brintos

brintos / llvm-project-archived public Read only

0
0
Text · 147.0 KiB · b761f68 Raw
2777 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 42; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -o - %s | FileCheck -check-prefixes=GFX9,GFX9-SDAG %s3; RUN: llc -O0 -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -o - %s | FileCheck -check-prefixes=GFX9-O0,GFX9-SDAG-O0 %s4 5; FIXME: GlobalISel missing the power-of-2 cases in legalization. https://github.com/llvm/llvm-project/issues/806716; xUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -o - %s | FileCheck -check-prefixes=GFX9,GFX9 %s7; xUN: llc -O0 -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -o - %s | FileCheck -check-prefixes=GFX9-O0,GFX9-O0 %s}}8 9define i128 @v_srem_i128_vv(i128 %lhs, i128 %rhs) {10; GFX9-LABEL: v_srem_i128_vv:11; GFX9:       ; %bb.0: ; %_udiv-special-cases12; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)13; GFX9-NEXT:    v_sub_co_u32_e32 v8, vcc, 0, v014; GFX9-NEXT:    v_subb_co_u32_e32 v9, vcc, 0, v1, vcc15; GFX9-NEXT:    v_subb_co_u32_e32 v10, vcc, 0, v2, vcc16; GFX9-NEXT:    v_subb_co_u32_e32 v11, vcc, 0, v3, vcc17; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, 0, v[2:3]18; GFX9-NEXT:    v_ashrrev_i32_e32 v20, 31, v319; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc20; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc21; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc22; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc23; GFX9-NEXT:    v_sub_co_u32_e32 v8, vcc, 0, v424; GFX9-NEXT:    v_subb_co_u32_e32 v9, vcc, 0, v5, vcc25; GFX9-NEXT:    v_subb_co_u32_e32 v10, vcc, 0, v6, vcc26; GFX9-NEXT:    v_subb_co_u32_e32 v11, vcc, 0, v7, vcc27; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, 0, v[6:7]28; GFX9-NEXT:    v_mov_b32_e32 v21, v2029; GFX9-NEXT:    v_cndmask_b32_e32 v22, v5, v9, vcc30; GFX9-NEXT:    v_cndmask_b32_e32 v23, v4, v8, vcc31; GFX9-NEXT:    v_cndmask_b32_e32 v5, v7, v11, vcc32; GFX9-NEXT:    v_cndmask_b32_e32 v4, v6, v10, vcc33; GFX9-NEXT:    v_or_b32_e32 v7, v22, v534; GFX9-NEXT:    v_or_b32_e32 v6, v23, v435; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[6:7]36; GFX9-NEXT:    v_or_b32_e32 v7, v1, v337; GFX9-NEXT:    v_or_b32_e32 v6, v0, v238; GFX9-NEXT:    v_cmp_eq_u64_e64 s[4:5], 0, v[6:7]39; GFX9-NEXT:    v_ffbh_u32_e32 v6, v440; GFX9-NEXT:    v_add_u32_e32 v6, 32, v641; GFX9-NEXT:    v_ffbh_u32_e32 v7, v542; GFX9-NEXT:    v_min_u32_e32 v6, v6, v743; GFX9-NEXT:    v_ffbh_u32_e32 v7, v2344; GFX9-NEXT:    v_add_u32_e32 v7, 32, v745; GFX9-NEXT:    v_ffbh_u32_e32 v8, v2246; GFX9-NEXT:    v_min_u32_e32 v7, v7, v847; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]48; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, 64, v749; GFX9-NEXT:    v_addc_co_u32_e64 v8, s[6:7], 0, 0, vcc50; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[4:5]51; GFX9-NEXT:    v_ffbh_u32_e32 v10, v352; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v6, vcc53; GFX9-NEXT:    v_ffbh_u32_e32 v7, v254; GFX9-NEXT:    v_add_u32_e32 v7, 32, v755; GFX9-NEXT:    v_min_u32_e32 v7, v7, v1056; GFX9-NEXT:    v_ffbh_u32_e32 v10, v057; GFX9-NEXT:    v_add_u32_e32 v10, 32, v1058; GFX9-NEXT:    v_ffbh_u32_e32 v11, v159; GFX9-NEXT:    v_min_u32_e32 v10, v10, v1160; GFX9-NEXT:    v_cndmask_b32_e64 v8, v8, 0, vcc61; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, 64, v1062; GFX9-NEXT:    v_addc_co_u32_e64 v11, s[6:7], 0, 0, vcc63; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[2:3]64; GFX9-NEXT:    v_mov_b32_e32 v9, 065; GFX9-NEXT:    v_cndmask_b32_e32 v7, v10, v7, vcc66; GFX9-NEXT:    v_cndmask_b32_e64 v11, v11, 0, vcc67; GFX9-NEXT:    v_sub_co_u32_e32 v6, vcc, v6, v768; GFX9-NEXT:    v_subb_co_u32_e32 v7, vcc, v8, v11, vcc69; GFX9-NEXT:    v_subb_co_u32_e32 v8, vcc, 0, v9, vcc70; GFX9-NEXT:    v_subb_co_u32_e32 v9, vcc, 0, v9, vcc71; GFX9-NEXT:    s_mov_b64 s[6:7], 0x7f72; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[6:7]73; GFX9-NEXT:    v_or_b32_e32 v12, v7, v974; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc75; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[8:9]76; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc77; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[8:9]78; GFX9-NEXT:    v_cndmask_b32_e32 v10, v11, v10, vcc79; GFX9-NEXT:    v_and_b32_e32 v10, 1, v1080; GFX9-NEXT:    v_xor_b32_e32 v11, 0x7f, v681; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v1082; GFX9-NEXT:    v_or_b32_e32 v11, v11, v883; GFX9-NEXT:    s_or_b64 s[4:5], s[4:5], vcc84; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[11:12]85; GFX9-NEXT:    s_xor_b64 s[6:7], s[4:5], -186; GFX9-NEXT:    v_cndmask_b32_e64 v10, v3, 0, s[4:5]87; GFX9-NEXT:    v_cndmask_b32_e64 v12, v2, 0, s[4:5]88; GFX9-NEXT:    v_cndmask_b32_e64 v11, v1, 0, s[4:5]89; GFX9-NEXT:    v_cndmask_b32_e64 v13, v0, 0, s[4:5]90; GFX9-NEXT:    s_and_b64 s[4:5], s[6:7], vcc91; GFX9-NEXT:    s_and_saveexec_b64 s[8:9], s[4:5]92; GFX9-NEXT:    s_cbranch_execz .LBB0_693; GFX9-NEXT:  ; %bb.1: ; %udiv-bb194; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, 1, v695; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v7, vcc96; GFX9-NEXT:    v_addc_co_u32_e32 v26, vcc, 0, v8, vcc97; GFX9-NEXT:    v_sub_u32_e32 v13, 0x7f, v698; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, 0, v9, vcc99; GFX9-NEXT:    v_sub_u32_e32 v11, 64, v13100; GFX9-NEXT:    v_or_b32_e32 v8, v25, v27101; GFX9-NEXT:    v_or_b32_e32 v7, v24, v26102; GFX9-NEXT:    v_lshlrev_b64 v[9:10], v13, v[2:3]103; GFX9-NEXT:    v_lshrrev_b64 v[11:12], v11, v[0:1]104; GFX9-NEXT:    v_sub_u32_e32 v6, 63, v6105; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[7:8]106; GFX9-NEXT:    v_lshlrev_b64 v[6:7], v6, v[0:1]107; GFX9-NEXT:    v_or_b32_e32 v8, v10, v12108; GFX9-NEXT:    v_or_b32_e32 v9, v9, v11109; GFX9-NEXT:    v_cmp_gt_u32_e64 s[4:5], 64, v13110; GFX9-NEXT:    v_lshlrev_b64 v[10:11], v13, v[0:1]111; GFX9-NEXT:    v_cndmask_b32_e64 v7, v7, v8, s[4:5]112; GFX9-NEXT:    v_cmp_eq_u32_e64 s[6:7], 0, v13113; GFX9-NEXT:    v_cndmask_b32_e64 v6, v6, v9, s[4:5]114; GFX9-NEXT:    v_mov_b32_e32 v8, 0115; GFX9-NEXT:    v_mov_b32_e32 v12, 0116; GFX9-NEXT:    v_cndmask_b32_e64 v7, v7, v3, s[6:7]117; GFX9-NEXT:    v_cndmask_b32_e64 v6, v6, v2, s[6:7]118; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, v11, s[4:5]119; GFX9-NEXT:    v_mov_b32_e32 v9, 0120; GFX9-NEXT:    v_mov_b32_e32 v13, 0121; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, v10, s[4:5]122; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc123; GFX9-NEXT:    s_xor_b64 s[6:7], exec, s[4:5]124; GFX9-NEXT:    s_cbranch_execz .LBB0_5125; GFX9-NEXT:  ; %bb.2: ; %udiv-preheader126; GFX9-NEXT:    v_sub_u32_e32 v12, 64, v24127; GFX9-NEXT:    v_lshrrev_b64 v[8:9], v24, v[0:1]128; GFX9-NEXT:    v_lshlrev_b64 v[12:13], v12, v[2:3]129; GFX9-NEXT:    v_cmp_gt_u32_e32 vcc, 64, v24130; GFX9-NEXT:    v_or_b32_e32 v12, v8, v12131; GFX9-NEXT:    v_subrev_u32_e32 v8, 64, v24132; GFX9-NEXT:    v_or_b32_e32 v13, v9, v13133; GFX9-NEXT:    v_lshrrev_b64 v[8:9], v8, v[2:3]134; GFX9-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v24135; GFX9-NEXT:    v_cndmask_b32_e32 v9, v9, v13, vcc136; GFX9-NEXT:    v_cndmask_b32_e64 v15, v9, v1, s[4:5]137; GFX9-NEXT:    v_cndmask_b32_e32 v12, v8, v12, vcc138; GFX9-NEXT:    v_lshrrev_b64 v[8:9], v24, v[2:3]139; GFX9-NEXT:    v_cndmask_b32_e64 v14, v12, v0, s[4:5]140; GFX9-NEXT:    v_cndmask_b32_e32 v17, 0, v9, vcc141; GFX9-NEXT:    v_cndmask_b32_e32 v16, 0, v8, vcc142; GFX9-NEXT:    v_add_co_u32_e32 v28, vcc, -1, v23143; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, -1, v22, vcc144; GFX9-NEXT:    v_addc_co_u32_e32 v30, vcc, -1, v4, vcc145; GFX9-NEXT:    v_mov_b32_e32 v12, 0146; GFX9-NEXT:    v_mov_b32_e32 v18, 0147; GFX9-NEXT:    v_addc_co_u32_e32 v31, vcc, -1, v5, vcc148; GFX9-NEXT:    v_mov_b32_e32 v13, 0149; GFX9-NEXT:    s_mov_b64 s[4:5], 0150; GFX9-NEXT:    v_mov_b32_e32 v19, 0151; GFX9-NEXT:    v_mov_b32_e32 v9, 0152; GFX9-NEXT:  .LBB0_3: ; %udiv-do-while153; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1154; GFX9-NEXT:    v_lshrrev_b32_e32 v32, 31, v15155; GFX9-NEXT:    v_lshlrev_b64 v[14:15], 1, v[14:15]156; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 31, v7157; GFX9-NEXT:    v_lshlrev_b64 v[6:7], 1, v[6:7]158; GFX9-NEXT:    v_lshrrev_b32_e32 v8, 31, v11159; GFX9-NEXT:    v_lshlrev_b64 v[16:17], 1, v[16:17]160; GFX9-NEXT:    v_or_b32_e32 v14, v14, v33161; GFX9-NEXT:    v_or3_b32 v6, v6, v8, v12162; GFX9-NEXT:    v_sub_co_u32_e32 v8, vcc, v28, v14163; GFX9-NEXT:    v_or_b32_e32 v16, v16, v32164; GFX9-NEXT:    v_subb_co_u32_e32 v8, vcc, v29, v15, vcc165; GFX9-NEXT:    v_subb_co_u32_e32 v8, vcc, v30, v16, vcc166; GFX9-NEXT:    v_lshlrev_b64 v[10:11], 1, v[10:11]167; GFX9-NEXT:    v_subb_co_u32_e32 v8, vcc, v31, v17, vcc168; GFX9-NEXT:    v_ashrrev_i32_e32 v8, 31, v8169; GFX9-NEXT:    v_or_b32_e32 v10, v18, v10170; GFX9-NEXT:    v_and_b32_e32 v18, v8, v23171; GFX9-NEXT:    v_or_b32_e32 v11, v19, v11172; GFX9-NEXT:    v_and_b32_e32 v19, v8, v22173; GFX9-NEXT:    v_sub_co_u32_e32 v14, vcc, v14, v18174; GFX9-NEXT:    v_and_b32_e32 v32, v8, v4175; GFX9-NEXT:    v_subb_co_u32_e32 v15, vcc, v15, v19, vcc176; GFX9-NEXT:    v_and_b32_e32 v33, v8, v5177; GFX9-NEXT:    v_subb_co_u32_e32 v16, vcc, v16, v32, vcc178; GFX9-NEXT:    v_subb_co_u32_e32 v17, vcc, v17, v33, vcc179; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, -1, v24180; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, -1, v25, vcc181; GFX9-NEXT:    v_addc_co_u32_e32 v26, vcc, -1, v26, vcc182; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, -1, v27, vcc183; GFX9-NEXT:    v_or_b32_e32 v18, v24, v26184; GFX9-NEXT:    v_or_b32_e32 v19, v25, v27185; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[18:19]186; GFX9-NEXT:    v_and_b32_e32 v8, 1, v8187; GFX9-NEXT:    v_mov_b32_e32 v19, v9188; GFX9-NEXT:    v_or3_b32 v7, v7, 0, v13189; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]190; GFX9-NEXT:    v_mov_b32_e32 v18, v8191; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]192; GFX9-NEXT:    s_cbranch_execnz .LBB0_3193; GFX9-NEXT:  ; %bb.4: ; %Flow194; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]195; GFX9-NEXT:  .LBB0_5: ; %Flow2196; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]197; GFX9-NEXT:    v_lshlrev_b64 v[14:15], 1, v[10:11]198; GFX9-NEXT:    v_lshlrev_b64 v[6:7], 1, v[6:7]199; GFX9-NEXT:    v_lshrrev_b32_e32 v11, 31, v11200; GFX9-NEXT:    v_or3_b32 v10, v7, 0, v13201; GFX9-NEXT:    v_or3_b32 v12, v6, v11, v12202; GFX9-NEXT:    v_or_b32_e32 v11, v9, v15203; GFX9-NEXT:    v_or_b32_e32 v13, v8, v14204; GFX9-NEXT:  .LBB0_6: ; %Flow3205; GFX9-NEXT:    s_or_b64 exec, exec, s[8:9]206; GFX9-NEXT:    v_mul_lo_u32 v16, v13, v5207; GFX9-NEXT:    v_mad_u64_u32 v[5:6], s[4:5], v23, v13, 0208; GFX9-NEXT:    v_mov_b32_e32 v7, 0209; GFX9-NEXT:    v_mad_u64_u32 v[8:9], s[4:5], v13, v4, 0210; GFX9-NEXT:    v_mad_u64_u32 v[13:14], s[4:5], v22, v13, v[6:7]211; GFX9-NEXT:    v_mul_lo_u32 v15, v11, v4212; GFX9-NEXT:    v_mul_lo_u32 v4, v12, v22213; GFX9-NEXT:    v_mov_b32_e32 v6, v13214; GFX9-NEXT:    v_mad_u64_u32 v[6:7], s[4:5], v23, v11, v[6:7]215; GFX9-NEXT:    v_add3_u32 v9, v9, v16, v15216; GFX9-NEXT:    v_mad_u64_u32 v[8:9], s[4:5], v12, v23, v[8:9]217; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, v14, v7218; GFX9-NEXT:    v_addc_co_u32_e64 v13, s[4:5], 0, 0, vcc219; GFX9-NEXT:    v_mul_lo_u32 v15, v10, v23220; GFX9-NEXT:    v_mad_u64_u32 v[10:11], s[4:5], v22, v11, v[12:13]221; GFX9-NEXT:    v_add3_u32 v4, v15, v9, v4222; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v10, v8223; GFX9-NEXT:    v_addc_co_u32_e32 v4, vcc, v11, v4, vcc224; GFX9-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v5225; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v6, vcc226; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v7, vcc227; GFX9-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v4, vcc228; GFX9-NEXT:    v_xor_b32_e32 v0, v0, v20229; GFX9-NEXT:    v_xor_b32_e32 v1, v1, v21230; GFX9-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v20231; GFX9-NEXT:    v_xor_b32_e32 v2, v2, v20232; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v21, vcc233; GFX9-NEXT:    v_xor_b32_e32 v3, v3, v21234; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v20, vcc235; GFX9-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v21, vcc236; GFX9-NEXT:    s_setpc_b64 s[30:31]237;238; GFX9-O0-LABEL: v_srem_i128_vv:239; GFX9-O0:       ; %bb.0: ; %_udiv-special-cases240; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)241; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[4:5], -1242; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill243; GFX9-O0-NEXT:    s_mov_b64 exec, s[4:5]244; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v6245; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v4246; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v2247; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v1248; GFX9-O0-NEXT:    ; kill: def $vgpr17 killed $vgpr17 def $vgpr17_vgpr18 killed $exec249; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v7250; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec251; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v5252; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec253; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v2254; GFX9-O0-NEXT:    ; kill: def $vgpr12 killed $vgpr12 def $vgpr12_vgpr13 killed $exec255; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v3256; GFX9-O0-NEXT:    s_mov_b32 s4, 63257; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v12258; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v13259; GFX9-O0-NEXT:    v_ashrrev_i64 v[2:3], s4, v[2:3]260; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill261; GFX9-O0-NEXT:    s_nop 0262; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill263; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill264; GFX9-O0-NEXT:    s_nop 0265; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill266; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v0267; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v1268; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v12269; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v13270; GFX9-O0-NEXT:    s_mov_b64 s[6:7], 0271; GFX9-O0-NEXT:    ; implicit-def: $vgpr30 : SGPR spill to VGPR lane272; GFX9-O0-NEXT:    v_writelane_b32 v30, s6, 0273; GFX9-O0-NEXT:    v_writelane_b32 v30, s7, 1274; GFX9-O0-NEXT:    s_mov_b32 s10, s6275; GFX9-O0-NEXT:    v_writelane_b32 v30, s10, 2276; GFX9-O0-NEXT:    s_mov_b32 s11, s7277; GFX9-O0-NEXT:    v_writelane_b32 v30, s11, 3278; GFX9-O0-NEXT:    v_sub_co_u32_e32 v10, vcc, s10, v2279; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s11280; GFX9-O0-NEXT:    v_subb_co_u32_e32 v5, vcc, v4, v3, vcc281; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s10282; GFX9-O0-NEXT:    v_subb_co_u32_e32 v4, vcc, v4, v0, vcc283; GFX9-O0-NEXT:    v_mov_b32_e32 v6, s11284; GFX9-O0-NEXT:    v_subb_co_u32_e32 v6, vcc, v6, v1, vcc285; GFX9-O0-NEXT:    ; kill: def $vgpr10 killed $vgpr10 def $vgpr10_vgpr11 killed $exec286; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v5287; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v11288; GFX9-O0-NEXT:    s_mov_b64 s[4:5], s[6:7]289; GFX9-O0-NEXT:    v_cmp_lt_i64_e64 s[4:5], v[12:13], s[4:5]290; GFX9-O0-NEXT:    v_cndmask_b32_e64 v3, v3, v5, s[4:5]291; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v10292; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v2, v5, s[4:5]293; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v2294; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v3295; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec296; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v6297; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v5298; GFX9-O0-NEXT:    v_cndmask_b32_e64 v1, v1, v6, s[4:5]299; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 killed $vgpr4_vgpr5 killed $exec300; GFX9-O0-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s[4:5]301; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v0302; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v1303; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v8304; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v9305; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v17306; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v18307; GFX9-O0-NEXT:    v_sub_co_u32_e32 v12, vcc, s10, v7308; GFX9-O0-NEXT:    v_mov_b32_e32 v9, s11309; GFX9-O0-NEXT:    v_subb_co_u32_e32 v14, vcc, v9, v10, vcc310; GFX9-O0-NEXT:    v_mov_b32_e32 v9, s10311; GFX9-O0-NEXT:    v_subb_co_u32_e32 v11, vcc, v9, v4, vcc312; GFX9-O0-NEXT:    v_mov_b32_e32 v9, s11313; GFX9-O0-NEXT:    v_subb_co_u32_e32 v9, vcc, v9, v8, vcc314; GFX9-O0-NEXT:    ; kill: def $vgpr12 killed $vgpr12 def $vgpr12_vgpr13 killed $exec315; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v14316; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v13317; GFX9-O0-NEXT:    s_mov_b64 s[4:5], s[6:7]318; GFX9-O0-NEXT:    v_cmp_lt_i64_e64 s[4:5], v[17:18], s[4:5]319; GFX9-O0-NEXT:    v_cndmask_b32_e64 v10, v10, v14, s[4:5]320; GFX9-O0-NEXT:    ; kill: def $vgpr12 killed $vgpr12 killed $vgpr12_vgpr13 killed $exec321; GFX9-O0-NEXT:    v_cndmask_b32_e64 v7, v7, v12, s[4:5]322; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v7323; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v10324; GFX9-O0-NEXT:    ; kill: def $vgpr11 killed $vgpr11 def $vgpr11_vgpr12 killed $exec325; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v9326; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v12327; GFX9-O0-NEXT:    v_cndmask_b32_e64 v8, v8, v9, s[4:5]328; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v11329; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v4, v9, s[4:5]330; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v4331; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v8332; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v6333; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v5334; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill335; GFX9-O0-NEXT:    s_nop 0336; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill337; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v15338; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v16339; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill340; GFX9-O0-NEXT:    s_nop 0341; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill342; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v12343; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v11344; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill345; GFX9-O0-NEXT:    s_nop 0346; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill347; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v17348; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v18349; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill350; GFX9-O0-NEXT:    s_nop 0351; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill352; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v12353; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v11354; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill355; GFX9-O0-NEXT:    s_nop 0356; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill357; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v17358; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v18359; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill360; GFX9-O0-NEXT:    s_nop 0361; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill362; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v6363; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v5364; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill365; GFX9-O0-NEXT:    s_nop 0366; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill367; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v15368; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v16369; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill370; GFX9-O0-NEXT:    s_nop 0371; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill372; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v12373; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v18374; GFX9-O0-NEXT:    v_or_b32_e64 v9, v9, v13375; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v11376; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v17377; GFX9-O0-NEXT:    v_or_b32_e64 v13, v13, v14378; GFX9-O0-NEXT:    ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14 killed $exec379; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v9380; GFX9-O0-NEXT:    v_cmp_eq_u64_e64 s[4:5], v[13:14], s[6:7]381; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v6382; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v16383; GFX9-O0-NEXT:    v_or_b32_e64 v9, v9, v13384; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v5385; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v15386; GFX9-O0-NEXT:    v_or_b32_e64 v13, v13, v14387; GFX9-O0-NEXT:    ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14 killed $exec388; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v9389; GFX9-O0-NEXT:    v_cmp_eq_u64_e64 s[8:9], v[13:14], s[6:7]390; GFX9-O0-NEXT:    s_or_b64 s[4:5], s[4:5], s[8:9]391; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[8:9], v[11:12], s[6:7]392; GFX9-O0-NEXT:    v_ffbh_u32_e64 v4, v4393; GFX9-O0-NEXT:    s_mov_b32 s12, 32394; GFX9-O0-NEXT:    v_add_u32_e64 v4, v4, s12395; GFX9-O0-NEXT:    v_ffbh_u32_e64 v8, v8396; GFX9-O0-NEXT:    v_min_u32_e64 v8, v4, v8397; GFX9-O0-NEXT:    v_mov_b32_e32 v4, 0398; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec399; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v4400; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v9401; GFX9-O0-NEXT:    v_ffbh_u32_e64 v7, v7402; GFX9-O0-NEXT:    v_add_u32_e64 v7, v7, s12403; GFX9-O0-NEXT:    v_ffbh_u32_e64 v10, v10404; GFX9-O0-NEXT:    v_min_u32_e64 v13, v7, v10405; GFX9-O0-NEXT:    ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14 killed $exec406; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v4407; GFX9-O0-NEXT:    s_mov_b64 s[14:15], 64408; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v13409; GFX9-O0-NEXT:    s_mov_b32 s16, s14410; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v14411; GFX9-O0-NEXT:    s_mov_b32 s13, s15412; GFX9-O0-NEXT:    v_add_co_u32_e64 v10, s[16:17], v10, s16413; GFX9-O0-NEXT:    v_mov_b32_e32 v11, s13414; GFX9-O0-NEXT:    v_addc_co_u32_e64 v7, s[16:17], v7, v11, s[16:17]415; GFX9-O0-NEXT:    ; kill: def $vgpr10 killed $vgpr10 def $vgpr10_vgpr11 killed $exec416; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v7417; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v11418; GFX9-O0-NEXT:    v_cndmask_b32_e64 v7, v7, v12, s[8:9]419; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v8420; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v10421; GFX9-O0-NEXT:    v_cndmask_b32_e64 v8, v8, v9, s[8:9]422; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec423; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v7424; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[8:9], v[5:6], s[6:7]425; GFX9-O0-NEXT:    v_ffbh_u32_e64 v5, v0426; GFX9-O0-NEXT:    v_add_u32_e64 v5, v5, s12427; GFX9-O0-NEXT:    v_ffbh_u32_e64 v6, v1428; GFX9-O0-NEXT:    v_min_u32_e64 v5, v5, v6429; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec430; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v4431; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v6432; GFX9-O0-NEXT:    v_ffbh_u32_e64 v10, v2433; GFX9-O0-NEXT:    v_add_u32_e64 v10, v10, s12434; GFX9-O0-NEXT:    v_ffbh_u32_e64 v11, v3435; GFX9-O0-NEXT:    v_min_u32_e64 v11, v10, v11436; GFX9-O0-NEXT:    ; kill: def $vgpr11 killed $vgpr11 def $vgpr11_vgpr12 killed $exec437; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v4438; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v11439; GFX9-O0-NEXT:    s_mov_b32 s12, s14440; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v12441; GFX9-O0-NEXT:    s_mov_b32 s14, s15442; GFX9-O0-NEXT:    v_add_co_u32_e64 v10, s[12:13], v10, s12443; GFX9-O0-NEXT:    v_mov_b32_e32 v11, s14444; GFX9-O0-NEXT:    v_addc_co_u32_e64 v4, s[12:13], v4, v11, s[12:13]445; GFX9-O0-NEXT:    ; kill: def $vgpr10 killed $vgpr10 def $vgpr10_vgpr11 killed $exec446; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v4447; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v11448; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v4, v7, s[8:9]449; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v5450; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v10451; GFX9-O0-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s[8:9]452; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec453; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v4454; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v5455; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v8456; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr5_vgpr6 killed $exec457; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v9458; GFX9-O0-NEXT:    v_sub_co_u32_e32 v4, vcc, v4, v7459; GFX9-O0-NEXT:    v_subb_co_u32_e32 v8, vcc, v5, v6, vcc460; GFX9-O0-NEXT:    v_mov_b32_e32 v6, s10461; GFX9-O0-NEXT:    v_mov_b32_e32 v5, s10462; GFX9-O0-NEXT:    v_subb_co_u32_e32 v7, vcc, v5, v6, vcc463; GFX9-O0-NEXT:    v_mov_b32_e32 v6, s11464; GFX9-O0-NEXT:    v_mov_b32_e32 v5, s11465; GFX9-O0-NEXT:    v_subb_co_u32_e32 v6, vcc, v5, v6, vcc466; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec467; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v8468; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill469; GFX9-O0-NEXT:    s_nop 0470; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill471; GFX9-O0-NEXT:    ; kill: def $vgpr7 killed $vgpr7 def $vgpr7_vgpr8 killed $exec472; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v6473; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill474; GFX9-O0-NEXT:    s_nop 0475; GFX9-O0-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill476; GFX9-O0-NEXT:    v_cmp_eq_u64_e64 s[8:9], v[7:8], s[6:7]477; GFX9-O0-NEXT:    s_mov_b64 s[12:13], 0x7f478; GFX9-O0-NEXT:    v_cmp_gt_u64_e64 s[14:15], v[4:5], s[12:13]479; GFX9-O0-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[14:15]480; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[14:15], v[7:8], s[6:7]481; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[14:15]482; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, v6, v9, s[8:9]483; GFX9-O0-NEXT:    v_and_b32_e64 v6, 1, v6484; GFX9-O0-NEXT:    v_cmp_eq_u32_e64 s[8:9], v6, 1485; GFX9-O0-NEXT:    s_or_b64 s[8:9], s[4:5], s[8:9]486; GFX9-O0-NEXT:    s_mov_b64 s[14:15], -1487; GFX9-O0-NEXT:    s_mov_b64 s[4:5], s[8:9]488; GFX9-O0-NEXT:    s_xor_b64 s[4:5], s[4:5], s[14:15]489; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v5490; GFX9-O0-NEXT:    s_mov_b32 s14, s13491; GFX9-O0-NEXT:    v_xor_b32_e64 v6, v6, s14492; GFX9-O0-NEXT:    ; kill: def $sgpr12 killed $sgpr12 killed $sgpr12_sgpr13493; GFX9-O0-NEXT:    v_xor_b32_e64 v4, v4, s12494; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec495; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v6496; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v5497; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v8498; GFX9-O0-NEXT:    v_or_b32_e64 v6, v6, v9499; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 killed $vgpr4_vgpr5 killed $exec500; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v7501; GFX9-O0-NEXT:    v_or_b32_e64 v4, v4, v5502; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec503; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v6504; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[6:7], v[4:5], s[6:7]505; GFX9-O0-NEXT:    s_mov_b64 s[12:13], s[8:9]506; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s11507; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v1, v4, s[12:13]508; GFX9-O0-NEXT:    s_mov_b64 s[12:13], s[8:9]509; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s10510; GFX9-O0-NEXT:    v_cndmask_b32_e64 v0, v0, v1, s[12:13]511; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec512; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v4513; GFX9-O0-NEXT:    s_mov_b64 s[12:13], s[8:9]514; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s11515; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v3, v4, s[12:13]516; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s10517; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[8:9]518; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec519; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v4520; GFX9-O0-NEXT:    s_and_b64 s[6:7], s[4:5], s[6:7]521; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill522; GFX9-O0-NEXT:    s_nop 0523; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill524; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill525; GFX9-O0-NEXT:    s_nop 0526; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill527; GFX9-O0-NEXT:    s_mov_b64 s[4:5], exec528; GFX9-O0-NEXT:    v_writelane_b32 v30, s4, 4529; GFX9-O0-NEXT:    v_writelane_b32 v30, s5, 5530; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -1531; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill532; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]533; GFX9-O0-NEXT:    s_and_b64 s[4:5], s[4:5], s[6:7]534; GFX9-O0-NEXT:    s_mov_b64 exec, s[4:5]535; GFX9-O0-NEXT:    s_cbranch_execz .LBB0_3536; GFX9-O0-NEXT:    s_branch .LBB0_8537; GFX9-O0-NEXT:  .LBB0_1: ; %Flow538; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -1539; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload540; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]541; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)542; GFX9-O0-NEXT:    v_readlane_b32 s4, v30, 6543; GFX9-O0-NEXT:    v_readlane_b32 s5, v30, 7544; GFX9-O0-NEXT:    s_or_b64 exec, exec, s[4:5]545; GFX9-O0-NEXT:  ; %bb.2: ; %Flow546; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload547; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload548; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload549; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload550; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload551; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload552; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload553; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload554; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)555; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill556; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)557; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill558; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)559; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill560; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)561; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill562; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)563; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill564; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)565; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill566; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)567; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill568; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)569; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill570; GFX9-O0-NEXT:    s_branch .LBB0_5571; GFX9-O0-NEXT:  .LBB0_3: ; %Flow2572; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -1573; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload574; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]575; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)576; GFX9-O0-NEXT:    v_readlane_b32 s4, v30, 4577; GFX9-O0-NEXT:    v_readlane_b32 s5, v30, 5578; GFX9-O0-NEXT:    s_or_b64 exec, exec, s[4:5]579; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload580; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload581; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload582; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload583; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)584; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill585; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)586; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill587; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill588; GFX9-O0-NEXT:    s_nop 0589; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill590; GFX9-O0-NEXT:    s_branch .LBB0_9591; GFX9-O0-NEXT:  .LBB0_4: ; %udiv-loop-exit592; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload593; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload594; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload595; GFX9-O0-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload596; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload597; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload598; GFX9-O0-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload599; GFX9-O0-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload600; GFX9-O0-NEXT:    s_mov_b32 s4, 1601; GFX9-O0-NEXT:    s_waitcnt vmcnt(2)602; GFX9-O0-NEXT:    v_lshlrev_b64 v[2:3], s4, v[0:1]603; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)604; GFX9-O0-NEXT:    v_lshlrev_b64 v[9:10], s4, v[9:10]605; GFX9-O0-NEXT:    s_mov_b32 s4, 63606; GFX9-O0-NEXT:    v_lshrrev_b64 v[0:1], s4, v[0:1]607; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v1608; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v10609; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v8610; GFX9-O0-NEXT:    v_or3_b32 v4, v4, v11, v12611; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v0612; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v9613; GFX9-O0-NEXT:    v_or3_b32 v0, v0, v1, v7614; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec615; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v4616; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v3617; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v6618; GFX9-O0-NEXT:    v_or_b32_e64 v4, v4, v7619; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v2620; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v5621; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v3622; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec623; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v4624; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill625; GFX9-O0-NEXT:    s_nop 0626; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill627; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill628; GFX9-O0-NEXT:    s_nop 0629; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill630; GFX9-O0-NEXT:    s_branch .LBB0_3631; GFX9-O0-NEXT:  .LBB0_5: ; %Flow1632; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -1633; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload634; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]635; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)636; GFX9-O0-NEXT:    v_readlane_b32 s4, v30, 8637; GFX9-O0-NEXT:    v_readlane_b32 s5, v30, 9638; GFX9-O0-NEXT:    s_or_b64 exec, exec, s[4:5]639; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload640; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload641; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload642; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload643; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload644; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload645; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload646; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload647; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)648; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill649; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)650; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill651; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill652; GFX9-O0-NEXT:    s_nop 0653; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill654; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill655; GFX9-O0-NEXT:    s_nop 0656; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill657; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill658; GFX9-O0-NEXT:    s_nop 0659; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill660; GFX9-O0-NEXT:    s_branch .LBB0_4661; GFX9-O0-NEXT:  .LBB0_6: ; %udiv-do-while662; GFX9-O0-NEXT:    ; =>This Inner Loop Header: Depth=1663; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -1664; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload665; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]666; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)667; GFX9-O0-NEXT:    v_readlane_b32 s6, v30, 10668; GFX9-O0-NEXT:    v_readlane_b32 s7, v30, 11669; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload670; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload671; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload672; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload673; GFX9-O0-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload674; GFX9-O0-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload675; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload676; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload677; GFX9-O0-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload678; GFX9-O0-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload679; GFX9-O0-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload680; GFX9-O0-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload681; GFX9-O0-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload682; GFX9-O0-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload683; GFX9-O0-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload684; GFX9-O0-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload685; GFX9-O0-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload686; GFX9-O0-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload687; GFX9-O0-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload688; GFX9-O0-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload689; GFX9-O0-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload690; GFX9-O0-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload691; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload692; GFX9-O0-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload693; GFX9-O0-NEXT:    s_mov_b32 s4, 63694; GFX9-O0-NEXT:    s_waitcnt vmcnt(16)695; GFX9-O0-NEXT:    v_lshrrev_b64 v[28:29], s4, v[2:3]696; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v29697; GFX9-O0-NEXT:    s_mov_b32 s5, 1698; GFX9-O0-NEXT:    v_lshlrev_b64 v[22:23], s5, v[22:23]699; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v23700; GFX9-O0-NEXT:    v_or_b32_e64 v4, v4, v5701; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v28702; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v22703; GFX9-O0-NEXT:    v_or_b32_e64 v22, v5, v10704; GFX9-O0-NEXT:    ; kill: def $vgpr22 killed $vgpr22 def $vgpr22_vgpr23 killed $exec705; GFX9-O0-NEXT:    v_mov_b32_e32 v23, v4706; GFX9-O0-NEXT:    v_lshlrev_b64 v[28:29], s5, v[2:3]707; GFX9-O0-NEXT:    v_lshrrev_b64 v[4:5], s4, v[6:7]708; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v29709; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v5710; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v3711; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v28712; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 killed $vgpr4_vgpr5 killed $exec713; GFX9-O0-NEXT:    v_or_b32_e64 v4, v3, v4714; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec715; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v2716; GFX9-O0-NEXT:    v_lshlrev_b64 v[2:3], s5, v[0:1]717; GFX9-O0-NEXT:    v_lshlrev_b64 v[28:29], s5, v[6:7]718; GFX9-O0-NEXT:    v_lshrrev_b64 v[0:1], s4, v[0:1]719; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v1720; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v29721; GFX9-O0-NEXT:    s_waitcnt vmcnt(10)722; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v27723; GFX9-O0-NEXT:    v_or3_b32 v6, v6, v7, v10724; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v0725; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v28726; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v26727; GFX9-O0-NEXT:    v_or3_b32 v0, v0, v1, v7728; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec729; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v6730; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v3731; GFX9-O0-NEXT:    s_waitcnt vmcnt(8)732; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v25733; GFX9-O0-NEXT:    v_or_b32_e64 v6, v6, v7734; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v2735; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v24736; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v3737; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec738; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v6739; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v4740; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v5741; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v22742; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v23743; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)744; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v11745; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v14746; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v15747; GFX9-O0-NEXT:    v_sub_co_u32_e32 v13, vcc, v13, v6748; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)749; GFX9-O0-NEXT:    v_subb_co_u32_e32 v12, vcc, v12, v10, vcc750; GFX9-O0-NEXT:    v_subb_co_u32_e32 v11, vcc, v11, v4, vcc751; GFX9-O0-NEXT:    v_subb_co_u32_e32 v7, vcc, v7, v5, vcc752; GFX9-O0-NEXT:    ; kill: def $vgpr11 killed $vgpr11 def $vgpr11_vgpr12 killed $exec753; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v7754; GFX9-O0-NEXT:    v_ashrrev_i64 v[13:14], s4, v[11:12]755; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v14756; GFX9-O0-NEXT:    s_mov_b64 s[4:5], 1757; GFX9-O0-NEXT:    s_mov_b32 s8, s5758; GFX9-O0-NEXT:    v_and_b32_e64 v12, v7, s8759; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v13760; GFX9-O0-NEXT:    ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5761; GFX9-O0-NEXT:    v_and_b32_e64 v14, v11, s4762; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec763; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v12764; GFX9-O0-NEXT:    v_mov_b32_e32 v12, 0765; GFX9-O0-NEXT:    v_mov_b32_e32 v13, 0766; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v21767; GFX9-O0-NEXT:    v_and_b32_e64 v22, v7, v22768; GFX9-O0-NEXT:    v_and_b32_e64 v20, v11, v20769; GFX9-O0-NEXT:    ; kill: def $vgpr20 killed $vgpr20 def $vgpr20_vgpr21 killed $exec770; GFX9-O0-NEXT:    v_mov_b32_e32 v21, v22771; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v19772; GFX9-O0-NEXT:    v_and_b32_e64 v7, v7, v22773; GFX9-O0-NEXT:    v_and_b32_e64 v22, v11, v18774; GFX9-O0-NEXT:    ; kill: def $vgpr22 killed $vgpr22 def $vgpr22_vgpr23 killed $exec775; GFX9-O0-NEXT:    v_mov_b32_e32 v23, v7776; GFX9-O0-NEXT:    v_mov_b32_e32 v19, v22777; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v23778; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v20779; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v21780; GFX9-O0-NEXT:    v_sub_co_u32_e32 v6, vcc, v6, v19781; GFX9-O0-NEXT:    v_subb_co_u32_e32 v10, vcc, v10, v18, vcc782; GFX9-O0-NEXT:    v_subb_co_u32_e32 v4, vcc, v4, v11, vcc783; GFX9-O0-NEXT:    v_subb_co_u32_e32 v7, vcc, v5, v7, vcc784; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec785; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v7786; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec787; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v10788; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v8789; GFX9-O0-NEXT:    ; kill: def $vgpr9 killed $vgpr9 killed $vgpr8_vgpr9 killed $exec790; GFX9-O0-NEXT:    s_mov_b64 s[8:9], -1791; GFX9-O0-NEXT:    s_mov_b32 s5, s8792; GFX9-O0-NEXT:    s_mov_b32 s4, s9793; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v16794; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v17795; GFX9-O0-NEXT:    v_mov_b32_e32 v16, s5796; GFX9-O0-NEXT:    v_add_co_u32_e32 v19, vcc, v11, v16797; GFX9-O0-NEXT:    v_mov_b32_e32 v11, s4798; GFX9-O0-NEXT:    v_addc_co_u32_e32 v9, vcc, v9, v11, vcc799; GFX9-O0-NEXT:    v_mov_b32_e32 v11, s5800; GFX9-O0-NEXT:    v_addc_co_u32_e32 v16, vcc, v10, v11, vcc801; GFX9-O0-NEXT:    v_mov_b32_e32 v10, s4802; GFX9-O0-NEXT:    v_addc_co_u32_e32 v8, vcc, v8, v10, vcc803; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec804; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v9805; GFX9-O0-NEXT:    ; kill: def $vgpr16 killed $vgpr16 def $vgpr16_vgpr17 killed $exec806; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v8807; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v16808; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v17809; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v19810; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v20811; GFX9-O0-NEXT:    v_mov_b32_e32 v21, v17812; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v20813; GFX9-O0-NEXT:    v_or_b32_e64 v18, v18, v21814; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v16815; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v19816; GFX9-O0-NEXT:    v_or_b32_e64 v16, v16, v17817; GFX9-O0-NEXT:    ; kill: def $vgpr16 killed $vgpr16 def $vgpr16_vgpr17 killed $exec818; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v18819; GFX9-O0-NEXT:    v_cmp_eq_u64_e64 s[4:5], v[16:17], v[12:13]820; GFX9-O0-NEXT:    s_or_b64 s[4:5], s[4:5], s[6:7]821; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v3822; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v2823; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill824; GFX9-O0-NEXT:    s_nop 0825; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill826; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v1827; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v0828; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill829; GFX9-O0-NEXT:    s_nop 0830; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill831; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v15832; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v14833; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill834; GFX9-O0-NEXT:    s_nop 0835; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill836; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v13837; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v12838; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill839; GFX9-O0-NEXT:    s_nop 0840; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill841; GFX9-O0-NEXT:    s_mov_b64 s[6:7], s[4:5]842; GFX9-O0-NEXT:    v_writelane_b32 v30, s6, 6843; GFX9-O0-NEXT:    v_writelane_b32 v30, s7, 7844; GFX9-O0-NEXT:    s_mov_b64 s[6:7], s[4:5]845; GFX9-O0-NEXT:    v_writelane_b32 v30, s6, 10846; GFX9-O0-NEXT:    v_writelane_b32 v30, s7, 11847; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -1848; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill849; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]850; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill851; GFX9-O0-NEXT:    s_nop 0852; GFX9-O0-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill853; GFX9-O0-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill854; GFX9-O0-NEXT:    s_nop 0855; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill856; GFX9-O0-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill857; GFX9-O0-NEXT:    s_nop 0858; GFX9-O0-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill859; GFX9-O0-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill860; GFX9-O0-NEXT:    s_nop 0861; GFX9-O0-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill862; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill863; GFX9-O0-NEXT:    s_nop 0864; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill865; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill866; GFX9-O0-NEXT:    s_nop 0867; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill868; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill869; GFX9-O0-NEXT:    s_nop 0870; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill871; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill872; GFX9-O0-NEXT:    s_nop 0873; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill874; GFX9-O0-NEXT:    s_andn2_b64 exec, exec, s[4:5]875; GFX9-O0-NEXT:    s_cbranch_execnz .LBB0_6876; GFX9-O0-NEXT:    s_branch .LBB0_1877; GFX9-O0-NEXT:  .LBB0_7: ; %udiv-preheader878; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -1879; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload880; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]881; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload882; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload883; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload884; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload885; GFX9-O0-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload886; GFX9-O0-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload887; GFX9-O0-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload888; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload889; GFX9-O0-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload890; GFX9-O0-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload891; GFX9-O0-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload892; GFX9-O0-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload893; GFX9-O0-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload894; GFX9-O0-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload895; GFX9-O0-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload896; GFX9-O0-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload897; GFX9-O0-NEXT:    s_waitcnt vmcnt(9)898; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v10899; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)900; GFX9-O0-NEXT:    v_lshrrev_b64 v[6:7], v4, v[18:19]901; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v7902; GFX9-O0-NEXT:    s_mov_b32 s4, 64903; GFX9-O0-NEXT:    v_sub_u32_e64 v20, s4, v4904; GFX9-O0-NEXT:    v_lshlrev_b64 v[20:21], v20, v[14:15]905; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v21906; GFX9-O0-NEXT:    v_or_b32_e64 v5, v5, v22907; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec908; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v20909; GFX9-O0-NEXT:    v_or_b32_e64 v6, v6, v7910; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec911; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v5912; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v7913; GFX9-O0-NEXT:    v_cmp_lt_u32_e64 s[6:7], v4, s4914; GFX9-O0-NEXT:    v_sub_u32_e64 v5, v4, s4915; GFX9-O0-NEXT:    v_lshrrev_b64 v[20:21], v5, v[14:15]916; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v21917; GFX9-O0-NEXT:    v_cndmask_b32_e64 v5, v5, v22, s[6:7]918; GFX9-O0-NEXT:    s_mov_b32 s4, 0919; GFX9-O0-NEXT:    v_cmp_eq_u32_e64 s[4:5], v4, s4920; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v19921; GFX9-O0-NEXT:    v_cndmask_b32_e64 v5, v5, v22, s[4:5]922; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v6923; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v20924; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, v6, v7, s[6:7]925; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v18926; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, v6, v7, s[4:5]927; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec928; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v5929; GFX9-O0-NEXT:    v_lshrrev_b64 v[4:5], v4, v[14:15]930; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v5931; GFX9-O0-NEXT:    s_mov_b64 s[4:5], 0932; GFX9-O0-NEXT:    s_mov_b32 s8, s5933; GFX9-O0-NEXT:    v_mov_b32_e32 v14, s8934; GFX9-O0-NEXT:    v_cndmask_b32_e64 v14, v14, v15, s[6:7]935; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v4936; GFX9-O0-NEXT:    s_mov_b32 s8, s4937; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s8938; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v4, v5, s[6:7]939; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec940; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v14941; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v12942; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v13943; GFX9-O0-NEXT:    s_mov_b64 s[8:9], -1944; GFX9-O0-NEXT:    s_mov_b32 s7, s8945; GFX9-O0-NEXT:    s_mov_b32 s6, s9946; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v16947; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v17948; GFX9-O0-NEXT:    v_mov_b32_e32 v16, s7949; GFX9-O0-NEXT:    v_add_co_u32_e32 v16, vcc, v15, v16950; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s6951; GFX9-O0-NEXT:    v_addc_co_u32_e32 v12, vcc, v12, v15, vcc952; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s7953; GFX9-O0-NEXT:    v_addc_co_u32_e32 v18, vcc, v14, v15, vcc954; GFX9-O0-NEXT:    v_mov_b32_e32 v14, s6955; GFX9-O0-NEXT:    v_addc_co_u32_e32 v13, vcc, v13, v14, vcc956; GFX9-O0-NEXT:    ; kill: def $vgpr18 killed $vgpr18 def $vgpr18_vgpr19 killed $exec957; GFX9-O0-NEXT:    v_mov_b32_e32 v19, v13958; GFX9-O0-NEXT:    ; kill: def $vgpr16 killed $vgpr16 def $vgpr16_vgpr17 killed $exec959; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v12960; GFX9-O0-NEXT:    v_mov_b32_e32 v13, s5961; GFX9-O0-NEXT:    v_mov_b32_e32 v12, s4962; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s5963; GFX9-O0-NEXT:    v_mov_b32_e32 v14, s4964; GFX9-O0-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill965; GFX9-O0-NEXT:    s_nop 0966; GFX9-O0-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill967; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill968; GFX9-O0-NEXT:    s_nop 0969; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill970; GFX9-O0-NEXT:    v_writelane_b32 v30, s4, 10971; GFX9-O0-NEXT:    v_writelane_b32 v30, s5, 11972; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -1973; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill974; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]975; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill976; GFX9-O0-NEXT:    s_nop 0977; GFX9-O0-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill978; GFX9-O0-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill979; GFX9-O0-NEXT:    s_nop 0980; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill981; GFX9-O0-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill982; GFX9-O0-NEXT:    s_nop 0983; GFX9-O0-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill984; GFX9-O0-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill985; GFX9-O0-NEXT:    s_nop 0986; GFX9-O0-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill987; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill988; GFX9-O0-NEXT:    s_nop 0989; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill990; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill991; GFX9-O0-NEXT:    s_nop 0992; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill993; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill994; GFX9-O0-NEXT:    s_nop 0995; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill996; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill997; GFX9-O0-NEXT:    s_nop 0998; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill999; GFX9-O0-NEXT:    s_branch .LBB0_61000; GFX9-O0-NEXT:  .LBB0_8: ; %udiv-bb11001; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -11002; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload1003; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]1004; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload1005; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload1006; GFX9-O0-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload1007; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload1008; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload1009; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload1010; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload1011; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload1012; GFX9-O0-NEXT:    s_mov_b64 s[6:7], 11013; GFX9-O0-NEXT:    s_mov_b32 s5, s61014; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)1015; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v01016; GFX9-O0-NEXT:    s_mov_b32 s4, s71017; GFX9-O0-NEXT:    s_mov_b64 s[6:7], 01018; GFX9-O0-NEXT:    s_mov_b32 s8, s61019; GFX9-O0-NEXT:    s_mov_b32 s9, s71020; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v41021; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v51022; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s51023; GFX9-O0-NEXT:    v_add_co_u32_e32 v8, vcc, v3, v41024; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s41025; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)1026; GFX9-O0-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v4, vcc1027; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s81028; GFX9-O0-NEXT:    v_addc_co_u32_e32 v0, vcc, v0, v4, vcc1029; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s91030; GFX9-O0-NEXT:    v_addc_co_u32_e32 v2, vcc, v2, v4, vcc1031; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec1032; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v11033; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1034; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v21035; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v11036; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v01037; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill1038; GFX9-O0-NEXT:    s_nop 01039; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill1040; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v81041; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v91042; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill1043; GFX9-O0-NEXT:    s_nop 01044; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill1045; GFX9-O0-NEXT:    s_mov_b32 s4, 0x7f1046; GFX9-O0-NEXT:    v_sub_u32_e64 v2, s4, v31047; GFX9-O0-NEXT:    v_lshlrev_b64 v[4:5], v2, v[10:11]1048; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v51049; GFX9-O0-NEXT:    s_mov_b32 s4, 641050; GFX9-O0-NEXT:    v_sub_u32_e64 v13, s4, v21051; GFX9-O0-NEXT:    v_lshrrev_b64 v[13:14], v13, v[6:7]1052; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v141053; GFX9-O0-NEXT:    v_or_b32_e64 v12, v12, v151054; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 killed $vgpr4_vgpr5 killed $exec1055; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v131056; GFX9-O0-NEXT:    v_or_b32_e64 v4, v4, v51057; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec1058; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v121059; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v51060; GFX9-O0-NEXT:    v_cmp_lt_u32_e64 s[4:5], v2, s41061; GFX9-O0-NEXT:    s_mov_b32 s10, 631062; GFX9-O0-NEXT:    v_sub_u32_e64 v3, s10, v31063; GFX9-O0-NEXT:    v_lshlrev_b64 v[12:13], v3, v[6:7]1064; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v131065; GFX9-O0-NEXT:    v_cndmask_b32_e64 v3, v3, v14, s[4:5]1066; GFX9-O0-NEXT:    s_mov_b32 s10, 01067; GFX9-O0-NEXT:    v_cmp_eq_u32_e64 s[10:11], v2, s101068; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v111069; GFX9-O0-NEXT:    v_cndmask_b32_e64 v3, v3, v14, s[10:11]1070; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v41071; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v121072; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v4, v5, s[4:5]1073; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v101074; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v4, v5, s[10:11]1075; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec1076; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v31077; GFX9-O0-NEXT:    v_lshlrev_b64 v[6:7], v2, v[6:7]1078; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v71079; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s91080; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[4:5]1081; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec1082; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s81083; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, v3, v6, s[4:5]1084; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec1085; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v21086; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill1087; GFX9-O0-NEXT:    s_nop 01088; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill1089; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill1090; GFX9-O0-NEXT:    s_nop 01091; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill1092; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v11093; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v91094; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v31095; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v01096; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v81097; GFX9-O0-NEXT:    v_or_b32_e64 v0, v0, v11098; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1099; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v21100; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[4:5], v[0:1], s[6:7]1101; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s61102; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s71103; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s61104; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s71105; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill1106; GFX9-O0-NEXT:    s_nop 01107; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill1108; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill1109; GFX9-O0-NEXT:    s_nop 01110; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill1111; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill1112; GFX9-O0-NEXT:    s_nop 01113; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill1114; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill1115; GFX9-O0-NEXT:    s_nop 01116; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill1117; GFX9-O0-NEXT:    s_mov_b64 s[6:7], exec1118; GFX9-O0-NEXT:    s_and_b64 s[4:5], s[6:7], s[4:5]1119; GFX9-O0-NEXT:    s_xor_b64 s[6:7], s[4:5], s[6:7]1120; GFX9-O0-NEXT:    v_writelane_b32 v30, s6, 81121; GFX9-O0-NEXT:    v_writelane_b32 v30, s7, 91122; GFX9-O0-NEXT:    s_or_saveexec_b64 s[20:21], -11123; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill1124; GFX9-O0-NEXT:    s_mov_b64 exec, s[20:21]1125; GFX9-O0-NEXT:    s_mov_b64 exec, s[4:5]1126; GFX9-O0-NEXT:    s_cbranch_execz .LBB0_51127; GFX9-O0-NEXT:    s_branch .LBB0_71128; GFX9-O0-NEXT:  .LBB0_9: ; %udiv-end1129; GFX9-O0-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload1130; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload1131; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload1132; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload1133; GFX9-O0-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload1134; GFX9-O0-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload1135; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload1136; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload1137; GFX9-O0-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload1138; GFX9-O0-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload1139; GFX9-O0-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload1140; GFX9-O0-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload1141; GFX9-O0-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload1142; GFX9-O0-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload1143; GFX9-O0-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload1144; GFX9-O0-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload1145; GFX9-O0-NEXT:    s_mov_b32 s4, 321146; GFX9-O0-NEXT:    s_waitcnt vmcnt(2)1147; GFX9-O0-NEXT:    v_lshrrev_b64 v[0:1], s4, v[17:18]1148; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v01149; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)1150; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v211151; GFX9-O0-NEXT:    v_mul_lo_u32 v8, v1, v01152; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)1153; GFX9-O0-NEXT:    v_lshrrev_b64 v[21:22], s4, v[21:22]1154; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v211155; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v171156; GFX9-O0-NEXT:    v_mul_lo_u32 v2, v5, v21157; GFX9-O0-NEXT:    v_mad_u64_u32 v[17:18], s[6:7], v5, v0, 01158; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v181159; GFX9-O0-NEXT:    v_add3_u32 v8, v0, v2, v81160; GFX9-O0-NEXT:    ; implicit-def: $sgpr51161; GFX9-O0-NEXT:    ; implicit-def: $sgpr61162; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec1163; GFX9-O0-NEXT:    v_mov_b32_e32 v9, s51164; GFX9-O0-NEXT:    v_lshlrev_b64 v[8:9], s4, v[8:9]1165; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v91166; GFX9-O0-NEXT:    ; kill: def $vgpr17 killed $vgpr17 killed $vgpr17_vgpr18 killed $exec1167; GFX9-O0-NEXT:    v_mov_b32_e32 v2, 01168; GFX9-O0-NEXT:    ; kill: def $vgpr17 killed $vgpr17 def $vgpr17_vgpr18 killed $exec1169; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v21170; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v181171; GFX9-O0-NEXT:    v_or_b32_e64 v0, v0, v141172; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v81173; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v171174; GFX9-O0-NEXT:    v_or_b32_e64 v17, v8, v91175; GFX9-O0-NEXT:    ; kill: def $vgpr17 killed $vgpr17 def $vgpr17_vgpr18 killed $exec1176; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v01177; GFX9-O0-NEXT:    v_lshrrev_b64 v[8:9], s4, v[19:20]1178; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 killed $vgpr8_vgpr9 killed $exec1179; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v151180; GFX9-O0-NEXT:    v_mul_lo_u32 v14, v9, v81181; GFX9-O0-NEXT:    v_lshrrev_b64 v[15:16], s4, v[15:16]1182; GFX9-O0-NEXT:    ; kill: def $vgpr15 killed $vgpr15 killed $vgpr15_vgpr16 killed $exec1183; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v191184; GFX9-O0-NEXT:    v_mul_lo_u32 v15, v15, v01185; GFX9-O0-NEXT:    v_mad_u64_u32 v[19:20], s[6:7], v9, v0, 01186; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v201187; GFX9-O0-NEXT:    v_add3_u32 v14, v9, v14, v151188; GFX9-O0-NEXT:    ; implicit-def: $sgpr51189; GFX9-O0-NEXT:    ; implicit-def: $sgpr61190; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec1191; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s51192; GFX9-O0-NEXT:    v_lshlrev_b64 v[14:15], s4, v[14:15]1193; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v151194; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 killed $vgpr19_vgpr20 killed $exec1195; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec1196; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v21197; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v201198; GFX9-O0-NEXT:    v_or_b32_e64 v9, v9, v161199; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v141200; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v191201; GFX9-O0-NEXT:    v_or_b32_e64 v19, v14, v151202; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec1203; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v91204; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v191205; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v171206; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v201207; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v181208; GFX9-O0-NEXT:    v_add_co_u32_e64 v17, s[6:7], v15, v161209; GFX9-O0-NEXT:    v_addc_co_u32_e64 v9, s[6:7], v9, v14, s[6:7]1210; GFX9-O0-NEXT:    ; kill: def $vgpr17 killed $vgpr17 def $vgpr17_vgpr18 killed $exec1211; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v91212; GFX9-O0-NEXT:    v_mad_u64_u32 v[14:15], s[6:7], v8, v1, 01213; GFX9-O0-NEXT:    v_mov_b32_e32 v19, v141214; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec1215; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v21216; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v201217; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v151218; GFX9-O0-NEXT:    ; implicit-def: $sgpr51219; GFX9-O0-NEXT:    ; implicit-def: $sgpr61220; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec1221; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s51222; GFX9-O0-NEXT:    v_lshlrev_b64 v[15:16], s4, v[14:15]1223; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v161224; GFX9-O0-NEXT:    v_or_b32_e64 v9, v9, v141225; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v191226; GFX9-O0-NEXT:    ; kill: def $vgpr15 killed $vgpr15 killed $vgpr15_vgpr16 killed $exec1227; GFX9-O0-NEXT:    v_or_b32_e64 v21, v14, v151228; GFX9-O0-NEXT:    ; kill: def $vgpr21 killed $vgpr21 def $vgpr21_vgpr22 killed $exec1229; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v91230; GFX9-O0-NEXT:    v_mad_u64_u32 v[14:15], s[6:7], v8, v5, 01231; GFX9-O0-NEXT:    v_mov_b32_e32 v19, v141232; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec1233; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v21234; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v201235; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v151236; GFX9-O0-NEXT:    ; implicit-def: $sgpr51237; GFX9-O0-NEXT:    ; implicit-def: $sgpr61238; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec1239; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s51240; GFX9-O0-NEXT:    v_lshlrev_b64 v[14:15], s4, v[14:15]1241; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v151242; GFX9-O0-NEXT:    v_or_b32_e64 v8, v8, v91243; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v191244; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 killed $vgpr14_vgpr15 killed $exec1245; GFX9-O0-NEXT:    v_or_b32_e64 v23, v9, v141246; GFX9-O0-NEXT:    ; kill: def $vgpr23 killed $vgpr23 def $vgpr23_vgpr24 killed $exec1247; GFX9-O0-NEXT:    v_mov_b32_e32 v24, v81248; GFX9-O0-NEXT:    v_mad_u64_u32 v[14:15], s[6:7], v0, v5, 01249; GFX9-O0-NEXT:    v_mov_b32_e32 v19, v151250; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec1251; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v21252; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v231253; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v191254; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v241255; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v201256; GFX9-O0-NEXT:    v_add_co_u32_e64 v8, s[6:7], v8, v161257; GFX9-O0-NEXT:    v_addc_co_u32_e64 v5, s[6:7], v5, v9, s[6:7]1258; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec1259; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v51260; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v91261; GFX9-O0-NEXT:    s_mov_b64 s[6:7], 0xffffffff1262; GFX9-O0-NEXT:    s_mov_b32 s5, s71263; GFX9-O0-NEXT:    v_and_b32_e64 v5, v5, s51264; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v81265; GFX9-O0-NEXT:    s_mov_b32 s5, s61266; GFX9-O0-NEXT:    v_and_b32_e64 v19, v16, s51267; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec1268; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v51269; GFX9-O0-NEXT:    v_mad_u64_u32 v[23:24], s[6:7], v0, v1, 01270; GFX9-O0-NEXT:    v_mov_b32_e32 v25, v231271; GFX9-O0-NEXT:    ; kill: def $vgpr25 killed $vgpr25 def $vgpr25_vgpr26 killed $exec1272; GFX9-O0-NEXT:    v_mov_b32_e32 v26, v21273; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v261274; GFX9-O0-NEXT:    v_mov_b32_e32 v23, v241275; GFX9-O0-NEXT:    ; implicit-def: $sgpr51276; GFX9-O0-NEXT:    ; implicit-def: $sgpr61277; GFX9-O0-NEXT:    ; kill: def $vgpr23 killed $vgpr23 def $vgpr23_vgpr24 killed $exec1278; GFX9-O0-NEXT:    v_mov_b32_e32 v24, s51279; GFX9-O0-NEXT:    v_lshlrev_b64 v[23:24], s4, v[23:24]1280; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v241281; GFX9-O0-NEXT:    v_or_b32_e64 v0, v0, v11282; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v251283; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v231284; GFX9-O0-NEXT:    v_or_b32_e64 v23, v1, v51285; GFX9-O0-NEXT:    ; kill: def $vgpr23 killed $vgpr23 def $vgpr23_vgpr24 killed $exec1286; GFX9-O0-NEXT:    v_mov_b32_e32 v24, v01287; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v231288; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v191289; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v241290; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v201291; GFX9-O0-NEXT:    v_add_co_u32_e64 v0, s[6:7], v0, v161292; GFX9-O0-NEXT:    v_addc_co_u32_e64 v5, s[6:7], v1, v5, s[6:7]1293; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1294; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v51295; GFX9-O0-NEXT:    v_lshrrev_b64 v[19:20], s4, v[0:1]1296; GFX9-O0-NEXT:    v_lshrrev_b64 v[23:24], s4, v[8:9]1297; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v231298; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v191299; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v241300; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v201301; GFX9-O0-NEXT:    v_add_co_u32_e64 v19, s[6:7], v9, v161302; GFX9-O0-NEXT:    v_addc_co_u32_e64 v5, s[6:7], v5, v8, s[6:7]1303; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec1304; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v51305; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v211306; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v191307; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v221308; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v201309; GFX9-O0-NEXT:    v_add_co_u32_e64 v19, s[6:7], v9, v161310; GFX9-O0-NEXT:    v_addc_co_u32_e64 v5, s[6:7], v5, v8, s[6:7]1311; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec1312; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v51313; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v191314; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v171315; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v201316; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v181317; GFX9-O0-NEXT:    v_add_co_u32_e64 v8, s[6:7], v8, v161318; GFX9-O0-NEXT:    v_addc_co_u32_e64 v5, s[6:7], v5, v9, s[6:7]1319; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec1320; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v51321; GFX9-O0-NEXT:    v_lshlrev_b64 v[0:1], s4, v[0:1]1322; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v11323; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 killed $vgpr14_vgpr15 killed $exec1324; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec1325; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v21326; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v151327; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v51328; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v01329; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v141330; GFX9-O0-NEXT:    v_or_b32_e64 v0, v0, v11331; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1332; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v21333; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v81334; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v91335; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v01336; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v11337; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v31338; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v41339; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v121340; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v131341; GFX9-O0-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v91342; GFX9-O0-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v8, vcc1343; GFX9-O0-NEXT:    v_subb_co_u32_e32 v4, vcc, v4, v5, vcc1344; GFX9-O0-NEXT:    v_subb_co_u32_e32 v2, vcc, v1, v2, vcc1345; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1346; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v31347; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec1348; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v21349; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v51350; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v71351; GFX9-O0-NEXT:    v_xor_b32_e64 v3, v3, v21352; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v41353; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v61354; GFX9-O0-NEXT:    v_xor_b32_e64 v8, v5, v41355; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec1356; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v31357; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v11358; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v111359; GFX9-O0-NEXT:    v_xor_b32_e64 v3, v3, v61360; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec1361; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v101362; GFX9-O0-NEXT:    v_xor_b32_e64 v0, v0, v71363; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1364; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v31365; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v01366; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v11367; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v81368; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v91369; GFX9-O0-NEXT:    v_sub_co_u32_e32 v5, vcc, v5, v71370; GFX9-O0-NEXT:    v_subb_co_u32_e32 v0, vcc, v0, v6, vcc1371; GFX9-O0-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v4, vcc1372; GFX9-O0-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v2, vcc1373; GFX9-O0-NEXT:    ; kill: def $vgpr3 killed $vgpr3 def $vgpr3_vgpr4 killed $exec1374; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v11375; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec1376; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v01377; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v51378; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v31379; GFX9-O0-NEXT:    v_lshrrev_b64 v[5:6], s4, v[5:6]1380; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v51381; GFX9-O0-NEXT:    v_lshrrev_b64 v[3:4], s4, v[3:4]1382; GFX9-O0-NEXT:    ; kill: def $vgpr3 killed $vgpr3 killed $vgpr3_vgpr4 killed $exec1383; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[4:5], -11384; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload1385; GFX9-O0-NEXT:    s_mov_b64 exec, s[4:5]1386; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)1387; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]1388  %div = srem i128 %lhs, %rhs1389  ret i128 %div1390}1391 1392define i128 @v_urem_i128_vv(i128 %lhs, i128 %rhs) {1393; GFX9-LABEL: v_urem_i128_vv:1394; GFX9:       ; %bb.0: ; %_udiv-special-cases1395; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1396; GFX9-NEXT:    v_or_b32_e32 v9, v5, v71397; GFX9-NEXT:    v_or_b32_e32 v8, v4, v61398; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[8:9]1399; GFX9-NEXT:    v_or_b32_e32 v9, v1, v31400; GFX9-NEXT:    v_or_b32_e32 v8, v0, v21401; GFX9-NEXT:    v_cmp_eq_u64_e64 s[4:5], 0, v[8:9]1402; GFX9-NEXT:    v_ffbh_u32_e32 v8, v61403; GFX9-NEXT:    v_add_u32_e32 v8, 32, v81404; GFX9-NEXT:    v_ffbh_u32_e32 v9, v71405; GFX9-NEXT:    v_min_u32_e32 v8, v8, v91406; GFX9-NEXT:    v_ffbh_u32_e32 v9, v41407; GFX9-NEXT:    v_add_u32_e32 v9, 32, v91408; GFX9-NEXT:    v_ffbh_u32_e32 v10, v51409; GFX9-NEXT:    v_min_u32_e32 v9, v9, v101410; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]1411; GFX9-NEXT:    v_add_co_u32_e32 v9, vcc, 64, v91412; GFX9-NEXT:    v_addc_co_u32_e64 v10, s[6:7], 0, 0, vcc1413; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[6:7]1414; GFX9-NEXT:    v_ffbh_u32_e32 v11, v31415; GFX9-NEXT:    v_cndmask_b32_e32 v8, v9, v8, vcc1416; GFX9-NEXT:    v_ffbh_u32_e32 v9, v21417; GFX9-NEXT:    v_add_u32_e32 v9, 32, v91418; GFX9-NEXT:    v_min_u32_e32 v9, v9, v111419; GFX9-NEXT:    v_ffbh_u32_e32 v11, v01420; GFX9-NEXT:    v_add_u32_e32 v11, 32, v111421; GFX9-NEXT:    v_ffbh_u32_e32 v12, v11422; GFX9-NEXT:    v_min_u32_e32 v11, v11, v121423; GFX9-NEXT:    v_cndmask_b32_e64 v10, v10, 0, vcc1424; GFX9-NEXT:    v_add_co_u32_e32 v11, vcc, 64, v111425; GFX9-NEXT:    v_addc_co_u32_e64 v12, s[6:7], 0, 0, vcc1426; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[2:3]1427; GFX9-NEXT:    s_mov_b64 s[6:7], 0x7f1428; GFX9-NEXT:    v_cndmask_b32_e32 v9, v11, v9, vcc1429; GFX9-NEXT:    v_cndmask_b32_e64 v12, v12, 0, vcc1430; GFX9-NEXT:    v_sub_co_u32_e32 v8, vcc, v8, v91431; GFX9-NEXT:    v_subb_co_u32_e32 v9, vcc, v10, v12, vcc1432; GFX9-NEXT:    v_mov_b32_e32 v11, 01433; GFX9-NEXT:    v_subb_co_u32_e32 v10, vcc, 0, v11, vcc1434; GFX9-NEXT:    v_subb_co_u32_e32 v11, vcc, 0, v11, vcc1435; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[8:9]1436; GFX9-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc1437; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[10:11]1438; GFX9-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc1439; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[10:11]1440; GFX9-NEXT:    v_cndmask_b32_e32 v12, v13, v12, vcc1441; GFX9-NEXT:    v_and_b32_e32 v12, 1, v121442; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v121443; GFX9-NEXT:    v_xor_b32_e32 v12, 0x7f, v81444; GFX9-NEXT:    v_or_b32_e32 v13, v9, v111445; GFX9-NEXT:    v_or_b32_e32 v12, v12, v101446; GFX9-NEXT:    s_or_b64 s[4:5], s[4:5], vcc1447; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[12:13]1448; GFX9-NEXT:    s_xor_b64 s[6:7], s[4:5], -11449; GFX9-NEXT:    v_cndmask_b32_e64 v15, v3, 0, s[4:5]1450; GFX9-NEXT:    v_cndmask_b32_e64 v14, v2, 0, s[4:5]1451; GFX9-NEXT:    v_cndmask_b32_e64 v13, v1, 0, s[4:5]1452; GFX9-NEXT:    v_cndmask_b32_e64 v12, v0, 0, s[4:5]1453; GFX9-NEXT:    s_and_b64 s[4:5], s[6:7], vcc1454; GFX9-NEXT:    s_and_saveexec_b64 s[8:9], s[4:5]1455; GFX9-NEXT:    s_cbranch_execz .LBB1_61456; GFX9-NEXT:  ; %bb.1: ; %udiv-bb11457; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, 1, v81458; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, 0, v9, vcc1459; GFX9-NEXT:    v_addc_co_u32_e32 v24, vcc, 0, v10, vcc1460; GFX9-NEXT:    v_sub_u32_e32 v15, 0x7f, v81461; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v11, vcc1462; GFX9-NEXT:    v_sub_u32_e32 v13, 64, v151463; GFX9-NEXT:    v_or_b32_e32 v10, v23, v251464; GFX9-NEXT:    v_or_b32_e32 v9, v22, v241465; GFX9-NEXT:    v_lshlrev_b64 v[11:12], v15, v[2:3]1466; GFX9-NEXT:    v_lshrrev_b64 v[13:14], v13, v[0:1]1467; GFX9-NEXT:    v_sub_u32_e32 v8, 63, v81468; GFX9-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[9:10]1469; GFX9-NEXT:    v_lshlrev_b64 v[8:9], v8, v[0:1]1470; GFX9-NEXT:    v_or_b32_e32 v10, v12, v141471; GFX9-NEXT:    v_or_b32_e32 v11, v11, v131472; GFX9-NEXT:    v_cmp_gt_u32_e64 s[4:5], 64, v151473; GFX9-NEXT:    v_cndmask_b32_e64 v9, v9, v10, s[4:5]1474; GFX9-NEXT:    v_cndmask_b32_e64 v8, v8, v11, s[4:5]1475; GFX9-NEXT:    v_lshlrev_b64 v[10:11], v15, v[0:1]1476; GFX9-NEXT:    v_cmp_eq_u32_e64 s[6:7], 0, v151477; GFX9-NEXT:    v_mov_b32_e32 v12, 01478; GFX9-NEXT:    v_mov_b32_e32 v14, 01479; GFX9-NEXT:    v_cndmask_b32_e64 v9, v9, v3, s[6:7]1480; GFX9-NEXT:    v_cndmask_b32_e64 v8, v8, v2, s[6:7]1481; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, v11, s[4:5]1482; GFX9-NEXT:    v_mov_b32_e32 v13, 01483; GFX9-NEXT:    v_mov_b32_e32 v15, 01484; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, v10, s[4:5]1485; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc1486; GFX9-NEXT:    s_xor_b64 s[6:7], exec, s[4:5]1487; GFX9-NEXT:    s_cbranch_execz .LBB1_51488; GFX9-NEXT:  ; %bb.2: ; %udiv-preheader1489; GFX9-NEXT:    v_sub_u32_e32 v14, 64, v221490; GFX9-NEXT:    v_lshrrev_b64 v[12:13], v22, v[0:1]1491; GFX9-NEXT:    v_lshlrev_b64 v[14:15], v14, v[2:3]1492; GFX9-NEXT:    v_cmp_gt_u32_e32 vcc, 64, v221493; GFX9-NEXT:    v_or_b32_e32 v14, v12, v141494; GFX9-NEXT:    v_subrev_u32_e32 v12, 64, v221495; GFX9-NEXT:    v_or_b32_e32 v15, v13, v151496; GFX9-NEXT:    v_lshrrev_b64 v[12:13], v12, v[2:3]1497; GFX9-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v221498; GFX9-NEXT:    v_cndmask_b32_e32 v13, v13, v15, vcc1499; GFX9-NEXT:    v_cndmask_b32_e64 v17, v13, v1, s[4:5]1500; GFX9-NEXT:    v_cndmask_b32_e32 v14, v12, v14, vcc1501; GFX9-NEXT:    v_lshrrev_b64 v[12:13], v22, v[2:3]1502; GFX9-NEXT:    v_cndmask_b32_e64 v16, v14, v0, s[4:5]1503; GFX9-NEXT:    v_cndmask_b32_e32 v19, 0, v13, vcc1504; GFX9-NEXT:    v_cndmask_b32_e32 v18, 0, v12, vcc1505; GFX9-NEXT:    v_add_co_u32_e32 v26, vcc, -1, v41506; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, -1, v5, vcc1507; GFX9-NEXT:    v_addc_co_u32_e32 v28, vcc, -1, v6, vcc1508; GFX9-NEXT:    v_mov_b32_e32 v14, 01509; GFX9-NEXT:    v_mov_b32_e32 v20, 01510; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, -1, v7, vcc1511; GFX9-NEXT:    v_mov_b32_e32 v15, 01512; GFX9-NEXT:    s_mov_b64 s[4:5], 01513; GFX9-NEXT:    v_mov_b32_e32 v21, 01514; GFX9-NEXT:    v_mov_b32_e32 v13, 01515; GFX9-NEXT:  .LBB1_3: ; %udiv-do-while1516; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=11517; GFX9-NEXT:    v_lshlrev_b64 v[30:31], 1, v[10:11]1518; GFX9-NEXT:    v_lshrrev_b32_e32 v12, 31, v111519; GFX9-NEXT:    v_or_b32_e32 v10, v20, v301520; GFX9-NEXT:    v_lshrrev_b32_e32 v20, 31, v171521; GFX9-NEXT:    v_lshlrev_b64 v[16:17], 1, v[16:17]1522; GFX9-NEXT:    v_or_b32_e32 v11, v21, v311523; GFX9-NEXT:    v_lshlrev_b64 v[18:19], 1, v[18:19]1524; GFX9-NEXT:    v_lshrrev_b32_e32 v21, 31, v91525; GFX9-NEXT:    v_or_b32_e32 v16, v16, v211526; GFX9-NEXT:    v_or_b32_e32 v18, v18, v201527; GFX9-NEXT:    v_sub_co_u32_e32 v20, vcc, v26, v161528; GFX9-NEXT:    v_subb_co_u32_e32 v20, vcc, v27, v17, vcc1529; GFX9-NEXT:    v_subb_co_u32_e32 v20, vcc, v28, v18, vcc1530; GFX9-NEXT:    v_subb_co_u32_e32 v20, vcc, v29, v19, vcc1531; GFX9-NEXT:    v_ashrrev_i32_e32 v30, 31, v201532; GFX9-NEXT:    v_and_b32_e32 v20, v30, v41533; GFX9-NEXT:    v_lshlrev_b64 v[8:9], 1, v[8:9]1534; GFX9-NEXT:    v_sub_co_u32_e32 v16, vcc, v16, v201535; GFX9-NEXT:    v_and_b32_e32 v20, v30, v51536; GFX9-NEXT:    v_subb_co_u32_e32 v17, vcc, v17, v20, vcc1537; GFX9-NEXT:    v_or3_b32 v8, v8, v12, v141538; GFX9-NEXT:    v_and_b32_e32 v12, v30, v61539; GFX9-NEXT:    v_and_b32_e32 v20, v30, v71540; GFX9-NEXT:    v_subb_co_u32_e32 v18, vcc, v18, v12, vcc1541; GFX9-NEXT:    v_subb_co_u32_e32 v19, vcc, v19, v20, vcc1542; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, -1, v221543; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, -1, v23, vcc1544; GFX9-NEXT:    v_addc_co_u32_e32 v24, vcc, -1, v24, vcc1545; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, -1, v25, vcc1546; GFX9-NEXT:    v_or_b32_e32 v20, v22, v241547; GFX9-NEXT:    v_or_b32_e32 v21, v23, v251548; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[20:21]1549; GFX9-NEXT:    v_and_b32_e32 v12, 1, v301550; GFX9-NEXT:    v_mov_b32_e32 v21, v131551; GFX9-NEXT:    v_or3_b32 v9, v9, 0, v151552; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]1553; GFX9-NEXT:    v_mov_b32_e32 v20, v121554; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]1555; GFX9-NEXT:    s_cbranch_execnz .LBB1_31556; GFX9-NEXT:  ; %bb.4: ; %Flow1557; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]1558; GFX9-NEXT:  .LBB1_5: ; %Flow21559; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]1560; GFX9-NEXT:    v_lshlrev_b64 v[16:17], 1, v[10:11]1561; GFX9-NEXT:    v_lshlrev_b64 v[8:9], 1, v[8:9]1562; GFX9-NEXT:    v_lshrrev_b32_e32 v10, 31, v111563; GFX9-NEXT:    v_or3_b32 v15, v9, 0, v151564; GFX9-NEXT:    v_or3_b32 v14, v8, v10, v141565; GFX9-NEXT:    v_or_b32_e32 v13, v13, v171566; GFX9-NEXT:    v_or_b32_e32 v12, v12, v161567; GFX9-NEXT:  .LBB1_6: ; %Flow31568; GFX9-NEXT:    s_or_b64 exec, exec, s[8:9]1569; GFX9-NEXT:    v_mul_lo_u32 v19, v12, v71570; GFX9-NEXT:    v_mad_u64_u32 v[7:8], s[4:5], v4, v12, 01571; GFX9-NEXT:    v_mov_b32_e32 v9, 01572; GFX9-NEXT:    v_mul_lo_u32 v18, v13, v61573; GFX9-NEXT:    v_mad_u64_u32 v[16:17], s[4:5], v5, v12, v[8:9]1574; GFX9-NEXT:    v_mad_u64_u32 v[10:11], s[4:5], v12, v6, 01575; GFX9-NEXT:    v_mov_b32_e32 v8, v161576; GFX9-NEXT:    v_mad_u64_u32 v[8:9], s[4:5], v4, v13, v[8:9]1577; GFX9-NEXT:    v_add3_u32 v11, v11, v19, v181578; GFX9-NEXT:    v_mad_u64_u32 v[10:11], s[4:5], v14, v4, v[10:11]1579; GFX9-NEXT:    v_mul_lo_u32 v6, v14, v51580; GFX9-NEXT:    v_add_co_u32_e32 v14, vcc, v17, v91581; GFX9-NEXT:    v_mul_lo_u32 v12, v15, v41582; GFX9-NEXT:    v_addc_co_u32_e64 v15, s[4:5], 0, 0, vcc1583; GFX9-NEXT:    v_mad_u64_u32 v[4:5], s[4:5], v5, v13, v[14:15]1584; GFX9-NEXT:    v_add3_u32 v6, v12, v11, v61585; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, v4, v101586; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, v5, v6, vcc1587; GFX9-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v71588; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v8, vcc1589; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v4, vcc1590; GFX9-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v5, vcc1591; GFX9-NEXT:    s_setpc_b64 s[30:31]1592;1593; GFX9-O0-LABEL: v_urem_i128_vv:1594; GFX9-O0:       ; %bb.0: ; %_udiv-special-cases1595; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1596; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[4:5], -11597; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill1598; GFX9-O0-NEXT:    s_mov_b64 exec, s[4:5]1599; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v61600; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill1601; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v21602; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v01603; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload1604; GFX9-O0-NEXT:    ; kill: def $vgpr9 killed $vgpr9 def $vgpr9_vgpr10 killed $exec1605; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v11606; GFX9-O0-NEXT:    ; kill: def $vgpr12 killed $vgpr12 def $vgpr12_vgpr13 killed $exec1607; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v31608; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1609; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v51610; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec1611; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v71612; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v121613; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v131614; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill1615; GFX9-O0-NEXT:    s_nop 01616; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill1617; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v91618; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v101619; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill1620; GFX9-O0-NEXT:    s_nop 01621; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill1622; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v141623; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v151624; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill1625; GFX9-O0-NEXT:    s_nop 01626; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill1627; GFX9-O0-NEXT:    s_waitcnt vmcnt(6)1628; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v11629; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v01630; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill1631; GFX9-O0-NEXT:    s_nop 01632; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill1633; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v141634; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v151635; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill1636; GFX9-O0-NEXT:    s_nop 01637; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill1638; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v11639; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v01640; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill1641; GFX9-O0-NEXT:    s_nop 01642; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill1643; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v121644; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v131645; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill1646; GFX9-O0-NEXT:    s_nop 01647; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill1648; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v91649; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v101650; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill1651; GFX9-O0-NEXT:    s_nop 01652; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill1653; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v151654; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v11655; GFX9-O0-NEXT:    v_or_b32_e64 v2, v8, v61656; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v141657; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v01658; GFX9-O0-NEXT:    v_or_b32_e64 v0, v5, v41659; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1660; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v21661; GFX9-O0-NEXT:    s_mov_b64 s[6:7], 01662; GFX9-O0-NEXT:    ; implicit-def: $vgpr30 : SGPR spill to VGPR lane1663; GFX9-O0-NEXT:    v_writelane_b32 v30, s6, 01664; GFX9-O0-NEXT:    v_writelane_b32 v30, s7, 11665; GFX9-O0-NEXT:    v_cmp_eq_u64_e64 s[4:5], v[0:1], s[6:7]1666; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v131667; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v101668; GFX9-O0-NEXT:    v_or_b32_e64 v7, v3, v11669; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v121670; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v91671; GFX9-O0-NEXT:    v_or_b32_e64 v9, v2, v01672; GFX9-O0-NEXT:    ; kill: def $vgpr9 killed $vgpr9 def $vgpr9_vgpr10 killed $exec1673; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v71674; GFX9-O0-NEXT:    v_cmp_eq_u64_e64 s[8:9], v[9:10], s[6:7]1675; GFX9-O0-NEXT:    s_or_b64 s[4:5], s[4:5], s[8:9]1676; GFX9-O0-NEXT:    v_ffbh_u32_e64 v4, v41677; GFX9-O0-NEXT:    s_mov_b32 s8, 321678; GFX9-O0-NEXT:    v_add_u32_e64 v4, v4, s81679; GFX9-O0-NEXT:    v_ffbh_u32_e64 v6, v61680; GFX9-O0-NEXT:    v_min_u32_e64 v6, v4, v61681; GFX9-O0-NEXT:    v_mov_b32_e32 v4, 01682; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec1683; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v41684; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v71685; GFX9-O0-NEXT:    v_ffbh_u32_e64 v5, v51686; GFX9-O0-NEXT:    v_add_u32_e64 v5, v5, s81687; GFX9-O0-NEXT:    v_ffbh_u32_e64 v8, v81688; GFX9-O0-NEXT:    v_min_u32_e64 v16, v5, v81689; GFX9-O0-NEXT:    ; kill: def $vgpr16 killed $vgpr16 def $vgpr16_vgpr17 killed $exec1690; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v41691; GFX9-O0-NEXT:    s_mov_b64 s[10:11], 641692; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v161693; GFX9-O0-NEXT:    s_mov_b32 s12, s101694; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v171695; GFX9-O0-NEXT:    s_mov_b32 s9, s111696; GFX9-O0-NEXT:    v_add_co_u32_e64 v8, s[12:13], v8, s121697; GFX9-O0-NEXT:    v_mov_b32_e32 v9, s91698; GFX9-O0-NEXT:    v_addc_co_u32_e64 v5, s[12:13], v5, v9, s[12:13]1699; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec1700; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v51701; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v91702; GFX9-O0-NEXT:    s_mov_b64 s[12:13], s[6:7]1703; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[12:13], v[14:15], s[12:13]1704; GFX9-O0-NEXT:    v_cndmask_b32_e64 v5, v5, v10, s[12:13]1705; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v61706; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v81707; GFX9-O0-NEXT:    v_cndmask_b32_e64 v8, v6, v7, s[12:13]1708; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec1709; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v51710; GFX9-O0-NEXT:    v_ffbh_u32_e64 v5, v01711; GFX9-O0-NEXT:    v_add_u32_e64 v5, v5, s81712; GFX9-O0-NEXT:    v_ffbh_u32_e64 v6, v11713; GFX9-O0-NEXT:    v_min_u32_e64 v5, v5, v61714; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec1715; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v41716; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v61717; GFX9-O0-NEXT:    v_ffbh_u32_e64 v10, v21718; GFX9-O0-NEXT:    v_add_u32_e64 v10, v10, s81719; GFX9-O0-NEXT:    v_ffbh_u32_e64 v11, v31720; GFX9-O0-NEXT:    v_min_u32_e64 v14, v10, v111721; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec1722; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v41723; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v141724; GFX9-O0-NEXT:    s_mov_b32 s8, s101725; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v151726; GFX9-O0-NEXT:    s_mov_b32 s10, s111727; GFX9-O0-NEXT:    v_add_co_u32_e64 v10, s[8:9], v10, s81728; GFX9-O0-NEXT:    v_mov_b32_e32 v11, s101729; GFX9-O0-NEXT:    v_addc_co_u32_e64 v4, s[8:9], v4, v11, s[8:9]1730; GFX9-O0-NEXT:    ; kill: def $vgpr10 killed $vgpr10 def $vgpr10_vgpr11 killed $exec1731; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v41732; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v111733; GFX9-O0-NEXT:    s_mov_b64 s[8:9], s[6:7]1734; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[8:9], v[12:13], s[8:9]1735; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v4, v7, s[8:9]1736; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v51737; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v101738; GFX9-O0-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s[8:9]1739; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec1740; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v41741; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v51742; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v81743; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr5_vgpr6 killed $exec1744; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v91745; GFX9-O0-NEXT:    s_mov_b32 s10, s61746; GFX9-O0-NEXT:    s_mov_b32 s11, s71747; GFX9-O0-NEXT:    v_sub_co_u32_e32 v4, vcc, v4, v71748; GFX9-O0-NEXT:    v_subb_co_u32_e32 v8, vcc, v5, v6, vcc1749; GFX9-O0-NEXT:    v_mov_b32_e32 v6, s101750; GFX9-O0-NEXT:    v_mov_b32_e32 v5, s101751; GFX9-O0-NEXT:    v_subb_co_u32_e32 v7, vcc, v5, v6, vcc1752; GFX9-O0-NEXT:    v_mov_b32_e32 v6, s111753; GFX9-O0-NEXT:    v_mov_b32_e32 v5, s111754; GFX9-O0-NEXT:    v_subb_co_u32_e32 v6, vcc, v5, v6, vcc1755; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec1756; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v81757; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill1758; GFX9-O0-NEXT:    s_nop 01759; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill1760; GFX9-O0-NEXT:    ; kill: def $vgpr7 killed $vgpr7 def $vgpr7_vgpr8 killed $exec1761; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v61762; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill1763; GFX9-O0-NEXT:    s_nop 01764; GFX9-O0-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill1765; GFX9-O0-NEXT:    v_cmp_eq_u64_e64 s[8:9], v[7:8], s[6:7]1766; GFX9-O0-NEXT:    s_mov_b64 s[12:13], 0x7f1767; GFX9-O0-NEXT:    v_cmp_gt_u64_e64 s[14:15], v[4:5], s[12:13]1768; GFX9-O0-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[14:15]1769; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[14:15], v[7:8], s[6:7]1770; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[14:15]1771; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, v6, v9, s[8:9]1772; GFX9-O0-NEXT:    v_and_b32_e64 v6, 1, v61773; GFX9-O0-NEXT:    v_cmp_eq_u32_e64 s[8:9], v6, 11774; GFX9-O0-NEXT:    s_or_b64 s[8:9], s[4:5], s[8:9]1775; GFX9-O0-NEXT:    s_mov_b64 s[4:5], -11776; GFX9-O0-NEXT:    s_xor_b64 s[4:5], s[8:9], s[4:5]1777; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v51778; GFX9-O0-NEXT:    s_mov_b32 s14, s131779; GFX9-O0-NEXT:    v_xor_b32_e64 v6, v6, s141780; GFX9-O0-NEXT:    ; kill: def $sgpr12 killed $sgpr12 killed $sgpr12_sgpr131781; GFX9-O0-NEXT:    v_xor_b32_e64 v4, v4, s121782; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec1783; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v61784; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v51785; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v81786; GFX9-O0-NEXT:    v_or_b32_e64 v6, v6, v91787; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 killed $vgpr4_vgpr5 killed $exec1788; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v71789; GFX9-O0-NEXT:    v_or_b32_e64 v4, v4, v51790; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec1791; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v61792; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[6:7], v[4:5], s[6:7]1793; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s111794; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v1, v4, s[8:9]1795; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s101796; GFX9-O0-NEXT:    v_cndmask_b32_e64 v0, v0, v1, s[8:9]1797; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1798; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v41799; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s111800; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v3, v4, s[8:9]1801; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s101802; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[8:9]1803; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec1804; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v41805; GFX9-O0-NEXT:    s_and_b64 s[6:7], s[4:5], s[6:7]1806; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill1807; GFX9-O0-NEXT:    s_nop 01808; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill1809; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill1810; GFX9-O0-NEXT:    s_nop 01811; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill1812; GFX9-O0-NEXT:    s_mov_b64 s[4:5], exec1813; GFX9-O0-NEXT:    v_writelane_b32 v30, s4, 21814; GFX9-O0-NEXT:    v_writelane_b32 v30, s5, 31815; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -11816; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill1817; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]1818; GFX9-O0-NEXT:    s_and_b64 s[4:5], s[4:5], s[6:7]1819; GFX9-O0-NEXT:    s_mov_b64 exec, s[4:5]1820; GFX9-O0-NEXT:    s_cbranch_execz .LBB1_31821; GFX9-O0-NEXT:    s_branch .LBB1_81822; GFX9-O0-NEXT:  .LBB1_1: ; %Flow1823; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -11824; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload1825; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]1826; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)1827; GFX9-O0-NEXT:    v_readlane_b32 s4, v30, 41828; GFX9-O0-NEXT:    v_readlane_b32 s5, v30, 51829; GFX9-O0-NEXT:    s_or_b64 exec, exec, s[4:5]1830; GFX9-O0-NEXT:  ; %bb.2: ; %Flow1831; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload1832; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload1833; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload1834; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload1835; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload1836; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload1837; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload1838; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload1839; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)1840; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill1841; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)1842; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill1843; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)1844; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill1845; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)1846; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill1847; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)1848; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill1849; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)1850; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill1851; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)1852; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill1853; GFX9-O0-NEXT:    s_waitcnt vmcnt(7)1854; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill1855; GFX9-O0-NEXT:    s_branch .LBB1_51856; GFX9-O0-NEXT:  .LBB1_3: ; %Flow21857; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -11858; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload1859; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]1860; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)1861; GFX9-O0-NEXT:    v_readlane_b32 s4, v30, 21862; GFX9-O0-NEXT:    v_readlane_b32 s5, v30, 31863; GFX9-O0-NEXT:    s_or_b64 exec, exec, s[4:5]1864; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload1865; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload1866; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload1867; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload1868; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)1869; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill1870; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)1871; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill1872; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill1873; GFX9-O0-NEXT:    s_nop 01874; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill1875; GFX9-O0-NEXT:    s_branch .LBB1_91876; GFX9-O0-NEXT:  .LBB1_4: ; %udiv-loop-exit1877; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload1878; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload1879; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload1880; GFX9-O0-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload1881; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload1882; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload1883; GFX9-O0-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload1884; GFX9-O0-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload1885; GFX9-O0-NEXT:    s_mov_b32 s4, 11886; GFX9-O0-NEXT:    s_waitcnt vmcnt(2)1887; GFX9-O0-NEXT:    v_lshlrev_b64 v[2:3], s4, v[0:1]1888; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)1889; GFX9-O0-NEXT:    v_lshlrev_b64 v[9:10], s4, v[9:10]1890; GFX9-O0-NEXT:    s_mov_b32 s4, 631891; GFX9-O0-NEXT:    v_lshrrev_b64 v[0:1], s4, v[0:1]1892; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v11893; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v101894; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v81895; GFX9-O0-NEXT:    v_or3_b32 v4, v4, v11, v121896; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v01897; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v91898; GFX9-O0-NEXT:    v_or3_b32 v0, v0, v1, v71899; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec1900; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v41901; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v31902; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v61903; GFX9-O0-NEXT:    v_or_b32_e64 v4, v4, v71904; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v21905; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v51906; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v31907; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec1908; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v41909; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill1910; GFX9-O0-NEXT:    s_nop 01911; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill1912; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill1913; GFX9-O0-NEXT:    s_nop 01914; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill1915; GFX9-O0-NEXT:    s_branch .LBB1_31916; GFX9-O0-NEXT:  .LBB1_5: ; %Flow11917; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -11918; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload1919; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]1920; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)1921; GFX9-O0-NEXT:    v_readlane_b32 s4, v30, 61922; GFX9-O0-NEXT:    v_readlane_b32 s5, v30, 71923; GFX9-O0-NEXT:    s_or_b64 exec, exec, s[4:5]1924; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload1925; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload1926; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload1927; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload1928; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload1929; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload1930; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload1931; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload1932; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)1933; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill1934; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)1935; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill1936; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill1937; GFX9-O0-NEXT:    s_nop 01938; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill1939; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill1940; GFX9-O0-NEXT:    s_nop 01941; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill1942; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill1943; GFX9-O0-NEXT:    s_nop 01944; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill1945; GFX9-O0-NEXT:    s_branch .LBB1_41946; GFX9-O0-NEXT:  .LBB1_6: ; %udiv-do-while1947; GFX9-O0-NEXT:    ; =>This Inner Loop Header: Depth=11948; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -11949; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload1950; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]1951; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)1952; GFX9-O0-NEXT:    v_readlane_b32 s6, v30, 81953; GFX9-O0-NEXT:    v_readlane_b32 s7, v30, 91954; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload1955; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload1956; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload1957; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload1958; GFX9-O0-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload1959; GFX9-O0-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload1960; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload1961; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload1962; GFX9-O0-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload1963; GFX9-O0-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload1964; GFX9-O0-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload1965; GFX9-O0-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload1966; GFX9-O0-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload1967; GFX9-O0-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload1968; GFX9-O0-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload1969; GFX9-O0-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload1970; GFX9-O0-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload1971; GFX9-O0-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload1972; GFX9-O0-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload1973; GFX9-O0-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload1974; GFX9-O0-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload1975; GFX9-O0-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload1976; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload1977; GFX9-O0-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload1978; GFX9-O0-NEXT:    s_mov_b32 s4, 631979; GFX9-O0-NEXT:    s_waitcnt vmcnt(16)1980; GFX9-O0-NEXT:    v_lshrrev_b64 v[28:29], s4, v[2:3]1981; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v291982; GFX9-O0-NEXT:    s_mov_b32 s5, 11983; GFX9-O0-NEXT:    v_lshlrev_b64 v[22:23], s5, v[22:23]1984; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v231985; GFX9-O0-NEXT:    v_or_b32_e64 v4, v4, v51986; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v281987; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v221988; GFX9-O0-NEXT:    v_or_b32_e64 v22, v5, v101989; GFX9-O0-NEXT:    ; kill: def $vgpr22 killed $vgpr22 def $vgpr22_vgpr23 killed $exec1990; GFX9-O0-NEXT:    v_mov_b32_e32 v23, v41991; GFX9-O0-NEXT:    v_lshlrev_b64 v[28:29], s5, v[2:3]1992; GFX9-O0-NEXT:    v_lshrrev_b64 v[4:5], s4, v[6:7]1993; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v291994; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v51995; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v31996; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v281997; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 killed $vgpr4_vgpr5 killed $exec1998; GFX9-O0-NEXT:    v_or_b32_e64 v4, v3, v41999; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec2000; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v22001; GFX9-O0-NEXT:    v_lshlrev_b64 v[2:3], s5, v[0:1]2002; GFX9-O0-NEXT:    v_lshlrev_b64 v[28:29], s5, v[6:7]2003; GFX9-O0-NEXT:    v_lshrrev_b64 v[0:1], s4, v[0:1]2004; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v12005; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v292006; GFX9-O0-NEXT:    s_waitcnt vmcnt(10)2007; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v272008; GFX9-O0-NEXT:    v_or3_b32 v6, v6, v7, v102009; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v02010; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v282011; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v262012; GFX9-O0-NEXT:    v_or3_b32 v0, v0, v1, v72013; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec2014; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v62015; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v32016; GFX9-O0-NEXT:    s_waitcnt vmcnt(8)2017; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v252018; GFX9-O0-NEXT:    v_or_b32_e64 v6, v6, v72019; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v22020; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v242021; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v32022; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec2023; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v62024; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v42025; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v52026; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v222027; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v232028; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)2029; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v112030; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v142031; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v152032; GFX9-O0-NEXT:    v_sub_co_u32_e32 v13, vcc, v13, v62033; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)2034; GFX9-O0-NEXT:    v_subb_co_u32_e32 v12, vcc, v12, v10, vcc2035; GFX9-O0-NEXT:    v_subb_co_u32_e32 v11, vcc, v11, v4, vcc2036; GFX9-O0-NEXT:    v_subb_co_u32_e32 v7, vcc, v7, v5, vcc2037; GFX9-O0-NEXT:    ; kill: def $vgpr11 killed $vgpr11 def $vgpr11_vgpr12 killed $exec2038; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v72039; GFX9-O0-NEXT:    v_ashrrev_i64 v[13:14], s4, v[11:12]2040; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v142041; GFX9-O0-NEXT:    s_mov_b64 s[4:5], 12042; GFX9-O0-NEXT:    s_mov_b32 s8, s52043; GFX9-O0-NEXT:    v_and_b32_e64 v12, v7, s82044; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v132045; GFX9-O0-NEXT:    ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr52046; GFX9-O0-NEXT:    v_and_b32_e64 v14, v11, s42047; GFX9-O0-NEXT:    ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec2048; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v122049; GFX9-O0-NEXT:    v_mov_b32_e32 v12, 02050; GFX9-O0-NEXT:    v_mov_b32_e32 v13, 02051; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v212052; GFX9-O0-NEXT:    v_and_b32_e64 v22, v7, v222053; GFX9-O0-NEXT:    v_and_b32_e64 v20, v11, v202054; GFX9-O0-NEXT:    ; kill: def $vgpr20 killed $vgpr20 def $vgpr20_vgpr21 killed $exec2055; GFX9-O0-NEXT:    v_mov_b32_e32 v21, v222056; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v192057; GFX9-O0-NEXT:    v_and_b32_e64 v7, v7, v222058; GFX9-O0-NEXT:    v_and_b32_e64 v22, v11, v182059; GFX9-O0-NEXT:    ; kill: def $vgpr22 killed $vgpr22 def $vgpr22_vgpr23 killed $exec2060; GFX9-O0-NEXT:    v_mov_b32_e32 v23, v72061; GFX9-O0-NEXT:    v_mov_b32_e32 v19, v222062; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v232063; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v202064; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v212065; GFX9-O0-NEXT:    v_sub_co_u32_e32 v6, vcc, v6, v192066; GFX9-O0-NEXT:    v_subb_co_u32_e32 v10, vcc, v10, v18, vcc2067; GFX9-O0-NEXT:    v_subb_co_u32_e32 v4, vcc, v4, v11, vcc2068; GFX9-O0-NEXT:    v_subb_co_u32_e32 v7, vcc, v5, v7, vcc2069; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec2070; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v72071; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec2072; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v102073; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v82074; GFX9-O0-NEXT:    ; kill: def $vgpr9 killed $vgpr9 killed $vgpr8_vgpr9 killed $exec2075; GFX9-O0-NEXT:    s_mov_b64 s[8:9], -12076; GFX9-O0-NEXT:    s_mov_b32 s5, s82077; GFX9-O0-NEXT:    s_mov_b32 s4, s92078; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v162079; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v172080; GFX9-O0-NEXT:    v_mov_b32_e32 v16, s52081; GFX9-O0-NEXT:    v_add_co_u32_e32 v19, vcc, v11, v162082; GFX9-O0-NEXT:    v_mov_b32_e32 v11, s42083; GFX9-O0-NEXT:    v_addc_co_u32_e32 v9, vcc, v9, v11, vcc2084; GFX9-O0-NEXT:    v_mov_b32_e32 v11, s52085; GFX9-O0-NEXT:    v_addc_co_u32_e32 v16, vcc, v10, v11, vcc2086; GFX9-O0-NEXT:    v_mov_b32_e32 v10, s42087; GFX9-O0-NEXT:    v_addc_co_u32_e32 v8, vcc, v8, v10, vcc2088; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec2089; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v92090; GFX9-O0-NEXT:    ; kill: def $vgpr16 killed $vgpr16 def $vgpr16_vgpr17 killed $exec2091; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v82092; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v162093; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v172094; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v192095; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v202096; GFX9-O0-NEXT:    v_mov_b32_e32 v21, v172097; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v202098; GFX9-O0-NEXT:    v_or_b32_e64 v18, v18, v212099; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v162100; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v192101; GFX9-O0-NEXT:    v_or_b32_e64 v16, v16, v172102; GFX9-O0-NEXT:    ; kill: def $vgpr16 killed $vgpr16 def $vgpr16_vgpr17 killed $exec2103; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v182104; GFX9-O0-NEXT:    v_cmp_eq_u64_e64 s[4:5], v[16:17], v[12:13]2105; GFX9-O0-NEXT:    s_or_b64 s[4:5], s[4:5], s[6:7]2106; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v32107; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v22108; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill2109; GFX9-O0-NEXT:    s_nop 02110; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill2111; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v12112; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v02113; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill2114; GFX9-O0-NEXT:    s_nop 02115; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill2116; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v152117; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v142118; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill2119; GFX9-O0-NEXT:    s_nop 02120; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill2121; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v132122; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v122123; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill2124; GFX9-O0-NEXT:    s_nop 02125; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill2126; GFX9-O0-NEXT:    s_mov_b64 s[6:7], s[4:5]2127; GFX9-O0-NEXT:    v_writelane_b32 v30, s6, 42128; GFX9-O0-NEXT:    v_writelane_b32 v30, s7, 52129; GFX9-O0-NEXT:    s_mov_b64 s[6:7], s[4:5]2130; GFX9-O0-NEXT:    v_writelane_b32 v30, s6, 82131; GFX9-O0-NEXT:    v_writelane_b32 v30, s7, 92132; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -12133; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill2134; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]2135; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill2136; GFX9-O0-NEXT:    s_nop 02137; GFX9-O0-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill2138; GFX9-O0-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill2139; GFX9-O0-NEXT:    s_nop 02140; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill2141; GFX9-O0-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill2142; GFX9-O0-NEXT:    s_nop 02143; GFX9-O0-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill2144; GFX9-O0-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill2145; GFX9-O0-NEXT:    s_nop 02146; GFX9-O0-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill2147; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill2148; GFX9-O0-NEXT:    s_nop 02149; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill2150; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill2151; GFX9-O0-NEXT:    s_nop 02152; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill2153; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill2154; GFX9-O0-NEXT:    s_nop 02155; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill2156; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill2157; GFX9-O0-NEXT:    s_nop 02158; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill2159; GFX9-O0-NEXT:    s_andn2_b64 exec, exec, s[4:5]2160; GFX9-O0-NEXT:    s_cbranch_execnz .LBB1_62161; GFX9-O0-NEXT:    s_branch .LBB1_12162; GFX9-O0-NEXT:  .LBB1_7: ; %udiv-preheader2163; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -12164; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload2165; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]2166; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload2167; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload2168; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload2169; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload2170; GFX9-O0-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload2171; GFX9-O0-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload2172; GFX9-O0-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload2173; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload2174; GFX9-O0-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload2175; GFX9-O0-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload2176; GFX9-O0-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload2177; GFX9-O0-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload2178; GFX9-O0-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload2179; GFX9-O0-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload2180; GFX9-O0-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload2181; GFX9-O0-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload2182; GFX9-O0-NEXT:    s_waitcnt vmcnt(9)2183; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v102184; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)2185; GFX9-O0-NEXT:    v_lshrrev_b64 v[6:7], v4, v[18:19]2186; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v72187; GFX9-O0-NEXT:    s_mov_b32 s4, 642188; GFX9-O0-NEXT:    v_sub_u32_e64 v20, s4, v42189; GFX9-O0-NEXT:    v_lshlrev_b64 v[20:21], v20, v[14:15]2190; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v212191; GFX9-O0-NEXT:    v_or_b32_e64 v5, v5, v222192; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec2193; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v202194; GFX9-O0-NEXT:    v_or_b32_e64 v6, v6, v72195; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec2196; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v52197; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v72198; GFX9-O0-NEXT:    v_cmp_lt_u32_e64 s[6:7], v4, s42199; GFX9-O0-NEXT:    v_sub_u32_e64 v5, v4, s42200; GFX9-O0-NEXT:    v_lshrrev_b64 v[20:21], v5, v[14:15]2201; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v212202; GFX9-O0-NEXT:    v_cndmask_b32_e64 v5, v5, v22, s[6:7]2203; GFX9-O0-NEXT:    s_mov_b32 s4, 02204; GFX9-O0-NEXT:    v_cmp_eq_u32_e64 s[4:5], v4, s42205; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v192206; GFX9-O0-NEXT:    v_cndmask_b32_e64 v5, v5, v22, s[4:5]2207; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v62208; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v202209; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, v6, v7, s[6:7]2210; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v182211; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, v6, v7, s[4:5]2212; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec2213; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v52214; GFX9-O0-NEXT:    v_lshrrev_b64 v[4:5], v4, v[14:15]2215; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v52216; GFX9-O0-NEXT:    s_mov_b64 s[4:5], 02217; GFX9-O0-NEXT:    s_mov_b32 s8, s52218; GFX9-O0-NEXT:    v_mov_b32_e32 v14, s82219; GFX9-O0-NEXT:    v_cndmask_b32_e64 v14, v14, v15, s[6:7]2220; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v42221; GFX9-O0-NEXT:    s_mov_b32 s8, s42222; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s82223; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v4, v5, s[6:7]2224; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec2225; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v142226; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v122227; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v132228; GFX9-O0-NEXT:    s_mov_b64 s[8:9], -12229; GFX9-O0-NEXT:    s_mov_b32 s7, s82230; GFX9-O0-NEXT:    s_mov_b32 s6, s92231; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v162232; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v172233; GFX9-O0-NEXT:    v_mov_b32_e32 v16, s72234; GFX9-O0-NEXT:    v_add_co_u32_e32 v16, vcc, v15, v162235; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s62236; GFX9-O0-NEXT:    v_addc_co_u32_e32 v12, vcc, v12, v15, vcc2237; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s72238; GFX9-O0-NEXT:    v_addc_co_u32_e32 v18, vcc, v14, v15, vcc2239; GFX9-O0-NEXT:    v_mov_b32_e32 v14, s62240; GFX9-O0-NEXT:    v_addc_co_u32_e32 v13, vcc, v13, v14, vcc2241; GFX9-O0-NEXT:    ; kill: def $vgpr18 killed $vgpr18 def $vgpr18_vgpr19 killed $exec2242; GFX9-O0-NEXT:    v_mov_b32_e32 v19, v132243; GFX9-O0-NEXT:    ; kill: def $vgpr16 killed $vgpr16 def $vgpr16_vgpr17 killed $exec2244; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v122245; GFX9-O0-NEXT:    v_mov_b32_e32 v13, s52246; GFX9-O0-NEXT:    v_mov_b32_e32 v12, s42247; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s52248; GFX9-O0-NEXT:    v_mov_b32_e32 v14, s42249; GFX9-O0-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill2250; GFX9-O0-NEXT:    s_nop 02251; GFX9-O0-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill2252; GFX9-O0-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill2253; GFX9-O0-NEXT:    s_nop 02254; GFX9-O0-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill2255; GFX9-O0-NEXT:    v_writelane_b32 v30, s4, 82256; GFX9-O0-NEXT:    v_writelane_b32 v30, s5, 92257; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -12258; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill2259; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]2260; GFX9-O0-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill2261; GFX9-O0-NEXT:    s_nop 02262; GFX9-O0-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill2263; GFX9-O0-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill2264; GFX9-O0-NEXT:    s_nop 02265; GFX9-O0-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill2266; GFX9-O0-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill2267; GFX9-O0-NEXT:    s_nop 02268; GFX9-O0-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill2269; GFX9-O0-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill2270; GFX9-O0-NEXT:    s_nop 02271; GFX9-O0-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill2272; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill2273; GFX9-O0-NEXT:    s_nop 02274; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill2275; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill2276; GFX9-O0-NEXT:    s_nop 02277; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill2278; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill2279; GFX9-O0-NEXT:    s_nop 02280; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill2281; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill2282; GFX9-O0-NEXT:    s_nop 02283; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill2284; GFX9-O0-NEXT:    s_branch .LBB1_62285; GFX9-O0-NEXT:  .LBB1_8: ; %udiv-bb12286; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -12287; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 ; 4-byte Folded Reload2288; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]2289; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload2290; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload2291; GFX9-O0-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload2292; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload2293; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload2294; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload2295; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload2296; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload2297; GFX9-O0-NEXT:    s_mov_b64 s[6:7], 12298; GFX9-O0-NEXT:    s_mov_b32 s5, s62299; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)2300; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v02301; GFX9-O0-NEXT:    s_mov_b32 s4, s72302; GFX9-O0-NEXT:    s_mov_b64 s[6:7], 02303; GFX9-O0-NEXT:    s_mov_b32 s8, s62304; GFX9-O0-NEXT:    s_mov_b32 s9, s72305; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v42306; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v52307; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s52308; GFX9-O0-NEXT:    v_add_co_u32_e32 v8, vcc, v3, v42309; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s42310; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)2311; GFX9-O0-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v4, vcc2312; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s82313; GFX9-O0-NEXT:    v_addc_co_u32_e32 v0, vcc, v0, v4, vcc2314; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s92315; GFX9-O0-NEXT:    v_addc_co_u32_e32 v2, vcc, v2, v4, vcc2316; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec2317; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v12318; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec2319; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v22320; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v12321; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v02322; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill2323; GFX9-O0-NEXT:    s_nop 02324; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill2325; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v82326; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v92327; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill2328; GFX9-O0-NEXT:    s_nop 02329; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill2330; GFX9-O0-NEXT:    s_mov_b32 s4, 0x7f2331; GFX9-O0-NEXT:    v_sub_u32_e64 v2, s4, v32332; GFX9-O0-NEXT:    v_lshlrev_b64 v[4:5], v2, v[10:11]2333; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v52334; GFX9-O0-NEXT:    s_mov_b32 s4, 642335; GFX9-O0-NEXT:    v_sub_u32_e64 v13, s4, v22336; GFX9-O0-NEXT:    v_lshrrev_b64 v[13:14], v13, v[6:7]2337; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v142338; GFX9-O0-NEXT:    v_or_b32_e64 v12, v12, v152339; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 killed $vgpr4_vgpr5 killed $exec2340; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v132341; GFX9-O0-NEXT:    v_or_b32_e64 v4, v4, v52342; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec2343; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v122344; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v52345; GFX9-O0-NEXT:    v_cmp_lt_u32_e64 s[4:5], v2, s42346; GFX9-O0-NEXT:    s_mov_b32 s10, 632347; GFX9-O0-NEXT:    v_sub_u32_e64 v3, s10, v32348; GFX9-O0-NEXT:    v_lshlrev_b64 v[12:13], v3, v[6:7]2349; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v132350; GFX9-O0-NEXT:    v_cndmask_b32_e64 v3, v3, v14, s[4:5]2351; GFX9-O0-NEXT:    s_mov_b32 s10, 02352; GFX9-O0-NEXT:    v_cmp_eq_u32_e64 s[10:11], v2, s102353; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v112354; GFX9-O0-NEXT:    v_cndmask_b32_e64 v3, v3, v14, s[10:11]2355; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v42356; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v122357; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v4, v5, s[4:5]2358; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v102359; GFX9-O0-NEXT:    v_cndmask_b32_e64 v4, v4, v5, s[10:11]2360; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec2361; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v32362; GFX9-O0-NEXT:    v_lshlrev_b64 v[6:7], v2, v[6:7]2363; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v72364; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s92365; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[4:5]2366; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec2367; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s82368; GFX9-O0-NEXT:    v_cndmask_b32_e64 v6, v3, v6, s[4:5]2369; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec2370; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v22371; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill2372; GFX9-O0-NEXT:    s_nop 02373; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill2374; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill2375; GFX9-O0-NEXT:    s_nop 02376; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill2377; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v12378; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v92379; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v32380; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v02381; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v82382; GFX9-O0-NEXT:    v_or_b32_e64 v0, v0, v12383; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec2384; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v22385; GFX9-O0-NEXT:    v_cmp_ne_u64_e64 s[4:5], v[0:1], s[6:7]2386; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s62387; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s72388; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s62389; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s72390; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill2391; GFX9-O0-NEXT:    s_nop 02392; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill2393; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill2394; GFX9-O0-NEXT:    s_nop 02395; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill2396; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill2397; GFX9-O0-NEXT:    s_nop 02398; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill2399; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill2400; GFX9-O0-NEXT:    s_nop 02401; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill2402; GFX9-O0-NEXT:    s_mov_b64 s[6:7], exec2403; GFX9-O0-NEXT:    s_and_b64 s[4:5], s[6:7], s[4:5]2404; GFX9-O0-NEXT:    s_xor_b64 s[6:7], s[4:5], s[6:7]2405; GFX9-O0-NEXT:    v_writelane_b32 v30, s6, 62406; GFX9-O0-NEXT:    v_writelane_b32 v30, s7, 72407; GFX9-O0-NEXT:    s_or_saveexec_b64 s[18:19], -12408; GFX9-O0-NEXT:    buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill2409; GFX9-O0-NEXT:    s_mov_b64 exec, s[18:19]2410; GFX9-O0-NEXT:    s_mov_b64 exec, s[4:5]2411; GFX9-O0-NEXT:    s_cbranch_execz .LBB1_52412; GFX9-O0-NEXT:    s_branch .LBB1_72413; GFX9-O0-NEXT:  .LBB1_9: ; %udiv-end2414; GFX9-O0-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload2415; GFX9-O0-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload2416; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload2417; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload2418; GFX9-O0-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload2419; GFX9-O0-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload2420; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload2421; GFX9-O0-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload2422; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload2423; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload2424; GFX9-O0-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload2425; GFX9-O0-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload2426; GFX9-O0-NEXT:    s_mov_b32 s4, 322427; GFX9-O0-NEXT:    s_waitcnt vmcnt(2)2428; GFX9-O0-NEXT:    v_lshrrev_b64 v[2:3], s4, v[4:5]2429; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v22430; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)2431; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v132432; GFX9-O0-NEXT:    v_mul_lo_u32 v10, v6, v22433; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)2434; GFX9-O0-NEXT:    v_lshrrev_b64 v[13:14], s4, v[13:14]2435; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v132436; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v42437; GFX9-O0-NEXT:    v_mul_lo_u32 v3, v7, v32438; GFX9-O0-NEXT:    v_mad_u64_u32 v[4:5], s[6:7], v7, v2, 02439; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v52440; GFX9-O0-NEXT:    v_add3_u32 v2, v2, v3, v102441; GFX9-O0-NEXT:    ; implicit-def: $sgpr52442; GFX9-O0-NEXT:    ; implicit-def: $sgpr62443; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec2444; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s52445; GFX9-O0-NEXT:    v_lshlrev_b64 v[17:18], s4, v[2:3]2446; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v182447; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v42448; GFX9-O0-NEXT:    v_mov_b32_e32 v4, 02449; GFX9-O0-NEXT:    ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14 killed $exec2450; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v42451; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v142452; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v32453; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v172454; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v132455; GFX9-O0-NEXT:    v_or_b32_e64 v13, v3, v52456; GFX9-O0-NEXT:    ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14 killed $exec2457; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v22458; GFX9-O0-NEXT:    v_lshrrev_b64 v[2:3], s4, v[15:16]2459; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v22460; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v112461; GFX9-O0-NEXT:    v_mul_lo_u32 v3, v2, v102462; GFX9-O0-NEXT:    v_lshrrev_b64 v[11:12], s4, v[11:12]2463; GFX9-O0-NEXT:    ; kill: def $vgpr11 killed $vgpr11 killed $vgpr11_vgpr12 killed $exec2464; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v152465; GFX9-O0-NEXT:    v_mul_lo_u32 v11, v11, v52466; GFX9-O0-NEXT:    v_mad_u64_u32 v[15:16], s[6:7], v2, v5, 02467; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v162468; GFX9-O0-NEXT:    v_add3_u32 v2, v2, v3, v112469; GFX9-O0-NEXT:    ; implicit-def: $sgpr52470; GFX9-O0-NEXT:    ; implicit-def: $sgpr62471; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec2472; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s52473; GFX9-O0-NEXT:    v_lshlrev_b64 v[2:3], s4, v[2:3]2474; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v32475; GFX9-O0-NEXT:    ; kill: def $vgpr15 killed $vgpr15 killed $vgpr15_vgpr16 killed $exec2476; GFX9-O0-NEXT:    ; kill: def $vgpr15 killed $vgpr15 def $vgpr15_vgpr16 killed $exec2477; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v42478; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v162479; GFX9-O0-NEXT:    v_or_b32_e64 v11, v11, v122480; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v22481; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v152482; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v32483; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec2484; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v112485; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v22486; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v132487; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v32488; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v142489; GFX9-O0-NEXT:    v_add_co_u32_e64 v13, s[6:7], v11, v122490; GFX9-O0-NEXT:    v_addc_co_u32_e64 v2, s[6:7], v2, v3, s[6:7]2491; GFX9-O0-NEXT:    ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14 killed $exec2492; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v22493; GFX9-O0-NEXT:    v_mad_u64_u32 v[15:16], s[6:7], v10, v6, 02494; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v152495; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec2496; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v42497; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v32498; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v162499; GFX9-O0-NEXT:    ; implicit-def: $sgpr52500; GFX9-O0-NEXT:    ; implicit-def: $sgpr62501; GFX9-O0-NEXT:    ; kill: def $vgpr15 killed $vgpr15 def $vgpr15_vgpr16 killed $exec2502; GFX9-O0-NEXT:    v_mov_b32_e32 v16, s52503; GFX9-O0-NEXT:    v_lshlrev_b64 v[15:16], s4, v[15:16]2504; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v162505; GFX9-O0-NEXT:    v_or_b32_e64 v11, v11, v122506; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 killed $vgpr2_vgpr3 killed $exec2507; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v152508; GFX9-O0-NEXT:    v_or_b32_e64 v2, v2, v32509; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec2510; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v112511; GFX9-O0-NEXT:    v_mad_u64_u32 v[15:16], s[6:7], v10, v7, 02512; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v152513; GFX9-O0-NEXT:    ; kill: def $vgpr11 killed $vgpr11 def $vgpr11_vgpr12 killed $exec2514; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v42515; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v122516; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v162517; GFX9-O0-NEXT:    ; implicit-def: $sgpr52518; GFX9-O0-NEXT:    ; implicit-def: $sgpr62519; GFX9-O0-NEXT:    ; kill: def $vgpr15 killed $vgpr15 def $vgpr15_vgpr16 killed $exec2520; GFX9-O0-NEXT:    v_mov_b32_e32 v16, s52521; GFX9-O0-NEXT:    v_lshlrev_b64 v[15:16], s4, v[15:16]2522; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v162523; GFX9-O0-NEXT:    v_or_b32_e64 v10, v10, v172524; GFX9-O0-NEXT:    ; kill: def $vgpr11 killed $vgpr11 killed $vgpr11_vgpr12 killed $exec2525; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v152526; GFX9-O0-NEXT:    v_or_b32_e64 v19, v11, v122527; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec2528; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v102529; GFX9-O0-NEXT:    v_mad_u64_u32 v[10:11], s[6:7], v5, v7, 02530; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v112531; GFX9-O0-NEXT:    ; kill: def $vgpr17 killed $vgpr17 def $vgpr17_vgpr18 killed $exec2532; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v42533; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v192534; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v172535; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v202536; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v182537; GFX9-O0-NEXT:    v_add_co_u32_e64 v15, s[6:7], v15, v162538; GFX9-O0-NEXT:    v_addc_co_u32_e64 v7, s[6:7], v7, v12, s[6:7]2539; GFX9-O0-NEXT:    ; kill: def $vgpr15 killed $vgpr15 def $vgpr15_vgpr16 killed $exec2540; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v72541; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v162542; GFX9-O0-NEXT:    s_mov_b64 s[6:7], 0xffffffff2543; GFX9-O0-NEXT:    s_mov_b32 s5, s72544; GFX9-O0-NEXT:    v_and_b32_e64 v7, v7, s52545; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v152546; GFX9-O0-NEXT:    s_mov_b32 s5, s62547; GFX9-O0-NEXT:    v_and_b32_e64 v17, v12, s52548; GFX9-O0-NEXT:    ; kill: def $vgpr17 killed $vgpr17 def $vgpr17_vgpr18 killed $exec2549; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v72550; GFX9-O0-NEXT:    v_mad_u64_u32 v[19:20], s[6:7], v5, v6, 02551; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v192552; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec2553; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v42554; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v72555; GFX9-O0-NEXT:    v_mov_b32_e32 v19, v202556; GFX9-O0-NEXT:    ; implicit-def: $sgpr52557; GFX9-O0-NEXT:    ; implicit-def: $sgpr62558; GFX9-O0-NEXT:    ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20 killed $exec2559; GFX9-O0-NEXT:    v_mov_b32_e32 v20, s52560; GFX9-O0-NEXT:    v_lshlrev_b64 v[19:20], s4, v[19:20]2561; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v202562; GFX9-O0-NEXT:    v_or_b32_e64 v5, v5, v122563; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec2564; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v192565; GFX9-O0-NEXT:    v_or_b32_e64 v6, v6, v72566; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec2567; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v52568; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v62569; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v172570; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v72571; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v182572; GFX9-O0-NEXT:    v_add_co_u32_e64 v5, s[6:7], v5, v122573; GFX9-O0-NEXT:    v_addc_co_u32_e64 v7, s[6:7], v6, v7, s[6:7]2574; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec2575; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v72576; GFX9-O0-NEXT:    v_lshrrev_b64 v[17:18], s4, v[5:6]2577; GFX9-O0-NEXT:    v_lshrrev_b64 v[19:20], s4, v[15:16]2578; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v192579; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v172580; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v202581; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v182582; GFX9-O0-NEXT:    v_add_co_u32_e64 v15, s[6:7], v15, v162583; GFX9-O0-NEXT:    v_addc_co_u32_e64 v7, s[6:7], v7, v12, s[6:7]2584; GFX9-O0-NEXT:    ; kill: def $vgpr15 killed $vgpr15 def $vgpr15_vgpr16 killed $exec2585; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v72586; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v22587; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v152588; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v32589; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v162590; GFX9-O0-NEXT:    v_add_co_u32_e64 v15, s[6:7], v7, v122591; GFX9-O0-NEXT:    v_addc_co_u32_e64 v2, s[6:7], v2, v3, s[6:7]2592; GFX9-O0-NEXT:    ; kill: def $vgpr15 killed $vgpr15 def $vgpr15_vgpr16 killed $exec2593; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v22594; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v152595; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v132596; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v162597; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v142598; GFX9-O0-NEXT:    v_add_co_u32_e64 v2, s[6:7], v2, v122599; GFX9-O0-NEXT:    v_addc_co_u32_e64 v7, s[6:7], v3, v7, s[6:7]2600; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec2601; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v72602; GFX9-O0-NEXT:    v_lshlrev_b64 v[5:6], s4, v[5:6]2603; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v62604; GFX9-O0-NEXT:    ; kill: def $vgpr10 killed $vgpr10 killed $vgpr10_vgpr11 killed $exec2605; GFX9-O0-NEXT:    ; kill: def $vgpr10 killed $vgpr10 def $vgpr10_vgpr11 killed $exec2606; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v42607; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v112608; GFX9-O0-NEXT:    v_or_b32_e64 v4, v4, v72609; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v52610; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v102611; GFX9-O0-NEXT:    v_or_b32_e64 v5, v5, v62612; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec2613; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v42614; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v22615; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v32616; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v52617; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr5_vgpr6 killed $exec2618; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v02619; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v12620; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v82621; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v92622; GFX9-O0-NEXT:    v_sub_co_u32_e32 v5, vcc, v5, v72623; GFX9-O0-NEXT:    v_subb_co_u32_e32 v0, vcc, v0, v6, vcc2624; GFX9-O0-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v4, vcc2625; GFX9-O0-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v2, vcc2626; GFX9-O0-NEXT:    ; kill: def $vgpr3 killed $vgpr3 def $vgpr3_vgpr4 killed $exec2627; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v12628; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec2629; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v02630; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v52631; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v32632; GFX9-O0-NEXT:    v_lshrrev_b64 v[5:6], s4, v[5:6]2633; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v52634; GFX9-O0-NEXT:    v_lshrrev_b64 v[3:4], s4, v[3:4]2635; GFX9-O0-NEXT:    ; kill: def $vgpr3 killed $vgpr3 killed $vgpr3_vgpr4 killed $exec2636; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[4:5], -12637; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload2638; GFX9-O0-NEXT:    s_mov_b64 exec, s[4:5]2639; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)2640; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]2641  %div = urem i128 %lhs, %rhs2642  ret i128 %div2643}2644 2645define i128 @v_srem_i128_v_pow2k(i128 %lhs) {2646; GFX9-LABEL: v_srem_i128_v_pow2k:2647; GFX9:       ; %bb.0:2648; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2649; GFX9-NEXT:    v_ashrrev_i32_e32 v4, 31, v32650; GFX9-NEXT:    v_mov_b32_e32 v5, v42651; GFX9-NEXT:    v_lshrrev_b64 v[4:5], 31, v[4:5]2652; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, v0, v42653; GFX9-NEXT:    v_addc_co_u32_e32 v4, vcc, v1, v5, vcc2654; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v2, vcc2655; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v3, vcc2656; GFX9-NEXT:    v_and_b32_e32 v4, -2, v42657; GFX9-NEXT:    v_subrev_co_u32_e32 v0, vcc, 0, v02658; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v4, vcc2659; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v5, vcc2660; GFX9-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v6, vcc2661; GFX9-NEXT:    s_setpc_b64 s[30:31]2662;2663; GFX9-O0-LABEL: v_srem_i128_v_pow2k:2664; GFX9-O0:       ; %bb.0:2665; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2666; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v22667; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v12668; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec2669; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v32670; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec2671; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v22672; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v02673; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v12674; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v62675; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v72676; GFX9-O0-NEXT:    s_mov_b32 s4, 632677; GFX9-O0-NEXT:    v_ashrrev_i64 v[6:7], s4, v[6:7]2678; GFX9-O0-NEXT:    s_mov_b32 s4, 312679; GFX9-O0-NEXT:    v_lshrrev_b64 v[6:7], s4, v[6:7]2680; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v62681; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v72682; GFX9-O0-NEXT:    s_mov_b64 s[6:7], 02683; GFX9-O0-NEXT:    s_mov_b32 s5, s62684; GFX9-O0-NEXT:    s_mov_b32 s4, s72685; GFX9-O0-NEXT:    v_add_co_u32_e32 v6, vcc, v5, v42686; GFX9-O0-NEXT:    v_addc_co_u32_e32 v4, vcc, v0, v2, vcc2687; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s52688; GFX9-O0-NEXT:    v_addc_co_u32_e32 v8, vcc, v3, v2, vcc2689; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s42690; GFX9-O0-NEXT:    v_addc_co_u32_e32 v2, vcc, v1, v2, vcc2691; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec2692; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v42693; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v72694; GFX9-O0-NEXT:    s_mov_b32 s6, -22695; GFX9-O0-NEXT:    s_mov_b32 s4, 02696; GFX9-O0-NEXT:    ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr52697; GFX9-O0-NEXT:    s_mov_b32 s5, s62698; GFX9-O0-NEXT:    s_mov_b32 s6, s52699; GFX9-O0-NEXT:    v_and_b32_e64 v4, v4, s62700; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec2701; GFX9-O0-NEXT:    ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr52702; GFX9-O0-NEXT:    v_and_b32_e64 v9, v6, s42703; GFX9-O0-NEXT:    ; kill: def $vgpr9 killed $vgpr9 def $vgpr9_vgpr10 killed $exec2704; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v42705; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v92706; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v102707; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec2708; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v22709; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v82710; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v92711; GFX9-O0-NEXT:    v_sub_co_u32_e32 v5, vcc, v5, v72712; GFX9-O0-NEXT:    v_subb_co_u32_e32 v0, vcc, v0, v6, vcc2713; GFX9-O0-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v4, vcc2714; GFX9-O0-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v2, vcc2715; GFX9-O0-NEXT:    ; kill: def $vgpr3 killed $vgpr3 def $vgpr3_vgpr4 killed $exec2716; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v12717; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec2718; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v02719; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v52720; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v32721; GFX9-O0-NEXT:    s_mov_b32 s4, 322722; GFX9-O0-NEXT:    v_lshrrev_b64 v[5:6], s4, v[5:6]2723; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v52724; GFX9-O0-NEXT:    v_lshrrev_b64 v[3:4], s4, v[3:4]2725; GFX9-O0-NEXT:    ; kill: def $vgpr3 killed $vgpr3 killed $vgpr3_vgpr4 killed $exec2726; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]2727  %div = srem i128 %lhs, 85899345922728  ret i128 %div2729}2730 2731define i128 @v_urem_i128_v_pow2k(i128 %lhs) {2732; GFX9-LABEL: v_urem_i128_v_pow2k:2733; GFX9:       ; %bb.0:2734; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2735; GFX9-NEXT:    v_and_b32_e32 v1, 1, v12736; GFX9-NEXT:    v_mov_b32_e32 v2, 02737; GFX9-NEXT:    v_mov_b32_e32 v3, 02738; GFX9-NEXT:    s_setpc_b64 s[30:31]2739;2740; GFX9-O0-LABEL: v_urem_i128_v_pow2k:2741; GFX9-O0:       ; %bb.0:2742; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2743; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill2744; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v22745; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v12746; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload2747; GFX9-O0-NEXT:    ; kill: def $vgpr3 killed $vgpr3 def $vgpr3_vgpr4 killed $exec2748; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr1 killed $exec2749; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec2750; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)2751; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v22752; GFX9-O0-NEXT:    s_mov_b32 s6, 12753; GFX9-O0-NEXT:    s_mov_b32 s4, -12754; GFX9-O0-NEXT:    ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr52755; GFX9-O0-NEXT:    s_mov_b32 s5, s62756; GFX9-O0-NEXT:    s_mov_b32 s6, s52757; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v12758; GFX9-O0-NEXT:    v_and_b32_e64 v3, v2, s62759; GFX9-O0-NEXT:    ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr52760; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec2761; GFX9-O0-NEXT:    v_and_b32_e64 v1, v0, s42762; GFX9-O0-NEXT:    ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec2763; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v32764; GFX9-O0-NEXT:    s_mov_b32 s4, 322765; GFX9-O0-NEXT:    v_lshrrev_b64 v[1:2], s4, v[1:2]2766; GFX9-O0-NEXT:    ; kill: def $vgpr1 killed $vgpr1 killed $vgpr1_vgpr2 killed $exec2767; GFX9-O0-NEXT:    v_mov_b32_e32 v3, 02768; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v32769; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]2770  %div = urem i128 %lhs, 85899345922771  ret i128 %div2772}2773 2774;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:2775; GFX9-SDAG: {{.*}}2776; GFX9-SDAG-O0: {{.*}}2777