI-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13804; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13805; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload13806; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload13807; VI-NEXT:    ; implicit-def: $vgpr4213808; VI-NEXT:    s_waitcnt vmcnt(0)13809; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13810; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload13811; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload13812; VI-NEXT:    s_waitcnt vmcnt(0)13813; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13814; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13815; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload13816; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload13817; VI-NEXT:    s_waitcnt vmcnt(0)13818; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13819; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload13820; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload13821; VI-NEXT:    s_waitcnt vmcnt(0)13822; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13823; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13824; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload13825; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload13826; VI-NEXT:    s_waitcnt vmcnt(0)13827; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13828; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload13829; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload13830; VI-NEXT:    s_waitcnt vmcnt(0)13831; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13832; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13833; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload13834; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload13835; VI-NEXT:    s_waitcnt vmcnt(0)13836; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13837; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload13838; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload13839; VI-NEXT:    s_waitcnt vmcnt(0)13840; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13841; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13842; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload13843; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload13844; VI-NEXT:    s_waitcnt vmcnt(0)13845; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13846; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload13847; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload13848; VI-NEXT:    s_waitcnt vmcnt(0)13849; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13850; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13851; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload13852; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload13853; VI-NEXT:    s_waitcnt vmcnt(0)13854; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13855; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload13856; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload13857; VI-NEXT:    s_waitcnt vmcnt(0)13858; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13859; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13860; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload13861; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload13862; VI-NEXT:    s_waitcnt vmcnt(0)13863; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13864; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload13865; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload13866; VI-NEXT:    s_waitcnt vmcnt(0)13867; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13868; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13869; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload13870; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload13871; VI-NEXT:    s_waitcnt vmcnt(0)13872; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13873; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload13874; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload13875; VI-NEXT:    s_waitcnt vmcnt(0)13876; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13877; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13878; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload13879; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload13880; VI-NEXT:    s_waitcnt vmcnt(0)13881; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13882; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload13883; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload13884; VI-NEXT:    s_waitcnt vmcnt(0)13885; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13886; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13887; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload13888; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload13889; VI-NEXT:    s_waitcnt vmcnt(0)13890; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13891; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload13892; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload13893; VI-NEXT:    s_waitcnt vmcnt(0)13894; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13895; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13896; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload13897; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload13898; VI-NEXT:    s_waitcnt vmcnt(0)13899; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13900; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload13901; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload13902; VI-NEXT:    s_waitcnt vmcnt(0)13903; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13904; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13905; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload13906; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload13907; VI-NEXT:    s_waitcnt vmcnt(0)13908; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13909; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload13910; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload13911; VI-NEXT:    s_waitcnt vmcnt(0)13912; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13913; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13914; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload13915; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload13916; VI-NEXT:    s_waitcnt vmcnt(0)13917; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13918; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload13919; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload13920; VI-NEXT:    s_waitcnt vmcnt(0)13921; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13922; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13923; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload13924; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload13925; VI-NEXT:    s_waitcnt vmcnt(0)13926; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13927; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload13928; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload13929; VI-NEXT:    s_waitcnt vmcnt(0)13930; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13931; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13932; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload13933; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload13934; VI-NEXT:    s_waitcnt vmcnt(0)13935; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13936; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload13937; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload13938; VI-NEXT:    s_waitcnt vmcnt(0)13939; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13940; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13941; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload13942; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload13943; VI-NEXT:    s_waitcnt vmcnt(0)13944; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13945; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload13946; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload13947; VI-NEXT:    s_waitcnt vmcnt(0)13948; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13949; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD13950; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload13951; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload13952; VI-NEXT:    s_waitcnt vmcnt(0)13953; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13954; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload13955; VI-NEXT:    s_waitcnt vmcnt(0)13956; VI-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD13957; VI-NEXT:    ; implicit-def: $vgpr6313958; VI-NEXT:    ; kill: killed $vgpr6313959; VI-NEXT:    ; implicit-def: $vgpr6313960; VI-NEXT:    ; kill: killed $vgpr6313961; VI-NEXT:    ; implicit-def: $vgpr6313962; VI-NEXT:    ; kill: killed $vgpr6313963; VI-NEXT:    ; implicit-def: $vgpr6313964; VI-NEXT:    ; kill: killed $vgpr6313965; VI-NEXT:    ; implicit-def: $vgpr6313966; VI-NEXT:    ; kill: killed $vgpr6313967; VI-NEXT:    ; implicit-def: $vgpr6313968; VI-NEXT:    ; kill: killed $vgpr6313969; VI-NEXT:    ; implicit-def: $vgpr6313970; VI-NEXT:    ; kill: killed $vgpr6313971; VI-NEXT:    ; implicit-def: $vgpr6313972; VI-NEXT:    ; kill: killed $vgpr6313973; VI-NEXT:    ; implicit-def: $vgpr6313974; VI-NEXT:    ; kill: killed $vgpr6313975; VI-NEXT:    ; implicit-def: $vgpr6313976; VI-NEXT:    ; kill: killed $vgpr6313977; VI-NEXT:    ; implicit-def: $vgpr6313978; VI-NEXT:    ; kill: killed $vgpr6313979; VI-NEXT:    ; implicit-def: $vgpr6313980; VI-NEXT:    ; kill: killed $vgpr6313981; VI-NEXT:    ; implicit-def: $vgpr6313982; VI-NEXT:    ; kill: killed $vgpr6313983; VI-NEXT:    ; implicit-def: $vgpr6313984; VI-NEXT:    ; kill: killed $vgpr6313985; VI-NEXT:    ; implicit-def: $vgpr6313986; VI-NEXT:    ; kill: killed $vgpr6313987; VI-NEXT:    ; implicit-def: $vgpr6313988; VI-NEXT:    ; kill: killed $vgpr6313989; VI-NEXT:    ; implicit-def: $vgpr6313990; VI-NEXT:    ; kill: killed $vgpr6313991; VI-NEXT:    ; implicit-def: $vgpr6313992; VI-NEXT:    ; kill: killed $vgpr6313993; VI-NEXT:    ; implicit-def: $vgpr6313994; VI-NEXT:    ; kill: killed $vgpr6313995; VI-NEXT:    ; implicit-def: $vgpr6313996; VI-NEXT:    ; kill: killed $vgpr6313997; VI-NEXT:    ; implicit-def: $vgpr6313998; VI-NEXT:    ; kill: killed $vgpr6313999; VI-NEXT:    ; implicit-def: $vgpr6314000; VI-NEXT:    ; kill: killed $vgpr6314001; VI-NEXT:    ; implicit-def: $vgpr6314002; VI-NEXT:    ; kill: killed $vgpr6314003; VI-NEXT:    ; implicit-def: $vgpr6314004; VI-NEXT:    ; kill: killed $vgpr6314005; VI-NEXT:    ; implicit-def: $vgpr6314006; VI-NEXT:    ; kill: killed $vgpr6314007; VI-NEXT:    ; implicit-def: $vgpr6314008; VI-NEXT:    ; kill: killed $vgpr6314009; VI-NEXT:    ; implicit-def: $vgpr6314010; VI-NEXT:    ; kill: killed $vgpr6314011; VI-NEXT:    ; implicit-def: $vgpr6314012; VI-NEXT:    ; kill: killed $vgpr6314013; VI-NEXT:    ; implicit-def: $vgpr6314014; VI-NEXT:    ; kill: killed $vgpr6314015; VI-NEXT:    ; implicit-def: $vgpr6314016; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14017; VI-NEXT:    ; implicit-def: $vgpr3214018; VI-NEXT:    ; kill: killed $vgpr6314019; VI-NEXT:    ; implicit-def: $vgpr6314020; VI-NEXT:    ; kill: killed $vgpr3214021; VI-NEXT:    ; implicit-def: $vgpr3214022; VI-NEXT:    ; kill: killed $vgpr6314023; VI-NEXT:    ; implicit-def: $vgpr6314024; VI-NEXT:    ; kill: killed $vgpr3214025; VI-NEXT:    ; implicit-def: $vgpr3214026; VI-NEXT:    ; kill: killed $vgpr6314027; VI-NEXT:    ; implicit-def: $vgpr6314028; VI-NEXT:    ; kill: killed $vgpr3214029; VI-NEXT:    ; implicit-def: $vgpr3214030; VI-NEXT:    ; kill: killed $vgpr6314031; VI-NEXT:    ; implicit-def: $vgpr6314032; VI-NEXT:    ; kill: killed $vgpr3214033; VI-NEXT:    ; implicit-def: $vgpr3214034; VI-NEXT:    ; kill: killed $vgpr6314035; VI-NEXT:    ; implicit-def: $vgpr6314036; VI-NEXT:    ; kill: killed $vgpr3214037; VI-NEXT:    ; implicit-def: $vgpr3214038; VI-NEXT:    ; kill: killed $vgpr6314039; VI-NEXT:    ; implicit-def: $vgpr6314040; VI-NEXT:    ; kill: killed $vgpr3214041; VI-NEXT:    ; implicit-def: $vgpr3214042; VI-NEXT:    ; kill: killed $vgpr6314043; VI-NEXT:    ; implicit-def: $vgpr6314044; VI-NEXT:    ; kill: killed $vgpr3214045; VI-NEXT:    ; implicit-def: $vgpr3214046; VI-NEXT:    ; kill: killed $vgpr6314047; VI-NEXT:    ; implicit-def: $vgpr6314048; VI-NEXT:    ; kill: killed $vgpr3214049; VI-NEXT:    ; implicit-def: $vgpr3214050; VI-NEXT:    ; kill: killed $vgpr6314051; VI-NEXT:    ; implicit-def: $vgpr6314052; VI-NEXT:    ; kill: killed $vgpr3214053; VI-NEXT:    ; implicit-def: $vgpr3214054; VI-NEXT:    ; kill: killed $vgpr6314055; VI-NEXT:    ; implicit-def: $vgpr6314056; VI-NEXT:    ; kill: killed $vgpr3214057; VI-NEXT:    ; implicit-def: $vgpr3214058; VI-NEXT:    ; kill: killed $vgpr6314059; VI-NEXT:    ; implicit-def: $vgpr6314060; VI-NEXT:    ; kill: killed $vgpr3214061; VI-NEXT:    ; implicit-def: $vgpr3214062; VI-NEXT:    ; kill: killed $vgpr6314063; VI-NEXT:    ; implicit-def: $vgpr6314064; VI-NEXT:    ; kill: killed $vgpr3214065; VI-NEXT:    ; implicit-def: $vgpr3214066; VI-NEXT:    ; kill: killed $vgpr6314067; VI-NEXT:    ; implicit-def: $vgpr6314068; VI-NEXT:    ; kill: killed $vgpr3214069; VI-NEXT:    ; implicit-def: $vgpr3214070; VI-NEXT:    ; kill: killed $vgpr6314071; VI-NEXT:    ; implicit-def: $vgpr6314072; VI-NEXT:    ; kill: killed $vgpr3214073; VI-NEXT:    ; implicit-def: $vgpr3214074; VI-NEXT:    ; kill: killed $vgpr6314075; VI-NEXT:    ; implicit-def: $vgpr6314076; VI-NEXT:    ; kill: killed $vgpr3214077; VI-NEXT:    ; implicit-def: $vgpr3214078; VI-NEXT:    ; kill: killed $vgpr6314079; VI-NEXT:    ; implicit-def: $vgpr6314080; VI-NEXT:    ; kill: killed $vgpr3214081; VI-NEXT:    ; implicit-def: $vgpr3214082; VI-NEXT:    ; kill: killed $vgpr6314083; VI-NEXT:    ; implicit-def: $vgpr6314084; VI-NEXT:  .LBB14_2: ; %Flow14085; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]14086; VI-NEXT:    s_cbranch_execz .LBB14_414087; VI-NEXT:  ; %bb.3: ; %cmp.true14088; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload14089; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload14090; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload14091; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload14092; VI-NEXT:    v_add_u16_e32 v9, 3, v6114093; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload14094; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload14095; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload14096; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload14097; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload14098; VI-NEXT:    v_mov_b32_e32 v31, 0x30014099; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload14100; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload14101; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload14102; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload14103; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload14104; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload14105; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload14106; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload14107; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload14108; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload14109; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload14110; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload14111; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload14112; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload14113; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload14114; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload14115; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload14116; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload14117; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload14118; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload14119; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload14120; VI-NEXT:    s_waitcnt vmcnt(14)14121; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014122; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload14123; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14124; VI-NEXT:    v_add_u16_e32 v8, 3, v814125; VI-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014126; VI-NEXT:    v_add_u16_e32 v0, 3, v014127; VI-NEXT:    v_add_u16_e32 v2, 3, v214128; VI-NEXT:    v_add_u16_e32 v3, 3, v314129; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014130; VI-NEXT:    v_add_u16_e32 v1, 3, v114131; VI-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014132; VI-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014133; VI-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014134; VI-NEXT:    v_add_u16_e32 v0, 0x300, v014135; VI-NEXT:    v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14136; VI-NEXT:    v_add_u16_e32 v2, 0x300, v214137; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14138; VI-NEXT:    v_or_b32_e32 v0, v0, v114139; VI-NEXT:    v_or_b32_e32 v1, v2, v314140; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload14141; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload14142; VI-NEXT:    v_add_u16_e32 v4, 3, v414143; VI-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014144; VI-NEXT:    v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14145; VI-NEXT:    s_waitcnt vmcnt(14)14146; VI-NEXT:    v_add_u16_e32 v5, 3, v514147; VI-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014148; VI-NEXT:    v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14149; VI-NEXT:    s_waitcnt vmcnt(13)14150; VI-NEXT:    v_add_u16_e32 v6, 3, v614151; VI-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014152; VI-NEXT:    v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14153; VI-NEXT:    s_waitcnt vmcnt(10)14154; VI-NEXT:    v_add_u16_e32 v7, 3, v714155; VI-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014156; VI-NEXT:    v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14157; VI-NEXT:    v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14158; VI-NEXT:    s_waitcnt vmcnt(1)14159; VI-NEXT:    v_add_u16_e32 v2, 3, v214160; VI-NEXT:    s_waitcnt vmcnt(0)14161; VI-NEXT:    v_add_u16_e32 v3, 3, v314162; VI-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014163; VI-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014164; VI-NEXT:    v_add_u16_e32 v2, 0x300, v214165; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14166; VI-NEXT:    v_or_b32_e32 v2, v2, v314167; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload14168; VI-NEXT:    s_waitcnt vmcnt(0)14169; VI-NEXT:    v_add_u16_e32 v3, 3, v314170; VI-NEXT:    v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014171; VI-NEXT:    v_add_u16_e32 v3, 0x300, v314172; VI-NEXT:    v_or_b32_e32 v3, v3, v414173; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload14174; VI-NEXT:    s_waitcnt vmcnt(0)14175; VI-NEXT:    v_add_u16_e32 v4, 3, v414176; VI-NEXT:    v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014177; VI-NEXT:    v_add_u16_e32 v4, 0x300, v414178; VI-NEXT:    v_or_b32_e32 v4, v4, v514179; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload14180; VI-NEXT:    s_waitcnt vmcnt(0)14181; VI-NEXT:    v_add_u16_e32 v5, 3, v514182; VI-NEXT:    v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014183; VI-NEXT:    v_add_u16_e32 v5, 0x300, v514184; VI-NEXT:    v_or_b32_e32 v5, v5, v614185; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload14186; VI-NEXT:    s_waitcnt vmcnt(0)14187; VI-NEXT:    v_add_u16_e32 v6, 3, v614188; VI-NEXT:    v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014189; VI-NEXT:    v_add_u16_e32 v6, 0x300, v614190; VI-NEXT:    v_or_b32_e32 v6, v6, v714191; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload14192; VI-NEXT:    s_waitcnt vmcnt(0)14193; VI-NEXT:    v_add_u16_e32 v7, 3, v714194; VI-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014195; VI-NEXT:    v_add_u16_e32 v7, 0x300, v714196; VI-NEXT:    v_or_b32_e32 v7, v7, v814197; VI-NEXT:    v_add_u16_e32 v8, 3, v6214198; VI-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014199; VI-NEXT:    v_add_u16_e32 v8, 0x300, v814200; VI-NEXT:    v_or_b32_e32 v8, v8, v914201; VI-NEXT:    v_add_u16_e32 v9, 3, v3214202; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014203; VI-NEXT:    v_add_u16_e32 v10, 3, v6014204; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014205; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload14206; VI-NEXT:    v_add_u16_e32 v9, 0x300, v914207; VI-NEXT:    v_add_u16_sdwa v10, v10, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14208; VI-NEXT:    v_or_b32_e32 v9, v9, v1014209; VI-NEXT:    v_add_u16_e32 v10, 3, v5814210; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload14211; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload14212; VI-NEXT:    s_waitcnt vmcnt(2)14213; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014214; VI-NEXT:    v_add_u16_e32 v11, 3, v5614215; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014216; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload14217; VI-NEXT:    v_add_u16_e32 v10, 0x300, v1014218; VI-NEXT:    v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14219; VI-NEXT:    v_or_b32_e32 v10, v10, v1114220; VI-NEXT:    v_add_u16_e32 v11, 3, v5914221; VI-NEXT:    s_waitcnt vmcnt(0)14222; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014223; VI-NEXT:    v_add_u16_e32 v12, 3, v5714224; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014225; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload14226; VI-NEXT:    v_add_u16_e32 v11, 0x300, v1114227; VI-NEXT:    v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14228; VI-NEXT:    v_or_b32_e32 v11, v11, v1214229; VI-NEXT:    v_add_u16_e32 v12, 3, v4714230; VI-NEXT:    s_waitcnt vmcnt(0)14231; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014232; VI-NEXT:    v_add_u16_e32 v13, 3, v4614233; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014234; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload14235; VI-NEXT:    v_add_u16_e32 v12, 0x300, v1214236; VI-NEXT:    v_add_u16_sdwa v13, v13, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14237; VI-NEXT:    v_or_b32_e32 v12, v12, v1314238; VI-NEXT:    v_add_u16_e32 v13, 3, v4514239; VI-NEXT:    s_waitcnt vmcnt(0)14240; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014241; VI-NEXT:    v_add_u16_e32 v14, 3, v4414242; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014243; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload14244; VI-NEXT:    v_add_u16_e32 v13, 0x300, v1314245; VI-NEXT:    v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14246; VI-NEXT:    v_or_b32_e32 v13, v13, v1414247; VI-NEXT:    v_add_u16_e32 v14, 3, v4214248; VI-NEXT:    s_waitcnt vmcnt(0)14249; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014250; VI-NEXT:    v_add_u16_e32 v15, 3, v4014251; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014252; VI-NEXT:    v_add_u16_e32 v14, 0x300, v1414253; VI-NEXT:    v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14254; VI-NEXT:    v_or_b32_e32 v14, v14, v1514255; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload14256; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload14257; VI-NEXT:    s_waitcnt vmcnt(1)14258; VI-NEXT:    v_add_u16_e32 v15, 3, v1514259; VI-NEXT:    s_waitcnt vmcnt(0)14260; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014261; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload14262; VI-NEXT:    v_add_u16_e32 v15, 0x300, v1514263; VI-NEXT:    s_waitcnt vmcnt(0)14264; VI-NEXT:    v_add_u16_e32 v16, 3, v1614265; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014266; VI-NEXT:    v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14267; VI-NEXT:    v_or_b32_e32 v15, v15, v1614268; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload14269; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload14270; VI-NEXT:    s_waitcnt vmcnt(1)14271; VI-NEXT:    v_add_u16_e32 v16, 3, v1614272; VI-NEXT:    s_waitcnt vmcnt(0)14273; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014274; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload14275; VI-NEXT:    v_add_u16_e32 v16, 0x300, v1614276; VI-NEXT:    s_waitcnt vmcnt(0)14277; VI-NEXT:    v_add_u16_e32 v17, 3, v1714278; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014279; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14280; VI-NEXT:    v_or_b32_e32 v16, v16, v1714281; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload14282; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload14283; VI-NEXT:    s_waitcnt vmcnt(1)14284; VI-NEXT:    v_add_u16_e32 v17, 3, v1714285; VI-NEXT:    s_waitcnt vmcnt(0)14286; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014287; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload14288; VI-NEXT:    v_add_u16_e32 v17, 0x300, v1714289; VI-NEXT:    s_waitcnt vmcnt(0)14290; VI-NEXT:    v_add_u16_e32 v18, 3, v1814291; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014292; VI-NEXT:    v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14293; VI-NEXT:    v_or_b32_e32 v17, v17, v1814294; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload14295; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload14296; VI-NEXT:    s_waitcnt vmcnt(1)14297; VI-NEXT:    v_add_u16_e32 v18, 3, v1814298; VI-NEXT:    s_waitcnt vmcnt(0)14299; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014300; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload14301; VI-NEXT:    v_add_u16_e32 v18, 0x300, v1814302; VI-NEXT:    s_waitcnt vmcnt(0)14303; VI-NEXT:    v_add_u16_e32 v19, 3, v1914304; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014305; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14306; VI-NEXT:    v_or_b32_e32 v18, v18, v1914307; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload14308; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload14309; VI-NEXT:    s_waitcnt vmcnt(1)14310; VI-NEXT:    v_add_u16_e32 v19, 3, v1914311; VI-NEXT:    s_waitcnt vmcnt(0)14312; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014313; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload14314; VI-NEXT:    v_add_u16_e32 v19, 0x300, v1914315; VI-NEXT:    s_waitcnt vmcnt(0)14316; VI-NEXT:    v_add_u16_e32 v20, 3, v2014317; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014318; VI-NEXT:    v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14319; VI-NEXT:    v_or_b32_e32 v19, v19, v2014320; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload14321; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload14322; VI-NEXT:    s_waitcnt vmcnt(1)14323; VI-NEXT:    v_add_u16_e32 v20, 3, v2014324; VI-NEXT:    s_waitcnt vmcnt(0)14325; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014326; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload14327; VI-NEXT:    v_add_u16_e32 v20, 0x300, v2014328; VI-NEXT:    s_waitcnt vmcnt(0)14329; VI-NEXT:    v_add_u16_e32 v21, 3, v2114330; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014331; VI-NEXT:    v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14332; VI-NEXT:    v_or_b32_e32 v20, v20, v2114333; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload14334; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload14335; VI-NEXT:    s_waitcnt vmcnt(1)14336; VI-NEXT:    v_add_u16_e32 v21, 3, v2114337; VI-NEXT:    s_waitcnt vmcnt(0)14338; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014339; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload14340; VI-NEXT:    v_add_u16_e32 v21, 0x300, v2114341; VI-NEXT:    s_waitcnt vmcnt(0)14342; VI-NEXT:    v_add_u16_e32 v22, 3, v2214343; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014344; VI-NEXT:    v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14345; VI-NEXT:    v_or_b32_e32 v21, v21, v2214346; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload14347; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload14348; VI-NEXT:    s_waitcnt vmcnt(1)14349; VI-NEXT:    v_add_u16_e32 v22, 3, v2214350; VI-NEXT:    s_waitcnt vmcnt(0)14351; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014352; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload14353; VI-NEXT:    v_add_u16_e32 v22, 0x300, v2214354; VI-NEXT:    s_waitcnt vmcnt(0)14355; VI-NEXT:    v_add_u16_e32 v23, 3, v2314356; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014357; VI-NEXT:    v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14358; VI-NEXT:    v_or_b32_e32 v22, v22, v2314359; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload14360; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload14361; VI-NEXT:    s_waitcnt vmcnt(1)14362; VI-NEXT:    v_add_u16_e32 v23, 3, v2314363; VI-NEXT:    s_waitcnt vmcnt(0)14364; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014365; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload14366; VI-NEXT:    v_add_u16_e32 v23, 0x300, v2314367; VI-NEXT:    s_waitcnt vmcnt(0)14368; VI-NEXT:    v_add_u16_e32 v24, 3, v2414369; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014370; VI-NEXT:    v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14371; VI-NEXT:    v_or_b32_e32 v23, v23, v2414372; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload14373; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload14374; VI-NEXT:    s_waitcnt vmcnt(1)14375; VI-NEXT:    v_add_u16_e32 v24, 3, v2414376; VI-NEXT:    s_waitcnt vmcnt(0)14377; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014378; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload14379; VI-NEXT:    v_add_u16_e32 v24, 0x300, v2414380; VI-NEXT:    s_waitcnt vmcnt(0)14381; VI-NEXT:    v_add_u16_e32 v25, 3, v2514382; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014383; VI-NEXT:    v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14384; VI-NEXT:    v_or_b32_e32 v24, v24, v2514385; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload14386; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload14387; VI-NEXT:    s_waitcnt vmcnt(1)14388; VI-NEXT:    v_add_u16_e32 v25, 3, v2514389; VI-NEXT:    s_waitcnt vmcnt(0)14390; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014391; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload14392; VI-NEXT:    v_add_u16_e32 v25, 0x300, v2514393; VI-NEXT:    s_waitcnt vmcnt(0)14394; VI-NEXT:    v_add_u16_e32 v26, 3, v2614395; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014396; VI-NEXT:    v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14397; VI-NEXT:    v_or_b32_e32 v25, v25, v2614398; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload14399; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload14400; VI-NEXT:    s_waitcnt vmcnt(1)14401; VI-NEXT:    v_add_u16_e32 v26, 3, v2614402; VI-NEXT:    s_waitcnt vmcnt(0)14403; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014404; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload14405; VI-NEXT:    v_add_u16_e32 v26, 0x300, v2614406; VI-NEXT:    s_waitcnt vmcnt(0)14407; VI-NEXT:    v_add_u16_e32 v27, 3, v2714408; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014409; VI-NEXT:    v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14410; VI-NEXT:    v_or_b32_e32 v26, v26, v2714411; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload14412; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload14413; VI-NEXT:    s_waitcnt vmcnt(1)14414; VI-NEXT:    v_add_u16_e32 v27, 3, v2714415; VI-NEXT:    s_waitcnt vmcnt(0)14416; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014417; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload14418; VI-NEXT:    v_add_u16_e32 v27, 0x300, v2714419; VI-NEXT:    s_waitcnt vmcnt(0)14420; VI-NEXT:    v_add_u16_e32 v28, 3, v2814421; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014422; VI-NEXT:    v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14423; VI-NEXT:    v_or_b32_e32 v27, v27, v2814424; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload14425; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload14426; VI-NEXT:    s_waitcnt vmcnt(1)14427; VI-NEXT:    v_add_u16_e32 v28, 3, v2814428; VI-NEXT:    s_waitcnt vmcnt(0)14429; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014430; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload14431; VI-NEXT:    v_add_u16_e32 v28, 0x300, v2814432; VI-NEXT:    s_waitcnt vmcnt(0)14433; VI-NEXT:    v_add_u16_e32 v29, 3, v2914434; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014435; VI-NEXT:    v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14436; VI-NEXT:    v_or_b32_e32 v28, v28, v2914437; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload14438; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload14439; VI-NEXT:    s_waitcnt vmcnt(1)14440; VI-NEXT:    v_add_u16_e32 v29, 3, v2914441; VI-NEXT:    s_waitcnt vmcnt(0)14442; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014443; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload14444; VI-NEXT:    v_add_u16_e32 v29, 0x300, v2914445; VI-NEXT:    s_waitcnt vmcnt(0)14446; VI-NEXT:    v_add_u16_e32 v30, 3, v3014447; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014448; VI-NEXT:    v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14449; VI-NEXT:    v_or_b32_e32 v29, v29, v3014450; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload14451; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload14452; VI-NEXT:    s_waitcnt vmcnt(1)14453; VI-NEXT:    v_add_u16_e32 v30, 3, v3014454; VI-NEXT:    s_waitcnt vmcnt(0)14455; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014456; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload14457; VI-NEXT:    v_add_u16_e32 v30, 0x300, v3014458; VI-NEXT:    s_waitcnt vmcnt(0)14459; VI-NEXT:    v_add_u16_e32 v32, 3, v3214460; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014461; VI-NEXT:    v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14462; VI-NEXT:    v_or_b32_e32 v30, v30, v3214463; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload14464; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload14465; VI-NEXT:    s_waitcnt vmcnt(1)14466; VI-NEXT:    v_add_u16_e32 v32, 3, v3214467; VI-NEXT:    s_waitcnt vmcnt(0)14468; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014469; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload14470; VI-NEXT:    v_add_u16_e32 v32, 0x300, v3214471; VI-NEXT:    s_waitcnt vmcnt(0)14472; VI-NEXT:    v_add_u16_e32 v33, 3, v3314473; VI-NEXT:    v_or_b32_sdwa v33, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_014474; VI-NEXT:    v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD14475; VI-NEXT:    v_or_b32_e32 v31, v32, v3114476; VI-NEXT:  .LBB14_4: ; %end14477; VI-NEXT:    s_or_b64 exec, exec, s[4:5]14478; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload14479; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload14480; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload14481; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload14482; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload14483; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload14484; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload14485; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload14486; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload14487; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload14488; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload14489; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload14490; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload14491; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload14492; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload14493; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload14494; VI-NEXT:    s_waitcnt vmcnt(0)14495; VI-NEXT:    s_setpc_b64 s[30:31]14496;14497; GFX9-LABEL: bitcast_v128i8_to_v32i32:14498; GFX9:       ; %bb.0:14499; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)14500; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill14501; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill14502; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill14503; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill14504; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill14505; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill14506; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill14507; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill14508; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill14509; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill14510; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill14511; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill14512; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill14513; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill14514; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill14515; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill14516; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill14517; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill14518; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill14519; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill14520; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill14521; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill14522; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill14523; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill14524; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill14525; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill14526; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill14527; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill14528; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill14529; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill14530; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill14531; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill14532; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:38814533; GFX9-NEXT:    s_nop 014534; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s3214535; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:814536; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:1614537; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:2414538; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:3214539; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:4014540; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:4814541; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:5614542; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:6414543; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:7214544; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:8014545; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:8814546; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:9614547; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:10414548; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:11214549; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:12014550; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:12814551; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:13614552; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:14414553; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:15214554; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:16014555; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:16814556; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:17614557; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:18414558; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v114559; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v314560; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v514561; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v714562; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v914563; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v1114564; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v1314565; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v1514566; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v1714567; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v1914568; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v2114569; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v2314570; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v2514571; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v2714572; GFX9-NEXT:    v_lshlrev_b16_e32 v35, 8, v2914573; GFX9-NEXT:    s_waitcnt vmcnt(24)14574; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3014575; GFX9-NEXT:    s_waitcnt vmcnt(23)14576; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v014577; GFX9-NEXT:    s_waitcnt vmcnt(22)14578; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v214579; GFX9-NEXT:    s_waitcnt vmcnt(21)14580; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v414581; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill14582; GFX9-NEXT:    s_waitcnt vmcnt(21)14583; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v614584; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill14585; GFX9-NEXT:    s_waitcnt vmcnt(21)14586; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v814587; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill14588; GFX9-NEXT:    s_waitcnt vmcnt(21)14589; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1014590; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill14591; GFX9-NEXT:    s_waitcnt vmcnt(21)14592; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1214593; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill14594; GFX9-NEXT:    s_waitcnt vmcnt(21)14595; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1414596; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill14597; GFX9-NEXT:    s_waitcnt vmcnt(21)14598; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1614599; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill14600; GFX9-NEXT:    s_waitcnt vmcnt(21)14601; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1814602; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill14603; GFX9-NEXT:    s_waitcnt vmcnt(21)14604; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2014605; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill14606; GFX9-NEXT:    s_waitcnt vmcnt(21)14607; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2214608; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill14609; GFX9-NEXT:    s_waitcnt vmcnt(21)14610; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2414611; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill14612; GFX9-NEXT:    s_waitcnt vmcnt(21)14613; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2614614; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill14615; GFX9-NEXT:    s_waitcnt vmcnt(21)14616; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2814617; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill14618; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:11614619; GFX9-NEXT:    s_waitcnt vmcnt(0)14620; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill14621; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3114622; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill14623; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:12414624; GFX9-NEXT:    s_waitcnt vmcnt(0)14625; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill14626; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:13214627; GFX9-NEXT:    s_waitcnt vmcnt(0)14628; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill14629; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:14014630; GFX9-NEXT:    s_waitcnt vmcnt(0)14631; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill14632; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:14814633; GFX9-NEXT:    s_waitcnt vmcnt(0)14634; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill14635; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3214636; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill14637; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4014638; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill14639; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4214640; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill14641; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4414642; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill14643; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:15614644; GFX9-NEXT:    s_waitcnt vmcnt(0)14645; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill14646; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16414647; GFX9-NEXT:    s_waitcnt vmcnt(0)14648; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill14649; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:17214650; GFX9-NEXT:    s_waitcnt vmcnt(0)14651; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill14652; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:18014653; GFX9-NEXT:    s_waitcnt vmcnt(0)14654; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill14655; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4514656; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill14657; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4614658; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill14659; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4714660; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill14661; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v5614662; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill14663; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:19214664; GFX9-NEXT:    s_nop 014665; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:20014666; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:20814667; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:21614668; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:18814669; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:4414670; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:3614671; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:2814672; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:2014673; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:1214674; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:414675; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:10814676; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:10014677; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:9214678; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:8414679; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:7614680; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:6814681; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:6014682; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:5214683; GFX9-NEXT:    s_waitcnt vmcnt(18)14684; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v014685; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill14686; GFX9-NEXT:    s_waitcnt vmcnt(18)14687; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v114688; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill14689; GFX9-NEXT:    s_waitcnt vmcnt(16)14690; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill14691; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:19614692; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v214693; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill14694; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v314695; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill14696; GFX9-NEXT:    s_waitcnt vmcnt(2)14697; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill14698; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:20414699; GFX9-NEXT:    s_waitcnt vmcnt(0)14700; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill14701; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:21214702; GFX9-NEXT:    s_waitcnt vmcnt(0)14703; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill14704; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:22414705; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:23214706; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:24014707; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:24814708; GFX9-NEXT:    s_nop 014709; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:22014710; GFX9-NEXT:    s_waitcnt vmcnt(4)14711; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v014712; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill14713; GFX9-NEXT:    s_waitcnt vmcnt(4)14714; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v114715; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill14716; GFX9-NEXT:    s_waitcnt vmcnt(2)14717; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill14718; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:22814719; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v214720; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill14721; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v314722; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill14723; GFX9-NEXT:    s_waitcnt vmcnt(2)14724; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill14725; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:23614726; GFX9-NEXT:    s_waitcnt vmcnt(0)14727; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill14728; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:24414729; GFX9-NEXT:    s_waitcnt vmcnt(0)14730; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill14731; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:25614732; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:26414733; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:27214734; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:28014735; GFX9-NEXT:    s_nop 014736; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:25214737; GFX9-NEXT:    s_waitcnt vmcnt(4)14738; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v014739; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill14740; GFX9-NEXT:    s_waitcnt vmcnt(4)14741; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v114742; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill14743; GFX9-NEXT:    s_waitcnt vmcnt(2)14744; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill14745; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:26014746; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v214747; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill14748; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v314749; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill14750; GFX9-NEXT:    s_waitcnt vmcnt(2)14751; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill14752; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:26814753; GFX9-NEXT:    s_waitcnt vmcnt(0)14754; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill14755; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:27614756; GFX9-NEXT:    s_waitcnt vmcnt(0)14757; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill14758; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:28814759; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:29614760; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:30414761; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:31214762; GFX9-NEXT:    s_nop 014763; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:28414764; GFX9-NEXT:    s_waitcnt vmcnt(4)14765; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v014766; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill14767; GFX9-NEXT:    s_waitcnt vmcnt(4)14768; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v114769; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill14770; GFX9-NEXT:    s_waitcnt vmcnt(2)14771; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill14772; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:29214773; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v214774; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill14775; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v314776; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill14777; GFX9-NEXT:    s_waitcnt vmcnt(2)14778; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill14779; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:30014780; GFX9-NEXT:    s_waitcnt vmcnt(0)14781; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill14782; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:30814783; GFX9-NEXT:    s_waitcnt vmcnt(0)14784; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill14785; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:32014786; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:32814787; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:33614788; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:34414789; GFX9-NEXT:    s_nop 014790; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:31614791; GFX9-NEXT:    s_waitcnt vmcnt(4)14792; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v014793; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill14794; GFX9-NEXT:    s_waitcnt vmcnt(4)14795; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v114796; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill14797; GFX9-NEXT:    s_waitcnt vmcnt(2)14798; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill14799; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32414800; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v214801; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill14802; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v314803; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill14804; GFX9-NEXT:    s_waitcnt vmcnt(2)14805; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill14806; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:33214807; GFX9-NEXT:    s_waitcnt vmcnt(0)14808; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill14809; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:34014810; GFX9-NEXT:    s_waitcnt vmcnt(0)14811; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill14812; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:35214813; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:36014814; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:36814815; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:37614816; GFX9-NEXT:    s_nop 014817; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:34814818; GFX9-NEXT:    s_waitcnt vmcnt(4)14819; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v014820; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill14821; GFX9-NEXT:    s_waitcnt vmcnt(4)14822; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v114823; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill14824; GFX9-NEXT:    s_waitcnt vmcnt(2)14825; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill14826; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:35614827; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v214828; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill14829; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v314830; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill14831; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:38414832; GFX9-NEXT:    s_nop 014833; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:38014834; GFX9-NEXT:    s_waitcnt vmcnt(1)14835; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v014836; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill14837; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:36414838; GFX9-NEXT:    s_waitcnt vmcnt(2)14839; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill14840; GFX9-NEXT:    s_waitcnt vmcnt(1)14841; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill14842; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:37214843; GFX9-NEXT:    s_waitcnt vmcnt(0)14844; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill14845; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr3114846; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc14847; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]14848; GFX9-NEXT:    s_cbranch_execz .LBB14_214849; GFX9-NEXT:  ; %bb.1: ; %cmp.false14850; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload14851; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload14852; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload14853; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload14854; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload14855; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload14856; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload14857; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload14858; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload14859; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload14860; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload14861; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload14862; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload14863; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload14864; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload14865; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload14866; GFX9-NEXT:    s_waitcnt vmcnt(15)14867; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14868; GFX9-NEXT:    s_waitcnt vmcnt(14)14869; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14870; GFX9-NEXT:    s_waitcnt vmcnt(13)14871; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14872; GFX9-NEXT:    s_waitcnt vmcnt(12)14873; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14874; GFX9-NEXT:    s_waitcnt vmcnt(11)14875; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14876; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14877; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14878; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload14879; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload14880; GFX9-NEXT:    s_waitcnt vmcnt(10)14881; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14882; GFX9-NEXT:    s_waitcnt vmcnt(8)14883; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14884; GFX9-NEXT:    s_waitcnt vmcnt(6)14885; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14886; GFX9-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14887; GFX9-NEXT:    s_waitcnt vmcnt(4)14888; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14889; GFX9-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14890; GFX9-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14891; GFX9-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14892; GFX9-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14893; GFX9-NEXT:    ; implicit-def: $vgpr6114894; GFX9-NEXT:    ; implicit-def: $vgpr6014895; GFX9-NEXT:    ; implicit-def: $vgpr5614896; GFX9-NEXT:    ; implicit-def: $vgpr5714897; GFX9-NEXT:    ; implicit-def: $vgpr4614898; GFX9-NEXT:    ; implicit-def: $vgpr4414899; GFX9-NEXT:    ; implicit-def: $vgpr4314900; GFX9-NEXT:    ; implicit-def: $vgpr4114901; GFX9-NEXT:    ; implicit-def: $vgpr5514902; GFX9-NEXT:    ; implicit-def: $vgpr5414903; GFX9-NEXT:    ; implicit-def: $vgpr5014904; GFX9-NEXT:    ; implicit-def: $vgpr4814905; GFX9-NEXT:    ; implicit-def: $vgpr3814906; GFX9-NEXT:    ; implicit-def: $vgpr3614907; GFX9-NEXT:    s_waitcnt vmcnt(2)14908; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14909; GFX9-NEXT:    v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14910; GFX9-NEXT:    ; implicit-def: $vgpr4014911; GFX9-NEXT:    ; implicit-def: $vgpr3414912; GFX9-NEXT:    s_waitcnt vmcnt(1)14913; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14914; GFX9-NEXT:    s_waitcnt vmcnt(0)14915; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14916; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14917; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload14918; GFX9-NEXT:    ; implicit-def: $vgpr5314919; GFX9-NEXT:    ; implicit-def: $vgpr5214920; GFX9-NEXT:    s_waitcnt vmcnt(0)14921; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14922; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14923; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload14924; GFX9-NEXT:    ; implicit-def: $vgpr5114925; GFX9-NEXT:    s_waitcnt vmcnt(0)14926; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14927; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14928; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload14929; GFX9-NEXT:    ; implicit-def: $vgpr4914930; GFX9-NEXT:    s_waitcnt vmcnt(0)14931; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14932; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14933; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload14934; GFX9-NEXT:    ; implicit-def: $vgpr3914935; GFX9-NEXT:    s_waitcnt vmcnt(0)14936; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14937; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14938; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload14939; GFX9-NEXT:    ; implicit-def: $vgpr3714940; GFX9-NEXT:    s_waitcnt vmcnt(0)14941; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14942; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14943; GFX9-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD14944; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD14945; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload14946; GFX9-NEXT:    ; implicit-def: $vgpr3314947; GFX9-NEXT:    ; kill: killed $vgpr3314948; GFX9-NEXT:    ; implicit-def: $vgpr3314949; GFX9-NEXT:    ; kill: killed $vgpr3314950; GFX9-NEXT:    ; implicit-def: $vgpr3314951; GFX9-NEXT:    ; kill: killed $vgpr3314952; GFX9-NEXT:    ; implicit-def: $vgpr3314953; GFX9-NEXT:    ; kill: killed $vgpr3314954; GFX9-NEXT:    ; implicit-def: $vgpr3314955; GFX9-NEXT:    ; kill: killed $vgpr3314956; GFX9-NEXT:    ; implicit-def: $vgpr3314957; GFX9-NEXT:    ; kill: killed $vgpr3314958; GFX9-NEXT:    ; implicit-def: $vgpr3314959; GFX9-NEXT:    ; kill: killed $vgpr3314960; GFX9-NEXT:    ; implicit-def: $vgpr3314961; GFX9-NEXT:    ; kill: killed $vgpr3314962; GFX9-NEXT:    ; implicit-def: $vgpr3314963; GFX9-NEXT:    ; kill: killed $vgpr3314964; GFX9-NEXT:    ; implicit-def: $vgpr3314965; GFX9-NEXT:    ; kill: killed $vgpr3314966; GFX9-NEXT:    ; implicit-def: $vgpr3314967; GFX9-NEXT:    ; kill: killed $vgpr3314968; GFX9-NEXT:    ; implicit-def: $vgpr3314969; GFX9-NEXT:    ; kill: killed $vgpr3314970; GFX9-NEXT:    ; implicit-def: $vgpr3314971; GFX9-NEXT:    ; kill: killed $vgpr3314972; GFX9-NEXT:    ; implicit-def: $vgpr3314973; GFX9-NEXT:    ; kill: killed $vgpr3314974; GFX9-NEXT:    ; implicit-def: $vgpr3314975; GFX9-NEXT:    ; kill: killed $vgpr3314976; GFX9-NEXT:    ; implicit-def: $vgpr3314977; GFX9-NEXT:    ; kill: killed $vgpr3314978; GFX9-NEXT:    ; implicit-def: $vgpr3314979; GFX9-NEXT:    ; kill: killed $vgpr3314980; GFX9-NEXT:    ; implicit-def: $vgpr3314981; GFX9-NEXT:    ; kill: killed $vgpr3314982; GFX9-NEXT:    ; implicit-def: $vgpr3314983; GFX9-NEXT:    ; kill: killed $vgpr3314984; GFX9-NEXT:    ; implicit-def: $vgpr3314985; GFX9-NEXT:    ; kill: killed $vgpr3314986; GFX9-NEXT:    ; implicit-def: $vgpr3314987; GFX9-NEXT:    ; kill: killed $vgpr3314988; GFX9-NEXT:    ; implicit-def: $vgpr3314989; GFX9-NEXT:    ; kill: killed $vgpr3314990; GFX9-NEXT:    ; implicit-def: $vgpr3314991; GFX9-NEXT:    ; kill: killed $vgpr3314992; GFX9-NEXT:    ; implicit-def: $vgpr3314993; GFX9-NEXT:    ; kill: killed $vgpr3314994; GFX9-NEXT:    ; implicit-def: $vgpr3314995; GFX9-NEXT:    ; kill: killed $vgpr3314996; GFX9-NEXT:    ; implicit-def: $vgpr3314997; GFX9-NEXT:    ; kill: killed $vgpr3314998; GFX9-NEXT:    ; implicit-def: $vgpr3314999; GFX9-NEXT:    ; kill: killed $vgpr3315000; GFX9-NEXT:    ; implicit-def: $vgpr3315001; GFX9-NEXT:    ; kill: killed $vgpr3315002; GFX9-NEXT:    ; implicit-def: $vgpr3315003; GFX9-NEXT:    ; kill: killed $vgpr3315004; GFX9-NEXT:    ; implicit-def: $vgpr3315005; GFX9-NEXT:    ; kill: killed $vgpr3315006; GFX9-NEXT:    ; implicit-def: $vgpr3315007; GFX9-NEXT:    ; kill: killed $vgpr3315008; GFX9-NEXT:    ; implicit-def: $vgpr3315009; GFX9-NEXT:    ; kill: killed $vgpr3315010; GFX9-NEXT:    ; implicit-def: $vgpr3315011; GFX9-NEXT:    ; kill: killed $vgpr3315012; GFX9-NEXT:    ; implicit-def: $vgpr3315013; GFX9-NEXT:    ; implicit-def: $vgpr6215014; GFX9-NEXT:    ; kill: killed $vgpr3315015; GFX9-NEXT:    ; implicit-def: $vgpr3515016; GFX9-NEXT:    ; implicit-def: $vgpr3315017; GFX9-NEXT:    s_waitcnt vmcnt(0)15018; GFX9-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15019; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15020; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload15021; GFX9-NEXT:    s_waitcnt vmcnt(0)15022; GFX9-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15023; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15024; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload15025; GFX9-NEXT:    ; implicit-def: $vgpr5815026; GFX9-NEXT:    s_waitcnt vmcnt(0)15027; GFX9-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15028; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15029; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload15030; GFX9-NEXT:    ; implicit-def: $vgpr5915031; GFX9-NEXT:    s_waitcnt vmcnt(0)15032; GFX9-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15033; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15034; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload15035; GFX9-NEXT:    ; implicit-def: $vgpr4715036; GFX9-NEXT:    s_waitcnt vmcnt(0)15037; GFX9-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15038; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15039; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload15040; GFX9-NEXT:    ; implicit-def: $vgpr4515041; GFX9-NEXT:    s_waitcnt vmcnt(0)15042; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15043; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15044; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload15045; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload15046; GFX9-NEXT:    ; implicit-def: $vgpr4215047; GFX9-NEXT:    s_waitcnt vmcnt(0)15048; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15049; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload15050; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload15051; GFX9-NEXT:    s_waitcnt vmcnt(0)15052; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15053; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15054; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload15055; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload15056; GFX9-NEXT:    s_waitcnt vmcnt(0)15057; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15058; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload15059; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload15060; GFX9-NEXT:    s_waitcnt vmcnt(0)15061; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15062; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15063; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload15064; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload15065; GFX9-NEXT:    s_waitcnt vmcnt(0)15066; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15067; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload15068; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload15069; GFX9-NEXT:    s_waitcnt vmcnt(0)15070; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15071; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15072; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload15073; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload15074; GFX9-NEXT:    s_waitcnt vmcnt(0)15075; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15076; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload15077; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload15078; GFX9-NEXT:    s_waitcnt vmcnt(0)15079; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15080; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15081; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload15082; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload15083; GFX9-NEXT:    s_waitcnt vmcnt(0)15084; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15085; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload15086; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload15087; GFX9-NEXT:    s_waitcnt vmcnt(0)15088; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15089; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15090; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload15091; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload15092; GFX9-NEXT:    s_waitcnt vmcnt(0)15093; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15094; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload15095; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload15096; GFX9-NEXT:    s_waitcnt vmcnt(0)15097; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15098; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15099; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload15100; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload15101; GFX9-NEXT:    s_waitcnt vmcnt(0)15102; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15103; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload15104; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload15105; GFX9-NEXT:    s_waitcnt vmcnt(0)15106; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15107; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15108; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload15109; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload15110; GFX9-NEXT:    s_waitcnt vmcnt(0)15111; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15112; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload15113; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload15114; GFX9-NEXT:    s_waitcnt vmcnt(0)15115; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15116; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15117; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload15118; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload15119; GFX9-NEXT:    s_waitcnt vmcnt(0)15120; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15121; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload15122; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload15123; GFX9-NEXT:    s_waitcnt vmcnt(0)15124; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15125; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15126; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload15127; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload15128; GFX9-NEXT:    s_waitcnt vmcnt(0)15129; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15130; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload15131; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload15132; GFX9-NEXT:    s_waitcnt vmcnt(0)15133; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15134; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15135; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload15136; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload15137; GFX9-NEXT:    s_waitcnt vmcnt(0)15138; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15139; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload15140; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload15141; GFX9-NEXT:    s_waitcnt vmcnt(0)15142; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15143; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15144; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload15145; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload15146; GFX9-NEXT:    s_waitcnt vmcnt(0)15147; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15148; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload15149; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload15150; GFX9-NEXT:    s_waitcnt vmcnt(0)15151; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15152; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15153; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload15154; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload15155; GFX9-NEXT:    s_waitcnt vmcnt(0)15156; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15157; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload15158; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload15159; GFX9-NEXT:    s_waitcnt vmcnt(0)15160; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15161; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15162; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload15163; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload15164; GFX9-NEXT:    s_waitcnt vmcnt(0)15165; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15166; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload15167; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload15168; GFX9-NEXT:    s_waitcnt vmcnt(0)15169; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15170; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15171; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload15172; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload15173; GFX9-NEXT:    s_waitcnt vmcnt(0)15174; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15175; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload15176; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload15177; GFX9-NEXT:    s_waitcnt vmcnt(0)15178; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15179; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15180; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload15181; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload15182; GFX9-NEXT:    s_waitcnt vmcnt(0)15183; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15184; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload15185; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload15186; GFX9-NEXT:    s_waitcnt vmcnt(0)15187; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15188; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15189; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload15190; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload15191; GFX9-NEXT:    s_waitcnt vmcnt(0)15192; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15193; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload15194; GFX9-NEXT:    s_waitcnt vmcnt(0)15195; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD15196; GFX9-NEXT:    ; implicit-def: $vgpr6315197; GFX9-NEXT:    ; kill: killed $vgpr6315198; GFX9-NEXT:    ; implicit-def: $vgpr6315199; GFX9-NEXT:    ; kill: killed $vgpr6315200; GFX9-NEXT:    ; implicit-def: $vgpr6315201; GFX9-NEXT:    ; kill: killed $vgpr6315202; GFX9-NEXT:    ; implicit-def: $vgpr6315203; GFX9-NEXT:    ; kill: killed $vgpr6315204; GFX9-NEXT:    ; implicit-def: $vgpr6315205; GFX9-NEXT:    ; kill: killed $vgpr6315206; GFX9-NEXT:    ; implicit-def: $vgpr6315207; GFX9-NEXT:    ; kill: killed $vgpr6315208; GFX9-NEXT:    ; implicit-def: $vgpr6315209; GFX9-NEXT:    ; kill: killed $vgpr6315210; GFX9-NEXT:    ; implicit-def: $vgpr6315211; GFX9-NEXT:    ; kill: killed $vgpr6315212; GFX9-NEXT:    ; implicit-def: $vgpr6315213; GFX9-NEXT:    ; kill: killed $vgpr6315214; GFX9-NEXT:    ; implicit-def: $vgpr6315215; GFX9-NEXT:    ; kill: killed $vgpr6315216; GFX9-NEXT:    ; implicit-def: $vgpr6315217; GFX9-NEXT:    ; kill: killed $vgpr6315218; GFX9-NEXT:    ; implicit-def: $vgpr6315219; GFX9-NEXT:    ; kill: killed $vgpr6315220; GFX9-NEXT:    ; implicit-def: $vgpr6315221; GFX9-NEXT:    ; kill: killed $vgpr6315222; GFX9-NEXT:    ; implicit-def: $vgpr6315223; GFX9-NEXT:    ; kill: killed $vgpr6315224; GFX9-NEXT:    ; implicit-def: $vgpr6315225; GFX9-NEXT:    ; kill: killed $vgpr6315226; GFX9-NEXT:    ; implicit-def: $vgpr6315227; GFX9-NEXT:    ; kill: killed $vgpr6315228; GFX9-NEXT:    ; implicit-def: $vgpr6315229; GFX9-NEXT:    ; kill: killed $vgpr6315230; GFX9-NEXT:    ; implicit-def: $vgpr6315231; GFX9-NEXT:    ; kill: killed $vgpr6315232; GFX9-NEXT:    ; implicit-def: $vgpr6315233; GFX9-NEXT:    ; kill: killed $vgpr6315234; GFX9-NEXT:    ; implicit-def: $vgpr6315235; GFX9-NEXT:    ; kill: killed $vgpr6315236; GFX9-NEXT:    ; implicit-def: $vgpr6315237; GFX9-NEXT:    ; kill: killed $vgpr6315238; GFX9-NEXT:    ; implicit-def: $vgpr6315239; GFX9-NEXT:    ; kill: killed $vgpr6315240; GFX9-NEXT:    ; implicit-def: $vgpr6315241; GFX9-NEXT:    ; kill: killed $vgpr6315242; GFX9-NEXT:    ; implicit-def: $vgpr6315243; GFX9-NEXT:    ; kill: killed $vgpr6315244; GFX9-NEXT:    ; implicit-def: $vgpr6315245; GFX9-NEXT:    ; kill: killed $vgpr6315246; GFX9-NEXT:    ; implicit-def: $vgpr6315247; GFX9-NEXT:    ; kill: killed $vgpr6315248; GFX9-NEXT:    ; implicit-def: $vgpr6315249; GFX9-NEXT:    ; kill: killed $vgpr6315250; GFX9-NEXT:    ; implicit-def: $vgpr6315251; GFX9-NEXT:    ; kill: killed $vgpr6315252; GFX9-NEXT:    ; implicit-def: $vgpr6315253; GFX9-NEXT:    ; kill: killed $vgpr6315254; GFX9-NEXT:    ; implicit-def: $vgpr6315255; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD15256; GFX9-NEXT:    ; implicit-def: $vgpr3215257; GFX9-NEXT:    ; kill: killed $vgpr6315258; GFX9-NEXT:    ; implicit-def: $vgpr6315259; GFX9-NEXT:    ; kill: killed $vgpr3215260; GFX9-NEXT:    ; implicit-def: $vgpr3215261; GFX9-NEXT:    ; kill: killed $vgpr6315262; GFX9-NEXT:    ; implicit-def: $vgpr6315263; GFX9-NEXT:    ; kill: killed $vgpr3215264; GFX9-NEXT:    ; implicit-def: $vgpr3215265; GFX9-NEXT:    ; kill: killed $vgpr6315266; GFX9-NEXT:    ; implicit-def: $vgpr6315267; GFX9-NEXT:    ; kill: killed $vgpr3215268; GFX9-NEXT:    ; implicit-def: $vgpr3215269; GFX9-NEXT:    ; kill: killed $vgpr6315270; GFX9-NEXT:    ; implicit-def: $vgpr6315271; GFX9-NEXT:    ; kill: killed $vgpr3215272; GFX9-NEXT:    ; implicit-def: $vgpr3215273; GFX9-NEXT:    ; kill: killed $vgpr6315274; GFX9-NEXT:    ; implicit-def: $vgpr6315275; GFX9-NEXT:    ; kill: killed $vgpr3215276; GFX9-NEXT:    ; implicit-def: $vgpr3215277; GFX9-NEXT:    ; kill: killed $vgpr6315278; GFX9-NEXT:    ; implicit-def: $vgpr6315279; GFX9-NEXT:    ; kill: killed $vgpr3215280; GFX9-NEXT:    ; implicit-def: $vgpr3215281; GFX9-NEXT:    ; kill: killed $vgpr6315282; GFX9-NEXT:    ; implicit-def: $vgpr6315283; GFX9-NEXT:    ; kill: killed $vgpr3215284; GFX9-NEXT:    ; implicit-def: $vgpr3215285; GFX9-NEXT:    ; kill: killed $vgpr6315286; GFX9-NEXT:    ; implicit-def: $vgpr6315287; GFX9-NEXT:    ; kill: killed $vgpr3215288; GFX9-NEXT:    ; implicit-def: $vgpr3215289; GFX9-NEXT:    ; kill: killed $vgpr6315290; GFX9-NEXT:    ; implicit-def: $vgpr6315291; GFX9-NEXT:    ; kill: killed $vgpr3215292; GFX9-NEXT:    ; implicit-def: $vgpr3215293; GFX9-NEXT:    ; kill: killed $vgpr6315294; GFX9-NEXT:    ; implicit-def: $vgpr6315295; GFX9-NEXT:    ; kill: killed $vgpr3215296; GFX9-NEXT:    ; implicit-def: $vgpr3215297; GFX9-NEXT:    ; kill: killed $vgpr6315298; GFX9-NEXT:    ; implicit-def: $vgpr6315299; GFX9-NEXT:    ; kill: killed $vgpr3215300; GFX9-NEXT:    ; implicit-def: $vgpr3215301; GFX9-NEXT:    ; kill: killed $vgpr6315302; GFX9-NEXT:    ; implicit-def: $vgpr6315303; GFX9-NEXT:    ; kill: killed $vgpr3215304; GFX9-NEXT:    ; implicit-def: $vgpr3215305; GFX9-NEXT:    ; kill: killed $vgpr6315306; GFX9-NEXT:    ; implicit-def: $vgpr6315307; GFX9-NEXT:    ; kill: killed $vgpr3215308; GFX9-NEXT:    ; implicit-def: $vgpr3215309; GFX9-NEXT:    ; kill: killed $vgpr6315310; GFX9-NEXT:    ; implicit-def: $vgpr6315311; GFX9-NEXT:    ; kill: killed $vgpr3215312; GFX9-NEXT:    ; implicit-def: $vgpr3215313; GFX9-NEXT:    ; kill: killed $vgpr6315314; GFX9-NEXT:    ; implicit-def: $vgpr6315315; GFX9-NEXT:    ; kill: killed $vgpr3215316; GFX9-NEXT:    ; implicit-def: $vgpr3215317; GFX9-NEXT:    ; kill: killed $vgpr6315318; GFX9-NEXT:    ; implicit-def: $vgpr6315319; GFX9-NEXT:    ; kill: killed $vgpr3215320; GFX9-NEXT:    ; implicit-def: $vgpr3215321; GFX9-NEXT:    ; kill: killed $vgpr6315322; GFX9-NEXT:    ; implicit-def: $vgpr6315323; GFX9-NEXT:  .LBB14_2: ; %Flow15324; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]15325; GFX9-NEXT:    s_cbranch_execz .LBB14_415326; GFX9-NEXT:  ; %bb.3: ; %cmp.true15327; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload15328; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload15329; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload15330; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload15331; GFX9-NEXT:    v_add_u16_e32 v9, 3, v6115332; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload15333; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload15334; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload15335; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload15336; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload15337; GFX9-NEXT:    s_movk_i32 s6, 0x30015338; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload15339; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload15340; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload15341; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload15342; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload15343; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload15344; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload15345; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload15346; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload15347; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload15348; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload15349; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload15350; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload15351; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload15352; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload15353; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload15354; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload15355; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload15356; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload15357; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload15358; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload15359; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload15360; GFX9-NEXT:    s_waitcnt vmcnt(30)15361; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015362; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload15363; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15364; GFX9-NEXT:    s_waitcnt vmcnt(28)15365; GFX9-NEXT:    v_add_u16_e32 v8, 3, v815366; GFX9-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015367; GFX9-NEXT:    s_waitcnt vmcnt(27)15368; GFX9-NEXT:    v_add_u16_e32 v0, 3, v015369; GFX9-NEXT:    s_waitcnt vmcnt(26)15370; GFX9-NEXT:    v_add_u16_e32 v2, 3, v215371; GFX9-NEXT:    s_waitcnt vmcnt(25)15372; GFX9-NEXT:    v_add_u16_e32 v3, 3, v315373; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015374; GFX9-NEXT:    s_waitcnt vmcnt(23)15375; GFX9-NEXT:    v_add_u16_e32 v1, 3, v115376; GFX9-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015377; GFX9-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015378; GFX9-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015379; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v015380; GFX9-NEXT:    v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15381; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v215382; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15383; GFX9-NEXT:    v_or_b32_e32 v0, v0, v115384; GFX9-NEXT:    v_or_b32_e32 v1, v2, v315385; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload15386; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload15387; GFX9-NEXT:    s_waitcnt vmcnt(20)15388; GFX9-NEXT:    v_add_u16_e32 v4, 3, v415389; GFX9-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015390; GFX9-NEXT:    v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15391; GFX9-NEXT:    s_waitcnt vmcnt(17)15392; GFX9-NEXT:    v_add_u16_e32 v5, 3, v515393; GFX9-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015394; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15395; GFX9-NEXT:    s_waitcnt vmcnt(14)15396; GFX9-NEXT:    v_add_u16_e32 v6, 3, v615397; GFX9-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015398; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15399; GFX9-NEXT:    s_waitcnt vmcnt(11)15400; GFX9-NEXT:    v_add_u16_e32 v7, 3, v715401; GFX9-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015402; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15403; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15404; GFX9-NEXT:    s_waitcnt vmcnt(1)15405; GFX9-NEXT:    v_add_u16_e32 v2, 3, v215406; GFX9-NEXT:    s_waitcnt vmcnt(0)15407; GFX9-NEXT:    v_add_u16_e32 v3, 3, v315408; GFX9-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015409; GFX9-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015410; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v215411; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15412; GFX9-NEXT:    v_or_b32_e32 v2, v2, v315413; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload15414; GFX9-NEXT:    s_waitcnt vmcnt(0)15415; GFX9-NEXT:    v_add_u16_e32 v3, 3, v315416; GFX9-NEXT:    v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015417; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v315418; GFX9-NEXT:    v_or_b32_e32 v3, v3, v415419; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload15420; GFX9-NEXT:    s_waitcnt vmcnt(0)15421; GFX9-NEXT:    v_add_u16_e32 v4, 3, v415422; GFX9-NEXT:    v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015423; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v415424; GFX9-NEXT:    v_or_b32_e32 v4, v4, v515425; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload15426; GFX9-NEXT:    s_waitcnt vmcnt(0)15427; GFX9-NEXT:    v_add_u16_e32 v5, 3, v515428; GFX9-NEXT:    v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015429; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v515430; GFX9-NEXT:    v_or_b32_e32 v5, v5, v615431; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload15432; GFX9-NEXT:    s_waitcnt vmcnt(0)15433; GFX9-NEXT:    v_add_u16_e32 v6, 3, v615434; GFX9-NEXT:    v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015435; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v615436; GFX9-NEXT:    v_or_b32_e32 v6, v6, v715437; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload15438; GFX9-NEXT:    s_waitcnt vmcnt(0)15439; GFX9-NEXT:    v_add_u16_e32 v7, 3, v715440; GFX9-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015441; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v715442; GFX9-NEXT:    v_or_b32_e32 v7, v7, v815443; GFX9-NEXT:    v_add_u16_e32 v8, 3, v6215444; GFX9-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015445; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v815446; GFX9-NEXT:    v_or_b32_e32 v8, v8, v915447; GFX9-NEXT:    v_add_u16_e32 v9, 3, v3215448; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015449; GFX9-NEXT:    v_add_u16_e32 v10, 3, v6015450; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015451; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload15452; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v915453; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15454; GFX9-NEXT:    v_or_b32_e32 v9, v9, v1015455; GFX9-NEXT:    v_add_u16_e32 v10, 3, v5815456; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload15457; GFX9-NEXT:    s_waitcnt vmcnt(1)15458; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015459; GFX9-NEXT:    v_add_u16_e32 v11, 3, v5615460; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015461; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload15462; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v1015463; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15464; GFX9-NEXT:    v_or_b32_e32 v10, v10, v1115465; GFX9-NEXT:    v_add_u16_e32 v11, 3, v5915466; GFX9-NEXT:    s_waitcnt vmcnt(0)15467; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015468; GFX9-NEXT:    v_add_u16_e32 v12, 3, v5715469; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015470; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload15471; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v1115472; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15473; GFX9-NEXT:    v_or_b32_e32 v11, v11, v1215474; GFX9-NEXT:    v_add_u16_e32 v12, 3, v4715475; GFX9-NEXT:    s_waitcnt vmcnt(0)15476; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015477; GFX9-NEXT:    v_add_u16_e32 v13, 3, v4615478; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015479; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload15480; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v1215481; GFX9-NEXT:    v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15482; GFX9-NEXT:    v_or_b32_e32 v12, v12, v1315483; GFX9-NEXT:    v_add_u16_e32 v13, 3, v4515484; GFX9-NEXT:    s_waitcnt vmcnt(0)15485; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015486; GFX9-NEXT:    v_add_u16_e32 v14, 3, v4415487; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015488; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload15489; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v1315490; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15491; GFX9-NEXT:    v_or_b32_e32 v13, v13, v1415492; GFX9-NEXT:    v_add_u16_e32 v14, 3, v4215493; GFX9-NEXT:    s_waitcnt vmcnt(0)15494; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015495; GFX9-NEXT:    v_add_u16_e32 v15, 3, v4015496; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015497; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v1415498; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15499; GFX9-NEXT:    v_or_b32_e32 v14, v14, v1515500; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload15501; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload15502; GFX9-NEXT:    s_waitcnt vmcnt(1)15503; GFX9-NEXT:    v_add_u16_e32 v15, 3, v1515504; GFX9-NEXT:    s_waitcnt vmcnt(0)15505; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015506; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload15507; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v1515508; GFX9-NEXT:    s_waitcnt vmcnt(0)15509; GFX9-NEXT:    v_add_u16_e32 v16, 3, v1615510; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015511; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15512; GFX9-NEXT:    v_or_b32_e32 v15, v15, v1615513; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload15514; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload15515; GFX9-NEXT:    s_waitcnt vmcnt(1)15516; GFX9-NEXT:    v_add_u16_e32 v16, 3, v1615517; GFX9-NEXT:    s_waitcnt vmcnt(0)15518; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015519; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload15520; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v1615521; GFX9-NEXT:    s_waitcnt vmcnt(0)15522; GFX9-NEXT:    v_add_u16_e32 v17, 3, v1715523; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015524; GFX9-NEXT:    v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15525; GFX9-NEXT:    v_or_b32_e32 v16, v16, v1715526; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload15527; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload15528; GFX9-NEXT:    s_waitcnt vmcnt(1)15529; GFX9-NEXT:    v_add_u16_e32 v17, 3, v1715530; GFX9-NEXT:    s_waitcnt vmcnt(0)15531; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015532; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload15533; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v1715534; GFX9-NEXT:    s_waitcnt vmcnt(0)15535; GFX9-NEXT:    v_add_u16_e32 v18, 3, v1815536; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015537; GFX9-NEXT:    v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15538; GFX9-NEXT:    v_or_b32_e32 v17, v17, v1815539; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload15540; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload15541; GFX9-NEXT:    s_waitcnt vmcnt(1)15542; GFX9-NEXT:    v_add_u16_e32 v18, 3, v1815543; GFX9-NEXT:    s_waitcnt vmcnt(0)15544; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015545; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload15546; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v1815547; GFX9-NEXT:    s_waitcnt vmcnt(0)15548; GFX9-NEXT:    v_add_u16_e32 v19, 3, v1915549; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015550; GFX9-NEXT:    v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15551; GFX9-NEXT:    v_or_b32_e32 v18, v18, v1915552; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload15553; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload15554; GFX9-NEXT:    s_waitcnt vmcnt(1)15555; GFX9-NEXT:    v_add_u16_e32 v19, 3, v1915556; GFX9-NEXT:    s_waitcnt vmcnt(0)15557; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015558; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload15559; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v1915560; GFX9-NEXT:    s_waitcnt vmcnt(0)15561; GFX9-NEXT:    v_add_u16_e32 v20, 3, v2015562; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015563; GFX9-NEXT:    v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15564; GFX9-NEXT:    v_or_b32_e32 v19, v19, v2015565; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload15566; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload15567; GFX9-NEXT:    s_waitcnt vmcnt(1)15568; GFX9-NEXT:    v_add_u16_e32 v20, 3, v2015569; GFX9-NEXT:    s_waitcnt vmcnt(0)15570; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015571; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload15572; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v2015573; GFX9-NEXT:    s_waitcnt vmcnt(0)15574; GFX9-NEXT:    v_add_u16_e32 v21, 3, v2115575; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015576; GFX9-NEXT:    v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15577; GFX9-NEXT:    v_or_b32_e32 v20, v20, v2115578; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload15579; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload15580; GFX9-NEXT:    s_waitcnt vmcnt(1)15581; GFX9-NEXT:    v_add_u16_e32 v21, 3, v2115582; GFX9-NEXT:    s_waitcnt vmcnt(0)15583; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015584; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload15585; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v2115586; GFX9-NEXT:    s_waitcnt vmcnt(0)15587; GFX9-NEXT:    v_add_u16_e32 v22, 3, v2215588; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015589; GFX9-NEXT:    v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15590; GFX9-NEXT:    v_or_b32_e32 v21, v21, v2215591; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload15592; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload15593; GFX9-NEXT:    s_waitcnt vmcnt(1)15594; GFX9-NEXT:    v_add_u16_e32 v22, 3, v2215595; GFX9-NEXT:    s_waitcnt vmcnt(0)15596; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015597; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload15598; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v2215599; GFX9-NEXT:    s_waitcnt vmcnt(0)15600; GFX9-NEXT:    v_add_u16_e32 v23, 3, v2315601; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015602; GFX9-NEXT:    v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15603; GFX9-NEXT:    v_or_b32_e32 v22, v22, v2315604; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload15605; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload15606; GFX9-NEXT:    s_waitcnt vmcnt(1)15607; GFX9-NEXT:    v_add_u16_e32 v23, 3, v2315608; GFX9-NEXT:    s_waitcnt vmcnt(0)15609; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015610; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload15611; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v2315612; GFX9-NEXT:    s_waitcnt vmcnt(0)15613; GFX9-NEXT:    v_add_u16_e32 v24, 3, v2415614; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015615; GFX9-NEXT:    v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15616; GFX9-NEXT:    v_or_b32_e32 v23, v23, v2415617; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload15618; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload15619; GFX9-NEXT:    s_waitcnt vmcnt(1)15620; GFX9-NEXT:    v_add_u16_e32 v24, 3, v2415621; GFX9-NEXT:    s_waitcnt vmcnt(0)15622; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015623; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload15624; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v2415625; GFX9-NEXT:    s_waitcnt vmcnt(0)15626; GFX9-NEXT:    v_add_u16_e32 v25, 3, v2515627; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015628; GFX9-NEXT:    v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15629; GFX9-NEXT:    v_or_b32_e32 v24, v24, v2515630; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload15631; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload15632; GFX9-NEXT:    s_waitcnt vmcnt(1)15633; GFX9-NEXT:    v_add_u16_e32 v25, 3, v2515634; GFX9-NEXT:    s_waitcnt vmcnt(0)15635; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015636; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload15637; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v2515638; GFX9-NEXT:    s_waitcnt vmcnt(0)15639; GFX9-NEXT:    v_add_u16_e32 v26, 3, v2615640; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015641; GFX9-NEXT:    v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15642; GFX9-NEXT:    v_or_b32_e32 v25, v25, v2615643; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload15644; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload15645; GFX9-NEXT:    s_waitcnt vmcnt(1)15646; GFX9-NEXT:    v_add_u16_e32 v26, 3, v2615647; GFX9-NEXT:    s_waitcnt vmcnt(0)15648; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015649; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload15650; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v2615651; GFX9-NEXT:    s_waitcnt vmcnt(0)15652; GFX9-NEXT:    v_add_u16_e32 v27, 3, v2715653; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015654; GFX9-NEXT:    v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15655; GFX9-NEXT:    v_or_b32_e32 v26, v26, v2715656; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload15657; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload15658; GFX9-NEXT:    s_waitcnt vmcnt(1)15659; GFX9-NEXT:    v_add_u16_e32 v27, 3, v2715660; GFX9-NEXT:    s_waitcnt vmcnt(0)15661; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015662; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload15663; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v2715664; GFX9-NEXT:    s_waitcnt vmcnt(0)15665; GFX9-NEXT:    v_add_u16_e32 v28, 3, v2815666; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015667; GFX9-NEXT:    v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15668; GFX9-NEXT:    v_or_b32_e32 v27, v27, v2815669; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload15670; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload15671; GFX9-NEXT:    s_waitcnt vmcnt(1)15672; GFX9-NEXT:    v_add_u16_e32 v28, 3, v2815673; GFX9-NEXT:    s_waitcnt vmcnt(0)15674; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015675; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload15676; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v2815677; GFX9-NEXT:    s_waitcnt vmcnt(0)15678; GFX9-NEXT:    v_add_u16_e32 v29, 3, v2915679; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015680; GFX9-NEXT:    v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15681; GFX9-NEXT:    v_or_b32_e32 v28, v28, v2915682; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload15683; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload15684; GFX9-NEXT:    s_waitcnt vmcnt(1)15685; GFX9-NEXT:    v_add_u16_e32 v29, 3, v2915686; GFX9-NEXT:    s_waitcnt vmcnt(0)15687; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015688; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload15689; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v2915690; GFX9-NEXT:    s_waitcnt vmcnt(0)15691; GFX9-NEXT:    v_add_u16_e32 v30, 3, v3015692; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015693; GFX9-NEXT:    v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15694; GFX9-NEXT:    v_or_b32_e32 v29, v29, v3015695; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload15696; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload15697; GFX9-NEXT:    s_waitcnt vmcnt(1)15698; GFX9-NEXT:    v_add_u16_e32 v30, 3, v3015699; GFX9-NEXT:    s_waitcnt vmcnt(0)15700; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015701; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload15702; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v3015703; GFX9-NEXT:    s_waitcnt vmcnt(0)15704; GFX9-NEXT:    v_add_u16_e32 v31, 3, v3115705; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015706; GFX9-NEXT:    v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15707; GFX9-NEXT:    v_or_b32_e32 v30, v30, v3115708; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload15709; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload15710; GFX9-NEXT:    s_waitcnt vmcnt(1)15711; GFX9-NEXT:    v_add_u16_e32 v31, 3, v3115712; GFX9-NEXT:    s_waitcnt vmcnt(0)15713; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015714; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload15715; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v3115716; GFX9-NEXT:    s_waitcnt vmcnt(0)15717; GFX9-NEXT:    v_add_u16_e32 v32, 3, v3215718; GFX9-NEXT:    v_or_b32_sdwa v32, v63, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_015719; GFX9-NEXT:    v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD15720; GFX9-NEXT:    v_or_b32_e32 v31, v31, v3215721; GFX9-NEXT:  .LBB14_4: ; %end15722; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]15723; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload15724; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload15725; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload15726; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload15727; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload15728; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload15729; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload15730; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload15731; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload15732; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload15733; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload15734; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload15735; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload15736; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload15737; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload15738; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload15739; GFX9-NEXT:    s_waitcnt vmcnt(0)15740; GFX9-NEXT:    s_setpc_b64 s[30:31]15741;15742; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v32i32:15743; GFX11-TRUE16:       ; %bb.0:15744; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)15745; GFX11-TRUE16-NEXT:    s_clause 0x1f15746; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:38415747; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:38015748; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v31, off, s32 offset:37615749; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v32, off, s32 offset:37215750; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v51, off, s32 offset:36815751; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:36415752; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:36015753; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v33, off, s32 offset:35615754; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v50, off, s32 offset:35215755; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:34815756; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:34415757; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v34, off, s32 offset:34015758; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v52, off, s32 offset:33615759; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:33215760; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v54, off, s32 offset:32815761; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v35, off, s32 offset:32415762; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:32015763; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:31615764; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v53, off, s32 offset:31215765; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v36, off, s32 offset:30815766; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:30415767; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:30015768; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v55, off, s32 offset:29615769; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v37, off, s32 offset:29215770; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:28815771; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:28415772; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v66, off, s32 offset:28015773; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v38, off, s32 offset:27615774; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:27215775; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:26815776; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v67, off, s32 offset:26415777; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v39, off, s32 offset:26015778; GFX11-TRUE16-NEXT:    s_clause 0x1f15779; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:25615780; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:25215781; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v71, off, s32 offset:24815782; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v48, off, s32 offset:24415783; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:24015784; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:23615785; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v70, off, s32 offset:23215786; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v49, off, s32 offset:22815787; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:22415788; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:22015789; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v80, off, s32 offset:21615790; GFX11-TRUE16-NEXT:    scratch_load_b32 v113, off, s32 offset:38815791; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v83, off, s3215792; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:815793; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v84, off, s32 offset:1615794; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:2415795; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v85, off, s32 offset:3215796; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:4015797; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v96, off, s32 offset:4815798; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:5615799; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v97, off, s32 offset:6415800; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:7215801; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v100, off, s32 offset:8015802; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:8815803; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v101, off, s32 offset:9615804; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:10415805; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v102, off, s32 offset:11215806; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:12015807; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v160, off, s32 offset:12815808; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:13615809; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v161, off, s32 offset:14415810; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:15215811; GFX11-TRUE16-NEXT:    s_clause 0x1f15812; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v162, off, s32 offset:16015813; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:16815814; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v163, off, s32 offset:17615815; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:18415816; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v164, off, s32 offset:19215817; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:20015818; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v165, off, s32 offset:20815819; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:21215820; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:20415821; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v64, off, s32 offset:19615822; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:18815823; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v65, off, s32 offset:18015824; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v68, off, s32 offset:17215825; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:16415826; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v69, off, s32 offset:15615827; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:14815828; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:14015829; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v81, off, s32 offset:13215830; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:12415831; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v82, off, s32 offset:11615832; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:10815833; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:10015834; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v86, off, s32 offset:9215835; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:8415836; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v87, off, s32 offset:7615837; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v98, off, s32 offset:6815838; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:6015839; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v99, off, s32 offset:5215840; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:4415841; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:3615842; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v103, off, s32 offset:2815843; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:2015844; GFX11-TRUE16-NEXT:    s_clause 0x115845; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v112, off, s32 offset:1215846; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v115, off, s32 offset:415847; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v116.l, v30.l15848; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v116.h, v28.l15849; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v119.h, v26.l15850; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, v24.l15851; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.h, v22.l15852; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v129.l, v20.l15853; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v129.h, v18.l15854; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v132.h, v16.l15855; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.l, v14.l15856; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.h, v12.l15857; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v134.l, v10.l15858; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v145.l, v8.l15859; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v145.h, v6.l15860; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v146.l, v4.l15861; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v149.l, v2.l15862; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v149.h, v0.l15863; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v150.l, 8, v1.l15864; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v150.h, 8, v3.l15865; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v151.l, 8, v5.l15866; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v151.h, 8, v7.l15867; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v146.h, 8, v9.l15868; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v147.l, 8, v11.l15869; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v147.h, 8, v13.l15870; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v148.l, 8, v15.l15871; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v148.h, 8, v17.l15872; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v134.h, 8, v19.l15873; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v135.l, 8, v21.l15874; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v135.h, 8, v23.l15875; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v144.l, 8, v25.l15876; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v144.h, 8, v27.l15877; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v130.l, 8, v29.l15878; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)15879; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v51.h, 8, v51.h15880; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v51.l, 8, v51.l15881; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(56)15882; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v50.h, 8, v50.h15883; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v54.h, 8, v54.h15884; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v54.l, 8, v54.l15885; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(26)15886; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v53.h, 8, v53.h15887; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v67.l, 8, v67.l15888; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v66.h, 8, v66.h15889; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v66.l, 8, v66.l15890; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v71.h, 8, v71.h15891; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v71.l, 8, v71.l15892; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(18)15893; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v70.h, 8, v70.h15894; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v11315895; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)15896; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v130.h, 8, v83.h15897; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v131.l, 8, v84.l15898; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v131.h, 8, v84.h15899; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v132.l, 8, v85.l15900; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v117.l, 8, v85.h15901; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v117.h, 8, v96.l15902; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v118.l, 8, v96.h15903; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v118.h, 8, v97.l15904; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v119.l, 8, v97.h15905; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(8)15906; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v113.l, 8, v98.l15907; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(5)15908; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v113.h, 8, v100.h15909; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v114.l, 8, v101.l15910; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v114.h, 8, v101.h15911; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)15912; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v115.l, 8, v102.l15913; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v100.h, 8, v102.h15914; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v101.l, 8, v160.l15915; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v101.h, 8, v160.h15916; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v102.l, 8, v161.l15917; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v102.h, 8, v161.h15918; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v96.l, 8, v162.l15919; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v96.h, 8, v162.h15920; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v97.l, 8, v163.l15921; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v97.h, 8, v163.h15922; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v98.l, 8, v164.l15923; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v83.h, 8, v164.h15924; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v84.l, 8, v165.l15925; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v84.h, 8, v165.h15926; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v85.l, 8, v80.h15927; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v85.h, 8, v80.l15928; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v80.l, 8, v68.l15929; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v80.h, 8, v67.h15930; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v67.h, 8, v55.h15931; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v68.l, 8, v55.l15932; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v55.l, 8, v52.h15933; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v55.h, 8, v52.l15934; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v52.l, 8, v31.h15935; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v52.h, 8, v31.l15936; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr3115937; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo15938; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)15939; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s015940; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB14_315941; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow15942; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s015943; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB14_415944; GFX11-TRUE16-NEXT:  .LBB14_2: ; %end15945; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s015946; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]15947; GFX11-TRUE16-NEXT:  .LBB14_3: ; %cmp.false15948; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v149.h15949; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v149.l15950; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v146.l15951; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v145.h15952; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v145.l15953; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v134.l15954; GFX11-TRUE16-NEXT:    v_and_b16 v3.l, 0xff, v133.h15955; GFX11-TRUE16-NEXT:    v_and_b16 v3.h, 0xff, v133.l15956; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v132.h15957; GFX11-TRUE16-NEXT:    v_and_b16 v4.h, 0xff, v129.h15958; GFX11-TRUE16-NEXT:    v_and_b16 v5.l, 0xff, v129.l15959; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v128.h15960; GFX11-TRUE16-NEXT:    v_and_b16 v6.l, 0xff, v128.l15961; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v119.h15962; GFX11-TRUE16-NEXT:    v_and_b16 v7.l, 0xff, v116.h15963; GFX11-TRUE16-NEXT:    v_and_b16 v7.h, 0xff, v116.l15964; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v115.h15965; GFX11-TRUE16-NEXT:    v_and_b16 v8.h, 0xff, v112.h15966; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v112.l15967; GFX11-TRUE16-NEXT:    v_and_b16 v9.h, 0xff, v103.h15968; GFX11-TRUE16-NEXT:    v_and_b16 v10.l, 0xff, v103.l15969; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v100.l15970; GFX11-TRUE16-NEXT:    v_and_b16 v11.l, 0xff, v99.h15971; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v99.l15972; GFX11-TRUE16-NEXT:    v_and_b16 v12.l, 0xff, v98.h15973; GFX11-TRUE16-NEXT:    v_and_b16 v12.h, 0xff, v87.h15974; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v87.l15975; GFX11-TRUE16-NEXT:    v_and_b16 v13.h, 0xff, v86.h15976; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v86.l15977; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v83.l15978; GFX11-TRUE16-NEXT:    v_and_b16 v15.l, 0xff, v82.h15979; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v82.l15980; GFX11-TRUE16-NEXT:    v_and_b16 v16.l, 0xff, v81.h15981; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v81.l15982; GFX11-TRUE16-NEXT:    v_and_b16 v17.l, 0xff, v70.l15983; GFX11-TRUE16-NEXT:    v_and_b16 v17.h, 0xff, v69.h15984; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v69.l15985; GFX11-TRUE16-NEXT:    v_and_b16 v18.h, 0xff, v68.h15986; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v65.h15987; GFX11-TRUE16-NEXT:    v_and_b16 v19.h, 0xff, v65.l15988; GFX11-TRUE16-NEXT:    v_and_b16 v20.l, 0xff, v64.h15989; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v64.l15990; GFX11-TRUE16-NEXT:    v_and_b16 v21.l, 0xff, v53.l15991; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v50.l15992; GFX11-TRUE16-NEXT:    v_and_b16 v22.l, 0xff, v49.h15993; GFX11-TRUE16-NEXT:    v_and_b16 v22.h, 0xff, v49.l15994; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v48.h15995; GFX11-TRUE16-NEXT:    v_and_b16 v23.h, 0xff, v48.l15996; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v39.h15997; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v39.l15998; GFX11-TRUE16-NEXT:    v_and_b16 v25.l, 0xff, v38.h15999; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v38.l16000; GFX11-TRUE16-NEXT:    v_and_b16 v26.l, 0xff, v37.h16001; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v37.l16002; GFX11-TRUE16-NEXT:    v_and_b16 v27.l, 0xff, v36.h16003; GFX11-TRUE16-NEXT:    v_and_b16 v27.h, 0xff, v36.l16004; GFX11-TRUE16-NEXT:    v_and_b16 v28.l, 0xff, v35.h16005; GFX11-TRUE16-NEXT:    v_and_b16 v28.h, 0xff, v35.l16006; GFX11-TRUE16-NEXT:    v_and_b16 v29.l, 0xff, v34.h16007; GFX11-TRUE16-NEXT:    v_and_b16 v29.h, 0xff, v34.l16008; GFX11-TRUE16-NEXT:    v_and_b16 v30.l, 0xff, v33.h16009; GFX11-TRUE16-NEXT:    v_and_b16 v30.h, 0xff, v33.l16010; GFX11-TRUE16-NEXT:    v_and_b16 v31.l, 0xff, v32.h16011; GFX11-TRUE16-NEXT:    v_and_b16 v31.h, 0xff, v32.l16012; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v0.l, v150.l16013; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v0.h, v150.h16014; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.l, v151.l16015; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v1.h, v151.h16016; GFX11-TRUE16-NEXT:    v_or_b16 v2.l, v2.l, v146.h16017; GFX11-TRUE16-NEXT:    v_or_b16 v2.h, v2.h, v147.l16018; GFX11-TRUE16-NEXT:    v_or_b16 v3.l, v3.l, v147.h16019; GFX11-TRUE16-NEXT:    v_or_b16 v3.h, v3.h, v148.l16020; GFX11-TRUE16-NEXT:    v_or_b16 v4.l, v4.l, v148.h16021; GFX11-TRUE16-NEXT:    v_or_b16 v4.h, v4.h, v134.h16022; GFX11-TRUE16-NEXT:    v_or_b16 v5.l, v5.l, v135.l16023; GFX11-TRUE16-NEXT:    v_or_b16 v5.h, v5.h, v135.h16024; GFX11-TRUE16-NEXT:    v_or_b16 v6.l, v6.l, v144.l16025; GFX11-TRUE16-NEXT:    v_or_b16 v6.h, v6.h, v144.h16026; GFX11-TRUE16-NEXT:    v_or_b16 v7.l, v7.l, v130.l16027; GFX11-TRUE16-NEXT:    v_or_b16 v7.h, v7.h, v130.h16028; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.l, v131.l16029; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v8.h, v131.h16030; GFX11-TRUE16-NEXT:    v_or_b16 v9.l, v9.l, v132.l16031; GFX11-TRUE16-NEXT:    v_or_b16 v9.h, v9.h, v117.l16032; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v10.l, v117.h16033; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v10.h, v118.l16034; GFX11-TRUE16-NEXT:    v_or_b16 v11.l, v11.l, v118.h16035; GFX11-TRUE16-NEXT:    v_or_b16 v11.h, v11.h, v119.l16036; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v12.l, v113.l16037; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v12.h, v113.h16038; GFX11-TRUE16-NEXT:    v_or_b16 v13.l, v13.l, v114.l16039; GFX11-TRUE16-NEXT:    v_or_b16 v13.h, v13.h, v114.h16040; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v14.l, v115.l16041; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v14.h, v100.h16042; GFX11-TRUE16-NEXT:    v_or_b16 v15.l, v15.l, v101.l16043; GFX11-TRUE16-NEXT:    v_or_b16 v15.h, v15.h, v101.h16044; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v16.l, v102.l16045; GFX11-TRUE16-NEXT:    v_or_b16 v16.h, v16.h, v102.h16046; GFX11-TRUE16-NEXT:    v_or_b16 v17.l, v17.l, v96.l16047; GFX11-TRUE16-NEXT:    v_or_b16 v17.h, v17.h, v96.h16048; GFX11-TRUE16-NEXT:    v_or_b16 v18.l, v18.l, v97.l16049; GFX11-TRUE16-NEXT:    v_or_b16 v18.h, v18.h, v97.h16050; GFX11-TRUE16-NEXT:    v_or_b16 v19.l, v19.l, v98.l16051; GFX11-TRUE16-NEXT:    v_or_b16 v19.h, v19.h, v83.h16052; GFX11-TRUE16-NEXT:    v_or_b16 v20.l, v20.l, v84.l16053; GFX11-TRUE16-NEXT:    v_or_b16 v20.h, v20.h, v84.h16054; GFX11-TRUE16-NEXT:    v_or_b16 v21.l, v21.l, v85.l16055; GFX11-TRUE16-NEXT:    v_or_b16 v21.h, v21.h, v85.h16056; GFX11-TRUE16-NEXT:    v_or_b16 v22.l, v22.l, v70.h16057; GFX11-TRUE16-NEXT:    v_or_b16 v22.h, v22.h, v71.l16058; GFX11-TRUE16-NEXT:    v_or_b16 v23.l, v23.l, v71.h16059; GFX11-TRUE16-NEXT:    v_or_b16 v23.h, v23.h, v80.l16060; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v24.l, v80.h16061; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v24.h, v66.l16062; GFX11-TRUE16-NEXT:    v_or_b16 v25.l, v25.l, v66.h16063; GFX11-TRUE16-NEXT:    v_or_b16 v25.h, v25.h, v67.l16064; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v26.l, v67.h16065; GFX11-TRUE16-NEXT:    v_or_b16 v26.h, v26.h, v68.l16066; GFX11-TRUE16-NEXT:    v_or_b16 v27.l, v27.l, v53.h16067; GFX11-TRUE16-NEXT:    v_or_b16 v27.h, v27.h, v54.l16068; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v28.l, v54.h16069; GFX11-TRUE16-NEXT:    v_or_b16 v28.h, v28.h, v55.l16070; GFX11-TRUE16-NEXT:    v_or_b16 v29.l, v29.l, v55.h16071; GFX11-TRUE16-NEXT:    v_or_b16 v29.h, v29.h, v50.h16072; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v30.l, v51.l16073; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v30.h, v51.h16074; GFX11-TRUE16-NEXT:    v_or_b16 v31.l, v31.l, v52.l16075; GFX11-TRUE16-NEXT:    v_or_b16 v31.h, v31.h, v52.h16076; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_hi1616077; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo1616078; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo1616079; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_hi1616080; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo1616081; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo1616082; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_hi1616083; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo1616084; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_hi1616085; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_hi1616086; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo1616087; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_hi1616088; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo1616089; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_hi1616090; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_hi1616091; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo1616092; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_hi1616093; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_hi1616094; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo1616095; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_hi1616096; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo1616097; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo1616098; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_hi1616099; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo1616100; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_hi1616101; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_hi1616102; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo1616103; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_hi1616104; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo1616105; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo1616106; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_hi1616107; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo1616108; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_hi1616109; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo1616110; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo1616111; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_hi1616112; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo1616113; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_hi1616114; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_hi1616115; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo1616116; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_hi1616117; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo1616118; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo1616119; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo1616120; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_hi1616121; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo1616122; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_hi1616123; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo1616124; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_hi1616125; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo1616126; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_hi1616127; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo1616128; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_hi1616129; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo1616130; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_hi1616131; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo1616132; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_hi1616133; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo1616134; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_hi1616135; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo1616136; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_hi1616137; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo1616138; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_hi1616139; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo1616140; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo1616141; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_hi1616142; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo1616143; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_hi1616144; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_hi1616145; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo1616146; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_hi1616147; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo1616148; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_hi1616149; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_hi1616150; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo1616151; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_hi1616152; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo1616153; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_hi1616154; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo1616155; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_hi1616156; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo1616157; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_hi1616158; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo1616159; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo1616160; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_hi1616161; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo1616162; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_hi1616163; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo1616164; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo1616165; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_hi1616166; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo1616167; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_hi1616168; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo1616169; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_hi1616170; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo1616171; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_hi1616172; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo1616173; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_hi1616174; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo1616175; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_hi1616176; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo1616177; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_hi1616178; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo1616179; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_hi1616180; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo1616181; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_hi1616182; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo1616183; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_hi1616184; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_hi1616185; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo1616186; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_hi1616187; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo1616188; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_hi1616189; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo1616190; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_hi1616191; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo1616192; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_hi1616193; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo1616194; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_hi1616195; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo1616196; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_hi1616197; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo1616198; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_hi1616199; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_hi1616200; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo1616201; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_hi1616202; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo1616203; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_hi1616204; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s016205; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB14_216206; GFX11-TRUE16-NEXT:  .LBB14_4: ; %cmp.true16207; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v149.h, 316208; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v149.l, 316209; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, v146.l, 316210; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v145.h, 316211; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v145.l, 316212; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v134.l, 316213; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v133.h, 316214; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.h, v133.l, 316215; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, v132.h, 316216; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, v129.h, 316217; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, v129.l, 316218; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, v128.h, 316219; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.l, v128.l, 316220; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, v119.h, 316221; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, v116.h, 316222; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, v116.l, 316223; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, v115.h, 316224; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.h, v112.h, 316225; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, v112.l, 316226; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, v103.h, 316227; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, v103.l, 316228; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, v100.l, 316229; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.l, v99.h, 316230; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, v99.l, 316231; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, v98.h, 316232; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, v87.h, 316233; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, v87.l, 316234; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.h, v86.h, 316235; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, v86.l, 316236; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, v83.l, 316237; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, v82.h, 316238; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, v82.l, 316239; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.l, v81.h, 316240; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, v81.l, 316241; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, v70.l, 316242; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, v69.h, 316243; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, v69.l, 316244; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.h, v68.h, 316245; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, v65.h, 316246; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, v65.l, 316247; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, v64.h, 316248; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, v64.l, 316249; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.l, v53.l, 316250; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, v50.l, 316251; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, v49.h, 316252; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, v49.l, 316253; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, v48.h, 316254; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.h, v48.l, 316255; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, v39.h, 316256; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, v39.l, 316257; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, v38.h, 316258; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, v38.l, 316259; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.l, v37.h, 316260; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, v37.l, 316261; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, v36.h, 316262; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v36.l, 316263; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, v35.h, 316264; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, v35.l, 316265; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, v34.h, 316266; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, v34.l, 316267; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.l, v33.h, 316268; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, v33.l, 316269; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, v32.h, 316270; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, v32.l, 316271; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l16272; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h16273; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.l16274; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h16275; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l16276; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v2.h16277; GFX11-TRUE16-NEXT:    v_and_b16 v3.l, 0xff, v3.l16278; GFX11-TRUE16-NEXT:    v_and_b16 v3.h, 0xff, v3.h16279; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v4.l16280; GFX11-TRUE16-NEXT:    v_and_b16 v4.h, 0xff, v4.h16281; GFX11-TRUE16-NEXT:    v_and_b16 v5.l, 0xff, v5.l16282; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v5.h16283; GFX11-TRUE16-NEXT:    v_and_b16 v6.l, 0xff, v6.l16284; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v6.h16285; GFX11-TRUE16-NEXT:    v_and_b16 v7.l, 0xff, v7.l16286; GFX11-TRUE16-NEXT:    v_and_b16 v7.h, 0xff, v7.h16287; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v8.l16288; GFX11-TRUE16-NEXT:    v_and_b16 v8.h, 0xff, v8.h16289; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v9.l16290; GFX11-TRUE16-NEXT:    v_and_b16 v9.h, 0xff, v9.h16291; GFX11-TRUE16-NEXT:    v_and_b16 v10.l, 0xff, v10.l16292; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v10.h16293; GFX11-TRUE16-NEXT:    v_and_b16 v11.l, 0xff, v11.l16294; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v11.h16295; GFX11-TRUE16-NEXT:    v_and_b16 v12.l, 0xff, v12.l16296; GFX11-TRUE16-NEXT:    v_and_b16 v12.h, 0xff, v12.h16297; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v13.l16298; GFX11-TRUE16-NEXT:    v_and_b16 v13.h, 0xff, v13.h16299; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v14.l16300; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v14.h16301; GFX11-TRUE16-NEXT:    v_and_b16 v15.l, 0xff, v15.l16302; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v15.h16303; GFX11-TRUE16-NEXT:    v_and_b16 v16.l, 0xff, v16.l16304; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v16.h16305; GFX11-TRUE16-NEXT:    v_and_b16 v17.l, 0xff, v17.l16306; GFX11-TRUE16-NEXT:    v_and_b16 v17.h, 0xff, v17.h16307; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v18.l16308; GFX11-TRUE16-NEXT:    v_and_b16 v18.h, 0xff, v18.h16309; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v19.l16310; GFX11-TRUE16-NEXT:    v_and_b16 v19.h, 0xff, v19.h16311; GFX11-TRUE16-NEXT:    v_and_b16 v20.l, 0xff, v20.l16312; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v20.h16313; GFX11-TRUE16-NEXT:    v_and_b16 v21.l, 0xff, v21.l16314; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v21.h16315; GFX11-TRUE16-NEXT:    v_and_b16 v22.l, 0xff, v22.l16316; GFX11-TRUE16-NEXT:    v_and_b16 v22.h, 0xff, v22.h16317; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v23.l16318; GFX11-TRUE16-NEXT:    v_and_b16 v23.h, 0xff, v23.h16319; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v24.l16320; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v24.h16321; GFX11-TRUE16-NEXT:    v_and_b16 v25.l, 0xff, v25.l16322; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v25.h16323; GFX11-TRUE16-NEXT:    v_and_b16 v26.l, 0xff, v26.l16324; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v26.h16325; GFX11-TRUE16-NEXT:    v_and_b16 v27.l, 0xff, v27.l16326; GFX11-TRUE16-NEXT:    v_and_b16 v27.h, 0xff, v27.h16327; GFX11-TRUE16-NEXT:    v_and_b16 v28.l, 0xff, v28.l16328; GFX11-TRUE16-NEXT:    v_and_b16 v28.h, 0xff, v28.h16329; GFX11-TRUE16-NEXT:    v_and_b16 v29.l, 0xff, v29.l16330; GFX11-TRUE16-NEXT:    v_and_b16 v29.h, 0xff, v29.h16331; GFX11-TRUE16-NEXT:    v_and_b16 v30.l, 0xff, v30.l16332; GFX11-TRUE16-NEXT:    v_and_b16 v30.h, 0xff, v30.h16333; GFX11-TRUE16-NEXT:    v_and_b16 v31.l, 0xff, v31.l16334; GFX11-TRUE16-NEXT:    v_and_b16 v31.h, 0xff, v31.h16335; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v150.l, v0.l16336; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v150.h, v0.h16337; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v151.l, v1.l16338; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v151.h, v1.h16339; GFX11-TRUE16-NEXT:    v_or_b16 v2.l, v146.h, v2.l16340; GFX11-TRUE16-NEXT:    v_or_b16 v2.h, v147.l, v2.h16341; GFX11-TRUE16-NEXT:    v_or_b16 v3.l, v147.h, v3.l16342; GFX11-TRUE16-NEXT:    v_or_b16 v3.h, v148.l, v3.h16343; GFX11-TRUE16-NEXT:    v_or_b16 v4.l, v148.h, v4.l16344; GFX11-TRUE16-NEXT:    v_or_b16 v4.h, v134.h, v4.h16345; GFX11-TRUE16-NEXT:    v_or_b16 v5.l, v135.l, v5.l16346; GFX11-TRUE16-NEXT:    v_or_b16 v5.h, v135.h, v5.h16347; GFX11-TRUE16-NEXT:    v_or_b16 v6.l, v144.l, v6.l16348; GFX11-TRUE16-NEXT:    v_or_b16 v6.h, v144.h, v6.h16349; GFX11-TRUE16-NEXT:    v_or_b16 v7.l, v130.l, v7.l16350; GFX11-TRUE16-NEXT:    v_or_b16 v7.h, v130.h, v7.h16351; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v131.l, v8.l16352; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v131.h, v8.h16353; GFX11-TRUE16-NEXT:    v_or_b16 v9.l, v132.l, v9.l16354; GFX11-TRUE16-NEXT:    v_or_b16 v9.h, v117.l, v9.h16355; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v117.h, v10.l16356; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v118.l, v10.h16357; GFX11-TRUE16-NEXT:    v_or_b16 v11.l, v118.h, v11.l16358; GFX11-TRUE16-NEXT:    v_or_b16 v11.h, v119.l, v11.h16359; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v113.l, v12.l16360; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v113.h, v12.h16361; GFX11-TRUE16-NEXT:    v_or_b16 v13.l, v114.l, v13.l16362; GFX11-TRUE16-NEXT:    v_or_b16 v13.h, v114.h, v13.h16363; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v115.l, v14.l16364; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v100.h, v14.h16365; GFX11-TRUE16-NEXT:    v_or_b16 v15.l, v101.l, v15.l16366; GFX11-TRUE16-NEXT:    v_or_b16 v15.h, v101.h, v15.h16367; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v102.l, v16.l16368; GFX11-TRUE16-NEXT:    v_or_b16 v16.h, v102.h, v16.h16369; GFX11-TRUE16-NEXT:    v_or_b16 v17.l, v96.l, v17.l16370; GFX11-TRUE16-NEXT:    v_or_b16 v17.h, v96.h, v17.h16371; GFX11-TRUE16-NEXT:    v_or_b16 v18.l, v97.l, v18.l16372; GFX11-TRUE16-NEXT:    v_or_b16 v18.h, v97.h, v18.h16373; GFX11-TRUE16-NEXT:    v_or_b16 v19.l, v98.l, v19.l16374; GFX11-TRUE16-NEXT:    v_or_b16 v19.h, v83.h, v19.h16375; GFX11-TRUE16-NEXT:    v_or_b16 v20.l, v84.l, v20.l16376; GFX11-TRUE16-NEXT:    v_or_b16 v20.h, v84.h, v20.h16377; GFX11-TRUE16-NEXT:    v_or_b16 v21.l, v85.l, v21.l16378; GFX11-TRUE16-NEXT:    v_or_b16 v21.h, v85.h, v21.h16379; GFX11-TRUE16-NEXT:    v_or_b16 v22.l, v70.h, v22.l16380; GFX11-TRUE16-NEXT:    v_or_b16 v22.h, v71.l, v22.h16381; GFX11-TRUE16-NEXT:    v_or_b16 v23.l, v71.h, v23.l16382; GFX11-TRUE16-NEXT:    v_or_b16 v23.h, v80.l, v23.h16383; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v80.h, v24.l16384; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v66.l, v24.h16385; GFX11-TRUE16-NEXT:    v_or_b16 v25.l, v66.h, v25.l16386; GFX11-TRUE16-NEXT:    v_or_b16 v25.h, v67.l, v25.h16387; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v67.h, v26.l16388; GFX11-TRUE16-NEXT:    v_or_b16 v26.h, v68.l, v26.h16389; GFX11-TRUE16-NEXT:    v_or_b16 v27.l, v53.h, v27.l16390; GFX11-TRUE16-NEXT:    v_or_b16 v27.h, v54.l, v27.h16391; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v54.h, v28.l16392; GFX11-TRUE16-NEXT:    v_or_b16 v28.h, v55.l, v28.h16393; GFX11-TRUE16-NEXT:    v_or_b16 v29.l, v55.h, v29.l16394; GFX11-TRUE16-NEXT:    v_or_b16 v29.h, v50.h, v29.h16395; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v51.l, v30.l16396; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v51.h, v30.h16397; GFX11-TRUE16-NEXT:    v_or_b16 v31.l, v52.l, v31.l16398; GFX11-TRUE16-NEXT:    v_or_b16 v31.h, v52.h, v31.h16399; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, 0x300, v0.l16400; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, 0x300, v0.h16401; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, 0x300, v1.l16402; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, 0x300, v1.h16403; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, 0x300, v2.l16404; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, 0x300, v2.h16405; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, 0x300, v3.l16406; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.h, 0x300, v3.h16407; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, 0x300, v4.l16408; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, 0x300, v4.h16409; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, 0x300, v5.l16410; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, 0x300, v5.h16411; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.l, 0x300, v6.l16412; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, 0x300, v6.h16413; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, 0x300, v7.l16414; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, 0x300, v7.h16415; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, 0x300, v8.l16416; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.h, 0x300, v8.h16417; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, 0x300, v9.l16418; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, 0x300, v9.h16419; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, 0x300, v10.l16420; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, 0x300, v10.h16421; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.l, 0x300, v11.l16422; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, 0x300, v11.h16423; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, 0x300, v12.l16424; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, 0x300, v12.h16425; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, 0x300, v13.l16426; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.h, 0x300, v13.h16427; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, 0x300, v14.l16428; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, 0x300, v14.h16429; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, 0x300, v15.l16430; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, 0x300, v15.h16431; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.l, 0x300, v16.l16432; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, 0x300, v16.h16433; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, 0x300, v17.l16434; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, 0x300, v17.h16435; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, 0x300, v18.l16436; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.h, 0x300, v18.h16437; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, 0x300, v19.l16438; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, 0x300, v19.h16439; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, 0x300, v20.l16440; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, 0x300, v20.h16441; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.l, 0x300, v21.l16442; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, 0x300, v21.h16443; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, 0x300, v22.l16444; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, 0x300, v22.h16445; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, 0x300, v23.l16446; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.h, 0x300, v23.h16447; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, 0x300, v24.l16448; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, 0x300, v24.h16449; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, 0x300, v25.l16450; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, 0x300, v25.h16451; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.l, 0x300, v26.l16452; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, 0x300, v26.h16453; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, 0x300, v27.l16454; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, 0x300, v27.h16455; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, 0x300, v28.l16456; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, 0x300, v28.h16457; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, 0x300, v29.l16458; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, 0x300, v29.h16459; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.l, 0x300, v30.l16460; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, 0x300, v30.h16461; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, 0x300, v31.l16462; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, 0x300, v31.h16463; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s016464; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]16465;16466; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v32i32:16467; GFX11-FAKE16:       ; %bb.0:16468; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)16469; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill16470; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:59216471; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:58816472; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:58416473; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:58016474; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:57616475; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:57216476; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:56816477; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:56416478; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:56016479; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:55616480; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:55216481; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:54816482; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:54416483; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:54016484; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:53616485; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:53216486; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:52816487; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:52416488; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:52016489; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:51616490; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:51216491; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:50816492; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:50416493; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:50016494; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:49616495; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:49216496; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:48816497; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:48416498; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:48016499; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:47616500; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:47216501; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:46816502; GFX11-FAKE16-NEXT:    s_clause 0x12 ; 76-byte Folded Spill16503; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:46416504; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:46016505; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:45616506; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:45216507; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:44816508; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:44416509; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:44016510; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:43616511; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:43216512; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:42816513; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:42416514; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:42016515; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:41616516; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:41216517; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:40816518; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:40416519; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v136, s32 offset:40016520; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v137, s32 offset:39616521; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v138, s32 offset:39216522; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v2816523; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, v26 :: v_dual_mov_b32 v35, v2416524; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, v22 :: v_dual_mov_b32 v37, v2016525; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, v18 :: v_dual_mov_b32 v39, v1616526; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, v14 :: v_dual_mov_b32 v49, v1216527; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v816528; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, v6 :: v_dual_mov_b32 v53, v416529; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, v2 :: v_dual_mov_b32 v55, v016530; GFX11-FAKE16-NEXT:    s_clause 0x1f16531; GFX11-FAKE16-NEXT:    scratch_load_u16 v0, off, s32 offset:38416532; GFX11-FAKE16-NEXT:    scratch_load_u16 v64, off, s32 offset:38016533; GFX11-FAKE16-NEXT:    scratch_load_u16 v2, off, s32 offset:37616534; GFX11-FAKE16-NEXT:    scratch_load_u16 v65, off, s32 offset:37216535; GFX11-FAKE16-NEXT:    scratch_load_u16 v4, off, s32 offset:36816536; GFX11-FAKE16-NEXT:    scratch_load_u16 v66, off, s32 offset:36416537; GFX11-FAKE16-NEXT:    scratch_load_u16 v6, off, s32 offset:36016538; GFX11-FAKE16-NEXT:    scratch_load_u16 v67, off, s32 offset:35616539; GFX11-FAKE16-NEXT:    scratch_load_u16 v8, off, s32 offset:35216540; GFX11-FAKE16-NEXT:    scratch_load_u16 v68, off, s32 offset:34816541; GFX11-FAKE16-NEXT:    scratch_load_u16 v10, off, s32 offset:34416542; GFX11-FAKE16-NEXT:    scratch_load_u16 v69, off, s32 offset:34016543; GFX11-FAKE16-NEXT:    scratch_load_u16 v12, off, s32 offset:33616544; GFX11-FAKE16-NEXT:    scratch_load_u16 v70, off, s32 offset:33216545; GFX11-FAKE16-NEXT:    scratch_load_u16 v14, off, s32 offset:32816546; GFX11-FAKE16-NEXT:    scratch_load_u16 v71, off, s32 offset:32416547; GFX11-FAKE16-NEXT:    scratch_load_u16 v16, off, s32 offset:32016548; GFX11-FAKE16-NEXT:    scratch_load_u16 v80, off, s32 offset:31616549; GFX11-FAKE16-NEXT:    scratch_load_u16 v18, off, s32 offset:31216550; GFX11-FAKE16-NEXT:    scratch_load_u16 v81, off, s32 offset:30816551; GFX11-FAKE16-NEXT:    scratch_load_u16 v20, off, s32 offset:30416552; GFX11-FAKE16-NEXT:    scratch_load_u16 v82, off, s32 offset:30016553; GFX11-FAKE16-NEXT:    scratch_load_u16 v22, off, s32 offset:29616554; GFX11-FAKE16-NEXT:    scratch_load_u16 v83, off, s32 offset:29216555; GFX11-FAKE16-NEXT:    scratch_load_u16 v24, off, s32 offset:28816556; GFX11-FAKE16-NEXT:    scratch_load_u16 v84, off, s32 offset:28416557; GFX11-FAKE16-NEXT:    scratch_load_u16 v26, off, s32 offset:28016558; GFX11-FAKE16-NEXT:    scratch_load_u16 v85, off, s32 offset:27616559; GFX11-FAKE16-NEXT:    scratch_load_u16 v28, off, s32 offset:27216560; GFX11-FAKE16-NEXT:    scratch_load_u16 v86, off, s32 offset:26816561; GFX11-FAKE16-NEXT:    scratch_load_u16 v30, off, s32 offset:26416562; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:26016563; GFX11-FAKE16-NEXT:    s_clause 0x1f16564; GFX11-FAKE16-NEXT:    scratch_load_u16 v31, off, s32 offset:25616565; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:25216566; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:24816567; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:24416568; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:24016569; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:23616570; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:23216571; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:22816572; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:22416573; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:22016574; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:21616575; GFX11-FAKE16-NEXT:    scratch_load_b32 v114, off, s32 offset:38816576; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s3216577; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:816578; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:1616579; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:2416580; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:3216581; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:4016582; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:4816583; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:5616584; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:6416585; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:7216586; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:8016587; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:8816588; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:9616589; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:10416590; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:11216591; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:12016592; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:12816593; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:13616594; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:14416595; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:15216596; GFX11-FAKE16-NEXT:    s_clause 0x1f16597; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:16016598; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:16816599; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:17616600; GFX11-FAKE16-NEXT:    scratch_load_u16 v183, off, s32 offset:18416601; GFX11-FAKE16-NEXT:    scratch_load_u16 v136, off, s32 offset:19216602; GFX11-FAKE16-NEXT:    scratch_load_u16 v137, off, s32 offset:20016603; GFX11-FAKE16-NEXT:    scratch_load_u16 v138, off, s32 offset:20816604; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:21216605; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:20416606; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:19616607; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:18816608; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:18016609; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:17216610; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:16416611; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:15616612; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:14816613; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:14016614; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:13216615; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:12416616; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:11616617; GFX11-FAKE16-NEXT:    scratch_load_u16 v40, off, s32 offset:10816618; GFX11-FAKE16-NEXT:    scratch_load_u16 v41, off, s32 offset:10016619; GFX11-FAKE16-NEXT:    scratch_load_u16 v42, off, s32 offset:9216620; GFX11-FAKE16-NEXT:    scratch_load_u16 v43, off, s32 offset:8416621; GFX11-FAKE16-NEXT:    scratch_load_u16 v44, off, s32 offset:7616622; GFX11-FAKE16-NEXT:    scratch_load_u16 v58, off, s32 offset:6816623; GFX11-FAKE16-NEXT:    scratch_load_u16 v59, off, s32 offset:6016624; GFX11-FAKE16-NEXT:    scratch_load_u16 v60, off, s32 offset:5216625; GFX11-FAKE16-NEXT:    scratch_load_u16 v74, off, s32 offset:4416626; GFX11-FAKE16-NEXT:    scratch_load_u16 v75, off, s32 offset:3616627; GFX11-FAKE16-NEXT:    scratch_load_u16 v76, off, s32 offset:2816628; GFX11-FAKE16-NEXT:    scratch_load_u16 v77, off, s32 offset:2016629; GFX11-FAKE16-NEXT:    s_clause 0x116630; GFX11-FAKE16-NEXT:    scratch_load_u16 v78, off, s32 offset:1216631; GFX11-FAKE16-NEXT:    scratch_load_u16 v92, off, s32 offset:416632; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v124, 8, v116633; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v125, 8, v316634; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v126, 8, v516635; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v127, 8, v716636; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v111, 8, v916637; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v120, 8, v1116638; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v121, 8, v1316639; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v122, 8, v1516640; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v123, 8, v1716641; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v106, 8, v1916642; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v107, 8, v2116643; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v108, 8, v2316644; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v109, 8, v2516645; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v110, 8, v2716646; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v93, 8, v2916647; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(54)16648; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v11416649; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(53)16650; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v94, 8, v11516651; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(52)16652; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v95, 8, v11616653; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(51)16654; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v104, 8, v11716655; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(50)16656; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v105, 8, v11816657; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(49)16658; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v79, 8, v11916659; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(48)16660; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v88, 8, v12816661; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(47)16662; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v89, 8, v12916663; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(46)16664; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v90, 8, v13016665; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(45)16666; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v91, 8, v13116667; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(44)16668; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v61, 8, v14416669; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(43)16670; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v62, 8, v14516671; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(42)16672; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v63, 8, v14616673; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(41)16674; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v72, 8, v14716675; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(40)16676; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v73, 8, v14816677; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(39)16678; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v45, 8, v16216679; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(38)16680; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v46, 8, v16316681; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(37)16682; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v47, 8, v16416683; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(36)16684; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v56, 8, v16516685; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(35)16686; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v57, 8, v16616687; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(34)16688; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v179, 8, v17916689; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(33)16690; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v180, 8, v18016691; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(32)16692; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v181, 8, v18116693; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(31)16694; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v182, 8, v18216695; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(30)16696; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v183, 8, v18316697; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(29)16698; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v162, 8, v13616699; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(28)16700; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v163, 8, v13716701; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(27)16702; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v164, 8, v13816703; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v165, 8, v10316704; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v166, 8, v10216705; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v144, 8, v10116706; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v145, 8, v10016707; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v146, 8, v9916708; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v147, 8, v3116709; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v148, 8, v3016710; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v119, 8, v2816711; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v128, 8, v2616712; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v129, 8, v2416713; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v130, 8, v2216714; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v131, 8, v2016715; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v114, 8, v1816716; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v115, 8, v1616717; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v116, 8, v1416718; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v117, 8, v1216719; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v118, 8, v1016720; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v99, 8, v816721; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v100, 8, v616722; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v101, 8, v416723; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v102, 8, v216724; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v103, 8, v016725; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr3116726; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo16727; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)16728; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s016729; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB14_216730; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false16731; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v5516732; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v5416733; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5316734; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5216735; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v5116736; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v12416737; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v12516738; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v12616739; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v12716740; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5016741; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v016742; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v116743; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v216744; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v316745; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v3916746; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v3716747; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v116748; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v3616749; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v2, v316750; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4916751; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4816752; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v3516753; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v3416754; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v11116755; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v12116756; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v12016757; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v12216758; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v12316759; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v10716760; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v216761; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3816762; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10816763; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v10916764; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v11016765; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v416766; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v10616767; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v516768; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v316769; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v616770; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v816771; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v216772; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v916773; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v1016774; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v1116775; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v516776; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v7, v316777; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v6, v1216778; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v8, v916779; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v10, v1116780; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v3316781; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v3216782; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)16783; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9216784; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v7816785; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v7716786; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v7616787; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v7516788; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v7416789; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v6016790; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v5916791; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v9316792; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v9416793; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v9516794; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v10416795; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v10516796; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v7916797; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v8816798; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v8916799; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v9016800; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v9116801; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v716802; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v816803; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v916804; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v1016805; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v1116806; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v1216807; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v1316808; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v1416809; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v1516810; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v1616811; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v816812; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v1016813; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v1216814; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v13, v1416815; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v15, v1616816; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v5816817; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v4416818; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v4316819; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v4216820; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v4116821; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v4016822; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v17816823; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v17716824; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v17616825; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v16716826; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v6116827; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v6216828; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v6316829; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v7216830; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v7316831; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v4516832; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v18, v4616833; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v4716834; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v5616835; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v5716836; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v1216837; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v1316838; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v1416839; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v1516840; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v1616841; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v1716842; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v1816843; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v1916844; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v2016845; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v2116846; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v1316847; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v1516848; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v16, v1716849; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v18, v1916850; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v2116851; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v16116852; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v16016853; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v15116854; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v15016855; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v14916856; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v13516857; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v13416858; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v13316859; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v13216860; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v11316861; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v17916862; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v18, v18016863; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v18116864; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v18216865; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v18316866; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v16216867; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v23, v16316868; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v16416869; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v25, v16516870; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v26, v16616871; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v1716872; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v1816873; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v1916874; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v2016875; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v2116876; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v2216877; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v2316878; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v2416879; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v2516880; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v2616881; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v1816882; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v2016883; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v2216884; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v23, v2416885; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v25, v2616886; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v11216887; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v9816888; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v9716889; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v9616890; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v8716891; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v8616892; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v8516893; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v8416894; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v8316895; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v8216896; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v14416897; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v23, v14516898; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v14616899; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v25, v14716900; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v26, v14816901; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v11916902; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v12816903; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v29, v12916904; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v30, v13016905; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v31, v13116906; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v2216907; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v2316908; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v2416909; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v2516910; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v2616911; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v2716912; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v2816913; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v2916914; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v3016915; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v3116916; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v2316917; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v2516918; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v2716919; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v28, v2916920; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v30, v3116921; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v8116922; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v8016923; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v7116924; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v7016925; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v6916926; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v6816927; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v6716928; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xff, v6616929; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xff, v6516930; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v6416931; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v11416932; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v11516933; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v29, v11616934; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v30, v11716935; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v31, v11816936; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v9916937; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v10016938; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v34, v10116939; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v35, v10216940; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v36, v10316941; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v2716942; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v2816943; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v2916944; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v3016945; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v3116946; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v3216947; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v3316948; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v3416949; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v3516950; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v3616951; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v2816952; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v3016953; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v3216954; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v3416955; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v35, v3616956; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5516957; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5416958; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5316959; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5216960; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5116961; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5016962; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4916963; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4816964; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr3916965; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr3816966; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr3716967; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr3616968; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr3516969; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr3416970; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr3316971; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr3216972; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9216973; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7816974; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7716975; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7616976; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7516977; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7416978; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6016979; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5916980; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5816981; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4416982; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4316983; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4216984; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4116985; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4016986; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr17816987; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr17716988; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr17616989; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16716990; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16116991; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16016992; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr15116993; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr15016994; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr14916995; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13516996; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13416997; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13316998; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13216999; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11317000; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11217001; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9817002; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9717003; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9617004; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8717005; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8617006; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8517007; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8417008; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8317009; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8217010; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8117011; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8017012; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7117013; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7017014; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6917015; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6817016; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6717017; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6617018; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6517019; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6417020; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12417021; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12517022; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12617023; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12717024; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11117025; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12017026; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12117027; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12217028; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12317029; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10617030; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10717031; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10817032; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10917033; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11017034; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9317035; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9417036; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9517037; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10417038; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10517039; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7917040; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8817041; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8917042; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9017043; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9117044; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6117045; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6217046; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6317047; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7217048; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr7317049; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4517050; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4617051; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4717052; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5617053; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr5717054; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr17917055; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr18017056; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr18117057; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr18217058; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr18317059; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16217060; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16317061; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16417062; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16517063; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16617064; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr14417065; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr14517066; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr14617067; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr14717068; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr14817069; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11917070; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12817071; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12917072; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13017073; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13117074; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11417075; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11517076; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11617077; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11717078; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11817079; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9917080; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10017081; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10117082; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10217083; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10317084; GFX11-FAKE16-NEXT:  .LBB14_2: ; %Flow17085; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s017086; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB14_417087; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true17088; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v55, 317089; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v54, 317090; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v53, 317091; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v52, 317092; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v51, 317093; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v017094; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v117095; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v217096; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v317097; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v50, 317098; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v124, v017099; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v125, v117100; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v126, v217101; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v127, v317102; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v49, 317103; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, 0x300, v017104; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, 0x300, v117105; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v217106; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v317107; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v48, 317108; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v017109; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v117110; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v217111; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v317112; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v37, 317113; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v36, 317114; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v117115; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, v35, 317116; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v2, v317117; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v39, 317118; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v417119; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v517120; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v617121; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v717122; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v217123; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, v34, 317124; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v817125; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v917126; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v1017127; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v123, v217128; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v1117129; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v111, v317130; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v120, v417131; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v121, v517132; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v217133; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v122, v617134; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v107, v817135; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v108, v917136; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v109, v1017137; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v217138; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v38, 317139; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v110, v1117140; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v317141; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v417142; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v517143; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v217144; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v617145; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v817146; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v917147; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v1017148; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v106, v217149; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v1117150; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v317151; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v417152; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v517153; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v217154; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v617155; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v817156; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v917157; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v1017158; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v217159; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v1117160; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v417161; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v5, v617162; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v8, v917163; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v7, v1217164; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v10, v1117165; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v33, 317166; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v32, 317167; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)17168; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v92, 317169; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, v78, 317170; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, v77, 317171; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v76, 317172; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v75, 317173; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v74, 317174; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, v60, 317175; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, v59, 317176; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v717177; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v817178; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v917179; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v1017180; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v1117181; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v1217182; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v1317183; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v1417184; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v1517185; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v1617186; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v93, v717187; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v94, v817188; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v95, v917189; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v104, v1017190; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v105, v1117191; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v79, v1217192; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v88, v1317193; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v89, v1417194; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v90, v1517195; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v91, v1617196; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v717197; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v817198; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v917199; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v1017200; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v1117201; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v1217202; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v1317203; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v1417204; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v1517205; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v1617206; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v717207; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v817208; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v917209; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v1017210; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v1117211; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v1217212; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v1317213; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v1417214; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v1517215; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v1617216; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v817217; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v1017218; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v1217219; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v13, v1417220; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v15, v1617221; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v58, 317222; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v44, 317223; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v43, 317224; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, v42, 317225; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, v41, 317226; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v40, 317227; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v178, 317228; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v177, 317229; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, v176, 317230; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, v167, 317231; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v1217232; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v1317233; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v1417234; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v1517235; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v1617236; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v1717237; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v1817238; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v1917239; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v2017240; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v2117241; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v61, v1217242; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v62, v1317243; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v63, v1417244; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v72, v1517245; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v73, v1617246; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v45, v1717247; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v46, v1817248; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v47, v1917249; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v56, v2017250; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v57, v2117251; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v1217252; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v1317253; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v1417254; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v1517255; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v1617256; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v1717257; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v1817258; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v1917259; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v2017260; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v2117261; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v1217262; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v1317263; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v1417264; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v1517265; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v1617266; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v1717267; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v1817268; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v1917269; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v2017270; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v2117271; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v1317272; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v1517273; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v16, v1717274; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v18, v1917275; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v2117276; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v161, 317277; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v160, 317278; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v151, 317279; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, v150, 317280; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, v149, 317281; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v135, 317282; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v134, 317283; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v133, 317284; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, v132, 317285; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, v113, 317286; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v1717287; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v1817288; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v1917289; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v2017290; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v2117291; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v2217292; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v2317293; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v2417294; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v2517295; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v2617296; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v179, v1717297; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v180, v1817298; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v181, v1917299; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v182, v2017300; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v183, v2117301; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v162, v2217302; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v163, v2317303; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v164, v2417304; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v165, v2517305; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v166, v2617306; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v1717307; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v1817308; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v1917309; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v2017310; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v2117311; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v2217312; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v2317313; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v2417314; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v2517315; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v2617316; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v1717317; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v1817318; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v1917319; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v2017320; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v2117321; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v2217322; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v2317323; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v2417324; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v2517325; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v2617326; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v1817327; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v2017328; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v2217329; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v23, v2417330; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v25, v2617331; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v112, 317332; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v98, 317333; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v97, 317334; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, v96, 317335; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, v87, 317336; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v86, 317337; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v85, 317338; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, v84, 317339; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v83, 317340; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v82, 317341; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v2217342; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v2317343; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v2417344; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v2517345; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v2617346; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v2717347; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v2817348; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v2917349; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v3017350; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v3117351; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v144, v2217352; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v145, v2317353; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v146, v2417354; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v147, v2517355; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v148, v2617356; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v119, v2717357; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v128, v2817358; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v129, v2917359; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v130, v3017360; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v131, v3117361; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v2217362; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v2317363; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v2417364; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v2517365; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v2617366; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v2717367; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v2817368; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, 0x300, v2917369; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, 0x300, v3017370; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, 0x300, v3117371; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v2217372; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v2317373; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v2417374; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v2517375; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v2617376; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v2717377; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v2817378; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v2917379; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v3017380; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v3117381; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v2317382; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v2517383; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v2717384; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v28, v2917385; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v30, v3117386; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v81, 317387; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v80, 317388; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, v71, 317389; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v70, 317390; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v69, 317391; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, v68, 317392; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, v67, 317393; GFX11-FAKE16-NEXT:    v_add_nc_u16 v34, v66, 317394; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, v65, 317395; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, v64, 317396; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v2717397; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v2817398; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v2917399; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v3017400; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v3117401; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v3217402; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v3317403; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xff, v3417404; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xff, v3517405; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v3617406; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v114, v2717407; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v115, v2817408; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v116, v2917409; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v117, v3017410; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v118, v3117411; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v99, v3217412; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v100, v3317413; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v101, v3417414; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v102, v3517415; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v103, v3617416; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v2717417; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v2817418; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, 0x300, v2917419; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, 0x300, v3017420; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, 0x300, v3117421; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, 0x300, v3217422; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, 0x300, v3317423; GFX11-FAKE16-NEXT:    v_add_nc_u16 v34, 0x300, v3417424; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, 0x300, v3517425; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, 0x300, v3617426; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v2717427; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v2817428; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v2917429; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v3017430; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v3117431; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v3217432; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v3317433; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v3417434; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v3517435; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v3617436; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v2817437; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v3017438; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v3217439; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v3417440; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v35, v3617441; GFX11-FAKE16-NEXT:  .LBB14_4: ; %end17442; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s017443; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload17444; GFX11-FAKE16-NEXT:    scratch_load_b32 v138, off, s32 offset:39217445; GFX11-FAKE16-NEXT:    scratch_load_b32 v137, off, s32 offset:39617446; GFX11-FAKE16-NEXT:    scratch_load_b32 v136, off, s32 offset:40017447; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:40417448; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:40817449; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:41217450; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:41617451; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:42017452; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:42417453; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:42817454; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:43217455; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:43617456; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:44017457; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:44417458; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:44817459; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:45217460; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:45617461; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:46017462; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:46417463; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:46817464; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:47217465; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:47617466; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:48017467; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:48417468; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:48817469; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:49217470; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:49617471; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:50017472; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:50417473; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:50817474; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:51217475; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:51617476; GFX11-FAKE16-NEXT:    s_clause 0x12 ; 76-byte Folded Reload17477; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:52017478; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:52417479; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:52817480; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:53217481; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:53617482; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:54017483; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:54417484; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:54817485; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:55217486; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:55617487; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:56017488; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:56417489; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:56817490; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:57217491; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:57617492; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:58017493; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:58417494; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:58817495; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:59217496; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)17497; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]17498  %cmp = icmp eq i32 %b, 017499  br i1 %cmp, label %cmp.true, label %cmp.false17500 17501cmp.true:17502  %a1 = add <128 x i8> %a, splat (i8 3)17503  %a2 = bitcast <128 x i8> %a1 to <32 x i32>17504  br label %end17505 17506cmp.false:17507  %a3 = bitcast <128 x i8> %a to <32 x i32>17508  br label %end17509 17510end:17511  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]17512  ret <32 x i32> %phi17513}17514 17515define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i32 inreg %b) {17516; SI-LABEL: bitcast_v128i8_to_v32i32_scalar:17517; SI:       ; %bb.0:17518; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)17519; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill17520; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill17521; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill17522; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill17523; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill17524; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill17525; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill17526; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill17527; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill17528; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill17529; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill17530; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill17531; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill17532; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill17533; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill17534; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill17535; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill17536; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill17537; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill17538; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill17539; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill17540; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill17541; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill17542; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill17543; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill17544; SI-NEXT:    s_waitcnt expcnt(0)17545; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:33217546; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s3217547; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:817548; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:1617549; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:2417550; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:3217551; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:4017552; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:4817553; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:5617554; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:6417555; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:7217556; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:8017557; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:8817558; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:9617559; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:10417560; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:11217561; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:12017562; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:12817563; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:13617564; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:14417565; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:15217566; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:16017567; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:16817568; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:17617569; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v117570; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill17571; SI-NEXT:    v_lshlrev_b32_e32 v20, 8, v317572; SI-NEXT:    v_lshlrev_b32_e32 v62, 24, v517573; SI-NEXT:    v_lshlrev_b32_e32 v22, 8, v717574; SI-NEXT:    v_lshlrev_b32_e32 v26, 24, v917575; SI-NEXT:    v_lshlrev_b32_e32 v30, 8, v1117576; SI-NEXT:    v_lshlrev_b32_e32 v47, 24, v1317577; SI-NEXT:    v_lshlrev_b32_e32 v61, 8, v1517578; SI-NEXT:    v_lshlrev_b32_e32 v63, 24, v2117579; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v2317580; SI-NEXT:    v_lshlrev_b32_e32 v23, 8, v2717581; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v1717582; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v1917583; SI-NEXT:    v_lshlrev_b32_e32 v59, 24, v2517584; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v2917585; SI-NEXT:    s_waitcnt vmcnt(14)17586; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v017587; SI-NEXT:    v_lshlrev_b32_e32 v15, 8, v4517588; SI-NEXT:    v_lshlrev_b32_e32 v29, 24, v4417589; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v4317590; SI-NEXT:    v_lshlrev_b32_e32 v44, 24, v4217591; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v4117592; SI-NEXT:    v_lshlrev_b32_e32 v25, 24, v5517593; SI-NEXT:    s_and_b64 s[4:5], vcc, exec17594; SI-NEXT:    v_lshlrev_b32_e32 v21, 24, v5317595; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v4017596; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v5017597; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill17598; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17599; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4817600; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill17601; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17602; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v5417603; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill17604; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17605; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5217606; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill17607; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17608; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v5117609; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill17610; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17611; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3617612; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill17613; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17614; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3117615; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill17616; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17617; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3317618; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill17619; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17620; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3417621; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill17622; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17623; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3517624; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill17625; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17626; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3717627; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill17628; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17629; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3817630; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill17631; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17632; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3917633; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill17634; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)17635; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4917636; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill17637; SI-NEXT:    s_waitcnt expcnt(0)17638; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:18417639; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:19217640; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:20017641; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:20817642; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:21617643; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:22417644; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:23217645; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:24017646; SI-NEXT:    v_lshlrev_b32_e32 v32, 8, v3217647; SI-NEXT:    s_waitcnt vmcnt(7)17648; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v017649; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill17650; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17651; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v117652; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill17653; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17654; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v1317655; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill17656; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17657; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v317658; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill17659; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17660; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v1117661; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill17662; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17663; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v517664; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill17665; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17666; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v917667; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill17668; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17669; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v717670; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill17671; SI-NEXT:    s_waitcnt expcnt(0)17672; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:24817673; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:25617674; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:26417675; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:27217676; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:28017677; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:28817678; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:29617679; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:30417680; SI-NEXT:    s_waitcnt vmcnt(7)17681; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v017682; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill17683; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17684; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v117685; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill17686; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17687; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v1317688; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill17689; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17690; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v317691; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill17692; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17693; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v1117694; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill17695; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)17696; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v517697; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill17698; SI-NEXT:    s_waitcnt expcnt(0)17699; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:31217700; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:32017701; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:32817702; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:417703; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:1217704; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:2017705; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:2817706; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:3617707; SI-NEXT:    s_waitcnt vmcnt(14)17708; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v917709; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v717710; SI-NEXT:    s_waitcnt vmcnt(7)17711; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v017712; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill17713; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:4417714; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:5217715; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:6017716; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:6817717; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:7617718; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:8417719; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:9217720; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:10017721; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:10817722; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:11617723; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:12417724; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:13217725; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:14017726; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:14817727; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:15617728; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:16417729; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:17217730; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:18017731; SI-NEXT:    s_waitcnt expcnt(0)17732; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:18817733; SI-NEXT:    s_waitcnt vmcnt(14)17734; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v117735; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v317736; SI-NEXT:    s_waitcnt vmcnt(0)17737; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill17738; SI-NEXT:    s_waitcnt expcnt(0)17739; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:19617740; SI-NEXT:    s_waitcnt vmcnt(0)17741; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill17742; SI-NEXT:    s_waitcnt expcnt(0)17743; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:20417744; SI-NEXT:    s_waitcnt vmcnt(0)17745; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill17746; SI-NEXT:    s_waitcnt expcnt(0)17747; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:21217748; SI-NEXT:    s_waitcnt vmcnt(0)17749; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill17750; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:22017751; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:22817752; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:23617753; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:24417754; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:25217755; SI-NEXT:    s_waitcnt expcnt(0)17756; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:26017757; SI-NEXT:    s_waitcnt vmcnt(0)17758; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill17759; SI-NEXT:    s_waitcnt expcnt(0)17760; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:26817761; SI-NEXT:    s_waitcnt vmcnt(0)17762; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill17763; SI-NEXT:    s_waitcnt expcnt(0)17764; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:27617765; SI-NEXT:    s_waitcnt vmcnt(0)17766; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill17767; SI-NEXT:    s_waitcnt expcnt(0)17768; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28417769; SI-NEXT:    s_waitcnt vmcnt(0)17770; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill17771; SI-NEXT:    s_waitcnt expcnt(0)17772; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:29217773; SI-NEXT:    s_waitcnt vmcnt(0)17774; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill17775; SI-NEXT:    s_waitcnt expcnt(0)17776; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:30017777; SI-NEXT:    s_waitcnt vmcnt(0)17778; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill17779; SI-NEXT:    s_waitcnt expcnt(0)17780; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:30817781; SI-NEXT:    s_waitcnt vmcnt(0)17782; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill17783; SI-NEXT:    s_waitcnt expcnt(0)17784; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:31617785; SI-NEXT:    s_waitcnt vmcnt(0)17786; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill17787; SI-NEXT:    s_waitcnt expcnt(0)17788; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32417789; SI-NEXT:    s_waitcnt vmcnt(0)17790; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill17791; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill17792; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill17793; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill17794; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill17795; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill17796; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill17797; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill17798; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill17799; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill17800; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill17801; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill17802; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill17803; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill17804; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill17805; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill17806; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill17807; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill17808; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill17809; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill17810; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill17811; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill17812; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill17813; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill17814; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill17815; SI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill17816; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill17817; SI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill17818; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill17819; SI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill17820; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill17821; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill17822; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill17823; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill17824; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill17825; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill17826; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill17827; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill17828; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill17829; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill17830; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill17831; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill17832; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill17833; SI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill17834; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill17835; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill17836; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill17837; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill17838; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill17839; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill17840; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill17841; SI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill17842; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill17843; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill17844; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill17845; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill17846; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill17847; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill17848; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill17849; SI-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill17850; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill17851; SI-NEXT:    s_cbranch_scc0 .LBB15_217852; SI-NEXT:  ; %bb.1: ; %cmp.false17853; SI-NEXT:    v_and_b32_e32 v0, 0xff, v217854; SI-NEXT:    v_and_b32_e32 v1, 0xff, v417855; SI-NEXT:    v_or_b32_e32 v0, v0, v2017856; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117857; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017858; SI-NEXT:    v_or_b32_e32 v1, v62, v117859; SI-NEXT:    v_or_b32_e32 v4, v0, v117860; SI-NEXT:    v_and_b32_e32 v0, 0xff, v1017861; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1217862; SI-NEXT:    v_or_b32_e32 v0, v0, v3017863; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117864; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017865; SI-NEXT:    v_or_b32_e32 v1, v47, v117866; SI-NEXT:    v_and_b32_e32 v2, 0xff, v617867; SI-NEXT:    v_or_b32_e32 v6, v0, v117868; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload17869; SI-NEXT:    s_waitcnt expcnt(4)17870; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload17871; SI-NEXT:    v_and_b32_e32 v3, 0xff, v817872; SI-NEXT:    v_and_b32_e32 v0, 0xff, v1417873; SI-NEXT:    v_or_b32_e32 v2, v2, v2217874; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v317875; SI-NEXT:    v_or_b32_e32 v0, v0, v6117876; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v217877; SI-NEXT:    v_or_b32_e32 v3, v26, v317878; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017879; SI-NEXT:    v_or_b32_e32 v5, v2, v317880; SI-NEXT:    v_mov_b32_e32 v3, v717881; SI-NEXT:    v_mov_b32_e32 v2, v917882; SI-NEXT:    s_and_b32 s4, s28, 0xff17883; SI-NEXT:    s_lshl_b32 s5, s29, 817884; SI-NEXT:    s_or_b32 s4, s4, s517885; SI-NEXT:    s_and_b32 s4, s4, 0xffff17886; SI-NEXT:    s_lshl_b32 s5, s17, 817887; SI-NEXT:    s_lshl_b32 s6, s19, 2417888; SI-NEXT:    s_lshl_b32 s7, s23, 2417889; SI-NEXT:    s_lshl_b32 s8, s27, 2417890; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload17891; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload17892; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload17893; SI-NEXT:    s_waitcnt vmcnt(4)17894; SI-NEXT:    v_and_b32_e32 v1, 0xff, v117895; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117896; SI-NEXT:    v_or_b32_e32 v1, v58, v117897; SI-NEXT:    v_or_b32_e32 v7, v0, v117898; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload17899; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload17900; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload17901; SI-NEXT:    s_waitcnt vmcnt(2)17902; SI-NEXT:    v_and_b32_e32 v0, 0xff, v017903; SI-NEXT:    s_waitcnt vmcnt(1)17904; SI-NEXT:    v_and_b32_e32 v1, 0xff, v117905; SI-NEXT:    v_or_b32_e32 v0, v0, v5717906; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117907; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017908; SI-NEXT:    v_or_b32_e32 v1, v63, v117909; SI-NEXT:    v_or_b32_e32 v8, v0, v117910; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload17911; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload17912; SI-NEXT:    s_waitcnt vmcnt(1)17913; SI-NEXT:    v_and_b32_e32 v0, 0xff, v017914; SI-NEXT:    s_waitcnt vmcnt(0)17915; SI-NEXT:    v_and_b32_e32 v1, 0xff, v117916; SI-NEXT:    v_or_b32_e32 v0, v0, v6017917; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117918; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017919; SI-NEXT:    v_or_b32_e32 v1, v59, v117920; SI-NEXT:    v_or_b32_e32 v9, v0, v117921; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload17922; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload17923; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload17924; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload17925; SI-NEXT:    s_waitcnt vmcnt(3)17926; SI-NEXT:    v_and_b32_e32 v0, 0xff, v017927; SI-NEXT:    s_waitcnt vmcnt(2)17928; SI-NEXT:    v_and_b32_e32 v1, 0xff, v117929; SI-NEXT:    v_or_b32_e32 v0, v0, v2317930; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117931; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017932; SI-NEXT:    v_or_b32_e32 v1, v19, v117933; SI-NEXT:    v_or_b32_e32 v10, v0, v117934; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload17935; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload17936; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload17937; SI-NEXT:    s_waitcnt expcnt(2)17938; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload17939; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1117940; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117941; SI-NEXT:    v_or_b32_e32 v1, v29, v117942; SI-NEXT:    s_waitcnt vmcnt(3)17943; SI-NEXT:    v_and_b32_e32 v0, 0xff, v017944; SI-NEXT:    v_or_b32_e32 v0, v0, v1517945; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload17946; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017947; SI-NEXT:    s_waitcnt expcnt(0)17948; SI-NEXT:    v_or_b32_e32 v11, v0, v117949; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5017950; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1617951; SI-NEXT:    v_or_b32_e32 v0, v0, v1717952; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117953; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017954; SI-NEXT:    v_or_b32_e32 v1, v44, v117955; SI-NEXT:    v_or_b32_e32 v12, v0, v117956; SI-NEXT:    v_and_b32_e32 v0, 0xff, v4517957; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1317958; SI-NEXT:    v_or_b32_e32 v0, v0, v4117959; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117960; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017961; SI-NEXT:    v_or_b32_e32 v1, v25, v117962; SI-NEXT:    v_or_b32_e32 v13, v0, v117963; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5117964; SI-NEXT:    v_and_b32_e32 v1, 0xff, v3617965; SI-NEXT:    v_or_b32_e32 v0, v0, v4017966; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117967; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017968; SI-NEXT:    v_or_b32_e32 v1, v21, v117969; SI-NEXT:    v_or_b32_e32 v14, v0, v117970; SI-NEXT:    v_and_b32_e32 v0, 0xff, v2717971; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1817972; SI-NEXT:    v_or_b32_e32 v0, v0, v3217973; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117974; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017975; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload17976; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload17977; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload17978; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload17979; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload17980; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload17981; SI-NEXT:    s_waitcnt vmcnt(6)17982; SI-NEXT:    v_or_b32_e32 v1, v15, v117983; SI-NEXT:    v_or_b32_e32 v15, v0, v117984; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload17985; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5317986; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload17987; SI-NEXT:    s_waitcnt vmcnt(7)17988; SI-NEXT:    v_mov_b32_e32 v50, v1617989; SI-NEXT:    s_waitcnt vmcnt(1)17990; SI-NEXT:    v_or_b32_e32 v0, v0, v117991; SI-NEXT:    v_and_b32_e32 v1, 0xff, v4817992; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v117993; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v017994; SI-NEXT:    v_or_b32_e32 v1, v16, v117995; SI-NEXT:    v_or_b32_e32 v16, v0, v117996; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload17997; SI-NEXT:    v_and_b32_e32 v0, 0xff, v3117998; SI-NEXT:    v_mov_b32_e32 v48, v2217999; SI-NEXT:    s_waitcnt vmcnt(0)18000; SI-NEXT:    v_mov_b32_e32 v32, v118001; SI-NEXT:    v_or_b32_e32 v0, v0, v118002; SI-NEXT:    v_and_b32_e32 v1, 0xff, v3418003; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118004; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018005; SI-NEXT:    v_or_b32_e32 v1, v17, v118006; SI-NEXT:    v_or_b32_e32 v17, v0, v118007; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload18008; SI-NEXT:    v_and_b32_e32 v0, 0xff, v3318009; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload18010; SI-NEXT:    s_waitcnt vmcnt(1)18011; SI-NEXT:    v_mov_b32_e32 v40, v118012; SI-NEXT:    v_or_b32_e32 v0, v0, v118013; SI-NEXT:    v_and_b32_e32 v1, 0xff, v3518014; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118015; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018016; SI-NEXT:    v_or_b32_e32 v1, v18, v118017; SI-NEXT:    v_or_b32_e32 v18, v0, v118018; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload18019; SI-NEXT:    v_and_b32_e32 v0, 0xff, v3918020; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload18021; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload18022; SI-NEXT:    s_waitcnt vmcnt(2)18023; SI-NEXT:    v_mov_b32_e32 v57, v118024; SI-NEXT:    v_or_b32_e32 v0, v0, v118025; SI-NEXT:    v_and_b32_e32 v1, 0xff, v3818026; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118027; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018028; SI-NEXT:    v_or_b32_e32 v1, v19, v118029; SI-NEXT:    v_or_b32_e32 v19, v0, v118030; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload18031; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5418032; SI-NEXT:    v_mov_b32_e32 v54, v2318033; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload18034; SI-NEXT:    s_waitcnt vmcnt(1)18035; SI-NEXT:    v_or_b32_e32 v0, v0, v118036; SI-NEXT:    v_and_b32_e32 v1, 0xff, v3718037; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118038; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018039; SI-NEXT:    v_or_b32_e32 v1, v20, v118040; SI-NEXT:    v_or_b32_e32 v20, v0, v118041; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload18042; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5218043; SI-NEXT:    s_waitcnt vmcnt(0)18044; SI-NEXT:    v_or_b32_e32 v0, v0, v118045; SI-NEXT:    v_and_b32_e32 v1, 0xff, v2418046; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118047; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018048; SI-NEXT:    v_or_b32_e32 v1, v21, v118049; SI-NEXT:    v_or_b32_e32 v21, v0, v118050; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload18051; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload18052; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload18053; SI-NEXT:    v_and_b32_e32 v0, 0xff, v2818054; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload18055; SI-NEXT:    s_waitcnt vmcnt(3)18056; SI-NEXT:    v_or_b32_e32 v0, v0, v118057; SI-NEXT:    v_and_b32_e32 v1, 0xff, v4918058; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118059; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018060; SI-NEXT:    v_or_b32_e32 v1, v22, v118061; SI-NEXT:    v_or_b32_e32 v22, v0, v118062; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload18063; SI-NEXT:    v_and_b32_e32 v0, 0xff, v6118064; SI-NEXT:    s_waitcnt vmcnt(3)18065; SI-NEXT:    v_mov_b32_e32 v45, v2418066; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload18067; SI-NEXT:    s_waitcnt vmcnt(1)18068; SI-NEXT:    v_mov_b32_e32 v34, v118069; SI-NEXT:    v_or_b32_e32 v0, v0, v118070; SI-NEXT:    v_and_b32_e32 v1, 0xff, v6018071; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118072; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018073; SI-NEXT:    v_or_b32_e32 v1, v23, v118074; SI-NEXT:    v_or_b32_e32 v23, v0, v118075; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload18076; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5918077; SI-NEXT:    s_waitcnt vmcnt(0)18078; SI-NEXT:    v_or_b32_e32 v0, v0, v118079; SI-NEXT:    v_and_b32_e32 v1, 0xff, v5818080; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118081; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018082; SI-NEXT:    v_or_b32_e32 v1, v24, v118083; SI-NEXT:    v_or_b32_e32 v24, v0, v118084; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload18085; SI-NEXT:    v_and_b32_e32 v0, 0xff, v4318086; SI-NEXT:    s_waitcnt vmcnt(0)18087; SI-NEXT:    v_mov_b32_e32 v43, v118088; SI-NEXT:    v_or_b32_e32 v0, v0, v118089; SI-NEXT:    v_and_b32_e32 v1, 0xff, v5518090; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118091; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018092; SI-NEXT:    v_or_b32_e32 v1, v25, v118093; SI-NEXT:    v_or_b32_e32 v25, v0, v118094; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload18095; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5618096; SI-NEXT:    v_mov_b32_e32 v55, v5618097; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload18098; SI-NEXT:    s_waitcnt vmcnt(1)18099; SI-NEXT:    v_or_b32_e32 v0, v0, v118100; SI-NEXT:    v_and_b32_e32 v1, 0xff, v4218101; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118102; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018103; SI-NEXT:    v_or_b32_e32 v1, v33, v118104; SI-NEXT:    v_or_b32_e32 v26, v0, v118105; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload18106; SI-NEXT:    v_and_b32_e32 v0, 0xff, v4618107; SI-NEXT:    v_mov_b32_e32 v42, v4618108; SI-NEXT:    v_mov_b32_e32 v46, v3318109; SI-NEXT:    s_waitcnt vmcnt(0)18110; SI-NEXT:    v_mov_b32_e32 v37, v118111; SI-NEXT:    v_or_b32_e32 v0, v0, v118112; SI-NEXT:    v_and_b32_e32 v1, 0xff, v4118113; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118114; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018115; SI-NEXT:    v_or_b32_e32 v1, v27, v118116; SI-NEXT:    v_or_b32_e32 v27, v0, v118117; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload18118; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5318119; SI-NEXT:    s_waitcnt vmcnt(0)18120; SI-NEXT:    v_or_b32_e32 v0, v0, v118121; SI-NEXT:    v_and_b32_e32 v1, 0xff, v4718122; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118123; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018124; SI-NEXT:    v_or_b32_e32 v1, v28, v118125; SI-NEXT:    v_or_b32_e32 v28, v0, v118126; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload18127; SI-NEXT:    v_and_b32_e32 v0, 0xff, v6218128; SI-NEXT:    s_waitcnt vmcnt(0)18129; SI-NEXT:    v_mov_b32_e32 v36, v118130; SI-NEXT:    v_or_b32_e32 v0, v0, v118131; SI-NEXT:    v_and_b32_e32 v1, 0xff, v5618132; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118133; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018134; SI-NEXT:    v_or_b32_e32 v1, v2, v118135; SI-NEXT:    v_or_b32_e32 v29, v0, v118136; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload18137; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload18138; SI-NEXT:    v_and_b32_e32 v0, 0xff, v4918139; SI-NEXT:    s_waitcnt vmcnt(1)18140; SI-NEXT:    v_or_b32_e32 v0, v0, v118141; SI-NEXT:    v_and_b32_e32 v1, 0xff, v3918142; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118143; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018144; SI-NEXT:    s_waitcnt vmcnt(0)18145; SI-NEXT:    v_or_b32_e32 v1, v2, v118146; SI-NEXT:    v_or_b32_e32 v30, v0, v118147; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload18148; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload18149; SI-NEXT:    v_and_b32_e32 v0, 0xff, v3818150; SI-NEXT:    v_or_b32_e32 v0, v0, v318151; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018152; SI-NEXT:    s_waitcnt vmcnt(1)18153; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118154; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118155; SI-NEXT:    s_waitcnt vmcnt(0)18156; SI-NEXT:    v_or_b32_e32 v1, v2, v118157; SI-NEXT:    v_or_b32_e32 v31, v0, v118158; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload18159; SI-NEXT:    v_and_b32_e32 v0, 0xff, v4418160; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v018161; SI-NEXT:    s_waitcnt vmcnt(0)18162; SI-NEXT:    v_or_b32_e32 v0, v1, v018163; SI-NEXT:    v_or_b32_e32 v3, s4, v018164; SI-NEXT:    s_and_b32 s4, s16, 0xff18165; SI-NEXT:    s_or_b32 s4, s4, s518166; SI-NEXT:    s_and_b32 s5, s18, 0xff18167; SI-NEXT:    s_lshl_b32 s5, s5, 1618168; SI-NEXT:    s_and_b32 s4, s4, 0xffff18169; SI-NEXT:    s_or_b32 s5, s6, s518170; SI-NEXT:    s_or_b32 s4, s4, s518171; SI-NEXT:    s_and_b32 s5, s20, 0xff18172; SI-NEXT:    s_lshl_b32 s6, s21, 818173; SI-NEXT:    s_or_b32 s5, s5, s618174; SI-NEXT:    s_and_b32 s6, s22, 0xff18175; SI-NEXT:    s_lshl_b32 s6, s6, 1618176; SI-NEXT:    s_and_b32 s5, s5, 0xffff18177; SI-NEXT:    s_or_b32 s6, s7, s618178; SI-NEXT:    s_or_b32 s5, s5, s618179; SI-NEXT:    s_and_b32 s6, s24, 0xff18180; SI-NEXT:    s_lshl_b32 s7, s25, 818181; SI-NEXT:    s_or_b32 s6, s6, s718182; SI-NEXT:    s_and_b32 s7, s26, 0xff18183; SI-NEXT:    s_lshl_b32 s7, s7, 1618184; SI-NEXT:    s_and_b32 s6, s6, 0xffff18185; SI-NEXT:    s_or_b32 s7, s8, s718186; SI-NEXT:    s_or_b32 s6, s6, s718187; SI-NEXT:    v_mov_b32_e32 v0, s418188; SI-NEXT:    v_mov_b32_e32 v1, s518189; SI-NEXT:    v_mov_b32_e32 v2, s618190; SI-NEXT:    s_mov_b64 s[4:5], 018191; SI-NEXT:    s_branch .LBB15_318192; SI-NEXT:  .LBB15_2:18193; SI-NEXT:    v_mov_b32_e32 v55, v5618194; SI-NEXT:    v_mov_b32_e32 v42, v4618195; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload18196; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload18197; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload18198; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload18199; SI-NEXT:    s_waitcnt expcnt(4)18200; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload18201; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload18202; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload18203; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload18204; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload18205; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload18206; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload18207; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload18208; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload18209; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload18210; SI-NEXT:    s_waitcnt expcnt(1)18211; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload18212; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload18213; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload18214; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload18215; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload18216; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload18217; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload18218; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload18219; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload18220; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload18221; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload18222; SI-NEXT:    s_mov_b64 s[4:5], -118223; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr3118224; SI-NEXT:  .LBB15_3: ; %Flow18225; SI-NEXT:    s_waitcnt vmcnt(7)18226; SI-NEXT:    v_mov_b32_e32 v35, v5718227; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload18228; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload18229; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload18230; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]18231; SI-NEXT:    s_cbranch_vccnz .LBB15_518232; SI-NEXT:  ; %bb.4: ; %cmp.true18233; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload18234; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload18235; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v4418236; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018237; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v018238; SI-NEXT:    s_add_i32 s28, s28, 318239; SI-NEXT:    s_and_b32 s4, s28, 0xff18240; SI-NEXT:    s_lshl_b32 s5, s29, 818241; SI-NEXT:    s_or_b32 s4, s5, s418242; SI-NEXT:    s_addk_i32 s4, 0x30018243; SI-NEXT:    s_and_b32 s4, s4, 0xffff18244; SI-NEXT:    s_add_i32 s16, s16, 318245; SI-NEXT:    s_lshl_b32 s5, s17, 818246; SI-NEXT:    s_add_i32 s18, s18, 318247; SI-NEXT:    s_and_b32 s6, s18, 0xff18248; SI-NEXT:    s_lshl_b32 s6, s6, 1618249; SI-NEXT:    s_add_i32 s20, s20, 318250; SI-NEXT:    s_add_i32 s22, s22, 318251; SI-NEXT:    s_and_b32 s7, s22, 0xff18252; SI-NEXT:    s_lshl_b32 s7, s7, 1618253; SI-NEXT:    s_add_i32 s24, s24, 318254; SI-NEXT:    s_add_i32 s26, s26, 318255; SI-NEXT:    s_and_b32 s8, s26, 0xff18256; SI-NEXT:    s_lshl_b32 s8, s8, 1618257; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload18258; SI-NEXT:    s_waitcnt vmcnt(2)18259; SI-NEXT:    v_or_b32_e32 v0, v1, v018260; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload18261; SI-NEXT:    v_or_b32_e32 v0, s4, v018262; SI-NEXT:    s_and_b32 s4, s16, 0xff18263; SI-NEXT:    s_or_b32 s4, s5, s418264; SI-NEXT:    s_addk_i32 s4, 0x30018265; SI-NEXT:    s_lshl_b32 s5, s19, 2418266; SI-NEXT:    s_and_b32 s4, s4, 0xffff18267; SI-NEXT:    s_or_b32 s5, s5, s618268; SI-NEXT:    s_or_b32 s4, s5, s418269; SI-NEXT:    s_and_b32 s5, s20, 0xff18270; SI-NEXT:    s_lshl_b32 s6, s21, 818271; SI-NEXT:    s_or_b32 s5, s6, s518272; SI-NEXT:    s_addk_i32 s5, 0x30018273; SI-NEXT:    s_lshl_b32 s6, s23, 2418274; SI-NEXT:    s_and_b32 s5, s5, 0xffff18275; SI-NEXT:    s_or_b32 s6, s6, s718276; SI-NEXT:    s_or_b32 s5, s6, s518277; SI-NEXT:    s_and_b32 s6, s24, 0xff18278; SI-NEXT:    s_lshl_b32 s7, s25, 818279; SI-NEXT:    s_or_b32 s6, s7, s618280; SI-NEXT:    s_addk_i32 s6, 0x30018281; SI-NEXT:    s_lshl_b32 s7, s27, 2418282; SI-NEXT:    s_and_b32 s6, s6, 0xffff18283; SI-NEXT:    s_or_b32 s7, s7, s818284; SI-NEXT:    s_or_b32 s6, s7, s618285; SI-NEXT:    s_add_i32 s4, s4, 0x300000018286; SI-NEXT:    s_add_i32 s5, s5, 0x300000018287; SI-NEXT:    s_add_i32 s6, s6, 0x300000018288; SI-NEXT:    s_waitcnt vmcnt(0)18289; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118290; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118291; SI-NEXT:    v_or_b32_e32 v1, v2, v118292; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload18293; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v118294; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v118295; SI-NEXT:    s_waitcnt vmcnt(0)18296; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v218297; SI-NEXT:    v_and_b32_e32 v2, 0xff, v218298; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v218299; SI-NEXT:    v_or_b32_e32 v2, v3, v218300; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x3000000, v018301; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload18302; SI-NEXT:    v_or_b32_e32 v1, v2, v118303; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x3000000, v118304; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload18305; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload18306; SI-NEXT:    s_waitcnt vmcnt(2)18307; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018308; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018309; SI-NEXT:    s_waitcnt vmcnt(1)18310; SI-NEXT:    v_or_b32_e32 v0, v1, v018311; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload18312; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018313; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018314; SI-NEXT:    s_waitcnt vmcnt(0)18315; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118316; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118317; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118318; SI-NEXT:    v_or_b32_e32 v1, v2, v118319; SI-NEXT:    v_or_b32_e32 v0, v1, v018320; SI-NEXT:    v_add_i32_e32 v5, vcc, 0x3000000, v018321; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload18322; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload18323; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload18324; SI-NEXT:    s_waitcnt vmcnt(2)18325; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018326; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018327; SI-NEXT:    s_waitcnt vmcnt(1)18328; SI-NEXT:    v_or_b32_e32 v0, v1, v018329; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload18330; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018331; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018332; SI-NEXT:    s_waitcnt vmcnt(0)18333; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118334; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118335; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118336; SI-NEXT:    v_or_b32_e32 v1, v2, v118337; SI-NEXT:    v_or_b32_e32 v0, v1, v018338; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x3000000, v018339; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload18340; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload18341; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload18342; SI-NEXT:    s_waitcnt vmcnt(2)18343; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018344; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018345; SI-NEXT:    s_waitcnt vmcnt(1)18346; SI-NEXT:    v_or_b32_e32 v0, v1, v018347; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload18348; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018349; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018350; SI-NEXT:    s_waitcnt vmcnt(0)18351; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118352; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118353; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118354; SI-NEXT:    v_or_b32_e32 v1, v2, v118355; SI-NEXT:    v_or_b32_e32 v0, v1, v018356; SI-NEXT:    v_add_i32_e32 v7, vcc, 0x3000000, v018357; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload18358; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload18359; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload18360; SI-NEXT:    s_waitcnt vmcnt(2)18361; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018362; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018363; SI-NEXT:    s_waitcnt vmcnt(1)18364; SI-NEXT:    v_or_b32_e32 v0, v1, v018365; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload18366; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018367; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018368; SI-NEXT:    s_waitcnt vmcnt(0)18369; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118370; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118371; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118372; SI-NEXT:    v_or_b32_e32 v1, v2, v118373; SI-NEXT:    v_or_b32_e32 v0, v1, v018374; SI-NEXT:    v_add_i32_e32 v8, vcc, 0x3000000, v018375; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload18376; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload18377; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload18378; SI-NEXT:    s_waitcnt vmcnt(2)18379; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018380; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018381; SI-NEXT:    s_waitcnt vmcnt(1)18382; SI-NEXT:    v_or_b32_e32 v0, v1, v018383; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload18384; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018385; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018386; SI-NEXT:    s_waitcnt vmcnt(0)18387; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118388; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118389; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118390; SI-NEXT:    v_or_b32_e32 v1, v2, v118391; SI-NEXT:    v_or_b32_e32 v0, v1, v018392; SI-NEXT:    v_add_i32_e32 v9, vcc, 0x3000000, v018393; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload18394; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload18395; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload18396; SI-NEXT:    s_waitcnt vmcnt(2)18397; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018398; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018399; SI-NEXT:    s_waitcnt vmcnt(1)18400; SI-NEXT:    v_or_b32_e32 v0, v1, v018401; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload18402; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018403; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018404; SI-NEXT:    s_waitcnt vmcnt(0)18405; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118406; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118407; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118408; SI-NEXT:    v_or_b32_e32 v1, v2, v118409; SI-NEXT:    v_or_b32_e32 v0, v1, v018410; SI-NEXT:    v_add_i32_e32 v10, vcc, 0x3000000, v018411; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload18412; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload18413; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload18414; SI-NEXT:    s_waitcnt vmcnt(2)18415; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018416; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018417; SI-NEXT:    s_waitcnt vmcnt(1)18418; SI-NEXT:    v_or_b32_e32 v0, v1, v018419; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload18420; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018421; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018422; SI-NEXT:    s_waitcnt vmcnt(0)18423; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118424; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118425; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118426; SI-NEXT:    v_or_b32_e32 v1, v2, v118427; SI-NEXT:    v_or_b32_e32 v0, v1, v018428; SI-NEXT:    s_waitcnt expcnt(0)18429; SI-NEXT:    v_add_i32_e32 v11, vcc, 0x3000000, v018430; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload18431; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload18432; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload18433; SI-NEXT:    s_waitcnt vmcnt(2)18434; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018435; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018436; SI-NEXT:    s_waitcnt vmcnt(1)18437; SI-NEXT:    v_or_b32_e32 v0, v1, v018438; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload18439; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018440; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018441; SI-NEXT:    s_waitcnt vmcnt(0)18442; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118443; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118444; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118445; SI-NEXT:    v_or_b32_e32 v1, v2, v118446; SI-NEXT:    v_or_b32_e32 v0, v1, v018447; SI-NEXT:    v_add_i32_e32 v12, vcc, 0x3000000, v018448; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload18449; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload18450; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload18451; SI-NEXT:    s_waitcnt vmcnt(2)18452; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018453; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018454; SI-NEXT:    s_waitcnt vmcnt(1)18455; SI-NEXT:    v_or_b32_e32 v0, v1, v018456; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload18457; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018458; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018459; SI-NEXT:    s_waitcnt vmcnt(0)18460; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118461; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118462; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118463; SI-NEXT:    v_or_b32_e32 v1, v2, v118464; SI-NEXT:    v_or_b32_e32 v0, v1, v018465; SI-NEXT:    v_add_i32_e32 v13, vcc, 0x3000000, v018466; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload18467; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload18468; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload18469; SI-NEXT:    s_waitcnt vmcnt(2)18470; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018471; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018472; SI-NEXT:    s_waitcnt vmcnt(1)18473; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118474; SI-NEXT:    v_or_b32_e32 v0, v63, v018475; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118476; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118477; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018478; SI-NEXT:    v_or_b32_e32 v1, v57, v118479; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018480; SI-NEXT:    v_or_b32_e32 v0, v1, v018481; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v018482; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload18483; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload18484; SI-NEXT:    s_waitcnt vmcnt(1)18485; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018486; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018487; SI-NEXT:    s_waitcnt vmcnt(0)18488; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118489; SI-NEXT:    v_or_b32_e32 v0, v33, v018490; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118491; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118492; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018493; SI-NEXT:    v_or_b32_e32 v1, v2, v118494; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018495; SI-NEXT:    v_or_b32_e32 v0, v1, v018496; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v018497; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload18498; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload18499; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload18500; SI-NEXT:    s_waitcnt vmcnt(2)18501; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018502; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018503; SI-NEXT:    s_waitcnt vmcnt(1)18504; SI-NEXT:    v_or_b32_e32 v0, v1, v018505; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload18506; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018507; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018508; SI-NEXT:    s_waitcnt vmcnt(0)18509; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118510; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118511; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118512; SI-NEXT:    v_or_b32_e32 v1, v50, v118513; SI-NEXT:    v_or_b32_e32 v0, v1, v018514; SI-NEXT:    v_add_i32_e32 v16, vcc, 0x3000000, v018515; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload18516; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload18517; SI-NEXT:    s_waitcnt vmcnt(1)18518; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018519; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018520; SI-NEXT:    s_waitcnt vmcnt(0)18521; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118522; SI-NEXT:    v_or_b32_e32 v0, v32, v018523; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118524; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118525; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018526; SI-NEXT:    v_or_b32_e32 v1, v2, v118527; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018528; SI-NEXT:    v_or_b32_e32 v0, v1, v018529; SI-NEXT:    v_add_i32_e32 v17, vcc, 0x3000000, v018530; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload18531; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload18532; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload18533; SI-NEXT:    s_waitcnt vmcnt(2)18534; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018535; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018536; SI-NEXT:    s_waitcnt vmcnt(1)18537; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118538; SI-NEXT:    v_or_b32_e32 v0, v40, v018539; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118540; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118541; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018542; SI-NEXT:    s_waitcnt vmcnt(0)18543; SI-NEXT:    v_or_b32_e32 v1, v2, v118544; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018545; SI-NEXT:    v_or_b32_e32 v0, v1, v018546; SI-NEXT:    v_add_i32_e32 v18, vcc, 0x3000000, v018547; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload18548; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload18549; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload18550; SI-NEXT:    s_waitcnt vmcnt(2)18551; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018552; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018553; SI-NEXT:    s_waitcnt vmcnt(1)18554; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118555; SI-NEXT:    v_or_b32_e32 v0, v35, v018556; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118557; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118558; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018559; SI-NEXT:    s_waitcnt vmcnt(0)18560; SI-NEXT:    v_or_b32_e32 v1, v2, v118561; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018562; SI-NEXT:    v_or_b32_e32 v0, v1, v018563; SI-NEXT:    v_add_i32_e32 v19, vcc, 0x3000000, v018564; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload18565; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload18566; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload18567; SI-NEXT:    s_waitcnt vmcnt(2)18568; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018569; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018570; SI-NEXT:    s_waitcnt vmcnt(1)18571; SI-NEXT:    v_or_b32_e32 v0, v1, v018572; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload18573; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018574; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018575; SI-NEXT:    s_waitcnt vmcnt(0)18576; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118577; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118578; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118579; SI-NEXT:    v_or_b32_e32 v1, v2, v118580; SI-NEXT:    v_or_b32_e32 v0, v1, v018581; SI-NEXT:    v_add_i32_e32 v20, vcc, 0x3000000, v018582; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload18583; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload18584; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload18585; SI-NEXT:    s_waitcnt vmcnt(2)18586; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018587; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018588; SI-NEXT:    s_waitcnt vmcnt(1)18589; SI-NEXT:    v_or_b32_e32 v0, v1, v018590; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload18591; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018592; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018593; SI-NEXT:    s_waitcnt vmcnt(0)18594; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118595; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118596; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118597; SI-NEXT:    v_or_b32_e32 v1, v2, v118598; SI-NEXT:    v_or_b32_e32 v0, v1, v018599; SI-NEXT:    v_add_i32_e32 v21, vcc, 0x3000000, v018600; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload18601; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload18602; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload18603; SI-NEXT:    s_waitcnt vmcnt(2)18604; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018605; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018606; SI-NEXT:    s_waitcnt vmcnt(1)18607; SI-NEXT:    v_or_b32_e32 v0, v1, v018608; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload18609; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018610; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018611; SI-NEXT:    s_waitcnt vmcnt(0)18612; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118613; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118614; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118615; SI-NEXT:    v_or_b32_e32 v1, v48, v118616; SI-NEXT:    v_or_b32_e32 v0, v1, v018617; SI-NEXT:    v_add_i32_e32 v22, vcc, 0x3000000, v018618; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v6118619; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018620; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v6018621; SI-NEXT:    v_or_b32_e32 v0, v34, v018622; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118623; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118624; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018625; SI-NEXT:    v_or_b32_e32 v1, v54, v118626; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018627; SI-NEXT:    v_or_b32_e32 v0, v1, v018628; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload18629; SI-NEXT:    v_add_i32_e32 v23, vcc, 0x3000000, v018630; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v5918631; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018632; SI-NEXT:    s_waitcnt vmcnt(0)18633; SI-NEXT:    v_or_b32_e32 v0, v1, v018634; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v5818635; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118636; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118637; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018638; SI-NEXT:    v_or_b32_e32 v1, v45, v118639; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018640; SI-NEXT:    v_or_b32_e32 v0, v1, v018641; SI-NEXT:    v_add_i32_e32 v24, vcc, 0x3000000, v018642; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload18643; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload18644; SI-NEXT:    s_waitcnt vmcnt(1)18645; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v018646; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018647; SI-NEXT:    s_waitcnt vmcnt(0)18648; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118649; SI-NEXT:    v_or_b32_e32 v0, v43, v018650; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118651; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018652; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118653; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018654; SI-NEXT:    v_or_b32_e32 v1, v2, v118655; SI-NEXT:    v_or_b32_e32 v0, v1, v018656; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload18657; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload18658; SI-NEXT:    v_add_i32_e32 v25, vcc, 0x3000000, v018659; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v5518660; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018661; SI-NEXT:    s_waitcnt vmcnt(1)18662; SI-NEXT:    v_or_b32_e32 v0, v1, v018663; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload18664; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018665; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018666; SI-NEXT:    s_waitcnt vmcnt(0)18667; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118668; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118669; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118670; SI-NEXT:    v_or_b32_e32 v1, v46, v118671; SI-NEXT:    v_or_b32_e32 v0, v1, v018672; SI-NEXT:    v_add_i32_e32 v26, vcc, 0x3000000, v018673; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v4218674; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018675; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v4118676; SI-NEXT:    v_or_b32_e32 v0, v37, v018677; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118678; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018679; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118680; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018681; SI-NEXT:    v_or_b32_e32 v1, v2, v118682; SI-NEXT:    v_or_b32_e32 v0, v1, v018683; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload18684; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload18685; SI-NEXT:    v_add_i32_e32 v27, vcc, 0x3000000, v018686; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v5318687; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018688; SI-NEXT:    s_waitcnt vmcnt(1)18689; SI-NEXT:    v_or_b32_e32 v0, v1, v018690; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v4718691; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118692; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118693; SI-NEXT:    s_waitcnt vmcnt(0)18694; SI-NEXT:    v_or_b32_e32 v1, v2, v118695; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload18696; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018697; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018698; SI-NEXT:    v_or_b32_e32 v0, v1, v018699; SI-NEXT:    v_add_i32_e32 v28, vcc, 0x3000000, v018700; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v6218701; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018702; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v5618703; SI-NEXT:    v_or_b32_e32 v0, v36, v018704; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118705; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018706; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118707; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018708; SI-NEXT:    s_waitcnt vmcnt(0)18709; SI-NEXT:    v_or_b32_e32 v1, v2, v118710; SI-NEXT:    v_or_b32_e32 v0, v1, v018711; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload18712; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload18713; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x3000000, v018714; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v4918715; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018716; SI-NEXT:    s_waitcnt vmcnt(1)18717; SI-NEXT:    v_or_b32_e32 v0, v1, v018718; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v3918719; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118720; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018721; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118722; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018723; SI-NEXT:    s_waitcnt vmcnt(0)18724; SI-NEXT:    v_or_b32_e32 v1, v2, v118725; SI-NEXT:    v_or_b32_e32 v0, v1, v018726; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload18727; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v018728; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v3818729; SI-NEXT:    v_and_b32_e32 v0, 0xff, v018730; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload18731; SI-NEXT:    s_waitcnt vmcnt(1)18732; SI-NEXT:    v_or_b32_e32 v0, v1, v018733; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload18734; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v018735; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v018736; SI-NEXT:    s_waitcnt vmcnt(0)18737; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v118738; SI-NEXT:    v_and_b32_e32 v1, 0xff, v118739; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v118740; SI-NEXT:    v_or_b32_e32 v1, v2, v118741; SI-NEXT:    v_or_b32_e32 v0, v1, v018742; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v018743; SI-NEXT:    v_mov_b32_e32 v0, s418744; SI-NEXT:    v_mov_b32_e32 v1, s518745; SI-NEXT:    v_mov_b32_e32 v2, s618746; SI-NEXT:  .LBB15_5: ; %end18747; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload18748; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload18749; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload18750; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload18751; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload18752; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload18753; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload18754; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload18755; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload18756; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload18757; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload18758; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload18759; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload18760; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload18761; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload18762; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload18763; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)18764; SI-NEXT:    s_setpc_b64 s[30:31]18765;18766; VI-LABEL: bitcast_v128i8_to_v32i32_scalar:18767; VI:       ; %bb.0:18768; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)18769; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill18770; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill18771; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill18772; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill18773; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill18774; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill18775; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill18776; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill18777; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill18778; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill18779; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill18780; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill18781; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill18782; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill18783; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill18784; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill18785; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill18786; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill18787; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill18788; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill18789; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill18790; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill18791; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill18792; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill18793; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill18794; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill18795; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill18796; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill18797; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill18798; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:33218799; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s3218800; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:818801; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:1618802; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:2418803; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:3218804; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:4018805; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:4818806; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:5618807; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:6418808; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:7218809; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:8018810; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:8818811; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:9618812; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:10418813; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:11218814; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:12018815; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:12818816; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:13618817; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:14418818; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:15218819; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:16018820; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:16818821; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:17618822; VI-NEXT:    v_lshlrev_b32_e32 v1, 8, v118823; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill18824; VI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1518825; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill18826; VI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1918827; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill18828; VI-NEXT:    v_lshlrev_b32_e32 v1, 8, v2118829; VI-NEXT:    v_lshlrev_b32_e32 v14, 8, v318830; VI-NEXT:    v_lshlrev_b32_e32 v16, 8, v518831; VI-NEXT:    v_lshlrev_b32_e32 v47, 8, v718832; VI-NEXT:    v_lshlrev_b32_e32 v46, 8, v918833; VI-NEXT:    v_lshlrev_b32_e32 v10, 8, v1118834; VI-NEXT:    v_lshlrev_b32_e32 v18, 8, v1318835; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill18836; VI-NEXT:    v_lshlrev_b32_e32 v8, 8, v1718837; VI-NEXT:    s_waitcnt vmcnt(14)18838; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v018839; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2318840; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill18841; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2518842; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill18843; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2718844; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill18845; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2918846; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill18847; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3418848; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill18849; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4418850; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill18851; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3218852; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill18853; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4318854; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill18855; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4218856; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill18857; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4118858; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill18859; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4018860; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill18861; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5518862; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill18863; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5418864; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill18865; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5318866; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill18867; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5218868; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill18869; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5118870; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill18871; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5018872; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill18873; VI-NEXT:    s_waitcnt vmcnt(14)18874; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4918875; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill18876; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3018877; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill18878; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3118879; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill18880; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3318881; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill18882; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3518883; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill18884; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3618885; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill18886; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3718887; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill18888; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3818889; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill18890; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3918891; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill18892; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4818893; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill18894; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:18418895; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:19218896; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:20018897; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:20818898; VI-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:21618899; VI-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:22418900; VI-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:23218901; VI-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:24018902; VI-NEXT:    s_and_b64 s[4:5], vcc, exec18903; VI-NEXT:    s_waitcnt vmcnt(7)18904; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v018905; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill18906; VI-NEXT:    s_waitcnt vmcnt(7)18907; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v118908; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill18909; VI-NEXT:    s_waitcnt vmcnt(7)18910; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1318911; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill18912; VI-NEXT:    s_waitcnt vmcnt(7)18913; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v318914; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill18915; VI-NEXT:    s_waitcnt vmcnt(7)18916; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1118917; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill18918; VI-NEXT:    s_waitcnt vmcnt(7)18919; VI-NEXT:    v_lshlrev_b32_e32 v38, 8, v518920; VI-NEXT:    s_waitcnt vmcnt(6)18921; VI-NEXT:    v_lshlrev_b32_e32 v25, 8, v918922; VI-NEXT:    s_waitcnt vmcnt(5)18923; VI-NEXT:    v_lshlrev_b32_e32 v39, 8, v718924; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:24818925; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:25618926; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:26418927; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:27218928; VI-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:28018929; VI-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:28818930; VI-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:29618931; VI-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:30418932; VI-NEXT:    s_waitcnt vmcnt(7)18933; VI-NEXT:    v_lshlrev_b32_e32 v33, 8, v018934; VI-NEXT:    s_waitcnt vmcnt(6)18935; VI-NEXT:    v_lshlrev_b32_e32 v48, 8, v118936; VI-NEXT:    s_waitcnt vmcnt(5)18937; VI-NEXT:    v_lshlrev_b32_e32 v27, 8, v1318938; VI-NEXT:    s_waitcnt vmcnt(4)18939; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v318940; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill18941; VI-NEXT:    s_waitcnt vmcnt(4)18942; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1118943; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill18944; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:31218945; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:32018946; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:32818947; VI-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:418948; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:1218949; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:2018950; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:2818951; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:3618952; VI-NEXT:    s_waitcnt vmcnt(12)18953; VI-NEXT:    v_lshlrev_b32_e32 v31, 8, v518954; VI-NEXT:    s_waitcnt vmcnt(11)18955; VI-NEXT:    v_lshlrev_b32_e32 v29, 8, v918956; VI-NEXT:    s_waitcnt vmcnt(10)18957; VI-NEXT:    v_lshlrev_b32_e32 v30, 8, v718958; VI-NEXT:    s_waitcnt vmcnt(7)18959; VI-NEXT:    v_lshlrev_b32_e32 v37, 8, v018960; VI-NEXT:    s_waitcnt vmcnt(6)18961; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v118962; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill18963; VI-NEXT:    s_waitcnt vmcnt(6)18964; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v318965; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill18966; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4418967; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:5218968; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:6018969; VI-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:6818970; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:7618971; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:8418972; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:9218973; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:10018974; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:10818975; VI-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:11618976; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:12418977; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:13218978; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:14018979; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:14818980; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:15618981; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:16418982; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:17218983; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:18018984; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:18818985; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:19618986; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:20418987; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:21218988; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:22018989; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:22818990; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:23618991; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:24418992; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:25218993; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:26018994; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:26818995; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:27618996; VI-NEXT:    s_waitcnt vmcnt(0)18997; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill18998; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:28418999; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:29219000; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:30019001; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:30819002; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:31619003; VI-NEXT:    s_waitcnt vmcnt(0)19004; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill19005; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:32419006; VI-NEXT:    s_waitcnt vmcnt(0)19007; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill19008; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill19009; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill19010; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill19011; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill19012; VI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill19013; VI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill19014; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill19015; VI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill19016; VI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill19017; VI-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill19018; VI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill19019; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill19020; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill19021; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill19022; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill19023; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill19024; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill19025; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill19026; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill19027; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill19028; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill19029; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill19030; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill19031; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill19032; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill19033; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill19034; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill19035; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill19036; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill19037; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill19038; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill19039; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill19040; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill19041; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill19042; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill19043; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill19044; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill19045; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill19046; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill19047; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill19048; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill19049; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill19050; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill19051; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill19052; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill19053; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill19054; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill19055; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill19056; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill19057; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill19058; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill19059; VI-NEXT:    s_cbranch_scc0 .LBB15_219060; VI-NEXT:  ; %bb.1: ; %cmp.false19061; VI-NEXT:    v_or_b32_sdwa v0, v2, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19062; VI-NEXT:    v_or_b32_sdwa v1, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19063; VI-NEXT:    v_or_b32_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19064; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload19065; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload19066; VI-NEXT:    v_or_b32_sdwa v2, v6, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19067; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload19068; VI-NEXT:    s_and_b32 s4, s28, 0xff19069; VI-NEXT:    s_lshl_b32 s5, s29, 819070; VI-NEXT:    s_or_b32 s4, s4, s519071; VI-NEXT:    s_and_b32 s4, s4, 0xffff19072; VI-NEXT:    s_lshl_b32 s5, s17, 819073; VI-NEXT:    s_lshl_b32 s6, s19, 819074; VI-NEXT:    s_lshl_b32 s7, s23, 819075; VI-NEXT:    s_lshl_b32 s8, s27, 819076; VI-NEXT:    s_waitcnt vmcnt(2)19077; VI-NEXT:    v_or_b32_sdwa v0, v0, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19078; VI-NEXT:    s_waitcnt vmcnt(1)19079; VI-NEXT:    v_or_b32_sdwa v1, v1, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19080; VI-NEXT:    v_or_b32_sdwa v6, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19081; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload19082; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload19083; VI-NEXT:    s_waitcnt vmcnt(0)19084; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19085; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload19086; VI-NEXT:    v_or_b32_sdwa v3, v3, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19087; VI-NEXT:    v_or_b32_sdwa v5, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19088; VI-NEXT:    s_waitcnt vmcnt(0)19089; VI-NEXT:    v_or_b32_sdwa v1, v1, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19090; VI-NEXT:    v_or_b32_sdwa v7, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19091; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload19092; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload19093; VI-NEXT:    s_waitcnt vmcnt(0)19094; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19095; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload19096; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload19097; VI-NEXT:    s_waitcnt vmcnt(0)19098; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19099; VI-NEXT:    v_or_b32_sdwa v8, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19100; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload19101; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload19102; VI-NEXT:    s_waitcnt vmcnt(0)19103; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19104; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload19105; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload19106; VI-NEXT:    s_waitcnt vmcnt(0)19107; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19108; VI-NEXT:    v_or_b32_sdwa v9, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19109; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload19110; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload19111; VI-NEXT:    s_waitcnt vmcnt(0)19112; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19113; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload19114; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload19115; VI-NEXT:    s_waitcnt vmcnt(0)19116; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19117; VI-NEXT:    v_or_b32_sdwa v10, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19118; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload19119; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload19120; VI-NEXT:    s_waitcnt vmcnt(0)19121; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19122; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload19123; VI-NEXT:    s_waitcnt vmcnt(0)19124; VI-NEXT:    v_or_b32_sdwa v1, v11, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19125; VI-NEXT:    v_or_b32_sdwa v11, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19126; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload19127; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload19128; VI-NEXT:    s_waitcnt vmcnt(1)19129; VI-NEXT:    v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19130; VI-NEXT:    s_waitcnt vmcnt(0)19131; VI-NEXT:    v_or_b32_sdwa v1, v12, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19132; VI-NEXT:    v_or_b32_sdwa v12, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19133; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload19134; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload19135; VI-NEXT:    s_waitcnt vmcnt(1)19136; VI-NEXT:    v_or_b32_sdwa v0, v20, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19137; VI-NEXT:    s_waitcnt vmcnt(0)19138; VI-NEXT:    v_or_b32_sdwa v1, v13, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19139; VI-NEXT:    v_or_b32_sdwa v13, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19140; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload19141; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload19142; VI-NEXT:    s_waitcnt vmcnt(1)19143; VI-NEXT:    v_or_b32_sdwa v0, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19144; VI-NEXT:    s_waitcnt vmcnt(0)19145; VI-NEXT:    v_mov_b32_e32 v47, v119146; VI-NEXT:    v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19147; VI-NEXT:    v_or_b32_sdwa v14, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19148; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload19149; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload19150; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload19151; VI-NEXT:    s_waitcnt vmcnt(2)19152; VI-NEXT:    v_or_b32_sdwa v0, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19153; VI-NEXT:    s_waitcnt vmcnt(1)19154; VI-NEXT:    v_or_b32_sdwa v1, v15, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19155; VI-NEXT:    v_or_b32_sdwa v15, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19156; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload19157; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload19158; VI-NEXT:    s_waitcnt vmcnt(1)19159; VI-NEXT:    v_mov_b32_e32 v46, v019160; VI-NEXT:    v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19161; VI-NEXT:    s_waitcnt vmcnt(0)19162; VI-NEXT:    v_or_b32_sdwa v1, v23, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19163; VI-NEXT:    v_or_b32_sdwa v16, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19164; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload19165; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload19166; VI-NEXT:    s_waitcnt vmcnt(1)19167; VI-NEXT:    v_mov_b32_e32 v62, v019168; VI-NEXT:    v_or_b32_sdwa v0, v26, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19169; VI-NEXT:    s_waitcnt vmcnt(0)19170; VI-NEXT:    v_or_b32_sdwa v1, v17, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19171; VI-NEXT:    v_or_b32_sdwa v17, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19172; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload19173; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload19174; VI-NEXT:    s_waitcnt vmcnt(1)19175; VI-NEXT:    v_or_b32_sdwa v0, v24, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19176; VI-NEXT:    s_waitcnt vmcnt(0)19177; VI-NEXT:    v_mov_b32_e32 v63, v119178; VI-NEXT:    v_or_b32_sdwa v1, v21, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19179; VI-NEXT:    v_or_b32_sdwa v18, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19180; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload19181; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload19182; VI-NEXT:    s_waitcnt vmcnt(1)19183; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19184; VI-NEXT:    s_waitcnt vmcnt(0)19185; VI-NEXT:    v_or_b32_sdwa v1, v19, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19186; VI-NEXT:    v_or_b32_sdwa v19, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19187; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload19188; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload19189; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload19190; VI-NEXT:    s_waitcnt vmcnt(2)19191; VI-NEXT:    v_mov_b32_e32 v60, v019192; VI-NEXT:    v_or_b32_sdwa v0, v28, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19193; VI-NEXT:    s_waitcnt vmcnt(1)19194; VI-NEXT:    v_or_b32_sdwa v1, v35, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19195; VI-NEXT:    v_or_b32_sdwa v20, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19196; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload19197; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload19198; VI-NEXT:    s_waitcnt vmcnt(1)19199; VI-NEXT:    v_mov_b32_e32 v35, v019200; VI-NEXT:    v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19201; VI-NEXT:    s_waitcnt vmcnt(0)19202; VI-NEXT:    v_or_b32_sdwa v1, v58, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19203; VI-NEXT:    v_or_b32_sdwa v21, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19204; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload19205; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload19206; VI-NEXT:    v_or_b32_sdwa v1, v34, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19207; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload19208; VI-NEXT:    s_waitcnt vmcnt(2)19209; VI-NEXT:    v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19210; VI-NEXT:    v_or_b32_sdwa v22, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19211; VI-NEXT:    v_or_b32_sdwa v0, v59, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19212; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload19213; VI-NEXT:    s_waitcnt vmcnt(1)19214; VI-NEXT:    v_or_b32_sdwa v1, v32, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19215; VI-NEXT:    v_mov_b32_e32 v32, v6119216; VI-NEXT:    v_or_b32_sdwa v23, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19217; VI-NEXT:    v_or_b32_sdwa v1, v55, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19218; VI-NEXT:    v_mov_b32_e32 v55, v4319219; VI-NEXT:    s_waitcnt vmcnt(0)19220; VI-NEXT:    v_or_b32_sdwa v0, v42, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19221; VI-NEXT:    v_or_b32_sdwa v24, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19222; VI-NEXT:    v_or_b32_sdwa v0, v54, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19223; VI-NEXT:    v_or_b32_sdwa v1, v53, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19224; VI-NEXT:    v_or_b32_sdwa v25, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19225; VI-NEXT:    v_or_b32_sdwa v0, v41, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19226; VI-NEXT:    v_or_b32_sdwa v1, v44, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19227; VI-NEXT:    v_mov_b32_e32 v41, v3319228; VI-NEXT:    v_or_b32_sdwa v26, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19229; VI-NEXT:    v_mov_b32_e32 v44, v5619230; VI-NEXT:    v_or_b32_sdwa v0, v56, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19231; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload19232; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload19233; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload19234; VI-NEXT:    v_mov_b32_e32 v38, v3919235; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload19236; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload19237; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload19238; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload19239; VI-NEXT:    v_or_b32_sdwa v1, v52, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19240; VI-NEXT:    v_or_b32_sdwa v27, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19241; VI-NEXT:    v_mov_b32_e32 v54, v5319242; VI-NEXT:    v_mov_b32_e32 v52, v3619243; VI-NEXT:    s_waitcnt vmcnt(6)19244; VI-NEXT:    v_or_b32_sdwa v0, v36, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19245; VI-NEXT:    s_waitcnt vmcnt(4)19246; VI-NEXT:    v_or_b32_sdwa v1, v33, v56 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19247; VI-NEXT:    v_or_b32_sdwa v28, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19248; VI-NEXT:    v_or_b32_sdwa v0, v50, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19249; VI-NEXT:    v_or_b32_sdwa v1, v40, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19250; VI-NEXT:    v_or_b32_sdwa v29, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19251; VI-NEXT:    v_or_b32_sdwa v0, v49, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19252; VI-NEXT:    v_or_b32_sdwa v1, v51, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19253; VI-NEXT:    v_or_b32_sdwa v30, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19254; VI-NEXT:    s_waitcnt vmcnt(2)19255; VI-NEXT:    v_or_b32_sdwa v0, v48, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19256; VI-NEXT:    s_waitcnt vmcnt(0)19257; VI-NEXT:    v_or_b32_sdwa v1, v39, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19258; VI-NEXT:    v_or_b32_sdwa v31, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD19259; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload19260; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload19261; VI-NEXT:    v_mov_b32_e32 v50, v4019262; VI-NEXT:    v_mov_b32_e32 v49, v5119263; VI-NEXT:    v_mov_b32_e32 v40, v3419264; VI-NEXT:    s_waitcnt vmcnt(0)19265; VI-NEXT:    v_or_b32_sdwa v0, v37, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD19266; VI-NEXT:    v_or_b32_e32 v3, s4, v019267; VI-NEXT:    s_and_b32 s4, s16, 0xff19268; VI-NEXT:    s_or_b32 s4, s4, s519269; VI-NEXT:    s_and_b32 s5, s18, 0xff19270; VI-NEXT:    s_or_b32 s5, s5, s619271; VI-NEXT:    s_and_b32 s4, s4, 0xffff19272; VI-NEXT:    s_lshl_b32 s5, s5, 1619273; VI-NEXT:    s_or_b32 s4, s4, s519274; VI-NEXT:    s_and_b32 s5, s20, 0xff19275; VI-NEXT:    s_lshl_b32 s6, s21, 819276; VI-NEXT:    s_or_b32 s5, s5, s619277; VI-NEXT:    s_and_b32 s6, s22, 0xff19278; VI-NEXT:    s_or_b32 s6, s6, s719279; VI-NEXT:    s_and_b32 s5, s5, 0xffff19280; VI-NEXT:    s_lshl_b32 s6, s6, 1619281; VI-NEXT:    s_or_b32 s5, s5, s619282; VI-NEXT:    s_and_b32 s6, s24, 0xff19283; VI-NEXT:    s_lshl_b32 s7, s25, 819284; VI-NEXT:    s_or_b32 s6, s6, s719285; VI-NEXT:    s_and_b32 s7, s26, 0xff19286; VI-NEXT:    s_or_b32 s7, s7, s819287; VI-NEXT:    s_and_b32 s6, s6, 0xffff19288; VI-NEXT:    s_lshl_b32 s7, s7, 1619289; VI-NEXT:    s_or_b32 s6, s6, s719290; VI-NEXT:    v_mov_b32_e32 v0, s419291; VI-NEXT:    v_mov_b32_e32 v1, s519292; VI-NEXT:    v_mov_b32_e32 v2, s619293; VI-NEXT:    s_mov_b64 s[4:5], 019294; VI-NEXT:    s_branch .LBB15_319295; VI-NEXT:  .LBB15_2:19296; VI-NEXT:    v_mov_b32_e32 v44, v5619297; VI-NEXT:    v_mov_b32_e32 v41, v3319298; VI-NEXT:    v_mov_b32_e32 v50, v4019299; VI-NEXT:    v_mov_b32_e32 v38, v3919300; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload19301; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload19302; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload19303; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload19304; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload19305; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload19306; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload19307; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload19308; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload19309; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload19310; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload19311; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload19312; VI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload19313; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload19314; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload19315; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload19316; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload19317; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload19318; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload19319; VI-NEXT:    v_mov_b32_e32 v54, v5319320; VI-NEXT:    v_mov_b32_e32 v52, v3619321; VI-NEXT:    v_mov_b32_e32 v49, v5119322; VI-NEXT:    s_mov_b64 s[4:5], -119323; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr3119324; VI-NEXT:  .LBB15_3: ; %Flow19325; VI-NEXT:    v_mov_b32_e32 v51, v4119326; VI-NEXT:    v_mov_b32_e32 v36, v4419327; VI-NEXT:    v_mov_b32_e32 v53, v5419328; VI-NEXT:    s_waitcnt vmcnt(1)19329; VI-NEXT:    v_mov_b32_e32 v54, v6019330; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload19331; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload19332; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload19333; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload19334; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload19335; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload19336; VI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]19337; VI-NEXT:    s_cbranch_vccnz .LBB15_519338; VI-NEXT:  ; %bb.4: ; %cmp.true19339; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload19340; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload19341; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v3719342; VI-NEXT:    s_add_i32 s28, s28, 319343; VI-NEXT:    s_and_b32 s4, s28, 0xff19344; VI-NEXT:    s_lshl_b32 s5, s29, 819345; VI-NEXT:    s_or_b32 s4, s5, s419346; VI-NEXT:    s_addk_i32 s4, 0x30019347; VI-NEXT:    s_and_b32 s4, s4, 0xffff19348; VI-NEXT:    s_add_i32 s16, s16, 319349; VI-NEXT:    s_lshl_b32 s5, s17, 819350; VI-NEXT:    s_add_i32 s18, s18, 319351; VI-NEXT:    s_lshl_b32 s6, s19, 819352; VI-NEXT:    s_add_i32 s20, s20, 319353; VI-NEXT:    s_add_i32 s22, s22, 319354; VI-NEXT:    s_lshl_b32 s7, s23, 819355; VI-NEXT:    s_add_i32 s24, s24, 319356; VI-NEXT:    s_add_i32 s26, s26, 319357; VI-NEXT:    s_lshl_b32 s8, s27, 819358; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload19359; VI-NEXT:    s_waitcnt vmcnt(2)19360; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019361; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload19362; VI-NEXT:    v_or_b32_e32 v0, s4, v019363; VI-NEXT:    s_and_b32 s4, s16, 0xff19364; VI-NEXT:    s_or_b32 s4, s5, s419365; VI-NEXT:    s_and_b32 s5, s18, 0xff19366; VI-NEXT:    s_addk_i32 s4, 0x30019367; VI-NEXT:    s_or_b32 s5, s6, s519368; VI-NEXT:    s_and_b32 s4, s4, 0xffff19369; VI-NEXT:    s_lshl_b32 s5, s5, 1619370; VI-NEXT:    s_or_b32 s4, s5, s419371; VI-NEXT:    s_and_b32 s5, s20, 0xff19372; VI-NEXT:    s_lshl_b32 s6, s21, 819373; VI-NEXT:    s_or_b32 s5, s6, s519374; VI-NEXT:    s_and_b32 s6, s22, 0xff19375; VI-NEXT:    s_addk_i32 s5, 0x30019376; VI-NEXT:    s_or_b32 s6, s7, s619377; VI-NEXT:    s_and_b32 s5, s5, 0xffff19378; VI-NEXT:    s_lshl_b32 s6, s6, 1619379; VI-NEXT:    s_or_b32 s5, s6, s519380; VI-NEXT:    s_and_b32 s6, s24, 0xff19381; VI-NEXT:    s_lshl_b32 s7, s25, 819382; VI-NEXT:    s_or_b32 s6, s7, s619383; VI-NEXT:    s_and_b32 s7, s26, 0xff19384; VI-NEXT:    s_addk_i32 s6, 0x30019385; VI-NEXT:    s_or_b32 s7, s8, s719386; VI-NEXT:    s_and_b32 s6, s6, 0xffff19387; VI-NEXT:    s_lshl_b32 s7, s7, 1619388; VI-NEXT:    s_or_b32 s6, s7, s619389; VI-NEXT:    s_add_i32 s4, s4, 0x300000019390; VI-NEXT:    s_add_i32 s5, s5, 0x300000019391; VI-NEXT:    s_add_i32 s6, s6, 0x300000019392; VI-NEXT:    s_waitcnt vmcnt(0)19393; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v119394; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019395; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload19396; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x300, v119397; VI-NEXT:    s_waitcnt vmcnt(0)19398; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v219399; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019400; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_019401; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x3000000, v019402; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload19403; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x3000000, v119404; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload19405; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload19406; VI-NEXT:    s_waitcnt vmcnt(2)19407; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v019408; VI-NEXT:    s_waitcnt vmcnt(1)19409; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019410; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload19411; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v019412; VI-NEXT:    s_waitcnt vmcnt(0)19413; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v119414; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019415; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_019416; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x3000000, v019417; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload19418; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload19419; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload19420; VI-NEXT:    s_waitcnt vmcnt(2)19421; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v019422; VI-NEXT:    s_waitcnt vmcnt(1)19423; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019424; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload19425; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v019426; VI-NEXT:    s_waitcnt vmcnt(0)19427; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v119428; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019429; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_019430; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x3000000, v019431; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload19432; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload19433; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload19434; VI-NEXT:    s_waitcnt vmcnt(2)19435; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v019436; VI-NEXT:    s_waitcnt vmcnt(1)19437; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019438; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload19439; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v019440; VI-NEXT:    s_waitcnt vmcnt(0)19441; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v119442; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019443; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_019444; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x3000000, v019445; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload19446; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload19447; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload19448; VI-NEXT:    s_waitcnt vmcnt(2)19449; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v019450; VI-NEXT:    s_waitcnt vmcnt(1)19451; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_019452; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload19453; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v019454; VI-NEXT:    s_waitcnt vmcnt(0)19455; VI-NEXT:    v_add_u32

File truncated to the first 1 MiB.