brintos

brintos / llvm-project-archived public Read only

0
0
Text · 972.8 KiB · cb68a98 Raw
15993 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 %s -o - | FileCheck %s3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -mattr=-unaligned-access-mode %s -o - | FileCheck -check-prefix=ALIGNED %s4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -amdgpu-memcpy-loop-unroll=3 %s -o - | FileCheck -check-prefix=UNROLL3 %s5 6; For checking that LowerMemIntrinsics lowers memcpy and memmove with large7; constant copy-sizes into loops with multiple load/store pairs.8 9 10; memcpy for address spaces 0, 1, 4, 511 12define void @memcpy_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0) align 1 readonly %src) {13; CHECK-LABEL: memcpy_p0_p0_sz2048:14; CHECK:       ; %bb.0: ; %entry15; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)16; CHECK-NEXT:    s_mov_b64 s[4:5], 017; CHECK-NEXT:  .LBB0_1: ; %load-store-loop18; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=119; CHECK-NEXT:    v_add_co_u32 v24, vcc_lo, v2, s420; CHECK-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v3, vcc_lo21; CHECK-NEXT:    s_clause 0x422; CHECK-NEXT:    flat_load_dwordx4 v[4:7], v[24:25] offset:6423; CHECK-NEXT:    flat_load_dwordx4 v[8:11], v[24:25] offset:3224; CHECK-NEXT:    flat_load_dwordx4 v[12:15], v[24:25]25; CHECK-NEXT:    flat_load_dwordx4 v[16:19], v[24:25] offset:1626; CHECK-NEXT:    flat_load_dwordx4 v[20:23], v[24:25] offset:12827; CHECK-NEXT:    v_add_co_u32 v80, vcc_lo, v24, 0x6028; CHECK-NEXT:    v_add_co_ci_u32_e64 v81, null, 0, v25, vcc_lo29; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v24, 4830; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v25, vcc_lo31; CHECK-NEXT:    v_add_co_u32 v84, vcc_lo, 0x50, v2432; CHECK-NEXT:    v_add_co_ci_u32_e64 v85, null, 0, v25, vcc_lo33; CHECK-NEXT:    s_clause 0xa34; CHECK-NEXT:    flat_load_dwordx4 v[24:27], v[80:81] offset:12835; CHECK-NEXT:    flat_load_dwordx4 v[28:31], v[80:81] offset:14436; CHECK-NEXT:    flat_load_dwordx4 v[32:35], v[80:81] offset:9637; CHECK-NEXT:    flat_load_dwordx4 v[36:39], v[80:81] offset:11238; CHECK-NEXT:    flat_load_dwordx4 v[48:51], v[80:81] offset:6439; CHECK-NEXT:    flat_load_dwordx4 v[52:55], v[80:81] offset:8040; CHECK-NEXT:    flat_load_dwordx4 v[64:67], v[80:81] offset:4841; CHECK-NEXT:    flat_load_dwordx4 v[68:71], v[80:81]42; CHECK-NEXT:    flat_load_dwordx4 v[80:83], v[80:81] offset:1643; CHECK-NEXT:    flat_load_dwordx4 v[84:87], v[84:85]44; CHECK-NEXT:    flat_load_dwordx4 v[96:99], v[96:97]45; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s446; CHECK-NEXT:    s_add_u32 s4, s4, 0x10047; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo48; CHECK-NEXT:    s_addc_u32 s5, s5, 049; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 4850; CHECK-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]51; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo52; CHECK-NEXT:    s_waitcnt vmcnt(15) lgkmcnt(15)53; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[4:7] offset:6454; CHECK-NEXT:    s_waitcnt vmcnt(14) lgkmcnt(15)55; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[8:11] offset:3256; CHECK-NEXT:    s_waitcnt vmcnt(13) lgkmcnt(15)57; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[12:15]58; CHECK-NEXT:    s_waitcnt vmcnt(12) lgkmcnt(15)59; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[16:19] offset:1660; CHECK-NEXT:    s_waitcnt vmcnt(11) lgkmcnt(15)61; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[20:23] offset:12862; CHECK-NEXT:    s_waitcnt vmcnt(10) lgkmcnt(15)63; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:17664; CHECK-NEXT:    s_waitcnt vmcnt(9) lgkmcnt(15)65; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:19266; CHECK-NEXT:    s_waitcnt vmcnt(8) lgkmcnt(15)67; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[32:35] offset:14468; CHECK-NEXT:    s_waitcnt vmcnt(7) lgkmcnt(15)69; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:16070; CHECK-NEXT:    s_waitcnt vmcnt(6) lgkmcnt(15)71; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:11272; CHECK-NEXT:    s_waitcnt vmcnt(5) lgkmcnt(15)73; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:12874; CHECK-NEXT:    s_waitcnt vmcnt(4) lgkmcnt(15)75; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:9676; CHECK-NEXT:    s_waitcnt vmcnt(3) lgkmcnt(15)77; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71] offset:4878; CHECK-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(15)79; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[80:83] offset:6480; CHECK-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(15)81; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[84:87] offset:3282; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(15)83; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[96:99]84; CHECK-NEXT:    s_and_b32 vcc_lo, exec_lo, s685; CHECK-NEXT:    s_cbranch_vccnz .LBB0_186; CHECK-NEXT:  ; %bb.2: ; %memcpy-split87; CHECK-NEXT:    s_waitcnt lgkmcnt(0)88; CHECK-NEXT:    s_setpc_b64 s[30:31]89;90; ALIGNED-LABEL: memcpy_p0_p0_sz2048:91; ALIGNED:       ; %bb.0: ; %entry92; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)93; ALIGNED-NEXT:    s_mov_b64 s[4:5], 094; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill95; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill96; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill97; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill98; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill99; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill100; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill101; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill102; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill103; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill104; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill105; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill106; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill107; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill108; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill109; ALIGNED-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill110; ALIGNED-NEXT:    buffer_store_dword v72, off, s[0:3], s32 ; 4-byte Folded Spill111; ALIGNED-NEXT:  .LBB0_1: ; %load-store-loop112; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1113; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, s4114; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, s5, v3, vcc_lo115; ALIGNED-NEXT:    v_add_co_u32 v6, vcc_lo, v4, 48116; ALIGNED-NEXT:    flat_load_dwordx4 v[20:23], v[4:5] offset:128117; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v5, vcc_lo118; ALIGNED-NEXT:    v_add_co_u32 v8, vcc_lo, 0x50, v4119; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v5, vcc_lo120; ALIGNED-NEXT:    v_add_co_u32 v12, vcc_lo, 0x60, v4121; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v5, vcc_lo122; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, 0x70, v4123; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v5, vcc_lo124; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, 0x90, v4125; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v5, vcc_lo126; ALIGNED-NEXT:    v_add_co_u32 v52, vcc_lo, 0xa0, v4127; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v53, null, 0, v5, vcc_lo128; ALIGNED-NEXT:    v_add_co_u32 v54, vcc_lo, 0xb0, v4129; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v55, null, 0, v5, vcc_lo130; ALIGNED-NEXT:    v_add_co_u32 v64, vcc_lo, 0xc0, v4131; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v65, null, 0, v5, vcc_lo132; ALIGNED-NEXT:    v_add_co_u32 v66, vcc_lo, 0xd0, v4133; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v67, null, 0, v5, vcc_lo134; ALIGNED-NEXT:    v_add_co_u32 v68, vcc_lo, 0xe0, v4135; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v69, null, 0, v5, vcc_lo136; ALIGNED-NEXT:    v_add_co_u32 v70, vcc_lo, 0xf0, v4137; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v71, null, 0, v5, vcc_lo138; ALIGNED-NEXT:    s_clause 0xe139; ALIGNED-NEXT:    flat_load_dwordx4 v[26:29], v[4:5]140; ALIGNED-NEXT:    flat_load_dwordx4 v[35:38], v[4:5] offset:16141; ALIGNED-NEXT:    flat_load_dwordx4 v[30:33], v[4:5] offset:32142; ALIGNED-NEXT:    flat_load_dwordx4 v[48:51], v[4:5] offset:64143; ALIGNED-NEXT:    flat_load_dwordx4 v[4:7], v[6:7]144; ALIGNED-NEXT:    flat_load_dwordx4 v[8:11], v[8:9]145; ALIGNED-NEXT:    flat_load_dwordx4 v[12:15], v[12:13]146; ALIGNED-NEXT:    flat_load_dwordx4 v[16:19], v[16:17]147; ALIGNED-NEXT:    flat_load_dwordx4 v[112:115], v[24:25]148; ALIGNED-NEXT:    flat_load_dwordx4 v[116:119], v[52:53]149; ALIGNED-NEXT:    flat_load_dwordx4 v[40:43], v[54:55]150; ALIGNED-NEXT:    flat_load_dwordx4 v[44:47], v[64:65]151; ALIGNED-NEXT:    flat_load_dwordx4 v[55:58], v[66:67]152; ALIGNED-NEXT:    flat_load_dwordx4 v[59:62], v[68:69]153; ALIGNED-NEXT:    flat_load_dwordx4 v[68:71], v[70:71]154; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v0, s4155; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v1, vcc_lo156; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100157; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0158; ALIGNED-NEXT:    s_waitcnt vmcnt(15) lgkmcnt(15)159; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:160160; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:164161; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168162; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:172163; ALIGNED-NEXT:    s_clause 0x3164; ALIGNED-NEXT:    buffer_load_dword v102, off, s[0:3], s32 offset:172165; ALIGNED-NEXT:    buffer_load_dword v103, off, s[0:3], s32 offset:168166; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:164167; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s32 offset:160168; ALIGNED-NEXT:    v_add_co_u32 v22, vcc_lo, v24, 6169; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v25, vcc_lo170; ALIGNED-NEXT:    v_add_co_u32 v20, vcc_lo, v24, 3171; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v25, vcc_lo172; ALIGNED-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]173; ALIGNED-NEXT:    s_waitcnt vmcnt(3)174; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v102 offset:136175; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:137176; ALIGNED-NEXT:    s_waitcnt vmcnt(2)177; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v103 offset:132178; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:133179; ALIGNED-NEXT:    s_waitcnt vmcnt(1)180; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v63 offset:128181; ALIGNED-NEXT:    flat_store_byte v[20:21], v63 offset:129182; ALIGNED-NEXT:    s_waitcnt vmcnt(0)183; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v72 offset:124184; ALIGNED-NEXT:    flat_store_byte v[24:25], v72 offset:128185; ALIGNED-NEXT:    s_waitcnt lgkmcnt(19)186; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:224187; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:228188; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:232189; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:236190; ALIGNED-NEXT:    s_clause 0x3191; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:236192; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:232193; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:228194; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:224195; ALIGNED-NEXT:    s_and_b32 vcc_lo, exec_lo, s6196; ALIGNED-NEXT:    s_waitcnt vmcnt(3)197; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v101 offset:72198; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:73199; ALIGNED-NEXT:    s_waitcnt vmcnt(2)200; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v100 offset:68201; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:69202; ALIGNED-NEXT:    s_waitcnt vmcnt(1)203; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v99 offset:64204; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:65205; ALIGNED-NEXT:    s_waitcnt vmcnt(0)206; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v98 offset:60207; ALIGNED-NEXT:    flat_store_byte v[24:25], v98 offset:64208; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:320209; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:324210; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:328211; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332212; ALIGNED-NEXT:    s_clause 0x3213; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:328214; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:332215; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:320216; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:324217; ALIGNED-NEXT:    s_waitcnt vmcnt(3)218; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v39 offset:36219; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:37220; ALIGNED-NEXT:    s_waitcnt vmcnt(2)221; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v34 offset:40222; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:41223; ALIGNED-NEXT:    s_waitcnt vmcnt(1)224; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v33 offset:28225; ALIGNED-NEXT:    flat_store_byte v[24:25], v33 offset:32226; ALIGNED-NEXT:    s_waitcnt vmcnt(0)227; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v32 offset:32228; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:33229; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:272230; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:276231; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:280232; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:284233; ALIGNED-NEXT:    s_clause 0x3234; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:284235; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:280236; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:276237; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:272238; ALIGNED-NEXT:    s_waitcnt vmcnt(3)239; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v67 offset:24240; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:25241; ALIGNED-NEXT:    s_waitcnt vmcnt(2)242; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v66 offset:20243; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:21244; ALIGNED-NEXT:    s_waitcnt vmcnt(1)245; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v53 offset:16246; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:17247; ALIGNED-NEXT:    s_waitcnt vmcnt(0)248; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v52 offset:12249; ALIGNED-NEXT:    flat_store_byte v[24:25], v52 offset:16250; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:288251; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:292252; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:296253; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:300254; ALIGNED-NEXT:    s_clause 0x3255; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:300256; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:296257; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:292258; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:288259; ALIGNED-NEXT:    s_waitcnt vmcnt(3)260; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v86 offset:8261; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:9262; ALIGNED-NEXT:    s_waitcnt vmcnt(2)263; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v82 offset:4264; ALIGNED-NEXT:    flat_store_byte v[24:25], v82 offset:8265; ALIGNED-NEXT:    s_waitcnt vmcnt(1)266; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v81267; ALIGNED-NEXT:    flat_store_byte v[24:25], v81 offset:4268; ALIGNED-NEXT:    s_waitcnt vmcnt(0)269; ALIGNED-NEXT:    flat_store_byte_d16_hi v[24:25], v80 offset:2270; ALIGNED-NEXT:    flat_store_byte v[24:25], v80271; ALIGNED-NEXT:    s_waitcnt lgkmcnt(40)272; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:112273; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:116274; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:120275; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:124276; ALIGNED-NEXT:    s_clause 0x3277; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:124278; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:120279; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:116280; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:112281; ALIGNED-NEXT:    s_waitcnt vmcnt(3)282; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v97 offset:248283; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:249284; ALIGNED-NEXT:    s_waitcnt vmcnt(2)285; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v87 offset:244286; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:245287; ALIGNED-NEXT:    s_waitcnt vmcnt(1)288; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v83 offset:240289; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:241290; ALIGNED-NEXT:    s_waitcnt vmcnt(0)291; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v68 offset:236292; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:237293; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:128294; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:132295; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:136296; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:140297; ALIGNED-NEXT:    s_clause 0x3298; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:140299; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:136300; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:132301; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128302; ALIGNED-NEXT:    s_waitcnt vmcnt(3)303; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v96 offset:232304; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:233305; ALIGNED-NEXT:    s_waitcnt vmcnt(2)306; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v84 offset:228307; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:229308; ALIGNED-NEXT:    s_waitcnt vmcnt(1)309; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v69 offset:224310; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:225311; ALIGNED-NEXT:    s_waitcnt vmcnt(0)312; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v54 offset:220313; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:221314; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:80315; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:84316; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:88317; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:92318; ALIGNED-NEXT:    s_clause 0x3319; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:92320; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:88321; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:84322; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:80323; ALIGNED-NEXT:    s_waitcnt vmcnt(3)324; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v85 offset:216325; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:217326; ALIGNED-NEXT:    s_waitcnt vmcnt(2)327; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v70 offset:212328; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:213329; ALIGNED-NEXT:    s_waitcnt vmcnt(1)330; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v55 offset:208331; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:209332; ALIGNED-NEXT:    s_waitcnt vmcnt(0)333; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v48 offset:204334; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:205335; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:96336; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:100337; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:104338; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:108339; ALIGNED-NEXT:    s_clause 0x3340; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:108341; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:104342; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:100343; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:96344; ALIGNED-NEXT:    s_waitcnt vmcnt(3)345; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v71 offset:200346; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:201347; ALIGNED-NEXT:    s_waitcnt vmcnt(2)348; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v64 offset:196349; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:197350; ALIGNED-NEXT:    s_waitcnt vmcnt(1)351; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v49 offset:192352; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:193353; ALIGNED-NEXT:    s_waitcnt vmcnt(0)354; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v35 offset:188355; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:189356; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:176357; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:180358; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:184359; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:188360; ALIGNED-NEXT:    s_clause 0x3361; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:188362; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:184363; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:180364; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:176365; ALIGNED-NEXT:    s_waitcnt vmcnt(3)366; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v65 offset:184367; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:185368; ALIGNED-NEXT:    s_waitcnt vmcnt(2)369; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v50 offset:180370; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:181371; ALIGNED-NEXT:    s_waitcnt vmcnt(1)372; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v36 offset:176373; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:177374; ALIGNED-NEXT:    s_waitcnt vmcnt(0)375; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v29 offset:172376; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:173377; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:192378; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:196379; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:200380; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:204381; ALIGNED-NEXT:    s_clause 0x3382; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:204383; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:200384; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:196385; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:192386; ALIGNED-NEXT:    s_waitcnt vmcnt(3)387; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v51 offset:168388; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:169389; ALIGNED-NEXT:    s_waitcnt vmcnt(2)390; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v37 offset:164391; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:165392; ALIGNED-NEXT:    s_waitcnt vmcnt(1)393; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v30 offset:160394; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:161395; ALIGNED-NEXT:    s_waitcnt vmcnt(0)396; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v27 offset:156397; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:157398; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:144399; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:148400; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:152401; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:156402; ALIGNED-NEXT:    s_clause 0x3403; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:156404; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152405; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:148406; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:144407; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v102408; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v102409; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v103410; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v103411; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v63412; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:140413; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v63414; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:138415; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v72416; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v72417; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:136418; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:134419; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:132420; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:130421; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:128422; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:126423; ALIGNED-NEXT:    s_waitcnt vmcnt(3)424; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v38 offset:152425; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:153426; ALIGNED-NEXT:    s_waitcnt vmcnt(2)427; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v31 offset:148428; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:149429; ALIGNED-NEXT:    s_waitcnt vmcnt(1)430; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v28 offset:144431; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:145432; ALIGNED-NEXT:    s_waitcnt vmcnt(0)433; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v26 offset:140434; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:141435; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:240436; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:244437; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:248438; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:252439; ALIGNED-NEXT:    s_clause 0x3440; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252441; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:248442; ALIGNED-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:244443; ALIGNED-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:240444; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v101445; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v101446; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v100447; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v100448; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v99449; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v99450; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v98451; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v98452; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v80453; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v86454; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v86455; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v80456; ALIGNED-NEXT:    s_waitcnt vmcnt(3)457; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v19 offset:120458; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:121459; ALIGNED-NEXT:    s_waitcnt vmcnt(2)460; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v18 offset:116461; ALIGNED-NEXT:    flat_store_byte v[20:21], v18 offset:117462; ALIGNED-NEXT:    s_waitcnt vmcnt(1)463; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v17 offset:112464; ALIGNED-NEXT:    flat_store_byte v[20:21], v17 offset:113465; ALIGNED-NEXT:    s_waitcnt vmcnt(0)466; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v16 offset:108467; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:109468; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:256469; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:260470; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:264471; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:268472; ALIGNED-NEXT:    s_clause 0x3473; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:268474; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:264475; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:260476; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:256477; ALIGNED-NEXT:    s_waitcnt vmcnt(3)478; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v15 offset:104479; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:105480; ALIGNED-NEXT:    s_waitcnt vmcnt(2)481; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v14 offset:100482; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:101483; ALIGNED-NEXT:    s_waitcnt vmcnt(1)484; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v13 offset:96485; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:97486; ALIGNED-NEXT:    s_waitcnt vmcnt(0)487; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v12 offset:92488; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:93489; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:208490; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:212491; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:216492; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:220493; ALIGNED-NEXT:    s_clause 0x3494; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:220495; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:216496; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212497; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208498; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:76499; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:74500; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:72501; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:70502; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:68503; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:66504; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:64505; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:62506; ALIGNED-NEXT:    s_waitcnt vmcnt(3)507; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v11 offset:88508; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:89509; ALIGNED-NEXT:    s_waitcnt vmcnt(2)510; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v10 offset:84511; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:85512; ALIGNED-NEXT:    s_waitcnt vmcnt(1)513; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v9 offset:80514; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:81515; ALIGNED-NEXT:    s_waitcnt vmcnt(0)516; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v8 offset:76517; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:77518; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:304519; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:308520; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:312521; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:316522; ALIGNED-NEXT:    s_clause 0x3523; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:316524; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:312525; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:308526; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:304527; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v32528; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v39529; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v39530; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v34531; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v34532; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v33533; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v33534; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v32535; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v67536; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v67537; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:36538; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v54539; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v54540; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v66541; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v66542; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v53543; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v53544; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v52545; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v52546; ALIGNED-NEXT:    flat_store_byte v[24:25], v114 offset:1547; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v82548; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v82549; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v81550; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v81551; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v97552; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v97553; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:40554; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v87555; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v87556; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:38557; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v83558; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v83559; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:44560; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v68561; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v68562; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:42563; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v96564; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v96565; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:32566; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v84567; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v84568; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:30569; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v69570; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v69571; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:34572; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v85573; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v85574; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:28575; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v70576; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v70577; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:26578; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 24, v55579; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:222580; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v8581; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v8582; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v55583; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:24584; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v48585; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v48586; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:22587; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v71588; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v71589; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:20590; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v64591; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v64592; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:18593; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 24, v49594; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v49595; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:16596; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v35597; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v35598; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:14599; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v65600; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v65601; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:12602; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v50603; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v50604; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:10605; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v36606; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v36607; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:8608; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v29609; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v29610; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:6611; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v51612; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v51613; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:4614; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v37615; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v37616; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:2617; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 24, v30618; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30619; ALIGNED-NEXT:    flat_store_byte v[20:21], v80620; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v27621; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v27622; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:252623; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v38624; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v38625; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:250626; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v31627; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v31628; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:248629; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v28630; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v28631; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:246632; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 24, v26633; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v26634; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:244635; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v19636; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v19637; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:242638; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 24, v18639; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v18640; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:240641; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v17642; ALIGNED-NEXT:    v_lshrrev_b32_e32 v17, 8, v17643; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:238644; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v16645; ALIGNED-NEXT:    v_lshrrev_b32_e32 v16, 8, v16646; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:236647; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v15648; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v15649; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:234650; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v14651; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v14652; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:232653; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v13654; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v13655; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:230656; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v12657; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v12658; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:228659; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v11660; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v11661; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:226662; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 24, v10663; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v10664; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:224665; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v9666; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v9667; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:220668; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:218669; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:216670; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:214671; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:212672; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:210673; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:208674; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:206675; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:204676; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:202677; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:200678; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:198679; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:196680; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:194681; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:192682; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:190683; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:188684; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:186685; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:184686; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:182687; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:180688; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:178689; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:176690; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:174691; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:172692; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:170693; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:168694; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:166695; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:164696; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:162697; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:160698; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:158699; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:156700; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:154701; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:152702; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:150703; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:148704; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:146705; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:144706; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:142707; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:124708; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:122709; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:120710; ALIGNED-NEXT:    flat_store_byte v[20:21], v18 offset:118711; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:116712; ALIGNED-NEXT:    flat_store_byte v[20:21], v17 offset:114713; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:112714; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:110715; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:108716; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:106717; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:104718; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:102719; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:100720; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:98721; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:96722; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:94723; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:92724; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:90725; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:88726; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:86727; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:84728; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:82729; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:80730; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:78731; ALIGNED-NEXT:    s_waitcnt vmcnt(3)732; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v7 offset:56733; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:57734; ALIGNED-NEXT:    s_waitcnt vmcnt(2)735; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v6 offset:52736; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:53737; ALIGNED-NEXT:    s_waitcnt vmcnt(1)738; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v5 offset:48739; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:49740; ALIGNED-NEXT:    s_waitcnt vmcnt(0)741; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v4 offset:44742; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:45743; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 24, v7744; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v7745; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 24, v6746; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v6747; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 24, v5748; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v5749; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 24, v4750; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v4751; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:60752; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:58753; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:56754; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:54755; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:52756; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:50757; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:48758; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:46759; ALIGNED-NEXT:    s_cbranch_vccnz .LBB0_1760; ALIGNED-NEXT:  ; %bb.2: ; %memcpy-split761; ALIGNED-NEXT:    s_clause 0x10 ; 68-byte Folded Reload762; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s32763; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:4764; ALIGNED-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:8765; ALIGNED-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:12766; ALIGNED-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:16767; ALIGNED-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:20768; ALIGNED-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:24769; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:28770; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:32771; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:36772; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:40773; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:44774; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:48775; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:52776; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:56777; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:60778; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:64779; ALIGNED-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)780; ALIGNED-NEXT:    s_setpc_b64 s[30:31]781;782; UNROLL3-LABEL: memcpy_p0_p0_sz2048:783; UNROLL3:       ; %bb.0: ; %entry784; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)785; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0786; UNROLL3-NEXT:    .p2align 6787; UNROLL3-NEXT:  .LBB0_1: ; %load-store-loop788; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1789; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s4790; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo791; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4792; UNROLL3-NEXT:    s_add_u32 s4, s4, 48793; UNROLL3-NEXT:    s_clause 0x2794; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[12:13]795; UNROLL3-NEXT:    flat_load_dwordx4 v[8:11], v[12:13] offset:16796; UNROLL3-NEXT:    flat_load_dwordx4 v[12:15], v[12:13] offset:32797; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo798; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0799; UNROLL3-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)800; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7]801; UNROLL3-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(2)802; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11] offset:16803; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(2)804; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:32805; UNROLL3-NEXT:    v_cmp_gt_u64_e64 s6, 0x7e0, s[4:5]806; UNROLL3-NEXT:    s_and_b32 vcc_lo, exec_lo, s6807; UNROLL3-NEXT:    s_cbranch_vccnz .LBB0_1808; UNROLL3-NEXT:  ; %bb.2: ; %memcpy-split809; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[2:3] offset:2016810; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)811; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:2016812; UNROLL3-NEXT:    flat_load_dwordx4 v[2:5], v[2:3] offset:2032813; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)814; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[2:5] offset:2032815; UNROLL3-NEXT:    s_waitcnt lgkmcnt(0)816; UNROLL3-NEXT:    s_setpc_b64 s[30:31]817entry:818  tail call void @llvm.memcpy.p0.p0.i64(ptr addrspace(0) noundef nonnull align 1 %dst, ptr addrspace(0) noundef nonnull align 1 %src, i64 2048, i1 false)819  ret void820}821 822define void @memcpy_p1_p1_sz2048(ptr addrspace(1) align 1 %dst, ptr addrspace(1) align 1 readonly %src) {823; CHECK-LABEL: memcpy_p1_p1_sz2048:824; CHECK:       ; %bb.0: ; %entry825; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)826; CHECK-NEXT:    s_mov_b64 s[4:5], 0827; CHECK-NEXT:  .LBB1_1: ; %load-store-loop828; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1829; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s4830; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo831; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4832; CHECK-NEXT:    s_add_u32 s4, s4, 0x100833; CHECK-NEXT:    s_clause 0xf834; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:224835; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:240836; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:192837; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:208838; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:160839; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:176840; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:128841; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:144842; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:96843; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:112844; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:64845; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:80846; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:32847; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:48848; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off849; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off offset:16850; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo851; CHECK-NEXT:    s_addc_u32 s5, s5, 0852; CHECK-NEXT:    s_waitcnt vmcnt(15)853; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[4:7], off offset:224854; CHECK-NEXT:    s_waitcnt vmcnt(14)855; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[8:11], off offset:240856; CHECK-NEXT:    s_waitcnt vmcnt(13)857; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[12:15], off offset:192858; CHECK-NEXT:    s_waitcnt vmcnt(12)859; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[16:19], off offset:208860; CHECK-NEXT:    s_waitcnt vmcnt(11)861; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[20:23], off offset:160862; CHECK-NEXT:    s_waitcnt vmcnt(10)863; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[24:27], off offset:176864; CHECK-NEXT:    s_waitcnt vmcnt(9)865; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[28:31], off offset:128866; CHECK-NEXT:    s_waitcnt vmcnt(8)867; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[32:35], off offset:144868; CHECK-NEXT:    s_waitcnt vmcnt(7)869; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[36:39], off offset:96870; CHECK-NEXT:    s_waitcnt vmcnt(6)871; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[48:51], off offset:112872; CHECK-NEXT:    s_waitcnt vmcnt(5)873; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[52:55], off offset:64874; CHECK-NEXT:    s_waitcnt vmcnt(4)875; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[64:67], off offset:80876; CHECK-NEXT:    s_waitcnt vmcnt(3)877; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[68:71], off offset:32878; CHECK-NEXT:    s_waitcnt vmcnt(2)879; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[80:83], off offset:48880; CHECK-NEXT:    s_waitcnt vmcnt(1)881; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[84:87], off882; CHECK-NEXT:    s_waitcnt vmcnt(0)883; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[96:99], off offset:16884; CHECK-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]885; CHECK-NEXT:    s_and_b32 vcc_lo, exec_lo, s6886; CHECK-NEXT:    s_cbranch_vccnz .LBB1_1887; CHECK-NEXT:  ; %bb.2: ; %memcpy-split888; CHECK-NEXT:    s_setpc_b64 s[30:31]889;890; ALIGNED-LABEL: memcpy_p1_p1_sz2048:891; ALIGNED:       ; %bb.0: ; %entry892; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)893; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0894; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill895; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill896; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill897; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill898; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill899; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill900; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill901; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill902; ALIGNED-NEXT:  .LBB1_1: ; %load-store-loop903; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1904; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v2, s4905; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v3, vcc_lo906; ALIGNED-NEXT:    s_clause 0xf907; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[24:25], off offset:240908; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[24:25], off offset:224909; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[24:25], off910; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[24:25], off offset:16911; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[24:25], off offset:32912; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[24:25], off offset:48913; ALIGNED-NEXT:    global_load_dwordx4 v[116:119], v[24:25], off offset:64914; ALIGNED-NEXT:    global_load_dwordx4 v[40:43], v[24:25], off offset:80915; ALIGNED-NEXT:    global_load_dwordx4 v[26:29], v[24:25], off offset:96916; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[24:25], off offset:112917; ALIGNED-NEXT:    global_load_dwordx4 v[44:47], v[24:25], off offset:128918; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[24:25], off offset:144919; ALIGNED-NEXT:    global_load_dwordx4 v[66:69], v[24:25], off offset:160920; ALIGNED-NEXT:    global_load_dwordx4 v[81:84], v[24:25], off offset:176921; ALIGNED-NEXT:    global_load_dwordx4 v[96:99], v[24:25], off offset:192922; ALIGNED-NEXT:    global_load_dwordx4 v[100:103], v[24:25], off offset:208923; ALIGNED-NEXT:    s_waitcnt vmcnt(15)924; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:64925; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:68926; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:72927; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:76928; ALIGNED-NEXT:    s_clause 0x3929; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76930; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72931; ALIGNED-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:68932; ALIGNED-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:64933; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4934; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo935; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100936; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0937; ALIGNED-NEXT:    s_waitcnt vmcnt(3)938; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v31, off offset:254939; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:252940; ALIGNED-NEXT:    s_waitcnt vmcnt(2)941; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v30, off offset:250942; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:248943; ALIGNED-NEXT:    s_waitcnt vmcnt(1)944; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v25, off offset:246945; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:244946; ALIGNED-NEXT:    s_waitcnt vmcnt(0)947; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v24, off offset:242948; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:240949; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:80950; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:84951; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:88952; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:92953; ALIGNED-NEXT:    s_clause 0x3954; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:92955; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:88956; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:84957; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:80958; ALIGNED-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]959; ALIGNED-NEXT:    s_waitcnt vmcnt(3)960; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v51, off offset:238961; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:236962; ALIGNED-NEXT:    s_waitcnt vmcnt(2)963; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v50, off offset:234964; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:232965; ALIGNED-NEXT:    s_waitcnt vmcnt(1)966; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v49, off offset:230967; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:228968; ALIGNED-NEXT:    s_waitcnt vmcnt(0)969; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v36, off offset:226970; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:224971; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:32972; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:36973; ALIGNED-NEXT:    buffer_store_dword v102, off, s[0:3], s32 offset:40974; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:44975; ALIGNED-NEXT:    s_clause 0x3976; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:44977; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:40978; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:36979; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:32980; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v31981; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v31982; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v30983; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30984; ALIGNED-NEXT:    s_and_b32 vcc_lo, exec_lo, s6985; ALIGNED-NEXT:    s_waitcnt vmcnt(3)986; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v71, off offset:222987; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:220988; ALIGNED-NEXT:    s_waitcnt vmcnt(2)989; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v70, off offset:218990; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:216991; ALIGNED-NEXT:    s_waitcnt vmcnt(1)992; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v65, off offset:214993; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:212994; ALIGNED-NEXT:    s_waitcnt vmcnt(0)995; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v64, off offset:210996; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:208997; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:48998; ALIGNED-NEXT:    buffer_store_dword v97, off, s[0:3], s32 offset:52999; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:561000; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:601001; ALIGNED-NEXT:    s_clause 0x31002; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:601003; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:561004; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:521005; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:481006; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1007; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v87, off offset:2061008; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:2041009; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1010; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v86, off offset:2021011; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:2001012; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1013; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v85, off offset:1981014; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:1961015; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1016; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v80, off offset:1941017; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:1921018; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:1281019; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:1321020; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:1361021; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:1401022; ALIGNED-NEXT:    s_clause 0x31023; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:1401024; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:1361025; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:1321026; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:1281027; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1028; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v101, off offset:1901029; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:1881030; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1031; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v99, off offset:1861032; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:1841033; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1034; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v96, off offset:1821035; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:1801036; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1037; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v81, off offset:1781038; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:1761039; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:1441040; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:1481041; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:1521042; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:1561043; ALIGNED-NEXT:    s_clause 0x31044; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:1561045; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:1521046; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:1481047; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:1441048; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1049; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v100, off offset:1741050; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:1721051; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1052; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v97, off offset:1701053; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:1681054; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1055; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v82, off offset:1661056; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:1641057; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1058; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v66, off offset:1621059; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:1601060; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:961061; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:1001062; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:1041063; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:1081064; ALIGNED-NEXT:    s_clause 0x31065; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:1081066; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:1041067; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:1001068; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:961069; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1070; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v98, off offset:1581071; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:1561072; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1073; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v83, off offset:1541074; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:1521075; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1076; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v67, off offset:1501077; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:1481078; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1079; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v52, off offset:1461080; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:1441081; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:1121082; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:1161083; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:1201084; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:1241085; ALIGNED-NEXT:    s_clause 0x31086; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:1241087; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:1201088; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:1161089; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:1121090; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1091; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v84, off offset:1421092; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:1401093; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1094; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v68, off offset:1381095; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:1361096; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1097; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v53, off offset:1341098; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:1321099; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1100; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v37, off offset:1301101; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:1281102; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:1921103; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:1961104; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:2001105; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:2041106; ALIGNED-NEXT:    s_clause 0x31107; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:2041108; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:2001109; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:1961110; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:1921111; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1112; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v69, off offset:1261113; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:1241114; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1115; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v54, off offset:1221116; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:1201117; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1118; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v38, off offset:1181119; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:1161120; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1121; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v32, off offset:1141122; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:1121123; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:2081124; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:2121125; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:2161126; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:2201127; ALIGNED-NEXT:    s_clause 0x31128; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:2201129; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:2161130; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:2121131; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:2081132; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1133; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v55, off offset:1101134; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:1081135; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1136; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v39, off offset:1061137; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:1041138; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1139; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v33, off offset:1021140; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:1001141; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1142; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v26, off offset:981143; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:961144; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:1601145; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:1641146; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:1681147; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:1721148; ALIGNED-NEXT:    s_clause 0x31149; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:1721150; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:1681151; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:1641152; ALIGNED-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:1601153; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1154; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v48, off offset:941155; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:921156; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1157; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v34, off offset:901158; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:881159; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1160; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v27, off offset:861161; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:841162; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1163; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v21, off offset:821164; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:801165; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:1761166; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:1801167; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:1841168; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:1881169; ALIGNED-NEXT:    s_clause 0x31170; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:1881171; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:1841172; ALIGNED-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:1801173; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:1761174; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1175; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v35, off offset:781176; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:761177; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1178; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v28, off offset:741179; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:721180; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1181; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v22, off offset:701182; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:681183; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1184; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v19, off offset:661185; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:641186; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:2561187; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:2601188; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:2641189; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:2681190; ALIGNED-NEXT:    s_clause 0x31191; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:2681192; ALIGNED-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:2641193; ALIGNED-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:2601194; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:2561195; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v251196; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v251197; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v241198; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 8, v241199; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v511200; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v511201; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v501202; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v501203; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1204; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v29, off offset:621205; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:601206; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1207; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v23, off offset:581208; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:561209; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1210; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v20, off offset:541211; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:521212; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1213; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v18, off offset:501214; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:481215; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:2721216; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:2761217; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:2801218; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:2841219; ALIGNED-NEXT:    s_clause 0x31220; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:2801221; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:2841222; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:2721223; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:2761224; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1225; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v15, off offset:421226; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:401227; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1228; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v14, off offset:461229; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:441230; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1231; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v13, off offset:341232; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:321233; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1234; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v12, off offset:381235; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:361236; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:2241237; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:2281238; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:2321239; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:2361240; ALIGNED-NEXT:    s_clause 0x31241; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:2361242; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:2321243; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:2281244; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:2241245; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1246; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v11, off offset:301247; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:281248; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1249; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v10, off offset:261250; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:241251; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1252; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v9, off offset:221253; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:201254; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1255; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v8, off offset:181256; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:161257; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:2401258; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:2441259; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:2481260; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:2521261; ALIGNED-NEXT:    s_clause 0x31262; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:2521263; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:2481264; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:2441265; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:2401266; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:2471267; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v651268; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:2551269; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v491270; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v491271; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:2531272; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v361273; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v361274; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:2511275; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v711276; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v711277; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:2491278; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v701279; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v701280; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v651281; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:2451282; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v641283; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v641284; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:2151285; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v671286; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v671287; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:2431288; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v871289; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v871290; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:2411291; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v861292; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v861293; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:2391294; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v851295; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v851296; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:2371297; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v801298; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v801299; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:2351300; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v1011301; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1011302; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:2331303; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v991304; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v991305; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:2311306; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v961307; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v961308; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:2291309; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v811310; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v811311; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:2271312; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v1001313; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v1001314; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:2251315; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v971316; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v971317; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:2231318; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v821319; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 8, v821320; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:2211321; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 24, v661322; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v661323; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:2191324; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v981325; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v981326; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:2171327; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v831328; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v831329; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:2131330; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 24, v521331; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v521332; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:2111333; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v841334; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v841335; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:2091336; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v681337; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:1491338; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 24, v81339; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v81340; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v681341; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:2071342; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v531343; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v531344; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:2051345; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 24, v371346; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v371347; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:2031348; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v691349; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v691350; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:2011351; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v541352; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v541353; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:1991354; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v381355; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v381356; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:1971357; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 24, v321358; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v321359; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:1951360; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v551361; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v551362; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:1931363; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v391364; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v391365; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:1911366; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v331367; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v331368; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:1891369; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v261370; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v261371; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:1871372; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v481373; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v481374; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:1851375; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v341376; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v341377; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:1831378; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v271379; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v271380; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:1811381; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v211382; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 8, v211383; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:1791384; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v351385; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v351386; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:1771387; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 24, v281388; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v281389; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:1751390; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v221391; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 8, v221392; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:1731393; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v191394; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v191395; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:1711396; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v291397; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v291398; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:1691399; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v231400; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 8, v231401; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:1671402; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v201403; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 8, v201404; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:1651405; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v181406; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v181407; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:1631408; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 24, v151409; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v151410; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:1611411; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v141412; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v141413; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:1591414; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v131415; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v131416; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:1571417; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v121418; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v121419; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:1551420; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v111421; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v111422; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:1531423; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 24, v101424; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v101425; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:1511426; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v91427; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v91428; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:1471429; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:1451430; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:1431431; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:1411432; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:1391433; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:1371434; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:1351435; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:1331436; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:1311437; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:1291438; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:1271439; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:1251440; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:1231441; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:1211442; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:1191443; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:1171444; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:1151445; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:1131446; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:1111447; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:1091448; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:1071449; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:1051450; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:1031451; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:1011452; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:991453; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:971454; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:951455; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:931456; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:911457; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:891458; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:871459; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:851460; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:831461; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:811462; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:791463; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:771464; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:751465; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:731466; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:711467; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:691468; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:671469; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:651470; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:631471; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:611472; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:591473; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:571474; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:551475; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:531476; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:511477; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:491478; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:431479; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:411480; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:471481; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:451482; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:351483; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:331484; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:391485; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:371486; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:311487; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:291488; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:271489; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:251490; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:231491; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:211492; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:191493; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:171494; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1495; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v7, off offset:141496; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:121497; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1498; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v6, off offset:101499; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:81500; ALIGNED-NEXT:    s_waitcnt vmcnt(1)1501; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v5, off offset:61502; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:41503; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1504; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v4, off offset:21505; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off1506; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 24, v71507; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v71508; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 24, v61509; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v61510; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 24, v51511; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v51512; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 24, v41513; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v41514; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:151515; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:131516; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:111517; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:91518; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:71519; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:51520; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:31521; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off offset:11522; ALIGNED-NEXT:    s_cbranch_vccnz .LBB1_11523; ALIGNED-NEXT:  ; %bb.2: ; %memcpy-split1524; ALIGNED-NEXT:    s_clause 0x7 ; 32-byte Folded Reload1525; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s321526; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:41527; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:81528; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:121529; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:161530; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:201531; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:241532; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:281533; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1534; ALIGNED-NEXT:    s_setpc_b64 s[30:31]1535;1536; UNROLL3-LABEL: memcpy_p1_p1_sz2048:1537; UNROLL3:       ; %bb.0: ; %entry1538; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1539; UNROLL3-NEXT:    s_mov_b64 s[4:5], 01540; UNROLL3-NEXT:    .p2align 61541; UNROLL3-NEXT:  .LBB1_1: ; %load-store-loop1542; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=11543; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s41544; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo1545; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s41546; UNROLL3-NEXT:    s_add_u32 s4, s4, 481547; UNROLL3-NEXT:    s_clause 0x21548; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off1549; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off offset:161550; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:321551; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo1552; UNROLL3-NEXT:    s_addc_u32 s5, s5, 01553; UNROLL3-NEXT:    s_waitcnt vmcnt(2)1554; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[4:7], off1555; UNROLL3-NEXT:    s_waitcnt vmcnt(1)1556; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[8:11], off offset:161557; UNROLL3-NEXT:    s_waitcnt vmcnt(0)1558; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[12:15], off offset:321559; UNROLL3-NEXT:    v_cmp_gt_u64_e64 s6, 0x7e0, s[4:5]1560; UNROLL3-NEXT:    s_and_b32 vcc_lo, exec_lo, s61561; UNROLL3-NEXT:    s_cbranch_vccnz .LBB1_11562; UNROLL3-NEXT:  ; %bb.2: ; %memcpy-split1563; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:20161564; UNROLL3-NEXT:    s_waitcnt vmcnt(0)1565; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[4:7], off offset:20161566; UNROLL3-NEXT:    global_load_dwordx4 v[2:5], v[2:3], off offset:20321567; UNROLL3-NEXT:    s_waitcnt vmcnt(0)1568; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:20321569; UNROLL3-NEXT:    s_setpc_b64 s[30:31]1570entry:1571  tail call void @llvm.memcpy.p1.p1.i64(ptr addrspace(1) noundef nonnull align 1 %dst, ptr addrspace(1) noundef nonnull align 1 %src, i64 2048, i1 false)1572  ret void1573}1574 1575define void @memcpy_p0_p4_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(4) align 1 readonly %src) {1576; CHECK-LABEL: memcpy_p0_p4_sz2048:1577; CHECK:       ; %bb.0: ; %entry1578; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1579; CHECK-NEXT:    s_mov_b64 s[4:5], 01580; CHECK-NEXT:  .LBB2_1: ; %load-store-loop1581; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=11582; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s41583; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo1584; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s41585; CHECK-NEXT:    s_add_u32 s4, s4, 0x1001586; CHECK-NEXT:    s_clause 0xf1587; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:2401588; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:2241589; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:2081590; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:1921591; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:1761592; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:1601593; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:1441594; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:1281595; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:1121596; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:961597; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:801598; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:641599; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:481600; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:321601; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off offset:161602; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off1603; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo1604; CHECK-NEXT:    s_addc_u32 s5, s5, 01605; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 481606; CHECK-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]1607; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo1608; CHECK-NEXT:    s_waitcnt vmcnt(15)1609; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[4:7] offset:1921610; CHECK-NEXT:    s_waitcnt vmcnt(14)1611; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[8:11] offset:1761612; CHECK-NEXT:    s_waitcnt vmcnt(13)1613; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[12:15] offset:1601614; CHECK-NEXT:    s_waitcnt vmcnt(12)1615; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[16:19] offset:1441616; CHECK-NEXT:    s_waitcnt vmcnt(11)1617; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[20:23] offset:1281618; CHECK-NEXT:    s_waitcnt vmcnt(10)1619; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:1121620; CHECK-NEXT:    s_waitcnt vmcnt(9)1621; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:961622; CHECK-NEXT:    s_waitcnt vmcnt(8)1623; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[32:35] offset:1281624; CHECK-NEXT:    s_waitcnt vmcnt(7)1625; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:641626; CHECK-NEXT:    s_waitcnt vmcnt(6)1627; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:481628; CHECK-NEXT:    s_waitcnt vmcnt(5)1629; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:321630; CHECK-NEXT:    s_waitcnt vmcnt(4)1631; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:641632; CHECK-NEXT:    s_waitcnt vmcnt(3)1633; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71]1634; CHECK-NEXT:    s_waitcnt vmcnt(2)1635; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:321636; CHECK-NEXT:    s_waitcnt vmcnt(1)1637; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:161638; CHECK-NEXT:    s_waitcnt vmcnt(0)1639; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]1640; CHECK-NEXT:    s_and_b32 vcc_lo, exec_lo, s61641; CHECK-NEXT:    s_cbranch_vccnz .LBB2_11642; CHECK-NEXT:  ; %bb.2: ; %memcpy-split1643; CHECK-NEXT:    s_waitcnt lgkmcnt(0)1644; CHECK-NEXT:    s_setpc_b64 s[30:31]1645;1646; ALIGNED-LABEL: memcpy_p0_p4_sz2048:1647; ALIGNED:       ; %bb.0: ; %entry1648; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1649; ALIGNED-NEXT:    s_mov_b64 s[4:5], 01650; ALIGNED-NEXT:  .LBB2_1: ; %load-store-loop1651; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=11652; ALIGNED-NEXT:    v_add_co_u32 v8, vcc_lo, v2, s41653; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v9, null, s5, v3, vcc_lo1654; ALIGNED-NEXT:    v_add_co_u32 v84, vcc_lo, v0, s41655; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v85, null, s5, v1, vcc_lo1656; ALIGNED-NEXT:    s_clause 0xf1657; ALIGNED-NEXT:    global_load_dwordx4 v[98:101], v[8:9], off offset:2401658; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[8:9], off offset:2241659; ALIGNED-NEXT:    global_load_dwordx4 v[80:83], v[8:9], off offset:2081660; ALIGNED-NEXT:    global_load_dwordx4 v[68:71], v[8:9], off offset:1921661; ALIGNED-NEXT:    global_load_dwordx4 v[64:67], v[8:9], off offset:1761662; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[8:9], off offset:1601663; ALIGNED-NEXT:    global_load_dwordx4 v[48:51], v[8:9], off offset:1441664; ALIGNED-NEXT:    global_load_dwordx4 v[36:39], v[8:9], off offset:1281665; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[8:9], off offset:1121666; ALIGNED-NEXT:    global_load_dwordx4 v[28:31], v[8:9], off offset:961667; ALIGNED-NEXT:    global_load_dwordx4 v[24:27], v[8:9], off offset:801668; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[8:9], off offset:641669; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[8:9], off offset:481670; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[8:9], off offset:321671; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[8:9], off1672; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[8:9], off offset:161673; ALIGNED-NEXT:    v_add_co_u32 v96, vcc_lo, v84, 61674; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v85, vcc_lo1675; ALIGNED-NEXT:    v_add_co_u32 v86, vcc_lo, v84, 31676; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v87, null, 0, v85, vcc_lo1677; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x1001678; ALIGNED-NEXT:    s_addc_u32 s5, s5, 01679; ALIGNED-NEXT:    s_waitcnt vmcnt(15)1680; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:1681681; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:1721682; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:1641683; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:1601684; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v100 offset:2441685; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v101 offset:2481686; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:2491687; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:2451688; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v99 offset:2401689; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:2411690; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v98 offset:2361691; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:2371692; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v1001693; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v1011694; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1011695; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v1001696; ALIGNED-NEXT:    v_lshrrev_b32_e32 v116, 24, v991697; ALIGNED-NEXT:    v_lshrrev_b32_e32 v117, 8, v991698; ALIGNED-NEXT:    v_lshrrev_b32_e32 v118, 24, v981699; ALIGNED-NEXT:    v_lshrrev_b32_e32 v119, 8, v981700; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:2481701; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:2461702; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:2521703; ALIGNED-NEXT:    s_waitcnt vmcnt(14)1704; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v1141705; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:2501706; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v1141707; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v1151708; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1151709; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v1131710; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v1131711; ALIGNED-NEXT:    flat_store_byte v[86:87], v116 offset:2441712; ALIGNED-NEXT:    flat_store_byte v[86:87], v117 offset:2421713; ALIGNED-NEXT:    flat_store_byte v[86:87], v118 offset:2401714; ALIGNED-NEXT:    flat_store_byte v[86:87], v119 offset:2381715; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:1841716; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:1881717; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:1801718; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:1761719; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v114 offset:2281720; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v115 offset:2321721; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:2331722; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:2291723; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v113 offset:2241724; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:2251725; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v112 offset:2201726; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:2211727; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v1121728; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v1121729; ALIGNED-NEXT:    s_waitcnt vmcnt(13)1730; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v821731; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v821732; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:2321733; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v831734; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:2301735; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v831736; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:2361737; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v811738; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:2341739; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v811740; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:2281741; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v801742; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:2261743; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v801744; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:2241745; ALIGNED-NEXT:    s_waitcnt vmcnt(12)1746; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v701747; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:2221748; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v701749; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:1361750; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:1401751; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:1321752; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:1281753; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v82 offset:2121754; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v83 offset:2161755; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:2171756; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:2131757; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v81 offset:2081758; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:2091759; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v80 offset:2041760; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:2051761; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v711762; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v711763; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v691764; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v691765; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:2161766; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v681767; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:2141768; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v681769; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:2201770; ALIGNED-NEXT:    s_waitcnt vmcnt(11)1771; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v661772; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:2181773; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v661774; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:2121775; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v671776; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:2101777; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v671778; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:2081779; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v651780; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:2061781; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v651782; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:1521783; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:1561784; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:1481785; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:1441786; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v70 offset:1961787; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v71 offset:2001788; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:2011789; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:1971790; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v69 offset:1921791; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:1931792; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v68 offset:1881793; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:1891794; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v641795; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v641796; ALIGNED-NEXT:    s_waitcnt vmcnt(10)1797; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v541798; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v541799; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:2001800; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v551801; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:1981802; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v551803; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:2041804; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v521805; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:2021806; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v521807; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:1961808; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v531809; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:1941810; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v531811; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:1921812; ALIGNED-NEXT:    s_waitcnt vmcnt(9)1813; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v501814; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:1901815; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v501816; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:2321817; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:2361818; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:2281819; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:2241820; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v66 offset:1801821; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v67 offset:1841822; ALIGNED-NEXT:    flat_store_byte v[86:87], v67 offset:1851823; ALIGNED-NEXT:    flat_store_byte v[86:87], v66 offset:1811824; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v65 offset:1761825; ALIGNED-NEXT:    flat_store_byte v[86:87], v65 offset:1771826; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v64 offset:1721827; ALIGNED-NEXT:    flat_store_byte v[86:87], v64 offset:1731828; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v511829; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v511830; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v491831; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v491832; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:1841833; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v481834; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:1821835; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v481836; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:1881837; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:1861838; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:1801839; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:1781840; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:1761841; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:1741842; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:2521843; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:2481844; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:2441845; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:2401846; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v54 offset:1641847; ALIGNED-NEXT:    flat_store_byte v[86:87], v54 offset:1651848; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v55 offset:1681849; ALIGNED-NEXT:    flat_store_byte v[86:87], v55 offset:1691850; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v52 offset:1561851; ALIGNED-NEXT:    flat_store_byte v[86:87], v52 offset:1571852; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v53 offset:1601853; ALIGNED-NEXT:    flat_store_byte v[86:87], v53 offset:1611854; ALIGNED-NEXT:    s_waitcnt vmcnt(8)1855; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v361856; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v381857; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v381858; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v391859; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v391860; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v371861; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v371862; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v361863; ALIGNED-NEXT:    s_waitcnt vmcnt(7)1864; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v341865; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v341866; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:1681867; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v351868; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:1661869; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v351870; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:1721871; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v331872; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:1701873; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v331874; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:1601875; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v321876; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:1581877; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v321878; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:1641879; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:1621880; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:2001881; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:2041882; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:1961883; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:1921884; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:1521885; ALIGNED-NEXT:    s_waitcnt vmcnt(6)1886; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v281887; ALIGNED-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]1888; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:1501889; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v281890; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v301891; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v301892; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v50 offset:1481893; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v51 offset:1521894; ALIGNED-NEXT:    flat_store_byte v[86:87], v51 offset:1531895; ALIGNED-NEXT:    flat_store_byte v[86:87], v50 offset:1491896; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v49 offset:1441897; ALIGNED-NEXT:    flat_store_byte v[86:87], v49 offset:1451898; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v48 offset:1401899; ALIGNED-NEXT:    flat_store_byte v[86:87], v48 offset:1411900; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v311901; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v311902; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v291903; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v291904; ALIGNED-NEXT:    flat_store_byte v[86:87], v64 offset:1561905; ALIGNED-NEXT:    s_waitcnt vmcnt(5)1906; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v261907; ALIGNED-NEXT:    flat_store_byte v[86:87], v65 offset:1541908; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v261909; ALIGNED-NEXT:    flat_store_byte v[86:87], v66 offset:1481910; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v271911; ALIGNED-NEXT:    flat_store_byte v[86:87], v67 offset:1461912; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:1441913; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:1421914; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:2161915; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:2201916; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:2121917; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:2081918; ALIGNED-NEXT:    flat_store_byte v[86:87], v52 offset:1281919; ALIGNED-NEXT:    s_waitcnt vmcnt(3)1920; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v181921; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v271922; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v38 offset:1321923; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v39 offset:1361924; ALIGNED-NEXT:    flat_store_byte v[86:87], v39 offset:1371925; ALIGNED-NEXT:    flat_store_byte v[86:87], v38 offset:1331926; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v37 offset:1281927; ALIGNED-NEXT:    flat_store_byte v[86:87], v37 offset:1291928; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v36 offset:1241929; ALIGNED-NEXT:    flat_store_byte v[84:85], v36 offset:1281930; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:1361931; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:1341932; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:1401933; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:1381934; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:1321935; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:1301936; ALIGNED-NEXT:    flat_store_byte v[86:87], v53 offset:1261937; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v181938; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:401939; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:441940; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:361941; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:321942; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v34 offset:1161943; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v35 offset:1201944; ALIGNED-NEXT:    flat_store_byte v[86:87], v35 offset:1211945; ALIGNED-NEXT:    flat_store_byte v[86:87], v34 offset:1171946; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v33 offset:1121947; ALIGNED-NEXT:    flat_store_byte v[86:87], v33 offset:1131948; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v32 offset:1081949; ALIGNED-NEXT:    flat_store_byte v[86:87], v32 offset:1091950; ALIGNED-NEXT:    flat_store_byte v[86:87], v54 offset:1201951; ALIGNED-NEXT:    flat_store_byte v[86:87], v55 offset:1181952; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:1241953; ALIGNED-NEXT:    s_waitcnt vmcnt(2)1954; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v141955; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:1221956; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:1161957; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:1141958; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:1121959; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:1101960; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:561961; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:601962; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:521963; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:481964; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v30 offset:1001965; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v31 offset:1041966; ALIGNED-NEXT:    flat_store_byte v[86:87], v31 offset:1051967; ALIGNED-NEXT:    flat_store_byte v[86:87], v30 offset:1011968; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v29 offset:961969; ALIGNED-NEXT:    flat_store_byte v[86:87], v29 offset:971970; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v28 offset:921971; ALIGNED-NEXT:    flat_store_byte v[86:87], v28 offset:931972; ALIGNED-NEXT:    s_waitcnt vmcnt(0)1973; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v101974; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:961975; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v61976; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v251977; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v191978; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v141979; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v101980; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:941981; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v61982; ALIGNED-NEXT:    s_and_b32 vcc_lo, exec_lo, s61983; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v251984; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v191985; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v241986; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v241987; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v171988; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v151989; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:1041990; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v111991; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:1021992; ALIGNED-NEXT:    flat_store_byte v[86:87], v48 offset:1081993; ALIGNED-NEXT:    flat_store_byte v[86:87], v49 offset:1061994; ALIGNED-NEXT:    flat_store_byte v[86:87], v50 offset:1001995; ALIGNED-NEXT:    flat_store_byte v[86:87], v51 offset:981996; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:81997; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:121998; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:41999; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s322000; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v26 offset:842001; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v27 offset:882002; ALIGNED-NEXT:    flat_store_byte v[86:87], v27 offset:892003; ALIGNED-NEXT:    flat_store_byte v[86:87], v26 offset:852004; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v25 offset:802005; ALIGNED-NEXT:    flat_store_byte v[86:87], v25 offset:812006; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v24 offset:762007; ALIGNED-NEXT:    flat_store_byte v[86:87], v24 offset:772008; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v72009; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v222010; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v222011; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v232012; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v232013; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v212014; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v212015; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v202016; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v202017; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v172018; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v162019; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v162020; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v152021; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v132022; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v132023; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v122024; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v122025; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v112026; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v92027; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v92028; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v82029; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v82030; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v72031; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v52032; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v52033; ALIGNED-NEXT:    flat_store_byte v[86:87], v64 offset:882034; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v42035; ALIGNED-NEXT:    flat_store_byte v[86:87], v65 offset:862036; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v42037; ALIGNED-NEXT:    flat_store_byte v[86:87], v66 offset:922038; ALIGNED-NEXT:    flat_store_byte v[86:87], v67 offset:902039; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:842040; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:822041; ALIGNED-NEXT:    flat_store_byte v[86:87], v36 offset:802042; ALIGNED-NEXT:    flat_store_byte v[86:87], v37 offset:782043; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:242044; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:282045; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:202046; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:162047; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v22 offset:682048; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v23 offset:722049; ALIGNED-NEXT:    flat_store_byte v[86:87], v23 offset:732050; ALIGNED-NEXT:    flat_store_byte v[86:87], v22 offset:692051; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v21 offset:642052; ALIGNED-NEXT:    flat_store_byte v[86:87], v21 offset:652053; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v20 offset:602054; ALIGNED-NEXT:    flat_store_byte v[84:85], v20 offset:642055; ALIGNED-NEXT:    flat_store_byte v[86:87], v38 offset:722056; ALIGNED-NEXT:    flat_store_byte v[86:87], v39 offset:702057; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:762058; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:742059; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:682060; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:662061; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:642062; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:622063; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:1042064; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:1082065; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:1002066; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:962067; ALIGNED-NEXT:    flat_store_byte v[86:87], v52 offset:562068; ALIGNED-NEXT:    flat_store_byte v[86:87], v53 offset:542069; ALIGNED-NEXT:    flat_store_byte v[86:87], v32 offset:602070; ALIGNED-NEXT:    flat_store_byte v[86:87], v33 offset:582071; ALIGNED-NEXT:    flat_store_byte v[86:87], v34 offset:522072; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v18 offset:522073; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v19 offset:562074; ALIGNED-NEXT:    flat_store_byte v[86:87], v19 offset:572075; ALIGNED-NEXT:    flat_store_byte v[86:87], v18 offset:532076; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v17 offset:482077; ALIGNED-NEXT:    flat_store_byte v[86:87], v35 offset:502078; ALIGNED-NEXT:    flat_store_byte v[86:87], v17 offset:492079; ALIGNED-NEXT:    flat_store_byte v[86:87], v54 offset:482080; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v16 offset:442081; ALIGNED-NEXT:    flat_store_byte v[86:87], v55 offset:462082; ALIGNED-NEXT:    flat_store_byte v[86:87], v16 offset:452083; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:1202084; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:1242085; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1162086; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1122087; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v14 offset:362088; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:402089; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:382090; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v15 offset:402091; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:442092; ALIGNED-NEXT:    flat_store_byte v[86:87], v15 offset:412093; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:422094; ALIGNED-NEXT:    flat_store_byte v[86:87], v14 offset:372095; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:362096; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v13 offset:322097; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:342098; ALIGNED-NEXT:    flat_store_byte v[86:87], v13 offset:332099; ALIGNED-NEXT:    flat_store_byte v[86:87], v28 offset:322100; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v12 offset:282101; ALIGNED-NEXT:    flat_store_byte v[86:87], v29 offset:302102; ALIGNED-NEXT:    flat_store_byte v[84:85], v12 offset:322103; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:722104; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:762105; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:682106; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:642107; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v10 offset:202108; ALIGNED-NEXT:    flat_store_byte v[86:87], v30 offset:242109; ALIGNED-NEXT:    flat_store_byte v[86:87], v31 offset:222110; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v11 offset:242111; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:282112; ALIGNED-NEXT:    flat_store_byte v[86:87], v11 offset:252113; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:262114; ALIGNED-NEXT:    flat_store_byte v[86:87], v10 offset:212115; ALIGNED-NEXT:    flat_store_byte v[86:87], v48 offset:202116; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v9 offset:162117; ALIGNED-NEXT:    flat_store_byte v[86:87], v49 offset:182118; ALIGNED-NEXT:    flat_store_byte v[86:87], v9 offset:172119; ALIGNED-NEXT:    flat_store_byte v[86:87], v50 offset:162120; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v8 offset:122121; ALIGNED-NEXT:    flat_store_byte v[86:87], v51 offset:142122; ALIGNED-NEXT:    flat_store_byte v[84:85], v8 offset:162123; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:882124; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:922125; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:842126; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:802127; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v6 offset:42128; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:82129; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:62130; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v7 offset:82131; ALIGNED-NEXT:    flat_store_byte v[86:87], v24 offset:122132; ALIGNED-NEXT:    flat_store_byte v[86:87], v7 offset:92133; ALIGNED-NEXT:    flat_store_byte v[86:87], v25 offset:102134; ALIGNED-NEXT:    flat_store_byte v[84:85], v6 offset:82135; ALIGNED-NEXT:    flat_store_byte v[86:87], v26 offset:42136; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v52137; ALIGNED-NEXT:    flat_store_byte v[86:87], v27 offset:22138; ALIGNED-NEXT:    flat_store_byte v[84:85], v5 offset:42139; ALIGNED-NEXT:    flat_store_byte v[86:87], v642140; ALIGNED-NEXT:    flat_store_byte_d16_hi v[84:85], v4 offset:22141; ALIGNED-NEXT:    flat_store_byte v[84:85], v65 offset:12142; ALIGNED-NEXT:    flat_store_byte v[84:85], v42143; ALIGNED-NEXT:    s_cbranch_vccnz .LBB2_12144; ALIGNED-NEXT:  ; %bb.2: ; %memcpy-split2145; ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)2146; ALIGNED-NEXT:    s_setpc_b64 s[30:31]2147;2148; UNROLL3-LABEL: memcpy_p0_p4_sz2048:2149; UNROLL3:       ; %bb.0: ; %entry2150; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2151; UNROLL3-NEXT:    s_mov_b64 s[4:5], 02152; UNROLL3-NEXT:    .p2align 62153; UNROLL3-NEXT:  .LBB2_1: ; %load-store-loop2154; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=12155; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s42156; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo2157; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s42158; UNROLL3-NEXT:    s_add_u32 s4, s4, 482159; UNROLL3-NEXT:    s_clause 0x22160; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off offset:162161; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off2162; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:322163; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo2164; UNROLL3-NEXT:    s_addc_u32 s5, s5, 02165; UNROLL3-NEXT:    s_waitcnt vmcnt(2)2166; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7] offset:162167; UNROLL3-NEXT:    s_waitcnt vmcnt(1)2168; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11]2169; UNROLL3-NEXT:    s_waitcnt vmcnt(0)2170; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:322171; UNROLL3-NEXT:    v_cmp_gt_u64_e64 s6, 0x7e0, s[4:5]2172; UNROLL3-NEXT:    s_and_b32 vcc_lo, exec_lo, s62173; UNROLL3-NEXT:    s_cbranch_vccnz .LBB2_12174; UNROLL3-NEXT:  ; %bb.2: ; %memcpy-split2175; UNROLL3-NEXT:    s_clause 0x12176; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:20162177; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:20322178; UNROLL3-NEXT:    s_waitcnt vmcnt(1)2179; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:20162180; UNROLL3-NEXT:    s_waitcnt vmcnt(0)2181; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[8:11] offset:20322182; UNROLL3-NEXT:    s_waitcnt lgkmcnt(0)2183; UNROLL3-NEXT:    s_setpc_b64 s[30:31]2184entry:2185  tail call void @llvm.memcpy.p0.p4.i64(ptr addrspace(0) noundef nonnull align 1 %dst, ptr addrspace(4) noundef nonnull align 1 %src, i64 2048, i1 false)2186  ret void2187}2188 2189define void @memcpy_p5_p5_sz2048(ptr addrspace(5) align 1 %dst, ptr addrspace(5) align 1 readonly %src) {2190; CHECK-LABEL: memcpy_p5_p5_sz2048:2191; CHECK:       ; %bb.0: ; %entry2192; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2193; CHECK-NEXT:    s_mov_b64 s[4:5], 02194; CHECK-NEXT:  .LBB3_1: ; %load-store-loop2195; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=12196; CHECK-NEXT:    s_clause 0x3e2197; CHECK-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:2522198; CHECK-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:2482199; CHECK-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:2442200; CHECK-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:2402201; CHECK-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:2362202; CHECK-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:2322203; CHECK-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:2282204; CHECK-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:2242205; CHECK-NEXT:    buffer_load_dword v10, v1, s[0:3], 0 offen offset:2202206; CHECK-NEXT:    buffer_load_dword v11, v1, s[0:3], 0 offen offset:2162207; CHECK-NEXT:    buffer_load_dword v12, v1, s[0:3], 0 offen offset:2122208; CHECK-NEXT:    buffer_load_dword v13, v1, s[0:3], 0 offen offset:2082209; CHECK-NEXT:    buffer_load_dword v14, v1, s[0:3], 0 offen offset:2042210; CHECK-NEXT:    buffer_load_dword v15, v1, s[0:3], 0 offen offset:2002211; CHECK-NEXT:    buffer_load_dword v16, v1, s[0:3], 0 offen offset:1962212; CHECK-NEXT:    buffer_load_dword v17, v1, s[0:3], 0 offen offset:1922213; CHECK-NEXT:    buffer_load_dword v18, v1, s[0:3], 0 offen offset:1882214; CHECK-NEXT:    buffer_load_dword v19, v1, s[0:3], 0 offen offset:1842215; CHECK-NEXT:    buffer_load_dword v20, v1, s[0:3], 0 offen offset:1802216; CHECK-NEXT:    buffer_load_dword v21, v1, s[0:3], 0 offen offset:1762217; CHECK-NEXT:    buffer_load_dword v22, v1, s[0:3], 0 offen offset:1722218; CHECK-NEXT:    buffer_load_dword v23, v1, s[0:3], 0 offen offset:1682219; CHECK-NEXT:    buffer_load_dword v24, v1, s[0:3], 0 offen offset:1642220; CHECK-NEXT:    buffer_load_dword v25, v1, s[0:3], 0 offen offset:1602221; CHECK-NEXT:    buffer_load_dword v26, v1, s[0:3], 0 offen offset:1562222; CHECK-NEXT:    buffer_load_dword v27, v1, s[0:3], 0 offen offset:1522223; CHECK-NEXT:    buffer_load_dword v28, v1, s[0:3], 0 offen offset:1482224; CHECK-NEXT:    buffer_load_dword v29, v1, s[0:3], 0 offen offset:1442225; CHECK-NEXT:    buffer_load_dword v30, v1, s[0:3], 0 offen offset:1402226; CHECK-NEXT:    buffer_load_dword v31, v1, s[0:3], 0 offen offset:1362227; CHECK-NEXT:    buffer_load_dword v32, v1, s[0:3], 0 offen offset:1322228; CHECK-NEXT:    buffer_load_dword v33, v1, s[0:3], 0 offen offset:1282229; CHECK-NEXT:    buffer_load_dword v34, v1, s[0:3], 0 offen offset:1242230; CHECK-NEXT:    buffer_load_dword v35, v1, s[0:3], 0 offen offset:1202231; CHECK-NEXT:    buffer_load_dword v36, v1, s[0:3], 0 offen offset:1162232; CHECK-NEXT:    buffer_load_dword v37, v1, s[0:3], 0 offen offset:1122233; CHECK-NEXT:    buffer_load_dword v38, v1, s[0:3], 0 offen offset:1082234; CHECK-NEXT:    buffer_load_dword v39, v1, s[0:3], 0 offen offset:1042235; CHECK-NEXT:    buffer_load_dword v48, v1, s[0:3], 0 offen offset:1002236; CHECK-NEXT:    buffer_load_dword v49, v1, s[0:3], 0 offen offset:962237; CHECK-NEXT:    buffer_load_dword v50, v1, s[0:3], 0 offen offset:922238; CHECK-NEXT:    buffer_load_dword v51, v1, s[0:3], 0 offen offset:882239; CHECK-NEXT:    buffer_load_dword v52, v1, s[0:3], 0 offen offset:842240; CHECK-NEXT:    buffer_load_dword v53, v1, s[0:3], 0 offen offset:802241; CHECK-NEXT:    buffer_load_dword v54, v1, s[0:3], 0 offen offset:762242; CHECK-NEXT:    buffer_load_dword v55, v1, s[0:3], 0 offen offset:722243; CHECK-NEXT:    buffer_load_dword v64, v1, s[0:3], 0 offen offset:682244; CHECK-NEXT:    buffer_load_dword v65, v1, s[0:3], 0 offen offset:642245; CHECK-NEXT:    buffer_load_dword v66, v1, s[0:3], 0 offen offset:602246; CHECK-NEXT:    buffer_load_dword v67, v1, s[0:3], 0 offen offset:562247; CHECK-NEXT:    buffer_load_dword v68, v1, s[0:3], 0 offen offset:522248; CHECK-NEXT:    buffer_load_dword v69, v1, s[0:3], 0 offen offset:482249; CHECK-NEXT:    buffer_load_dword v70, v1, s[0:3], 0 offen offset:442250; CHECK-NEXT:    buffer_load_dword v71, v1, s[0:3], 0 offen offset:402251; CHECK-NEXT:    buffer_load_dword v80, v1, s[0:3], 0 offen offset:362252; CHECK-NEXT:    buffer_load_dword v81, v1, s[0:3], 0 offen offset:322253; CHECK-NEXT:    buffer_load_dword v82, v1, s[0:3], 0 offen offset:282254; CHECK-NEXT:    buffer_load_dword v83, v1, s[0:3], 0 offen offset:242255; CHECK-NEXT:    buffer_load_dword v84, v1, s[0:3], 0 offen offset:202256; CHECK-NEXT:    buffer_load_dword v85, v1, s[0:3], 0 offen offset:162257; CHECK-NEXT:    buffer_load_dword v86, v1, s[0:3], 0 offen offset:122258; CHECK-NEXT:    buffer_load_dword v87, v1, s[0:3], 0 offen offset:82259; CHECK-NEXT:    buffer_load_dword v96, v1, s[0:3], 0 offen offset:42260; CHECK-NEXT:    buffer_load_dword v97, v1, s[0:3], 0 offen2261; CHECK-NEXT:    s_add_u32 s4, s4, 0x1002262; CHECK-NEXT:    s_addc_u32 s5, s5, 02263; CHECK-NEXT:    v_add_nc_u32_e32 v1, 0x100, v12264; CHECK-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]2265; CHECK-NEXT:    s_waitcnt vmcnt(62)2266; CHECK-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:2522267; CHECK-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:2482268; CHECK-NEXT:    s_waitcnt vmcnt(61)2269; CHECK-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:2442270; CHECK-NEXT:    s_waitcnt vmcnt(60)2271; CHECK-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:2402272; CHECK-NEXT:    s_waitcnt vmcnt(59)2273; CHECK-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:2362274; CHECK-NEXT:    s_waitcnt vmcnt(58)2275; CHECK-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:2322276; CHECK-NEXT:    s_waitcnt vmcnt(57)2277; CHECK-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:2282278; CHECK-NEXT:    s_waitcnt vmcnt(56)2279; CHECK-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:2242280; CHECK-NEXT:    s_waitcnt vmcnt(55)2281; CHECK-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:2202282; CHECK-NEXT:    s_waitcnt vmcnt(54)2283; CHECK-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:2162284; CHECK-NEXT:    s_waitcnt vmcnt(53)2285; CHECK-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:2122286; CHECK-NEXT:    s_waitcnt vmcnt(52)2287; CHECK-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:2082288; CHECK-NEXT:    s_waitcnt vmcnt(51)2289; CHECK-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:2042290; CHECK-NEXT:    s_waitcnt vmcnt(50)2291; CHECK-NEXT:    buffer_store_dword v15, v0, s[0:3], 0 offen offset:2002292; CHECK-NEXT:    s_waitcnt vmcnt(49)2293; CHECK-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:1962294; CHECK-NEXT:    s_waitcnt vmcnt(48)2295; CHECK-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:1922296; CHECK-NEXT:    s_waitcnt vmcnt(47)2297; CHECK-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:1882298; CHECK-NEXT:    s_waitcnt vmcnt(46)2299; CHECK-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:1842300; CHECK-NEXT:    s_waitcnt vmcnt(45)2301; CHECK-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen offset:1802302; CHECK-NEXT:    s_waitcnt vmcnt(44)2303; CHECK-NEXT:    buffer_store_dword v21, v0, s[0:3], 0 offen offset:1762304; CHECK-NEXT:    s_waitcnt vmcnt(43)2305; CHECK-NEXT:    buffer_store_dword v22, v0, s[0:3], 0 offen offset:1722306; CHECK-NEXT:    s_waitcnt vmcnt(42)2307; CHECK-NEXT:    buffer_store_dword v23, v0, s[0:3], 0 offen offset:1682308; CHECK-NEXT:    s_waitcnt vmcnt(41)2309; CHECK-NEXT:    buffer_store_dword v24, v0, s[0:3], 0 offen offset:1642310; CHECK-NEXT:    s_waitcnt vmcnt(40)2311; CHECK-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:1602312; CHECK-NEXT:    s_waitcnt vmcnt(39)2313; CHECK-NEXT:    buffer_store_dword v26, v0, s[0:3], 0 offen offset:1562314; CHECK-NEXT:    s_waitcnt vmcnt(38)2315; CHECK-NEXT:    buffer_store_dword v27, v0, s[0:3], 0 offen offset:1522316; CHECK-NEXT:    s_waitcnt vmcnt(37)2317; CHECK-NEXT:    buffer_store_dword v28, v0, s[0:3], 0 offen offset:1482318; CHECK-NEXT:    s_waitcnt vmcnt(36)2319; CHECK-NEXT:    buffer_store_dword v29, v0, s[0:3], 0 offen offset:1442320; CHECK-NEXT:    s_waitcnt vmcnt(35)2321; CHECK-NEXT:    buffer_store_dword v30, v0, s[0:3], 0 offen offset:1402322; CHECK-NEXT:    s_waitcnt vmcnt(34)2323; CHECK-NEXT:    buffer_store_dword v31, v0, s[0:3], 0 offen offset:1362324; CHECK-NEXT:    s_waitcnt vmcnt(33)2325; CHECK-NEXT:    buffer_store_dword v32, v0, s[0:3], 0 offen offset:1322326; CHECK-NEXT:    s_waitcnt vmcnt(32)2327; CHECK-NEXT:    buffer_store_dword v33, v0, s[0:3], 0 offen offset:1282328; CHECK-NEXT:    s_waitcnt vmcnt(31)2329; CHECK-NEXT:    buffer_store_dword v34, v0, s[0:3], 0 offen offset:1242330; CHECK-NEXT:    s_waitcnt vmcnt(30)2331; CHECK-NEXT:    buffer_store_dword v35, v0, s[0:3], 0 offen offset:1202332; CHECK-NEXT:    s_waitcnt vmcnt(29)2333; CHECK-NEXT:    buffer_store_dword v36, v0, s[0:3], 0 offen offset:1162334; CHECK-NEXT:    s_waitcnt vmcnt(28)2335; CHECK-NEXT:    buffer_store_dword v37, v0, s[0:3], 0 offen offset:1122336; CHECK-NEXT:    s_waitcnt vmcnt(27)2337; CHECK-NEXT:    buffer_store_dword v38, v0, s[0:3], 0 offen offset:1082338; CHECK-NEXT:    s_waitcnt vmcnt(26)2339; CHECK-NEXT:    buffer_store_dword v39, v0, s[0:3], 0 offen offset:1042340; CHECK-NEXT:    s_waitcnt vmcnt(25)2341; CHECK-NEXT:    buffer_store_dword v48, v0, s[0:3], 0 offen offset:1002342; CHECK-NEXT:    s_waitcnt vmcnt(24)2343; CHECK-NEXT:    buffer_store_dword v49, v0, s[0:3], 0 offen offset:962344; CHECK-NEXT:    s_waitcnt vmcnt(23)2345; CHECK-NEXT:    buffer_store_dword v50, v0, s[0:3], 0 offen offset:922346; CHECK-NEXT:    s_waitcnt vmcnt(22)2347; CHECK-NEXT:    buffer_store_dword v51, v0, s[0:3], 0 offen offset:882348; CHECK-NEXT:    s_waitcnt vmcnt(21)2349; CHECK-NEXT:    buffer_store_dword v52, v0, s[0:3], 0 offen offset:842350; CHECK-NEXT:    s_waitcnt vmcnt(20)2351; CHECK-NEXT:    buffer_store_dword v53, v0, s[0:3], 0 offen offset:802352; CHECK-NEXT:    s_waitcnt vmcnt(19)2353; CHECK-NEXT:    buffer_store_dword v54, v0, s[0:3], 0 offen offset:762354; CHECK-NEXT:    s_waitcnt vmcnt(18)2355; CHECK-NEXT:    buffer_store_dword v55, v0, s[0:3], 0 offen offset:722356; CHECK-NEXT:    s_waitcnt vmcnt(17)2357; CHECK-NEXT:    buffer_store_dword v64, v0, s[0:3], 0 offen offset:682358; CHECK-NEXT:    s_waitcnt vmcnt(16)2359; CHECK-NEXT:    buffer_store_dword v65, v0, s[0:3], 0 offen offset:642360; CHECK-NEXT:    s_waitcnt vmcnt(15)2361; CHECK-NEXT:    buffer_store_dword v66, v0, s[0:3], 0 offen offset:602362; CHECK-NEXT:    s_waitcnt vmcnt(14)2363; CHECK-NEXT:    buffer_store_dword v67, v0, s[0:3], 0 offen offset:562364; CHECK-NEXT:    s_waitcnt vmcnt(13)2365; CHECK-NEXT:    buffer_store_dword v68, v0, s[0:3], 0 offen offset:522366; CHECK-NEXT:    s_waitcnt vmcnt(12)2367; CHECK-NEXT:    buffer_store_dword v69, v0, s[0:3], 0 offen offset:482368; CHECK-NEXT:    s_waitcnt vmcnt(11)2369; CHECK-NEXT:    buffer_store_dword v70, v0, s[0:3], 0 offen offset:442370; CHECK-NEXT:    s_waitcnt vmcnt(10)2371; CHECK-NEXT:    buffer_store_dword v71, v0, s[0:3], 0 offen offset:402372; CHECK-NEXT:    s_waitcnt vmcnt(9)2373; CHECK-NEXT:    buffer_store_dword v80, v0, s[0:3], 0 offen offset:362374; CHECK-NEXT:    s_waitcnt vmcnt(8)2375; CHECK-NEXT:    buffer_store_dword v81, v0, s[0:3], 0 offen offset:322376; CHECK-NEXT:    s_waitcnt vmcnt(7)2377; CHECK-NEXT:    buffer_store_dword v82, v0, s[0:3], 0 offen offset:282378; CHECK-NEXT:    s_waitcnt vmcnt(6)2379; CHECK-NEXT:    buffer_store_dword v83, v0, s[0:3], 0 offen offset:242380; CHECK-NEXT:    s_waitcnt vmcnt(5)2381; CHECK-NEXT:    buffer_store_dword v84, v0, s[0:3], 0 offen offset:202382; CHECK-NEXT:    s_waitcnt vmcnt(4)2383; CHECK-NEXT:    buffer_store_dword v85, v0, s[0:3], 0 offen offset:162384; CHECK-NEXT:    s_waitcnt vmcnt(3)2385; CHECK-NEXT:    buffer_store_dword v86, v0, s[0:3], 0 offen offset:122386; CHECK-NEXT:    s_waitcnt vmcnt(2)2387; CHECK-NEXT:    buffer_store_dword v87, v0, s[0:3], 0 offen offset:82388; CHECK-NEXT:    s_waitcnt vmcnt(1)2389; CHECK-NEXT:    buffer_store_dword v96, v0, s[0:3], 0 offen offset:42390; CHECK-NEXT:    s_waitcnt vmcnt(0)2391; CHECK-NEXT:    buffer_store_dword v97, v0, s[0:3], 0 offen2392; CHECK-NEXT:    v_add_nc_u32_e32 v0, 0x100, v02393; CHECK-NEXT:    s_and_b32 vcc_lo, exec_lo, s62394; CHECK-NEXT:    s_cbranch_vccnz .LBB3_12395; CHECK-NEXT:  ; %bb.2: ; %memcpy-split2396; CHECK-NEXT:    s_setpc_b64 s[30:31]2397;2398; ALIGNED-LABEL: memcpy_p5_p5_sz2048:2399; ALIGNED:       ; %bb.0: ; %entry2400; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)2401; ALIGNED-NEXT:    s_mov_b64 s[4:5], 02402; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill2403; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill2404; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill2405; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill2406; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill2407; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill2408; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill2409; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill2410; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill2411; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill2412; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill2413; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill2414; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill2415; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill2416; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill2417; ALIGNED-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill2418; ALIGNED-NEXT:    buffer_store_dword v72, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill2419; ALIGNED-NEXT:    buffer_store_dword v73, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill2420; ALIGNED-NEXT:    buffer_store_dword v74, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill2421; ALIGNED-NEXT:    buffer_store_dword v75, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill2422; ALIGNED-NEXT:    buffer_store_dword v76, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill2423; ALIGNED-NEXT:    buffer_store_dword v77, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill2424; ALIGNED-NEXT:    buffer_store_dword v78, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill2425; ALIGNED-NEXT:    buffer_store_dword v79, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill2426; ALIGNED-NEXT:    buffer_store_dword v88, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill2427; ALIGNED-NEXT:    buffer_store_dword v89, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill2428; ALIGNED-NEXT:    buffer_store_dword v90, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill2429; ALIGNED-NEXT:    buffer_store_dword v91, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill2430; ALIGNED-NEXT:    buffer_store_dword v92, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill2431; ALIGNED-NEXT:    buffer_store_dword v93, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill2432; ALIGNED-NEXT:    buffer_store_dword v94, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill2433; ALIGNED-NEXT:    buffer_store_dword v95, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill2434; ALIGNED-NEXT:    buffer_store_dword v104, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill2435; ALIGNED-NEXT:    buffer_store_dword v105, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill2436; ALIGNED-NEXT:    buffer_store_dword v106, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill2437; ALIGNED-NEXT:    buffer_store_dword v107, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill2438; ALIGNED-NEXT:    buffer_store_dword v108, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill2439; ALIGNED-NEXT:    buffer_store_dword v109, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill2440; ALIGNED-NEXT:    buffer_store_dword v110, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill2441; ALIGNED-NEXT:    buffer_store_dword v111, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill2442; ALIGNED-NEXT:    buffer_store_dword v120, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill2443; ALIGNED-NEXT:    buffer_store_dword v121, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill2444; ALIGNED-NEXT:    buffer_store_dword v122, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill2445; ALIGNED-NEXT:    buffer_store_dword v123, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill2446; ALIGNED-NEXT:    buffer_store_dword v124, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill2447; ALIGNED-NEXT:    buffer_store_dword v125, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill2448; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill2449; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 ; 4-byte Folded Spill2450; ALIGNED-NEXT:  .LBB3_1: ; %load-store-loop2451; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=12452; ALIGNED-NEXT:    s_clause 0x342453; ALIGNED-NEXT:    buffer_load_ubyte v116, v1, s[0:3], 0 offen offset:2552454; ALIGNED-NEXT:    buffer_load_ubyte v117, v1, s[0:3], 0 offen offset:2542455; ALIGNED-NEXT:    buffer_load_ubyte v118, v1, s[0:3], 0 offen offset:2532456; ALIGNED-NEXT:    buffer_load_ubyte v119, v1, s[0:3], 0 offen offset:2522457; ALIGNED-NEXT:    buffer_load_ubyte v40, v1, s[0:3], 0 offen offset:2512458; ALIGNED-NEXT:    buffer_load_ubyte v41, v1, s[0:3], 0 offen offset:2502459; ALIGNED-NEXT:    buffer_load_ubyte v42, v1, s[0:3], 0 offen offset:2492460; ALIGNED-NEXT:    buffer_load_ubyte v43, v1, s[0:3], 0 offen offset:2482461; ALIGNED-NEXT:    buffer_load_ubyte v44, v1, s[0:3], 0 offen offset:2472462; ALIGNED-NEXT:    buffer_load_ubyte v45, v1, s[0:3], 0 offen offset:2462463; ALIGNED-NEXT:    buffer_load_ubyte v46, v1, s[0:3], 0 offen offset:2452464; ALIGNED-NEXT:    buffer_load_ubyte v47, v1, s[0:3], 0 offen offset:2442465; ALIGNED-NEXT:    buffer_load_ubyte v56, v1, s[0:3], 0 offen offset:2432466; ALIGNED-NEXT:    buffer_load_ubyte v57, v1, s[0:3], 0 offen offset:2422467; ALIGNED-NEXT:    buffer_load_ubyte v58, v1, s[0:3], 0 offen offset:2412468; ALIGNED-NEXT:    buffer_load_ubyte v59, v1, s[0:3], 0 offen offset:2402469; ALIGNED-NEXT:    buffer_load_ubyte v60, v1, s[0:3], 0 offen offset:2392470; ALIGNED-NEXT:    buffer_load_ubyte v61, v1, s[0:3], 0 offen offset:2382471; ALIGNED-NEXT:    buffer_load_ubyte v62, v1, s[0:3], 0 offen offset:2372472; ALIGNED-NEXT:    buffer_load_ubyte v63, v1, s[0:3], 0 offen offset:2362473; ALIGNED-NEXT:    buffer_load_ubyte v72, v1, s[0:3], 0 offen offset:2352474; ALIGNED-NEXT:    buffer_load_ubyte v73, v1, s[0:3], 0 offen offset:2342475; ALIGNED-NEXT:    buffer_load_ubyte v74, v1, s[0:3], 0 offen offset:2332476; ALIGNED-NEXT:    buffer_load_ubyte v75, v1, s[0:3], 0 offen offset:2322477; ALIGNED-NEXT:    buffer_load_ubyte v76, v1, s[0:3], 0 offen offset:2312478; ALIGNED-NEXT:    buffer_load_ubyte v77, v1, s[0:3], 0 offen offset:2302479; ALIGNED-NEXT:    buffer_load_ubyte v78, v1, s[0:3], 0 offen offset:2292480; ALIGNED-NEXT:    buffer_load_ubyte v79, v1, s[0:3], 0 offen offset:2282481; ALIGNED-NEXT:    buffer_load_ubyte v88, v1, s[0:3], 0 offen offset:2272482; ALIGNED-NEXT:    buffer_load_ubyte v89, v1, s[0:3], 0 offen offset:2262483; ALIGNED-NEXT:    buffer_load_ubyte v90, v1, s[0:3], 0 offen offset:2252484; ALIGNED-NEXT:    buffer_load_ubyte v91, v1, s[0:3], 0 offen offset:2242485; ALIGNED-NEXT:    buffer_load_ubyte v92, v1, s[0:3], 0 offen offset:2232486; ALIGNED-NEXT:    buffer_load_ubyte v93, v1, s[0:3], 0 offen offset:2222487; ALIGNED-NEXT:    buffer_load_ubyte v94, v1, s[0:3], 0 offen offset:2212488; ALIGNED-NEXT:    buffer_load_ubyte v95, v1, s[0:3], 0 offen offset:2202489; ALIGNED-NEXT:    buffer_load_ubyte v104, v1, s[0:3], 0 offen offset:2192490; ALIGNED-NEXT:    buffer_load_ubyte v105, v1, s[0:3], 0 offen offset:2182491; ALIGNED-NEXT:    buffer_load_ubyte v106, v1, s[0:3], 0 offen offset:2172492; ALIGNED-NEXT:    buffer_load_ubyte v107, v1, s[0:3], 0 offen offset:2162493; ALIGNED-NEXT:    buffer_load_ubyte v108, v1, s[0:3], 0 offen offset:2152494; ALIGNED-NEXT:    buffer_load_ubyte v109, v1, s[0:3], 0 offen offset:2142495; ALIGNED-NEXT:    buffer_load_ubyte v110, v1, s[0:3], 0 offen offset:2132496; ALIGNED-NEXT:    buffer_load_ubyte v111, v1, s[0:3], 0 offen offset:2122497; ALIGNED-NEXT:    buffer_load_ubyte v120, v1, s[0:3], 0 offen offset:2112498; ALIGNED-NEXT:    buffer_load_ubyte v121, v1, s[0:3], 0 offen offset:2102499; ALIGNED-NEXT:    buffer_load_ubyte v122, v1, s[0:3], 0 offen offset:2092500; ALIGNED-NEXT:    buffer_load_ubyte v123, v1, s[0:3], 0 offen offset:2082501; ALIGNED-NEXT:    buffer_load_ubyte v124, v1, s[0:3], 0 offen offset:2072502; ALIGNED-NEXT:    buffer_load_ubyte v125, v1, s[0:3], 0 offen offset:2062503; ALIGNED-NEXT:    buffer_load_ubyte v126, v1, s[0:3], 0 offen offset:2052504; ALIGNED-NEXT:    buffer_load_ubyte v127, v1, s[0:3], 0 offen offset:2042505; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2032506; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x1002507; ALIGNED-NEXT:    s_addc_u32 s5, s5, 02508; ALIGNED-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]2509; ALIGNED-NEXT:    s_and_b32 vcc_lo, exec_lo, s62510; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2511; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill2512; ALIGNED-NEXT:    s_clause 0x3e2513; ALIGNED-NEXT:    buffer_load_ubyte v3, v1, s[0:3], 0 offen offset:2022514; ALIGNED-NEXT:    buffer_load_ubyte v4, v1, s[0:3], 0 offen offset:2012515; ALIGNED-NEXT:    buffer_load_ubyte v5, v1, s[0:3], 0 offen offset:2002516; ALIGNED-NEXT:    buffer_load_ubyte v6, v1, s[0:3], 0 offen offset:1992517; ALIGNED-NEXT:    buffer_load_ubyte v7, v1, s[0:3], 0 offen offset:1982518; ALIGNED-NEXT:    buffer_load_ubyte v8, v1, s[0:3], 0 offen offset:1972519; ALIGNED-NEXT:    buffer_load_ubyte v9, v1, s[0:3], 0 offen offset:1962520; ALIGNED-NEXT:    buffer_load_ubyte v10, v1, s[0:3], 0 offen offset:1952521; ALIGNED-NEXT:    buffer_load_ubyte v11, v1, s[0:3], 0 offen offset:1942522; ALIGNED-NEXT:    buffer_load_ubyte v12, v1, s[0:3], 0 offen offset:1932523; ALIGNED-NEXT:    buffer_load_ubyte v13, v1, s[0:3], 0 offen offset:1922524; ALIGNED-NEXT:    buffer_load_ubyte v14, v1, s[0:3], 0 offen offset:1912525; ALIGNED-NEXT:    buffer_load_ubyte v15, v1, s[0:3], 0 offen offset:1902526; ALIGNED-NEXT:    buffer_load_ubyte v16, v1, s[0:3], 0 offen offset:1892527; ALIGNED-NEXT:    buffer_load_ubyte v17, v1, s[0:3], 0 offen offset:1882528; ALIGNED-NEXT:    buffer_load_ubyte v18, v1, s[0:3], 0 offen offset:1872529; ALIGNED-NEXT:    buffer_load_ubyte v19, v1, s[0:3], 0 offen offset:1862530; ALIGNED-NEXT:    buffer_load_ubyte v20, v1, s[0:3], 0 offen offset:1852531; ALIGNED-NEXT:    buffer_load_ubyte v21, v1, s[0:3], 0 offen offset:1842532; ALIGNED-NEXT:    buffer_load_ubyte v22, v1, s[0:3], 0 offen offset:1832533; ALIGNED-NEXT:    buffer_load_ubyte v23, v1, s[0:3], 0 offen offset:1822534; ALIGNED-NEXT:    buffer_load_ubyte v24, v1, s[0:3], 0 offen offset:1812535; ALIGNED-NEXT:    buffer_load_ubyte v25, v1, s[0:3], 0 offen offset:1802536; ALIGNED-NEXT:    buffer_load_ubyte v26, v1, s[0:3], 0 offen offset:1792537; ALIGNED-NEXT:    buffer_load_ubyte v27, v1, s[0:3], 0 offen offset:1782538; ALIGNED-NEXT:    buffer_load_ubyte v28, v1, s[0:3], 0 offen offset:1772539; ALIGNED-NEXT:    buffer_load_ubyte v29, v1, s[0:3], 0 offen offset:1762540; ALIGNED-NEXT:    buffer_load_ubyte v30, v1, s[0:3], 0 offen offset:1752541; ALIGNED-NEXT:    buffer_load_ubyte v31, v1, s[0:3], 0 offen offset:1742542; ALIGNED-NEXT:    buffer_load_ubyte v32, v1, s[0:3], 0 offen offset:1732543; ALIGNED-NEXT:    buffer_load_ubyte v33, v1, s[0:3], 0 offen offset:1722544; ALIGNED-NEXT:    buffer_load_ubyte v34, v1, s[0:3], 0 offen offset:1712545; ALIGNED-NEXT:    buffer_load_ubyte v35, v1, s[0:3], 0 offen offset:1702546; ALIGNED-NEXT:    buffer_load_ubyte v36, v1, s[0:3], 0 offen offset:1692547; ALIGNED-NEXT:    buffer_load_ubyte v37, v1, s[0:3], 0 offen offset:1682548; ALIGNED-NEXT:    buffer_load_ubyte v38, v1, s[0:3], 0 offen offset:1672549; ALIGNED-NEXT:    buffer_load_ubyte v39, v1, s[0:3], 0 offen offset:1662550; ALIGNED-NEXT:    buffer_load_ubyte v48, v1, s[0:3], 0 offen offset:1652551; ALIGNED-NEXT:    buffer_load_ubyte v49, v1, s[0:3], 0 offen offset:1642552; ALIGNED-NEXT:    buffer_load_ubyte v50, v1, s[0:3], 0 offen offset:1632553; ALIGNED-NEXT:    buffer_load_ubyte v51, v1, s[0:3], 0 offen offset:1622554; ALIGNED-NEXT:    buffer_load_ubyte v52, v1, s[0:3], 0 offen offset:1612555; ALIGNED-NEXT:    buffer_load_ubyte v53, v1, s[0:3], 0 offen offset:1602556; ALIGNED-NEXT:    buffer_load_ubyte v54, v1, s[0:3], 0 offen offset:1592557; ALIGNED-NEXT:    buffer_load_ubyte v55, v1, s[0:3], 0 offen offset:1582558; ALIGNED-NEXT:    buffer_load_ubyte v64, v1, s[0:3], 0 offen offset:1572559; ALIGNED-NEXT:    buffer_load_ubyte v65, v1, s[0:3], 0 offen offset:1562560; ALIGNED-NEXT:    buffer_load_ubyte v66, v1, s[0:3], 0 offen offset:1552561; ALIGNED-NEXT:    buffer_load_ubyte v67, v1, s[0:3], 0 offen offset:1542562; ALIGNED-NEXT:    buffer_load_ubyte v68, v1, s[0:3], 0 offen offset:1532563; ALIGNED-NEXT:    buffer_load_ubyte v69, v1, s[0:3], 0 offen offset:1522564; ALIGNED-NEXT:    buffer_load_ubyte v70, v1, s[0:3], 0 offen offset:1512565; ALIGNED-NEXT:    buffer_load_ubyte v71, v1, s[0:3], 0 offen offset:1502566; ALIGNED-NEXT:    buffer_load_ubyte v80, v1, s[0:3], 0 offen offset:1492567; ALIGNED-NEXT:    buffer_load_ubyte v81, v1, s[0:3], 0 offen offset:1482568; ALIGNED-NEXT:    buffer_load_ubyte v82, v1, s[0:3], 0 offen offset:1472569; ALIGNED-NEXT:    buffer_load_ubyte v83, v1, s[0:3], 0 offen offset:1462570; ALIGNED-NEXT:    buffer_load_ubyte v84, v1, s[0:3], 0 offen offset:1452571; ALIGNED-NEXT:    buffer_load_ubyte v85, v1, s[0:3], 0 offen offset:1442572; ALIGNED-NEXT:    buffer_load_ubyte v86, v1, s[0:3], 0 offen offset:1432573; ALIGNED-NEXT:    buffer_load_ubyte v87, v1, s[0:3], 0 offen offset:1422574; ALIGNED-NEXT:    buffer_load_ubyte v96, v1, s[0:3], 0 offen offset:1412575; ALIGNED-NEXT:    buffer_load_ubyte v97, v1, s[0:3], 0 offen offset:1402576; ALIGNED-NEXT:    s_clause 0xa2577; ALIGNED-NEXT:    buffer_load_ubyte v98, v1, s[0:3], 0 offen offset:1392578; ALIGNED-NEXT:    buffer_load_ubyte v99, v1, s[0:3], 0 offen offset:1382579; ALIGNED-NEXT:    buffer_load_ubyte v100, v1, s[0:3], 0 offen offset:1372580; ALIGNED-NEXT:    buffer_load_ubyte v101, v1, s[0:3], 0 offen offset:1362581; ALIGNED-NEXT:    buffer_load_ubyte v102, v1, s[0:3], 0 offen offset:1352582; ALIGNED-NEXT:    buffer_load_ubyte v103, v1, s[0:3], 0 offen offset:1342583; ALIGNED-NEXT:    buffer_load_ubyte v112, v1, s[0:3], 0 offen offset:1332584; ALIGNED-NEXT:    buffer_load_ubyte v113, v1, s[0:3], 0 offen offset:1322585; ALIGNED-NEXT:    buffer_load_ubyte v114, v1, s[0:3], 0 offen offset:1312586; ALIGNED-NEXT:    buffer_load_ubyte v115, v1, s[0:3], 0 offen offset:1302587; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1292588; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2589; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill2590; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1282591; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2592; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill2593; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1272594; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2595; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill2596; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1262597; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2598; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill2599; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1252600; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2601; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill2602; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1242603; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2604; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill2605; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1232606; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2607; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill2608; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1222609; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2610; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill2611; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1212612; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2613; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill2614; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1202615; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2616; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill2617; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1192618; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2619; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill2620; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1182621; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2622; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill2623; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1172624; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2625; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill2626; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1162627; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2628; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill2629; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1152630; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2631; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill2632; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1142633; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2634; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill2635; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1132636; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2637; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill2638; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1122639; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2640; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill2641; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1112642; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2643; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill2644; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1102645; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2646; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill2647; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1092648; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2649; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill2650; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1082651; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2652; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill2653; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1072654; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2655; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill2656; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1062657; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2658; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill2659; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1052660; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2661; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill2662; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1042663; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2664; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill2665; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1032666; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2667; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill2668; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1022669; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2670; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill2671; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1012672; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2673; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill2674; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1002675; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2676; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill2677; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:992678; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2679; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill2680; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:982681; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2682; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill2683; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:972684; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2685; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill2686; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:962687; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2688; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill2689; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:952690; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2691; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill2692; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:942693; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2694; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill2695; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:932696; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2697; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill2698; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:922699; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2700; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill2701; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:912702; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2703; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill2704; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:902705; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2706; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill2707; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:892708; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2709; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill2710; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:882711; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2712; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill2713; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:872714; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2715; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill2716; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:862717; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2718; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill2719; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:852720; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2721; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill2722; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:842723; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2724; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill2725; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:832726; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2727; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill2728; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:822729; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2730; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill2731; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:812732; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2733; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill2734; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:802735; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2736; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill2737; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:792738; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2739; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill2740; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:782741; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2742; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill2743; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:772744; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2745; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill2746; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:762747; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2748; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill2749; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:752750; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2751; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill2752; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:742753; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2754; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill2755; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:732756; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2757; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill2758; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:722759; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2760; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill2761; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:712762; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2763; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill2764; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:702765; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2766; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill2767; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:692768; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2769; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill2770; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:682771; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2772; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill2773; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:672774; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2775; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill2776; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:662777; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2778; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill2779; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:652780; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2781; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill2782; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:642783; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2784; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill2785; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:632786; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2787; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill2788; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:622789; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2790; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill2791; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:612792; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2793; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill2794; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:602795; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2796; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill2797; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:592798; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2799; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill2800; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:582801; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2802; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill2803; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:572804; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2805; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill2806; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:562807; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2808; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill2809; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:552810; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2811; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill2812; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:542813; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2814; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill2815; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:532816; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2817; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill2818; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:522819; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2820; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill2821; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:512822; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2823; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill2824; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:502825; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2826; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill2827; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:492828; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2829; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill2830; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:482831; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2832; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill2833; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:472834; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2835; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill2836; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:462837; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2838; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill2839; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:452840; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2841; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill2842; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:442843; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2844; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill2845; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:432846; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2847; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill2848; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:422849; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2850; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill2851; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:412852; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2853; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill2854; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:402855; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2856; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill2857; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:392858; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2859; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill2860; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:382861; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2862; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill2863; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:372864; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2865; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill2866; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:362867; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2868; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill2869; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:352870; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2871; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill2872; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:342873; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2874; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill2875; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:332876; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2877; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill2878; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:322879; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2880; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill2881; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:312882; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2883; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill2884; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:302885; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2886; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill2887; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:292888; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2889; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill2890; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:282891; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2892; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill2893; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:272894; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2895; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill2896; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:262897; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2898; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill2899; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:252900; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2901; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill2902; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:242903; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2904; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill2905; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:232906; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2907; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill2908; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:222909; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2910; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill2911; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:212912; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2913; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill2914; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:202915; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2916; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill2917; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:192918; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2919; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill2920; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:182921; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2922; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill2923; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:172924; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2925; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill2926; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:162927; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2928; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill2929; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:152930; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2931; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill2932; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:142933; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2934; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill2935; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:132936; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2937; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill2938; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:122939; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2940; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill2941; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:112942; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2943; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill2944; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:102945; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2946; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill2947; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:92948; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2949; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill2950; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:82951; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2952; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill2953; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:72954; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2955; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill2956; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:62957; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2958; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill2959; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:52960; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2961; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill2962; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:42963; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2964; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill2965; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:32966; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2967; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill2968; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22969; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2970; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill2971; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:12972; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2973; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill2974; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen2975; ALIGNED-NEXT:    v_add_nc_u32_e32 v1, 0x100, v12976; ALIGNED-NEXT:    s_waitcnt vmcnt(0)2977; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill2978; ALIGNED-NEXT:    buffer_store_byte v116, v0, s[0:3], 0 offen offset:2552979; ALIGNED-NEXT:    buffer_store_byte v117, v0, s[0:3], 0 offen offset:2542980; ALIGNED-NEXT:    buffer_store_byte v118, v0, s[0:3], 0 offen offset:2532981; ALIGNED-NEXT:    buffer_store_byte v119, v0, s[0:3], 0 offen offset:2522982; ALIGNED-NEXT:    buffer_store_byte v40, v0, s[0:3], 0 offen offset:2512983; ALIGNED-NEXT:    buffer_store_byte v41, v0, s[0:3], 0 offen offset:2502984; ALIGNED-NEXT:    buffer_store_byte v42, v0, s[0:3], 0 offen offset:2492985; ALIGNED-NEXT:    buffer_store_byte v43, v0, s[0:3], 0 offen offset:2482986; ALIGNED-NEXT:    buffer_store_byte v44, v0, s[0:3], 0 offen offset:2472987; ALIGNED-NEXT:    buffer_store_byte v45, v0, s[0:3], 0 offen offset:2462988; ALIGNED-NEXT:    buffer_store_byte v46, v0, s[0:3], 0 offen offset:2452989; ALIGNED-NEXT:    buffer_store_byte v47, v0, s[0:3], 0 offen offset:2442990; ALIGNED-NEXT:    buffer_store_byte v56, v0, s[0:3], 0 offen offset:2432991; ALIGNED-NEXT:    buffer_store_byte v57, v0, s[0:3], 0 offen offset:2422992; ALIGNED-NEXT:    buffer_store_byte v58, v0, s[0:3], 0 offen offset:2412993; ALIGNED-NEXT:    buffer_store_byte v59, v0, s[0:3], 0 offen offset:2402994; ALIGNED-NEXT:    buffer_store_byte v60, v0, s[0:3], 0 offen offset:2392995; ALIGNED-NEXT:    buffer_store_byte v61, v0, s[0:3], 0 offen offset:2382996; ALIGNED-NEXT:    buffer_store_byte v62, v0, s[0:3], 0 offen offset:2372997; ALIGNED-NEXT:    buffer_store_byte v63, v0, s[0:3], 0 offen offset:2362998; ALIGNED-NEXT:    buffer_store_byte v72, v0, s[0:3], 0 offen offset:2352999; ALIGNED-NEXT:    buffer_store_byte v73, v0, s[0:3], 0 offen offset:2343000; ALIGNED-NEXT:    buffer_store_byte v74, v0, s[0:3], 0 offen offset:2333001; ALIGNED-NEXT:    buffer_store_byte v75, v0, s[0:3], 0 offen offset:2323002; ALIGNED-NEXT:    buffer_store_byte v76, v0, s[0:3], 0 offen offset:2313003; ALIGNED-NEXT:    buffer_store_byte v77, v0, s[0:3], 0 offen offset:2303004; ALIGNED-NEXT:    buffer_store_byte v78, v0, s[0:3], 0 offen offset:2293005; ALIGNED-NEXT:    buffer_store_byte v79, v0, s[0:3], 0 offen offset:2283006; ALIGNED-NEXT:    buffer_store_byte v88, v0, s[0:3], 0 offen offset:2273007; ALIGNED-NEXT:    buffer_store_byte v89, v0, s[0:3], 0 offen offset:2263008; ALIGNED-NEXT:    buffer_store_byte v90, v0, s[0:3], 0 offen offset:2253009; ALIGNED-NEXT:    buffer_store_byte v91, v0, s[0:3], 0 offen offset:2243010; ALIGNED-NEXT:    buffer_store_byte v92, v0, s[0:3], 0 offen offset:2233011; ALIGNED-NEXT:    buffer_store_byte v93, v0, s[0:3], 0 offen offset:2223012; ALIGNED-NEXT:    buffer_store_byte v94, v0, s[0:3], 0 offen offset:2213013; ALIGNED-NEXT:    buffer_store_byte v95, v0, s[0:3], 0 offen offset:2203014; ALIGNED-NEXT:    buffer_store_byte v104, v0, s[0:3], 0 offen offset:2193015; ALIGNED-NEXT:    buffer_store_byte v105, v0, s[0:3], 0 offen offset:2183016; ALIGNED-NEXT:    buffer_store_byte v106, v0, s[0:3], 0 offen offset:2173017; ALIGNED-NEXT:    buffer_store_byte v107, v0, s[0:3], 0 offen offset:2163018; ALIGNED-NEXT:    buffer_store_byte v108, v0, s[0:3], 0 offen offset:2153019; ALIGNED-NEXT:    buffer_store_byte v109, v0, s[0:3], 0 offen offset:2143020; ALIGNED-NEXT:    buffer_store_byte v110, v0, s[0:3], 0 offen offset:2133021; ALIGNED-NEXT:    buffer_store_byte v111, v0, s[0:3], 0 offen offset:2123022; ALIGNED-NEXT:    buffer_store_byte v120, v0, s[0:3], 0 offen offset:2113023; ALIGNED-NEXT:    buffer_store_byte v121, v0, s[0:3], 0 offen offset:2103024; ALIGNED-NEXT:    buffer_store_byte v122, v0, s[0:3], 0 offen offset:2093025; ALIGNED-NEXT:    buffer_store_byte v123, v0, s[0:3], 0 offen offset:2083026; ALIGNED-NEXT:    buffer_store_byte v124, v0, s[0:3], 0 offen offset:2073027; ALIGNED-NEXT:    buffer_store_byte v125, v0, s[0:3], 0 offen offset:2063028; ALIGNED-NEXT:    buffer_store_byte v126, v0, s[0:3], 0 offen offset:2053029; ALIGNED-NEXT:    buffer_store_byte v127, v0, s[0:3], 0 offen offset:2043030; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload3031; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3032; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2033033; ALIGNED-NEXT:    buffer_store_byte v3, v0, s[0:3], 0 offen offset:2023034; ALIGNED-NEXT:    buffer_store_byte v4, v0, s[0:3], 0 offen offset:2013035; ALIGNED-NEXT:    buffer_store_byte v5, v0, s[0:3], 0 offen offset:2003036; ALIGNED-NEXT:    buffer_store_byte v6, v0, s[0:3], 0 offen offset:1993037; ALIGNED-NEXT:    buffer_store_byte v7, v0, s[0:3], 0 offen offset:1983038; ALIGNED-NEXT:    buffer_store_byte v8, v0, s[0:3], 0 offen offset:1973039; ALIGNED-NEXT:    buffer_store_byte v9, v0, s[0:3], 0 offen offset:1963040; ALIGNED-NEXT:    buffer_store_byte v10, v0, s[0:3], 0 offen offset:1953041; ALIGNED-NEXT:    buffer_store_byte v11, v0, s[0:3], 0 offen offset:1943042; ALIGNED-NEXT:    buffer_store_byte v12, v0, s[0:3], 0 offen offset:1933043; ALIGNED-NEXT:    buffer_store_byte v13, v0, s[0:3], 0 offen offset:1923044; ALIGNED-NEXT:    buffer_store_byte v14, v0, s[0:3], 0 offen offset:1913045; ALIGNED-NEXT:    buffer_store_byte v15, v0, s[0:3], 0 offen offset:1903046; ALIGNED-NEXT:    buffer_store_byte v16, v0, s[0:3], 0 offen offset:1893047; ALIGNED-NEXT:    buffer_store_byte v17, v0, s[0:3], 0 offen offset:1883048; ALIGNED-NEXT:    buffer_store_byte v18, v0, s[0:3], 0 offen offset:1873049; ALIGNED-NEXT:    buffer_store_byte v19, v0, s[0:3], 0 offen offset:1863050; ALIGNED-NEXT:    buffer_store_byte v20, v0, s[0:3], 0 offen offset:1853051; ALIGNED-NEXT:    buffer_store_byte v21, v0, s[0:3], 0 offen offset:1843052; ALIGNED-NEXT:    buffer_store_byte v22, v0, s[0:3], 0 offen offset:1833053; ALIGNED-NEXT:    buffer_store_byte v23, v0, s[0:3], 0 offen offset:1823054; ALIGNED-NEXT:    buffer_store_byte v24, v0, s[0:3], 0 offen offset:1813055; ALIGNED-NEXT:    buffer_store_byte v25, v0, s[0:3], 0 offen offset:1803056; ALIGNED-NEXT:    buffer_store_byte v26, v0, s[0:3], 0 offen offset:1793057; ALIGNED-NEXT:    buffer_store_byte v27, v0, s[0:3], 0 offen offset:1783058; ALIGNED-NEXT:    buffer_store_byte v28, v0, s[0:3], 0 offen offset:1773059; ALIGNED-NEXT:    buffer_store_byte v29, v0, s[0:3], 0 offen offset:1763060; ALIGNED-NEXT:    buffer_store_byte v30, v0, s[0:3], 0 offen offset:1753061; ALIGNED-NEXT:    buffer_store_byte v31, v0, s[0:3], 0 offen offset:1743062; ALIGNED-NEXT:    buffer_store_byte v32, v0, s[0:3], 0 offen offset:1733063; ALIGNED-NEXT:    buffer_store_byte v33, v0, s[0:3], 0 offen offset:1723064; ALIGNED-NEXT:    buffer_store_byte v34, v0, s[0:3], 0 offen offset:1713065; ALIGNED-NEXT:    buffer_store_byte v35, v0, s[0:3], 0 offen offset:1703066; ALIGNED-NEXT:    buffer_store_byte v36, v0, s[0:3], 0 offen offset:1693067; ALIGNED-NEXT:    buffer_store_byte v37, v0, s[0:3], 0 offen offset:1683068; ALIGNED-NEXT:    buffer_store_byte v38, v0, s[0:3], 0 offen offset:1673069; ALIGNED-NEXT:    buffer_store_byte v39, v0, s[0:3], 0 offen offset:1663070; ALIGNED-NEXT:    buffer_store_byte v48, v0, s[0:3], 0 offen offset:1653071; ALIGNED-NEXT:    buffer_store_byte v49, v0, s[0:3], 0 offen offset:1643072; ALIGNED-NEXT:    buffer_store_byte v50, v0, s[0:3], 0 offen offset:1633073; ALIGNED-NEXT:    buffer_store_byte v51, v0, s[0:3], 0 offen offset:1623074; ALIGNED-NEXT:    buffer_store_byte v52, v0, s[0:3], 0 offen offset:1613075; ALIGNED-NEXT:    buffer_store_byte v53, v0, s[0:3], 0 offen offset:1603076; ALIGNED-NEXT:    buffer_store_byte v54, v0, s[0:3], 0 offen offset:1593077; ALIGNED-NEXT:    buffer_store_byte v55, v0, s[0:3], 0 offen offset:1583078; ALIGNED-NEXT:    buffer_store_byte v64, v0, s[0:3], 0 offen offset:1573079; ALIGNED-NEXT:    buffer_store_byte v65, v0, s[0:3], 0 offen offset:1563080; ALIGNED-NEXT:    buffer_store_byte v66, v0, s[0:3], 0 offen offset:1553081; ALIGNED-NEXT:    buffer_store_byte v67, v0, s[0:3], 0 offen offset:1543082; ALIGNED-NEXT:    buffer_store_byte v68, v0, s[0:3], 0 offen offset:1533083; ALIGNED-NEXT:    buffer_store_byte v69, v0, s[0:3], 0 offen offset:1523084; ALIGNED-NEXT:    buffer_store_byte v70, v0, s[0:3], 0 offen offset:1513085; ALIGNED-NEXT:    buffer_store_byte v71, v0, s[0:3], 0 offen offset:1503086; ALIGNED-NEXT:    buffer_store_byte v80, v0, s[0:3], 0 offen offset:1493087; ALIGNED-NEXT:    buffer_store_byte v81, v0, s[0:3], 0 offen offset:1483088; ALIGNED-NEXT:    buffer_store_byte v82, v0, s[0:3], 0 offen offset:1473089; ALIGNED-NEXT:    buffer_store_byte v83, v0, s[0:3], 0 offen offset:1463090; ALIGNED-NEXT:    buffer_store_byte v84, v0, s[0:3], 0 offen offset:1453091; ALIGNED-NEXT:    buffer_store_byte v85, v0, s[0:3], 0 offen offset:1443092; ALIGNED-NEXT:    buffer_store_byte v86, v0, s[0:3], 0 offen offset:1433093; ALIGNED-NEXT:    buffer_store_byte v87, v0, s[0:3], 0 offen offset:1423094; ALIGNED-NEXT:    buffer_store_byte v96, v0, s[0:3], 0 offen offset:1413095; ALIGNED-NEXT:    buffer_store_byte v97, v0, s[0:3], 0 offen offset:1403096; ALIGNED-NEXT:    buffer_store_byte v98, v0, s[0:3], 0 offen offset:1393097; ALIGNED-NEXT:    buffer_store_byte v99, v0, s[0:3], 0 offen offset:1383098; ALIGNED-NEXT:    buffer_store_byte v100, v0, s[0:3], 0 offen offset:1373099; ALIGNED-NEXT:    buffer_store_byte v101, v0, s[0:3], 0 offen offset:1363100; ALIGNED-NEXT:    buffer_store_byte v102, v0, s[0:3], 0 offen offset:1353101; ALIGNED-NEXT:    buffer_store_byte v103, v0, s[0:3], 0 offen offset:1343102; ALIGNED-NEXT:    buffer_store_byte v112, v0, s[0:3], 0 offen offset:1333103; ALIGNED-NEXT:    buffer_store_byte v113, v0, s[0:3], 0 offen offset:1323104; ALIGNED-NEXT:    buffer_store_byte v114, v0, s[0:3], 0 offen offset:1313105; ALIGNED-NEXT:    buffer_store_byte v115, v0, s[0:3], 0 offen offset:1303106; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload3107; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3108; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1293109; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload3110; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3111; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1283112; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload3113; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3114; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1273115; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload3116; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3117; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1263118; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload3119; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3120; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1253121; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload3122; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3123; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1243124; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload3125; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3126; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1233127; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload3128; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3129; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1223130; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload3131; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3132; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1213133; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload3134; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3135; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1203136; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload3137; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3138; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1193139; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload3140; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3141; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1183142; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload3143; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3144; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1173145; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload3146; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3147; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1163148; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload3149; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3150; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1153151; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload3152; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3153; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1143154; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload3155; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3156; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1133157; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload3158; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3159; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1123160; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload3161; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3162; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1113163; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload3164; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3165; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1103166; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload3167; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3168; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1093169; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload3170; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3171; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1083172; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload3173; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3174; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1073175; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload3176; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3177; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1063178; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload3179; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3180; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1053181; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload3182; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3183; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1043184; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload3185; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3186; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1033187; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload3188; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3189; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1023190; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload3191; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3192; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1013193; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload3194; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3195; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1003196; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload3197; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3198; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:993199; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload3200; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3201; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:983202; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload3203; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3204; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:973205; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload3206; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3207; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:963208; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload3209; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3210; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:953211; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload3212; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3213; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:943214; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload3215; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3216; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:933217; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload3218; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3219; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:923220; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload3221; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3222; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:913223; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload3224; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3225; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:903226; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload3227; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3228; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:893229; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload3230; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3231; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:883232; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload3233; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3234; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:873235; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload3236; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3237; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:863238; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload3239; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3240; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:853241; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload3242; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3243; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:843244; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload3245; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3246; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:833247; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload3248; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3249; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:823250; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload3251; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3252; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:813253; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload3254; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3255; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:803256; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload3257; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3258; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:793259; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload3260; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3261; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:783262; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload3263; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3264; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:773265; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload3266; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3267; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:763268; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload3269; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3270; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:753271; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload3272; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3273; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:743274; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload3275; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3276; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:733277; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload3278; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3279; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:723280; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload3281; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3282; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:713283; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload3284; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3285; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:703286; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload3287; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3288; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:693289; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload3290; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3291; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:683292; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload3293; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3294; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:673295; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload3296; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3297; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:663298; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload3299; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3300; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:653301; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload3302; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3303; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:643304; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload3305; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3306; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:633307; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload3308; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3309; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:623310; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload3311; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3312; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:613313; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload3314; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3315; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:603316; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload3317; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3318; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:593319; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload3320; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3321; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:583322; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload3323; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3324; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:573325; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload3326; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3327; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:563328; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload3329; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3330; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:553331; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload3332; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3333; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:543334; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload3335; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3336; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:533337; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload3338; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3339; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:523340; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload3341; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3342; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:513343; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload3344; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3345; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:503346; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload3347; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3348; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:493349; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload3350; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3351; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:483352; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload3353; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3354; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:473355; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload3356; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3357; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:463358; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload3359; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3360; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:453361; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload3362; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3363; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:443364; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload3365; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3366; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:433367; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload3368; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3369; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:423370; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload3371; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3372; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:413373; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload3374; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3375; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:403376; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload3377; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3378; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:393379; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload3380; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3381; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:383382; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload3383; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3384; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:373385; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload3386; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3387; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:363388; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload3389; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3390; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:353391; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload3392; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3393; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:343394; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload3395; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3396; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:333397; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload3398; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3399; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:323400; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload3401; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3402; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:313403; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload3404; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3405; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:303406; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload3407; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3408; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:293409; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload3410; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3411; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:283412; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload3413; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3414; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:273415; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload3416; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3417; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:263418; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload3419; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3420; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:253421; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload3422; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3423; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:243424; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload3425; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3426; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:233427; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload3428; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3429; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:223430; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload3431; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3432; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:213433; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload3434; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3435; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:203436; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload3437; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3438; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:193439; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload3440; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3441; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:183442; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload3443; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3444; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:173445; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload3446; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3447; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:163448; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload3449; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3450; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:153451; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload3452; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3453; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:143454; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload3455; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3456; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:133457; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload3458; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3459; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:123460; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload3461; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3462; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:113463; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload3464; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3465; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:103466; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload3467; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3468; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:93469; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload3470; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3471; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:83472; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload3473; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3474; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:73475; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload3476; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3477; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:63478; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload3479; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3480; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:53481; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload3482; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3483; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:43484; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload3485; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3486; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:33487; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload3488; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3489; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23490; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload3491; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3492; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:13493; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload3494; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3495; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen3496; ALIGNED-NEXT:    v_add_nc_u32_e32 v0, 0x100, v03497; ALIGNED-NEXT:    s_cbranch_vccnz .LBB3_13498; ALIGNED-NEXT:  ; %bb.2: ; %memcpy-split3499; ALIGNED-NEXT:    s_clause 0x2f ; 192-byte Folded Reload3500; ALIGNED-NEXT:    buffer_load_dword v127, off, s[0:3], s323501; ALIGNED-NEXT:    buffer_load_dword v126, off, s[0:3], s32 offset:43502; ALIGNED-NEXT:    buffer_load_dword v125, off, s[0:3], s32 offset:83503; ALIGNED-NEXT:    buffer_load_dword v124, off, s[0:3], s32 offset:123504; ALIGNED-NEXT:    buffer_load_dword v123, off, s[0:3], s32 offset:163505; ALIGNED-NEXT:    buffer_load_dword v122, off, s[0:3], s32 offset:203506; ALIGNED-NEXT:    buffer_load_dword v121, off, s[0:3], s32 offset:243507; ALIGNED-NEXT:    buffer_load_dword v120, off, s[0:3], s32 offset:283508; ALIGNED-NEXT:    buffer_load_dword v111, off, s[0:3], s32 offset:323509; ALIGNED-NEXT:    buffer_load_dword v110, off, s[0:3], s32 offset:363510; ALIGNED-NEXT:    buffer_load_dword v109, off, s[0:3], s32 offset:403511; ALIGNED-NEXT:    buffer_load_dword v108, off, s[0:3], s32 offset:443512; ALIGNED-NEXT:    buffer_load_dword v107, off, s[0:3], s32 offset:483513; ALIGNED-NEXT:    buffer_load_dword v106, off, s[0:3], s32 offset:523514; ALIGNED-NEXT:    buffer_load_dword v105, off, s[0:3], s32 offset:563515; ALIGNED-NEXT:    buffer_load_dword v104, off, s[0:3], s32 offset:603516; ALIGNED-NEXT:    buffer_load_dword v95, off, s[0:3], s32 offset:643517; ALIGNED-NEXT:    buffer_load_dword v94, off, s[0:3], s32 offset:683518; ALIGNED-NEXT:    buffer_load_dword v93, off, s[0:3], s32 offset:723519; ALIGNED-NEXT:    buffer_load_dword v92, off, s[0:3], s32 offset:763520; ALIGNED-NEXT:    buffer_load_dword v91, off, s[0:3], s32 offset:803521; ALIGNED-NEXT:    buffer_load_dword v90, off, s[0:3], s32 offset:843522; ALIGNED-NEXT:    buffer_load_dword v89, off, s[0:3], s32 offset:883523; ALIGNED-NEXT:    buffer_load_dword v88, off, s[0:3], s32 offset:923524; ALIGNED-NEXT:    buffer_load_dword v79, off, s[0:3], s32 offset:963525; ALIGNED-NEXT:    buffer_load_dword v78, off, s[0:3], s32 offset:1003526; ALIGNED-NEXT:    buffer_load_dword v77, off, s[0:3], s32 offset:1043527; ALIGNED-NEXT:    buffer_load_dword v76, off, s[0:3], s32 offset:1083528; ALIGNED-NEXT:    buffer_load_dword v75, off, s[0:3], s32 offset:1123529; ALIGNED-NEXT:    buffer_load_dword v74, off, s[0:3], s32 offset:1163530; ALIGNED-NEXT:    buffer_load_dword v73, off, s[0:3], s32 offset:1203531; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s32 offset:1243532; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:1283533; ALIGNED-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:1323534; ALIGNED-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:1363535; ALIGNED-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:1403536; ALIGNED-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:1443537; ALIGNED-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:1483538; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:1523539; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:1563540; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:1603541; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:1643542; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:1683543; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:1723544; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:1763545; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:1803546; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:1843547; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:1883548; ALIGNED-NEXT:    s_waitcnt vmcnt(0)3549; ALIGNED-NEXT:    s_setpc_b64 s[30:31]3550;3551; UNROLL3-LABEL: memcpy_p5_p5_sz2048:3552; UNROLL3:       ; %bb.0: ; %entry3553; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)3554; UNROLL3-NEXT:    v_mov_b32_e32 v2, v13555; UNROLL3-NEXT:    v_mov_b32_e32 v3, v03556; UNROLL3-NEXT:    s_mov_b64 s[4:5], 03557; UNROLL3-NEXT:  .LBB3_1: ; %load-store-loop3558; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=13559; UNROLL3-NEXT:    s_clause 0xb3560; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:443561; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:403562; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:363563; UNROLL3-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:323564; UNROLL3-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:283565; UNROLL3-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:243566; UNROLL3-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:203567; UNROLL3-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:163568; UNROLL3-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:123569; UNROLL3-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:83570; UNROLL3-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:43571; UNROLL3-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen3572; UNROLL3-NEXT:    s_add_u32 s4, s4, 483573; UNROLL3-NEXT:    s_addc_u32 s5, s5, 03574; UNROLL3-NEXT:    v_add_nc_u32_e32 v2, 48, v23575; UNROLL3-NEXT:    v_cmp_gt_u64_e64 s6, 0x7e0, s[4:5]3576; UNROLL3-NEXT:    s_waitcnt vmcnt(11)3577; UNROLL3-NEXT:    buffer_store_dword v4, v3, s[0:3], 0 offen offset:443578; UNROLL3-NEXT:    s_waitcnt vmcnt(10)3579; UNROLL3-NEXT:    buffer_store_dword v5, v3, s[0:3], 0 offen offset:403580; UNROLL3-NEXT:    s_waitcnt vmcnt(9)3581; UNROLL3-NEXT:    buffer_store_dword v6, v3, s[0:3], 0 offen offset:363582; UNROLL3-NEXT:    s_waitcnt vmcnt(8)3583; UNROLL3-NEXT:    buffer_store_dword v7, v3, s[0:3], 0 offen offset:323584; UNROLL3-NEXT:    s_waitcnt vmcnt(7)3585; UNROLL3-NEXT:    buffer_store_dword v8, v3, s[0:3], 0 offen offset:283586; UNROLL3-NEXT:    s_waitcnt vmcnt(6)3587; UNROLL3-NEXT:    buffer_store_dword v9, v3, s[0:3], 0 offen offset:243588; UNROLL3-NEXT:    s_waitcnt vmcnt(5)3589; UNROLL3-NEXT:    buffer_store_dword v10, v3, s[0:3], 0 offen offset:203590; UNROLL3-NEXT:    s_waitcnt vmcnt(4)3591; UNROLL3-NEXT:    buffer_store_dword v11, v3, s[0:3], 0 offen offset:163592; UNROLL3-NEXT:    s_waitcnt vmcnt(3)3593; UNROLL3-NEXT:    buffer_store_dword v12, v3, s[0:3], 0 offen offset:123594; UNROLL3-NEXT:    s_waitcnt vmcnt(2)3595; UNROLL3-NEXT:    buffer_store_dword v13, v3, s[0:3], 0 offen offset:83596; UNROLL3-NEXT:    s_waitcnt vmcnt(1)3597; UNROLL3-NEXT:    buffer_store_dword v14, v3, s[0:3], 0 offen offset:43598; UNROLL3-NEXT:    s_waitcnt vmcnt(0)3599; UNROLL3-NEXT:    buffer_store_dword v15, v3, s[0:3], 0 offen3600; UNROLL3-NEXT:    v_add_nc_u32_e32 v3, 48, v33601; UNROLL3-NEXT:    s_and_b32 vcc_lo, exec_lo, s63602; UNROLL3-NEXT:    s_cbranch_vccnz .LBB3_13603; UNROLL3-NEXT:  ; %bb.2: ; %memcpy-split3604; UNROLL3-NEXT:    s_clause 0x33605; UNROLL3-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:20283606; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:20243607; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:20203608; UNROLL3-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:20163609; UNROLL3-NEXT:    s_waitcnt vmcnt(3)3610; UNROLL3-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:20283611; UNROLL3-NEXT:    s_waitcnt vmcnt(2)3612; UNROLL3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:20243613; UNROLL3-NEXT:    s_waitcnt vmcnt(1)3614; UNROLL3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:20203615; UNROLL3-NEXT:    s_waitcnt vmcnt(0)3616; UNROLL3-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:20163617; UNROLL3-NEXT:    s_clause 0x33618; UNROLL3-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:20443619; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:20403620; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:20363621; UNROLL3-NEXT:    buffer_load_dword v1, v1, s[0:3], 0 offen offset:20323622; UNROLL3-NEXT:    s_waitcnt vmcnt(3)3623; UNROLL3-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:20443624; UNROLL3-NEXT:    s_waitcnt vmcnt(2)3625; UNROLL3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:20403626; UNROLL3-NEXT:    s_waitcnt vmcnt(1)3627; UNROLL3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:20363628; UNROLL3-NEXT:    s_waitcnt vmcnt(0)3629; UNROLL3-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:20323630; UNROLL3-NEXT:    s_setpc_b64 s[30:31]3631entry:3632  tail call void @llvm.memcpy.p5.p5.i64(ptr addrspace(5) noundef nonnull align 1 %dst, ptr addrspace(5) noundef nonnull align 1 %src, i64 2048, i1 false)3633  ret void3634}3635 3636define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5) align 1 readonly %src) {3637; CHECK-LABEL: memcpy_p0_p5_sz2048:3638; CHECK:       ; %bb.0: ; %entry3639; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)3640; CHECK-NEXT:    s_mov_b64 s[4:5], 03641; CHECK-NEXT:  .LBB4_1: ; %load-store-loop3642; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=13643; CHECK-NEXT:    s_clause 0x3e3644; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:323645; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:363646; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:403647; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:443648; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:483649; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:523650; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:563651; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:603652; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:763653; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:923654; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:883655; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:843656; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:803657; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:723658; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:683659; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:643660; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:2363661; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:2523662; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:2483663; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:2443664; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:2403665; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:2323666; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:2283667; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:2243668; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:2203669; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:2163670; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:2123671; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:2083672; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:2043673; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:2003674; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:1963675; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:1923676; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:1083677; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:1243678; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:1203679; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:1163680; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:1123681; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:1043682; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:1003683; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:963684; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:1883685; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:1843686; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:1803687; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:1763688; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:1723689; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:1683690; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:1643691; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:1603692; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:1563693; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:1523694; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:1483695; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:1443696; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:1403697; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:1363698; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:1323699; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:1283700; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen3701; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:43702; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:83703; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:163704; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:203705; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:243706; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:283707; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:123708; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s43709; CHECK-NEXT:    s_add_u32 s4, s4, 0x1003710; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo3711; CHECK-NEXT:    s_addc_u32 s5, s5, 03712; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 483713; CHECK-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]3714; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0x100, v23715; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo3716; CHECK-NEXT:    s_waitcnt vmcnt(43)3717; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[23:26] offset:1923718; CHECK-NEXT:    s_waitcnt vmcnt(40)3719; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[19:22] offset:1763720; CHECK-NEXT:    s_waitcnt vmcnt(36)3721; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[27:30] offset:1603722; CHECK-NEXT:    s_waitcnt vmcnt(32)3723; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[31:34] offset:1443724; CHECK-NEXT:    s_waitcnt vmcnt(20)3725; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:1283726; CHECK-NEXT:    s_waitcnt vmcnt(16)3727; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:1123728; CHECK-NEXT:    s_waitcnt vmcnt(12)3729; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71] offset:963730; CHECK-NEXT:    s_waitcnt vmcnt(8)3731; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:1283732; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:643733; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[35:38] offset:483734; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[15:18] offset:323735; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:643736; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[7:10]3737; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:323738; CHECK-NEXT:    s_waitcnt vmcnt(1)3739; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99] offset:163740; CHECK-NEXT:    s_waitcnt vmcnt(0)3741; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87]3742; CHECK-NEXT:    s_and_b32 vcc_lo, exec_lo, s63743; CHECK-NEXT:    s_cbranch_vccnz .LBB4_13744; CHECK-NEXT:  ; %bb.2: ; %memcpy-split3745; CHECK-NEXT:    s_waitcnt lgkmcnt(0)3746; CHECK-NEXT:    s_setpc_b64 s[30:31]3747;3748; ALIGNED-LABEL: memcpy_p0_p5_sz2048:3749; ALIGNED:       ; %bb.0: ; %entry3750; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)3751; ALIGNED-NEXT:    s_mov_b64 s[4:5], 03752; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill3753; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill3754; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill3755; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill3756; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill3757; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill3758; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill3759; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill3760; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill3761; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill3762; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill3763; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill3764; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill3765; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill3766; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill3767; ALIGNED-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill3768; ALIGNED-NEXT:    buffer_store_dword v72, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill3769; ALIGNED-NEXT:    buffer_store_dword v73, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill3770; ALIGNED-NEXT:    buffer_store_dword v74, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill3771; ALIGNED-NEXT:    buffer_store_dword v75, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill3772; ALIGNED-NEXT:    buffer_store_dword v76, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill3773; ALIGNED-NEXT:    buffer_store_dword v77, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill3774; ALIGNED-NEXT:    buffer_store_dword v78, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill3775; ALIGNED-NEXT:    buffer_store_dword v79, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill3776; ALIGNED-NEXT:    buffer_store_dword v88, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill3777; ALIGNED-NEXT:    buffer_store_dword v89, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill3778; ALIGNED-NEXT:    buffer_store_dword v90, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill3779; ALIGNED-NEXT:    buffer_store_dword v91, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill3780; ALIGNED-NEXT:    buffer_store_dword v92, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill3781; ALIGNED-NEXT:    buffer_store_dword v93, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill3782; ALIGNED-NEXT:    buffer_store_dword v94, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill3783; ALIGNED-NEXT:    buffer_store_dword v95, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill3784; ALIGNED-NEXT:    buffer_store_dword v104, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill3785; ALIGNED-NEXT:    buffer_store_dword v105, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill3786; ALIGNED-NEXT:    buffer_store_dword v106, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill3787; ALIGNED-NEXT:    buffer_store_dword v107, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill3788; ALIGNED-NEXT:    buffer_store_dword v108, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill3789; ALIGNED-NEXT:    buffer_store_dword v109, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill3790; ALIGNED-NEXT:    buffer_store_dword v110, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill3791; ALIGNED-NEXT:    buffer_store_dword v111, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill3792; ALIGNED-NEXT:    buffer_store_dword v120, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill3793; ALIGNED-NEXT:    buffer_store_dword v121, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill3794; ALIGNED-NEXT:    buffer_store_dword v122, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill3795; ALIGNED-NEXT:    buffer_store_dword v123, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill3796; ALIGNED-NEXT:    buffer_store_dword v124, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill3797; ALIGNED-NEXT:    buffer_store_dword v125, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill3798; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill3799; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 ; 4-byte Folded Spill3800; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1224 ; 4-byte Folded Spill3801; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Spill3802; ALIGNED-NEXT:  .LBB4_1: ; %load-store-loop3803; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=13804; ALIGNED-NEXT:    s_clause 0x3e3805; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:203806; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:213807; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:223808; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:233809; ALIGNED-NEXT:    buffer_load_ubyte v126, v2, s[0:3], 0 offen offset:193810; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:283811; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:293812; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:303813; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:313814; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:243815; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:253816; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:263817; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:273818; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:323819; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:333820; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:343821; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:353822; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:363823; ALIGNED-NEXT:    buffer_load_ubyte v16, v2, s[0:3], 0 offen offset:373824; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:383825; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:393826; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:403827; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:413828; ALIGNED-NEXT:    buffer_load_ubyte v27, v2, s[0:3], 0 offen offset:423829; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:433830; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:443831; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:453832; ALIGNED-NEXT:    buffer_load_ubyte v24, v2, s[0:3], 0 offen offset:463833; ALIGNED-NEXT:    buffer_load_ubyte v25, v2, s[0:3], 0 offen offset:473834; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:483835; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:493836; ALIGNED-NEXT:    buffer_load_ubyte v33, v2, s[0:3], 0 offen offset:503837; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:513838; ALIGNED-NEXT:    buffer_load_ubyte v32, v2, s[0:3], 0 offen offset:523839; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:533840; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:543841; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:553842; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:603843; ALIGNED-NEXT:    buffer_load_ubyte v50, v2, s[0:3], 0 offen offset:613844; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:623845; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:633846; ALIGNED-NEXT:    buffer_load_ubyte v38, v2, s[0:3], 0 offen offset:563847; ALIGNED-NEXT:    buffer_load_ubyte v51, v2, s[0:3], 0 offen offset:573848; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:583849; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:593850; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:643851; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:653852; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:663853; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:673854; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:683855; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:693856; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:703857; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:713858; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:763859; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:773860; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:783861; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:793862; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:753863; ALIGNED-NEXT:    buffer_load_ubyte v125, v2, s[0:3], 0 offen offset:1593864; ALIGNED-NEXT:    buffer_load_ubyte v124, v2, s[0:3], 0 offen offset:1553865; ALIGNED-NEXT:    buffer_load_ubyte v123, v2, s[0:3], 0 offen offset:1523866; ALIGNED-NEXT:    buffer_load_ubyte v120, v2, s[0:3], 0 offen offset:1533867; ALIGNED-NEXT:    buffer_load_ubyte v110, v2, s[0:3], 0 offen offset:1543868; ALIGNED-NEXT:    s_clause 0x353869; ALIGNED-NEXT:    buffer_load_ubyte v108, v2, s[0:3], 0 offen offset:1603870; ALIGNED-NEXT:    buffer_load_ubyte v104, v2, s[0:3], 0 offen offset:1613871; ALIGNED-NEXT:    buffer_load_ubyte v93, v2, s[0:3], 0 offen offset:1623872; ALIGNED-NEXT:    buffer_load_ubyte v92, v2, s[0:3], 0 offen offset:1633873; ALIGNED-NEXT:    buffer_load_ubyte v105, v2, s[0:3], 0 offen offset:1643874; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:1653875; ALIGNED-NEXT:    buffer_load_ubyte v94, v2, s[0:3], 0 offen offset:1663876; ALIGNED-NEXT:    buffer_load_ubyte v91, v2, s[0:3], 0 offen offset:1673877; ALIGNED-NEXT:    buffer_load_ubyte v78, v2, s[0:3], 0 offen offset:1723878; ALIGNED-NEXT:    buffer_load_ubyte v77, v2, s[0:3], 0 offen offset:1733879; ALIGNED-NEXT:    buffer_load_ubyte v76, v2, s[0:3], 0 offen offset:1743880; ALIGNED-NEXT:    buffer_load_ubyte v75, v2, s[0:3], 0 offen offset:1753881; ALIGNED-NEXT:    buffer_load_ubyte v74, v2, s[0:3], 0 offen offset:1713882; ALIGNED-NEXT:    buffer_load_ubyte v73, v2, s[0:3], 0 offen offset:1683883; ALIGNED-NEXT:    buffer_load_ubyte v63, v2, s[0:3], 0 offen offset:1693884; ALIGNED-NEXT:    buffer_load_ubyte v62, v2, s[0:3], 0 offen offset:1703885; ALIGNED-NEXT:    buffer_load_ubyte v61, v2, s[0:3], 0 offen offset:1763886; ALIGNED-NEXT:    buffer_load_ubyte v59, v2, s[0:3], 0 offen offset:1773887; ALIGNED-NEXT:    buffer_load_ubyte v47, v2, s[0:3], 0 offen offset:1783888; ALIGNED-NEXT:    buffer_load_ubyte v56, v2, s[0:3], 0 offen offset:1793889; ALIGNED-NEXT:    buffer_load_ubyte v60, v2, s[0:3], 0 offen offset:1803890; ALIGNED-NEXT:    buffer_load_ubyte v57, v2, s[0:3], 0 offen offset:1813891; ALIGNED-NEXT:    buffer_load_ubyte v58, v2, s[0:3], 0 offen offset:1823892; ALIGNED-NEXT:    buffer_load_ubyte v46, v2, s[0:3], 0 offen offset:1833893; ALIGNED-NEXT:    buffer_load_ubyte v44, v2, s[0:3], 0 offen offset:1883894; ALIGNED-NEXT:    buffer_load_ubyte v43, v2, s[0:3], 0 offen offset:1893895; ALIGNED-NEXT:    buffer_load_ubyte v42, v2, s[0:3], 0 offen offset:1903896; ALIGNED-NEXT:    buffer_load_ubyte v41, v2, s[0:3], 0 offen offset:1913897; ALIGNED-NEXT:    buffer_load_ubyte v40, v2, s[0:3], 0 offen offset:1873898; ALIGNED-NEXT:    buffer_load_ubyte v119, v2, s[0:3], 0 offen offset:1843899; ALIGNED-NEXT:    buffer_load_ubyte v118, v2, s[0:3], 0 offen offset:1853900; ALIGNED-NEXT:    buffer_load_ubyte v116, v2, s[0:3], 0 offen offset:1863901; ALIGNED-NEXT:    buffer_load_ubyte v115, v2, s[0:3], 0 offen offset:1923902; ALIGNED-NEXT:    buffer_load_ubyte v112, v2, s[0:3], 0 offen offset:1933903; ALIGNED-NEXT:    buffer_load_ubyte v101, v2, s[0:3], 0 offen offset:1943904; ALIGNED-NEXT:    buffer_load_ubyte v100, v2, s[0:3], 0 offen offset:1953905; ALIGNED-NEXT:    buffer_load_ubyte v113, v2, s[0:3], 0 offen offset:1963906; ALIGNED-NEXT:    buffer_load_ubyte v103, v2, s[0:3], 0 offen offset:1973907; ALIGNED-NEXT:    buffer_load_ubyte v102, v2, s[0:3], 0 offen offset:1983908; ALIGNED-NEXT:    buffer_load_ubyte v99, v2, s[0:3], 0 offen offset:1993909; ALIGNED-NEXT:    buffer_load_ubyte v97, v2, s[0:3], 0 offen offset:2043910; ALIGNED-NEXT:    buffer_load_ubyte v87, v2, s[0:3], 0 offen offset:2053911; ALIGNED-NEXT:    buffer_load_ubyte v96, v2, s[0:3], 0 offen offset:2063912; ALIGNED-NEXT:    buffer_load_ubyte v86, v2, s[0:3], 0 offen offset:2073913; ALIGNED-NEXT:    buffer_load_ubyte v85, v2, s[0:3], 0 offen offset:2033914; ALIGNED-NEXT:    buffer_load_ubyte v84, v2, s[0:3], 0 offen offset:2003915; ALIGNED-NEXT:    buffer_load_ubyte v83, v2, s[0:3], 0 offen offset:2013916; ALIGNED-NEXT:    buffer_load_ubyte v82, v2, s[0:3], 0 offen offset:2023917; ALIGNED-NEXT:    buffer_load_ubyte v121, v2, s[0:3], 0 offen offset:23918; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:33919; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:43920; ALIGNED-NEXT:    buffer_load_ubyte v109, v2, s[0:3], 0 offen offset:53921; ALIGNED-NEXT:    buffer_load_ubyte v111, v2, s[0:3], 0 offen offset:63922; ALIGNED-NEXT:    buffer_load_ubyte v122, v2, s[0:3], 0 offen offset:73923; ALIGNED-NEXT:    s_waitcnt vmcnt(62)3924; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill3925; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill3926; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill3927; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill3928; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v03929; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill3930; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill3931; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill3932; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill3933; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill3934; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill3935; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill3936; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill3937; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill3938; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v33939; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v53940; ALIGNED-NEXT:    v_lshl_or_b32 v4, v9, 8, v83941; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill3942; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 8, v63943; ALIGNED-NEXT:    v_lshl_or_b32 v6, v11, 8, v123944; ALIGNED-NEXT:    v_lshl_or_b32 v7, v15, 8, v143945; ALIGNED-NEXT:    v_lshl_or_b32 v8, v20, 8, v193946; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v133947; ALIGNED-NEXT:    v_lshl_or_b32 v10, v18, 8, v173948; ALIGNED-NEXT:    v_lshl_or_b32 v11, v26, 8, v223949; ALIGNED-NEXT:    v_lshl_or_b32 v12, v28, 8, v273950; ALIGNED-NEXT:    v_lshl_or_b32 v13, v23, 8, v213951; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v03952; ALIGNED-NEXT:    v_lshl_or_b32 v14, v25, 8, v243953; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v33954; ALIGNED-NEXT:    v_lshl_or_b32 v3, v6, 16, v53955; ALIGNED-NEXT:    v_lshl_or_b32 v4, v8, 16, v73956; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 16, v93957; ALIGNED-NEXT:    v_lshl_or_b32 v6, v12, 16, v113958; ALIGNED-NEXT:    v_lshl_or_b32 v7, v14, 16, v133959; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill3960; ALIGNED-NEXT:    v_lshl_or_b32 v15, v31, 8, v303961; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill3962; ALIGNED-NEXT:    v_lshl_or_b32 v0, v34, 8, v333963; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill3964; ALIGNED-NEXT:    v_lshl_or_b32 v1, v37, 8, v323965; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill3966; ALIGNED-NEXT:    v_lshl_or_b32 v3, v36, 8, v353967; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill3968; ALIGNED-NEXT:    v_lshl_or_b32 v4, v50, 8, v393969; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill3970; ALIGNED-NEXT:    v_lshl_or_b32 v5, v49, 8, v483971; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill3972; ALIGNED-NEXT:    v_lshl_or_b32 v6, v51, 8, v383973; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill3974; ALIGNED-NEXT:    v_lshl_or_b32 v7, v53, 8, v523975; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 16, v153976; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v13977; ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 16, v43978; ALIGNED-NEXT:    s_clause 0x13979; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:853980; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:873981; ALIGNED-NEXT:    v_lshl_or_b32 v4, v7, 16, v63982; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill3983; ALIGNED-NEXT:    v_lshl_or_b32 v0, v55, 8, v293984; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill3985; ALIGNED-NEXT:    v_lshl_or_b32 v1, v67, 8, v663986; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill3987; ALIGNED-NEXT:    v_lshl_or_b32 v3, v64, 8, v543988; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill3989; ALIGNED-NEXT:    v_lshl_or_b32 v4, v68, 8, v653990; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v03991; ALIGNED-NEXT:    s_clause 0x13992; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:823993; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:863994; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill3995; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v33996; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill3997; ALIGNED-NEXT:    v_lshl_or_b32 v0, v70, 8, v693998; ALIGNED-NEXT:    s_clause 0x13999; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:744000; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:834001; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill4002; ALIGNED-NEXT:    s_waitcnt vmcnt(62)4003; ALIGNED-NEXT:    v_lshl_or_b32 v1, v80, 8, v714004; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill4005; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill4006; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill4007; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill4008; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04009; ALIGNED-NEXT:    s_clause 0x54010; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:734011; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:2124012; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:2134013; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:2144014; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:2154015; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:2114016; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill4017; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill4018; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill4019; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:724020; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill4021; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill4022; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill4023; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill4024; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill4025; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill4026; ALIGNED-NEXT:    s_clause 0x74027; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:2164028; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:2174029; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:2184030; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:2194031; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:2204032; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:2214033; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:2224034; ALIGNED-NEXT:    buffer_load_ubyte v51, v2, s[0:3], 0 offen offset:2234035; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill4036; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill4037; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill4038; ALIGNED-NEXT:    s_clause 0x24039; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:2084040; ALIGNED-NEXT:    buffer_load_ubyte v38, v2, s[0:3], 0 offen offset:2094041; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:2104042; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill4043; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill4044; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill4045; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill4046; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill4047; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill4048; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill4049; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill4050; ALIGNED-NEXT:    s_clause 0x74051; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:2244052; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:2254053; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:2264054; ALIGNED-NEXT:    buffer_load_ubyte v32, v2, s[0:3], 0 offen offset:2274055; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:2284056; ALIGNED-NEXT:    buffer_load_ubyte v33, v2, s[0:3], 0 offen offset:2294057; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:2304058; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:2314059; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill4060; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill4061; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill4062; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill4063; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill4064; ALIGNED-NEXT:    s_clause 0x44065; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:2364066; ALIGNED-NEXT:    buffer_load_ubyte v27, v2, s[0:3], 0 offen offset:2374067; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:2384068; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:2394069; ALIGNED-NEXT:    buffer_load_ubyte v25, v2, s[0:3], 0 offen offset:2354070; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill4071; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill4072; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill4073; ALIGNED-NEXT:    s_clause 0x24074; ALIGNED-NEXT:    buffer_load_ubyte v24, v2, s[0:3], 0 offen offset:2324075; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:2334076; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:2344077; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill4078; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill4079; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill4080; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill4081; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill4082; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill4083; ALIGNED-NEXT:    s_clause 0xc4084; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:2404085; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:2414086; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:2424087; ALIGNED-NEXT:    buffer_load_ubyte v16, v2, s[0:3], 0 offen offset:2434088; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:2444089; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:2454090; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:2464091; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:2474092; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:2524093; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:2534094; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:2544095; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:2554096; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:2514097; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill4098; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill4099; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill4100; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill4101; ALIGNED-NEXT:    s_waitcnt vmcnt(57)4102; ALIGNED-NEXT:    buffer_store_dword v90, off, s[0:3], s32 offset:1084 ; 4-byte Folded Spill4103; ALIGNED-NEXT:    v_lshl_or_b32 v90, v90, 8, v1214104; ALIGNED-NEXT:    s_waitcnt vmcnt(56)4105; ALIGNED-NEXT:    buffer_store_dword v106, off, s[0:3], s32 offset:1092 ; 4-byte Folded Spill4106; ALIGNED-NEXT:    buffer_store_dword v121, off, s[0:3], s32 offset:1096 ; 4-byte Folded Spill4107; ALIGNED-NEXT:    s_waitcnt vmcnt(55)4108; ALIGNED-NEXT:    buffer_store_dword v109, off, s[0:3], s32 offset:1108 ; 4-byte Folded Spill4109; ALIGNED-NEXT:    s_waitcnt vmcnt(54)4110; ALIGNED-NEXT:    buffer_store_dword v111, off, s[0:3], s32 offset:1112 ; 4-byte Folded Spill4111; ALIGNED-NEXT:    s_waitcnt vmcnt(53)4112; ALIGNED-NEXT:    buffer_store_dword v122, off, s[0:3], s32 offset:1124 ; 4-byte Folded Spill4113; ALIGNED-NEXT:    s_waitcnt vmcnt(52)4114; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill4115; ALIGNED-NEXT:    s_waitcnt vmcnt(51)4116; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill4117; ALIGNED-NEXT:    s_waitcnt vmcnt(49)4118; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill4119; ALIGNED-NEXT:    s_waitcnt vmcnt(48)4120; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill4121; ALIGNED-NEXT:    s_waitcnt vmcnt(46)4122; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill4123; ALIGNED-NEXT:    s_waitcnt vmcnt(40)4124; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill4125; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04126; ALIGNED-NEXT:    v_lshl_or_b32 v1, v81, 8, v34127; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:844128; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill4129; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill4130; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04131; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:814132; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill4133; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:804134; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4135; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill4136; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4137; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill4138; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4139; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill4140; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04141; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v74142; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:984143; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04144; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v64145; ALIGNED-NEXT:    s_clause 0x14146; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1024147; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1034148; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill4149; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v34150; ALIGNED-NEXT:    s_clause 0x14151; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:944152; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:954153; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04154; ALIGNED-NEXT:    s_clause 0x14155; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:934156; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:914157; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill4158; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:924159; ALIGNED-NEXT:    s_waitcnt vmcnt(6)4160; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill4161; ALIGNED-NEXT:    s_waitcnt vmcnt(5)4162; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill4163; ALIGNED-NEXT:    s_waitcnt vmcnt(4)4164; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill4165; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4166; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill4167; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4168; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill4169; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4170; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill4171; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04172; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v34173; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:904174; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill4175; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1014176; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04177; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:894178; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill4179; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:884180; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4181; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill4182; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4183; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill4184; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4185; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill4186; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4187; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill4188; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04189; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v34190; ALIGNED-NEXT:    s_clause 0x14191; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:994192; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1004193; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill4194; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04195; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:974196; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill4197; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:964198; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4199; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill4200; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4201; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill4202; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4203; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill4204; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4205; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill4206; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04207; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v74208; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1144209; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04210; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v64211; ALIGNED-NEXT:    s_clause 0x14212; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1184213; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1194214; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill4215; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v34216; ALIGNED-NEXT:    s_clause 0x14217; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1104218; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1114219; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04220; ALIGNED-NEXT:    s_clause 0x14221; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1094222; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1074223; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill4224; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1084225; ALIGNED-NEXT:    s_waitcnt vmcnt(6)4226; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill4227; ALIGNED-NEXT:    s_waitcnt vmcnt(5)4228; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill4229; ALIGNED-NEXT:    s_waitcnt vmcnt(4)4230; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill4231; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4232; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill4233; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4234; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill4235; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4236; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill4237; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04238; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v34239; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1064240; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill4241; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1174242; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04243; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1054244; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill4245; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1044246; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4247; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill4248; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4249; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill4250; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4251; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill4252; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4253; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill4254; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04255; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v34256; ALIGNED-NEXT:    s_clause 0x14257; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1154258; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1164259; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill4260; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04261; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1134262; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill4263; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1124264; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4265; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill4266; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4267; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill4268; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4269; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill4270; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4271; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill4272; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04273; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v74274; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1304275; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04276; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v64277; ALIGNED-NEXT:    s_clause 0x14278; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1344279; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1354280; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill4281; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v34282; ALIGNED-NEXT:    s_clause 0x14283; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1264284; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1274285; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04286; ALIGNED-NEXT:    s_clause 0x14287; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1254288; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1234289; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill4290; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1244291; ALIGNED-NEXT:    s_waitcnt vmcnt(6)4292; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1012 ; 4-byte Folded Spill4293; ALIGNED-NEXT:    s_waitcnt vmcnt(5)4294; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1016 ; 4-byte Folded Spill4295; ALIGNED-NEXT:    s_waitcnt vmcnt(4)4296; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill4297; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4298; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill4299; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4300; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill4301; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4302; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill4303; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04304; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v34305; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1224306; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill4307; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1334308; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04309; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1214310; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill4311; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1204312; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4313; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill4314; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4315; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1008 ; 4-byte Folded Spill4316; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4317; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill4318; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4319; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill4320; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04321; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v34322; ALIGNED-NEXT:    s_clause 0x14323; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1314324; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1324325; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1004 ; 4-byte Folded Spill4326; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04327; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1294328; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill4329; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1284330; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4331; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:996 ; 4-byte Folded Spill4332; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4333; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1000 ; 4-byte Folded Spill4334; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4335; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:992 ; 4-byte Folded Spill4336; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4337; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:988 ; 4-byte Folded Spill4338; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04339; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v74340; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1464341; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04342; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v64343; ALIGNED-NEXT:    s_clause 0x14344; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1504345; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1514346; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Spill4347; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v34348; ALIGNED-NEXT:    s_clause 0x14349; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1424350; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1434351; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04352; ALIGNED-NEXT:    s_clause 0x14353; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1414354; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1394355; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1028 ; 4-byte Folded Spill4356; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1404357; ALIGNED-NEXT:    s_waitcnt vmcnt(6)4358; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1120 ; 4-byte Folded Spill4359; ALIGNED-NEXT:    s_waitcnt vmcnt(5)4360; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1128 ; 4-byte Folded Spill4361; ALIGNED-NEXT:    s_waitcnt vmcnt(4)4362; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1036 ; 4-byte Folded Spill4363; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4364; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1040 ; 4-byte Folded Spill4365; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4366; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1032 ; 4-byte Folded Spill4367; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4368; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1024 ; 4-byte Folded Spill4369; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04370; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v34371; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1384372; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1056 ; 4-byte Folded Spill4373; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1494374; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04375; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1374376; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Spill4377; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1364378; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4379; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1060 ; 4-byte Folded Spill4380; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4381; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1116 ; 4-byte Folded Spill4382; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4383; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1052 ; 4-byte Folded Spill4384; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4385; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Spill4386; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04387; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v34388; ALIGNED-NEXT:    s_clause 0x14389; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1474390; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1484391; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1104 ; 4-byte Folded Spill4392; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04393; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1454394; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Spill4395; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1444396; ALIGNED-NEXT:    s_waitcnt vmcnt(3)4397; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1088 ; 4-byte Folded Spill4398; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4399; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1100 ; 4-byte Folded Spill4400; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4401; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1080 ; 4-byte Folded Spill4402; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4403; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1072 ; 4-byte Folded Spill4404; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04405; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v74406; ALIGNED-NEXT:    v_lshl_or_b32 v4, v86, 8, v964407; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04408; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v64409; ALIGNED-NEXT:    v_lshl_or_b32 v6, v10, 8, v124410; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1136 ; 4-byte Folded Spill4411; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v34412; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1584413; ALIGNED-NEXT:    v_lshl_or_b32 v5, v14, 8, v184414; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04415; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1574416; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Spill4417; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1564418; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4419; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1156 ; 4-byte Folded Spill4420; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4421; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1152 ; 4-byte Folded Spill4422; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4423; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Spill4424; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v04425; ALIGNED-NEXT:    v_lshl_or_b32 v1, v125, 8, v34426; ALIGNED-NEXT:    v_lshl_or_b32 v3, v87, 8, v974427; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04428; ALIGNED-NEXT:    v_lshl_or_b32 v1, v124, 8, v1104429; ALIGNED-NEXT:    v_lshl_or_b32 v79, v4, 16, v34430; ALIGNED-NEXT:    v_lshl_or_b32 v3, v83, 8, v844431; ALIGNED-NEXT:    v_lshl_or_b32 v4, v85, 8, v824432; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Spill4433; ALIGNED-NEXT:    v_lshl_or_b32 v0, v120, 8, v1234434; ALIGNED-NEXT:    v_lshl_or_b32 v72, v4, 16, v34435; ALIGNED-NEXT:    v_lshl_or_b32 v3, v67, 8, v804436; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04437; ALIGNED-NEXT:    v_lshl_or_b32 v1, v92, 8, v934438; ALIGNED-NEXT:    v_lshl_or_b32 v4, v64, 8, v704439; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Spill4440; ALIGNED-NEXT:    v_lshl_or_b32 v0, v104, 8, v1084441; ALIGNED-NEXT:    v_lshl_or_b32 v45, v4, 16, v34442; ALIGNED-NEXT:    v_lshl_or_b32 v3, v66, 8, v714443; ALIGNED-NEXT:    v_lshl_or_b32 v4, v52, 8, v534444; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04445; ALIGNED-NEXT:    v_lshl_or_b32 v1, v91, 8, v944446; ALIGNED-NEXT:    v_lshl_or_b32 v117, v4, 16, v34447; ALIGNED-NEXT:    v_lshl_or_b32 v3, v55, 8, v694448; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1192 ; 4-byte Folded Spill4449; ALIGNED-NEXT:    v_lshl_or_b32 v0, v95, 8, v1054450; ALIGNED-NEXT:    v_lshl_or_b32 v4, v51, 8, v544451; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04452; ALIGNED-NEXT:    v_lshl_or_b32 v1, v75, 8, v764453; ALIGNED-NEXT:    v_lshl_or_b32 v114, v4, 16, v34454; ALIGNED-NEXT:    v_lshl_or_b32 v3, v38, 8, v494455; ALIGNED-NEXT:    v_lshl_or_b32 v4, v65, 8, v394456; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1196 ; 4-byte Folded Spill4457; ALIGNED-NEXT:    v_lshl_or_b32 v0, v77, 8, v784458; ALIGNED-NEXT:    v_lshl_or_b32 v98, v4, 16, v34459; ALIGNED-NEXT:    v_lshl_or_b32 v3, v35, 8, v374460; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04461; ALIGNED-NEXT:    v_lshl_or_b32 v1, v74, 8, v624462; ALIGNED-NEXT:    v_lshl_or_b32 v4, v32, 8, v314463; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Spill4464; ALIGNED-NEXT:    v_lshl_or_b32 v0, v63, 8, v734465; ALIGNED-NEXT:    v_lshl_or_b32 v81, v4, 16, v34466; ALIGNED-NEXT:    v_lshl_or_b32 v3, v33, 8, v364467; ALIGNED-NEXT:    v_lshl_or_b32 v4, v30, 8, v344468; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04469; ALIGNED-NEXT:    v_lshl_or_b32 v1, v56, 8, v474470; ALIGNED-NEXT:    v_lshl_or_b32 v68, v4, 16, v34471; ALIGNED-NEXT:    v_lshl_or_b32 v3, v27, 8, v294472; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Spill4473; ALIGNED-NEXT:    v_lshl_or_b32 v0, v59, 8, v614474; ALIGNED-NEXT:    v_lshl_or_b32 v4, v26, 8, v284475; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04476; ALIGNED-NEXT:    v_lshl_or_b32 v1, v46, 8, v584477; ALIGNED-NEXT:    v_lshl_or_b32 v50, v4, 16, v34478; ALIGNED-NEXT:    v_lshl_or_b32 v3, v23, 8, v244479; ALIGNED-NEXT:    v_lshl_or_b32 v4, v25, 8, v224480; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Spill4481; ALIGNED-NEXT:    v_lshl_or_b32 v0, v57, 8, v604482; ALIGNED-NEXT:    v_lshl_or_b32 v48, v4, 16, v34483; ALIGNED-NEXT:    v_lshl_or_b32 v3, v19, 8, v214484; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04485; ALIGNED-NEXT:    v_lshl_or_b32 v1, v41, 8, v424486; ALIGNED-NEXT:    v_lshl_or_b32 v4, v16, 8, v154487; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1212 ; 4-byte Folded Spill4488; ALIGNED-NEXT:    v_lshl_or_b32 v0, v43, 8, v444489; ALIGNED-NEXT:    v_lshl_or_b32 v3, v4, 16, v34490; ALIGNED-NEXT:    v_lshl_or_b32 v4, v17, 8, v204491; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04492; ALIGNED-NEXT:    v_lshl_or_b32 v1, v40, 8, v1164493; ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 16, v44494; ALIGNED-NEXT:    v_lshl_or_b32 v5, v11, 8, v134495; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1216 ; 4-byte Folded Spill4496; ALIGNED-NEXT:    v_lshl_or_b32 v0, v118, 8, v1194497; ALIGNED-NEXT:    v_lshl_or_b32 v5, v6, 16, v54498; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v04499; ALIGNED-NEXT:    v_lshl_or_b32 v1, v100, 8, v1014500; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1220 ; 4-byte Folded Spill4501; ALIGNED-NEXT:    v_lshl_or_b32 v0, v112, 8, v1154502; ALIGNED-NEXT:    v_lshl_or_b32 v107, v1, 16, v04503; ALIGNED-NEXT:    v_lshl_or_b32 v0, v103, 8, v1134504; ALIGNED-NEXT:    v_lshl_or_b32 v1, v99, 8, v1024505; ALIGNED-NEXT:    v_lshl_or_b32 v88, v1, 16, v04506; ALIGNED-NEXT:    s_clause 0x34507; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:2484508; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:2494509; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:2504510; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen4511; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4512; ALIGNED-NEXT:    v_lshl_or_b32 v6, v7, 8, v84513; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4514; ALIGNED-NEXT:    v_lshl_or_b32 v89, v9, 8, v14515; ALIGNED-NEXT:    v_lshl_or_b32 v6, v89, 16, v64516; ALIGNED-NEXT:    buffer_load_ubyte v89, v2, s[0:3], 0 offen offset:14517; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4518; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1068 ; 4-byte Folded Spill4519; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4520; ALIGNED-NEXT:    buffer_store_dword v89, off, s[0:3], s32 offset:1076 ; 4-byte Folded Spill4521; ALIGNED-NEXT:    v_lshl_or_b32 v89, v89, 8, v04522; ALIGNED-NEXT:    v_lshl_or_b32 v0, v90, 16, v894523; ALIGNED-NEXT:    v_lshl_or_b32 v89, v109, 8, v1064524; ALIGNED-NEXT:    v_lshl_or_b32 v90, v122, 8, v1114525; ALIGNED-NEXT:    s_clause 0x14526; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:154527; ALIGNED-NEXT:    buffer_load_ubyte v121, v2, s[0:3], 0 offen offset:114528; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1132 ; 4-byte Folded Spill4529; ALIGNED-NEXT:    s_clause 0x24530; ALIGNED-NEXT:    buffer_load_ubyte v122, v2, s[0:3], 0 offen offset:84531; ALIGNED-NEXT:    buffer_load_ubyte v111, v2, s[0:3], 0 offen offset:94532; ALIGNED-NEXT:    buffer_load_ubyte v109, v2, s[0:3], 0 offen offset:104533; ALIGNED-NEXT:    v_lshl_or_b32 v0, v90, 16, v894534; ALIGNED-NEXT:    s_clause 0x14535; ALIGNED-NEXT:    buffer_load_ubyte v127, v2, s[0:3], 0 offen offset:134536; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:144537; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:1188 ; 4-byte Folded Spill4538; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1140 ; 4-byte Folded Spill4539; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:124540; ALIGNED-NEXT:    s_waitcnt vmcnt(7)4541; ALIGNED-NEXT:    buffer_store_dword v106, off, s[0:3], s32 offset:1168 ; 4-byte Folded Spill4542; ALIGNED-NEXT:    s_waitcnt vmcnt(1)4543; ALIGNED-NEXT:    buffer_store_dword v90, off, s[0:3], s32 offset:1164 ; 4-byte Folded Spill4544; ALIGNED-NEXT:    v_lshl_or_b32 v90, v106, 8, v904545; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4546; ALIGNED-NEXT:    v_lshl_or_b32 v89, v127, 8, v04547; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Spill4548; ALIGNED-NEXT:    v_lshl_or_b32 v0, v90, 16, v894549; ALIGNED-NEXT:    v_lshl_or_b32 v89, v111, 8, v1224550; ALIGNED-NEXT:    v_lshl_or_b32 v90, v121, 8, v1094551; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1176 ; 4-byte Folded Spill4552; ALIGNED-NEXT:    v_lshl_or_b32 v0, v90, 16, v894553; ALIGNED-NEXT:    s_clause 0x24554; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:184555; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:164556; ALIGNED-NEXT:    buffer_load_ubyte v89, v2, s[0:3], 0 offen offset:174557; ALIGNED-NEXT:    v_add_nc_u32_e32 v2, 0x100, v24558; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1184 ; 4-byte Folded Spill4559; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:2324560; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:2364561; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:2284562; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:2244563; ALIGNED-NEXT:    s_waitcnt vmcnt(2)4564; ALIGNED-NEXT:    v_lshl_or_b32 v0, v126, 8, v904565; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4566; ALIGNED-NEXT:    v_lshl_or_b32 v126, v89, 8, v1064567; ALIGNED-NEXT:    v_lshl_or_b32 v126, v0, 16, v1264568; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Reload4569; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4570; ALIGNED-NEXT:    v_add_co_u32 v3, vcc_lo, v0, s44571; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1224 ; 4-byte Folded Reload4572; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x1004573; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4574; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v4, null, s5, v0, vcc_lo4575; ALIGNED-NEXT:    v_add_co_u32 v5, vcc_lo, v3, 34576; ALIGNED-NEXT:    s_addc_u32 s5, s5, 04577; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v6, null, 0, v4, vcc_lo4578; ALIGNED-NEXT:    flat_store_byte v[5:6], v1 offset:2474579; ALIGNED-NEXT:    flat_store_byte v[5:6], v9 offset:2484580; ALIGNED-NEXT:    flat_store_byte v[5:6], v7 offset:2464581; ALIGNED-NEXT:    flat_store_byte v[5:6], v10 offset:2524582; ALIGNED-NEXT:    flat_store_byte v[5:6], v11 offset:2504583; ALIGNED-NEXT:    flat_store_byte v[5:6], v12 offset:2514584; ALIGNED-NEXT:    flat_store_byte v[5:6], v13 offset:2494585; ALIGNED-NEXT:    flat_store_byte v[5:6], v8 offset:2454586; ALIGNED-NEXT:    flat_store_byte v[5:6], v15 offset:2394587; ALIGNED-NEXT:    flat_store_byte v[5:6], v16 offset:2404588; ALIGNED-NEXT:    flat_store_byte v[5:6], v19 offset:2384589; ALIGNED-NEXT:    flat_store_byte v[5:6], v14 offset:2444590; ALIGNED-NEXT:    flat_store_byte v[5:6], v17 offset:2424591; ALIGNED-NEXT:    flat_store_byte v[5:6], v18 offset:2434592; ALIGNED-NEXT:    flat_store_byte v[5:6], v20 offset:2414593; ALIGNED-NEXT:    flat_store_byte v[5:6], v21 offset:2374594; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:2484595; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:2524596; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:2444597; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:2404598; ALIGNED-NEXT:    flat_store_byte v[5:6], v22 offset:2314599; ALIGNED-NEXT:    flat_store_byte v[5:6], v25 offset:2324600; ALIGNED-NEXT:    flat_store_byte v[5:6], v23 offset:2304601; ALIGNED-NEXT:    flat_store_byte v[5:6], v26 offset:2364602; ALIGNED-NEXT:    flat_store_byte v[5:6], v27 offset:2344603; ALIGNED-NEXT:    flat_store_byte v[5:6], v28 offset:2354604; ALIGNED-NEXT:    flat_store_byte v[5:6], v29 offset:2334605; ALIGNED-NEXT:    flat_store_byte v[5:6], v24 offset:2294606; ALIGNED-NEXT:    flat_store_byte v[5:6], v31 offset:2234607; ALIGNED-NEXT:    flat_store_byte v[5:6], v32 offset:2244608; ALIGNED-NEXT:    flat_store_byte v[5:6], v35 offset:2224609; ALIGNED-NEXT:    flat_store_byte v[5:6], v30 offset:2284610; ALIGNED-NEXT:    flat_store_byte v[5:6], v33 offset:2264611; ALIGNED-NEXT:    flat_store_byte v[5:6], v34 offset:2274612; ALIGNED-NEXT:    flat_store_byte v[5:6], v36 offset:2254613; ALIGNED-NEXT:    flat_store_byte v[5:6], v37 offset:2214614; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:1924615; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:2044616; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:2004617; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:1964618; ALIGNED-NEXT:    flat_store_byte v[5:6], v67 offset:2104619; ALIGNED-NEXT:    flat_store_byte v[5:6], v64 offset:2124620; ALIGNED-NEXT:    flat_store_byte v[5:6], v38 offset:2064621; ALIGNED-NEXT:    flat_store_byte v[5:6], v65 offset:2084622; ALIGNED-NEXT:    flat_store_byte v[5:6], v39 offset:2074623; ALIGNED-NEXT:    flat_store_byte v[5:6], v70 offset:2114624; ALIGNED-NEXT:    flat_store_byte v[5:6], v80 offset:2094625; ALIGNED-NEXT:    flat_store_byte v[5:6], v53 offset:2154626; ALIGNED-NEXT:    flat_store_byte v[5:6], v52 offset:2164627; ALIGNED-NEXT:    flat_store_byte v[5:6], v66 offset:2144628; ALIGNED-NEXT:    flat_store_byte v[5:6], v51 offset:2204629; ALIGNED-NEXT:    flat_store_byte v[5:6], v55 offset:2184630; ALIGNED-NEXT:    flat_store_byte v[5:6], v54 offset:2194631; ALIGNED-NEXT:    flat_store_byte v[5:6], v69 offset:2174632; ALIGNED-NEXT:    flat_store_byte v[5:6], v71 offset:2134633; ALIGNED-NEXT:    flat_store_byte v[5:6], v49 offset:2054634; ALIGNED-NEXT:    buffer_store_dword v72, off, s[0:3], s32 offset:2164635; ALIGNED-NEXT:    buffer_store_dword v79, off, s[0:3], s32 offset:2204636; ALIGNED-NEXT:    buffer_store_dword v88, off, s[0:3], s32 offset:2124637; ALIGNED-NEXT:    buffer_store_dword v107, off, s[0:3], s32 offset:2084638; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1220 ; 4-byte Folded Reload4639; ALIGNED-NEXT:    flat_store_byte v[5:6], v82 offset:1994640; ALIGNED-NEXT:    flat_store_byte v[5:6], v85 offset:2004641; ALIGNED-NEXT:    flat_store_byte v[5:6], v83 offset:1984642; ALIGNED-NEXT:    flat_store_byte v[5:6], v86 offset:2044643; ALIGNED-NEXT:    flat_store_byte v[5:6], v87 offset:2024644; ALIGNED-NEXT:    flat_store_byte v[5:6], v96 offset:2034645; ALIGNED-NEXT:    flat_store_byte v[5:6], v97 offset:2014646; ALIGNED-NEXT:    flat_store_byte v[5:6], v84 offset:1974647; ALIGNED-NEXT:    flat_store_byte v[5:6], v101 offset:1914648; ALIGNED-NEXT:    flat_store_byte v[5:6], v100 offset:1924649; ALIGNED-NEXT:    flat_store_byte v[5:6], v112 offset:1904650; ALIGNED-NEXT:    flat_store_byte v[5:6], v99 offset:1964651; ALIGNED-NEXT:    flat_store_byte v[5:6], v103 offset:1944652; ALIGNED-NEXT:    flat_store_byte v[5:6], v102 offset:1954653; ALIGNED-NEXT:    flat_store_byte v[5:6], v113 offset:1934654; ALIGNED-NEXT:    flat_store_byte v[5:6], v115 offset:1894655; ALIGNED-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]4656; ALIGNED-NEXT:    s_and_b32 vcc_lo, exec_lo, s64657; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4658; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2964659; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1216 ; 4-byte Folded Reload4660; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4661; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3004662; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1212 ; 4-byte Folded Reload4663; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4664; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2924665; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Reload4666; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4667; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2884668; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Reload4669; ALIGNED-NEXT:    flat_store_byte v[5:6], v116 offset:1834670; ALIGNED-NEXT:    flat_store_byte v[5:6], v40 offset:1844671; ALIGNED-NEXT:    flat_store_byte v[5:6], v118 offset:1824672; ALIGNED-NEXT:    flat_store_byte v[5:6], v41 offset:1884673; ALIGNED-NEXT:    flat_store_byte v[5:6], v43 offset:1864674; ALIGNED-NEXT:    flat_store_byte v[5:6], v42 offset:1874675; ALIGNED-NEXT:    flat_store_byte v[5:6], v44 offset:1854676; ALIGNED-NEXT:    flat_store_byte v[5:6], v119 offset:1814677; ALIGNED-NEXT:    flat_store_byte v[5:6], v47 offset:1754678; ALIGNED-NEXT:    flat_store_byte v[5:6], v56 offset:1764679; ALIGNED-NEXT:    flat_store_byte v[5:6], v59 offset:1744680; ALIGNED-NEXT:    flat_store_byte v[5:6], v46 offset:1804681; ALIGNED-NEXT:    flat_store_byte v[5:6], v57 offset:1784682; ALIGNED-NEXT:    flat_store_byte v[5:6], v58 offset:1794683; ALIGNED-NEXT:    flat_store_byte v[5:6], v60 offset:1774684; ALIGNED-NEXT:    flat_store_byte v[5:6], v61 offset:1734685; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4686; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3124687; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Reload4688; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4689; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3164690; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1196 ; 4-byte Folded Reload4691; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4692; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3084693; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1192 ; 4-byte Folded Reload4694; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4695; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3044696; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Reload4697; ALIGNED-NEXT:    flat_store_byte v[5:6], v62 offset:1674698; ALIGNED-NEXT:    flat_store_byte v[5:6], v74 offset:1684699; ALIGNED-NEXT:    flat_store_byte v[5:6], v63 offset:1664700; ALIGNED-NEXT:    flat_store_byte v[5:6], v75 offset:1724701; ALIGNED-NEXT:    flat_store_byte v[5:6], v77 offset:1704702; ALIGNED-NEXT:    flat_store_byte v[5:6], v76 offset:1714703; ALIGNED-NEXT:    flat_store_byte v[5:6], v78 offset:1694704; ALIGNED-NEXT:    flat_store_byte v[5:6], v73 offset:1654705; ALIGNED-NEXT:    flat_store_byte v[5:6], v93 offset:1594706; ALIGNED-NEXT:    flat_store_byte v[5:6], v92 offset:1604707; ALIGNED-NEXT:    flat_store_byte v[5:6], v104 offset:1584708; ALIGNED-NEXT:    flat_store_byte v[5:6], v91 offset:1644709; ALIGNED-NEXT:    flat_store_byte v[5:6], v95 offset:1624710; ALIGNED-NEXT:    flat_store_byte v[5:6], v94 offset:1634711; ALIGNED-NEXT:    flat_store_byte v[5:6], v105 offset:1614712; ALIGNED-NEXT:    flat_store_byte v[5:6], v108 offset:1574713; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4714; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2644715; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Reload4716; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4717; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2684718; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Reload4719; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4720; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2604721; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1136 ; 4-byte Folded Reload4722; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4723; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2564724; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Reload4725; ALIGNED-NEXT:    flat_store_byte v[5:6], v110 offset:1514726; ALIGNED-NEXT:    flat_store_byte v[5:6], v124 offset:1524727; ALIGNED-NEXT:    flat_store_byte v[5:6], v120 offset:1504728; ALIGNED-NEXT:    flat_store_byte v[5:6], v125 offset:1564729; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4730; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1544731; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1156 ; 4-byte Folded Reload4732; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4733; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1554734; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Reload4735; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4736; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1534737; ALIGNED-NEXT:    flat_store_byte v[5:6], v123 offset:1494738; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1104 ; 4-byte Folded Reload4739; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4740; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1434741; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Reload4742; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4743; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1444744; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Reload4745; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4746; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1424747; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Reload4748; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4749; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1484750; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1116 ; 4-byte Folded Reload4751; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4752; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1464753; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Reload4754; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4755; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1474756; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1100 ; 4-byte Folded Reload4757; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4758; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1454759; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1072 ; 4-byte Folded Reload4760; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4761; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1414762; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Reload4763; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4764; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2804765; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Reload4766; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4767; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2844768; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1028 ; 4-byte Folded Reload4769; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4770; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2764771; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Reload4772; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4773; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:2724774; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1060 ; 4-byte Folded Reload4775; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4776; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1354777; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1056 ; 4-byte Folded Reload4778; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4779; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1364780; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Reload4781; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4782; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1344783; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1040 ; 4-byte Folded Reload4784; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4785; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1404786; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1032 ; 4-byte Folded Reload4787; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4788; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1384789; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1036 ; 4-byte Folded Reload4790; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4791; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1394792; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1024 ; 4-byte Folded Reload4793; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4794; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1374795; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Reload4796; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4797; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1334798; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1004 ; 4-byte Folded Reload4799; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4800; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1274801; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:996 ; 4-byte Folded Reload4802; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4803; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1284804; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:992 ; 4-byte Folded Reload4805; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4806; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1264807; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1016 ; 4-byte Folded Reload4808; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4809; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1324810; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Reload4811; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4812; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1304813; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1012 ; 4-byte Folded Reload4814; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4815; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1314816; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1000 ; 4-byte Folded Reload4817; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4818; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1294819; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:988 ; 4-byte Folded Reload4820; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4821; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:1284822; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:984 ; 4-byte Folded Reload4823; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4824; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3604825; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload4826; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4827; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3644828; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload4829; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4830; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3564831; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload4832; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4833; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3524834; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:980 ; 4-byte Folded Reload4835; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4836; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1194837; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Reload4838; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4839; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1204840; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload4841; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4842; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1184843; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload4844; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4845; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1244846; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload4847; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4848; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1224849; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload4850; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4851; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1234852; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload4853; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4854; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1214855; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload4856; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4857; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1174858; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload4859; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4860; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1114861; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload4862; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4863; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1124864; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload4865; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4866; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1104867; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload4868; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4869; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1164870; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload4871; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4872; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1144873; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload4874; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4875; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1154876; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload4877; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4878; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1134879; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload4880; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4881; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1094882; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload4883; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4884; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3764885; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload4886; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4887; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3804888; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload4889; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4890; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3724891; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload4892; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4893; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3684894; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload4895; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4896; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1034897; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload4898; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4899; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1044900; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload4901; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4902; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1024903; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload4904; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4905; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1084906; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload4907; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4908; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1064909; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload4910; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4911; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1074912; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload4913; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4914; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1054915; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload4916; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4917; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1014918; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload4919; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4920; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:954921; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload4922; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4923; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:964924; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload4925; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4926; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:944927; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload4928; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4929; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1004930; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload4931; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4932; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:984933; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload4934; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4935; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:994936; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload4937; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4938; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:974939; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload4940; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4941; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:934942; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload4943; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4944; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3284945; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload4946; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4947; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3324948; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload4949; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4950; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3244951; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload4952; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4953; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3204954; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload4955; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4956; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:874957; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload4958; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4959; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:884960; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload4961; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4962; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:864963; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload4964; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4965; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:924966; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload4967; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4968; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:904969; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload4970; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4971; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:914972; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload4973; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4974; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:894975; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload4976; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4977; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:854978; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload4979; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4980; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:794981; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload4982; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4983; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:804984; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload4985; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4986; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:784987; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload4988; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4989; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:844990; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload4991; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4992; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:824993; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload4994; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4995; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:834996; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload4997; ALIGNED-NEXT:    s_waitcnt vmcnt(0)4998; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:814999; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload5000; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5001; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:775002; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload5003; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5004; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3445005; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload5006; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5007; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3485008; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload5009; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5010; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3405011; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload5012; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5013; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3365014; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload5015; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5016; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:715017; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload5018; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5019; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:725020; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload5021; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5022; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:705023; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload5024; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5025; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:765026; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload5027; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5028; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:745029; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload5030; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5031; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:755032; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload5033; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5034; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:735035; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload5036; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5037; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:695038; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload5039; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5040; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:635041; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload5042; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5043; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:645044; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload5045; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5046; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:625047; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload5048; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5049; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:685050; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload5051; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5052; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:665053; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload5054; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5055; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:675056; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload5057; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5058; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:655059; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload5060; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5061; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:645062; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload5063; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5064; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4245065; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload5066; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5067; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4285068; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload5069; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5070; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4205071; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload5072; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5073; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4165074; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload5075; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5076; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:585077; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload5078; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5079; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:555080; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload5081; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5082; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:565083; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload5084; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5085; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:545086; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload5087; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5088; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:605089; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload5090; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5091; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:595092; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload5093; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5094; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:575095; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload5096; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5097; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:535098; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload5099; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5100; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:505101; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload5102; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5103; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:475104; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload5105; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5106; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:485107; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload5108; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5109; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:465110; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload5111; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5112; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:525113; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload5114; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5115; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:515116; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload5117; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5118; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:495119; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload5120; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5121; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:455122; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload5123; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5124; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4445125; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload5126; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5127; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4405128; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload5129; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5130; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4365131; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload5132; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5133; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4325134; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload5135; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5136; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:405137; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload5138; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5139; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:395140; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload5141; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5142; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:385143; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload5144; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5145; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:375146; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload5147; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5148; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:445149; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload5150; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5151; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:435152; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload5153; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5154; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:425155; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload5156; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5157; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:415158; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload5159; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5160; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:325161; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload5162; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5163; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:315164; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload5165; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5166; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:305167; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload5168; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5169; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:325170; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload5171; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5172; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:365173; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload5174; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5175; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:355176; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload5177; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5178; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:345179; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload5180; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5181; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:335182; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload5183; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5184; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3925185; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload5186; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5187; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3965188; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload5189; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5190; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:3885191; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:3845192; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload5193; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5194; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:235195; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload5196; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5197; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:245198; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload5199; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5200; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:225201; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload5202; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5203; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:285204; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload5205; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5206; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:265207; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload5208; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5209; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:275210; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload5211; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5212; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:255213; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload5214; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5215; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:215216; ALIGNED-NEXT:    flat_store_byte v[5:6], v90 offset:155217; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Reload5218; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5219; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:165220; ALIGNED-NEXT:    flat_store_byte v[5:6], v89 offset:145221; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload5222; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5223; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:205224; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload5225; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5226; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:185227; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload5228; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5229; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:195230; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload5231; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5232; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:175233; ALIGNED-NEXT:    flat_store_byte v[3:4], v106 offset:165234; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1184 ; 4-byte Folded Reload5235; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5236; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4085237; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1176 ; 4-byte Folded Reload5238; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5239; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4125240; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1140 ; 4-byte Folded Reload5241; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5242; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4045243; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1132 ; 4-byte Folded Reload5244; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5245; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4005246; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Reload5247; ALIGNED-NEXT:    flat_store_byte v[5:6], v109 offset:75248; ALIGNED-NEXT:    flat_store_byte v[5:6], v121 offset:85249; ALIGNED-NEXT:    flat_store_byte v[5:6], v127 offset:105250; ALIGNED-NEXT:    flat_store_byte v[5:6], v111 offset:65251; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5252; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:125253; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1164 ; 4-byte Folded Reload5254; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5255; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:115256; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Reload5257; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5258; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:95259; ALIGNED-NEXT:    flat_store_byte v[3:4], v122 offset:85260; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1096 ; 4-byte Folded Reload5261; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5262; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:25263; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Reload5264; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5265; ALIGNED-NEXT:    flat_store_byte v[5:6], v05266; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1076 ; 4-byte Folded Reload5267; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5268; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:15269; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Reload5270; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5271; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:45272; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1108 ; 4-byte Folded Reload5273; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5274; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:25275; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1112 ; 4-byte Folded Reload5276; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5277; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:35278; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Reload5279; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5280; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:45281; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1068 ; 4-byte Folded Reload5282; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5283; ALIGNED-NEXT:    flat_store_byte v[3:4], v05284; ALIGNED-NEXT:    s_cbranch_vccnz .LBB4_15285; ALIGNED-NEXT:  ; %bb.2: ; %memcpy-split5286; ALIGNED-NEXT:    s_clause 0x2f ; 192-byte Folded Reload5287; ALIGNED-NEXT:    buffer_load_dword v127, off, s[0:3], s325288; ALIGNED-NEXT:    buffer_load_dword v126, off, s[0:3], s32 offset:45289; ALIGNED-NEXT:    buffer_load_dword v125, off, s[0:3], s32 offset:85290; ALIGNED-NEXT:    buffer_load_dword v124, off, s[0:3], s32 offset:125291; ALIGNED-NEXT:    buffer_load_dword v123, off, s[0:3], s32 offset:165292; ALIGNED-NEXT:    buffer_load_dword v122, off, s[0:3], s32 offset:205293; ALIGNED-NEXT:    buffer_load_dword v121, off, s[0:3], s32 offset:245294; ALIGNED-NEXT:    buffer_load_dword v120, off, s[0:3], s32 offset:285295; ALIGNED-NEXT:    buffer_load_dword v111, off, s[0:3], s32 offset:325296; ALIGNED-NEXT:    buffer_load_dword v110, off, s[0:3], s32 offset:365297; ALIGNED-NEXT:    buffer_load_dword v109, off, s[0:3], s32 offset:405298; ALIGNED-NEXT:    buffer_load_dword v108, off, s[0:3], s32 offset:445299; ALIGNED-NEXT:    buffer_load_dword v107, off, s[0:3], s32 offset:485300; ALIGNED-NEXT:    buffer_load_dword v106, off, s[0:3], s32 offset:525301; ALIGNED-NEXT:    buffer_load_dword v105, off, s[0:3], s32 offset:565302; ALIGNED-NEXT:    buffer_load_dword v104, off, s[0:3], s32 offset:605303; ALIGNED-NEXT:    buffer_load_dword v95, off, s[0:3], s32 offset:645304; ALIGNED-NEXT:    buffer_load_dword v94, off, s[0:3], s32 offset:685305; ALIGNED-NEXT:    buffer_load_dword v93, off, s[0:3], s32 offset:725306; ALIGNED-NEXT:    buffer_load_dword v92, off, s[0:3], s32 offset:765307; ALIGNED-NEXT:    buffer_load_dword v91, off, s[0:3], s32 offset:805308; ALIGNED-NEXT:    buffer_load_dword v90, off, s[0:3], s32 offset:845309; ALIGNED-NEXT:    buffer_load_dword v89, off, s[0:3], s32 offset:885310; ALIGNED-NEXT:    buffer_load_dword v88, off, s[0:3], s32 offset:925311; ALIGNED-NEXT:    buffer_load_dword v79, off, s[0:3], s32 offset:965312; ALIGNED-NEXT:    buffer_load_dword v78, off, s[0:3], s32 offset:1005313; ALIGNED-NEXT:    buffer_load_dword v77, off, s[0:3], s32 offset:1045314; ALIGNED-NEXT:    buffer_load_dword v76, off, s[0:3], s32 offset:1085315; ALIGNED-NEXT:    buffer_load_dword v75, off, s[0:3], s32 offset:1125316; ALIGNED-NEXT:    buffer_load_dword v74, off, s[0:3], s32 offset:1165317; ALIGNED-NEXT:    buffer_load_dword v73, off, s[0:3], s32 offset:1205318; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s32 offset:1245319; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:1285320; ALIGNED-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:1325321; ALIGNED-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:1365322; ALIGNED-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:1405323; ALIGNED-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:1445324; ALIGNED-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:1485325; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:1525326; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:1565327; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:1605328; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:1645329; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:1685330; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:1725331; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:1765332; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:1805333; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:1845334; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:1885335; ALIGNED-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)5336; ALIGNED-NEXT:    s_setpc_b64 s[30:31]5337;5338; UNROLL3-LABEL: memcpy_p0_p5_sz2048:5339; UNROLL3:       ; %bb.0: ; %entry5340; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)5341; UNROLL3-NEXT:    v_mov_b32_e32 v3, v25342; UNROLL3-NEXT:    s_mov_b64 s[4:5], 05343; UNROLL3-NEXT:    s_inst_prefetch 0x15344; UNROLL3-NEXT:    .p2align 65345; UNROLL3-NEXT:  .LBB4_1: ; %load-store-loop5346; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=15347; UNROLL3-NEXT:    s_clause 0xb5348; UNROLL3-NEXT:    buffer_load_dword v4, v3, s[0:3], 0 offen5349; UNROLL3-NEXT:    buffer_load_dword v5, v3, s[0:3], 0 offen offset:45350; UNROLL3-NEXT:    buffer_load_dword v6, v3, s[0:3], 0 offen offset:85351; UNROLL3-NEXT:    buffer_load_dword v7, v3, s[0:3], 0 offen offset:125352; UNROLL3-NEXT:    buffer_load_dword v8, v3, s[0:3], 0 offen offset:165353; UNROLL3-NEXT:    buffer_load_dword v9, v3, s[0:3], 0 offen offset:205354; UNROLL3-NEXT:    buffer_load_dword v10, v3, s[0:3], 0 offen offset:245355; UNROLL3-NEXT:    buffer_load_dword v11, v3, s[0:3], 0 offen offset:285356; UNROLL3-NEXT:    buffer_load_dword v12, v3, s[0:3], 0 offen offset:325357; UNROLL3-NEXT:    buffer_load_dword v13, v3, s[0:3], 0 offen offset:365358; UNROLL3-NEXT:    buffer_load_dword v14, v3, s[0:3], 0 offen offset:405359; UNROLL3-NEXT:    buffer_load_dword v15, v3, s[0:3], 0 offen offset:445360; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s45361; UNROLL3-NEXT:    s_add_u32 s4, s4, 485362; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo5363; UNROLL3-NEXT:    s_addc_u32 s5, s5, 05364; UNROLL3-NEXT:    v_add_nc_u32_e32 v3, 48, v35365; UNROLL3-NEXT:    v_cmp_gt_u64_e64 s6, 0x7e0, s[4:5]5366; UNROLL3-NEXT:    s_waitcnt vmcnt(4)5367; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11] offset:165368; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7]5369; UNROLL3-NEXT:    s_waitcnt vmcnt(0)5370; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:325371; UNROLL3-NEXT:    s_and_b32 vcc_lo, exec_lo, s65372; UNROLL3-NEXT:    s_cbranch_vccnz .LBB4_15373; UNROLL3-NEXT:  ; %bb.2: ; %memcpy-split5374; UNROLL3-NEXT:    s_inst_prefetch 0x25375; UNROLL3-NEXT:    s_clause 0x35376; UNROLL3-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:20165377; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:20205378; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:20245379; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:20285380; UNROLL3-NEXT:    s_waitcnt vmcnt(0)5381; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:20165382; UNROLL3-NEXT:    s_clause 0x35383; UNROLL3-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:20325384; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:20365385; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:20405386; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:20445387; UNROLL3-NEXT:    s_waitcnt vmcnt(0)5388; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:20325389; UNROLL3-NEXT:    s_waitcnt lgkmcnt(0)5390; UNROLL3-NEXT:    s_setpc_b64 s[30:31]5391entry:5392  tail call void @llvm.memcpy.p0.p5.i64(ptr addrspace(0) noundef nonnull align 1 %dst, ptr addrspace(5) noundef nonnull align 1 %src, i64 2048, i1 false)5393  ret void5394}5395 5396 5397; memmove for address spaces 0, 1, 4, 55398 5399define void @memmove_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0) align 1 readonly %src) {5400; CHECK-LABEL: memmove_p0_p0_sz2048:5401; CHECK:       ; %bb.0: ; %entry5402; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)5403; CHECK-NEXT:    s_mov_b32 s4, exec_lo5404; CHECK-NEXT:    v_cmpx_ge_u64_e64 v[2:3], v[0:1]5405; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s45406; CHECK-NEXT:    s_cbranch_execz .LBB5_35407; CHECK-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader5408; CHECK-NEXT:    s_mov_b64 s[4:5], 05409; CHECK-NEXT:  .LBB5_2: ; %memmove_fwd_loop5410; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=15411; CHECK-NEXT:    v_add_co_u32 v52, vcc_lo, v2, s45412; CHECK-NEXT:    v_add_co_ci_u32_e64 v53, null, s5, v3, vcc_lo5413; CHECK-NEXT:    s_clause 0x15414; CHECK-NEXT:    flat_load_dwordx4 v[4:7], v[52:53] offset:165415; CHECK-NEXT:    flat_load_dwordx4 v[8:11], v[52:53]5416; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v52, 485417; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v53, vcc_lo5418; CHECK-NEXT:    v_add_co_u32 v80, vcc_lo, v52, 0x605419; CHECK-NEXT:    v_add_co_ci_u32_e64 v81, null, 0, v53, vcc_lo5420; CHECK-NEXT:    v_add_co_u32 v84, vcc_lo, 0x50, v525421; CHECK-NEXT:    v_add_co_ci_u32_e64 v85, null, 0, v53, vcc_lo5422; CHECK-NEXT:    s_clause 0xd5423; CHECK-NEXT:    flat_load_dwordx4 v[12:15], v[52:53] offset:325424; CHECK-NEXT:    flat_load_dwordx4 v[16:19], v[52:53] offset:645425; CHECK-NEXT:    flat_load_dwordx4 v[20:23], v[80:81] offset:1285426; CHECK-NEXT:    flat_load_dwordx4 v[24:27], v[80:81] offset:1445427; CHECK-NEXT:    flat_load_dwordx4 v[28:31], v[80:81] offset:965428; CHECK-NEXT:    flat_load_dwordx4 v[32:35], v[80:81] offset:1125429; CHECK-NEXT:    flat_load_dwordx4 v[36:39], v[80:81] offset:645430; CHECK-NEXT:    flat_load_dwordx4 v[48:51], v[80:81] offset:805431; CHECK-NEXT:    flat_load_dwordx4 v[52:55], v[52:53] offset:1285432; CHECK-NEXT:    flat_load_dwordx4 v[64:67], v[80:81] offset:485433; CHECK-NEXT:    flat_load_dwordx4 v[68:71], v[80:81]5434; CHECK-NEXT:    flat_load_dwordx4 v[80:83], v[80:81] offset:165435; CHECK-NEXT:    flat_load_dwordx4 v[84:87], v[84:85]5436; CHECK-NEXT:    flat_load_dwordx4 v[96:99], v[96:97]5437; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s45438; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo5439; CHECK-NEXT:    s_add_u32 s4, s4, 0x1005440; CHECK-NEXT:    s_addc_u32 s5, s5, 05441; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 485442; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0x8005443; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo5444; CHECK-NEXT:    s_waitcnt vmcnt(12) lgkmcnt(12)5445; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[16:19] offset:645446; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[12:15] offset:325447; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[8:11]5448; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[4:7] offset:165449; CHECK-NEXT:    s_waitcnt vmcnt(11) lgkmcnt(15)5450; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[20:23] offset:1765451; CHECK-NEXT:    s_waitcnt vmcnt(10) lgkmcnt(15)5452; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:1925453; CHECK-NEXT:    s_waitcnt vmcnt(9) lgkmcnt(15)5454; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:1445455; CHECK-NEXT:    s_waitcnt vmcnt(8) lgkmcnt(15)5456; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[32:35] offset:1605457; CHECK-NEXT:    s_waitcnt vmcnt(7) lgkmcnt(15)5458; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:1125459; CHECK-NEXT:    s_waitcnt vmcnt(6) lgkmcnt(15)5460; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:1285461; CHECK-NEXT:    s_waitcnt vmcnt(5) lgkmcnt(15)5462; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:1285463; CHECK-NEXT:    s_waitcnt vmcnt(4) lgkmcnt(15)5464; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:965465; CHECK-NEXT:    s_waitcnt vmcnt(3) lgkmcnt(15)5466; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71] offset:485467; CHECK-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(15)5468; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[80:83] offset:645469; CHECK-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(15)5470; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[84:87] offset:325471; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(15)5472; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[96:99]5473; CHECK-NEXT:    s_cbranch_scc1 .LBB5_25474; CHECK-NEXT:  .LBB5_3: ; %Flow55475; CHECK-NEXT:    s_andn2_saveexec_b32 s8, s65476; CHECK-NEXT:    s_cbranch_execz .LBB5_65477; CHECK-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader5478; CHECK-NEXT:    s_movk_i32 s6, 0xff005479; CHECK-NEXT:    s_mov_b64 s[4:5], 0x7005480; CHECK-NEXT:    s_mov_b32 s7, -15481; CHECK-NEXT:  .LBB5_5: ; %memmove_bwd_loop5482; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=15483; CHECK-NEXT:    v_add_co_u32 v24, vcc_lo, v2, s45484; CHECK-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v3, vcc_lo5485; CHECK-NEXT:    s_clause 0x45486; CHECK-NEXT:    flat_load_dwordx4 v[4:7], v[24:25] offset:645487; CHECK-NEXT:    flat_load_dwordx4 v[8:11], v[24:25] offset:325488; CHECK-NEXT:    flat_load_dwordx4 v[12:15], v[24:25]5489; CHECK-NEXT:    flat_load_dwordx4 v[16:19], v[24:25] offset:165490; CHECK-NEXT:    flat_load_dwordx4 v[20:23], v[24:25] offset:1285491; CHECK-NEXT:    v_add_co_u32 v80, vcc_lo, v24, 0x605492; CHECK-NEXT:    v_add_co_ci_u32_e64 v81, null, 0, v25, vcc_lo5493; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v24, 485494; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v25, vcc_lo5495; CHECK-NEXT:    v_add_co_u32 v84, vcc_lo, 0x50, v245496; CHECK-NEXT:    v_add_co_ci_u32_e64 v85, null, 0, v25, vcc_lo5497; CHECK-NEXT:    s_clause 0xa5498; CHECK-NEXT:    flat_load_dwordx4 v[24:27], v[80:81] offset:1285499; CHECK-NEXT:    flat_load_dwordx4 v[28:31], v[80:81] offset:1445500; CHECK-NEXT:    flat_load_dwordx4 v[32:35], v[80:81] offset:965501; CHECK-NEXT:    flat_load_dwordx4 v[36:39], v[80:81] offset:1125502; CHECK-NEXT:    flat_load_dwordx4 v[48:51], v[80:81] offset:645503; CHECK-NEXT:    flat_load_dwordx4 v[52:55], v[80:81] offset:805504; CHECK-NEXT:    flat_load_dwordx4 v[64:67], v[80:81] offset:485505; CHECK-NEXT:    flat_load_dwordx4 v[68:71], v[80:81]5506; CHECK-NEXT:    flat_load_dwordx4 v[80:83], v[80:81] offset:165507; CHECK-NEXT:    flat_load_dwordx4 v[84:87], v[84:85]5508; CHECK-NEXT:    flat_load_dwordx4 v[96:99], v[96:97]5509; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s45510; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo5511; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff005512; CHECK-NEXT:    s_addc_u32 s5, s5, -15513; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 485514; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]5515; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo5516; CHECK-NEXT:    s_waitcnt vmcnt(15) lgkmcnt(15)5517; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[4:7] offset:645518; CHECK-NEXT:    s_waitcnt vmcnt(14) lgkmcnt(15)5519; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[8:11] offset:325520; CHECK-NEXT:    s_waitcnt vmcnt(13) lgkmcnt(15)5521; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[12:15]5522; CHECK-NEXT:    s_waitcnt vmcnt(12) lgkmcnt(15)5523; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[16:19] offset:165524; CHECK-NEXT:    s_waitcnt vmcnt(11) lgkmcnt(15)5525; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[20:23] offset:1285526; CHECK-NEXT:    s_waitcnt vmcnt(10) lgkmcnt(15)5527; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:1765528; CHECK-NEXT:    s_waitcnt vmcnt(9) lgkmcnt(15)5529; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:1925530; CHECK-NEXT:    s_waitcnt vmcnt(8) lgkmcnt(15)5531; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[32:35] offset:1445532; CHECK-NEXT:    s_waitcnt vmcnt(7) lgkmcnt(15)5533; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:1605534; CHECK-NEXT:    s_waitcnt vmcnt(6) lgkmcnt(15)5535; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:1125536; CHECK-NEXT:    s_waitcnt vmcnt(5) lgkmcnt(15)5537; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:1285538; CHECK-NEXT:    s_waitcnt vmcnt(4) lgkmcnt(15)5539; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:965540; CHECK-NEXT:    s_waitcnt vmcnt(3) lgkmcnt(15)5541; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71] offset:485542; CHECK-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(15)5543; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[80:83] offset:645544; CHECK-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(15)5545; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[84:87] offset:325546; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(15)5547; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[96:99]5548; CHECK-NEXT:    s_cbranch_scc0 .LBB5_55549; CHECK-NEXT:  .LBB5_6: ; %Flow65550; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s85551; CHECK-NEXT:    s_waitcnt lgkmcnt(0)5552; CHECK-NEXT:    s_setpc_b64 s[30:31]5553;5554; ALIGNED-LABEL: memmove_p0_p0_sz2048:5555; ALIGNED:       ; %bb.0: ; %entry5556; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)5557; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill5558; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill5559; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill5560; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill5561; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill5562; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill5563; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill5564; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill5565; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill5566; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill5567; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill5568; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill5569; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill5570; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill5571; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill5572; ALIGNED-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill5573; ALIGNED-NEXT:    buffer_store_dword v72, off, s[0:3], s32 ; 4-byte Folded Spill5574; ALIGNED-NEXT:    s_mov_b32 s4, exec_lo5575; ALIGNED-NEXT:    v_cmpx_ge_u64_e64 v[2:3], v[0:1]5576; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s45577; ALIGNED-NEXT:    s_cbranch_execz .LBB5_35578; ALIGNED-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader5579; ALIGNED-NEXT:    s_mov_b64 s[4:5], 05580; ALIGNED-NEXT:  .LBB5_2: ; %memmove_fwd_loop5581; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=15582; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, s45583; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, s5, v3, vcc_lo5584; ALIGNED-NEXT:    v_add_co_u32 v6, vcc_lo, v4, 485585; ALIGNED-NEXT:    flat_load_dwordx4 v[20:23], v[4:5] offset:1285586; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v5, vcc_lo5587; ALIGNED-NEXT:    v_add_co_u32 v8, vcc_lo, 0x50, v45588; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v5, vcc_lo5589; ALIGNED-NEXT:    v_add_co_u32 v12, vcc_lo, 0x60, v45590; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v5, vcc_lo5591; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, 0x70, v45592; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v5, vcc_lo5593; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, 0x90, v45594; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v5, vcc_lo5595; ALIGNED-NEXT:    v_add_co_u32 v38, vcc_lo, 0xa0, v45596; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v39, null, 0, v5, vcc_lo5597; ALIGNED-NEXT:    v_add_co_u32 v52, vcc_lo, 0xb0, v45598; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v53, null, 0, v5, vcc_lo5599; ALIGNED-NEXT:    v_add_co_u32 v54, vcc_lo, 0xc0, v45600; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v55, null, 0, v5, vcc_lo5601; ALIGNED-NEXT:    v_add_co_u32 v64, vcc_lo, 0xd0, v45602; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v65, null, 0, v5, vcc_lo5603; ALIGNED-NEXT:    v_add_co_u32 v70, vcc_lo, 0xe0, v45604; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v71, null, 0, v5, vcc_lo5605; ALIGNED-NEXT:    v_add_co_u32 v80, vcc_lo, 0xf0, v45606; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v81, null, 0, v5, vcc_lo5607; ALIGNED-NEXT:    s_clause 0xe5608; ALIGNED-NEXT:    flat_load_dwordx4 v[34:37], v[4:5]5609; ALIGNED-NEXT:    flat_load_dwordx4 v[30:33], v[4:5] offset:165610; ALIGNED-NEXT:    flat_load_dwordx4 v[26:29], v[4:5] offset:325611; ALIGNED-NEXT:    flat_load_dwordx4 v[48:51], v[4:5] offset:645612; ALIGNED-NEXT:    flat_load_dwordx4 v[4:7], v[6:7]5613; ALIGNED-NEXT:    flat_load_dwordx4 v[8:11], v[8:9]5614; ALIGNED-NEXT:    flat_load_dwordx4 v[12:15], v[12:13]5615; ALIGNED-NEXT:    flat_load_dwordx4 v[16:19], v[16:17]5616; ALIGNED-NEXT:    flat_load_dwordx4 v[112:115], v[24:25]5617; ALIGNED-NEXT:    flat_load_dwordx4 v[82:85], v[38:39]5618; ALIGNED-NEXT:    flat_load_dwordx4 v[116:119], v[52:53]5619; ALIGNED-NEXT:    flat_load_dwordx4 v[66:69], v[54:55]5620; ALIGNED-NEXT:    flat_load_dwordx4 v[40:43], v[64:65]5621; ALIGNED-NEXT:    flat_load_dwordx4 v[52:55], v[70:71]5622; ALIGNED-NEXT:    flat_load_dwordx4 v[44:47], v[80:81]5623; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v0, s45624; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v1, vcc_lo5625; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x1005626; ALIGNED-NEXT:    s_addc_u32 s5, s5, 05627; ALIGNED-NEXT:    s_waitcnt vmcnt(15) lgkmcnt(15)5628; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:1605629; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:1645630; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:1685631; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:1725632; ALIGNED-NEXT:    s_clause 0x35633; ALIGNED-NEXT:    buffer_load_dword v102, off, s[0:3], s32 offset:1725634; ALIGNED-NEXT:    buffer_load_dword v103, off, s[0:3], s32 offset:1685635; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:1645636; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:1605637; ALIGNED-NEXT:    v_add_co_u32 v22, vcc_lo, v24, 65638; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v25, vcc_lo5639; ALIGNED-NEXT:    v_add_co_u32 v20, vcc_lo, v24, 35640; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v25, vcc_lo5641; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0x8005642; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5643; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v102 offset:1365644; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:1375645; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5646; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v103 offset:1325647; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:1335648; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5649; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v56 offset:1285650; ALIGNED-NEXT:    flat_store_byte v[20:21], v56 offset:1295651; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5652; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v57 offset:1245653; ALIGNED-NEXT:    flat_store_byte v[24:25], v57 offset:1285654; ALIGNED-NEXT:    s_waitcnt lgkmcnt(19)5655; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:2245656; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:2285657; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:2325658; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:2365659; ALIGNED-NEXT:    s_clause 0x35660; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:2365661; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:2325662; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:2285663; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:2245664; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5665; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v101 offset:725666; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:735667; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5668; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v100 offset:685669; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:695670; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5671; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v99 offset:645672; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:655673; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5674; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v98 offset:605675; ALIGNED-NEXT:    flat_store_byte v[24:25], v98 offset:645676; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:3205677; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:3245678; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:3285679; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:3325680; ALIGNED-NEXT:    s_clause 0x35681; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:3285682; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:3325683; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:3205684; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:3245685; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5686; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v29 offset:365687; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:375688; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5689; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v28 offset:405690; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:415691; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5692; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v27 offset:285693; ALIGNED-NEXT:    flat_store_byte v[24:25], v27 offset:325694; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5695; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v26 offset:325696; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:335697; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:2725698; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:2765699; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:2805700; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:2845701; ALIGNED-NEXT:    s_clause 0x35702; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:2845703; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:2805704; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:2765705; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:2725706; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5707; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v33 offset:245708; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:255709; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5710; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v32 offset:205711; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:215712; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5713; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v31 offset:165714; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:175715; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5716; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v30 offset:125717; ALIGNED-NEXT:    flat_store_byte v[24:25], v30 offset:165718; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:2885719; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:2925720; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:2965721; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:3005722; ALIGNED-NEXT:    s_clause 0x35723; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:3005724; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:2965725; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:2925726; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:2885727; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5728; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v37 offset:85729; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:95730; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5731; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v36 offset:45732; ALIGNED-NEXT:    flat_store_byte v[24:25], v36 offset:85733; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5734; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v355735; ALIGNED-NEXT:    flat_store_byte v[24:25], v35 offset:45736; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5737; ALIGNED-NEXT:    flat_store_byte_d16_hi v[24:25], v34 offset:25738; ALIGNED-NEXT:    flat_store_byte v[24:25], v345739; ALIGNED-NEXT:    s_waitcnt lgkmcnt(40)5740; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:1125741; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:1165742; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:1205743; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:1245744; ALIGNED-NEXT:    s_clause 0x35745; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:1245746; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:1205747; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:1165748; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:1125749; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5750; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v49 offset:2485751; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:2495752; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5753; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v48 offset:2445754; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:2455755; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5756; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v39 offset:2405757; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:2415758; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5759; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v38 offset:2365760; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:2375761; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:1285762; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:1325763; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:1365764; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:1405765; ALIGNED-NEXT:    s_clause 0x35766; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:1405767; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:1365768; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:1325769; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:1285770; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5771; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v53 offset:2325772; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:2335773; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5774; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v52 offset:2285775; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:2295776; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5777; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v51 offset:2245778; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:2255779; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5780; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v50 offset:2205781; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:2215782; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:805783; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:845784; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:885785; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:925786; ALIGNED-NEXT:    s_clause 0x35787; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:925788; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:885789; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:845790; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:805791; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5792; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v65 offset:2165793; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:2175794; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5795; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v64 offset:2125796; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:2135797; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5798; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v55 offset:2085799; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:2095800; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5801; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v54 offset:2045802; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:2055803; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:965804; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:1005805; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:1045806; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:1085807; ALIGNED-NEXT:    s_clause 0x35808; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:1085809; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:1045810; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:1005811; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:965812; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5813; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v69 offset:2005814; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:2015815; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5816; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v68 offset:1965817; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:1975818; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5819; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v67 offset:1925820; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:1935821; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5822; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v66 offset:1885823; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:1895824; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:1765825; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:1805826; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:1845827; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:1885828; ALIGNED-NEXT:    s_clause 0x35829; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:1885830; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:1845831; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:1805832; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:1765833; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5834; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v81 offset:1845835; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:1855836; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5837; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v80 offset:1805838; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:1815839; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5840; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v71 offset:1765841; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:1775842; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5843; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v70 offset:1725844; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:1735845; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:1925846; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:1965847; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:2005848; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:2045849; ALIGNED-NEXT:    s_clause 0x35850; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:2045851; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:2005852; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:1965853; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:1925854; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5855; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v85 offset:1685856; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:1695857; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5858; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v84 offset:1645859; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:1655860; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5861; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v83 offset:1605862; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:1615863; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5864; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v82 offset:1565865; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:1575866; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:1445867; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:1485868; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:1525869; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:1565870; ALIGNED-NEXT:    s_clause 0x35871; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:1565872; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:1525873; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:1485874; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:1445875; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v1025876; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v1025877; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v1035878; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v1035879; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v565880; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:1405881; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v565882; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:1385883; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v575884; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v575885; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:1365886; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:1345887; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:1325888; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:1305889; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:1285890; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:1265891; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5892; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v97 offset:1525893; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:1535894; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5895; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v96 offset:1485896; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:1495897; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5898; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v87 offset:1445899; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:1455900; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5901; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v86 offset:1405902; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:1415903; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:2405904; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:2445905; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:2485906; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:2525907; ALIGNED-NEXT:    s_clause 0x35908; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:2525909; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:2485910; ALIGNED-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:2445911; ALIGNED-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:2405912; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v1015913; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1015914; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v1005915; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v1005916; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v995917; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v995918; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v985919; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v985920; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v345921; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v375922; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v375923; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v345924; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5925; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v19 offset:1205926; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:1215927; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5928; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v18 offset:1165929; ALIGNED-NEXT:    flat_store_byte v[20:21], v18 offset:1175930; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5931; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v17 offset:1125932; ALIGNED-NEXT:    flat_store_byte v[20:21], v17 offset:1135933; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5934; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v16 offset:1085935; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:1095936; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:2565937; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:2605938; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:2645939; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:2685940; ALIGNED-NEXT:    s_clause 0x35941; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:2685942; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:2645943; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:2605944; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:2565945; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5946; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v15 offset:1045947; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:1055948; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5949; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v14 offset:1005950; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:1015951; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5952; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v13 offset:965953; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:975954; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5955; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v12 offset:925956; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:935957; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:2085958; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:2125959; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:2165960; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:2205961; ALIGNED-NEXT:    s_clause 0x35962; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:2205963; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:2165964; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:2125965; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:2085966; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:765967; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:745968; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:725969; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:705970; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:685971; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:665972; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:645973; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:625974; ALIGNED-NEXT:    s_waitcnt vmcnt(3)5975; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v11 offset:885976; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:895977; ALIGNED-NEXT:    s_waitcnt vmcnt(2)5978; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v10 offset:845979; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:855980; ALIGNED-NEXT:    s_waitcnt vmcnt(1)5981; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v9 offset:805982; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:815983; ALIGNED-NEXT:    s_waitcnt vmcnt(0)5984; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v8 offset:765985; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:775986; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:3045987; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:3085988; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:3125989; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:3165990; ALIGNED-NEXT:    s_clause 0x35991; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:3165992; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:3125993; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:3085994; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:3045995; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v265996; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v295997; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v295998; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v285999; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v286000; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v276001; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v276002; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v266003; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v336004; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v336005; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:366006; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v506007; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v506008; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v326009; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v326010; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v316011; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v316012; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v306013; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v306014; ALIGNED-NEXT:    flat_store_byte v[24:25], v114 offset:16015; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v366016; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v366017; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v356018; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v356019; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v496020; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v496021; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:406022; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v486023; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v486024; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:386025; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v396026; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v396027; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:446028; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v386029; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v386030; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:426031; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v536032; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v536033; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:326034; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v526035; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v526036; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:306037; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v516038; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v516039; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:346040; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v656041; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v656042; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:286043; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v646044; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v646045; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:266046; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v556047; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:2226048; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v86049; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v86050; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v556051; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:246052; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v546053; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v546054; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:226055; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v696056; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v696057; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:206058; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v686059; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v686060; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:186061; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v676062; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v676063; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:166064; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v666065; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v666066; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:146067; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v816068; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v816069; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:126070; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v806071; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v806072; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:106073; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 24, v716074; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v716075; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:86076; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v706077; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v706078; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:66079; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v856080; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v856081; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:46082; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v846083; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v846084; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:26085; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 24, v836086; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v836087; ALIGNED-NEXT:    flat_store_byte v[20:21], v346088; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v826089; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 8, v826090; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:2526091; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v976092; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v976093; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:2506094; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v966095; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v966096; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:2486097; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v876098; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v876099; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:2466100; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v866101; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v866102; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:2446103; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v196104; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v196105; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:2426106; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v186107; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v186108; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:2406109; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v176110; ALIGNED-NEXT:    v_lshrrev_b32_e32 v17, 8, v176111; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:2386112; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v166113; ALIGNED-NEXT:    v_lshrrev_b32_e32 v16, 8, v166114; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:2366115; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v156116; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v156117; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:2346118; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 24, v146119; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v146120; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:2326121; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v136122; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v136123; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:2306124; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v126125; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v126126; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:2286127; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v116128; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v116129; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:2266130; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v106131; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v106132; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:2246133; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v96134; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v96135; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:2206136; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:2186137; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:2166138; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:2146139; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:2126140; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:2106141; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:2086142; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:2066143; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:2046144; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:2026145; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:2006146; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:1986147; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:1966148; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:1946149; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:1926150; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:1906151; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:1886152; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:1866153; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:1846154; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:1826155; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:1806156; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:1786157; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:1766158; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:1746159; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:1726160; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:1706161; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:1686162; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:1666163; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:1646164; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:1626165; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:1606166; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:1586167; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:1566168; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:1546169; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:1526170; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:1506171; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:1486172; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:1466173; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:1446174; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:1426175; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:1246176; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:1226177; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:1206178; ALIGNED-NEXT:    flat_store_byte v[20:21], v18 offset:1186179; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:1166180; ALIGNED-NEXT:    flat_store_byte v[20:21], v17 offset:1146181; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:1126182; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:1106183; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:1086184; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:1066185; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:1046186; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:1026187; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:1006188; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:986189; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:966190; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:946191; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:926192; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:906193; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:886194; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:866195; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:846196; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:826197; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:806198; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:786199; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6200; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v7 offset:566201; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:576202; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6203; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v6 offset:526204; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:536205; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6206; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v5 offset:486207; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:496208; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6209; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v4 offset:446210; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:456211; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 24, v76212; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v76213; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 24, v66214; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v66215; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 24, v56216; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v56217; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 24, v46218; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v46219; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:606220; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:586221; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:566222; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:546223; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:526224; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:506225; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:486226; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:466227; ALIGNED-NEXT:    s_cbranch_scc1 .LBB5_26228; ALIGNED-NEXT:  .LBB5_3: ; %Flow56229; ALIGNED-NEXT:    s_andn2_saveexec_b32 s8, s66230; ALIGNED-NEXT:    s_cbranch_execz .LBB5_66231; ALIGNED-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader6232; ALIGNED-NEXT:    s_movk_i32 s6, 0xff006233; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x7006234; ALIGNED-NEXT:    s_mov_b32 s7, -16235; ALIGNED-NEXT:  .LBB5_5: ; %memmove_bwd_loop6236; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=16237; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, s46238; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, s5, v3, vcc_lo6239; ALIGNED-NEXT:    v_add_co_u32 v6, vcc_lo, v4, 486240; ALIGNED-NEXT:    flat_load_dwordx4 v[20:23], v[4:5] offset:1286241; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v5, vcc_lo6242; ALIGNED-NEXT:    v_add_co_u32 v8, vcc_lo, 0x50, v46243; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v5, vcc_lo6244; ALIGNED-NEXT:    v_add_co_u32 v12, vcc_lo, 0x60, v46245; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v5, vcc_lo6246; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, 0x70, v46247; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v5, vcc_lo6248; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, 0x90, v46249; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v5, vcc_lo6250; ALIGNED-NEXT:    v_add_co_u32 v52, vcc_lo, 0xa0, v46251; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v53, null, 0, v5, vcc_lo6252; ALIGNED-NEXT:    v_add_co_u32 v54, vcc_lo, 0xb0, v46253; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v55, null, 0, v5, vcc_lo6254; ALIGNED-NEXT:    v_add_co_u32 v64, vcc_lo, 0xc0, v46255; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v65, null, 0, v5, vcc_lo6256; ALIGNED-NEXT:    v_add_co_u32 v66, vcc_lo, 0xd0, v46257; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v67, null, 0, v5, vcc_lo6258; ALIGNED-NEXT:    v_add_co_u32 v68, vcc_lo, 0xe0, v46259; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v69, null, 0, v5, vcc_lo6260; ALIGNED-NEXT:    v_add_co_u32 v70, vcc_lo, 0xf0, v46261; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v71, null, 0, v5, vcc_lo6262; ALIGNED-NEXT:    s_clause 0xe6263; ALIGNED-NEXT:    flat_load_dwordx4 v[26:29], v[4:5]6264; ALIGNED-NEXT:    flat_load_dwordx4 v[35:38], v[4:5] offset:166265; ALIGNED-NEXT:    flat_load_dwordx4 v[30:33], v[4:5] offset:326266; ALIGNED-NEXT:    flat_load_dwordx4 v[48:51], v[4:5] offset:646267; ALIGNED-NEXT:    flat_load_dwordx4 v[4:7], v[6:7]6268; ALIGNED-NEXT:    flat_load_dwordx4 v[8:11], v[8:9]6269; ALIGNED-NEXT:    flat_load_dwordx4 v[12:15], v[12:13]6270; ALIGNED-NEXT:    flat_load_dwordx4 v[16:19], v[16:17]6271; ALIGNED-NEXT:    flat_load_dwordx4 v[112:115], v[24:25]6272; ALIGNED-NEXT:    flat_load_dwordx4 v[116:119], v[52:53]6273; ALIGNED-NEXT:    flat_load_dwordx4 v[40:43], v[54:55]6274; ALIGNED-NEXT:    flat_load_dwordx4 v[44:47], v[64:65]6275; ALIGNED-NEXT:    flat_load_dwordx4 v[55:58], v[66:67]6276; ALIGNED-NEXT:    flat_load_dwordx4 v[59:62], v[68:69]6277; ALIGNED-NEXT:    flat_load_dwordx4 v[68:71], v[70:71]6278; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v0, s46279; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v1, vcc_lo6280; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff006281; ALIGNED-NEXT:    s_addc_u32 s5, s5, -16282; ALIGNED-NEXT:    s_waitcnt vmcnt(15) lgkmcnt(15)6283; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:4166284; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:4206285; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:4246286; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:4286287; ALIGNED-NEXT:    s_clause 0x36288; ALIGNED-NEXT:    buffer_load_dword v102, off, s[0:3], s32 offset:4286289; ALIGNED-NEXT:    buffer_load_dword v103, off, s[0:3], s32 offset:4246290; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:4206291; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s32 offset:4166292; ALIGNED-NEXT:    v_add_co_u32 v22, vcc_lo, v24, 66293; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v25, vcc_lo6294; ALIGNED-NEXT:    v_add_co_u32 v20, vcc_lo, v24, 36295; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v25, vcc_lo6296; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]6297; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6298; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v102 offset:1366299; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:1376300; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6301; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v103 offset:1326302; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:1336303; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6304; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v63 offset:1286305; ALIGNED-NEXT:    flat_store_byte v[20:21], v63 offset:1296306; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6307; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v72 offset:1246308; ALIGNED-NEXT:    flat_store_byte v[24:25], v72 offset:1286309; ALIGNED-NEXT:    s_waitcnt lgkmcnt(19)6310; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:4806311; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:4846312; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:4886313; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:4926314; ALIGNED-NEXT:    s_clause 0x36315; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:4926316; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:4886317; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:4846318; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:4806319; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6320; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v101 offset:726321; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:736322; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6323; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v100 offset:686324; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:696325; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6326; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v99 offset:646327; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:656328; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6329; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v98 offset:606330; ALIGNED-NEXT:    flat_store_byte v[24:25], v98 offset:646331; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:5766332; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:5806333; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:5846334; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:5886335; ALIGNED-NEXT:    s_clause 0x36336; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:5846337; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:5886338; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:5766339; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:5806340; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6341; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v39 offset:366342; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:376343; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6344; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v34 offset:406345; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:416346; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6347; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v33 offset:286348; ALIGNED-NEXT:    flat_store_byte v[24:25], v33 offset:326349; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6350; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v32 offset:326351; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:336352; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:5286353; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:5326354; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:5366355; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:5406356; ALIGNED-NEXT:    s_clause 0x36357; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:5406358; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:5366359; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:5326360; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:5286361; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6362; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v67 offset:246363; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:256364; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6365; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v66 offset:206366; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:216367; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6368; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v53 offset:166369; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:176370; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6371; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v52 offset:126372; ALIGNED-NEXT:    flat_store_byte v[24:25], v52 offset:166373; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:5446374; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:5486375; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:5526376; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:5566377; ALIGNED-NEXT:    s_clause 0x36378; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:5566379; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:5526380; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:5486381; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:5446382; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6383; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v86 offset:86384; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:96385; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6386; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v82 offset:46387; ALIGNED-NEXT:    flat_store_byte v[24:25], v82 offset:86388; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6389; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v816390; ALIGNED-NEXT:    flat_store_byte v[24:25], v81 offset:46391; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6392; ALIGNED-NEXT:    flat_store_byte_d16_hi v[24:25], v80 offset:26393; ALIGNED-NEXT:    flat_store_byte v[24:25], v806394; ALIGNED-NEXT:    s_waitcnt lgkmcnt(40)6395; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:3686396; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:3726397; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:3766398; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:3806399; ALIGNED-NEXT:    s_clause 0x36400; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:3806401; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:3766402; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:3726403; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:3686404; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6405; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v97 offset:2486406; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:2496407; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6408; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v87 offset:2446409; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:2456410; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6411; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v83 offset:2406412; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:2416413; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6414; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v68 offset:2366415; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:2376416; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:3846417; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:3886418; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:3926419; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:3966420; ALIGNED-NEXT:    s_clause 0x36421; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:3966422; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:3926423; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:3886424; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:3846425; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6426; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v96 offset:2326427; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:2336428; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6429; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v84 offset:2286430; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:2296431; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6432; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v69 offset:2246433; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:2256434; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6435; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v54 offset:2206436; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:2216437; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:3366438; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:3406439; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:3446440; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:3486441; ALIGNED-NEXT:    s_clause 0x36442; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:3486443; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:3446444; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:3406445; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:3366446; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6447; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v85 offset:2166448; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:2176449; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6450; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v70 offset:2126451; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:2136452; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6453; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v55 offset:2086454; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:2096455; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6456; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v48 offset:2046457; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:2056458; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:3526459; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:3566460; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:3606461; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:3646462; ALIGNED-NEXT:    s_clause 0x36463; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:3646464; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:3606465; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:3566466; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:3526467; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6468; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v71 offset:2006469; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:2016470; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6471; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v64 offset:1966472; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:1976473; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6474; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v49 offset:1926475; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:1936476; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6477; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v35 offset:1886478; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:1896479; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:4326480; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:4366481; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:4406482; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:4446483; ALIGNED-NEXT:    s_clause 0x36484; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:4446485; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:4406486; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:4366487; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:4326488; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6489; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v65 offset:1846490; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:1856491; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6492; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v50 offset:1806493; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:1816494; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6495; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v36 offset:1766496; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:1776497; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6498; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v29 offset:1726499; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:1736500; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:4486501; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:4526502; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:4566503; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:4606504; ALIGNED-NEXT:    s_clause 0x36505; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:4606506; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:4566507; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:4526508; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:4486509; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6510; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v51 offset:1686511; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:1696512; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6513; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v37 offset:1646514; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:1656515; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6516; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v30 offset:1606517; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:1616518; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6519; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v27 offset:1566520; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:1576521; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:4006522; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:4046523; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:4086524; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:4126525; ALIGNED-NEXT:    s_clause 0x36526; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:4126527; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4086528; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:4046529; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:4006530; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v1026531; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v1026532; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v1036533; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v1036534; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v636535; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:1406536; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v636537; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:1386538; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v726539; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v726540; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:1366541; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:1346542; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:1326543; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:1306544; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:1286545; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:1266546; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6547; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v38 offset:1526548; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:1536549; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6550; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v31 offset:1486551; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:1496552; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6553; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v28 offset:1446554; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:1456555; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6556; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v26 offset:1406557; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:1416558; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:4966559; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:5006560; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:5046561; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:5086562; ALIGNED-NEXT:    s_clause 0x36563; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:5086564; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:5046565; ALIGNED-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:5006566; ALIGNED-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:4966567; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v1016568; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1016569; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v1006570; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v1006571; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v996572; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v996573; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v986574; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v986575; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v806576; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v866577; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v866578; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v806579; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6580; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v19 offset:1206581; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:1216582; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6583; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v18 offset:1166584; ALIGNED-NEXT:    flat_store_byte v[20:21], v18 offset:1176585; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6586; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v17 offset:1126587; ALIGNED-NEXT:    flat_store_byte v[20:21], v17 offset:1136588; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6589; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v16 offset:1086590; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:1096591; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:5126592; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:5166593; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:5206594; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:5246595; ALIGNED-NEXT:    s_clause 0x36596; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:5246597; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:5206598; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:5166599; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:5126600; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6601; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v15 offset:1046602; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:1056603; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6604; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v14 offset:1006605; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:1016606; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6607; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v13 offset:966608; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:976609; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6610; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v12 offset:926611; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:936612; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:4646613; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:4686614; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:4726615; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:4766616; ALIGNED-NEXT:    s_clause 0x36617; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:4766618; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:4726619; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:4686620; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:4646621; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:766622; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:746623; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:726624; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:706625; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:686626; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:666627; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:646628; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:626629; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6630; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v11 offset:886631; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:896632; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6633; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v10 offset:846634; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:856635; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6636; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v9 offset:806637; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:816638; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6639; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v8 offset:766640; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:776641; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:5606642; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:5646643; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:5686644; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:5726645; ALIGNED-NEXT:    s_clause 0x36646; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:5726647; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:5686648; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:5646649; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:5606650; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v326651; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v396652; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v396653; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v346654; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v346655; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v336656; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v336657; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v326658; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v676659; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v676660; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:366661; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v546662; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v546663; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v666664; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v666665; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v536666; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v536667; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v526668; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v526669; ALIGNED-NEXT:    flat_store_byte v[24:25], v114 offset:16670; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v826671; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v826672; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v816673; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v816674; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v976675; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v976676; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:406677; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v876678; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v876679; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:386680; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v836681; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v836682; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:446683; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v686684; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v686685; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:426686; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v966687; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v966688; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:326689; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v846690; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v846691; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:306692; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v696693; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v696694; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:346695; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v856696; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v856697; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:286698; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v706699; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v706700; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:266701; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 24, v556702; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:2226703; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v86704; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v86705; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v556706; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:246707; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v486708; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v486709; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:226710; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v716711; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v716712; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:206713; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v646714; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v646715; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:186716; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 24, v496717; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v496718; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:166719; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v356720; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v356721; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:146722; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v656723; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v656724; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:126725; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v506726; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v506727; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:106728; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v366729; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v366730; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:86731; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v296732; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v296733; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:66734; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v516735; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v516736; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:46737; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v376738; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v376739; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:26740; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 24, v306741; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v306742; ALIGNED-NEXT:    flat_store_byte v[20:21], v806743; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v276744; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v276745; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:2526746; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v386747; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v386748; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:2506749; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v316750; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v316751; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:2486752; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v286753; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v286754; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:2466755; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 24, v266756; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v266757; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:2446758; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v196759; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v196760; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:2426761; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 24, v186762; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v186763; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:2406764; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v176765; ALIGNED-NEXT:    v_lshrrev_b32_e32 v17, 8, v176766; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:2386767; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v166768; ALIGNED-NEXT:    v_lshrrev_b32_e32 v16, 8, v166769; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:2366770; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v156771; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v156772; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:2346773; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v146774; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v146775; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:2326776; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v136777; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v136778; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:2306779; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v126780; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v126781; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:2286782; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v116783; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v116784; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:2266785; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 24, v106786; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v106787; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:2246788; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v96789; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v96790; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:2206791; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:2186792; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:2166793; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:2146794; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:2126795; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:2106796; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:2086797; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:2066798; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:2046799; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:2026800; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:2006801; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:1986802; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:1966803; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:1946804; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:1926805; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:1906806; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:1886807; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:1866808; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:1846809; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:1826810; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:1806811; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:1786812; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:1766813; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:1746814; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:1726815; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:1706816; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:1686817; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:1666818; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:1646819; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:1626820; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:1606821; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:1586822; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:1566823; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:1546824; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:1526825; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:1506826; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:1486827; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:1466828; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:1446829; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:1426830; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:1246831; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:1226832; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:1206833; ALIGNED-NEXT:    flat_store_byte v[20:21], v18 offset:1186834; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:1166835; ALIGNED-NEXT:    flat_store_byte v[20:21], v17 offset:1146836; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:1126837; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:1106838; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:1086839; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:1066840; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:1046841; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:1026842; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:1006843; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:986844; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:966845; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:946846; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:926847; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:906848; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:886849; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:866850; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:846851; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:826852; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:806853; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:786854; ALIGNED-NEXT:    s_waitcnt vmcnt(3)6855; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v7 offset:566856; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:576857; ALIGNED-NEXT:    s_waitcnt vmcnt(2)6858; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v6 offset:526859; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:536860; ALIGNED-NEXT:    s_waitcnt vmcnt(1)6861; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v5 offset:486862; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:496863; ALIGNED-NEXT:    s_waitcnt vmcnt(0)6864; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v4 offset:446865; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:456866; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 24, v76867; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v76868; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 24, v66869; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v66870; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 24, v56871; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v56872; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 24, v46873; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v46874; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:606875; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:586876; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:566877; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:546878; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:526879; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:506880; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:486881; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:466882; ALIGNED-NEXT:    s_cbranch_scc0 .LBB5_56883; ALIGNED-NEXT:  .LBB5_6: ; %Flow66884; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s86885; ALIGNED-NEXT:    s_clause 0x10 ; 68-byte Folded Reload6886; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s326887; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:46888; ALIGNED-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:86889; ALIGNED-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:126890; ALIGNED-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:166891; ALIGNED-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:206892; ALIGNED-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:246893; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:286894; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:326895; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:366896; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:406897; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:446898; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:486899; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:526900; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:566901; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:606902; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:646903; ALIGNED-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)6904; ALIGNED-NEXT:    s_setpc_b64 s[30:31]6905;6906; UNROLL3-LABEL: memmove_p0_p0_sz2048:6907; UNROLL3:       ; %bb.0: ; %entry6908; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)6909; UNROLL3-NEXT:    s_mov_b32 s4, exec_lo6910; UNROLL3-NEXT:    v_cmpx_ge_u64_e64 v[2:3], v[0:1]6911; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s46912; UNROLL3-NEXT:    s_cbranch_execz .LBB5_46913; UNROLL3-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader6914; UNROLL3-NEXT:    s_mov_b64 s[4:5], 06915; UNROLL3-NEXT:    .p2align 66916; UNROLL3-NEXT:  .LBB5_2: ; %memmove_fwd_loop6917; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=16918; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s46919; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo6920; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s46921; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo6922; UNROLL3-NEXT:    s_clause 0x26923; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[12:13]6924; UNROLL3-NEXT:    flat_load_dwordx4 v[8:11], v[12:13] offset:166925; UNROLL3-NEXT:    flat_load_dwordx4 v[12:15], v[12:13] offset:326926; UNROLL3-NEXT:    s_add_u32 s4, s4, 486927; UNROLL3-NEXT:    s_addc_u32 s5, s5, 06928; UNROLL3-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)6929; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7]6930; UNROLL3-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(2)6931; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11] offset:166932; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(2)6933; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:326934; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0x7e06935; UNROLL3-NEXT:    s_cbranch_scc1 .LBB5_26936; UNROLL3-NEXT:  ; %bb.3: ; %memmove_fwd_residual6937; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[2:3] offset:20166938; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)6939; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:20166940; UNROLL3-NEXT:    flat_load_dwordx4 v[2:5], v[2:3] offset:20326941; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)6942; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[2:5] offset:20326943; UNROLL3-NEXT:    ; implicit-def: $vgpr2_vgpr36944; UNROLL3-NEXT:    ; implicit-def: $vgpr0_vgpr16945; UNROLL3-NEXT:  .LBB5_4: ; %Flow36946; UNROLL3-NEXT:    s_andn2_saveexec_b32 s8, s66947; UNROLL3-NEXT:    s_cbranch_execz .LBB5_76948; UNROLL3-NEXT:  ; %bb.5: ; %memmove_bwd_residual6949; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[2:3] offset:20326950; UNROLL3-NEXT:    s_movk_i32 s6, 0xffd06951; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7b06952; UNROLL3-NEXT:    s_mov_b32 s7, -16953; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)6954; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:20326955; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[2:3] offset:20166956; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)6957; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:20166958; UNROLL3-NEXT:    .p2align 66959; UNROLL3-NEXT:  .LBB5_6: ; %memmove_bwd_loop6960; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=16961; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s46962; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo6963; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s46964; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo6965; UNROLL3-NEXT:    s_clause 0x26966; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[12:13]6967; UNROLL3-NEXT:    flat_load_dwordx4 v[8:11], v[12:13] offset:166968; UNROLL3-NEXT:    flat_load_dwordx4 v[12:15], v[12:13] offset:326969; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd06970; UNROLL3-NEXT:    s_addc_u32 s5, s5, -16971; UNROLL3-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)6972; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7]6973; UNROLL3-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(2)6974; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11] offset:166975; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(2)6976; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:326977; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]6978; UNROLL3-NEXT:    s_cbranch_scc0 .LBB5_66979; UNROLL3-NEXT:  .LBB5_7: ; %Flow46980; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s86981; UNROLL3-NEXT:    s_waitcnt lgkmcnt(0)6982; UNROLL3-NEXT:    s_setpc_b64 s[30:31]6983entry:6984  tail call void @llvm.memmove.p0.p0.i64(ptr addrspace(0) noundef nonnull align 1 %dst, ptr addrspace(0) noundef nonnull align 1 %src, i64 2048, i1 false)6985  ret void6986}6987 6988define void @memmove_p1_p1_sz2048(ptr addrspace(1) align 1 %dst, ptr addrspace(1) align 1 readonly %src) {6989; CHECK-LABEL: memmove_p1_p1_sz2048:6990; CHECK:       ; %bb.0: ; %entry6991; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)6992; CHECK-NEXT:    s_mov_b32 s4, exec_lo6993; CHECK-NEXT:    v_cmpx_ge_u64_e64 v[2:3], v[0:1]6994; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s46995; CHECK-NEXT:    s_cbranch_execz .LBB6_36996; CHECK-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader6997; CHECK-NEXT:    s_mov_b64 s[4:5], 06998; CHECK-NEXT:  .LBB6_2: ; %memmove_fwd_loop6999; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=17000; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s47001; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo7002; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s47003; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo7004; CHECK-NEXT:    s_clause 0xf7005; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:2247006; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:2407007; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:1927008; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:2087009; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:1607010; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:1767011; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:1287012; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:1447013; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:967014; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:1127015; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:647016; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:807017; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:327018; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:487019; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off7020; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off offset:167021; CHECK-NEXT:    s_add_u32 s4, s4, 0x1007022; CHECK-NEXT:    s_addc_u32 s5, s5, 07023; CHECK-NEXT:    s_waitcnt vmcnt(15)7024; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[4:7], off offset:2247025; CHECK-NEXT:    s_waitcnt vmcnt(14)7026; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[8:11], off offset:2407027; CHECK-NEXT:    s_waitcnt vmcnt(13)7028; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[12:15], off offset:1927029; CHECK-NEXT:    s_waitcnt vmcnt(12)7030; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[16:19], off offset:2087031; CHECK-NEXT:    s_waitcnt vmcnt(11)7032; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[20:23], off offset:1607033; CHECK-NEXT:    s_waitcnt vmcnt(10)7034; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[24:27], off offset:1767035; CHECK-NEXT:    s_waitcnt vmcnt(9)7036; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[28:31], off offset:1287037; CHECK-NEXT:    s_waitcnt vmcnt(8)7038; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[32:35], off offset:1447039; CHECK-NEXT:    s_waitcnt vmcnt(7)7040; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[36:39], off offset:967041; CHECK-NEXT:    s_waitcnt vmcnt(6)7042; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[48:51], off offset:1127043; CHECK-NEXT:    s_waitcnt vmcnt(5)7044; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[52:55], off offset:647045; CHECK-NEXT:    s_waitcnt vmcnt(4)7046; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[64:67], off offset:807047; CHECK-NEXT:    s_waitcnt vmcnt(3)7048; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[68:71], off offset:327049; CHECK-NEXT:    s_waitcnt vmcnt(2)7050; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[80:83], off offset:487051; CHECK-NEXT:    s_waitcnt vmcnt(1)7052; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[84:87], off7053; CHECK-NEXT:    s_waitcnt vmcnt(0)7054; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[96:99], off offset:167055; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0x8007056; CHECK-NEXT:    s_cbranch_scc1 .LBB6_27057; CHECK-NEXT:  .LBB6_3: ; %Flow77058; CHECK-NEXT:    s_andn2_saveexec_b32 s8, s67059; CHECK-NEXT:    s_cbranch_execz .LBB6_67060; CHECK-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader7061; CHECK-NEXT:    s_movk_i32 s6, 0xff007062; CHECK-NEXT:    s_mov_b64 s[4:5], 0x7007063; CHECK-NEXT:    s_mov_b32 s7, -17064; CHECK-NEXT:  .LBB6_5: ; %memmove_bwd_loop7065; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=17066; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s47067; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo7068; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s47069; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo7070; CHECK-NEXT:    s_clause 0xf7071; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:2247072; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:2407073; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:1927074; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:2087075; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:1607076; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:1767077; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:1287078; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:1447079; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:967080; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:1127081; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:647082; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:807083; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:327084; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:487085; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off7086; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off offset:167087; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff007088; CHECK-NEXT:    s_addc_u32 s5, s5, -17089; CHECK-NEXT:    s_waitcnt vmcnt(15)7090; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[4:7], off offset:2247091; CHECK-NEXT:    s_waitcnt vmcnt(14)7092; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[8:11], off offset:2407093; CHECK-NEXT:    s_waitcnt vmcnt(13)7094; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[12:15], off offset:1927095; CHECK-NEXT:    s_waitcnt vmcnt(12)7096; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[16:19], off offset:2087097; CHECK-NEXT:    s_waitcnt vmcnt(11)7098; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[20:23], off offset:1607099; CHECK-NEXT:    s_waitcnt vmcnt(10)7100; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[24:27], off offset:1767101; CHECK-NEXT:    s_waitcnt vmcnt(9)7102; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[28:31], off offset:1287103; CHECK-NEXT:    s_waitcnt vmcnt(8)7104; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[32:35], off offset:1447105; CHECK-NEXT:    s_waitcnt vmcnt(7)7106; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[36:39], off offset:967107; CHECK-NEXT:    s_waitcnt vmcnt(6)7108; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[48:51], off offset:1127109; CHECK-NEXT:    s_waitcnt vmcnt(5)7110; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[52:55], off offset:647111; CHECK-NEXT:    s_waitcnt vmcnt(4)7112; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[64:67], off offset:807113; CHECK-NEXT:    s_waitcnt vmcnt(3)7114; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[68:71], off offset:327115; CHECK-NEXT:    s_waitcnt vmcnt(2)7116; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[80:83], off offset:487117; CHECK-NEXT:    s_waitcnt vmcnt(1)7118; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[84:87], off7119; CHECK-NEXT:    s_waitcnt vmcnt(0)7120; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[96:99], off offset:167121; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]7122; CHECK-NEXT:    s_cbranch_scc0 .LBB6_57123; CHECK-NEXT:  .LBB6_6: ; %Flow87124; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s87125; CHECK-NEXT:    s_setpc_b64 s[30:31]7126;7127; ALIGNED-LABEL: memmove_p1_p1_sz2048:7128; ALIGNED:       ; %bb.0: ; %entry7129; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)7130; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill7131; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill7132; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill7133; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill7134; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill7135; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill7136; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill7137; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill7138; ALIGNED-NEXT:    s_mov_b32 s4, exec_lo7139; ALIGNED-NEXT:    v_cmpx_ge_u64_e64 v[2:3], v[0:1]7140; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s47141; ALIGNED-NEXT:    s_cbranch_execz .LBB6_37142; ALIGNED-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader7143; ALIGNED-NEXT:    s_mov_b64 s[4:5], 07144; ALIGNED-NEXT:  .LBB6_2: ; %memmove_fwd_loop7145; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=17146; ALIGNED-NEXT:    v_add_co_u32 v20, vcc_lo, v2, s47147; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v21, null, s5, v3, vcc_lo7148; ALIGNED-NEXT:    s_clause 0xf7149; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[20:21], off offset:2407150; ALIGNED-NEXT:    global_load_dwordx4 v[22:25], v[20:21], off offset:2247151; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[20:21], off7152; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[20:21], off offset:167153; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[20:21], off offset:327154; ALIGNED-NEXT:    global_load_dwordx4 v[98:101], v[20:21], off offset:487155; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[20:21], off offset:647156; ALIGNED-NEXT:    global_load_dwordx4 v[82:85], v[20:21], off offset:807157; ALIGNED-NEXT:    global_load_dwordx4 v[116:119], v[20:21], off offset:967158; ALIGNED-NEXT:    global_load_dwordx4 v[66:69], v[20:21], off offset:1127159; ALIGNED-NEXT:    global_load_dwordx4 v[40:43], v[20:21], off offset:1287160; ALIGNED-NEXT:    global_load_dwordx4 v[50:53], v[20:21], off offset:1447161; ALIGNED-NEXT:    global_load_dwordx4 v[44:47], v[20:21], off offset:1607162; ALIGNED-NEXT:    global_load_dwordx4 v[34:37], v[20:21], off offset:1767163; ALIGNED-NEXT:    global_load_dwordx4 v[30:33], v[20:21], off offset:1927164; ALIGNED-NEXT:    global_load_dwordx4 v[26:29], v[20:21], off offset:2087165; ALIGNED-NEXT:    s_waitcnt vmcnt(15)7166; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:647167; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:687168; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:727169; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:767170; ALIGNED-NEXT:    s_clause 0x37171; ALIGNED-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:767172; ALIGNED-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:727173; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:687174; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:647175; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s47176; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo7177; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x1007178; ALIGNED-NEXT:    s_addc_u32 s5, s5, 07179; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7180; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v21, off offset:2547181; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:2527182; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7183; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v20, off offset:2507184; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:2487185; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7186; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v19, off offset:2467187; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:2447188; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7189; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v18, off offset:2427190; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:2407191; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:807192; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:847193; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:887194; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:927195; ALIGNED-NEXT:    s_clause 0x37196; ALIGNED-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:927197; ALIGNED-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:887198; ALIGNED-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:847199; ALIGNED-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:807200; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v217201; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 8, v217202; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v207203; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 8, v207204; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0x8007205; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7206; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v25, off offset:2387207; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:2367208; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7209; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v24, off offset:2347210; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:2327211; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7212; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v23, off offset:2307213; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:2287214; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7215; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v22, off offset:2267216; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:2247217; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:327218; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:367219; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:407220; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:447221; ALIGNED-NEXT:    s_clause 0x37222; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:447223; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:407224; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:367225; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:327226; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7227; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v29, off offset:2227228; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:2207229; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7230; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v28, off offset:2187231; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:2167232; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7233; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v27, off offset:2147234; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:2127235; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7236; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v26, off offset:2107237; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:2087238; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:487239; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:527240; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:567241; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:607242; ALIGNED-NEXT:    s_clause 0x37243; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:607244; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:567245; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:527246; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:487247; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7248; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v33, off offset:2067249; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:2047250; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7251; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v32, off offset:2027252; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:2007253; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7254; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v31, off offset:1987255; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:1967256; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7257; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v30, off offset:1947258; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:1927259; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:1287260; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:1327261; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:1367262; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:1407263; ALIGNED-NEXT:    s_clause 0x37264; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:1407265; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:1367266; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:1327267; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:1287268; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7269; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v37, off offset:1907270; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:1887271; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7272; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v36, off offset:1867273; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:1847274; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7275; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v35, off offset:1827276; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:1807277; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7278; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v34, off offset:1787279; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:1767280; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:1447281; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:1487282; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:1527283; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:1567284; ALIGNED-NEXT:    s_clause 0x37285; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:1567286; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:1527287; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:1487288; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:1447289; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7290; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v49, off offset:1747291; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:1727292; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7293; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v48, off offset:1707294; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:1687295; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7296; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v39, off offset:1667297; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:1647298; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7299; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v38, off offset:1627300; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:1607301; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:967302; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:1007303; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:1047304; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:1087305; ALIGNED-NEXT:    s_clause 0x37306; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:1087307; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:1047308; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:1007309; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:967310; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7311; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v53, off offset:1587312; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:1567313; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7314; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v52, off offset:1547315; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:1527316; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7317; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v51, off offset:1507318; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:1487319; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7320; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v50, off offset:1467321; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:1447322; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:1127323; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:1167324; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:1207325; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:1247326; ALIGNED-NEXT:    s_clause 0x37327; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:1247328; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:1207329; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:1167330; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:1127331; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7332; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v65, off offset:1427333; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:1407334; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7335; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v64, off offset:1387336; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:1367337; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7338; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v55, off offset:1347339; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:1327340; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7341; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v54, off offset:1307342; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:1287343; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:1927344; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:1967345; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:2007346; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:2047347; ALIGNED-NEXT:    s_clause 0x37348; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:2047349; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:2007350; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:1967351; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:1927352; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7353; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v69, off offset:1267354; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:1247355; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7356; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v68, off offset:1227357; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:1207358; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7359; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v67, off offset:1187360; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:1167361; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7362; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v66, off offset:1147363; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:1127364; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:2087365; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:2127366; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:2167367; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:2207368; ALIGNED-NEXT:    s_clause 0x37369; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:2207370; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:2167371; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:2127372; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:2087373; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7374; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v81, off offset:1107375; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:1087376; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7377; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v80, off offset:1067378; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:1047379; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7380; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v71, off offset:1027381; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:1007382; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7383; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v70, off offset:987384; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:967385; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:1607386; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:1647387; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:1687388; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:1727389; ALIGNED-NEXT:    s_clause 0x37390; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:1727391; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:1687392; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:1647393; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:1607394; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7395; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v85, off offset:947396; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:927397; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7398; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v84, off offset:907399; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:887400; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7401; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v83, off offset:867402; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:847403; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7404; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v82, off offset:827405; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:807406; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:1767407; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:1807408; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:1847409; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:1887410; ALIGNED-NEXT:    s_clause 0x37411; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:1887412; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:1847413; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:1807414; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:1767415; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v197416; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v197417; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v187418; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v187419; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v257420; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v257421; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v247422; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 8, v247423; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7424; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v97, off offset:787425; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:767426; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7427; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v96, off offset:747428; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:727429; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7430; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v87, off offset:707431; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:687432; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7433; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v86, off offset:667434; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:647435; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:2567436; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:2607437; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:2647438; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:2687439; ALIGNED-NEXT:    s_clause 0x37440; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:2687441; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:2647442; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:2607443; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:2567444; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7445; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v101, off offset:627446; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:607447; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7448; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v100, off offset:587449; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:567450; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7451; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v99, off offset:547452; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:527453; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7454; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v98, off offset:507455; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:487456; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:2727457; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:2767458; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:2807459; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:2847460; ALIGNED-NEXT:    s_clause 0x37461; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:2807462; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:2847463; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:2727464; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:2767465; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7466; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v15, off offset:427467; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:407468; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7469; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v14, off offset:467470; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:447471; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7472; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v13, off offset:347473; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:327474; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7475; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v12, off offset:387476; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:367477; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:2247478; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:2287479; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:2327480; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:2367481; ALIGNED-NEXT:    s_clause 0x37482; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:2367483; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:2327484; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:2287485; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:2247486; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7487; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v11, off offset:307488; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:287489; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7490; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v10, off offset:267491; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:247492; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7493; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v9, off offset:227494; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:207495; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7496; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v8, off offset:187497; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:167498; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:2407499; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:2447500; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:2487501; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:2527502; ALIGNED-NEXT:    s_clause 0x37503; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:2527504; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:2487505; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:2447506; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:2407507; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:2477508; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v277509; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:2557510; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v237511; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 8, v237512; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:2537513; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 24, v227514; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 8, v227515; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:2517516; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v297517; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v297518; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:2497519; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 24, v287520; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v287521; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v277522; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:2457523; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 24, v267524; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v267525; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:2157526; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v517527; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v517528; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:2437529; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v337530; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v337531; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:2417532; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 24, v327533; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v327534; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:2397535; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v317536; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v317537; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:2377538; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v307539; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v307540; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:2357541; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v377542; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v377543; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:2337544; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v367545; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v367546; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:2317547; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v357548; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v357549; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:2297550; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 24, v347551; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v347552; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:2277553; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 24, v497554; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v497555; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:2257556; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 24, v487557; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v487558; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:2237559; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v397560; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v397561; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:2217562; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v387563; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v387564; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:2197565; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 24, v537566; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v537567; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:2177568; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v527569; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v527570; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:2137571; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v507572; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v507573; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:2117574; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 24, v657575; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v657576; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:2097577; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v647578; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:1497579; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v87580; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v87581; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v647582; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:2077583; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v557584; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v557585; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:2057586; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v547587; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v547588; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:2037589; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 24, v697590; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v697591; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:2017592; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v687593; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v687594; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:1997595; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v677596; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v677597; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:1977598; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v667599; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v667600; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:1957601; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v817602; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v817603; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:1937604; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v807605; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v807606; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:1917607; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v717608; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v717609; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:1897610; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 24, v707611; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v707612; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:1877613; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v857614; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v857615; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:1857616; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v847617; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v847618; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:1837619; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v837620; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v837621; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:1817622; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 24, v827623; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 8, v827624; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:1797625; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 24, v977626; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v977627; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:1777628; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v967629; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v967630; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:1757631; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 24, v877632; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v877633; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:1737634; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v867635; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v867636; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:1717637; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 24, v1017638; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1017639; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:1697640; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v1007641; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v1007642; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:1677643; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v997644; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v997645; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:1657646; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v987647; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v987648; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:1637649; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v157650; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v157651; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:1617652; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v147653; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v147654; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:1597655; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 24, v137656; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v137657; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:1577658; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 24, v127659; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v127660; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:1557661; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v117662; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v117663; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:1537664; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v107665; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v107666; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:1517667; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v97668; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v97669; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:1477670; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:1457671; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:1437672; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:1417673; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:1397674; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:1377675; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:1357676; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:1337677; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:1317678; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:1297679; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:1277680; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:1257681; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:1237682; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:1217683; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:1197684; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:1177685; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:1157686; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:1137687; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:1117688; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:1097689; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:1077690; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:1057691; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:1037692; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:1017693; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:997694; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:977695; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:957696; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:937697; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:917698; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:897699; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:877700; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:857701; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:837702; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:817703; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:797704; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:777705; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:757706; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:737707; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:717708; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:697709; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:677710; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:657711; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:637712; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:617713; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:597714; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:577715; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:557716; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:537717; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:517718; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:497719; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:437720; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:417721; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:477722; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:457723; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:357724; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:337725; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:397726; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:377727; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:317728; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:297729; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:277730; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:257731; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:237732; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:217733; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:197734; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:177735; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7736; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v7, off offset:147737; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:127738; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7739; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v6, off offset:107740; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:87741; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7742; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v5, off offset:67743; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:47744; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7745; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v4, off offset:27746; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off7747; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 24, v77748; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v77749; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 24, v67750; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v67751; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 24, v57752; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v57753; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 24, v47754; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v47755; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:157756; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:137757; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:117758; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:97759; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:77760; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:57761; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:37762; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off offset:17763; ALIGNED-NEXT:    s_cbranch_scc1 .LBB6_27764; ALIGNED-NEXT:  .LBB6_3: ; %Flow77765; ALIGNED-NEXT:    s_andn2_saveexec_b32 s8, s67766; ALIGNED-NEXT:    s_cbranch_execz .LBB6_67767; ALIGNED-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader7768; ALIGNED-NEXT:    s_movk_i32 s6, 0xff007769; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x7007770; ALIGNED-NEXT:    s_mov_b32 s7, -17771; ALIGNED-NEXT:  .LBB6_5: ; %memmove_bwd_loop7772; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=17773; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v2, s47774; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v3, vcc_lo7775; ALIGNED-NEXT:    s_clause 0xf7776; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[24:25], off offset:2407777; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[24:25], off offset:2247778; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[24:25], off7779; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[24:25], off offset:167780; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[24:25], off offset:327781; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[24:25], off offset:487782; ALIGNED-NEXT:    global_load_dwordx4 v[116:119], v[24:25], off offset:647783; ALIGNED-NEXT:    global_load_dwordx4 v[40:43], v[24:25], off offset:807784; ALIGNED-NEXT:    global_load_dwordx4 v[26:29], v[24:25], off offset:967785; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[24:25], off offset:1127786; ALIGNED-NEXT:    global_load_dwordx4 v[44:47], v[24:25], off offset:1287787; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[24:25], off offset:1447788; ALIGNED-NEXT:    global_load_dwordx4 v[66:69], v[24:25], off offset:1607789; ALIGNED-NEXT:    global_load_dwordx4 v[81:84], v[24:25], off offset:1767790; ALIGNED-NEXT:    global_load_dwordx4 v[96:99], v[24:25], off offset:1927791; ALIGNED-NEXT:    global_load_dwordx4 v[100:103], v[24:25], off offset:2087792; ALIGNED-NEXT:    s_waitcnt vmcnt(15)7793; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:3207794; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:3247795; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:3287796; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:3327797; ALIGNED-NEXT:    s_clause 0x37798; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:3327799; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:3287800; ALIGNED-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:3247801; ALIGNED-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:3207802; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s47803; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo7804; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff007805; ALIGNED-NEXT:    s_addc_u32 s5, s5, -17806; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7807; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v31, off offset:2547808; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:2527809; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7810; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v30, off offset:2507811; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:2487812; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7813; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v25, off offset:2467814; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:2447815; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7816; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v24, off offset:2427817; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:2407818; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:3367819; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:3407820; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:3447821; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:3487822; ALIGNED-NEXT:    s_clause 0x37823; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:3487824; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:3447825; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:3407826; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:3367827; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]7828; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7829; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v51, off offset:2387830; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:2367831; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7832; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v50, off offset:2347833; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:2327834; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7835; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v49, off offset:2307836; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:2287837; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7838; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v36, off offset:2267839; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:2247840; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:2887841; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:2927842; ALIGNED-NEXT:    buffer_store_dword v102, off, s[0:3], s32 offset:2967843; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:3007844; ALIGNED-NEXT:    s_clause 0x37845; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:3007846; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:2967847; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:2927848; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:2887849; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v317850; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v317851; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v307852; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v307853; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7854; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v71, off offset:2227855; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:2207856; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7857; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v70, off offset:2187858; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:2167859; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7860; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v65, off offset:2147861; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:2127862; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7863; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v64, off offset:2107864; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:2087865; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:3047866; ALIGNED-NEXT:    buffer_store_dword v97, off, s[0:3], s32 offset:3087867; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:3127868; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:3167869; ALIGNED-NEXT:    s_clause 0x37870; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:3167871; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:3127872; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:3087873; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:3047874; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7875; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v87, off offset:2067876; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:2047877; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7878; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v86, off offset:2027879; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:2007880; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7881; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v85, off offset:1987882; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:1967883; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7884; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v80, off offset:1947885; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:1927886; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:3847887; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:3887888; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:3927889; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:3967890; ALIGNED-NEXT:    s_clause 0x37891; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:3967892; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:3927893; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:3887894; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:3847895; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7896; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v101, off offset:1907897; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:1887898; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7899; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v99, off offset:1867900; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:1847901; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7902; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v96, off offset:1827903; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:1807904; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7905; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v81, off offset:1787906; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:1767907; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:4007908; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:4047909; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:4087910; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:4127911; ALIGNED-NEXT:    s_clause 0x37912; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:4127913; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:4087914; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:4047915; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:4007916; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7917; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v100, off offset:1747918; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:1727919; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7920; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v97, off offset:1707921; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:1687922; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7923; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v82, off offset:1667924; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:1647925; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7926; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v66, off offset:1627927; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:1607928; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:3527929; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:3567930; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:3607931; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:3647932; ALIGNED-NEXT:    s_clause 0x37933; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:3647934; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:3607935; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:3567936; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:3527937; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7938; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v98, off offset:1587939; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:1567940; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7941; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v83, off offset:1547942; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:1527943; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7944; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v67, off offset:1507945; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:1487946; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7947; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v52, off offset:1467948; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:1447949; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:3687950; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:3727951; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:3767952; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:3807953; ALIGNED-NEXT:    s_clause 0x37954; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:3807955; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:3767956; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:3727957; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:3687958; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7959; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v84, off offset:1427960; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:1407961; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7962; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v68, off offset:1387963; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:1367964; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7965; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v53, off offset:1347966; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:1327967; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7968; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v37, off offset:1307969; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:1287970; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:4487971; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:4527972; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:4567973; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:4607974; ALIGNED-NEXT:    s_clause 0x37975; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:4607976; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:4567977; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:4527978; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4487979; ALIGNED-NEXT:    s_waitcnt vmcnt(3)7980; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v69, off offset:1267981; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:1247982; ALIGNED-NEXT:    s_waitcnt vmcnt(2)7983; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v54, off offset:1227984; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:1207985; ALIGNED-NEXT:    s_waitcnt vmcnt(1)7986; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v38, off offset:1187987; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:1167988; ALIGNED-NEXT:    s_waitcnt vmcnt(0)7989; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v32, off offset:1147990; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:1127991; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:4647992; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:4687993; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:4727994; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:4767995; ALIGNED-NEXT:    s_clause 0x37996; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:4767997; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:4727998; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4687999; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:4648000; ALIGNED-NEXT:    s_waitcnt vmcnt(3)8001; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v55, off offset:1108002; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:1088003; ALIGNED-NEXT:    s_waitcnt vmcnt(2)8004; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v39, off offset:1068005; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:1048006; ALIGNED-NEXT:    s_waitcnt vmcnt(1)8007; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v33, off offset:1028008; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:1008009; ALIGNED-NEXT:    s_waitcnt vmcnt(0)8010; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v26, off offset:988011; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:968012; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:4168013; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:4208014; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:4248015; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:4288016; ALIGNED-NEXT:    s_clause 0x38017; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:4288018; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:4248019; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:4208020; ALIGNED-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:4168021; ALIGNED-NEXT:    s_waitcnt vmcnt(3)8022; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v48, off offset:948023; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:928024; ALIGNED-NEXT:    s_waitcnt vmcnt(2)8025; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v34, off offset:908026; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:888027; ALIGNED-NEXT:    s_waitcnt vmcnt(1)8028; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v27, off offset:868029; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:848030; ALIGNED-NEXT:    s_waitcnt vmcnt(0)8031; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v21, off offset:828032; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:808033; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:4328034; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:4368035; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:4408036; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:4448037; ALIGNED-NEXT:    s_clause 0x38038; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4448039; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:4408040; ALIGNED-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:4368041; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:4328042; ALIGNED-NEXT:    s_waitcnt vmcnt(3)8043; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v35, off offset:788044; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:768045; ALIGNED-NEXT:    s_waitcnt vmcnt(2)8046; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v28, off offset:748047; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:728048; ALIGNED-NEXT:    s_waitcnt vmcnt(1)8049; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v22, off offset:708050; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:688051; ALIGNED-NEXT:    s_waitcnt vmcnt(0)8052; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v19, off offset:668053; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:648054; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:5128055; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:5168056; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:5208057; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:5248058; ALIGNED-NEXT:    s_clause 0x38059; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:5248060; ALIGNED-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:5208061; ALIGNED-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:5168062; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:5128063; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v258064; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v258065; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v248066; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 8, v248067; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v518068; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v518069; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v508070; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v508071; ALIGNED-NEXT:    s_waitcnt vmcnt(3)8072; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v29, off offset:628073; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:608074; ALIGNED-NEXT:    s_waitcnt vmcnt(2)8075; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v23, off offset:588076; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:568077; ALIGNED-NEXT:    s_waitcnt vmcnt(1)8078; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v20, off offset:548079; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:528080; ALIGNED-NEXT:    s_waitcnt vmcnt(0)8081; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v18, off offset:508082; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:488083; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:5288084; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:5328085; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:5368086; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:5408087; ALIGNED-NEXT:    s_clause 0x38088; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:5368089; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:5408090; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:5288091; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:5328092; ALIGNED-NEXT:    s_waitcnt vmcnt(3)8093; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v15, off offset:428094; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:408095; ALIGNED-NEXT:    s_waitcnt vmcnt(2)8096; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v14, off offset:468097; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:448098; ALIGNED-NEXT:    s_waitcnt vmcnt(1)8099; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v13, off offset:348100; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:328101; ALIGNED-NEXT:    s_waitcnt vmcnt(0)8102; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v12, off offset:388103; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:368104; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:4808105; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:4848106; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:4888107; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:4928108; ALIGNED-NEXT:    s_clause 0x38109; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:4928110; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:4888111; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:4848112; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:4808113; ALIGNED-NEXT:    s_waitcnt vmcnt(3)8114; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v11, off offset:308115; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:288116; ALIGNED-NEXT:    s_waitcnt vmcnt(2)8117; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v10, off offset:268118; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:248119; ALIGNED-NEXT:    s_waitcnt vmcnt(1)8120; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v9, off offset:228121; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:208122; ALIGNED-NEXT:    s_waitcnt vmcnt(0)8123; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v8, off offset:188124; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:168125; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:4968126; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:5008127; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:5048128; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:5088129; ALIGNED-NEXT:    s_clause 0x38130; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:5088131; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:5048132; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:5008133; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:4968134; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:2478135; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v658136; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:2558137; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v498138; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v498139; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:2538140; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v368141; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v368142; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:2518143; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v718144; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v718145; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:2498146; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v708147; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v708148; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v658149; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:2458150; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v648151; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v648152; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:2158153; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v678154; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v678155; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:2438156; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v878157; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v878158; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:2418159; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v868160; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v868161; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:2398162; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v858163; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v858164; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:2378165; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v808166; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v808167; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:2358168; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v1018169; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1018170; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:2338171; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v998172; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v998173; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:2318174; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v968175; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v968176; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:2298177; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v818178; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v818179; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:2278180; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v1008181; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v1008182; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:2258183; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v978184; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v978185; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:2238186; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v828187; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 8, v828188; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:2218189; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 24, v668190; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v668191; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:2198192; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v988193; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v988194; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:2178195; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v838196; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v838197; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:2138198; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 24, v528199; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v528200; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:2118201; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v848202; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v848203; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:2098204; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v688205; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:1498206; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 24, v88207; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v88208; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v688209; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:2078210; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v538211; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v538212; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:2058213; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 24, v378214; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v378215; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:2038216; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v698217; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v698218; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:2018219; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v548220; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v548221; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:1998222; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v388223; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v388224; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:1978225; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 24, v328226; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v328227; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:1958228; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v558229; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v558230; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:1938231; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v398232; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v398233; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:1918234; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v338235; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v338236; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:1898237; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v268238; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v268239; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:1878240; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v488241; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v488242; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:1858243; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v348244; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v348245; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:1838246; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v278247; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v278248; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:1818249; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v218250; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 8, v218251; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:1798252; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v358253; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v358254; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:1778255; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 24, v288256; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v288257; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:1758258; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v228259; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 8, v228260; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:1738261; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v198262; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v198263; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:1718264; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v298265; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v298266; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:1698267; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v238268; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 8, v238269; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:1678270; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v208271; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 8, v208272; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:1658273; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v188274; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v188275; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:1638276; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 24, v158277; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v158278; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:1618279; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v148280; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v148281; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:1598282; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v138283; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v138284; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:1578285; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v128286; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v128287; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:1558288; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v118289; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v118290; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:1538291; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 24, v108292; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v108293; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:1518294; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v98295; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v98296; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:1478297; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:1458298; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:1438299; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:1418300; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:1398301; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:1378302; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:1358303; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:1338304; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:1318305; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:1298306; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:1278307; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:1258308; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:1238309; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:1218310; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:1198311; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:1178312; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:1158313; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:1138314; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:1118315; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:1098316; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:1078317; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:1058318; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:1038319; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:1018320; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:998321; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:978322; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:958323; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:938324; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:918325; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:898326; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:878327; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:858328; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:838329; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:818330; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:798331; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:778332; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:758333; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:738334; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:718335; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:698336; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:678337; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:658338; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:638339; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:618340; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:598341; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:578342; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:558343; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:538344; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:518345; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:498346; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:438347; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:418348; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:478349; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:458350; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:358351; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:338352; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:398353; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:378354; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:318355; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:298356; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:278357; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:258358; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:238359; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:218360; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:198361; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:178362; ALIGNED-NEXT:    s_waitcnt vmcnt(3)8363; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v7, off offset:148364; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:128365; ALIGNED-NEXT:    s_waitcnt vmcnt(2)8366; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v6, off offset:108367; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:88368; ALIGNED-NEXT:    s_waitcnt vmcnt(1)8369; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v5, off offset:68370; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:48371; ALIGNED-NEXT:    s_waitcnt vmcnt(0)8372; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v4, off offset:28373; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off8374; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 24, v78375; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v78376; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 24, v68377; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v68378; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 24, v58379; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v58380; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 24, v48381; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v48382; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:158383; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:138384; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:118385; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:98386; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:78387; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:58388; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:38389; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off offset:18390; ALIGNED-NEXT:    s_cbranch_scc0 .LBB6_58391; ALIGNED-NEXT:  .LBB6_6: ; %Flow88392; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s88393; ALIGNED-NEXT:    s_clause 0x7 ; 32-byte Folded Reload8394; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s328395; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48396; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:88397; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:128398; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:168399; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:208400; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:248401; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:288402; ALIGNED-NEXT:    s_waitcnt vmcnt(0)8403; ALIGNED-NEXT:    s_setpc_b64 s[30:31]8404;8405; UNROLL3-LABEL: memmove_p1_p1_sz2048:8406; UNROLL3:       ; %bb.0: ; %entry8407; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)8408; UNROLL3-NEXT:    s_mov_b32 s4, exec_lo8409; UNROLL3-NEXT:    v_cmpx_ge_u64_e64 v[2:3], v[0:1]8410; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s48411; UNROLL3-NEXT:    s_cbranch_execz .LBB6_48412; UNROLL3-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader8413; UNROLL3-NEXT:    s_mov_b64 s[4:5], 08414; UNROLL3-NEXT:    .p2align 68415; UNROLL3-NEXT:  .LBB6_2: ; %memmove_fwd_loop8416; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=18417; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s48418; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo8419; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s48420; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo8421; UNROLL3-NEXT:    s_clause 0x28422; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off8423; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off offset:168424; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:328425; UNROLL3-NEXT:    s_add_u32 s4, s4, 488426; UNROLL3-NEXT:    s_addc_u32 s5, s5, 08427; UNROLL3-NEXT:    s_waitcnt vmcnt(2)8428; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[4:7], off8429; UNROLL3-NEXT:    s_waitcnt vmcnt(1)8430; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[8:11], off offset:168431; UNROLL3-NEXT:    s_waitcnt vmcnt(0)8432; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[12:15], off offset:328433; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0x7e08434; UNROLL3-NEXT:    s_cbranch_scc1 .LBB6_28435; UNROLL3-NEXT:  ; %bb.3: ; %memmove_fwd_residual8436; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:20168437; UNROLL3-NEXT:    s_waitcnt vmcnt(0)8438; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[4:7], off offset:20168439; UNROLL3-NEXT:    global_load_dwordx4 v[2:5], v[2:3], off offset:20328440; UNROLL3-NEXT:    s_waitcnt vmcnt(0)8441; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:20328442; UNROLL3-NEXT:    ; implicit-def: $vgpr2_vgpr38443; UNROLL3-NEXT:    ; implicit-def: $vgpr0_vgpr18444; UNROLL3-NEXT:  .LBB6_4: ; %Flow58445; UNROLL3-NEXT:    s_andn2_saveexec_b32 s8, s68446; UNROLL3-NEXT:    s_cbranch_execz .LBB6_78447; UNROLL3-NEXT:  ; %bb.5: ; %memmove_bwd_residual8448; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:20328449; UNROLL3-NEXT:    s_movk_i32 s6, 0xffd08450; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7b08451; UNROLL3-NEXT:    s_mov_b32 s7, -18452; UNROLL3-NEXT:    s_waitcnt vmcnt(0)8453; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[4:7], off offset:20328454; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:20168455; UNROLL3-NEXT:    s_waitcnt vmcnt(0)8456; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[4:7], off offset:20168457; UNROLL3-NEXT:    .p2align 68458; UNROLL3-NEXT:  .LBB6_6: ; %memmove_bwd_loop8459; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=18460; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s48461; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo8462; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s48463; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo8464; UNROLL3-NEXT:    s_clause 0x28465; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off8466; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off offset:168467; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:328468; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd08469; UNROLL3-NEXT:    s_addc_u32 s5, s5, -18470; UNROLL3-NEXT:    s_waitcnt vmcnt(2)8471; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[4:7], off8472; UNROLL3-NEXT:    s_waitcnt vmcnt(1)8473; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[8:11], off offset:168474; UNROLL3-NEXT:    s_waitcnt vmcnt(0)8475; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[12:15], off offset:328476; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]8477; UNROLL3-NEXT:    s_cbranch_scc0 .LBB6_68478; UNROLL3-NEXT:  .LBB6_7: ; %Flow68479; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s88480; UNROLL3-NEXT:    s_setpc_b64 s[30:31]8481entry:8482  tail call void @llvm.memmove.p1.p1.i64(ptr addrspace(1) noundef nonnull align 1 %dst, ptr addrspace(1) noundef nonnull align 1 %src, i64 2048, i1 false)8483  ret void8484}8485 8486define void @memmove_p0_p4_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(4) align 1 readonly %src) {8487; CHECK-LABEL: memmove_p0_p4_sz2048:8488; CHECK:       ; %bb.0: ; %entry8489; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)8490; CHECK-NEXT:    s_mov_b32 s4, exec_lo8491; CHECK-NEXT:    v_cmpx_ge_u64_e64 v[2:3], v[0:1]8492; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s48493; CHECK-NEXT:    s_cbranch_execz .LBB7_38494; CHECK-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader8495; CHECK-NEXT:    s_mov_b64 s[4:5], 08496; CHECK-NEXT:  .LBB7_2: ; %memmove_fwd_loop8497; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=18498; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s48499; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo8500; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s48501; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo8502; CHECK-NEXT:    s_clause 0xf8503; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:2408504; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:2248505; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:2088506; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:1928507; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:1768508; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:1608509; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:1448510; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:1288511; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:1128512; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:968513; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:808514; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:648515; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:488516; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:328517; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off offset:168518; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off8519; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 488520; CHECK-NEXT:    s_add_u32 s4, s4, 0x1008521; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo8522; CHECK-NEXT:    s_addc_u32 s5, s5, 08523; CHECK-NEXT:    s_waitcnt vmcnt(15)8524; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[4:7] offset:1928525; CHECK-NEXT:    s_waitcnt vmcnt(14)8526; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[8:11] offset:1768527; CHECK-NEXT:    s_waitcnt vmcnt(13)8528; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[12:15] offset:1608529; CHECK-NEXT:    s_waitcnt vmcnt(12)8530; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[16:19] offset:1448531; CHECK-NEXT:    s_waitcnt vmcnt(11)8532; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[20:23] offset:1288533; CHECK-NEXT:    s_waitcnt vmcnt(10)8534; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:1128535; CHECK-NEXT:    s_waitcnt vmcnt(9)8536; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:968537; CHECK-NEXT:    s_waitcnt vmcnt(8)8538; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[32:35] offset:1288539; CHECK-NEXT:    s_waitcnt vmcnt(7)8540; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:648541; CHECK-NEXT:    s_waitcnt vmcnt(6)8542; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:488543; CHECK-NEXT:    s_waitcnt vmcnt(5)8544; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:328545; CHECK-NEXT:    s_waitcnt vmcnt(4)8546; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:648547; CHECK-NEXT:    s_waitcnt vmcnt(3)8548; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71]8549; CHECK-NEXT:    s_waitcnt vmcnt(2)8550; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:328551; CHECK-NEXT:    s_waitcnt vmcnt(1)8552; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:168553; CHECK-NEXT:    s_waitcnt vmcnt(0)8554; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]8555; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0x8008556; CHECK-NEXT:    s_cbranch_scc1 .LBB7_28557; CHECK-NEXT:  .LBB7_3: ; %Flow68558; CHECK-NEXT:    s_andn2_saveexec_b32 s8, s68559; CHECK-NEXT:    s_cbranch_execz .LBB7_68560; CHECK-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader8561; CHECK-NEXT:    s_movk_i32 s6, 0xff008562; CHECK-NEXT:    s_mov_b64 s[4:5], 0x7008563; CHECK-NEXT:    s_mov_b32 s7, -18564; CHECK-NEXT:  .LBB7_5: ; %memmove_bwd_loop8565; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=18566; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s48567; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo8568; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s48569; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo8570; CHECK-NEXT:    s_clause 0xf8571; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:2408572; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:2248573; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:2088574; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:1928575; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:1768576; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:1608577; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:1448578; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:1288579; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:1128580; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:968581; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:808582; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:648583; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:488584; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:328585; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off offset:168586; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off8587; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 488588; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff008589; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo8590; CHECK-NEXT:    s_addc_u32 s5, s5, -18591; CHECK-NEXT:    s_waitcnt vmcnt(15)8592; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[4:7] offset:1928593; CHECK-NEXT:    s_waitcnt vmcnt(14)8594; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[8:11] offset:1768595; CHECK-NEXT:    s_waitcnt vmcnt(13)8596; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[12:15] offset:1608597; CHECK-NEXT:    s_waitcnt vmcnt(12)8598; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[16:19] offset:1448599; CHECK-NEXT:    s_waitcnt vmcnt(11)8600; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[20:23] offset:1288601; CHECK-NEXT:    s_waitcnt vmcnt(10)8602; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:1128603; CHECK-NEXT:    s_waitcnt vmcnt(9)8604; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:968605; CHECK-NEXT:    s_waitcnt vmcnt(8)8606; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[32:35] offset:1288607; CHECK-NEXT:    s_waitcnt vmcnt(7)8608; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:648609; CHECK-NEXT:    s_waitcnt vmcnt(6)8610; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:488611; CHECK-NEXT:    s_waitcnt vmcnt(5)8612; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:328613; CHECK-NEXT:    s_waitcnt vmcnt(4)8614; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:648615; CHECK-NEXT:    s_waitcnt vmcnt(3)8616; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71]8617; CHECK-NEXT:    s_waitcnt vmcnt(2)8618; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:328619; CHECK-NEXT:    s_waitcnt vmcnt(1)8620; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:168621; CHECK-NEXT:    s_waitcnt vmcnt(0)8622; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]8623; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]8624; CHECK-NEXT:    s_cbranch_scc0 .LBB7_58625; CHECK-NEXT:  .LBB7_6: ; %Flow78626; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s88627; CHECK-NEXT:    s_waitcnt lgkmcnt(0)8628; CHECK-NEXT:    s_setpc_b64 s[30:31]8629;8630; ALIGNED-LABEL: memmove_p0_p4_sz2048:8631; ALIGNED:       ; %bb.0: ; %entry8632; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)8633; ALIGNED-NEXT:    s_mov_b32 s4, exec_lo8634; ALIGNED-NEXT:    v_cmpx_ge_u64_e64 v[2:3], v[0:1]8635; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s48636; ALIGNED-NEXT:    s_cbranch_execz .LBB7_38637; ALIGNED-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader8638; ALIGNED-NEXT:    s_mov_b64 s[4:5], 08639; ALIGNED-NEXT:  .LBB7_2: ; %memmove_fwd_loop8640; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=18641; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, s48642; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, s5, v3, vcc_lo8643; ALIGNED-NEXT:    v_add_co_u32 v84, vcc_lo, v0, s48644; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v85, null, s5, v1, vcc_lo8645; ALIGNED-NEXT:    s_clause 0xf8646; ALIGNED-NEXT:    global_load_dwordx4 v[98:101], v[4:5], off offset:2408647; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[4:5], off offset:2248648; ALIGNED-NEXT:    global_load_dwordx4 v[80:83], v[4:5], off offset:2088649; ALIGNED-NEXT:    global_load_dwordx4 v[68:71], v[4:5], off offset:1928650; ALIGNED-NEXT:    global_load_dwordx4 v[64:67], v[4:5], off offset:1768651; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[4:5], off offset:1608652; ALIGNED-NEXT:    global_load_dwordx4 v[48:51], v[4:5], off offset:1448653; ALIGNED-NEXT:    global_load_dwordx4 v[36:39], v[4:5], off offset:1288654; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[4:5], off offset:1128655; ALIGNED-NEXT:    global_load_dwordx4 v[28:31], v[4:5], off offset:968656; ALIGNED-NEXT:    global_load_dwordx4 v[24:27], v[4:5], off offset:808657; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[4:5], off offset:648658; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[4:5], off offset:488659; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[4:5], off offset:328660; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[4:5], off offset:168661; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[4:5], off8662; ALIGNED-NEXT:    v_add_co_u32 v86, vcc_lo, v84, 68663; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v87, null, 0, v85, vcc_lo8664; ALIGNED-NEXT:    v_add_co_u32 v96, vcc_lo, v84, 38665; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v85, vcc_lo8666; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x1008667; ALIGNED-NEXT:    s_addc_u32 s5, s5, 08668; ALIGNED-NEXT:    s_waitcnt vmcnt(15)8669; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:1688670; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:1728671; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:1648672; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:1608673; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v100 offset:2448674; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v1008675; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v1008676; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v101 offset:2488677; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:2498678; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:2458679; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v99 offset:2408680; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:2418681; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v98 offset:2368682; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:2378683; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v1018684; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1018685; ALIGNED-NEXT:    v_lshrrev_b32_e32 v116, 24, v998686; ALIGNED-NEXT:    v_lshrrev_b32_e32 v117, 8, v998687; ALIGNED-NEXT:    v_lshrrev_b32_e32 v118, 24, v988688; ALIGNED-NEXT:    v_lshrrev_b32_e32 v119, 8, v988689; ALIGNED-NEXT:    s_waitcnt vmcnt(14)8690; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v1148691; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v1148692; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:2488693; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:2468694; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:2528695; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:2508696; ALIGNED-NEXT:    flat_store_byte v[96:97], v116 offset:2448697; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v1158698; ALIGNED-NEXT:    flat_store_byte v[96:97], v117 offset:2428699; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1158700; ALIGNED-NEXT:    flat_store_byte v[96:97], v118 offset:2408701; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v1138702; ALIGNED-NEXT:    flat_store_byte v[96:97], v119 offset:2388703; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v1138704; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:1848705; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:1888706; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:1808707; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:1768708; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v114 offset:2288709; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v115 offset:2328710; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:2338711; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:2298712; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v113 offset:2248713; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:2258714; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v112 offset:2208715; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:2218716; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v1128717; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v1128718; ALIGNED-NEXT:    s_waitcnt vmcnt(13)8719; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v828720; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v828721; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:2328722; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v838723; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:2308724; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v838725; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:2368726; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v818727; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:2348728; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v818729; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:2288730; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v808731; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:2268732; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v808733; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:2248734; ALIGNED-NEXT:    s_waitcnt vmcnt(12)8735; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v708736; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:2228737; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v708738; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:1368739; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:1408740; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:1328741; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:1288742; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v82 offset:2128743; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v83 offset:2168744; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:2178745; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:2138746; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v81 offset:2088747; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:2098748; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v80 offset:2048749; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:2058750; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v718751; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v718752; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v698753; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v698754; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:2168755; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v688756; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:2148757; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v688758; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:2208759; ALIGNED-NEXT:    s_waitcnt vmcnt(11)8760; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v668761; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:2188762; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v668763; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:2128764; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v678765; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:2108766; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v678767; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:2088768; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v658769; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:2068770; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v658771; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:1528772; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:1568773; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:1488774; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:1448775; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v70 offset:1968776; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v71 offset:2008777; ALIGNED-NEXT:    flat_store_byte v[96:97], v71 offset:2018778; ALIGNED-NEXT:    flat_store_byte v[96:97], v70 offset:1978779; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v69 offset:1928780; ALIGNED-NEXT:    flat_store_byte v[96:97], v69 offset:1938781; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v68 offset:1888782; ALIGNED-NEXT:    flat_store_byte v[96:97], v68 offset:1898783; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v648784; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v648785; ALIGNED-NEXT:    s_waitcnt vmcnt(10)8786; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v548787; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v548788; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:2008789; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v558790; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:1988791; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v558792; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:2048793; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v528794; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:2028795; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v528796; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:1968797; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v538798; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:1948799; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v538800; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:1928801; ALIGNED-NEXT:    s_waitcnt vmcnt(9)8802; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v508803; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:1908804; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v508805; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:2328806; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:2368807; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:2288808; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:2248809; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v66 offset:1808810; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v67 offset:1848811; ALIGNED-NEXT:    flat_store_byte v[96:97], v67 offset:1858812; ALIGNED-NEXT:    flat_store_byte v[96:97], v66 offset:1818813; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v65 offset:1768814; ALIGNED-NEXT:    flat_store_byte v[96:97], v65 offset:1778815; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v64 offset:1728816; ALIGNED-NEXT:    flat_store_byte v[96:97], v64 offset:1738817; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v518818; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v518819; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v498820; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v498821; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:1848822; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v488823; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:1828824; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v488825; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:1888826; ALIGNED-NEXT:    s_waitcnt vmcnt(8)8827; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v388828; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:1868829; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v388830; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:1808831; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v398832; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:1788833; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v398834; ALIGNED-NEXT:    flat_store_byte v[96:97], v68 offset:1768835; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v378836; ALIGNED-NEXT:    flat_store_byte v[96:97], v69 offset:1748837; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v378838; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:2528839; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:2488840; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:2448841; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:2408842; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v54 offset:1648843; ALIGNED-NEXT:    flat_store_byte v[96:97], v54 offset:1658844; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v55 offset:1688845; ALIGNED-NEXT:    flat_store_byte v[96:97], v55 offset:1698846; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v52 offset:1568847; ALIGNED-NEXT:    flat_store_byte v[96:97], v52 offset:1578848; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v53 offset:1608849; ALIGNED-NEXT:    flat_store_byte v[96:97], v53 offset:1618850; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v368851; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v368852; ALIGNED-NEXT:    s_waitcnt vmcnt(7)8853; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v348854; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v348855; ALIGNED-NEXT:    flat_store_byte v[96:97], v70 offset:1688856; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v358857; ALIGNED-NEXT:    flat_store_byte v[96:97], v71 offset:1668858; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v358859; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:1728860; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v338861; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:1708862; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v338863; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:1608864; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v328865; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:1588866; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v328867; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:1648868; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:1628869; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:2008870; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:2048871; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:1968872; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:1928873; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:1528874; ALIGNED-NEXT:    s_waitcnt vmcnt(6)8875; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v288876; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:1508877; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v288878; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v308879; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v308880; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v50 offset:1488881; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v51 offset:1528882; ALIGNED-NEXT:    flat_store_byte v[96:97], v51 offset:1538883; ALIGNED-NEXT:    flat_store_byte v[96:97], v50 offset:1498884; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v49 offset:1448885; ALIGNED-NEXT:    flat_store_byte v[96:97], v49 offset:1458886; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v48 offset:1408887; ALIGNED-NEXT:    flat_store_byte v[96:97], v48 offset:1418888; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v318889; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v318890; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v298891; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v298892; ALIGNED-NEXT:    flat_store_byte v[96:97], v64 offset:1568893; ALIGNED-NEXT:    s_waitcnt vmcnt(5)8894; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v268895; ALIGNED-NEXT:    flat_store_byte v[96:97], v65 offset:1548896; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v268897; ALIGNED-NEXT:    flat_store_byte v[96:97], v66 offset:1488898; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v278899; ALIGNED-NEXT:    flat_store_byte v[96:97], v67 offset:1468900; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v278901; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:1448902; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v258903; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:1428904; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v258905; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:2168906; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:2208907; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:2128908; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:2088909; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v38 offset:1328910; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v39 offset:1368911; ALIGNED-NEXT:    flat_store_byte v[96:97], v39 offset:1378912; ALIGNED-NEXT:    flat_store_byte v[96:97], v38 offset:1338913; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v37 offset:1288914; ALIGNED-NEXT:    flat_store_byte v[96:97], v37 offset:1298915; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v36 offset:1248916; ALIGNED-NEXT:    flat_store_byte v[84:85], v36 offset:1288917; ALIGNED-NEXT:    s_waitcnt vmcnt(4)8918; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v228919; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:1368920; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:1348921; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:1408922; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:1388923; ALIGNED-NEXT:    flat_store_byte v[96:97], v68 offset:1328924; ALIGNED-NEXT:    flat_store_byte v[96:97], v69 offset:1308925; ALIGNED-NEXT:    flat_store_byte v[96:97], v52 offset:1288926; ALIGNED-NEXT:    s_waitcnt vmcnt(3)8927; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v188928; ALIGNED-NEXT:    flat_store_byte v[96:97], v53 offset:1268929; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:408930; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:448931; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368932; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:328933; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v34 offset:1168934; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v35 offset:1208935; ALIGNED-NEXT:    flat_store_byte v[96:97], v35 offset:1218936; ALIGNED-NEXT:    flat_store_byte v[96:97], v34 offset:1178937; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v33 offset:1128938; ALIGNED-NEXT:    flat_store_byte v[96:97], v33 offset:1138939; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v32 offset:1088940; ALIGNED-NEXT:    flat_store_byte v[96:97], v32 offset:1098941; ALIGNED-NEXT:    flat_store_byte v[96:97], v54 offset:1208942; ALIGNED-NEXT:    flat_store_byte v[96:97], v55 offset:1188943; ALIGNED-NEXT:    flat_store_byte v[96:97], v70 offset:1248944; ALIGNED-NEXT:    s_waitcnt vmcnt(2)8945; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v148946; ALIGNED-NEXT:    flat_store_byte v[96:97], v71 offset:1228947; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:1168948; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:1148949; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:1128950; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:1108951; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:568952; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:608953; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:528954; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:488955; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v30 offset:1008956; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v31 offset:1048957; ALIGNED-NEXT:    flat_store_byte v[96:97], v31 offset:1058958; ALIGNED-NEXT:    flat_store_byte v[96:97], v30 offset:1018959; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v29 offset:968960; ALIGNED-NEXT:    flat_store_byte v[96:97], v29 offset:978961; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v28 offset:928962; ALIGNED-NEXT:    flat_store_byte v[96:97], v28 offset:938963; ALIGNED-NEXT:    s_waitcnt vmcnt(1)8964; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v108965; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:968966; ALIGNED-NEXT:    s_waitcnt vmcnt(0)8967; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v68968; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v228969; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v188970; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v148971; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v108972; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:948973; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v68974; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0x8008975; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v238976; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v198977; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v158978; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:1048979; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v118980; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:1028981; ALIGNED-NEXT:    flat_store_byte v[96:97], v48 offset:1088982; ALIGNED-NEXT:    flat_store_byte v[96:97], v49 offset:1068983; ALIGNED-NEXT:    flat_store_byte v[96:97], v50 offset:1008984; ALIGNED-NEXT:    flat_store_byte v[96:97], v51 offset:988985; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:88986; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:128987; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:48988; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s328989; ALIGNED-NEXT:    flat_store_byte v[96:97], v64 offset:888990; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v78991; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v248992; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v248993; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v238994; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v218995; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v218996; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v208997; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v208998; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v198999; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v179000; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v179001; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v169002; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v169003; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v159004; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v139005; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v139006; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v129007; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v129008; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v119009; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v99010; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v99011; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v89012; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v89013; ALIGNED-NEXT:    flat_store_byte v[96:97], v65 offset:869014; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v79015; ALIGNED-NEXT:    flat_store_byte v[96:97], v66 offset:929016; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v59017; ALIGNED-NEXT:    flat_store_byte v[96:97], v67 offset:909018; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v59019; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:849020; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v49021; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:829022; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v49023; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v26 offset:849024; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v27 offset:889025; ALIGNED-NEXT:    flat_store_byte v[96:97], v27 offset:899026; ALIGNED-NEXT:    flat_store_byte v[96:97], v26 offset:859027; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v25 offset:809028; ALIGNED-NEXT:    flat_store_byte v[96:97], v25 offset:819029; ALIGNED-NEXT:    flat_store_byte v[96:97], v36 offset:809030; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v24 offset:769031; ALIGNED-NEXT:    flat_store_byte v[96:97], v37 offset:789032; ALIGNED-NEXT:    flat_store_byte v[96:97], v24 offset:779033; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:249034; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:289035; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:209036; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:169037; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v22 offset:689038; ALIGNED-NEXT:    flat_store_byte v[96:97], v38 offset:729039; ALIGNED-NEXT:    flat_store_byte v[96:97], v39 offset:709040; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v23 offset:729041; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:769042; ALIGNED-NEXT:    flat_store_byte v[96:97], v23 offset:739043; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:749044; ALIGNED-NEXT:    flat_store_byte v[96:97], v22 offset:699045; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:689046; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v21 offset:649047; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:669048; ALIGNED-NEXT:    flat_store_byte v[96:97], v21 offset:659049; ALIGNED-NEXT:    flat_store_byte v[96:97], v68 offset:649050; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v20 offset:609051; ALIGNED-NEXT:    flat_store_byte v[96:97], v69 offset:629052; ALIGNED-NEXT:    flat_store_byte v[84:85], v20 offset:649053; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:1049054; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:1089055; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:1009056; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:969057; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v18 offset:529058; ALIGNED-NEXT:    flat_store_byte v[96:97], v52 offset:569059; ALIGNED-NEXT:    flat_store_byte v[96:97], v53 offset:549060; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v19 offset:569061; ALIGNED-NEXT:    flat_store_byte v[96:97], v32 offset:609062; ALIGNED-NEXT:    flat_store_byte v[96:97], v19 offset:579063; ALIGNED-NEXT:    flat_store_byte v[96:97], v33 offset:589064; ALIGNED-NEXT:    flat_store_byte v[96:97], v18 offset:539065; ALIGNED-NEXT:    flat_store_byte v[96:97], v34 offset:529066; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v17 offset:489067; ALIGNED-NEXT:    flat_store_byte v[96:97], v35 offset:509068; ALIGNED-NEXT:    flat_store_byte v[96:97], v17 offset:499069; ALIGNED-NEXT:    flat_store_byte v[96:97], v54 offset:489070; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v16 offset:449071; ALIGNED-NEXT:    flat_store_byte v[96:97], v55 offset:469072; ALIGNED-NEXT:    flat_store_byte v[96:97], v16 offset:459073; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:1209074; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:1249075; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1169076; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1129077; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v14 offset:369078; ALIGNED-NEXT:    flat_store_byte v[96:97], v70 offset:409079; ALIGNED-NEXT:    flat_store_byte v[96:97], v71 offset:389080; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v15 offset:409081; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:449082; ALIGNED-NEXT:    flat_store_byte v[96:97], v15 offset:419083; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:429084; ALIGNED-NEXT:    flat_store_byte v[96:97], v14 offset:379085; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:369086; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v13 offset:329087; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:349088; ALIGNED-NEXT:    flat_store_byte v[96:97], v13 offset:339089; ALIGNED-NEXT:    flat_store_byte v[96:97], v28 offset:329090; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v12 offset:289091; ALIGNED-NEXT:    flat_store_byte v[96:97], v29 offset:309092; ALIGNED-NEXT:    flat_store_byte v[84:85], v12 offset:329093; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:729094; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:769095; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:689096; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:649097; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v10 offset:209098; ALIGNED-NEXT:    flat_store_byte v[96:97], v30 offset:249099; ALIGNED-NEXT:    flat_store_byte v[96:97], v31 offset:229100; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v11 offset:249101; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:289102; ALIGNED-NEXT:    flat_store_byte v[96:97], v11 offset:259103; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:269104; ALIGNED-NEXT:    flat_store_byte v[96:97], v10 offset:219105; ALIGNED-NEXT:    flat_store_byte v[96:97], v48 offset:209106; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v9 offset:169107; ALIGNED-NEXT:    flat_store_byte v[96:97], v49 offset:189108; ALIGNED-NEXT:    flat_store_byte v[96:97], v9 offset:179109; ALIGNED-NEXT:    flat_store_byte v[96:97], v50 offset:169110; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v8 offset:129111; ALIGNED-NEXT:    flat_store_byte v[96:97], v51 offset:149112; ALIGNED-NEXT:    flat_store_byte v[84:85], v8 offset:169113; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:889114; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:929115; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:849116; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:809117; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v6 offset:49118; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:89119; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:69120; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v7 offset:89121; ALIGNED-NEXT:    flat_store_byte v[96:97], v64 offset:129122; ALIGNED-NEXT:    flat_store_byte v[96:97], v7 offset:99123; ALIGNED-NEXT:    flat_store_byte v[96:97], v65 offset:109124; ALIGNED-NEXT:    flat_store_byte v[84:85], v6 offset:89125; ALIGNED-NEXT:    flat_store_byte v[96:97], v66 offset:49126; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v59127; ALIGNED-NEXT:    flat_store_byte v[96:97], v67 offset:29128; ALIGNED-NEXT:    flat_store_byte v[84:85], v5 offset:49129; ALIGNED-NEXT:    flat_store_byte v[96:97], v989130; ALIGNED-NEXT:    flat_store_byte_d16_hi v[84:85], v4 offset:29131; ALIGNED-NEXT:    flat_store_byte v[84:85], v99 offset:19132; ALIGNED-NEXT:    flat_store_byte v[84:85], v49133; ALIGNED-NEXT:    s_cbranch_scc1 .LBB7_29134; ALIGNED-NEXT:  .LBB7_3: ; %Flow69135; ALIGNED-NEXT:    s_andn2_saveexec_b32 s8, s69136; ALIGNED-NEXT:    s_cbranch_execz .LBB7_69137; ALIGNED-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader9138; ALIGNED-NEXT:    s_movk_i32 s6, 0xff009139; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x7009140; ALIGNED-NEXT:    s_mov_b32 s7, -19141; ALIGNED-NEXT:  .LBB7_5: ; %memmove_bwd_loop9142; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=19143; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, s49144; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, s5, v3, vcc_lo9145; ALIGNED-NEXT:    v_add_co_u32 v84, vcc_lo, v0, s49146; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v85, null, s5, v1, vcc_lo9147; ALIGNED-NEXT:    s_clause 0xf9148; ALIGNED-NEXT:    global_load_dwordx4 v[98:101], v[4:5], off offset:2409149; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[4:5], off offset:2249150; ALIGNED-NEXT:    global_load_dwordx4 v[80:83], v[4:5], off offset:2089151; ALIGNED-NEXT:    global_load_dwordx4 v[68:71], v[4:5], off offset:1929152; ALIGNED-NEXT:    global_load_dwordx4 v[64:67], v[4:5], off offset:1769153; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[4:5], off offset:1609154; ALIGNED-NEXT:    global_load_dwordx4 v[48:51], v[4:5], off offset:1449155; ALIGNED-NEXT:    global_load_dwordx4 v[36:39], v[4:5], off offset:1289156; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[4:5], off offset:1129157; ALIGNED-NEXT:    global_load_dwordx4 v[28:31], v[4:5], off offset:969158; ALIGNED-NEXT:    global_load_dwordx4 v[24:27], v[4:5], off offset:809159; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[4:5], off offset:649160; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[4:5], off offset:489161; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[4:5], off offset:329162; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[4:5], off offset:169163; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[4:5], off9164; ALIGNED-NEXT:    v_add_co_u32 v96, vcc_lo, v84, 69165; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v85, vcc_lo9166; ALIGNED-NEXT:    v_add_co_u32 v86, vcc_lo, v84, 39167; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v87, null, 0, v85, vcc_lo9168; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff009169; ALIGNED-NEXT:    s_addc_u32 s5, s5, -19170; ALIGNED-NEXT:    s_waitcnt vmcnt(15)9171; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:4249172; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:4289173; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:4209174; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:4169175; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v100 offset:2449176; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v101 offset:2489177; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:2499178; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:2459179; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v99 offset:2409180; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:2419181; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v98 offset:2369182; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:2379183; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v1009184; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v1019185; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1019186; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v1009187; ALIGNED-NEXT:    v_lshrrev_b32_e32 v116, 24, v999188; ALIGNED-NEXT:    v_lshrrev_b32_e32 v117, 8, v999189; ALIGNED-NEXT:    v_lshrrev_b32_e32 v118, 24, v989190; ALIGNED-NEXT:    v_lshrrev_b32_e32 v119, 8, v989191; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:2489192; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:2469193; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:2529194; ALIGNED-NEXT:    s_waitcnt vmcnt(14)9195; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v1149196; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:2509197; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v1149198; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v1159199; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v1159200; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v1139201; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v1139202; ALIGNED-NEXT:    flat_store_byte v[86:87], v116 offset:2449203; ALIGNED-NEXT:    flat_store_byte v[86:87], v117 offset:2429204; ALIGNED-NEXT:    flat_store_byte v[86:87], v118 offset:2409205; ALIGNED-NEXT:    flat_store_byte v[86:87], v119 offset:2389206; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:4409207; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:4449208; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:4369209; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:4329210; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v114 offset:2289211; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v115 offset:2329212; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:2339213; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:2299214; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v113 offset:2249215; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:2259216; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v112 offset:2209217; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:2219218; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v1129219; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v1129220; ALIGNED-NEXT:    s_waitcnt vmcnt(13)9221; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v829222; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v829223; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:2329224; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v839225; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:2309226; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v839227; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:2369228; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v819229; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:2349230; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v819231; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:2289232; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v809233; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:2269234; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v809235; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:2249236; ALIGNED-NEXT:    s_waitcnt vmcnt(12)9237; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v709238; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:2229239; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v709240; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:3929241; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:3969242; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:3889243; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:3849244; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v82 offset:2129245; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v83 offset:2169246; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:2179247; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:2139248; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v81 offset:2089249; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:2099250; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v80 offset:2049251; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:2059252; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v719253; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v719254; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v699255; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v699256; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:2169257; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v689258; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:2149259; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v689260; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:2209261; ALIGNED-NEXT:    s_waitcnt vmcnt(11)9262; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v669263; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:2189264; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v669265; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:2129266; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v679267; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:2109268; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v679269; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:2089270; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v659271; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:2069272; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v659273; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:4089274; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:4129275; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:4049276; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:4009277; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v70 offset:1969278; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v71 offset:2009279; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:2019280; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:1979281; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v69 offset:1929282; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:1939283; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v68 offset:1889284; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:1899285; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v649286; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v649287; ALIGNED-NEXT:    s_waitcnt vmcnt(10)9288; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v549289; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v549290; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:2009291; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v559292; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:1989293; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v559294; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:2049295; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v529296; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:2029297; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v529298; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:1969299; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v539300; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:1949301; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v539302; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:1929303; ALIGNED-NEXT:    s_waitcnt vmcnt(9)9304; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v509305; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:1909306; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v509307; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:4889308; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:4929309; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:4849310; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:4809311; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v66 offset:1809312; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v67 offset:1849313; ALIGNED-NEXT:    flat_store_byte v[86:87], v67 offset:1859314; ALIGNED-NEXT:    flat_store_byte v[86:87], v66 offset:1819315; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v65 offset:1769316; ALIGNED-NEXT:    flat_store_byte v[86:87], v65 offset:1779317; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v64 offset:1729318; ALIGNED-NEXT:    flat_store_byte v[86:87], v64 offset:1739319; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v519320; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v519321; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v499322; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v499323; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:1849324; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v489325; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:1829326; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v489327; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:1889328; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:1869329; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:1809330; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:1789331; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:1769332; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:1749333; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:5089334; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:5049335; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:5009336; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:4969337; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v54 offset:1649338; ALIGNED-NEXT:    flat_store_byte v[86:87], v54 offset:1659339; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v55 offset:1689340; ALIGNED-NEXT:    flat_store_byte v[86:87], v55 offset:1699341; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v52 offset:1569342; ALIGNED-NEXT:    flat_store_byte v[86:87], v52 offset:1579343; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v53 offset:1609344; ALIGNED-NEXT:    flat_store_byte v[86:87], v53 offset:1619345; ALIGNED-NEXT:    s_waitcnt vmcnt(8)9346; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v369347; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v389348; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v389349; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v399350; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v399351; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v379352; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v379353; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v369354; ALIGNED-NEXT:    s_waitcnt vmcnt(7)9355; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v349356; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v349357; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:1689358; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v359359; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:1669360; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v359361; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:1729362; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v339363; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:1709364; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v339365; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:1609366; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v329367; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:1589368; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v329369; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:1649370; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:1629371; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:4569372; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:4609373; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:4529374; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:4489375; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:1529376; ALIGNED-NEXT:    s_waitcnt vmcnt(6)9377; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v289378; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:1509379; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v289380; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v309381; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v309382; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v50 offset:1489383; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v51 offset:1529384; ALIGNED-NEXT:    flat_store_byte v[86:87], v51 offset:1539385; ALIGNED-NEXT:    flat_store_byte v[86:87], v50 offset:1499386; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v49 offset:1449387; ALIGNED-NEXT:    flat_store_byte v[86:87], v49 offset:1459388; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v48 offset:1409389; ALIGNED-NEXT:    flat_store_byte v[86:87], v48 offset:1419390; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v319391; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v319392; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v299393; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v299394; ALIGNED-NEXT:    flat_store_byte v[86:87], v64 offset:1569395; ALIGNED-NEXT:    s_waitcnt vmcnt(5)9396; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v269397; ALIGNED-NEXT:    flat_store_byte v[86:87], v65 offset:1549398; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v269399; ALIGNED-NEXT:    flat_store_byte v[86:87], v66 offset:1489400; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v279401; ALIGNED-NEXT:    flat_store_byte v[86:87], v67 offset:1469402; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:1449403; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:1429404; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:4729405; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:4769406; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:4689407; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:4649408; ALIGNED-NEXT:    flat_store_byte v[86:87], v52 offset:1289409; ALIGNED-NEXT:    s_waitcnt vmcnt(3)9410; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v189411; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v279412; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v38 offset:1329413; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v39 offset:1369414; ALIGNED-NEXT:    flat_store_byte v[86:87], v39 offset:1379415; ALIGNED-NEXT:    flat_store_byte v[86:87], v38 offset:1339416; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v37 offset:1289417; ALIGNED-NEXT:    flat_store_byte v[86:87], v37 offset:1299418; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v36 offset:1249419; ALIGNED-NEXT:    flat_store_byte v[84:85], v36 offset:1289420; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:1369421; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:1349422; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:1409423; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:1389424; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:1329425; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:1309426; ALIGNED-NEXT:    flat_store_byte v[86:87], v53 offset:1269427; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v189428; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:2969429; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:3009430; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:2929431; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:2889432; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v34 offset:1169433; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v35 offset:1209434; ALIGNED-NEXT:    flat_store_byte v[86:87], v35 offset:1219435; ALIGNED-NEXT:    flat_store_byte v[86:87], v34 offset:1179436; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v33 offset:1129437; ALIGNED-NEXT:    flat_store_byte v[86:87], v33 offset:1139438; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v32 offset:1089439; ALIGNED-NEXT:    flat_store_byte v[86:87], v32 offset:1099440; ALIGNED-NEXT:    flat_store_byte v[86:87], v54 offset:1209441; ALIGNED-NEXT:    flat_store_byte v[86:87], v55 offset:1189442; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:1249443; ALIGNED-NEXT:    s_waitcnt vmcnt(2)9444; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v149445; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:1229446; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:1169447; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:1149448; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:1129449; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:1109450; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:3129451; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:3169452; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:3089453; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:3049454; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v30 offset:1009455; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v31 offset:1049456; ALIGNED-NEXT:    flat_store_byte v[86:87], v31 offset:1059457; ALIGNED-NEXT:    flat_store_byte v[86:87], v30 offset:1019458; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v29 offset:969459; ALIGNED-NEXT:    flat_store_byte v[86:87], v29 offset:979460; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v28 offset:929461; ALIGNED-NEXT:    flat_store_byte v[86:87], v28 offset:939462; ALIGNED-NEXT:    s_waitcnt vmcnt(1)9463; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v109464; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:969465; ALIGNED-NEXT:    s_waitcnt vmcnt(0)9466; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v69467; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v259468; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v199469; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v149470; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v109471; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:949472; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v69473; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]9474; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v259475; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v199476; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v249477; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v249478; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v179479; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v159480; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:1049481; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v119482; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:1029483; ALIGNED-NEXT:    flat_store_byte v[86:87], v48 offset:1089484; ALIGNED-NEXT:    flat_store_byte v[86:87], v49 offset:1069485; ALIGNED-NEXT:    flat_store_byte v[86:87], v50 offset:1009486; ALIGNED-NEXT:    flat_store_byte v[86:87], v51 offset:989487; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:2649488; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:2689489; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:2609490; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:2569491; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v26 offset:849492; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v27 offset:889493; ALIGNED-NEXT:    flat_store_byte v[86:87], v27 offset:899494; ALIGNED-NEXT:    flat_store_byte v[86:87], v26 offset:859495; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v25 offset:809496; ALIGNED-NEXT:    flat_store_byte v[86:87], v25 offset:819497; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v24 offset:769498; ALIGNED-NEXT:    flat_store_byte v[86:87], v24 offset:779499; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v79500; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v229501; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v229502; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v239503; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v239504; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v219505; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v219506; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v209507; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v209508; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v179509; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v169510; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v169511; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v159512; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v139513; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v139514; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v129515; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v129516; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v119517; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v99518; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v99519; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v89520; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v89521; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v79522; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v59523; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v59524; ALIGNED-NEXT:    flat_store_byte v[86:87], v64 offset:889525; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v49526; ALIGNED-NEXT:    flat_store_byte v[86:87], v65 offset:869527; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v49528; ALIGNED-NEXT:    flat_store_byte v[86:87], v66 offset:929529; ALIGNED-NEXT:    flat_store_byte v[86:87], v67 offset:909530; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:849531; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:829532; ALIGNED-NEXT:    flat_store_byte v[86:87], v36 offset:809533; ALIGNED-NEXT:    flat_store_byte v[86:87], v37 offset:789534; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:2809535; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:2849536; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:2769537; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:2729538; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v22 offset:689539; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v23 offset:729540; ALIGNED-NEXT:    flat_store_byte v[86:87], v23 offset:739541; ALIGNED-NEXT:    flat_store_byte v[86:87], v22 offset:699542; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v21 offset:649543; ALIGNED-NEXT:    flat_store_byte v[86:87], v21 offset:659544; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v20 offset:609545; ALIGNED-NEXT:    flat_store_byte v[84:85], v20 offset:649546; ALIGNED-NEXT:    flat_store_byte v[86:87], v38 offset:729547; ALIGNED-NEXT:    flat_store_byte v[86:87], v39 offset:709548; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:769549; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:749550; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:689551; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:669552; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:649553; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:629554; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:3609555; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:3649556; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:3569557; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:3529558; ALIGNED-NEXT:    flat_store_byte v[86:87], v52 offset:569559; ALIGNED-NEXT:    flat_store_byte v[86:87], v53 offset:549560; ALIGNED-NEXT:    flat_store_byte v[86:87], v32 offset:609561; ALIGNED-NEXT:    flat_store_byte v[86:87], v33 offset:589562; ALIGNED-NEXT:    flat_store_byte v[86:87], v34 offset:529563; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v18 offset:529564; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v19 offset:569565; ALIGNED-NEXT:    flat_store_byte v[86:87], v19 offset:579566; ALIGNED-NEXT:    flat_store_byte v[86:87], v18 offset:539567; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v17 offset:489568; ALIGNED-NEXT:    flat_store_byte v[86:87], v35 offset:509569; ALIGNED-NEXT:    flat_store_byte v[86:87], v17 offset:499570; ALIGNED-NEXT:    flat_store_byte v[86:87], v54 offset:489571; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v16 offset:449572; ALIGNED-NEXT:    flat_store_byte v[86:87], v55 offset:469573; ALIGNED-NEXT:    flat_store_byte v[86:87], v16 offset:459574; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:3769575; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:3809576; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:3729577; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:3689578; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v14 offset:369579; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:409580; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:389581; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v15 offset:409582; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:449583; ALIGNED-NEXT:    flat_store_byte v[86:87], v15 offset:419584; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:429585; ALIGNED-NEXT:    flat_store_byte v[86:87], v14 offset:379586; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:369587; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v13 offset:329588; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:349589; ALIGNED-NEXT:    flat_store_byte v[86:87], v13 offset:339590; ALIGNED-NEXT:    flat_store_byte v[86:87], v28 offset:329591; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v12 offset:289592; ALIGNED-NEXT:    flat_store_byte v[86:87], v29 offset:309593; ALIGNED-NEXT:    flat_store_byte v[84:85], v12 offset:329594; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:3289595; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:3329596; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:3249597; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:3209598; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v10 offset:209599; ALIGNED-NEXT:    flat_store_byte v[86:87], v30 offset:249600; ALIGNED-NEXT:    flat_store_byte v[86:87], v31 offset:229601; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v11 offset:249602; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:289603; ALIGNED-NEXT:    flat_store_byte v[86:87], v11 offset:259604; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:269605; ALIGNED-NEXT:    flat_store_byte v[86:87], v10 offset:219606; ALIGNED-NEXT:    flat_store_byte v[86:87], v48 offset:209607; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v9 offset:169608; ALIGNED-NEXT:    flat_store_byte v[86:87], v49 offset:189609; ALIGNED-NEXT:    flat_store_byte v[86:87], v9 offset:179610; ALIGNED-NEXT:    flat_store_byte v[86:87], v50 offset:169611; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v8 offset:129612; ALIGNED-NEXT:    flat_store_byte v[86:87], v51 offset:149613; ALIGNED-NEXT:    flat_store_byte v[84:85], v8 offset:169614; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:3449615; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:3489616; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:3409617; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:3369618; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v6 offset:49619; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:89620; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:69621; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v7 offset:89622; ALIGNED-NEXT:    flat_store_byte v[86:87], v24 offset:129623; ALIGNED-NEXT:    flat_store_byte v[86:87], v7 offset:99624; ALIGNED-NEXT:    flat_store_byte v[86:87], v25 offset:109625; ALIGNED-NEXT:    flat_store_byte v[84:85], v6 offset:89626; ALIGNED-NEXT:    flat_store_byte v[86:87], v26 offset:49627; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v59628; ALIGNED-NEXT:    flat_store_byte v[86:87], v27 offset:29629; ALIGNED-NEXT:    flat_store_byte v[84:85], v5 offset:49630; ALIGNED-NEXT:    flat_store_byte v[86:87], v649631; ALIGNED-NEXT:    flat_store_byte_d16_hi v[84:85], v4 offset:29632; ALIGNED-NEXT:    flat_store_byte v[84:85], v65 offset:19633; ALIGNED-NEXT:    flat_store_byte v[84:85], v49634; ALIGNED-NEXT:    s_cbranch_scc0 .LBB7_59635; ALIGNED-NEXT:  .LBB7_6: ; %Flow79636; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s89637; ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)9638; ALIGNED-NEXT:    s_setpc_b64 s[30:31]9639;9640; UNROLL3-LABEL: memmove_p0_p4_sz2048:9641; UNROLL3:       ; %bb.0: ; %entry9642; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)9643; UNROLL3-NEXT:    s_mov_b32 s4, exec_lo9644; UNROLL3-NEXT:    v_cmpx_ge_u64_e64 v[2:3], v[0:1]9645; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s49646; UNROLL3-NEXT:    s_cbranch_execz .LBB7_49647; UNROLL3-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader9648; UNROLL3-NEXT:    s_mov_b64 s[4:5], 09649; UNROLL3-NEXT:    .p2align 69650; UNROLL3-NEXT:  .LBB7_2: ; %memmove_fwd_loop9651; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=19652; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s49653; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo9654; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s49655; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo9656; UNROLL3-NEXT:    s_clause 0x29657; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off offset:169658; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off9659; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:329660; UNROLL3-NEXT:    s_add_u32 s4, s4, 489661; UNROLL3-NEXT:    s_addc_u32 s5, s5, 09662; UNROLL3-NEXT:    s_waitcnt vmcnt(2)9663; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7] offset:169664; UNROLL3-NEXT:    s_waitcnt vmcnt(1)9665; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11]9666; UNROLL3-NEXT:    s_waitcnt vmcnt(0)9667; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:329668; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0x7e09669; UNROLL3-NEXT:    s_cbranch_scc1 .LBB7_29670; UNROLL3-NEXT:  ; %bb.3: ; %memmove_fwd_residual9671; UNROLL3-NEXT:    s_clause 0x19672; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:20169673; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:20329674; UNROLL3-NEXT:    ; implicit-def: $vgpr2_vgpr39675; UNROLL3-NEXT:    s_waitcnt vmcnt(1)9676; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:20169677; UNROLL3-NEXT:    s_waitcnt vmcnt(0)9678; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[8:11] offset:20329679; UNROLL3-NEXT:    ; implicit-def: $vgpr0_vgpr19680; UNROLL3-NEXT:  .LBB7_4: ; %Flow49681; UNROLL3-NEXT:    s_andn2_saveexec_b32 s8, s69682; UNROLL3-NEXT:    s_cbranch_execz .LBB7_79683; UNROLL3-NEXT:  ; %bb.5: ; %memmove_bwd_residual9684; UNROLL3-NEXT:    s_clause 0x19685; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:20329686; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:20169687; UNROLL3-NEXT:    s_movk_i32 s6, 0xffd09688; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7b09689; UNROLL3-NEXT:    s_mov_b32 s7, -19690; UNROLL3-NEXT:    s_waitcnt vmcnt(1)9691; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:20329692; UNROLL3-NEXT:    s_waitcnt vmcnt(0)9693; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[8:11] offset:20169694; UNROLL3-NEXT:    .p2align 69695; UNROLL3-NEXT:  .LBB7_6: ; %memmove_bwd_loop9696; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=19697; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s49698; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo9699; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s49700; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo9701; UNROLL3-NEXT:    s_clause 0x29702; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off offset:169703; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off9704; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:329705; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd09706; UNROLL3-NEXT:    s_addc_u32 s5, s5, -19707; UNROLL3-NEXT:    s_waitcnt vmcnt(2)9708; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7] offset:169709; UNROLL3-NEXT:    s_waitcnt vmcnt(1)9710; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11]9711; UNROLL3-NEXT:    s_waitcnt vmcnt(0)9712; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:329713; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]9714; UNROLL3-NEXT:    s_cbranch_scc0 .LBB7_69715; UNROLL3-NEXT:  .LBB7_7: ; %Flow59716; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s89717; UNROLL3-NEXT:    s_waitcnt lgkmcnt(0)9718; UNROLL3-NEXT:    s_setpc_b64 s[30:31]9719entry:9720  tail call void @llvm.memmove.p0.p4.i64(ptr addrspace(0) noundef nonnull align 1 %dst, ptr addrspace(4) noundef nonnull align 1 %src, i64 2048, i1 false)9721  ret void9722}9723 9724define void @memmove_p5_p5_sz2048(ptr addrspace(5) align 1 %dst, ptr addrspace(5) align 1 readonly %src) {9725; CHECK-LABEL: memmove_p5_p5_sz2048:9726; CHECK:       ; %bb.0: ; %entry9727; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)9728; CHECK-NEXT:    s_mov_b32 s4, exec_lo9729; CHECK-NEXT:    v_cmpx_ge_u32_e64 v1, v09730; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s49731; CHECK-NEXT:    s_cbranch_execz .LBB8_39732; CHECK-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader9733; CHECK-NEXT:    s_mov_b64 s[4:5], 0x8009734; CHECK-NEXT:  .LBB8_2: ; %memmove_fwd_loop9735; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=19736; CHECK-NEXT:    s_clause 0x3e9737; CHECK-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:2529738; CHECK-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:2489739; CHECK-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:2449740; CHECK-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:2409741; CHECK-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:2369742; CHECK-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:2329743; CHECK-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:2289744; CHECK-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:2249745; CHECK-NEXT:    buffer_load_dword v10, v1, s[0:3], 0 offen offset:2209746; CHECK-NEXT:    buffer_load_dword v11, v1, s[0:3], 0 offen offset:2169747; CHECK-NEXT:    buffer_load_dword v12, v1, s[0:3], 0 offen offset:2129748; CHECK-NEXT:    buffer_load_dword v13, v1, s[0:3], 0 offen offset:2089749; CHECK-NEXT:    buffer_load_dword v14, v1, s[0:3], 0 offen offset:2049750; CHECK-NEXT:    buffer_load_dword v15, v1, s[0:3], 0 offen offset:2009751; CHECK-NEXT:    buffer_load_dword v16, v1, s[0:3], 0 offen offset:1969752; CHECK-NEXT:    buffer_load_dword v17, v1, s[0:3], 0 offen offset:1929753; CHECK-NEXT:    buffer_load_dword v18, v1, s[0:3], 0 offen offset:1889754; CHECK-NEXT:    buffer_load_dword v19, v1, s[0:3], 0 offen offset:1849755; CHECK-NEXT:    buffer_load_dword v20, v1, s[0:3], 0 offen offset:1809756; CHECK-NEXT:    buffer_load_dword v21, v1, s[0:3], 0 offen offset:1769757; CHECK-NEXT:    buffer_load_dword v22, v1, s[0:3], 0 offen offset:1729758; CHECK-NEXT:    buffer_load_dword v23, v1, s[0:3], 0 offen offset:1689759; CHECK-NEXT:    buffer_load_dword v24, v1, s[0:3], 0 offen offset:1649760; CHECK-NEXT:    buffer_load_dword v25, v1, s[0:3], 0 offen offset:1609761; CHECK-NEXT:    buffer_load_dword v26, v1, s[0:3], 0 offen offset:1569762; CHECK-NEXT:    buffer_load_dword v27, v1, s[0:3], 0 offen offset:1529763; CHECK-NEXT:    buffer_load_dword v28, v1, s[0:3], 0 offen offset:1489764; CHECK-NEXT:    buffer_load_dword v29, v1, s[0:3], 0 offen offset:1449765; CHECK-NEXT:    buffer_load_dword v30, v1, s[0:3], 0 offen offset:1409766; CHECK-NEXT:    buffer_load_dword v31, v1, s[0:3], 0 offen offset:1369767; CHECK-NEXT:    buffer_load_dword v32, v1, s[0:3], 0 offen offset:1329768; CHECK-NEXT:    buffer_load_dword v33, v1, s[0:3], 0 offen offset:1289769; CHECK-NEXT:    buffer_load_dword v34, v1, s[0:3], 0 offen offset:1249770; CHECK-NEXT:    buffer_load_dword v35, v1, s[0:3], 0 offen offset:1209771; CHECK-NEXT:    buffer_load_dword v36, v1, s[0:3], 0 offen offset:1169772; CHECK-NEXT:    buffer_load_dword v37, v1, s[0:3], 0 offen offset:1129773; CHECK-NEXT:    buffer_load_dword v38, v1, s[0:3], 0 offen offset:1089774; CHECK-NEXT:    buffer_load_dword v39, v1, s[0:3], 0 offen offset:1049775; CHECK-NEXT:    buffer_load_dword v48, v1, s[0:3], 0 offen offset:1009776; CHECK-NEXT:    buffer_load_dword v49, v1, s[0:3], 0 offen offset:969777; CHECK-NEXT:    buffer_load_dword v50, v1, s[0:3], 0 offen offset:929778; CHECK-NEXT:    buffer_load_dword v51, v1, s[0:3], 0 offen offset:889779; CHECK-NEXT:    buffer_load_dword v52, v1, s[0:3], 0 offen offset:849780; CHECK-NEXT:    buffer_load_dword v53, v1, s[0:3], 0 offen offset:809781; CHECK-NEXT:    buffer_load_dword v54, v1, s[0:3], 0 offen offset:769782; CHECK-NEXT:    buffer_load_dword v55, v1, s[0:3], 0 offen offset:729783; CHECK-NEXT:    buffer_load_dword v64, v1, s[0:3], 0 offen offset:689784; CHECK-NEXT:    buffer_load_dword v65, v1, s[0:3], 0 offen offset:649785; CHECK-NEXT:    buffer_load_dword v66, v1, s[0:3], 0 offen offset:609786; CHECK-NEXT:    buffer_load_dword v67, v1, s[0:3], 0 offen offset:569787; CHECK-NEXT:    buffer_load_dword v68, v1, s[0:3], 0 offen offset:529788; CHECK-NEXT:    buffer_load_dword v69, v1, s[0:3], 0 offen offset:489789; CHECK-NEXT:    buffer_load_dword v70, v1, s[0:3], 0 offen offset:449790; CHECK-NEXT:    buffer_load_dword v71, v1, s[0:3], 0 offen offset:409791; CHECK-NEXT:    buffer_load_dword v80, v1, s[0:3], 0 offen offset:369792; CHECK-NEXT:    buffer_load_dword v81, v1, s[0:3], 0 offen offset:329793; CHECK-NEXT:    buffer_load_dword v82, v1, s[0:3], 0 offen offset:289794; CHECK-NEXT:    buffer_load_dword v83, v1, s[0:3], 0 offen offset:249795; CHECK-NEXT:    buffer_load_dword v84, v1, s[0:3], 0 offen offset:209796; CHECK-NEXT:    buffer_load_dword v85, v1, s[0:3], 0 offen offset:169797; CHECK-NEXT:    buffer_load_dword v86, v1, s[0:3], 0 offen offset:129798; CHECK-NEXT:    buffer_load_dword v87, v1, s[0:3], 0 offen offset:89799; CHECK-NEXT:    buffer_load_dword v96, v1, s[0:3], 0 offen offset:49800; CHECK-NEXT:    buffer_load_dword v97, v1, s[0:3], 0 offen9801; CHECK-NEXT:    v_add_nc_u32_e32 v1, 0x100, v19802; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff009803; CHECK-NEXT:    s_addc_u32 s5, s5, -19804; CHECK-NEXT:    s_waitcnt vmcnt(62)9805; CHECK-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:2529806; CHECK-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:2489807; CHECK-NEXT:    s_waitcnt vmcnt(61)9808; CHECK-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:2449809; CHECK-NEXT:    s_waitcnt vmcnt(60)9810; CHECK-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:2409811; CHECK-NEXT:    s_waitcnt vmcnt(59)9812; CHECK-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:2369813; CHECK-NEXT:    s_waitcnt vmcnt(58)9814; CHECK-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:2329815; CHECK-NEXT:    s_waitcnt vmcnt(57)9816; CHECK-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:2289817; CHECK-NEXT:    s_waitcnt vmcnt(56)9818; CHECK-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:2249819; CHECK-NEXT:    s_waitcnt vmcnt(55)9820; CHECK-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:2209821; CHECK-NEXT:    s_waitcnt vmcnt(54)9822; CHECK-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:2169823; CHECK-NEXT:    s_waitcnt vmcnt(53)9824; CHECK-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:2129825; CHECK-NEXT:    s_waitcnt vmcnt(52)9826; CHECK-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:2089827; CHECK-NEXT:    s_waitcnt vmcnt(51)9828; CHECK-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:2049829; CHECK-NEXT:    s_waitcnt vmcnt(50)9830; CHECK-NEXT:    buffer_store_dword v15, v0, s[0:3], 0 offen offset:2009831; CHECK-NEXT:    s_waitcnt vmcnt(49)9832; CHECK-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:1969833; CHECK-NEXT:    s_waitcnt vmcnt(48)9834; CHECK-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:1929835; CHECK-NEXT:    s_waitcnt vmcnt(47)9836; CHECK-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:1889837; CHECK-NEXT:    s_waitcnt vmcnt(46)9838; CHECK-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:1849839; CHECK-NEXT:    s_waitcnt vmcnt(45)9840; CHECK-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen offset:1809841; CHECK-NEXT:    s_waitcnt vmcnt(44)9842; CHECK-NEXT:    buffer_store_dword v21, v0, s[0:3], 0 offen offset:1769843; CHECK-NEXT:    s_waitcnt vmcnt(43)9844; CHECK-NEXT:    buffer_store_dword v22, v0, s[0:3], 0 offen offset:1729845; CHECK-NEXT:    s_waitcnt vmcnt(42)9846; CHECK-NEXT:    buffer_store_dword v23, v0, s[0:3], 0 offen offset:1689847; CHECK-NEXT:    s_waitcnt vmcnt(41)9848; CHECK-NEXT:    buffer_store_dword v24, v0, s[0:3], 0 offen offset:1649849; CHECK-NEXT:    s_waitcnt vmcnt(40)9850; CHECK-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:1609851; CHECK-NEXT:    s_waitcnt vmcnt(39)9852; CHECK-NEXT:    buffer_store_dword v26, v0, s[0:3], 0 offen offset:1569853; CHECK-NEXT:    s_waitcnt vmcnt(38)9854; CHECK-NEXT:    buffer_store_dword v27, v0, s[0:3], 0 offen offset:1529855; CHECK-NEXT:    s_waitcnt vmcnt(37)9856; CHECK-NEXT:    buffer_store_dword v28, v0, s[0:3], 0 offen offset:1489857; CHECK-NEXT:    s_waitcnt vmcnt(36)9858; CHECK-NEXT:    buffer_store_dword v29, v0, s[0:3], 0 offen offset:1449859; CHECK-NEXT:    s_waitcnt vmcnt(35)9860; CHECK-NEXT:    buffer_store_dword v30, v0, s[0:3], 0 offen offset:1409861; CHECK-NEXT:    s_waitcnt vmcnt(34)9862; CHECK-NEXT:    buffer_store_dword v31, v0, s[0:3], 0 offen offset:1369863; CHECK-NEXT:    s_waitcnt vmcnt(33)9864; CHECK-NEXT:    buffer_store_dword v32, v0, s[0:3], 0 offen offset:1329865; CHECK-NEXT:    s_waitcnt vmcnt(32)9866; CHECK-NEXT:    buffer_store_dword v33, v0, s[0:3], 0 offen offset:1289867; CHECK-NEXT:    s_waitcnt vmcnt(31)9868; CHECK-NEXT:    buffer_store_dword v34, v0, s[0:3], 0 offen offset:1249869; CHECK-NEXT:    s_waitcnt vmcnt(30)9870; CHECK-NEXT:    buffer_store_dword v35, v0, s[0:3], 0 offen offset:1209871; CHECK-NEXT:    s_waitcnt vmcnt(29)9872; CHECK-NEXT:    buffer_store_dword v36, v0, s[0:3], 0 offen offset:1169873; CHECK-NEXT:    s_waitcnt vmcnt(28)9874; CHECK-NEXT:    buffer_store_dword v37, v0, s[0:3], 0 offen offset:1129875; CHECK-NEXT:    s_waitcnt vmcnt(27)9876; CHECK-NEXT:    buffer_store_dword v38, v0, s[0:3], 0 offen offset:1089877; CHECK-NEXT:    s_waitcnt vmcnt(26)9878; CHECK-NEXT:    buffer_store_dword v39, v0, s[0:3], 0 offen offset:1049879; CHECK-NEXT:    s_waitcnt vmcnt(25)9880; CHECK-NEXT:    buffer_store_dword v48, v0, s[0:3], 0 offen offset:1009881; CHECK-NEXT:    s_waitcnt vmcnt(24)9882; CHECK-NEXT:    buffer_store_dword v49, v0, s[0:3], 0 offen offset:969883; CHECK-NEXT:    s_waitcnt vmcnt(23)9884; CHECK-NEXT:    buffer_store_dword v50, v0, s[0:3], 0 offen offset:929885; CHECK-NEXT:    s_waitcnt vmcnt(22)9886; CHECK-NEXT:    buffer_store_dword v51, v0, s[0:3], 0 offen offset:889887; CHECK-NEXT:    s_waitcnt vmcnt(21)9888; CHECK-NEXT:    buffer_store_dword v52, v0, s[0:3], 0 offen offset:849889; CHECK-NEXT:    s_waitcnt vmcnt(20)9890; CHECK-NEXT:    buffer_store_dword v53, v0, s[0:3], 0 offen offset:809891; CHECK-NEXT:    s_waitcnt vmcnt(19)9892; CHECK-NEXT:    buffer_store_dword v54, v0, s[0:3], 0 offen offset:769893; CHECK-NEXT:    s_waitcnt vmcnt(18)9894; CHECK-NEXT:    buffer_store_dword v55, v0, s[0:3], 0 offen offset:729895; CHECK-NEXT:    s_waitcnt vmcnt(17)9896; CHECK-NEXT:    buffer_store_dword v64, v0, s[0:3], 0 offen offset:689897; CHECK-NEXT:    s_waitcnt vmcnt(16)9898; CHECK-NEXT:    buffer_store_dword v65, v0, s[0:3], 0 offen offset:649899; CHECK-NEXT:    s_waitcnt vmcnt(15)9900; CHECK-NEXT:    buffer_store_dword v66, v0, s[0:3], 0 offen offset:609901; CHECK-NEXT:    s_waitcnt vmcnt(14)9902; CHECK-NEXT:    buffer_store_dword v67, v0, s[0:3], 0 offen offset:569903; CHECK-NEXT:    s_waitcnt vmcnt(13)9904; CHECK-NEXT:    buffer_store_dword v68, v0, s[0:3], 0 offen offset:529905; CHECK-NEXT:    s_waitcnt vmcnt(12)9906; CHECK-NEXT:    buffer_store_dword v69, v0, s[0:3], 0 offen offset:489907; CHECK-NEXT:    s_waitcnt vmcnt(11)9908; CHECK-NEXT:    buffer_store_dword v70, v0, s[0:3], 0 offen offset:449909; CHECK-NEXT:    s_waitcnt vmcnt(10)9910; CHECK-NEXT:    buffer_store_dword v71, v0, s[0:3], 0 offen offset:409911; CHECK-NEXT:    s_waitcnt vmcnt(9)9912; CHECK-NEXT:    buffer_store_dword v80, v0, s[0:3], 0 offen offset:369913; CHECK-NEXT:    s_waitcnt vmcnt(8)9914; CHECK-NEXT:    buffer_store_dword v81, v0, s[0:3], 0 offen offset:329915; CHECK-NEXT:    s_waitcnt vmcnt(7)9916; CHECK-NEXT:    buffer_store_dword v82, v0, s[0:3], 0 offen offset:289917; CHECK-NEXT:    s_waitcnt vmcnt(6)9918; CHECK-NEXT:    buffer_store_dword v83, v0, s[0:3], 0 offen offset:249919; CHECK-NEXT:    s_waitcnt vmcnt(5)9920; CHECK-NEXT:    buffer_store_dword v84, v0, s[0:3], 0 offen offset:209921; CHECK-NEXT:    s_waitcnt vmcnt(4)9922; CHECK-NEXT:    buffer_store_dword v85, v0, s[0:3], 0 offen offset:169923; CHECK-NEXT:    s_waitcnt vmcnt(3)9924; CHECK-NEXT:    buffer_store_dword v86, v0, s[0:3], 0 offen offset:129925; CHECK-NEXT:    s_waitcnt vmcnt(2)9926; CHECK-NEXT:    buffer_store_dword v87, v0, s[0:3], 0 offen offset:89927; CHECK-NEXT:    s_waitcnt vmcnt(1)9928; CHECK-NEXT:    buffer_store_dword v96, v0, s[0:3], 0 offen offset:49929; CHECK-NEXT:    s_waitcnt vmcnt(0)9930; CHECK-NEXT:    buffer_store_dword v97, v0, s[0:3], 0 offen9931; CHECK-NEXT:    v_add_nc_u32_e32 v0, 0x100, v09932; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 09933; CHECK-NEXT:    s_cbranch_scc1 .LBB8_29934; CHECK-NEXT:  .LBB8_3: ; %Flow189935; CHECK-NEXT:    s_andn2_saveexec_b32 s6, s69936; CHECK-NEXT:    s_cbranch_execz .LBB8_69937; CHECK-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader9938; CHECK-NEXT:    v_add_nc_u32_e32 v1, 0x700, v19939; CHECK-NEXT:    v_add_nc_u32_e32 v0, 0x700, v09940; CHECK-NEXT:    s_movk_i32 s4, 0xf8009941; CHECK-NEXT:    s_mov_b32 s5, -19942; CHECK-NEXT:  .LBB8_5: ; %memmove_bwd_loop9943; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=19944; CHECK-NEXT:    s_clause 0x3e9945; CHECK-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:2529946; CHECK-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:2489947; CHECK-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:2449948; CHECK-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:2409949; CHECK-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:2369950; CHECK-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:2329951; CHECK-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:2289952; CHECK-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:2249953; CHECK-NEXT:    buffer_load_dword v10, v1, s[0:3], 0 offen offset:2209954; CHECK-NEXT:    buffer_load_dword v11, v1, s[0:3], 0 offen offset:2169955; CHECK-NEXT:    buffer_load_dword v12, v1, s[0:3], 0 offen offset:2129956; CHECK-NEXT:    buffer_load_dword v13, v1, s[0:3], 0 offen offset:2089957; CHECK-NEXT:    buffer_load_dword v14, v1, s[0:3], 0 offen offset:2049958; CHECK-NEXT:    buffer_load_dword v15, v1, s[0:3], 0 offen offset:2009959; CHECK-NEXT:    buffer_load_dword v16, v1, s[0:3], 0 offen offset:1969960; CHECK-NEXT:    buffer_load_dword v17, v1, s[0:3], 0 offen offset:1929961; CHECK-NEXT:    buffer_load_dword v18, v1, s[0:3], 0 offen offset:1889962; CHECK-NEXT:    buffer_load_dword v19, v1, s[0:3], 0 offen offset:1849963; CHECK-NEXT:    buffer_load_dword v20, v1, s[0:3], 0 offen offset:1809964; CHECK-NEXT:    buffer_load_dword v21, v1, s[0:3], 0 offen offset:1769965; CHECK-NEXT:    buffer_load_dword v22, v1, s[0:3], 0 offen offset:1729966; CHECK-NEXT:    buffer_load_dword v23, v1, s[0:3], 0 offen offset:1689967; CHECK-NEXT:    buffer_load_dword v24, v1, s[0:3], 0 offen offset:1649968; CHECK-NEXT:    buffer_load_dword v25, v1, s[0:3], 0 offen offset:1609969; CHECK-NEXT:    buffer_load_dword v26, v1, s[0:3], 0 offen offset:1569970; CHECK-NEXT:    buffer_load_dword v27, v1, s[0:3], 0 offen offset:1529971; CHECK-NEXT:    buffer_load_dword v28, v1, s[0:3], 0 offen offset:1489972; CHECK-NEXT:    buffer_load_dword v29, v1, s[0:3], 0 offen offset:1449973; CHECK-NEXT:    buffer_load_dword v30, v1, s[0:3], 0 offen offset:1409974; CHECK-NEXT:    buffer_load_dword v31, v1, s[0:3], 0 offen offset:1369975; CHECK-NEXT:    buffer_load_dword v32, v1, s[0:3], 0 offen offset:1329976; CHECK-NEXT:    buffer_load_dword v33, v1, s[0:3], 0 offen offset:1289977; CHECK-NEXT:    buffer_load_dword v34, v1, s[0:3], 0 offen offset:1249978; CHECK-NEXT:    buffer_load_dword v35, v1, s[0:3], 0 offen offset:1209979; CHECK-NEXT:    buffer_load_dword v36, v1, s[0:3], 0 offen offset:1169980; CHECK-NEXT:    buffer_load_dword v37, v1, s[0:3], 0 offen offset:1129981; CHECK-NEXT:    buffer_load_dword v38, v1, s[0:3], 0 offen offset:1089982; CHECK-NEXT:    buffer_load_dword v39, v1, s[0:3], 0 offen offset:1049983; CHECK-NEXT:    buffer_load_dword v48, v1, s[0:3], 0 offen offset:1009984; CHECK-NEXT:    buffer_load_dword v49, v1, s[0:3], 0 offen offset:969985; CHECK-NEXT:    buffer_load_dword v50, v1, s[0:3], 0 offen offset:929986; CHECK-NEXT:    buffer_load_dword v51, v1, s[0:3], 0 offen offset:889987; CHECK-NEXT:    buffer_load_dword v52, v1, s[0:3], 0 offen offset:849988; CHECK-NEXT:    buffer_load_dword v53, v1, s[0:3], 0 offen offset:809989; CHECK-NEXT:    buffer_load_dword v54, v1, s[0:3], 0 offen offset:769990; CHECK-NEXT:    buffer_load_dword v55, v1, s[0:3], 0 offen offset:729991; CHECK-NEXT:    buffer_load_dword v64, v1, s[0:3], 0 offen offset:689992; CHECK-NEXT:    buffer_load_dword v65, v1, s[0:3], 0 offen offset:649993; CHECK-NEXT:    buffer_load_dword v66, v1, s[0:3], 0 offen offset:609994; CHECK-NEXT:    buffer_load_dword v67, v1, s[0:3], 0 offen offset:569995; CHECK-NEXT:    buffer_load_dword v68, v1, s[0:3], 0 offen offset:529996; CHECK-NEXT:    buffer_load_dword v69, v1, s[0:3], 0 offen offset:489997; CHECK-NEXT:    buffer_load_dword v70, v1, s[0:3], 0 offen offset:449998; CHECK-NEXT:    buffer_load_dword v71, v1, s[0:3], 0 offen offset:409999; CHECK-NEXT:    buffer_load_dword v80, v1, s[0:3], 0 offen offset:3610000; CHECK-NEXT:    buffer_load_dword v81, v1, s[0:3], 0 offen offset:3210001; CHECK-NEXT:    buffer_load_dword v82, v1, s[0:3], 0 offen offset:2810002; CHECK-NEXT:    buffer_load_dword v83, v1, s[0:3], 0 offen offset:2410003; CHECK-NEXT:    buffer_load_dword v84, v1, s[0:3], 0 offen offset:2010004; CHECK-NEXT:    buffer_load_dword v85, v1, s[0:3], 0 offen offset:1610005; CHECK-NEXT:    buffer_load_dword v86, v1, s[0:3], 0 offen offset:1210006; CHECK-NEXT:    buffer_load_dword v87, v1, s[0:3], 0 offen offset:810007; CHECK-NEXT:    buffer_load_dword v96, v1, s[0:3], 0 offen offset:410008; CHECK-NEXT:    buffer_load_dword v97, v1, s[0:3], 0 offen10009; CHECK-NEXT:    v_add_nc_u32_e32 v1, 0xffffff00, v110010; CHECK-NEXT:    s_add_u32 s4, s4, 0x10010011; CHECK-NEXT:    s_addc_u32 s5, s5, 010012; CHECK-NEXT:    s_waitcnt vmcnt(62)10013; CHECK-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:25210014; CHECK-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:24810015; CHECK-NEXT:    s_waitcnt vmcnt(61)10016; CHECK-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:24410017; CHECK-NEXT:    s_waitcnt vmcnt(60)10018; CHECK-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:24010019; CHECK-NEXT:    s_waitcnt vmcnt(59)10020; CHECK-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:23610021; CHECK-NEXT:    s_waitcnt vmcnt(58)10022; CHECK-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:23210023; CHECK-NEXT:    s_waitcnt vmcnt(57)10024; CHECK-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:22810025; CHECK-NEXT:    s_waitcnt vmcnt(56)10026; CHECK-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:22410027; CHECK-NEXT:    s_waitcnt vmcnt(55)10028; CHECK-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:22010029; CHECK-NEXT:    s_waitcnt vmcnt(54)10030; CHECK-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:21610031; CHECK-NEXT:    s_waitcnt vmcnt(53)10032; CHECK-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:21210033; CHECK-NEXT:    s_waitcnt vmcnt(52)10034; CHECK-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:20810035; CHECK-NEXT:    s_waitcnt vmcnt(51)10036; CHECK-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:20410037; CHECK-NEXT:    s_waitcnt vmcnt(50)10038; CHECK-NEXT:    buffer_store_dword v15, v0, s[0:3], 0 offen offset:20010039; CHECK-NEXT:    s_waitcnt vmcnt(49)10040; CHECK-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:19610041; CHECK-NEXT:    s_waitcnt vmcnt(48)10042; CHECK-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:19210043; CHECK-NEXT:    s_waitcnt vmcnt(47)10044; CHECK-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:18810045; CHECK-NEXT:    s_waitcnt vmcnt(46)10046; CHECK-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:18410047; CHECK-NEXT:    s_waitcnt vmcnt(45)10048; CHECK-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen offset:18010049; CHECK-NEXT:    s_waitcnt vmcnt(44)10050; CHECK-NEXT:    buffer_store_dword v21, v0, s[0:3], 0 offen offset:17610051; CHECK-NEXT:    s_waitcnt vmcnt(43)10052; CHECK-NEXT:    buffer_store_dword v22, v0, s[0:3], 0 offen offset:17210053; CHECK-NEXT:    s_waitcnt vmcnt(42)10054; CHECK-NEXT:    buffer_store_dword v23, v0, s[0:3], 0 offen offset:16810055; CHECK-NEXT:    s_waitcnt vmcnt(41)10056; CHECK-NEXT:    buffer_store_dword v24, v0, s[0:3], 0 offen offset:16410057; CHECK-NEXT:    s_waitcnt vmcnt(40)10058; CHECK-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:16010059; CHECK-NEXT:    s_waitcnt vmcnt(39)10060; CHECK-NEXT:    buffer_store_dword v26, v0, s[0:3], 0 offen offset:15610061; CHECK-NEXT:    s_waitcnt vmcnt(38)10062; CHECK-NEXT:    buffer_store_dword v27, v0, s[0:3], 0 offen offset:15210063; CHECK-NEXT:    s_waitcnt vmcnt(37)10064; CHECK-NEXT:    buffer_store_dword v28, v0, s[0:3], 0 offen offset:14810065; CHECK-NEXT:    s_waitcnt vmcnt(36)10066; CHECK-NEXT:    buffer_store_dword v29, v0, s[0:3], 0 offen offset:14410067; CHECK-NEXT:    s_waitcnt vmcnt(35)10068; CHECK-NEXT:    buffer_store_dword v30, v0, s[0:3], 0 offen offset:14010069; CHECK-NEXT:    s_waitcnt vmcnt(34)10070; CHECK-NEXT:    buffer_store_dword v31, v0, s[0:3], 0 offen offset:13610071; CHECK-NEXT:    s_waitcnt vmcnt(33)10072; CHECK-NEXT:    buffer_store_dword v32, v0, s[0:3], 0 offen offset:13210073; CHECK-NEXT:    s_waitcnt vmcnt(32)10074; CHECK-NEXT:    buffer_store_dword v33, v0, s[0:3], 0 offen offset:12810075; CHECK-NEXT:    s_waitcnt vmcnt(31)10076; CHECK-NEXT:    buffer_store_dword v34, v0, s[0:3], 0 offen offset:12410077; CHECK-NEXT:    s_waitcnt vmcnt(30)10078; CHECK-NEXT:    buffer_store_dword v35, v0, s[0:3], 0 offen offset:12010079; CHECK-NEXT:    s_waitcnt vmcnt(29)10080; CHECK-NEXT:    buffer_store_dword v36, v0, s[0:3], 0 offen offset:11610081; CHECK-NEXT:    s_waitcnt vmcnt(28)10082; CHECK-NEXT:    buffer_store_dword v37, v0, s[0:3], 0 offen offset:11210083; CHECK-NEXT:    s_waitcnt vmcnt(27)10084; CHECK-NEXT:    buffer_store_dword v38, v0, s[0:3], 0 offen offset:10810085; CHECK-NEXT:    s_waitcnt vmcnt(26)10086; CHECK-NEXT:    buffer_store_dword v39, v0, s[0:3], 0 offen offset:10410087; CHECK-NEXT:    s_waitcnt vmcnt(25)10088; CHECK-NEXT:    buffer_store_dword v48, v0, s[0:3], 0 offen offset:10010089; CHECK-NEXT:    s_waitcnt vmcnt(24)10090; CHECK-NEXT:    buffer_store_dword v49, v0, s[0:3], 0 offen offset:9610091; CHECK-NEXT:    s_waitcnt vmcnt(23)10092; CHECK-NEXT:    buffer_store_dword v50, v0, s[0:3], 0 offen offset:9210093; CHECK-NEXT:    s_waitcnt vmcnt(22)10094; CHECK-NEXT:    buffer_store_dword v51, v0, s[0:3], 0 offen offset:8810095; CHECK-NEXT:    s_waitcnt vmcnt(21)10096; CHECK-NEXT:    buffer_store_dword v52, v0, s[0:3], 0 offen offset:8410097; CHECK-NEXT:    s_waitcnt vmcnt(20)10098; CHECK-NEXT:    buffer_store_dword v53, v0, s[0:3], 0 offen offset:8010099; CHECK-NEXT:    s_waitcnt vmcnt(19)10100; CHECK-NEXT:    buffer_store_dword v54, v0, s[0:3], 0 offen offset:7610101; CHECK-NEXT:    s_waitcnt vmcnt(18)10102; CHECK-NEXT:    buffer_store_dword v55, v0, s[0:3], 0 offen offset:7210103; CHECK-NEXT:    s_waitcnt vmcnt(17)10104; CHECK-NEXT:    buffer_store_dword v64, v0, s[0:3], 0 offen offset:6810105; CHECK-NEXT:    s_waitcnt vmcnt(16)10106; CHECK-NEXT:    buffer_store_dword v65, v0, s[0:3], 0 offen offset:6410107; CHECK-NEXT:    s_waitcnt vmcnt(15)10108; CHECK-NEXT:    buffer_store_dword v66, v0, s[0:3], 0 offen offset:6010109; CHECK-NEXT:    s_waitcnt vmcnt(14)10110; CHECK-NEXT:    buffer_store_dword v67, v0, s[0:3], 0 offen offset:5610111; CHECK-NEXT:    s_waitcnt vmcnt(13)10112; CHECK-NEXT:    buffer_store_dword v68, v0, s[0:3], 0 offen offset:5210113; CHECK-NEXT:    s_waitcnt vmcnt(12)10114; CHECK-NEXT:    buffer_store_dword v69, v0, s[0:3], 0 offen offset:4810115; CHECK-NEXT:    s_waitcnt vmcnt(11)10116; CHECK-NEXT:    buffer_store_dword v70, v0, s[0:3], 0 offen offset:4410117; CHECK-NEXT:    s_waitcnt vmcnt(10)10118; CHECK-NEXT:    buffer_store_dword v71, v0, s[0:3], 0 offen offset:4010119; CHECK-NEXT:    s_waitcnt vmcnt(9)10120; CHECK-NEXT:    buffer_store_dword v80, v0, s[0:3], 0 offen offset:3610121; CHECK-NEXT:    s_waitcnt vmcnt(8)10122; CHECK-NEXT:    buffer_store_dword v81, v0, s[0:3], 0 offen offset:3210123; CHECK-NEXT:    s_waitcnt vmcnt(7)10124; CHECK-NEXT:    buffer_store_dword v82, v0, s[0:3], 0 offen offset:2810125; CHECK-NEXT:    s_waitcnt vmcnt(6)10126; CHECK-NEXT:    buffer_store_dword v83, v0, s[0:3], 0 offen offset:2410127; CHECK-NEXT:    s_waitcnt vmcnt(5)10128; CHECK-NEXT:    buffer_store_dword v84, v0, s[0:3], 0 offen offset:2010129; CHECK-NEXT:    s_waitcnt vmcnt(4)10130; CHECK-NEXT:    buffer_store_dword v85, v0, s[0:3], 0 offen offset:1610131; CHECK-NEXT:    s_waitcnt vmcnt(3)10132; CHECK-NEXT:    buffer_store_dword v86, v0, s[0:3], 0 offen offset:1210133; CHECK-NEXT:    s_waitcnt vmcnt(2)10134; CHECK-NEXT:    buffer_store_dword v87, v0, s[0:3], 0 offen offset:810135; CHECK-NEXT:    s_waitcnt vmcnt(1)10136; CHECK-NEXT:    buffer_store_dword v96, v0, s[0:3], 0 offen offset:410137; CHECK-NEXT:    s_waitcnt vmcnt(0)10138; CHECK-NEXT:    buffer_store_dword v97, v0, s[0:3], 0 offen10139; CHECK-NEXT:    v_add_nc_u32_e32 v0, 0xffffff00, v010140; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], 010141; CHECK-NEXT:    s_cbranch_scc0 .LBB8_510142; CHECK-NEXT:  .LBB8_6: ; %Flow1910143; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s610144; CHECK-NEXT:    s_setpc_b64 s[30:31]10145;10146; ALIGNED-LABEL: memmove_p5_p5_sz2048:10147; ALIGNED:       ; %bb.0: ; %entry10148; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)10149; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill10150; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill10151; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill10152; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill10153; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill10154; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill10155; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill10156; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill10157; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill10158; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill10159; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill10160; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill10161; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill10162; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill10163; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill10164; ALIGNED-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill10165; ALIGNED-NEXT:    buffer_store_dword v72, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill10166; ALIGNED-NEXT:    buffer_store_dword v73, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill10167; ALIGNED-NEXT:    buffer_store_dword v74, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill10168; ALIGNED-NEXT:    buffer_store_dword v75, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill10169; ALIGNED-NEXT:    buffer_store_dword v76, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill10170; ALIGNED-NEXT:    buffer_store_dword v77, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill10171; ALIGNED-NEXT:    buffer_store_dword v78, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill10172; ALIGNED-NEXT:    buffer_store_dword v79, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill10173; ALIGNED-NEXT:    buffer_store_dword v88, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill10174; ALIGNED-NEXT:    buffer_store_dword v89, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill10175; ALIGNED-NEXT:    buffer_store_dword v90, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill10176; ALIGNED-NEXT:    buffer_store_dword v91, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill10177; ALIGNED-NEXT:    buffer_store_dword v92, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill10178; ALIGNED-NEXT:    buffer_store_dword v93, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill10179; ALIGNED-NEXT:    buffer_store_dword v94, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill10180; ALIGNED-NEXT:    buffer_store_dword v95, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill10181; ALIGNED-NEXT:    buffer_store_dword v104, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill10182; ALIGNED-NEXT:    buffer_store_dword v105, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill10183; ALIGNED-NEXT:    buffer_store_dword v106, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill10184; ALIGNED-NEXT:    buffer_store_dword v107, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill10185; ALIGNED-NEXT:    buffer_store_dword v108, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill10186; ALIGNED-NEXT:    buffer_store_dword v109, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill10187; ALIGNED-NEXT:    buffer_store_dword v110, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill10188; ALIGNED-NEXT:    buffer_store_dword v111, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill10189; ALIGNED-NEXT:    buffer_store_dword v120, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill10190; ALIGNED-NEXT:    buffer_store_dword v121, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill10191; ALIGNED-NEXT:    buffer_store_dword v122, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill10192; ALIGNED-NEXT:    buffer_store_dword v123, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill10193; ALIGNED-NEXT:    buffer_store_dword v124, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill10194; ALIGNED-NEXT:    buffer_store_dword v125, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill10195; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill10196; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 ; 4-byte Folded Spill10197; ALIGNED-NEXT:    s_mov_b32 s4, exec_lo10198; ALIGNED-NEXT:    v_cmpx_ge_u32_e64 v1, v010199; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s410200; ALIGNED-NEXT:    s_cbranch_execz .LBB8_310201; ALIGNED-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader10202; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x80010203; ALIGNED-NEXT:  .LBB8_2: ; %memmove_fwd_loop10204; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=110205; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25510206; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff0010207; ALIGNED-NEXT:    s_addc_u32 s5, s5, -110208; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 010209; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10210; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill10211; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25410212; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10213; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill10214; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25310215; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10216; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill10217; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25210218; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10219; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill10220; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25110221; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10222; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill10223; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25010224; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10225; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill10226; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24910227; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10228; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill10229; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24810230; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10231; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill10232; ALIGNED-NEXT:    s_clause 0x3e10233; ALIGNED-NEXT:    buffer_load_ubyte v123, v1, s[0:3], 0 offen offset:24710234; ALIGNED-NEXT:    buffer_load_ubyte v122, v1, s[0:3], 0 offen offset:24610235; ALIGNED-NEXT:    buffer_load_ubyte v121, v1, s[0:3], 0 offen offset:24510236; ALIGNED-NEXT:    buffer_load_ubyte v120, v1, s[0:3], 0 offen offset:24410237; ALIGNED-NEXT:    buffer_load_ubyte v111, v1, s[0:3], 0 offen offset:24310238; ALIGNED-NEXT:    buffer_load_ubyte v110, v1, s[0:3], 0 offen offset:24210239; ALIGNED-NEXT:    buffer_load_ubyte v109, v1, s[0:3], 0 offen offset:24110240; ALIGNED-NEXT:    buffer_load_ubyte v108, v1, s[0:3], 0 offen offset:24010241; ALIGNED-NEXT:    buffer_load_ubyte v107, v1, s[0:3], 0 offen offset:23910242; ALIGNED-NEXT:    buffer_load_ubyte v106, v1, s[0:3], 0 offen offset:23810243; ALIGNED-NEXT:    buffer_load_ubyte v105, v1, s[0:3], 0 offen offset:23710244; ALIGNED-NEXT:    buffer_load_ubyte v104, v1, s[0:3], 0 offen offset:23610245; ALIGNED-NEXT:    buffer_load_ubyte v95, v1, s[0:3], 0 offen offset:23510246; ALIGNED-NEXT:    buffer_load_ubyte v94, v1, s[0:3], 0 offen offset:23410247; ALIGNED-NEXT:    buffer_load_ubyte v93, v1, s[0:3], 0 offen offset:23310248; ALIGNED-NEXT:    buffer_load_ubyte v92, v1, s[0:3], 0 offen offset:23210249; ALIGNED-NEXT:    buffer_load_ubyte v91, v1, s[0:3], 0 offen offset:23110250; ALIGNED-NEXT:    buffer_load_ubyte v90, v1, s[0:3], 0 offen offset:23010251; ALIGNED-NEXT:    buffer_load_ubyte v89, v1, s[0:3], 0 offen offset:22910252; ALIGNED-NEXT:    buffer_load_ubyte v88, v1, s[0:3], 0 offen offset:22810253; ALIGNED-NEXT:    buffer_load_ubyte v79, v1, s[0:3], 0 offen offset:22710254; ALIGNED-NEXT:    buffer_load_ubyte v78, v1, s[0:3], 0 offen offset:22610255; ALIGNED-NEXT:    buffer_load_ubyte v77, v1, s[0:3], 0 offen offset:22510256; ALIGNED-NEXT:    buffer_load_ubyte v76, v1, s[0:3], 0 offen offset:22410257; ALIGNED-NEXT:    buffer_load_ubyte v75, v1, s[0:3], 0 offen offset:22310258; ALIGNED-NEXT:    buffer_load_ubyte v74, v1, s[0:3], 0 offen offset:22210259; ALIGNED-NEXT:    buffer_load_ubyte v73, v1, s[0:3], 0 offen offset:22110260; ALIGNED-NEXT:    buffer_load_ubyte v72, v1, s[0:3], 0 offen offset:22010261; ALIGNED-NEXT:    buffer_load_ubyte v63, v1, s[0:3], 0 offen offset:21910262; ALIGNED-NEXT:    buffer_load_ubyte v62, v1, s[0:3], 0 offen offset:21810263; ALIGNED-NEXT:    buffer_load_ubyte v61, v1, s[0:3], 0 offen offset:21710264; ALIGNED-NEXT:    buffer_load_ubyte v60, v1, s[0:3], 0 offen offset:21610265; ALIGNED-NEXT:    buffer_load_ubyte v59, v1, s[0:3], 0 offen offset:21510266; ALIGNED-NEXT:    buffer_load_ubyte v58, v1, s[0:3], 0 offen offset:21410267; ALIGNED-NEXT:    buffer_load_ubyte v57, v1, s[0:3], 0 offen offset:21310268; ALIGNED-NEXT:    buffer_load_ubyte v56, v1, s[0:3], 0 offen offset:21210269; ALIGNED-NEXT:    buffer_load_ubyte v47, v1, s[0:3], 0 offen offset:21110270; ALIGNED-NEXT:    buffer_load_ubyte v46, v1, s[0:3], 0 offen offset:21010271; ALIGNED-NEXT:    buffer_load_ubyte v45, v1, s[0:3], 0 offen offset:20910272; ALIGNED-NEXT:    buffer_load_ubyte v44, v1, s[0:3], 0 offen offset:20810273; ALIGNED-NEXT:    buffer_load_ubyte v43, v1, s[0:3], 0 offen offset:20710274; ALIGNED-NEXT:    buffer_load_ubyte v42, v1, s[0:3], 0 offen offset:20610275; ALIGNED-NEXT:    buffer_load_ubyte v41, v1, s[0:3], 0 offen offset:20510276; ALIGNED-NEXT:    buffer_load_ubyte v40, v1, s[0:3], 0 offen offset:20410277; ALIGNED-NEXT:    buffer_load_ubyte v119, v1, s[0:3], 0 offen offset:20310278; ALIGNED-NEXT:    buffer_load_ubyte v118, v1, s[0:3], 0 offen offset:20210279; ALIGNED-NEXT:    buffer_load_ubyte v117, v1, s[0:3], 0 offen offset:20110280; ALIGNED-NEXT:    buffer_load_ubyte v116, v1, s[0:3], 0 offen offset:20010281; ALIGNED-NEXT:    buffer_load_ubyte v115, v1, s[0:3], 0 offen offset:19910282; ALIGNED-NEXT:    buffer_load_ubyte v114, v1, s[0:3], 0 offen offset:19810283; ALIGNED-NEXT:    buffer_load_ubyte v113, v1, s[0:3], 0 offen offset:19710284; ALIGNED-NEXT:    buffer_load_ubyte v112, v1, s[0:3], 0 offen offset:19610285; ALIGNED-NEXT:    buffer_load_ubyte v103, v1, s[0:3], 0 offen offset:19510286; ALIGNED-NEXT:    buffer_load_ubyte v102, v1, s[0:3], 0 offen offset:19410287; ALIGNED-NEXT:    buffer_load_ubyte v101, v1, s[0:3], 0 offen offset:19310288; ALIGNED-NEXT:    buffer_load_ubyte v100, v1, s[0:3], 0 offen offset:19210289; ALIGNED-NEXT:    buffer_load_ubyte v99, v1, s[0:3], 0 offen offset:19110290; ALIGNED-NEXT:    buffer_load_ubyte v98, v1, s[0:3], 0 offen offset:19010291; ALIGNED-NEXT:    buffer_load_ubyte v97, v1, s[0:3], 0 offen offset:18910292; ALIGNED-NEXT:    buffer_load_ubyte v96, v1, s[0:3], 0 offen offset:18810293; ALIGNED-NEXT:    buffer_load_ubyte v87, v1, s[0:3], 0 offen offset:18710294; ALIGNED-NEXT:    buffer_load_ubyte v86, v1, s[0:3], 0 offen offset:18610295; ALIGNED-NEXT:    buffer_load_ubyte v85, v1, s[0:3], 0 offen offset:18510296; ALIGNED-NEXT:    s_clause 0x3a10297; ALIGNED-NEXT:    buffer_load_ubyte v84, v1, s[0:3], 0 offen offset:18410298; ALIGNED-NEXT:    buffer_load_ubyte v83, v1, s[0:3], 0 offen offset:18310299; ALIGNED-NEXT:    buffer_load_ubyte v82, v1, s[0:3], 0 offen offset:18210300; ALIGNED-NEXT:    buffer_load_ubyte v81, v1, s[0:3], 0 offen offset:18110301; ALIGNED-NEXT:    buffer_load_ubyte v80, v1, s[0:3], 0 offen offset:18010302; ALIGNED-NEXT:    buffer_load_ubyte v71, v1, s[0:3], 0 offen offset:17910303; ALIGNED-NEXT:    buffer_load_ubyte v70, v1, s[0:3], 0 offen offset:17810304; ALIGNED-NEXT:    buffer_load_ubyte v69, v1, s[0:3], 0 offen offset:17710305; ALIGNED-NEXT:    buffer_load_ubyte v68, v1, s[0:3], 0 offen offset:17610306; ALIGNED-NEXT:    buffer_load_ubyte v67, v1, s[0:3], 0 offen offset:17510307; ALIGNED-NEXT:    buffer_load_ubyte v66, v1, s[0:3], 0 offen offset:17410308; ALIGNED-NEXT:    buffer_load_ubyte v65, v1, s[0:3], 0 offen offset:17310309; ALIGNED-NEXT:    buffer_load_ubyte v64, v1, s[0:3], 0 offen offset:17210310; ALIGNED-NEXT:    buffer_load_ubyte v55, v1, s[0:3], 0 offen offset:17110311; ALIGNED-NEXT:    buffer_load_ubyte v54, v1, s[0:3], 0 offen offset:17010312; ALIGNED-NEXT:    buffer_load_ubyte v53, v1, s[0:3], 0 offen offset:16910313; ALIGNED-NEXT:    buffer_load_ubyte v52, v1, s[0:3], 0 offen offset:16810314; ALIGNED-NEXT:    buffer_load_ubyte v51, v1, s[0:3], 0 offen offset:16710315; ALIGNED-NEXT:    buffer_load_ubyte v50, v1, s[0:3], 0 offen offset:16610316; ALIGNED-NEXT:    buffer_load_ubyte v49, v1, s[0:3], 0 offen offset:16510317; ALIGNED-NEXT:    buffer_load_ubyte v48, v1, s[0:3], 0 offen offset:16410318; ALIGNED-NEXT:    buffer_load_ubyte v39, v1, s[0:3], 0 offen offset:16310319; ALIGNED-NEXT:    buffer_load_ubyte v38, v1, s[0:3], 0 offen offset:16210320; ALIGNED-NEXT:    buffer_load_ubyte v37, v1, s[0:3], 0 offen offset:16110321; ALIGNED-NEXT:    buffer_load_ubyte v36, v1, s[0:3], 0 offen offset:16010322; ALIGNED-NEXT:    buffer_load_ubyte v35, v1, s[0:3], 0 offen offset:15910323; ALIGNED-NEXT:    buffer_load_ubyte v34, v1, s[0:3], 0 offen offset:15810324; ALIGNED-NEXT:    buffer_load_ubyte v33, v1, s[0:3], 0 offen offset:15710325; ALIGNED-NEXT:    buffer_load_ubyte v32, v1, s[0:3], 0 offen offset:15610326; ALIGNED-NEXT:    buffer_load_ubyte v31, v1, s[0:3], 0 offen offset:15510327; ALIGNED-NEXT:    buffer_load_ubyte v30, v1, s[0:3], 0 offen offset:15410328; ALIGNED-NEXT:    buffer_load_ubyte v29, v1, s[0:3], 0 offen offset:15310329; ALIGNED-NEXT:    buffer_load_ubyte v28, v1, s[0:3], 0 offen offset:15210330; ALIGNED-NEXT:    buffer_load_ubyte v27, v1, s[0:3], 0 offen offset:15110331; ALIGNED-NEXT:    buffer_load_ubyte v26, v1, s[0:3], 0 offen offset:15010332; ALIGNED-NEXT:    buffer_load_ubyte v25, v1, s[0:3], 0 offen offset:14910333; ALIGNED-NEXT:    buffer_load_ubyte v24, v1, s[0:3], 0 offen offset:14810334; ALIGNED-NEXT:    buffer_load_ubyte v23, v1, s[0:3], 0 offen offset:14710335; ALIGNED-NEXT:    buffer_load_ubyte v22, v1, s[0:3], 0 offen offset:14610336; ALIGNED-NEXT:    buffer_load_ubyte v21, v1, s[0:3], 0 offen offset:14510337; ALIGNED-NEXT:    buffer_load_ubyte v20, v1, s[0:3], 0 offen offset:14410338; ALIGNED-NEXT:    buffer_load_ubyte v19, v1, s[0:3], 0 offen offset:14310339; ALIGNED-NEXT:    buffer_load_ubyte v18, v1, s[0:3], 0 offen offset:14210340; ALIGNED-NEXT:    buffer_load_ubyte v17, v1, s[0:3], 0 offen offset:14110341; ALIGNED-NEXT:    buffer_load_ubyte v16, v1, s[0:3], 0 offen offset:14010342; ALIGNED-NEXT:    buffer_load_ubyte v15, v1, s[0:3], 0 offen offset:13910343; ALIGNED-NEXT:    buffer_load_ubyte v14, v1, s[0:3], 0 offen offset:13810344; ALIGNED-NEXT:    buffer_load_ubyte v13, v1, s[0:3], 0 offen offset:13710345; ALIGNED-NEXT:    buffer_load_ubyte v12, v1, s[0:3], 0 offen offset:13610346; ALIGNED-NEXT:    buffer_load_ubyte v11, v1, s[0:3], 0 offen offset:13510347; ALIGNED-NEXT:    buffer_load_ubyte v10, v1, s[0:3], 0 offen offset:13410348; ALIGNED-NEXT:    buffer_load_ubyte v9, v1, s[0:3], 0 offen offset:13310349; ALIGNED-NEXT:    buffer_load_ubyte v8, v1, s[0:3], 0 offen offset:13210350; ALIGNED-NEXT:    buffer_load_ubyte v7, v1, s[0:3], 0 offen offset:13110351; ALIGNED-NEXT:    buffer_load_ubyte v6, v1, s[0:3], 0 offen offset:13010352; ALIGNED-NEXT:    buffer_load_ubyte v5, v1, s[0:3], 0 offen offset:12910353; ALIGNED-NEXT:    buffer_load_ubyte v4, v1, s[0:3], 0 offen offset:12810354; ALIGNED-NEXT:    buffer_load_ubyte v3, v1, s[0:3], 0 offen offset:12710355; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:12610356; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10357; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill10358; ALIGNED-NEXT:    s_clause 0x410359; ALIGNED-NEXT:    buffer_load_ubyte v127, v1, s[0:3], 0 offen offset:12510360; ALIGNED-NEXT:    buffer_load_ubyte v126, v1, s[0:3], 0 offen offset:12410361; ALIGNED-NEXT:    buffer_load_ubyte v125, v1, s[0:3], 0 offen offset:12310362; ALIGNED-NEXT:    buffer_load_ubyte v124, v1, s[0:3], 0 offen offset:12210363; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:12110364; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10365; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill10366; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:12010367; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10368; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill10369; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11910370; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10371; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill10372; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11810373; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10374; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill10375; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11710376; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10377; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill10378; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11610379; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10380; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill10381; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11510382; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10383; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill10384; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11410385; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10386; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill10387; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11310388; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10389; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill10390; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11210391; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10392; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill10393; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11110394; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10395; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill10396; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11010397; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10398; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill10399; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10910400; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10401; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill10402; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10810403; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10404; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill10405; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10710406; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10407; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill10408; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10610409; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10410; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill10411; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10510412; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10413; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill10414; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10410415; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10416; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill10417; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10310418; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10419; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill10420; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10210421; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10422; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill10423; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10110424; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10425; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill10426; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10010427; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10428; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill10429; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9910430; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10431; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill10432; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9810433; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10434; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill10435; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9710436; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10437; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill10438; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9610439; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10440; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill10441; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9510442; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10443; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill10444; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9410445; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10446; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill10447; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9310448; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10449; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill10450; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9210451; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10452; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill10453; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9110454; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10455; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill10456; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9010457; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10458; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill10459; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8910460; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10461; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill10462; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8810463; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10464; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill10465; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8710466; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10467; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill10468; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8610469; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10470; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill10471; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8510472; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10473; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill10474; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8410475; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10476; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill10477; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8310478; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10479; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill10480; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8210481; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10482; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill10483; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8110484; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10485; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill10486; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8010487; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10488; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill10489; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7910490; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10491; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill10492; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7810493; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10494; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill10495; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7710496; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10497; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill10498; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7610499; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10500; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill10501; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7510502; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10503; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill10504; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7410505; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10506; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill10507; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7310508; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10509; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill10510; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7210511; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10512; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill10513; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7110514; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10515; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill10516; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7010517; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10518; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill10519; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6910520; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10521; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill10522; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6810523; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10524; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill10525; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6710526; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10527; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill10528; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6610529; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10530; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill10531; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6510532; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10533; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill10534; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6410535; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10536; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill10537; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6310538; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10539; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill10540; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6210541; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10542; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill10543; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6110544; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10545; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill10546; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6010547; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10548; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill10549; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5910550; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10551; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill10552; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5810553; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10554; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill10555; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5710556; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10557; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill10558; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5610559; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10560; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill10561; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5510562; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10563; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill10564; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5410565; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10566; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill10567; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5310568; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10569; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill10570; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5210571; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10572; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill10573; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5110574; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10575; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill10576; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5010577; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10578; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill10579; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4910580; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10581; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill10582; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4810583; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10584; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill10585; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4710586; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10587; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill10588; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4610589; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10590; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill10591; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4510592; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10593; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill10594; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4410595; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10596; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill10597; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4310598; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10599; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill10600; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4210601; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10602; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill10603; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4110604; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10605; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill10606; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4010607; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10608; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill10609; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3910610; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10611; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill10612; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3810613; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10614; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill10615; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3710616; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10617; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill10618; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3610619; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10620; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill10621; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3510622; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10623; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill10624; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3410625; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10626; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill10627; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3310628; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10629; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill10630; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3210631; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10632; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill10633; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3110634; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10635; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill10636; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3010637; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10638; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill10639; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2910640; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10641; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill10642; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2810643; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10644; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill10645; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2710646; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10647; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill10648; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2610649; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10650; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill10651; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2510652; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10653; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill10654; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2410655; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10656; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill10657; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2310658; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10659; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill10660; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2210661; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10662; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill10663; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2110664; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10665; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill10666; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2010667; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10668; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill10669; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1910670; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10671; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill10672; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1810673; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10674; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill10675; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1710676; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10677; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill10678; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1610679; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10680; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill10681; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1510682; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10683; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill10684; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1410685; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10686; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill10687; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1310688; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10689; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill10690; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1210691; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10692; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill10693; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1110694; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10695; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill10696; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1010697; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10698; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill10699; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:910700; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10701; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill10702; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:810703; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10704; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill10705; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:710706; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10707; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill10708; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:610709; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10710; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill10711; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:510712; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10713; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill10714; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:410715; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10716; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill10717; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:310718; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10719; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill10720; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:210721; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10722; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill10723; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:110724; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10725; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill10726; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen10727; ALIGNED-NEXT:    v_add_nc_u32_e32 v1, 0x100, v110728; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10729; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill10730; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload10731; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10732; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25510733; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload10734; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10735; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25410736; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload10737; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10738; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25310739; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload10740; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10741; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25210742; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload10743; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10744; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25110745; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload10746; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10747; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25010748; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload10749; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10750; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24910751; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload10752; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10753; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24810754; ALIGNED-NEXT:    buffer_store_byte v123, v0, s[0:3], 0 offen offset:24710755; ALIGNED-NEXT:    buffer_store_byte v122, v0, s[0:3], 0 offen offset:24610756; ALIGNED-NEXT:    buffer_store_byte v121, v0, s[0:3], 0 offen offset:24510757; ALIGNED-NEXT:    buffer_store_byte v120, v0, s[0:3], 0 offen offset:24410758; ALIGNED-NEXT:    buffer_store_byte v111, v0, s[0:3], 0 offen offset:24310759; ALIGNED-NEXT:    buffer_store_byte v110, v0, s[0:3], 0 offen offset:24210760; ALIGNED-NEXT:    buffer_store_byte v109, v0, s[0:3], 0 offen offset:24110761; ALIGNED-NEXT:    buffer_store_byte v108, v0, s[0:3], 0 offen offset:24010762; ALIGNED-NEXT:    buffer_store_byte v107, v0, s[0:3], 0 offen offset:23910763; ALIGNED-NEXT:    buffer_store_byte v106, v0, s[0:3], 0 offen offset:23810764; ALIGNED-NEXT:    buffer_store_byte v105, v0, s[0:3], 0 offen offset:23710765; ALIGNED-NEXT:    buffer_store_byte v104, v0, s[0:3], 0 offen offset:23610766; ALIGNED-NEXT:    buffer_store_byte v95, v0, s[0:3], 0 offen offset:23510767; ALIGNED-NEXT:    buffer_store_byte v94, v0, s[0:3], 0 offen offset:23410768; ALIGNED-NEXT:    buffer_store_byte v93, v0, s[0:3], 0 offen offset:23310769; ALIGNED-NEXT:    buffer_store_byte v92, v0, s[0:3], 0 offen offset:23210770; ALIGNED-NEXT:    buffer_store_byte v91, v0, s[0:3], 0 offen offset:23110771; ALIGNED-NEXT:    buffer_store_byte v90, v0, s[0:3], 0 offen offset:23010772; ALIGNED-NEXT:    buffer_store_byte v89, v0, s[0:3], 0 offen offset:22910773; ALIGNED-NEXT:    buffer_store_byte v88, v0, s[0:3], 0 offen offset:22810774; ALIGNED-NEXT:    buffer_store_byte v79, v0, s[0:3], 0 offen offset:22710775; ALIGNED-NEXT:    buffer_store_byte v78, v0, s[0:3], 0 offen offset:22610776; ALIGNED-NEXT:    buffer_store_byte v77, v0, s[0:3], 0 offen offset:22510777; ALIGNED-NEXT:    buffer_store_byte v76, v0, s[0:3], 0 offen offset:22410778; ALIGNED-NEXT:    buffer_store_byte v75, v0, s[0:3], 0 offen offset:22310779; ALIGNED-NEXT:    buffer_store_byte v74, v0, s[0:3], 0 offen offset:22210780; ALIGNED-NEXT:    buffer_store_byte v73, v0, s[0:3], 0 offen offset:22110781; ALIGNED-NEXT:    buffer_store_byte v72, v0, s[0:3], 0 offen offset:22010782; ALIGNED-NEXT:    buffer_store_byte v63, v0, s[0:3], 0 offen offset:21910783; ALIGNED-NEXT:    buffer_store_byte v62, v0, s[0:3], 0 offen offset:21810784; ALIGNED-NEXT:    buffer_store_byte v61, v0, s[0:3], 0 offen offset:21710785; ALIGNED-NEXT:    buffer_store_byte v60, v0, s[0:3], 0 offen offset:21610786; ALIGNED-NEXT:    buffer_store_byte v59, v0, s[0:3], 0 offen offset:21510787; ALIGNED-NEXT:    buffer_store_byte v58, v0, s[0:3], 0 offen offset:21410788; ALIGNED-NEXT:    buffer_store_byte v57, v0, s[0:3], 0 offen offset:21310789; ALIGNED-NEXT:    buffer_store_byte v56, v0, s[0:3], 0 offen offset:21210790; ALIGNED-NEXT:    buffer_store_byte v47, v0, s[0:3], 0 offen offset:21110791; ALIGNED-NEXT:    buffer_store_byte v46, v0, s[0:3], 0 offen offset:21010792; ALIGNED-NEXT:    buffer_store_byte v45, v0, s[0:3], 0 offen offset:20910793; ALIGNED-NEXT:    buffer_store_byte v44, v0, s[0:3], 0 offen offset:20810794; ALIGNED-NEXT:    buffer_store_byte v43, v0, s[0:3], 0 offen offset:20710795; ALIGNED-NEXT:    buffer_store_byte v42, v0, s[0:3], 0 offen offset:20610796; ALIGNED-NEXT:    buffer_store_byte v41, v0, s[0:3], 0 offen offset:20510797; ALIGNED-NEXT:    buffer_store_byte v40, v0, s[0:3], 0 offen offset:20410798; ALIGNED-NEXT:    buffer_store_byte v119, v0, s[0:3], 0 offen offset:20310799; ALIGNED-NEXT:    buffer_store_byte v118, v0, s[0:3], 0 offen offset:20210800; ALIGNED-NEXT:    buffer_store_byte v117, v0, s[0:3], 0 offen offset:20110801; ALIGNED-NEXT:    buffer_store_byte v116, v0, s[0:3], 0 offen offset:20010802; ALIGNED-NEXT:    buffer_store_byte v115, v0, s[0:3], 0 offen offset:19910803; ALIGNED-NEXT:    buffer_store_byte v114, v0, s[0:3], 0 offen offset:19810804; ALIGNED-NEXT:    buffer_store_byte v113, v0, s[0:3], 0 offen offset:19710805; ALIGNED-NEXT:    buffer_store_byte v112, v0, s[0:3], 0 offen offset:19610806; ALIGNED-NEXT:    buffer_store_byte v103, v0, s[0:3], 0 offen offset:19510807; ALIGNED-NEXT:    buffer_store_byte v102, v0, s[0:3], 0 offen offset:19410808; ALIGNED-NEXT:    buffer_store_byte v101, v0, s[0:3], 0 offen offset:19310809; ALIGNED-NEXT:    buffer_store_byte v100, v0, s[0:3], 0 offen offset:19210810; ALIGNED-NEXT:    buffer_store_byte v99, v0, s[0:3], 0 offen offset:19110811; ALIGNED-NEXT:    buffer_store_byte v98, v0, s[0:3], 0 offen offset:19010812; ALIGNED-NEXT:    buffer_store_byte v97, v0, s[0:3], 0 offen offset:18910813; ALIGNED-NEXT:    buffer_store_byte v96, v0, s[0:3], 0 offen offset:18810814; ALIGNED-NEXT:    buffer_store_byte v87, v0, s[0:3], 0 offen offset:18710815; ALIGNED-NEXT:    buffer_store_byte v86, v0, s[0:3], 0 offen offset:18610816; ALIGNED-NEXT:    buffer_store_byte v85, v0, s[0:3], 0 offen offset:18510817; ALIGNED-NEXT:    buffer_store_byte v84, v0, s[0:3], 0 offen offset:18410818; ALIGNED-NEXT:    buffer_store_byte v83, v0, s[0:3], 0 offen offset:18310819; ALIGNED-NEXT:    buffer_store_byte v82, v0, s[0:3], 0 offen offset:18210820; ALIGNED-NEXT:    buffer_store_byte v81, v0, s[0:3], 0 offen offset:18110821; ALIGNED-NEXT:    buffer_store_byte v80, v0, s[0:3], 0 offen offset:18010822; ALIGNED-NEXT:    buffer_store_byte v71, v0, s[0:3], 0 offen offset:17910823; ALIGNED-NEXT:    buffer_store_byte v70, v0, s[0:3], 0 offen offset:17810824; ALIGNED-NEXT:    buffer_store_byte v69, v0, s[0:3], 0 offen offset:17710825; ALIGNED-NEXT:    buffer_store_byte v68, v0, s[0:3], 0 offen offset:17610826; ALIGNED-NEXT:    buffer_store_byte v67, v0, s[0:3], 0 offen offset:17510827; ALIGNED-NEXT:    buffer_store_byte v66, v0, s[0:3], 0 offen offset:17410828; ALIGNED-NEXT:    buffer_store_byte v65, v0, s[0:3], 0 offen offset:17310829; ALIGNED-NEXT:    buffer_store_byte v64, v0, s[0:3], 0 offen offset:17210830; ALIGNED-NEXT:    buffer_store_byte v55, v0, s[0:3], 0 offen offset:17110831; ALIGNED-NEXT:    buffer_store_byte v54, v0, s[0:3], 0 offen offset:17010832; ALIGNED-NEXT:    buffer_store_byte v53, v0, s[0:3], 0 offen offset:16910833; ALIGNED-NEXT:    buffer_store_byte v52, v0, s[0:3], 0 offen offset:16810834; ALIGNED-NEXT:    buffer_store_byte v51, v0, s[0:3], 0 offen offset:16710835; ALIGNED-NEXT:    buffer_store_byte v50, v0, s[0:3], 0 offen offset:16610836; ALIGNED-NEXT:    buffer_store_byte v49, v0, s[0:3], 0 offen offset:16510837; ALIGNED-NEXT:    buffer_store_byte v48, v0, s[0:3], 0 offen offset:16410838; ALIGNED-NEXT:    buffer_store_byte v39, v0, s[0:3], 0 offen offset:16310839; ALIGNED-NEXT:    buffer_store_byte v38, v0, s[0:3], 0 offen offset:16210840; ALIGNED-NEXT:    buffer_store_byte v37, v0, s[0:3], 0 offen offset:16110841; ALIGNED-NEXT:    buffer_store_byte v36, v0, s[0:3], 0 offen offset:16010842; ALIGNED-NEXT:    buffer_store_byte v35, v0, s[0:3], 0 offen offset:15910843; ALIGNED-NEXT:    buffer_store_byte v34, v0, s[0:3], 0 offen offset:15810844; ALIGNED-NEXT:    buffer_store_byte v33, v0, s[0:3], 0 offen offset:15710845; ALIGNED-NEXT:    buffer_store_byte v32, v0, s[0:3], 0 offen offset:15610846; ALIGNED-NEXT:    buffer_store_byte v31, v0, s[0:3], 0 offen offset:15510847; ALIGNED-NEXT:    buffer_store_byte v30, v0, s[0:3], 0 offen offset:15410848; ALIGNED-NEXT:    buffer_store_byte v29, v0, s[0:3], 0 offen offset:15310849; ALIGNED-NEXT:    buffer_store_byte v28, v0, s[0:3], 0 offen offset:15210850; ALIGNED-NEXT:    buffer_store_byte v27, v0, s[0:3], 0 offen offset:15110851; ALIGNED-NEXT:    buffer_store_byte v26, v0, s[0:3], 0 offen offset:15010852; ALIGNED-NEXT:    buffer_store_byte v25, v0, s[0:3], 0 offen offset:14910853; ALIGNED-NEXT:    buffer_store_byte v24, v0, s[0:3], 0 offen offset:14810854; ALIGNED-NEXT:    buffer_store_byte v23, v0, s[0:3], 0 offen offset:14710855; ALIGNED-NEXT:    buffer_store_byte v22, v0, s[0:3], 0 offen offset:14610856; ALIGNED-NEXT:    buffer_store_byte v21, v0, s[0:3], 0 offen offset:14510857; ALIGNED-NEXT:    buffer_store_byte v20, v0, s[0:3], 0 offen offset:14410858; ALIGNED-NEXT:    buffer_store_byte v19, v0, s[0:3], 0 offen offset:14310859; ALIGNED-NEXT:    buffer_store_byte v18, v0, s[0:3], 0 offen offset:14210860; ALIGNED-NEXT:    buffer_store_byte v17, v0, s[0:3], 0 offen offset:14110861; ALIGNED-NEXT:    buffer_store_byte v16, v0, s[0:3], 0 offen offset:14010862; ALIGNED-NEXT:    buffer_store_byte v15, v0, s[0:3], 0 offen offset:13910863; ALIGNED-NEXT:    buffer_store_byte v14, v0, s[0:3], 0 offen offset:13810864; ALIGNED-NEXT:    buffer_store_byte v13, v0, s[0:3], 0 offen offset:13710865; ALIGNED-NEXT:    buffer_store_byte v12, v0, s[0:3], 0 offen offset:13610866; ALIGNED-NEXT:    buffer_store_byte v11, v0, s[0:3], 0 offen offset:13510867; ALIGNED-NEXT:    buffer_store_byte v10, v0, s[0:3], 0 offen offset:13410868; ALIGNED-NEXT:    buffer_store_byte v9, v0, s[0:3], 0 offen offset:13310869; ALIGNED-NEXT:    buffer_store_byte v8, v0, s[0:3], 0 offen offset:13210870; ALIGNED-NEXT:    buffer_store_byte v7, v0, s[0:3], 0 offen offset:13110871; ALIGNED-NEXT:    buffer_store_byte v6, v0, s[0:3], 0 offen offset:13010872; ALIGNED-NEXT:    buffer_store_byte v5, v0, s[0:3], 0 offen offset:12910873; ALIGNED-NEXT:    buffer_store_byte v4, v0, s[0:3], 0 offen offset:12810874; ALIGNED-NEXT:    buffer_store_byte v3, v0, s[0:3], 0 offen offset:12710875; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload10876; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10877; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:12610878; ALIGNED-NEXT:    buffer_store_byte v127, v0, s[0:3], 0 offen offset:12510879; ALIGNED-NEXT:    buffer_store_byte v126, v0, s[0:3], 0 offen offset:12410880; ALIGNED-NEXT:    buffer_store_byte v125, v0, s[0:3], 0 offen offset:12310881; ALIGNED-NEXT:    buffer_store_byte v124, v0, s[0:3], 0 offen offset:12210882; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload10883; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10884; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:12110885; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload10886; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10887; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:12010888; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload10889; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10890; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11910891; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload10892; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10893; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11810894; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload10895; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10896; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11710897; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload10898; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10899; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11610900; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload10901; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10902; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11510903; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload10904; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10905; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11410906; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload10907; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10908; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11310909; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload10910; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10911; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11210912; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload10913; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10914; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11110915; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload10916; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10917; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11010918; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload10919; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10920; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10910921; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload10922; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10923; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10810924; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload10925; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10926; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10710927; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload10928; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10929; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10610930; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload10931; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10932; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10510933; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload10934; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10935; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10410936; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload10937; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10938; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10310939; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload10940; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10941; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10210942; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload10943; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10944; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10110945; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload10946; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10947; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10010948; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload10949; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10950; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9910951; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload10952; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10953; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9810954; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload10955; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10956; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9710957; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload10958; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10959; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9610960; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload10961; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10962; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9510963; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload10964; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10965; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9410966; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload10967; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10968; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9310969; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload10970; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10971; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9210972; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload10973; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10974; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9110975; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload10976; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10977; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9010978; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload10979; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10980; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8910981; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload10982; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10983; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8810984; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload10985; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10986; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8710987; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload10988; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10989; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8610990; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload10991; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10992; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8510993; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload10994; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10995; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8410996; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload10997; ALIGNED-NEXT:    s_waitcnt vmcnt(0)10998; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8310999; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload11000; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11001; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8211002; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload11003; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11004; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8111005; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload11006; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11007; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8011008; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload11009; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11010; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7911011; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload11012; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11013; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7811014; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload11015; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11016; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7711017; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload11018; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11019; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7611020; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload11021; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11022; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7511023; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload11024; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11025; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7411026; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload11027; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11028; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7311029; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload11030; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11031; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7211032; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload11033; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11034; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7111035; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload11036; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11037; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7011038; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload11039; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11040; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6911041; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload11042; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11043; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6811044; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload11045; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11046; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6711047; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload11048; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11049; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6611050; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload11051; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11052; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6511053; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload11054; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11055; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6411056; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload11057; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11058; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6311059; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload11060; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11061; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6211062; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload11063; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11064; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6111065; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload11066; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11067; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6011068; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload11069; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11070; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5911071; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload11072; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11073; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5811074; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload11075; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11076; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5711077; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload11078; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11079; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5611080; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload11081; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11082; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5511083; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload11084; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11085; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5411086; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload11087; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11088; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5311089; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload11090; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11091; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5211092; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload11093; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11094; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5111095; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload11096; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11097; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5011098; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload11099; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11100; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4911101; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload11102; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11103; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4811104; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload11105; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11106; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4711107; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload11108; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11109; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4611110; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload11111; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11112; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4511113; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload11114; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11115; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4411116; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload11117; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11118; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4311119; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload11120; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11121; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4211122; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload11123; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11124; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4111125; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload11126; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11127; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4011128; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload11129; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11130; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3911131; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload11132; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11133; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3811134; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload11135; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11136; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3711137; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload11138; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11139; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3611140; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload11141; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11142; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3511143; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload11144; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11145; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3411146; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload11147; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11148; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3311149; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload11150; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11151; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3211152; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload11153; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11154; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3111155; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload11156; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11157; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3011158; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload11159; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11160; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2911161; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload11162; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11163; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2811164; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload11165; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11166; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2711167; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload11168; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11169; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2611170; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload11171; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11172; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2511173; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload11174; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11175; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2411176; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload11177; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11178; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2311179; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload11180; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11181; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2211182; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload11183; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11184; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2111185; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload11186; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11187; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2011188; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload11189; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11190; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1911191; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload11192; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11193; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1811194; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload11195; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11196; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1711197; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload11198; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11199; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1611200; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload11201; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11202; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1511203; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload11204; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11205; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1411206; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload11207; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11208; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1311209; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload11210; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11211; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1211212; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload11213; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11214; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1111215; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload11216; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11217; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1011218; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload11219; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11220; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:911221; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload11222; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11223; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:811224; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload11225; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11226; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:711227; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload11228; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11229; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:611230; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload11231; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11232; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:511233; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload11234; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11235; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:411236; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload11237; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11238; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:311239; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload11240; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11241; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:211242; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload11243; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11244; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:111245; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload11246; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11247; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen11248; ALIGNED-NEXT:    v_add_nc_u32_e32 v0, 0x100, v011249; ALIGNED-NEXT:    s_cbranch_scc1 .LBB8_211250; ALIGNED-NEXT:  .LBB8_3: ; %Flow1811251; ALIGNED-NEXT:    s_andn2_saveexec_b32 s6, s611252; ALIGNED-NEXT:    s_cbranch_execz .LBB8_611253; ALIGNED-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader11254; ALIGNED-NEXT:    v_add_nc_u32_e32 v1, 0x700, v111255; ALIGNED-NEXT:    v_add_nc_u32_e32 v0, 0x700, v011256; ALIGNED-NEXT:    s_movk_i32 s4, 0xf80011257; ALIGNED-NEXT:    s_mov_b32 s5, -111258; ALIGNED-NEXT:  .LBB8_5: ; %memmove_bwd_loop11259; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=111260; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25511261; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x10011262; ALIGNED-NEXT:    s_addc_u32 s5, s5, 011263; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], 011264; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11265; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill11266; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25411267; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11268; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill11269; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25311270; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11271; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill11272; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25211273; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11274; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill11275; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25111276; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11277; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill11278; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25011279; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11280; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill11281; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24911282; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11283; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill11284; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24811285; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11286; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill11287; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24711288; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11289; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill11290; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24611291; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11292; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill11293; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24511294; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11295; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill11296; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24411297; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11298; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill11299; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24311300; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11301; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill11302; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24211303; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11304; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill11305; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24111306; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11307; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill11308; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24011309; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11310; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill11311; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23911312; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11313; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill11314; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23811315; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11316; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill11317; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23711318; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11319; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill11320; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23611321; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11322; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill11323; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23511324; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11325; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill11326; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23411327; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11328; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill11329; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23311330; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11331; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill11332; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23211333; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11334; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill11335; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23111336; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11337; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill11338; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23011339; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11340; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill11341; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22911342; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11343; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill11344; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22811345; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11346; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill11347; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22711348; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11349; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill11350; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22611351; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11352; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill11353; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22511354; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11355; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill11356; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22411357; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11358; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill11359; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22311360; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11361; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill11362; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22211363; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11364; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill11365; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22111366; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11367; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill11368; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22011369; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11370; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill11371; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21911372; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11373; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill11374; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21811375; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11376; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill11377; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21711378; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11379; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill11380; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21611381; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11382; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill11383; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21511384; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11385; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill11386; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21411387; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11388; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill11389; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21311390; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11391; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill11392; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21211393; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11394; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill11395; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21111396; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11397; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill11398; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21011399; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11400; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill11401; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20911402; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11403; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill11404; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20811405; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11406; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill11407; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20711408; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11409; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill11410; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20611411; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11412; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill11413; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20511414; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11415; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill11416; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20411417; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11418; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill11419; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20311420; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11421; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill11422; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20211423; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11424; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill11425; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20111426; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11427; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill11428; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20011429; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11430; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill11431; ALIGNED-NEXT:    s_clause 0x3e11432; ALIGNED-NEXT:    buffer_load_ubyte v115, v1, s[0:3], 0 offen offset:19911433; ALIGNED-NEXT:    buffer_load_ubyte v114, v1, s[0:3], 0 offen offset:19811434; ALIGNED-NEXT:    buffer_load_ubyte v113, v1, s[0:3], 0 offen offset:19711435; ALIGNED-NEXT:    buffer_load_ubyte v112, v1, s[0:3], 0 offen offset:19611436; ALIGNED-NEXT:    buffer_load_ubyte v103, v1, s[0:3], 0 offen offset:19511437; ALIGNED-NEXT:    buffer_load_ubyte v102, v1, s[0:3], 0 offen offset:19411438; ALIGNED-NEXT:    buffer_load_ubyte v101, v1, s[0:3], 0 offen offset:19311439; ALIGNED-NEXT:    buffer_load_ubyte v100, v1, s[0:3], 0 offen offset:19211440; ALIGNED-NEXT:    buffer_load_ubyte v99, v1, s[0:3], 0 offen offset:19111441; ALIGNED-NEXT:    buffer_load_ubyte v98, v1, s[0:3], 0 offen offset:19011442; ALIGNED-NEXT:    buffer_load_ubyte v97, v1, s[0:3], 0 offen offset:18911443; ALIGNED-NEXT:    buffer_load_ubyte v96, v1, s[0:3], 0 offen offset:18811444; ALIGNED-NEXT:    buffer_load_ubyte v87, v1, s[0:3], 0 offen offset:18711445; ALIGNED-NEXT:    buffer_load_ubyte v86, v1, s[0:3], 0 offen offset:18611446; ALIGNED-NEXT:    buffer_load_ubyte v85, v1, s[0:3], 0 offen offset:18511447; ALIGNED-NEXT:    buffer_load_ubyte v84, v1, s[0:3], 0 offen offset:18411448; ALIGNED-NEXT:    buffer_load_ubyte v83, v1, s[0:3], 0 offen offset:18311449; ALIGNED-NEXT:    buffer_load_ubyte v82, v1, s[0:3], 0 offen offset:18211450; ALIGNED-NEXT:    buffer_load_ubyte v81, v1, s[0:3], 0 offen offset:18111451; ALIGNED-NEXT:    buffer_load_ubyte v80, v1, s[0:3], 0 offen offset:18011452; ALIGNED-NEXT:    buffer_load_ubyte v71, v1, s[0:3], 0 offen offset:17911453; ALIGNED-NEXT:    buffer_load_ubyte v70, v1, s[0:3], 0 offen offset:17811454; ALIGNED-NEXT:    buffer_load_ubyte v69, v1, s[0:3], 0 offen offset:17711455; ALIGNED-NEXT:    buffer_load_ubyte v68, v1, s[0:3], 0 offen offset:17611456; ALIGNED-NEXT:    buffer_load_ubyte v67, v1, s[0:3], 0 offen offset:17511457; ALIGNED-NEXT:    buffer_load_ubyte v66, v1, s[0:3], 0 offen offset:17411458; ALIGNED-NEXT:    buffer_load_ubyte v65, v1, s[0:3], 0 offen offset:17311459; ALIGNED-NEXT:    buffer_load_ubyte v64, v1, s[0:3], 0 offen offset:17211460; ALIGNED-NEXT:    buffer_load_ubyte v55, v1, s[0:3], 0 offen offset:17111461; ALIGNED-NEXT:    buffer_load_ubyte v54, v1, s[0:3], 0 offen offset:17011462; ALIGNED-NEXT:    buffer_load_ubyte v53, v1, s[0:3], 0 offen offset:16911463; ALIGNED-NEXT:    buffer_load_ubyte v52, v1, s[0:3], 0 offen offset:16811464; ALIGNED-NEXT:    buffer_load_ubyte v51, v1, s[0:3], 0 offen offset:16711465; ALIGNED-NEXT:    buffer_load_ubyte v50, v1, s[0:3], 0 offen offset:16611466; ALIGNED-NEXT:    buffer_load_ubyte v49, v1, s[0:3], 0 offen offset:16511467; ALIGNED-NEXT:    buffer_load_ubyte v48, v1, s[0:3], 0 offen offset:16411468; ALIGNED-NEXT:    buffer_load_ubyte v39, v1, s[0:3], 0 offen offset:16311469; ALIGNED-NEXT:    buffer_load_ubyte v38, v1, s[0:3], 0 offen offset:16211470; ALIGNED-NEXT:    buffer_load_ubyte v37, v1, s[0:3], 0 offen offset:16111471; ALIGNED-NEXT:    buffer_load_ubyte v36, v1, s[0:3], 0 offen offset:16011472; ALIGNED-NEXT:    buffer_load_ubyte v35, v1, s[0:3], 0 offen offset:15911473; ALIGNED-NEXT:    buffer_load_ubyte v34, v1, s[0:3], 0 offen offset:15811474; ALIGNED-NEXT:    buffer_load_ubyte v33, v1, s[0:3], 0 offen offset:15711475; ALIGNED-NEXT:    buffer_load_ubyte v32, v1, s[0:3], 0 offen offset:15611476; ALIGNED-NEXT:    buffer_load_ubyte v31, v1, s[0:3], 0 offen offset:15511477; ALIGNED-NEXT:    buffer_load_ubyte v30, v1, s[0:3], 0 offen offset:15411478; ALIGNED-NEXT:    buffer_load_ubyte v29, v1, s[0:3], 0 offen offset:15311479; ALIGNED-NEXT:    buffer_load_ubyte v28, v1, s[0:3], 0 offen offset:15211480; ALIGNED-NEXT:    buffer_load_ubyte v27, v1, s[0:3], 0 offen offset:15111481; ALIGNED-NEXT:    buffer_load_ubyte v26, v1, s[0:3], 0 offen offset:15011482; ALIGNED-NEXT:    buffer_load_ubyte v25, v1, s[0:3], 0 offen offset:14911483; ALIGNED-NEXT:    buffer_load_ubyte v24, v1, s[0:3], 0 offen offset:14811484; ALIGNED-NEXT:    buffer_load_ubyte v23, v1, s[0:3], 0 offen offset:14711485; ALIGNED-NEXT:    buffer_load_ubyte v22, v1, s[0:3], 0 offen offset:14611486; ALIGNED-NEXT:    buffer_load_ubyte v21, v1, s[0:3], 0 offen offset:14511487; ALIGNED-NEXT:    buffer_load_ubyte v20, v1, s[0:3], 0 offen offset:14411488; ALIGNED-NEXT:    buffer_load_ubyte v19, v1, s[0:3], 0 offen offset:14311489; ALIGNED-NEXT:    buffer_load_ubyte v18, v1, s[0:3], 0 offen offset:14211490; ALIGNED-NEXT:    buffer_load_ubyte v17, v1, s[0:3], 0 offen offset:14111491; ALIGNED-NEXT:    buffer_load_ubyte v16, v1, s[0:3], 0 offen offset:14011492; ALIGNED-NEXT:    buffer_load_ubyte v15, v1, s[0:3], 0 offen offset:13911493; ALIGNED-NEXT:    buffer_load_ubyte v14, v1, s[0:3], 0 offen offset:13811494; ALIGNED-NEXT:    buffer_load_ubyte v13, v1, s[0:3], 0 offen offset:13711495; ALIGNED-NEXT:    s_clause 0xa11496; ALIGNED-NEXT:    buffer_load_ubyte v12, v1, s[0:3], 0 offen offset:13611497; ALIGNED-NEXT:    buffer_load_ubyte v11, v1, s[0:3], 0 offen offset:13511498; ALIGNED-NEXT:    buffer_load_ubyte v10, v1, s[0:3], 0 offen offset:13411499; ALIGNED-NEXT:    buffer_load_ubyte v9, v1, s[0:3], 0 offen offset:13311500; ALIGNED-NEXT:    buffer_load_ubyte v8, v1, s[0:3], 0 offen offset:13211501; ALIGNED-NEXT:    buffer_load_ubyte v7, v1, s[0:3], 0 offen offset:13111502; ALIGNED-NEXT:    buffer_load_ubyte v6, v1, s[0:3], 0 offen offset:13011503; ALIGNED-NEXT:    buffer_load_ubyte v5, v1, s[0:3], 0 offen offset:12911504; ALIGNED-NEXT:    buffer_load_ubyte v4, v1, s[0:3], 0 offen offset:12811505; ALIGNED-NEXT:    buffer_load_ubyte v3, v1, s[0:3], 0 offen offset:12711506; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:12611507; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11508; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill11509; ALIGNED-NEXT:    s_clause 0x3411510; ALIGNED-NEXT:    buffer_load_ubyte v127, v1, s[0:3], 0 offen offset:12511511; ALIGNED-NEXT:    buffer_load_ubyte v126, v1, s[0:3], 0 offen offset:12411512; ALIGNED-NEXT:    buffer_load_ubyte v125, v1, s[0:3], 0 offen offset:12311513; ALIGNED-NEXT:    buffer_load_ubyte v124, v1, s[0:3], 0 offen offset:12211514; ALIGNED-NEXT:    buffer_load_ubyte v123, v1, s[0:3], 0 offen offset:12111515; ALIGNED-NEXT:    buffer_load_ubyte v122, v1, s[0:3], 0 offen offset:12011516; ALIGNED-NEXT:    buffer_load_ubyte v121, v1, s[0:3], 0 offen offset:11911517; ALIGNED-NEXT:    buffer_load_ubyte v120, v1, s[0:3], 0 offen offset:11811518; ALIGNED-NEXT:    buffer_load_ubyte v111, v1, s[0:3], 0 offen offset:11711519; ALIGNED-NEXT:    buffer_load_ubyte v110, v1, s[0:3], 0 offen offset:11611520; ALIGNED-NEXT:    buffer_load_ubyte v109, v1, s[0:3], 0 offen offset:11511521; ALIGNED-NEXT:    buffer_load_ubyte v108, v1, s[0:3], 0 offen offset:11411522; ALIGNED-NEXT:    buffer_load_ubyte v107, v1, s[0:3], 0 offen offset:11311523; ALIGNED-NEXT:    buffer_load_ubyte v106, v1, s[0:3], 0 offen offset:11211524; ALIGNED-NEXT:    buffer_load_ubyte v105, v1, s[0:3], 0 offen offset:11111525; ALIGNED-NEXT:    buffer_load_ubyte v104, v1, s[0:3], 0 offen offset:11011526; ALIGNED-NEXT:    buffer_load_ubyte v95, v1, s[0:3], 0 offen offset:10911527; ALIGNED-NEXT:    buffer_load_ubyte v94, v1, s[0:3], 0 offen offset:10811528; ALIGNED-NEXT:    buffer_load_ubyte v93, v1, s[0:3], 0 offen offset:10711529; ALIGNED-NEXT:    buffer_load_ubyte v92, v1, s[0:3], 0 offen offset:10611530; ALIGNED-NEXT:    buffer_load_ubyte v91, v1, s[0:3], 0 offen offset:10511531; ALIGNED-NEXT:    buffer_load_ubyte v90, v1, s[0:3], 0 offen offset:10411532; ALIGNED-NEXT:    buffer_load_ubyte v89, v1, s[0:3], 0 offen offset:10311533; ALIGNED-NEXT:    buffer_load_ubyte v88, v1, s[0:3], 0 offen offset:10211534; ALIGNED-NEXT:    buffer_load_ubyte v79, v1, s[0:3], 0 offen offset:10111535; ALIGNED-NEXT:    buffer_load_ubyte v78, v1, s[0:3], 0 offen offset:10011536; ALIGNED-NEXT:    buffer_load_ubyte v77, v1, s[0:3], 0 offen offset:9911537; ALIGNED-NEXT:    buffer_load_ubyte v76, v1, s[0:3], 0 offen offset:9811538; ALIGNED-NEXT:    buffer_load_ubyte v75, v1, s[0:3], 0 offen offset:9711539; ALIGNED-NEXT:    buffer_load_ubyte v74, v1, s[0:3], 0 offen offset:9611540; ALIGNED-NEXT:    buffer_load_ubyte v73, v1, s[0:3], 0 offen offset:9511541; ALIGNED-NEXT:    buffer_load_ubyte v72, v1, s[0:3], 0 offen offset:9411542; ALIGNED-NEXT:    buffer_load_ubyte v63, v1, s[0:3], 0 offen offset:9311543; ALIGNED-NEXT:    buffer_load_ubyte v62, v1, s[0:3], 0 offen offset:9211544; ALIGNED-NEXT:    buffer_load_ubyte v61, v1, s[0:3], 0 offen offset:9111545; ALIGNED-NEXT:    buffer_load_ubyte v60, v1, s[0:3], 0 offen offset:9011546; ALIGNED-NEXT:    buffer_load_ubyte v59, v1, s[0:3], 0 offen offset:8911547; ALIGNED-NEXT:    buffer_load_ubyte v58, v1, s[0:3], 0 offen offset:8811548; ALIGNED-NEXT:    buffer_load_ubyte v57, v1, s[0:3], 0 offen offset:8711549; ALIGNED-NEXT:    buffer_load_ubyte v56, v1, s[0:3], 0 offen offset:8611550; ALIGNED-NEXT:    buffer_load_ubyte v47, v1, s[0:3], 0 offen offset:8511551; ALIGNED-NEXT:    buffer_load_ubyte v46, v1, s[0:3], 0 offen offset:8411552; ALIGNED-NEXT:    buffer_load_ubyte v45, v1, s[0:3], 0 offen offset:8311553; ALIGNED-NEXT:    buffer_load_ubyte v44, v1, s[0:3], 0 offen offset:8211554; ALIGNED-NEXT:    buffer_load_ubyte v43, v1, s[0:3], 0 offen offset:8111555; ALIGNED-NEXT:    buffer_load_ubyte v42, v1, s[0:3], 0 offen offset:8011556; ALIGNED-NEXT:    buffer_load_ubyte v41, v1, s[0:3], 0 offen offset:7911557; ALIGNED-NEXT:    buffer_load_ubyte v40, v1, s[0:3], 0 offen offset:7811558; ALIGNED-NEXT:    buffer_load_ubyte v119, v1, s[0:3], 0 offen offset:7711559; ALIGNED-NEXT:    buffer_load_ubyte v118, v1, s[0:3], 0 offen offset:7611560; ALIGNED-NEXT:    buffer_load_ubyte v117, v1, s[0:3], 0 offen offset:7511561; ALIGNED-NEXT:    buffer_load_ubyte v116, v1, s[0:3], 0 offen offset:7411562; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7311563; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11564; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill11565; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7211566; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11567; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill11568; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7111569; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11570; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill11571; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7011572; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11573; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill11574; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6911575; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11576; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill11577; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6811578; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11579; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill11580; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6711581; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11582; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill11583; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6611584; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11585; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill11586; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6511587; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11588; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill11589; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6411590; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11591; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill11592; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6311593; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11594; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill11595; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6211596; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11597; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill11598; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6111599; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11600; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill11601; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6011602; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11603; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill11604; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5911605; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11606; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill11607; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5811608; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11609; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill11610; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5711611; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11612; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill11613; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5611614; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11615; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill11616; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5511617; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11618; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill11619; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5411620; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11621; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill11622; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5311623; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11624; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill11625; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5211626; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11627; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill11628; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5111629; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11630; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill11631; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5011632; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11633; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill11634; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4911635; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11636; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill11637; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4811638; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11639; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill11640; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4711641; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11642; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill11643; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4611644; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11645; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill11646; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4511647; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11648; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill11649; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4411650; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11651; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill11652; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4311653; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11654; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill11655; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4211656; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11657; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill11658; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4111659; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11660; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill11661; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4011662; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11663; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill11664; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3911665; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11666; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill11667; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3811668; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11669; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill11670; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3711671; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11672; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill11673; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3611674; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11675; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill11676; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3511677; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11678; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill11679; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3411680; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11681; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill11682; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3311683; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11684; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill11685; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3211686; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11687; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill11688; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3111689; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11690; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill11691; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3011692; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11693; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill11694; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2911695; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11696; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill11697; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2811698; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11699; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill11700; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2711701; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11702; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill11703; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2611704; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11705; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill11706; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2511707; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11708; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill11709; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2411710; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11711; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill11712; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2311713; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11714; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill11715; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2211716; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11717; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill11718; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2111719; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11720; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill11721; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2011722; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11723; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill11724; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1911725; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11726; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill11727; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1811728; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11729; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill11730; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1711731; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11732; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill11733; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1611734; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11735; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill11736; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1511737; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11738; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill11739; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1411740; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11741; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill11742; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1311743; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11744; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill11745; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1211746; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11747; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill11748; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1111749; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11750; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill11751; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1011752; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11753; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill11754; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:911755; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11756; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill11757; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:811758; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11759; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill11760; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:711761; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11762; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill11763; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:611764; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11765; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill11766; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:511767; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11768; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill11769; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:411770; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11771; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill11772; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:311773; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11774; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill11775; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:211776; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11777; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill11778; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:111779; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11780; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill11781; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen11782; ALIGNED-NEXT:    v_add_nc_u32_e32 v1, 0xffffff00, v111783; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11784; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill11785; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload11786; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11787; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25511788; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload11789; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11790; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25411791; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload11792; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11793; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25311794; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload11795; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11796; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25211797; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload11798; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11799; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25111800; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload11801; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11802; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25011803; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload11804; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11805; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24911806; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload11807; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11808; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24811809; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload11810; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11811; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24711812; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload11813; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11814; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24611815; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload11816; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11817; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24511818; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload11819; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11820; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24411821; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload11822; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11823; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24311824; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload11825; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11826; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24211827; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload11828; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11829; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24111830; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload11831; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11832; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24011833; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload11834; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11835; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23911836; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload11837; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11838; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23811839; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload11840; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11841; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23711842; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload11843; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11844; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23611845; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload11846; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11847; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23511848; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload11849; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11850; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23411851; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload11852; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11853; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23311854; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload11855; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11856; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23211857; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload11858; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11859; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23111860; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload11861; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11862; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23011863; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload11864; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11865; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22911866; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload11867; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11868; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22811869; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload11870; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11871; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22711872; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload11873; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11874; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22611875; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload11876; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11877; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22511878; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload11879; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11880; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22411881; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload11882; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11883; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22311884; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload11885; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11886; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22211887; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload11888; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11889; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22111890; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload11891; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11892; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22011893; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload11894; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11895; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21911896; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload11897; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11898; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21811899; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload11900; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11901; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21711902; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload11903; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11904; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21611905; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload11906; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11907; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21511908; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload11909; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11910; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21411911; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload11912; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11913; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21311914; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload11915; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11916; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21211917; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload11918; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11919; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21111920; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload11921; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11922; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21011923; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload11924; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11925; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20911926; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload11927; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11928; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20811929; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload11930; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11931; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20711932; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload11933; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11934; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20611935; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload11936; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11937; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20511938; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload11939; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11940; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20411941; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload11942; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11943; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20311944; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload11945; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11946; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20211947; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload11948; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11949; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20111950; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload11951; ALIGNED-NEXT:    s_waitcnt vmcnt(0)11952; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20011953; ALIGNED-NEXT:    buffer_store_byte v115, v0, s[0:3], 0 offen offset:19911954; ALIGNED-NEXT:    buffer_store_byte v114, v0, s[0:3], 0 offen offset:19811955; ALIGNED-NEXT:    buffer_store_byte v113, v0, s[0:3], 0 offen offset:19711956; ALIGNED-NEXT:    buffer_store_byte v112, v0, s[0:3], 0 offen offset:19611957; ALIGNED-NEXT:    buffer_store_byte v103, v0, s[0:3], 0 offen offset:19511958; ALIGNED-NEXT:    buffer_store_byte v102, v0, s[0:3], 0 offen offset:19411959; ALIGNED-NEXT:    buffer_store_byte v101, v0, s[0:3], 0 offen offset:19311960; ALIGNED-NEXT:    buffer_store_byte v100, v0, s[0:3], 0 offen offset:19211961; ALIGNED-NEXT:    buffer_store_byte v99, v0, s[0:3], 0 offen offset:19111962; ALIGNED-NEXT:    buffer_store_byte v98, v0, s[0:3], 0 offen offset:19011963; ALIGNED-NEXT:    buffer_store_byte v97, v0, s[0:3], 0 offen offset:18911964; ALIGNED-NEXT:    buffer_store_byte v96, v0, s[0:3], 0 offen offset:18811965; ALIGNED-NEXT:    buffer_store_byte v87, v0, s[0:3], 0 offen offset:18711966; ALIGNED-NEXT:    buffer_store_byte v86, v0, s[0:3], 0 offen offset:18611967; ALIGNED-NEXT:    buffer_store_byte v85, v0, s[0:3], 0 offen offset:18511968; ALIGNED-NEXT:    buffer_store_byte v84, v0, s[0:3], 0 offen offset:18411969; ALIGNED-NEXT:    buffer_store_byte v83, v0, s[0:3], 0 offen offset:18311970; ALIGNED-NEXT:    buffer_store_byte v82, v0, s[0:3], 0 offen offset:18211971; ALIGNED-NEXT:    buffer_store_byte v81, v0, s[0:3], 0 offen offset:18111972; ALIGNED-NEXT:    buffer_store_byte v80, v0, s[0:3], 0 offen offset:18011973; ALIGNED-NEXT:    buffer_store_byte v71, v0, s[0:3], 0 offen offset:17911974; ALIGNED-NEXT:    buffer_store_byte v70, v0, s[0:3], 0 offen offset:17811975; ALIGNED-NEXT:    buffer_store_byte v69, v0, s[0:3], 0 offen offset:17711976; ALIGNED-NEXT:    buffer_store_byte v68, v0, s[0:3], 0 offen offset:17611977; ALIGNED-NEXT:    buffer_store_byte v67, v0, s[0:3], 0 offen offset:17511978; ALIGNED-NEXT:    buffer_store_byte v66, v0, s[0:3], 0 offen offset:17411979; ALIGNED-NEXT:    buffer_store_byte v65, v0, s[0:3], 0 offen offset:17311980; ALIGNED-NEXT:    buffer_store_byte v64, v0, s[0:3], 0 offen offset:17211981; ALIGNED-NEXT:    buffer_store_byte v55, v0, s[0:3], 0 offen offset:17111982; ALIGNED-NEXT:    buffer_store_byte v54, v0, s[0:3], 0 offen offset:17011983; ALIGNED-NEXT:    buffer_store_byte v53, v0, s[0:3], 0 offen offset:16911984; ALIGNED-NEXT:    buffer_store_byte v52, v0, s[0:3], 0 offen offset:16811985; ALIGNED-NEXT:    buffer_store_byte v51, v0, s[0:3], 0 offen offset:16711986; ALIGNED-NEXT:    buffer_store_byte v50, v0, s[0:3], 0 offen offset:16611987; ALIGNED-NEXT:    buffer_store_byte v49, v0, s[0:3], 0 offen offset:16511988; ALIGNED-NEXT:    buffer_store_byte v48, v0, s[0:3], 0 offen offset:16411989; ALIGNED-NEXT:    buffer_store_byte v39, v0, s[0:3], 0 offen offset:16311990; ALIGNED-NEXT:    buffer_store_byte v38, v0, s[0:3], 0 offen offset:16211991; ALIGNED-NEXT:    buffer_store_byte v37, v0, s[0:3], 0 offen offset:16111992; ALIGNED-NEXT:    buffer_store_byte v36, v0, s[0:3], 0 offen offset:16011993; ALIGNED-NEXT:    buffer_store_byte v35, v0, s[0:3], 0 offen offset:15911994; ALIGNED-NEXT:    buffer_store_byte v34, v0, s[0:3], 0 offen offset:15811995; ALIGNED-NEXT:    buffer_store_byte v33, v0, s[0:3], 0 offen offset:15711996; ALIGNED-NEXT:    buffer_store_byte v32, v0, s[0:3], 0 offen offset:15611997; ALIGNED-NEXT:    buffer_store_byte v31, v0, s[0:3], 0 offen offset:15511998; ALIGNED-NEXT:    buffer_store_byte v30, v0, s[0:3], 0 offen offset:15411999; ALIGNED-NEXT:    buffer_store_byte v29, v0, s[0:3], 0 offen offset:15312000; ALIGNED-NEXT:    buffer_store_byte v28, v0, s[0:3], 0 offen offset:15212001; ALIGNED-NEXT:    buffer_store_byte v27, v0, s[0:3], 0 offen offset:15112002; ALIGNED-NEXT:    buffer_store_byte v26, v0, s[0:3], 0 offen offset:15012003; ALIGNED-NEXT:    buffer_store_byte v25, v0, s[0:3], 0 offen offset:14912004; ALIGNED-NEXT:    buffer_store_byte v24, v0, s[0:3], 0 offen offset:14812005; ALIGNED-NEXT:    buffer_store_byte v23, v0, s[0:3], 0 offen offset:14712006; ALIGNED-NEXT:    buffer_store_byte v22, v0, s[0:3], 0 offen offset:14612007; ALIGNED-NEXT:    buffer_store_byte v21, v0, s[0:3], 0 offen offset:14512008; ALIGNED-NEXT:    buffer_store_byte v20, v0, s[0:3], 0 offen offset:14412009; ALIGNED-NEXT:    buffer_store_byte v19, v0, s[0:3], 0 offen offset:14312010; ALIGNED-NEXT:    buffer_store_byte v18, v0, s[0:3], 0 offen offset:14212011; ALIGNED-NEXT:    buffer_store_byte v17, v0, s[0:3], 0 offen offset:14112012; ALIGNED-NEXT:    buffer_store_byte v16, v0, s[0:3], 0 offen offset:14012013; ALIGNED-NEXT:    buffer_store_byte v15, v0, s[0:3], 0 offen offset:13912014; ALIGNED-NEXT:    buffer_store_byte v14, v0, s[0:3], 0 offen offset:13812015; ALIGNED-NEXT:    buffer_store_byte v13, v0, s[0:3], 0 offen offset:13712016; ALIGNED-NEXT:    buffer_store_byte v12, v0, s[0:3], 0 offen offset:13612017; ALIGNED-NEXT:    buffer_store_byte v11, v0, s[0:3], 0 offen offset:13512018; ALIGNED-NEXT:    buffer_store_byte v10, v0, s[0:3], 0 offen offset:13412019; ALIGNED-NEXT:    buffer_store_byte v9, v0, s[0:3], 0 offen offset:13312020; ALIGNED-NEXT:    buffer_store_byte v8, v0, s[0:3], 0 offen offset:13212021; ALIGNED-NEXT:    buffer_store_byte v7, v0, s[0:3], 0 offen offset:13112022; ALIGNED-NEXT:    buffer_store_byte v6, v0, s[0:3], 0 offen offset:13012023; ALIGNED-NEXT:    buffer_store_byte v5, v0, s[0:3], 0 offen offset:12912024; ALIGNED-NEXT:    buffer_store_byte v4, v0, s[0:3], 0 offen offset:12812025; ALIGNED-NEXT:    buffer_store_byte v3, v0, s[0:3], 0 offen offset:12712026; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload12027; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12028; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:12612029; ALIGNED-NEXT:    buffer_store_byte v127, v0, s[0:3], 0 offen offset:12512030; ALIGNED-NEXT:    buffer_store_byte v126, v0, s[0:3], 0 offen offset:12412031; ALIGNED-NEXT:    buffer_store_byte v125, v0, s[0:3], 0 offen offset:12312032; ALIGNED-NEXT:    buffer_store_byte v124, v0, s[0:3], 0 offen offset:12212033; ALIGNED-NEXT:    buffer_store_byte v123, v0, s[0:3], 0 offen offset:12112034; ALIGNED-NEXT:    buffer_store_byte v122, v0, s[0:3], 0 offen offset:12012035; ALIGNED-NEXT:    buffer_store_byte v121, v0, s[0:3], 0 offen offset:11912036; ALIGNED-NEXT:    buffer_store_byte v120, v0, s[0:3], 0 offen offset:11812037; ALIGNED-NEXT:    buffer_store_byte v111, v0, s[0:3], 0 offen offset:11712038; ALIGNED-NEXT:    buffer_store_byte v110, v0, s[0:3], 0 offen offset:11612039; ALIGNED-NEXT:    buffer_store_byte v109, v0, s[0:3], 0 offen offset:11512040; ALIGNED-NEXT:    buffer_store_byte v108, v0, s[0:3], 0 offen offset:11412041; ALIGNED-NEXT:    buffer_store_byte v107, v0, s[0:3], 0 offen offset:11312042; ALIGNED-NEXT:    buffer_store_byte v106, v0, s[0:3], 0 offen offset:11212043; ALIGNED-NEXT:    buffer_store_byte v105, v0, s[0:3], 0 offen offset:11112044; ALIGNED-NEXT:    buffer_store_byte v104, v0, s[0:3], 0 offen offset:11012045; ALIGNED-NEXT:    buffer_store_byte v95, v0, s[0:3], 0 offen offset:10912046; ALIGNED-NEXT:    buffer_store_byte v94, v0, s[0:3], 0 offen offset:10812047; ALIGNED-NEXT:    buffer_store_byte v93, v0, s[0:3], 0 offen offset:10712048; ALIGNED-NEXT:    buffer_store_byte v92, v0, s[0:3], 0 offen offset:10612049; ALIGNED-NEXT:    buffer_store_byte v91, v0, s[0:3], 0 offen offset:10512050; ALIGNED-NEXT:    buffer_store_byte v90, v0, s[0:3], 0 offen offset:10412051; ALIGNED-NEXT:    buffer_store_byte v89, v0, s[0:3], 0 offen offset:10312052; ALIGNED-NEXT:    buffer_store_byte v88, v0, s[0:3], 0 offen offset:10212053; ALIGNED-NEXT:    buffer_store_byte v79, v0, s[0:3], 0 offen offset:10112054; ALIGNED-NEXT:    buffer_store_byte v78, v0, s[0:3], 0 offen offset:10012055; ALIGNED-NEXT:    buffer_store_byte v77, v0, s[0:3], 0 offen offset:9912056; ALIGNED-NEXT:    buffer_store_byte v76, v0, s[0:3], 0 offen offset:9812057; ALIGNED-NEXT:    buffer_store_byte v75, v0, s[0:3], 0 offen offset:9712058; ALIGNED-NEXT:    buffer_store_byte v74, v0, s[0:3], 0 offen offset:9612059; ALIGNED-NEXT:    buffer_store_byte v73, v0, s[0:3], 0 offen offset:9512060; ALIGNED-NEXT:    buffer_store_byte v72, v0, s[0:3], 0 offen offset:9412061; ALIGNED-NEXT:    buffer_store_byte v63, v0, s[0:3], 0 offen offset:9312062; ALIGNED-NEXT:    buffer_store_byte v62, v0, s[0:3], 0 offen offset:9212063; ALIGNED-NEXT:    buffer_store_byte v61, v0, s[0:3], 0 offen offset:9112064; ALIGNED-NEXT:    buffer_store_byte v60, v0, s[0:3], 0 offen offset:9012065; ALIGNED-NEXT:    buffer_store_byte v59, v0, s[0:3], 0 offen offset:8912066; ALIGNED-NEXT:    buffer_store_byte v58, v0, s[0:3], 0 offen offset:8812067; ALIGNED-NEXT:    buffer_store_byte v57, v0, s[0:3], 0 offen offset:8712068; ALIGNED-NEXT:    buffer_store_byte v56, v0, s[0:3], 0 offen offset:8612069; ALIGNED-NEXT:    buffer_store_byte v47, v0, s[0:3], 0 offen offset:8512070; ALIGNED-NEXT:    buffer_store_byte v46, v0, s[0:3], 0 offen offset:8412071; ALIGNED-NEXT:    buffer_store_byte v45, v0, s[0:3], 0 offen offset:8312072; ALIGNED-NEXT:    buffer_store_byte v44, v0, s[0:3], 0 offen offset:8212073; ALIGNED-NEXT:    buffer_store_byte v43, v0, s[0:3], 0 offen offset:8112074; ALIGNED-NEXT:    buffer_store_byte v42, v0, s[0:3], 0 offen offset:8012075; ALIGNED-NEXT:    buffer_store_byte v41, v0, s[0:3], 0 offen offset:7912076; ALIGNED-NEXT:    buffer_store_byte v40, v0, s[0:3], 0 offen offset:7812077; ALIGNED-NEXT:    buffer_store_byte v119, v0, s[0:3], 0 offen offset:7712078; ALIGNED-NEXT:    buffer_store_byte v118, v0, s[0:3], 0 offen offset:7612079; ALIGNED-NEXT:    buffer_store_byte v117, v0, s[0:3], 0 offen offset:7512080; ALIGNED-NEXT:    buffer_store_byte v116, v0, s[0:3], 0 offen offset:7412081; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload12082; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12083; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7312084; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload12085; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12086; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7212087; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload12088; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12089; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7112090; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload12091; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12092; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7012093; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload12094; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12095; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6912096; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload12097; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12098; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6812099; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload12100; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12101; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6712102; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload12103; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12104; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6612105; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload12106; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12107; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6512108; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload12109; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12110; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6412111; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload12112; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12113; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6312114; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload12115; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12116; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6212117; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload12118; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12119; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6112120; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload12121; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12122; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6012123; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload12124; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12125; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5912126; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload12127; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12128; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5812129; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload12130; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12131; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5712132; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload12133; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12134; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5612135; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload12136; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12137; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5512138; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload12139; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12140; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5412141; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload12142; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12143; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5312144; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload12145; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12146; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5212147; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload12148; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12149; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5112150; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload12151; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12152; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5012153; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload12154; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12155; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4912156; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload12157; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12158; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4812159; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload12160; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12161; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4712162; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload12163; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12164; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4612165; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload12166; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12167; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4512168; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload12169; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12170; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4412171; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload12172; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12173; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4312174; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload12175; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12176; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4212177; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload12178; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12179; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4112180; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload12181; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12182; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4012183; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload12184; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12185; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3912186; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload12187; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12188; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3812189; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload12190; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12191; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3712192; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload12193; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12194; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3612195; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload12196; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12197; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3512198; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload12199; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12200; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3412201; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload12202; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12203; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3312204; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload12205; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12206; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3212207; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload12208; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12209; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3112210; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload12211; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12212; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3012213; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload12214; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12215; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2912216; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload12217; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12218; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2812219; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload12220; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12221; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2712222; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload12223; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12224; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2612225; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload12226; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12227; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2512228; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload12229; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12230; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2412231; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload12232; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12233; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2312234; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload12235; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12236; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2212237; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload12238; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12239; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2112240; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload12241; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12242; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2012243; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload12244; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12245; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1912246; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload12247; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12248; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1812249; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload12250; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12251; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1712252; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload12253; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12254; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1612255; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload12256; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12257; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1512258; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload12259; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12260; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1412261; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload12262; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12263; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1312264; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload12265; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12266; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1212267; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload12268; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12269; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1112270; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload12271; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12272; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1012273; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload12274; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12275; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:912276; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload12277; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12278; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:812279; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload12280; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12281; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:712282; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload12283; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12284; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:612285; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload12286; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12287; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:512288; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload12289; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12290; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:412291; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload12292; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12293; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:312294; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload12295; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12296; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:212297; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload12298; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12299; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:112300; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload12301; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12302; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen12303; ALIGNED-NEXT:    v_add_nc_u32_e32 v0, 0xffffff00, v012304; ALIGNED-NEXT:    s_cbranch_scc0 .LBB8_512305; ALIGNED-NEXT:  .LBB8_6: ; %Flow1912306; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s612307; ALIGNED-NEXT:    s_clause 0x2f ; 192-byte Folded Reload12308; ALIGNED-NEXT:    buffer_load_dword v127, off, s[0:3], s3212309; ALIGNED-NEXT:    buffer_load_dword v126, off, s[0:3], s32 offset:412310; ALIGNED-NEXT:    buffer_load_dword v125, off, s[0:3], s32 offset:812311; ALIGNED-NEXT:    buffer_load_dword v124, off, s[0:3], s32 offset:1212312; ALIGNED-NEXT:    buffer_load_dword v123, off, s[0:3], s32 offset:1612313; ALIGNED-NEXT:    buffer_load_dword v122, off, s[0:3], s32 offset:2012314; ALIGNED-NEXT:    buffer_load_dword v121, off, s[0:3], s32 offset:2412315; ALIGNED-NEXT:    buffer_load_dword v120, off, s[0:3], s32 offset:2812316; ALIGNED-NEXT:    buffer_load_dword v111, off, s[0:3], s32 offset:3212317; ALIGNED-NEXT:    buffer_load_dword v110, off, s[0:3], s32 offset:3612318; ALIGNED-NEXT:    buffer_load_dword v109, off, s[0:3], s32 offset:4012319; ALIGNED-NEXT:    buffer_load_dword v108, off, s[0:3], s32 offset:4412320; ALIGNED-NEXT:    buffer_load_dword v107, off, s[0:3], s32 offset:4812321; ALIGNED-NEXT:    buffer_load_dword v106, off, s[0:3], s32 offset:5212322; ALIGNED-NEXT:    buffer_load_dword v105, off, s[0:3], s32 offset:5612323; ALIGNED-NEXT:    buffer_load_dword v104, off, s[0:3], s32 offset:6012324; ALIGNED-NEXT:    buffer_load_dword v95, off, s[0:3], s32 offset:6412325; ALIGNED-NEXT:    buffer_load_dword v94, off, s[0:3], s32 offset:6812326; ALIGNED-NEXT:    buffer_load_dword v93, off, s[0:3], s32 offset:7212327; ALIGNED-NEXT:    buffer_load_dword v92, off, s[0:3], s32 offset:7612328; ALIGNED-NEXT:    buffer_load_dword v91, off, s[0:3], s32 offset:8012329; ALIGNED-NEXT:    buffer_load_dword v90, off, s[0:3], s32 offset:8412330; ALIGNED-NEXT:    buffer_load_dword v89, off, s[0:3], s32 offset:8812331; ALIGNED-NEXT:    buffer_load_dword v88, off, s[0:3], s32 offset:9212332; ALIGNED-NEXT:    buffer_load_dword v79, off, s[0:3], s32 offset:9612333; ALIGNED-NEXT:    buffer_load_dword v78, off, s[0:3], s32 offset:10012334; ALIGNED-NEXT:    buffer_load_dword v77, off, s[0:3], s32 offset:10412335; ALIGNED-NEXT:    buffer_load_dword v76, off, s[0:3], s32 offset:10812336; ALIGNED-NEXT:    buffer_load_dword v75, off, s[0:3], s32 offset:11212337; ALIGNED-NEXT:    buffer_load_dword v74, off, s[0:3], s32 offset:11612338; ALIGNED-NEXT:    buffer_load_dword v73, off, s[0:3], s32 offset:12012339; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s32 offset:12412340; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12812341; ALIGNED-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:13212342; ALIGNED-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:13612343; ALIGNED-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:14012344; ALIGNED-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:14412345; ALIGNED-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:14812346; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:15212347; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:15612348; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:16012349; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:16412350; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:16812351; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:17212352; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:17612353; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:18012354; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:18412355; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:18812356; ALIGNED-NEXT:    s_waitcnt vmcnt(0)12357; ALIGNED-NEXT:    s_setpc_b64 s[30:31]12358;12359; UNROLL3-LABEL: memmove_p5_p5_sz2048:12360; UNROLL3:       ; %bb.0: ; %entry12361; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)12362; UNROLL3-NEXT:    s_mov_b32 s4, exec_lo12363; UNROLL3-NEXT:    v_cmpx_ge_u32_e64 v1, v012364; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s412365; UNROLL3-NEXT:    s_cbranch_execz .LBB8_412366; UNROLL3-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader12367; UNROLL3-NEXT:    v_mov_b32_e32 v2, v112368; UNROLL3-NEXT:    v_mov_b32_e32 v3, v012369; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7e012370; UNROLL3-NEXT:  .LBB8_2: ; %memmove_fwd_loop12371; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=112372; UNROLL3-NEXT:    s_clause 0xb12373; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4412374; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:4012375; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:3612376; UNROLL3-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:3212377; UNROLL3-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:2812378; UNROLL3-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:2412379; UNROLL3-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:2012380; UNROLL3-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:1612381; UNROLL3-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:1212382; UNROLL3-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:812383; UNROLL3-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:412384; UNROLL3-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen12385; UNROLL3-NEXT:    v_add_nc_u32_e32 v2, 48, v212386; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd012387; UNROLL3-NEXT:    s_addc_u32 s5, s5, -112388; UNROLL3-NEXT:    s_waitcnt vmcnt(11)12389; UNROLL3-NEXT:    buffer_store_dword v4, v3, s[0:3], 0 offen offset:4412390; UNROLL3-NEXT:    s_waitcnt vmcnt(10)12391; UNROLL3-NEXT:    buffer_store_dword v5, v3, s[0:3], 0 offen offset:4012392; UNROLL3-NEXT:    s_waitcnt vmcnt(9)12393; UNROLL3-NEXT:    buffer_store_dword v6, v3, s[0:3], 0 offen offset:3612394; UNROLL3-NEXT:    s_waitcnt vmcnt(8)12395; UNROLL3-NEXT:    buffer_store_dword v7, v3, s[0:3], 0 offen offset:3212396; UNROLL3-NEXT:    s_waitcnt vmcnt(7)12397; UNROLL3-NEXT:    buffer_store_dword v8, v3, s[0:3], 0 offen offset:2812398; UNROLL3-NEXT:    s_waitcnt vmcnt(6)12399; UNROLL3-NEXT:    buffer_store_dword v9, v3, s[0:3], 0 offen offset:2412400; UNROLL3-NEXT:    s_waitcnt vmcnt(5)12401; UNROLL3-NEXT:    buffer_store_dword v10, v3, s[0:3], 0 offen offset:2012402; UNROLL3-NEXT:    s_waitcnt vmcnt(4)12403; UNROLL3-NEXT:    buffer_store_dword v11, v3, s[0:3], 0 offen offset:1612404; UNROLL3-NEXT:    s_waitcnt vmcnt(3)12405; UNROLL3-NEXT:    buffer_store_dword v12, v3, s[0:3], 0 offen offset:1212406; UNROLL3-NEXT:    s_waitcnt vmcnt(2)12407; UNROLL3-NEXT:    buffer_store_dword v13, v3, s[0:3], 0 offen offset:812408; UNROLL3-NEXT:    s_waitcnt vmcnt(1)12409; UNROLL3-NEXT:    buffer_store_dword v14, v3, s[0:3], 0 offen offset:412410; UNROLL3-NEXT:    s_waitcnt vmcnt(0)12411; UNROLL3-NEXT:    buffer_store_dword v15, v3, s[0:3], 0 offen12412; UNROLL3-NEXT:    v_add_nc_u32_e32 v3, 48, v312413; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 012414; UNROLL3-NEXT:    s_cbranch_scc1 .LBB8_212415; UNROLL3-NEXT:  ; %bb.3: ; %memmove_fwd_residual12416; UNROLL3-NEXT:    s_clause 0x312417; UNROLL3-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:202812418; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:202412419; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:202012420; UNROLL3-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:201612421; UNROLL3-NEXT:    s_waitcnt vmcnt(3)12422; UNROLL3-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:202812423; UNROLL3-NEXT:    s_waitcnt vmcnt(2)12424; UNROLL3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:202412425; UNROLL3-NEXT:    s_waitcnt vmcnt(1)12426; UNROLL3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:202012427; UNROLL3-NEXT:    s_waitcnt vmcnt(0)12428; UNROLL3-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:201612429; UNROLL3-NEXT:    s_clause 0x312430; UNROLL3-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:204412431; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:204012432; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:203612433; UNROLL3-NEXT:    buffer_load_dword v1, v1, s[0:3], 0 offen offset:203212434; UNROLL3-NEXT:    s_waitcnt vmcnt(3)12435; UNROLL3-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:204412436; UNROLL3-NEXT:    s_waitcnt vmcnt(2)12437; UNROLL3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:204012438; UNROLL3-NEXT:    s_waitcnt vmcnt(1)12439; UNROLL3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:203612440; UNROLL3-NEXT:    s_waitcnt vmcnt(0)12441; UNROLL3-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:203212442; UNROLL3-NEXT:    ; implicit-def: $vgpr112443; UNROLL3-NEXT:    ; implicit-def: $vgpr012444; UNROLL3-NEXT:  .LBB8_4: ; %Flow1612445; UNROLL3-NEXT:    s_andn2_saveexec_b32 s6, s612446; UNROLL3-NEXT:    s_cbranch_execz .LBB8_712447; UNROLL3-NEXT:  ; %bb.5: ; %memmove_bwd_residual12448; UNROLL3-NEXT:    s_clause 0x312449; UNROLL3-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:204412450; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:204012451; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:203612452; UNROLL3-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:203212453; UNROLL3-NEXT:    s_movk_i32 s4, 0xf82012454; UNROLL3-NEXT:    s_mov_b32 s5, -112455; UNROLL3-NEXT:    s_waitcnt vmcnt(3)12456; UNROLL3-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:204412457; UNROLL3-NEXT:    s_waitcnt vmcnt(2)12458; UNROLL3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:204012459; UNROLL3-NEXT:    s_waitcnt vmcnt(1)12460; UNROLL3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:203612461; UNROLL3-NEXT:    s_waitcnt vmcnt(0)12462; UNROLL3-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:203212463; UNROLL3-NEXT:    s_clause 0x312464; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:202812465; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:202412466; UNROLL3-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:202012467; UNROLL3-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:201612468; UNROLL3-NEXT:    v_add_nc_u32_e32 v1, 0x7b0, v112469; UNROLL3-NEXT:    v_add_nc_u32_e32 v2, 0x7b0, v012470; UNROLL3-NEXT:    s_waitcnt vmcnt(3)12471; UNROLL3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:202812472; UNROLL3-NEXT:    s_waitcnt vmcnt(2)12473; UNROLL3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:202412474; UNROLL3-NEXT:    s_waitcnt vmcnt(1)12475; UNROLL3-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:202012476; UNROLL3-NEXT:    s_waitcnt vmcnt(0)12477; UNROLL3-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:201612478; UNROLL3-NEXT:  .LBB8_6: ; %memmove_bwd_loop12479; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=112480; UNROLL3-NEXT:    s_clause 0xb12481; UNROLL3-NEXT:    buffer_load_dword v0, v1, s[0:3], 0 offen offset:4412482; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:4012483; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:3612484; UNROLL3-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:3212485; UNROLL3-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:2812486; UNROLL3-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:2412487; UNROLL3-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:2012488; UNROLL3-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:1612489; UNROLL3-NEXT:    buffer_load_dword v10, v1, s[0:3], 0 offen offset:1212490; UNROLL3-NEXT:    buffer_load_dword v11, v1, s[0:3], 0 offen offset:812491; UNROLL3-NEXT:    buffer_load_dword v12, v1, s[0:3], 0 offen offset:412492; UNROLL3-NEXT:    buffer_load_dword v13, v1, s[0:3], 0 offen12493; UNROLL3-NEXT:    v_subrev_nc_u32_e32 v1, 48, v112494; UNROLL3-NEXT:    s_add_u32 s4, s4, 4812495; UNROLL3-NEXT:    s_addc_u32 s5, s5, 012496; UNROLL3-NEXT:    s_waitcnt vmcnt(11)12497; UNROLL3-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen offset:4412498; UNROLL3-NEXT:    s_waitcnt vmcnt(10)12499; UNROLL3-NEXT:    buffer_store_dword v3, v2, s[0:3], 0 offen offset:4012500; UNROLL3-NEXT:    s_waitcnt vmcnt(9)12501; UNROLL3-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen offset:3612502; UNROLL3-NEXT:    s_waitcnt vmcnt(8)12503; UNROLL3-NEXT:    buffer_store_dword v5, v2, s[0:3], 0 offen offset:3212504; UNROLL3-NEXT:    s_waitcnt vmcnt(7)12505; UNROLL3-NEXT:    buffer_store_dword v6, v2, s[0:3], 0 offen offset:2812506; UNROLL3-NEXT:    s_waitcnt vmcnt(6)12507; UNROLL3-NEXT:    buffer_store_dword v7, v2, s[0:3], 0 offen offset:2412508; UNROLL3-NEXT:    s_waitcnt vmcnt(5)12509; UNROLL3-NEXT:    buffer_store_dword v8, v2, s[0:3], 0 offen offset:2012510; UNROLL3-NEXT:    s_waitcnt vmcnt(4)12511; UNROLL3-NEXT:    buffer_store_dword v9, v2, s[0:3], 0 offen offset:1612512; UNROLL3-NEXT:    s_waitcnt vmcnt(3)12513; UNROLL3-NEXT:    buffer_store_dword v10, v2, s[0:3], 0 offen offset:1212514; UNROLL3-NEXT:    s_waitcnt vmcnt(2)12515; UNROLL3-NEXT:    buffer_store_dword v11, v2, s[0:3], 0 offen offset:812516; UNROLL3-NEXT:    s_waitcnt vmcnt(1)12517; UNROLL3-NEXT:    buffer_store_dword v12, v2, s[0:3], 0 offen offset:412518; UNROLL3-NEXT:    s_waitcnt vmcnt(0)12519; UNROLL3-NEXT:    buffer_store_dword v13, v2, s[0:3], 0 offen12520; UNROLL3-NEXT:    v_subrev_nc_u32_e32 v2, 48, v212521; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], 012522; UNROLL3-NEXT:    s_cbranch_scc0 .LBB8_612523; UNROLL3-NEXT:  .LBB8_7: ; %Flow1712524; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s612525; UNROLL3-NEXT:    s_setpc_b64 s[30:31]12526entry:12527  tail call void @llvm.memmove.p5.p5.i64(ptr addrspace(5) noundef nonnull align 1 %dst, ptr addrspace(5) noundef nonnull align 1 %src, i64 2048, i1 false)12528  ret void12529}12530 12531define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5) align 1 readonly %src) {12532; CHECK-LABEL: memmove_p0_p5_sz2048:12533; CHECK:       ; %bb.0: ; %entry12534; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)12535; CHECK-NEXT:    v_cmp_ne_u64_e32 vcc_lo, 0, v[0:1]12536; CHECK-NEXT:    s_mov_b64 s[4:5], 012537; CHECK-NEXT:    s_mov_b32 s6, exec_lo12538; CHECK-NEXT:    v_cndmask_b32_e32 v3, -1, v0, vcc_lo12539; CHECK-NEXT:    v_cmpx_ge_u32_e64 v2, v312540; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s612541; CHECK-NEXT:    s_cbranch_execz .LBB9_212542; CHECK-NEXT:  .LBB9_1: ; %memmove_fwd_loop12543; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=112544; CHECK-NEXT:    s_clause 0x3e12545; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:3212546; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:3612547; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:4012548; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:4412549; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:4812550; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:5212551; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:5612552; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:6012553; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:7612554; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:9212555; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:8812556; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:8412557; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:8012558; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:7212559; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:6812560; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:6412561; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:10812562; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:12412563; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:12012564; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:11612565; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:11212566; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:10412567; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:10012568; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:9612569; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:17212570; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:18812571; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:18412572; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:18012573; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:17612574; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:16812575; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:16412576; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:16012577; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:20412578; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:22012579; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:21612580; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:21212581; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:20812582; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:20012583; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:19612584; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:19212585; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:23612586; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:25212587; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:24812588; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:24412589; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:24012590; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:23212591; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:22812592; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:22412593; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:14012594; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:15612595; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:15212596; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:14812597; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:14412598; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:13612599; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:13212600; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:12812601; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen12602; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:412603; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:812604; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:1612605; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:2012606; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:2412607; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:2812608; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:1212609; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s412610; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo12611; CHECK-NEXT:    s_add_u32 s4, s4, 0x10012612; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 4812613; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0x100, v212614; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo12615; CHECK-NEXT:    s_addc_u32 s5, s5, 012616; CHECK-NEXT:    s_waitcnt vmcnt(19)12617; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:19212618; CHECK-NEXT:    s_waitcnt vmcnt(16)12619; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:17612620; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:16012621; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[35:38] offset:14412622; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[31:34] offset:12812623; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[27:30] offset:11212624; CHECK-NEXT:    s_waitcnt vmcnt(11)12625; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[80:83] offset:9612626; CHECK-NEXT:    s_waitcnt vmcnt(8)12627; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:12812628; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[23:26] offset:6412629; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[19:22] offset:4812630; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[15:18] offset:3212631; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:6412632; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[7:10]12633; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:3212634; CHECK-NEXT:    s_waitcnt vmcnt(1)12635; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99] offset:1612636; CHECK-NEXT:    s_waitcnt vmcnt(0)12637; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87]12638; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0x80012639; CHECK-NEXT:    s_cbranch_scc1 .LBB9_112640; CHECK-NEXT:  .LBB9_2: ; %Flow1012641; CHECK-NEXT:    s_andn2_saveexec_b32 s8, s612642; CHECK-NEXT:    s_cbranch_execz .LBB9_512643; CHECK-NEXT:  ; %bb.3: ; %memmove_bwd_loop.preheader12644; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0x700, v212645; CHECK-NEXT:    s_movk_i32 s6, 0xff0012646; CHECK-NEXT:    s_mov_b64 s[4:5], 0x70012647; CHECK-NEXT:    s_mov_b32 s7, -112648; CHECK-NEXT:  .LBB9_4: ; %memmove_bwd_loop12649; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=112650; CHECK-NEXT:    s_clause 0x3e12651; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:3212652; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:3612653; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:4012654; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:4412655; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:4812656; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:5212657; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:5612658; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:6012659; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:7612660; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:9212661; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:8812662; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:8412663; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:8012664; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:7212665; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:6812666; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:6412667; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:10812668; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:12412669; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:12012670; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:11612671; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:11212672; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:10412673; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:10012674; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:9612675; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:23612676; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:25212677; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:24812678; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:24412679; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:24012680; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:23212681; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:22812682; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:22412683; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:20412684; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:22012685; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:21612686; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:21212687; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:20812688; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:20012689; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:19612690; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:19212691; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:17212692; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:18812693; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:18412694; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:18012695; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:17612696; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:16812697; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:16412698; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:16012699; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:15612700; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:15212701; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:14812702; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:14412703; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:14012704; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:13612705; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:13212706; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:12812707; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen12708; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:412709; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:812710; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:1612711; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:2012712; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:2412713; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:2812714; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:1212715; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s412716; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo12717; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0xffffff00, v212718; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 4812719; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff0012720; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo12721; CHECK-NEXT:    s_addc_u32 s5, s5, -112722; CHECK-NEXT:    s_waitcnt vmcnt(35)12723; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[31:34] offset:19212724; CHECK-NEXT:    s_waitcnt vmcnt(32)12725; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[27:30] offset:17612726; CHECK-NEXT:    s_waitcnt vmcnt(27)12727; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:16012728; CHECK-NEXT:    s_waitcnt vmcnt(24)12729; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[35:38] offset:14412730; CHECK-NEXT:    s_waitcnt vmcnt(19)12731; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:12812732; CHECK-NEXT:    s_waitcnt vmcnt(16)12733; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:11212734; CHECK-NEXT:    s_waitcnt vmcnt(12)12735; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71] offset:9612736; CHECK-NEXT:    s_waitcnt vmcnt(8)12737; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:12812738; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[23:26] offset:6412739; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[19:22] offset:4812740; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[15:18] offset:3212741; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:6412742; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[7:10]12743; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:3212744; CHECK-NEXT:    s_waitcnt vmcnt(1)12745; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99] offset:1612746; CHECK-NEXT:    s_waitcnt vmcnt(0)12747; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87]12748; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]12749; CHECK-NEXT:    s_cbranch_scc0 .LBB9_412750; CHECK-NEXT:  .LBB9_5: ; %Flow1112751; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s812752; CHECK-NEXT:    s_waitcnt lgkmcnt(0)12753; CHECK-NEXT:    s_setpc_b64 s[30:31]12754;12755; ALIGNED-LABEL: memmove_p0_p5_sz2048:12756; ALIGNED:       ; %bb.0: ; %entry12757; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)12758; ALIGNED-NEXT:    v_cmp_ne_u64_e32 vcc_lo, 0, v[0:1]12759; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill12760; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill12761; ALIGNED-NEXT:    s_mov_b64 s[4:5], 012762; ALIGNED-NEXT:    s_mov_b32 s6, exec_lo12763; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill12764; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill12765; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill12766; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill12767; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill12768; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill12769; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill12770; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill12771; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill12772; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill12773; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill12774; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill12775; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill12776; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill12777; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill12778; ALIGNED-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill12779; ALIGNED-NEXT:    buffer_store_dword v72, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill12780; ALIGNED-NEXT:    buffer_store_dword v73, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill12781; ALIGNED-NEXT:    buffer_store_dword v74, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill12782; ALIGNED-NEXT:    buffer_store_dword v75, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill12783; ALIGNED-NEXT:    buffer_store_dword v76, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill12784; ALIGNED-NEXT:    buffer_store_dword v77, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill12785; ALIGNED-NEXT:    buffer_store_dword v78, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill12786; ALIGNED-NEXT:    buffer_store_dword v79, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill12787; ALIGNED-NEXT:    buffer_store_dword v88, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill12788; ALIGNED-NEXT:    buffer_store_dword v89, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill12789; ALIGNED-NEXT:    buffer_store_dword v90, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill12790; ALIGNED-NEXT:    buffer_store_dword v91, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill12791; ALIGNED-NEXT:    buffer_store_dword v92, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill12792; ALIGNED-NEXT:    buffer_store_dword v93, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill12793; ALIGNED-NEXT:    buffer_store_dword v94, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill12794; ALIGNED-NEXT:    buffer_store_dword v95, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill12795; ALIGNED-NEXT:    buffer_store_dword v104, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill12796; ALIGNED-NEXT:    buffer_store_dword v105, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill12797; ALIGNED-NEXT:    buffer_store_dword v106, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill12798; ALIGNED-NEXT:    buffer_store_dword v107, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill12799; ALIGNED-NEXT:    buffer_store_dword v108, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill12800; ALIGNED-NEXT:    buffer_store_dword v109, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill12801; ALIGNED-NEXT:    buffer_store_dword v110, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill12802; ALIGNED-NEXT:    buffer_store_dword v111, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill12803; ALIGNED-NEXT:    buffer_store_dword v120, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill12804; ALIGNED-NEXT:    buffer_store_dword v121, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill12805; ALIGNED-NEXT:    buffer_store_dword v122, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill12806; ALIGNED-NEXT:    buffer_store_dword v123, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill12807; ALIGNED-NEXT:    buffer_store_dword v124, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill12808; ALIGNED-NEXT:    buffer_store_dword v125, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill12809; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill12810; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 ; 4-byte Folded Spill12811; ALIGNED-NEXT:    v_cndmask_b32_e32 v0, -1, v0, vcc_lo12812; ALIGNED-NEXT:    v_cmpx_ge_u32_e64 v2, v012813; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s612814; ALIGNED-NEXT:    s_cbranch_execz .LBB9_212815; ALIGNED-NEXT:  .LBB9_1: ; %memmove_fwd_loop12816; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=112817; ALIGNED-NEXT:    s_clause 0x3e12818; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:2012819; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:2112820; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:2212821; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:2312822; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:2412823; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:2512824; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:2612825; ALIGNED-NEXT:    buffer_load_ubyte v127, v2, s[0:3], 0 offen offset:1912826; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:2812827; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:2912828; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:3012829; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:3112830; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:3212831; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:3312832; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:3412833; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:2712834; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:3512835; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:3612836; ALIGNED-NEXT:    buffer_load_ubyte v16, v2, s[0:3], 0 offen offset:3712837; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:3812838; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:3912839; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:4012840; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:4112841; ALIGNED-NEXT:    buffer_load_ubyte v25, v2, s[0:3], 0 offen offset:4212842; ALIGNED-NEXT:    buffer_load_ubyte v27, v2, s[0:3], 0 offen offset:4312843; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:4412844; ALIGNED-NEXT:    buffer_load_ubyte v24, v2, s[0:3], 0 offen offset:4512845; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:4612846; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:4712847; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:4812848; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:4912849; ALIGNED-NEXT:    buffer_load_ubyte v33, v2, s[0:3], 0 offen offset:5012850; ALIGNED-NEXT:    buffer_load_ubyte v32, v2, s[0:3], 0 offen offset:5112851; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:5212852; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:5312853; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:5412854; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:5512855; ALIGNED-NEXT:    buffer_load_ubyte v38, v2, s[0:3], 0 offen offset:5612856; ALIGNED-NEXT:    buffer_load_ubyte v50, v2, s[0:3], 0 offen offset:5712857; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:5812858; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:6012859; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:6112860; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:6212861; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:6312862; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:6412863; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:6512864; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:6612865; ALIGNED-NEXT:    buffer_load_ubyte v51, v2, s[0:3], 0 offen offset:5912866; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:6712867; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:6812868; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:6912869; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:7012870; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:7112871; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:7612872; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:7712873; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:7512874; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:7812875; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:7912876; ALIGNED-NEXT:    buffer_load_ubyte v126, v2, s[0:3], 0 offen offset:15912877; ALIGNED-NEXT:    buffer_load_ubyte v120, v2, s[0:3], 0 offen offset:15512878; ALIGNED-NEXT:    buffer_load_ubyte v123, v2, s[0:3], 0 offen offset:15212879; ALIGNED-NEXT:    buffer_load_ubyte v111, v2, s[0:3], 0 offen offset:15312880; ALIGNED-NEXT:    buffer_load_ubyte v110, v2, s[0:3], 0 offen offset:15412881; ALIGNED-NEXT:    s_clause 0x3112882; ALIGNED-NEXT:    buffer_load_ubyte v109, v2, s[0:3], 0 offen offset:16012883; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:16112884; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:16212885; ALIGNED-NEXT:    buffer_load_ubyte v94, v2, s[0:3], 0 offen offset:16312886; ALIGNED-NEXT:    buffer_load_ubyte v107, v2, s[0:3], 0 offen offset:16412887; ALIGNED-NEXT:    buffer_load_ubyte v105, v2, s[0:3], 0 offen offset:16512888; ALIGNED-NEXT:    buffer_load_ubyte v104, v2, s[0:3], 0 offen offset:16612889; ALIGNED-NEXT:    buffer_load_ubyte v93, v2, s[0:3], 0 offen offset:16712890; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:17212891; ALIGNED-NEXT:    buffer_load_ubyte v89, v2, s[0:3], 0 offen offset:17312892; ALIGNED-NEXT:    buffer_load_ubyte v88, v2, s[0:3], 0 offen offset:17412893; ALIGNED-NEXT:    buffer_load_ubyte v79, v2, s[0:3], 0 offen offset:17512894; ALIGNED-NEXT:    buffer_load_ubyte v78, v2, s[0:3], 0 offen offset:17112895; ALIGNED-NEXT:    buffer_load_ubyte v101, v2, s[0:3], 0 offen offset:20412896; ALIGNED-NEXT:    buffer_load_ubyte v99, v2, s[0:3], 0 offen offset:20512897; ALIGNED-NEXT:    buffer_load_ubyte v100, v2, s[0:3], 0 offen offset:20612898; ALIGNED-NEXT:    buffer_load_ubyte v98, v2, s[0:3], 0 offen offset:20712899; ALIGNED-NEXT:    buffer_load_ubyte v97, v2, s[0:3], 0 offen offset:20312900; ALIGNED-NEXT:    buffer_load_ubyte v96, v2, s[0:3], 0 offen offset:20012901; ALIGNED-NEXT:    buffer_load_ubyte v87, v2, s[0:3], 0 offen offset:20112902; ALIGNED-NEXT:    buffer_load_ubyte v86, v2, s[0:3], 0 offen offset:20212903; ALIGNED-NEXT:    buffer_load_ubyte v77, v2, s[0:3], 0 offen offset:16812904; ALIGNED-NEXT:    buffer_load_ubyte v75, v2, s[0:3], 0 offen offset:16912905; ALIGNED-NEXT:    buffer_load_ubyte v74, v2, s[0:3], 0 offen offset:17012906; ALIGNED-NEXT:    buffer_load_ubyte v73, v2, s[0:3], 0 offen offset:17612907; ALIGNED-NEXT:    buffer_load_ubyte v63, v2, s[0:3], 0 offen offset:17712908; ALIGNED-NEXT:    buffer_load_ubyte v59, v2, s[0:3], 0 offen offset:17812909; ALIGNED-NEXT:    buffer_load_ubyte v60, v2, s[0:3], 0 offen offset:17912910; ALIGNED-NEXT:    buffer_load_ubyte v72, v2, s[0:3], 0 offen offset:18012911; ALIGNED-NEXT:    buffer_load_ubyte v61, v2, s[0:3], 0 offen offset:18112912; ALIGNED-NEXT:    buffer_load_ubyte v62, v2, s[0:3], 0 offen offset:18212913; ALIGNED-NEXT:    buffer_load_ubyte v58, v2, s[0:3], 0 offen offset:18312914; ALIGNED-NEXT:    buffer_load_ubyte v56, v2, s[0:3], 0 offen offset:18812915; ALIGNED-NEXT:    buffer_load_ubyte v47, v2, s[0:3], 0 offen offset:18912916; ALIGNED-NEXT:    buffer_load_ubyte v46, v2, s[0:3], 0 offen offset:19012917; ALIGNED-NEXT:    buffer_load_ubyte v45, v2, s[0:3], 0 offen offset:19112918; ALIGNED-NEXT:    buffer_load_ubyte v44, v2, s[0:3], 0 offen offset:18712919; ALIGNED-NEXT:    buffer_load_ubyte v43, v2, s[0:3], 0 offen offset:18412920; ALIGNED-NEXT:    buffer_load_ubyte v42, v2, s[0:3], 0 offen offset:18512921; ALIGNED-NEXT:    buffer_load_ubyte v40, v2, s[0:3], 0 offen offset:18612922; ALIGNED-NEXT:    buffer_load_ubyte v119, v2, s[0:3], 0 offen offset:19212923; ALIGNED-NEXT:    buffer_load_ubyte v116, v2, s[0:3], 0 offen offset:19312924; ALIGNED-NEXT:    buffer_load_ubyte v113, v2, s[0:3], 0 offen offset:19412925; ALIGNED-NEXT:    buffer_load_ubyte v112, v2, s[0:3], 0 offen offset:19512926; ALIGNED-NEXT:    buffer_load_ubyte v117, v2, s[0:3], 0 offen offset:19612927; ALIGNED-NEXT:    buffer_load_ubyte v115, v2, s[0:3], 0 offen offset:19712928; ALIGNED-NEXT:    buffer_load_ubyte v114, v2, s[0:3], 0 offen offset:19812929; ALIGNED-NEXT:    buffer_load_ubyte v103, v2, s[0:3], 0 offen offset:19912930; ALIGNED-NEXT:    buffer_load_ubyte v124, v2, s[0:3], 0 offen offset:412931; ALIGNED-NEXT:    buffer_load_ubyte v125, v2, s[0:3], 0 offen offset:612932; ALIGNED-NEXT:    s_waitcnt vmcnt(62)12933; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill12934; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill12935; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill12936; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill12937; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill12938; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill12939; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill12940; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v012941; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill12942; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill12943; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill12944; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill12945; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill12946; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v312947; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v512948; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill12949; ALIGNED-NEXT:    v_lshl_or_b32 v4, v9, 8, v812950; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill12951; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 8, v612952; ALIGNED-NEXT:    v_lshl_or_b32 v6, v11, 8, v1212953; ALIGNED-NEXT:    v_lshl_or_b32 v7, v15, 8, v1412954; ALIGNED-NEXT:    v_lshl_or_b32 v8, v19, 8, v1712955; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v1312956; ALIGNED-NEXT:    v_lshl_or_b32 v10, v20, 8, v1812957; ALIGNED-NEXT:    v_lshl_or_b32 v11, v23, 8, v2212958; ALIGNED-NEXT:    v_lshl_or_b32 v12, v27, 8, v2512959; ALIGNED-NEXT:    v_lshl_or_b32 v13, v24, 8, v2112960; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v012961; ALIGNED-NEXT:    v_lshl_or_b32 v14, v28, 8, v2612962; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v312963; ALIGNED-NEXT:    v_lshl_or_b32 v3, v6, 16, v512964; ALIGNED-NEXT:    v_lshl_or_b32 v4, v8, 16, v712965; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 16, v912966; ALIGNED-NEXT:    v_lshl_or_b32 v6, v12, 16, v1112967; ALIGNED-NEXT:    v_lshl_or_b32 v7, v14, 16, v1312968; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill12969; ALIGNED-NEXT:    v_lshl_or_b32 v15, v30, 8, v2912970; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill12971; ALIGNED-NEXT:    v_lshl_or_b32 v0, v32, 8, v3312972; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill12973; ALIGNED-NEXT:    v_lshl_or_b32 v1, v34, 8, v3112974; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill12975; ALIGNED-NEXT:    v_lshl_or_b32 v3, v36, 8, v3512976; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill12977; ALIGNED-NEXT:    v_lshl_or_b32 v4, v39, 8, v3712978; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill12979; ALIGNED-NEXT:    v_lshl_or_b32 v5, v49, 8, v4812980; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill12981; ALIGNED-NEXT:    v_lshl_or_b32 v6, v50, 8, v3812982; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill12983; ALIGNED-NEXT:    v_lshl_or_b32 v7, v51, 8, v5212984; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 16, v1512985; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v112986; ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 16, v412987; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:8512988; ALIGNED-NEXT:    v_lshl_or_b32 v4, v7, 16, v612989; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill12990; ALIGNED-NEXT:    v_lshl_or_b32 v0, v54, 8, v5312991; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill12992; ALIGNED-NEXT:    v_lshl_or_b32 v1, v55, 8, v6512993; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill12994; ALIGNED-NEXT:    v_lshl_or_b32 v3, v66, 8, v6412995; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill12996; ALIGNED-NEXT:    s_waitcnt vmcnt(61)12997; ALIGNED-NEXT:    v_lshl_or_b32 v4, v68, 8, v6712998; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v012999; ALIGNED-NEXT:    s_clause 0x213000; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:8613001; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:8213002; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:8713003; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v313004; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill13005; ALIGNED-NEXT:    s_waitcnt vmcnt(62)13006; ALIGNED-NEXT:    v_lshl_or_b32 v0, v70, 8, v6913007; ALIGNED-NEXT:    s_clause 0x113008; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:8313009; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:7413010; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill13011; ALIGNED-NEXT:    s_waitcnt vmcnt(61)13012; ALIGNED-NEXT:    v_lshl_or_b32 v1, v80, 8, v7113013; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill13014; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill13015; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill13016; ALIGNED-NEXT:    s_clause 0x413017; ALIGNED-NEXT:    buffer_load_ubyte v84, v2, s[0:3], 0 offen offset:21213018; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:21313019; ALIGNED-NEXT:    buffer_load_ubyte v82, v2, s[0:3], 0 offen offset:21413020; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:21513021; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:21113022; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013023; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:7313024; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill13025; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill13026; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill13027; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:996 ; 4-byte Folded Spill13028; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:7213029; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill13030; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill13031; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill13032; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:1000 ; 4-byte Folded Spill13033; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill13034; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill13035; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill13036; ALIGNED-NEXT:    s_clause 0x213037; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:20813038; ALIGNED-NEXT:    buffer_load_ubyte v50, v2, s[0:3], 0 offen offset:20913039; ALIGNED-NEXT:    buffer_load_ubyte v51, v2, s[0:3], 0 offen offset:21013040; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill13041; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill13042; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill13043; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill13044; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill13045; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill13046; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill13047; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill13048; ALIGNED-NEXT:    s_clause 0x713049; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:22413050; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:22513051; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:22613052; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:22713053; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:22813054; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:22913055; ALIGNED-NEXT:    buffer_load_ubyte v38, v2, s[0:3], 0 offen offset:23013056; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:23113057; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill13058; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill13059; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill13060; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill13061; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill13062; ALIGNED-NEXT:    s_clause 0x413063; ALIGNED-NEXT:    buffer_load_ubyte v33, v2, s[0:3], 0 offen offset:23613064; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:23713065; ALIGNED-NEXT:    buffer_load_ubyte v32, v2, s[0:3], 0 offen offset:23813066; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:23913067; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:23513068; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill13069; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill13070; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill13071; ALIGNED-NEXT:    s_clause 0x213072; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:23213073; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:23313074; ALIGNED-NEXT:    buffer_load_ubyte v25, v2, s[0:3], 0 offen offset:23413075; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill13076; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill13077; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill13078; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill13079; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill13080; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill13081; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill13082; ALIGNED-NEXT:    s_clause 0xc13083; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:24013084; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:24113085; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:24213086; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:24313087; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:24413088; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:24513089; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:24613090; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:24713091; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:25213092; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:25313093; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:25413094; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:25513095; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:25113096; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill13097; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill13098; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill13099; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill13100; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill13101; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill13102; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:1428 ; 4-byte Folded Spill13103; ALIGNED-NEXT:    s_waitcnt vmcnt(44)13104; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1032 ; 4-byte Folded Spill13105; ALIGNED-NEXT:    s_waitcnt vmcnt(43)13106; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1036 ; 4-byte Folded Spill13107; ALIGNED-NEXT:    s_waitcnt vmcnt(42)13108; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1024 ; 4-byte Folded Spill13109; ALIGNED-NEXT:    s_waitcnt vmcnt(41)13110; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1040 ; 4-byte Folded Spill13111; ALIGNED-NEXT:    s_waitcnt vmcnt(40)13112; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1020 ; 4-byte Folded Spill13113; ALIGNED-NEXT:    s_waitcnt vmcnt(39)13114; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1004 ; 4-byte Folded Spill13115; ALIGNED-NEXT:    s_waitcnt vmcnt(33)13116; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:992 ; 4-byte Folded Spill13117; ALIGNED-NEXT:    s_waitcnt vmcnt(32)13118; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:988 ; 4-byte Folded Spill13119; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013120; ALIGNED-NEXT:    v_lshl_or_b32 v1, v81, 8, v313121; ALIGNED-NEXT:    s_clause 0x813122; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:8413123; ALIGNED-NEXT:    buffer_load_ubyte v83, v2, s[0:3], 0 offen offset:21613124; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:21713125; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:21813126; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:21913127; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:22013128; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:22113129; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:22213130; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:22313131; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013132; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:8113133; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Spill13134; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:8013135; ALIGNED-NEXT:    s_waitcnt vmcnt(10)13136; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1028 ; 4-byte Folded Spill13137; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13138; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1016 ; 4-byte Folded Spill13139; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13140; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1012 ; 4-byte Folded Spill13141; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013142; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v713143; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:9813144; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013145; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v613146; ALIGNED-NEXT:    s_clause 0x113147; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:10213148; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:10313149; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Spill13150; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v313151; ALIGNED-NEXT:    s_clause 0x113152; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:9413153; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:9513154; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013155; ALIGNED-NEXT:    s_clause 0x113156; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:9313157; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:9113158; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Spill13159; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:9213160; ALIGNED-NEXT:    s_waitcnt vmcnt(6)13161; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1116 ; 4-byte Folded Spill13162; ALIGNED-NEXT:    s_waitcnt vmcnt(5)13163; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1120 ; 4-byte Folded Spill13164; ALIGNED-NEXT:    s_waitcnt vmcnt(4)13165; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1056 ; 4-byte Folded Spill13166; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13167; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1060 ; 4-byte Folded Spill13168; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13169; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1052 ; 4-byte Folded Spill13170; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13171; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Spill13172; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013173; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v313174; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:9013175; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1072 ; 4-byte Folded Spill13176; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:10113177; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013178; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:8913179; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Spill13180; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:8813181; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13182; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1076 ; 4-byte Folded Spill13183; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13184; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1112 ; 4-byte Folded Spill13185; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13186; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1068 ; 4-byte Folded Spill13187; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13188; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Spill13189; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013190; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v313191; ALIGNED-NEXT:    s_clause 0x113192; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:9913193; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:10013194; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1104 ; 4-byte Folded Spill13195; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013196; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:9713197; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Spill13198; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:9613199; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13200; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1100 ; 4-byte Folded Spill13201; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13202; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1108 ; 4-byte Folded Spill13203; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13204; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1096 ; 4-byte Folded Spill13205; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13206; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Spill13207; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013208; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v713209; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:11413210; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013211; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v613212; ALIGNED-NEXT:    s_clause 0x113213; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:11813214; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:11913215; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Spill13216; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v313217; ALIGNED-NEXT:    s_clause 0x113218; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:11013219; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:11113220; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013221; ALIGNED-NEXT:    s_clause 0x113222; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:10913223; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:10713224; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1132 ; 4-byte Folded Spill13225; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:10813226; ALIGNED-NEXT:    s_waitcnt vmcnt(6)13227; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1196 ; 4-byte Folded Spill13228; ALIGNED-NEXT:    s_waitcnt vmcnt(5)13229; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1200 ; 4-byte Folded Spill13230; ALIGNED-NEXT:    s_waitcnt vmcnt(4)13231; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1140 ; 4-byte Folded Spill13232; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13233; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1144 ; 4-byte Folded Spill13234; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13235; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1136 ; 4-byte Folded Spill13236; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13237; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Spill13238; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013239; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v313240; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:10613241; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1160 ; 4-byte Folded Spill13242; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:11713243; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013244; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:10513245; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1156 ; 4-byte Folded Spill13246; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:10413247; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13248; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1164 ; 4-byte Folded Spill13249; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13250; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1192 ; 4-byte Folded Spill13251; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13252; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1152 ; 4-byte Folded Spill13253; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13254; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Spill13255; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013256; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v313257; ALIGNED-NEXT:    s_clause 0x113258; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:11513259; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:11613260; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1188 ; 4-byte Folded Spill13261; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013262; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:11313263; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Spill13264; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:11213265; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13266; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1180 ; 4-byte Folded Spill13267; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13268; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1184 ; 4-byte Folded Spill13269; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13270; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1176 ; 4-byte Folded Spill13271; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13272; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Spill13273; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013274; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v713275; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:13013276; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013277; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v613278; ALIGNED-NEXT:    s_clause 0x113279; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:13413280; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:13513281; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Spill13282; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v313283; ALIGNED-NEXT:    s_clause 0x113284; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:12613285; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:12713286; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013287; ALIGNED-NEXT:    s_clause 0x113288; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:12513289; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:12313290; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1212 ; 4-byte Folded Spill13291; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:12413292; ALIGNED-NEXT:    s_waitcnt vmcnt(6)13293; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1280 ; 4-byte Folded Spill13294; ALIGNED-NEXT:    s_waitcnt vmcnt(5)13295; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1284 ; 4-byte Folded Spill13296; ALIGNED-NEXT:    s_waitcnt vmcnt(4)13297; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1220 ; 4-byte Folded Spill13298; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13299; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1224 ; 4-byte Folded Spill13300; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13301; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1216 ; 4-byte Folded Spill13302; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13303; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Spill13304; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013305; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v313306; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:12213307; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1240 ; 4-byte Folded Spill13308; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:13313309; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013310; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:12113311; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Spill13312; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:12013313; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13314; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1244 ; 4-byte Folded Spill13315; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13316; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1276 ; 4-byte Folded Spill13317; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13318; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1236 ; 4-byte Folded Spill13319; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13320; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Spill13321; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013322; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v313323; ALIGNED-NEXT:    s_clause 0x113324; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:13113325; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:13213326; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1272 ; 4-byte Folded Spill13327; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013328; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:12913329; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Spill13330; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:12813331; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13332; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1260 ; 4-byte Folded Spill13333; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13334; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1264 ; 4-byte Folded Spill13335; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13336; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1256 ; 4-byte Folded Spill13337; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13338; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1252 ; 4-byte Folded Spill13339; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013340; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v713341; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:14613342; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013343; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v613344; ALIGNED-NEXT:    s_clause 0x113345; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:15013346; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:15113347; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1304 ; 4-byte Folded Spill13348; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v313349; ALIGNED-NEXT:    s_clause 0x113350; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:14213351; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:14313352; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013353; ALIGNED-NEXT:    s_clause 0x113354; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:14113355; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:13913356; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1316 ; 4-byte Folded Spill13357; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:14013358; ALIGNED-NEXT:    s_waitcnt vmcnt(6)13359; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1432 ; 4-byte Folded Spill13360; ALIGNED-NEXT:    s_waitcnt vmcnt(5)13361; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1436 ; 4-byte Folded Spill13362; ALIGNED-NEXT:    s_waitcnt vmcnt(4)13363; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1332 ; 4-byte Folded Spill13364; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13365; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1336 ; 4-byte Folded Spill13366; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13367; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1328 ; 4-byte Folded Spill13368; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13369; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Spill13370; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013371; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v313372; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:13813373; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1364 ; 4-byte Folded Spill13374; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:14913375; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013376; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:13713377; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1352 ; 4-byte Folded Spill13378; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:13613379; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13380; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1368 ; 4-byte Folded Spill13381; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13382; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1424 ; 4-byte Folded Spill13383; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13384; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1356 ; 4-byte Folded Spill13385; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13386; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1344 ; 4-byte Folded Spill13387; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013388; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v313389; ALIGNED-NEXT:    s_clause 0x113390; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:14713391; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:14813392; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1420 ; 4-byte Folded Spill13393; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013394; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:14513395; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1384 ; 4-byte Folded Spill13396; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:14413397; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13398; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1412 ; 4-byte Folded Spill13399; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13400; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1416 ; 4-byte Folded Spill13401; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13402; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1408 ; 4-byte Folded Spill13403; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13404; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Spill13405; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013406; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v713407; ALIGNED-NEXT:    v_lshl_or_b32 v4, v98, 8, v10013408; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013409; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v613410; ALIGNED-NEXT:    s_clause 0x213411; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:24813412; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:24913413; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:25013414; ALIGNED-NEXT:    v_lshl_or_b32 v6, v11, 8, v1313415; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1440 ; 4-byte Folded Spill13416; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v313417; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:15813418; ALIGNED-NEXT:    v_lshl_or_b32 v5, v15, 8, v2013419; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013420; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:15713421; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1448 ; 4-byte Folded Spill13422; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:15613423; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13424; ALIGNED-NEXT:    v_lshl_or_b32 v121, v9, 8, v713425; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13426; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1456 ; 4-byte Folded Spill13427; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13428; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1452 ; 4-byte Folded Spill13429; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13430; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Spill13431; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v013432; ALIGNED-NEXT:    v_lshl_or_b32 v1, v126, 8, v313433; ALIGNED-NEXT:    v_lshl_or_b32 v3, v99, 8, v10113434; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013435; ALIGNED-NEXT:    v_lshl_or_b32 v1, v120, 8, v11013436; ALIGNED-NEXT:    v_lshl_or_b32 v91, v4, 16, v313437; ALIGNED-NEXT:    v_lshl_or_b32 v3, v87, 8, v9613438; ALIGNED-NEXT:    v_lshl_or_b32 v4, v97, 8, v8613439; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1460 ; 4-byte Folded Spill13440; ALIGNED-NEXT:    v_lshl_or_b32 v0, v111, 8, v12313441; ALIGNED-NEXT:    v_lshl_or_b32 v76, v4, 16, v313442; ALIGNED-NEXT:    v_lshl_or_b32 v3, v70, 8, v8413443; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013444; ALIGNED-NEXT:    v_lshl_or_b32 v1, v94, 8, v9513445; ALIGNED-NEXT:    v_lshl_or_b32 v4, v68, 8, v8213446; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1464 ; 4-byte Folded Spill13447; ALIGNED-NEXT:    v_lshl_or_b32 v0, v106, 8, v10913448; ALIGNED-NEXT:    v_lshl_or_b32 v57, v4, 16, v313449; ALIGNED-NEXT:    v_lshl_or_b32 v3, v69, 8, v8313450; ALIGNED-NEXT:    v_lshl_or_b32 v4, v64, 8, v6513451; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013452; ALIGNED-NEXT:    v_lshl_or_b32 v1, v93, 8, v10413453; ALIGNED-NEXT:    v_lshl_or_b32 v41, v4, 16, v313454; ALIGNED-NEXT:    v_lshl_or_b32 v3, v67, 8, v8113455; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1468 ; 4-byte Folded Spill13456; ALIGNED-NEXT:    v_lshl_or_b32 v0, v105, 8, v10713457; ALIGNED-NEXT:    v_lshl_or_b32 v4, v55, 8, v6613458; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013459; ALIGNED-NEXT:    v_lshl_or_b32 v1, v79, 8, v8813460; ALIGNED-NEXT:    v_lshl_or_b32 v118, v4, 16, v313461; ALIGNED-NEXT:    v_lshl_or_b32 v3, v50, 8, v5313462; ALIGNED-NEXT:    v_lshl_or_b32 v4, v71, 8, v5113463; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1472 ; 4-byte Folded Spill13464; ALIGNED-NEXT:    v_lshl_or_b32 v0, v89, 8, v9013465; ALIGNED-NEXT:    v_lshl_or_b32 v102, v4, 16, v313466; ALIGNED-NEXT:    v_lshl_or_b32 v3, v39, 8, v4913467; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013468; ALIGNED-NEXT:    v_lshl_or_b32 v1, v78, 8, v7413469; ALIGNED-NEXT:    v_lshl_or_b32 v4, v36, 8, v3513470; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1476 ; 4-byte Folded Spill13471; ALIGNED-NEXT:    v_lshl_or_b32 v0, v75, 8, v7713472; ALIGNED-NEXT:    v_lshl_or_b32 v85, v4, 16, v313473; ALIGNED-NEXT:    v_lshl_or_b32 v3, v37, 8, v4813474; ALIGNED-NEXT:    v_lshl_or_b32 v4, v34, 8, v3813475; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013476; ALIGNED-NEXT:    v_lshl_or_b32 v1, v60, 8, v5913477; ALIGNED-NEXT:    v_lshl_or_b32 v80, v4, 16, v313478; ALIGNED-NEXT:    v_lshl_or_b32 v3, v31, 8, v3313479; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1480 ; 4-byte Folded Spill13480; ALIGNED-NEXT:    v_lshl_or_b32 v0, v63, 8, v7313481; ALIGNED-NEXT:    v_lshl_or_b32 v4, v30, 8, v3213482; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013483; ALIGNED-NEXT:    v_lshl_or_b32 v1, v58, 8, v6213484; ALIGNED-NEXT:    v_lshl_or_b32 v54, v4, 16, v313485; ALIGNED-NEXT:    v_lshl_or_b32 v3, v26, 8, v2813486; ALIGNED-NEXT:    v_lshl_or_b32 v4, v29, 8, v2513487; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Spill13488; ALIGNED-NEXT:    v_lshl_or_b32 v0, v61, 8, v7213489; ALIGNED-NEXT:    v_lshl_or_b32 v52, v4, 16, v313490; ALIGNED-NEXT:    v_lshl_or_b32 v3, v21, 8, v2313491; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013492; ALIGNED-NEXT:    v_lshl_or_b32 v1, v45, 8, v4613493; ALIGNED-NEXT:    v_lshl_or_b32 v4, v18, 8, v1713494; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1488 ; 4-byte Folded Spill13495; ALIGNED-NEXT:    v_lshl_or_b32 v0, v47, 8, v5613496; ALIGNED-NEXT:    v_lshl_or_b32 v27, v4, 16, v313497; ALIGNED-NEXT:    v_lshl_or_b32 v4, v19, 8, v2213498; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen13499; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013500; ALIGNED-NEXT:    v_lshl_or_b32 v1, v44, 8, v4013501; ALIGNED-NEXT:    v_lshl_or_b32 v24, v5, 16, v413502; ALIGNED-NEXT:    v_lshl_or_b32 v5, v12, 8, v1413503; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:713504; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1492 ; 4-byte Folded Spill13505; ALIGNED-NEXT:    v_lshl_or_b32 v0, v42, 8, v4313506; ALIGNED-NEXT:    buffer_store_dword v125, off, s[0:3], s32 offset:1324 ; 4-byte Folded Spill13507; ALIGNED-NEXT:    v_lshl_or_b32 v16, v6, 16, v513508; ALIGNED-NEXT:    v_lshl_or_b32 v6, v8, 8, v1013509; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:513510; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v013511; ALIGNED-NEXT:    v_lshl_or_b32 v1, v112, 8, v11313512; ALIGNED-NEXT:    buffer_store_dword v124, off, s[0:3], s32 offset:1296 ; 4-byte Folded Spill13513; ALIGNED-NEXT:    v_lshl_or_b32 v6, v121, 16, v613514; ALIGNED-NEXT:    buffer_load_ubyte v121, v2, s[0:3], 0 offen offset:113515; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1496 ; 4-byte Folded Spill13516; ALIGNED-NEXT:    v_lshl_or_b32 v0, v116, 8, v11913517; ALIGNED-NEXT:    v_lshl_or_b32 v108, v1, 16, v013518; ALIGNED-NEXT:    v_lshl_or_b32 v0, v115, 8, v11713519; ALIGNED-NEXT:    v_lshl_or_b32 v1, v103, 8, v11413520; ALIGNED-NEXT:    v_lshl_or_b32 v92, v1, 16, v013521; ALIGNED-NEXT:    s_clause 0x113522; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:213523; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:313524; ALIGNED-NEXT:    s_waitcnt vmcnt(5)13525; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1268 ; 4-byte Folded Spill13526; ALIGNED-NEXT:    s_waitcnt vmcnt(4)13527; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1340 ; 4-byte Folded Spill13528; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13529; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1308 ; 4-byte Folded Spill13530; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13531; ALIGNED-NEXT:    buffer_store_dword v121, off, s[0:3], s32 offset:1288 ; 4-byte Folded Spill13532; ALIGNED-NEXT:    v_lshl_or_b32 v121, v121, 8, v313533; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1213534; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13535; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1300 ; 4-byte Folded Spill13536; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13537; ALIGNED-NEXT:    v_lshl_or_b32 v122, v0, 8, v113538; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1292 ; 4-byte Folded Spill13539; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1513540; ALIGNED-NEXT:    v_lshl_or_b32 v0, v122, 16, v12113541; ALIGNED-NEXT:    v_lshl_or_b32 v121, v5, 8, v12413542; ALIGNED-NEXT:    v_lshl_or_b32 v122, v4, 8, v12513543; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1313544; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1320 ; 4-byte Folded Spill13545; ALIGNED-NEXT:    v_lshl_or_b32 v0, v122, 16, v12113546; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Spill13547; ALIGNED-NEXT:    s_clause 0x113548; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1413549; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1113550; ALIGNED-NEXT:    s_waitcnt vmcnt(4)13551; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1360 ; 4-byte Folded Spill13552; ALIGNED-NEXT:    s_waitcnt vmcnt(3)13553; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1376 ; 4-byte Folded Spill13554; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13555; ALIGNED-NEXT:    v_lshl_or_b32 v121, v4, 8, v313556; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:1013557; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1380 ; 4-byte Folded Spill13558; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13559; ALIGNED-NEXT:    v_lshl_or_b32 v122, v1, 8, v013560; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1372 ; 4-byte Folded Spill13561; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:913562; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13563; ALIGNED-NEXT:    v_mov_b32_e32 v4, v513564; ALIGNED-NEXT:    v_lshl_or_b32 v0, v122, 16, v12113565; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1396 ; 4-byte Folded Spill13566; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:813567; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13568; ALIGNED-NEXT:    v_lshl_or_b32 v122, v5, 8, v313569; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13570; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1392 ; 4-byte Folded Spill13571; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13572; ALIGNED-NEXT:    v_lshl_or_b32 v121, v1, 8, v013573; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1388 ; 4-byte Folded Spill13574; ALIGNED-NEXT:    s_clause 0x213575; ALIGNED-NEXT:    buffer_load_ubyte v124, v2, s[0:3], 0 offen offset:1813576; ALIGNED-NEXT:    buffer_load_ubyte v125, v2, s[0:3], 0 offen offset:1613577; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1713578; ALIGNED-NEXT:    v_add_nc_u32_e32 v2, 0x100, v213579; ALIGNED-NEXT:    v_lshl_or_b32 v0, v122, 16, v12113580; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Spill13581; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:23213582; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:23613583; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:22813584; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:22413585; ALIGNED-NEXT:    s_clause 0x1 ; 8-byte Folded Reload13586; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:70413587; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:70813588; ALIGNED-NEXT:    s_waitcnt vmcnt(4)13589; ALIGNED-NEXT:    v_lshl_or_b32 v0, v127, 8, v12413590; ALIGNED-NEXT:    s_waitcnt vmcnt(2)13591; ALIGNED-NEXT:    v_lshl_or_b32 v127, v1, 8, v12513592; ALIGNED-NEXT:    v_lshl_or_b32 v127, v0, 16, v12713593; ALIGNED-NEXT:    s_waitcnt vmcnt(1)13594; ALIGNED-NEXT:    v_add_co_u32 v121, vcc_lo, v5, s413595; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13596; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v122, null, s5, v6, vcc_lo13597; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x10013598; ALIGNED-NEXT:    v_add_co_u32 v5, vcc_lo, v121, 313599; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v6, null, 0, v122, vcc_lo13600; ALIGNED-NEXT:    flat_store_byte v[5:6], v7 offset:24713601; ALIGNED-NEXT:    flat_store_byte v[5:6], v9 offset:24813602; ALIGNED-NEXT:    flat_store_byte v[5:6], v8 offset:24613603; ALIGNED-NEXT:    flat_store_byte v[5:6], v11 offset:25213604; ALIGNED-NEXT:    flat_store_byte v[5:6], v12 offset:25013605; ALIGNED-NEXT:    flat_store_byte v[5:6], v13 offset:25113606; ALIGNED-NEXT:    flat_store_byte v[5:6], v14 offset:24913607; ALIGNED-NEXT:    flat_store_byte v[5:6], v10 offset:24513608; ALIGNED-NEXT:    flat_store_byte v[5:6], v17 offset:23913609; ALIGNED-NEXT:    flat_store_byte v[5:6], v18 offset:24013610; ALIGNED-NEXT:    flat_store_byte v[5:6], v21 offset:23813611; ALIGNED-NEXT:    flat_store_byte v[5:6], v15 offset:24413612; ALIGNED-NEXT:    flat_store_byte v[5:6], v19 offset:24213613; ALIGNED-NEXT:    flat_store_byte v[5:6], v20 offset:24313614; ALIGNED-NEXT:    flat_store_byte v[5:6], v22 offset:24113615; ALIGNED-NEXT:    flat_store_byte v[5:6], v23 offset:23713616; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:24813617; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:25213618; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:24413619; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:24013620; ALIGNED-NEXT:    flat_store_byte v[5:6], v25 offset:23113621; ALIGNED-NEXT:    flat_store_byte v[5:6], v29 offset:23213622; ALIGNED-NEXT:    flat_store_byte v[5:6], v26 offset:23013623; ALIGNED-NEXT:    flat_store_byte v[5:6], v30 offset:23613624; ALIGNED-NEXT:    flat_store_byte v[5:6], v31 offset:23413625; ALIGNED-NEXT:    flat_store_byte v[5:6], v32 offset:23513626; ALIGNED-NEXT:    flat_store_byte v[5:6], v33 offset:23313627; ALIGNED-NEXT:    flat_store_byte v[5:6], v28 offset:22913628; ALIGNED-NEXT:    flat_store_byte v[5:6], v35 offset:22313629; ALIGNED-NEXT:    flat_store_byte v[5:6], v36 offset:22413630; ALIGNED-NEXT:    flat_store_byte v[5:6], v39 offset:22213631; ALIGNED-NEXT:    flat_store_byte v[5:6], v34 offset:22813632; ALIGNED-NEXT:    flat_store_byte v[5:6], v37 offset:22613633; ALIGNED-NEXT:    flat_store_byte v[5:6], v38 offset:22713634; ALIGNED-NEXT:    flat_store_byte v[5:6], v48 offset:22513635; ALIGNED-NEXT:    flat_store_byte v[5:6], v49 offset:22113636; ALIGNED-NEXT:    buffer_store_dword v102, off, s[0:3], s32 offset:19213637; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:20413638; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:20013639; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:19613640; ALIGNED-NEXT:    flat_store_byte v[5:6], v70 offset:21013641; ALIGNED-NEXT:    flat_store_byte v[5:6], v68 offset:21213642; ALIGNED-NEXT:    flat_store_byte v[5:6], v50 offset:20613643; ALIGNED-NEXT:    flat_store_byte v[5:6], v71 offset:20813644; ALIGNED-NEXT:    flat_store_byte v[5:6], v51 offset:20713645; ALIGNED-NEXT:    flat_store_byte v[5:6], v82 offset:21113646; ALIGNED-NEXT:    flat_store_byte v[5:6], v84 offset:20913647; ALIGNED-NEXT:    flat_store_byte v[5:6], v65 offset:21513648; ALIGNED-NEXT:    flat_store_byte v[5:6], v64 offset:21613649; ALIGNED-NEXT:    flat_store_byte v[5:6], v69 offset:21413650; ALIGNED-NEXT:    flat_store_byte v[5:6], v55 offset:22013651; ALIGNED-NEXT:    flat_store_byte v[5:6], v67 offset:21813652; ALIGNED-NEXT:    flat_store_byte v[5:6], v66 offset:21913653; ALIGNED-NEXT:    flat_store_byte v[5:6], v81 offset:21713654; ALIGNED-NEXT:    flat_store_byte v[5:6], v83 offset:21313655; ALIGNED-NEXT:    flat_store_byte v[5:6], v53 offset:20513656; ALIGNED-NEXT:    buffer_store_dword v76, off, s[0:3], s32 offset:21613657; ALIGNED-NEXT:    buffer_store_dword v91, off, s[0:3], s32 offset:22013658; ALIGNED-NEXT:    buffer_store_dword v92, off, s[0:3], s32 offset:21213659; ALIGNED-NEXT:    buffer_store_dword v108, off, s[0:3], s32 offset:20813660; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1496 ; 4-byte Folded Reload13661; ALIGNED-NEXT:    flat_store_byte v[5:6], v86 offset:19913662; ALIGNED-NEXT:    flat_store_byte v[5:6], v97 offset:20013663; ALIGNED-NEXT:    flat_store_byte v[5:6], v87 offset:19813664; ALIGNED-NEXT:    flat_store_byte v[5:6], v98 offset:20413665; ALIGNED-NEXT:    flat_store_byte v[5:6], v99 offset:20213666; ALIGNED-NEXT:    flat_store_byte v[5:6], v100 offset:20313667; ALIGNED-NEXT:    flat_store_byte v[5:6], v101 offset:20113668; ALIGNED-NEXT:    flat_store_byte v[5:6], v96 offset:19713669; ALIGNED-NEXT:    flat_store_byte v[5:6], v113 offset:19113670; ALIGNED-NEXT:    flat_store_byte v[5:6], v112 offset:19213671; ALIGNED-NEXT:    flat_store_byte v[5:6], v116 offset:19013672; ALIGNED-NEXT:    flat_store_byte v[5:6], v103 offset:19613673; ALIGNED-NEXT:    flat_store_byte v[5:6], v115 offset:19413674; ALIGNED-NEXT:    flat_store_byte v[5:6], v114 offset:19513675; ALIGNED-NEXT:    flat_store_byte v[5:6], v117 offset:19313676; ALIGNED-NEXT:    flat_store_byte v[5:6], v119 offset:18913677; ALIGNED-NEXT:    s_addc_u32 s5, s5, 013678; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0x80013679; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13680; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:29613681; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1492 ; 4-byte Folded Reload13682; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13683; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:30013684; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1488 ; 4-byte Folded Reload13685; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13686; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:29213687; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Reload13688; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13689; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:28813690; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1480 ; 4-byte Folded Reload13691; ALIGNED-NEXT:    flat_store_byte v[5:6], v40 offset:18313692; ALIGNED-NEXT:    flat_store_byte v[5:6], v44 offset:18413693; ALIGNED-NEXT:    flat_store_byte v[5:6], v42 offset:18213694; ALIGNED-NEXT:    flat_store_byte v[5:6], v45 offset:18813695; ALIGNED-NEXT:    flat_store_byte v[5:6], v47 offset:18613696; ALIGNED-NEXT:    flat_store_byte v[5:6], v46 offset:18713697; ALIGNED-NEXT:    flat_store_byte v[5:6], v56 offset:18513698; ALIGNED-NEXT:    flat_store_byte v[5:6], v43 offset:18113699; ALIGNED-NEXT:    flat_store_byte v[5:6], v59 offset:17513700; ALIGNED-NEXT:    flat_store_byte v[5:6], v60 offset:17613701; ALIGNED-NEXT:    flat_store_byte v[5:6], v63 offset:17413702; ALIGNED-NEXT:    flat_store_byte v[5:6], v58 offset:18013703; ALIGNED-NEXT:    flat_store_byte v[5:6], v61 offset:17813704; ALIGNED-NEXT:    flat_store_byte v[5:6], v62 offset:17913705; ALIGNED-NEXT:    flat_store_byte v[5:6], v72 offset:17713706; ALIGNED-NEXT:    flat_store_byte v[5:6], v73 offset:17313707; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13708; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:31213709; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1476 ; 4-byte Folded Reload13710; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13711; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:31613712; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1472 ; 4-byte Folded Reload13713; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13714; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:30813715; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1468 ; 4-byte Folded Reload13716; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13717; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:30413718; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1464 ; 4-byte Folded Reload13719; ALIGNED-NEXT:    flat_store_byte v[5:6], v74 offset:16713720; ALIGNED-NEXT:    flat_store_byte v[5:6], v78 offset:16813721; ALIGNED-NEXT:    flat_store_byte v[5:6], v75 offset:16613722; ALIGNED-NEXT:    flat_store_byte v[5:6], v79 offset:17213723; ALIGNED-NEXT:    flat_store_byte v[5:6], v89 offset:17013724; ALIGNED-NEXT:    flat_store_byte v[5:6], v88 offset:17113725; ALIGNED-NEXT:    flat_store_byte v[5:6], v90 offset:16913726; ALIGNED-NEXT:    flat_store_byte v[5:6], v77 offset:16513727; ALIGNED-NEXT:    flat_store_byte v[5:6], v95 offset:15913728; ALIGNED-NEXT:    flat_store_byte v[5:6], v94 offset:16013729; ALIGNED-NEXT:    flat_store_byte v[5:6], v106 offset:15813730; ALIGNED-NEXT:    flat_store_byte v[5:6], v93 offset:16413731; ALIGNED-NEXT:    flat_store_byte v[5:6], v105 offset:16213732; ALIGNED-NEXT:    flat_store_byte v[5:6], v104 offset:16313733; ALIGNED-NEXT:    flat_store_byte v[5:6], v107 offset:16113734; ALIGNED-NEXT:    flat_store_byte v[5:6], v109 offset:15713735; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13736; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:26413737; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1460 ; 4-byte Folded Reload13738; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13739; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:26813740; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1448 ; 4-byte Folded Reload13741; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13742; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:26013743; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1440 ; 4-byte Folded Reload13744; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13745; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:25613746; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1452 ; 4-byte Folded Reload13747; ALIGNED-NEXT:    flat_store_byte v[5:6], v110 offset:15113748; ALIGNED-NEXT:    flat_store_byte v[5:6], v120 offset:15213749; ALIGNED-NEXT:    flat_store_byte v[5:6], v111 offset:15013750; ALIGNED-NEXT:    flat_store_byte v[5:6], v126 offset:15613751; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13752; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:15413753; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Reload13754; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13755; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:15513756; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Reload13757; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13758; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:15313759; ALIGNED-NEXT:    flat_store_byte v[5:6], v123 offset:14913760; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Reload13761; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13762; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:14313763; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1412 ; 4-byte Folded Reload13764; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13765; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:14413766; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Reload13767; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13768; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:14213769; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1436 ; 4-byte Folded Reload13770; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13771; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:14813772; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1424 ; 4-byte Folded Reload13773; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13774; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:14613775; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1432 ; 4-byte Folded Reload13776; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13777; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:14713778; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1416 ; 4-byte Folded Reload13779; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13780; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:14513781; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Reload13782; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13783; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:14113784; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1384 ; 4-byte Folded Reload13785; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13786; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:28013787; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1352 ; 4-byte Folded Reload13788; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13789; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:28413790; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1316 ; 4-byte Folded Reload13791; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13792; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:27613793; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1304 ; 4-byte Folded Reload13794; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13795; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:27213796; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1368 ; 4-byte Folded Reload13797; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13798; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13513799; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1364 ; 4-byte Folded Reload13800; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13801; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13613802; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1356 ; 4-byte Folded Reload13803; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13804; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13413805; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1336 ; 4-byte Folded Reload13806; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13807; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:14013808; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1328 ; 4-byte Folded Reload13809; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13810; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13813811; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1332 ; 4-byte Folded Reload13812; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13813; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13913814; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Reload13815; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13816; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13713817; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1344 ; 4-byte Folded Reload13818; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13819; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13313820; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1272 ; 4-byte Folded Reload13821; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13822; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:12713823; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1260 ; 4-byte Folded Reload13824; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13825; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:12813826; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1256 ; 4-byte Folded Reload13827; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13828; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:12613829; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1284 ; 4-byte Folded Reload13830; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13831; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13213832; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1276 ; 4-byte Folded Reload13833; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13834; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13013835; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1280 ; 4-byte Folded Reload13836; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13837; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:13113838; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1264 ; 4-byte Folded Reload13839; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13840; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:12913841; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1252 ; 4-byte Folded Reload13842; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13843; ALIGNED-NEXT:    flat_store_byte v[121:122], v0 offset:12813844; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Reload13845; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13846; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:36013847; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Reload13848; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13849; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:36413850; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1212 ; 4-byte Folded Reload13851; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13852; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:35613853; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Reload13854; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13855; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:35213856; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1244 ; 4-byte Folded Reload13857; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13858; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11913859; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1240 ; 4-byte Folded Reload13860; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13861; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:12013862; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1236 ; 4-byte Folded Reload13863; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13864; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11813865; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1224 ; 4-byte Folded Reload13866; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13867; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:12413868; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1216 ; 4-byte Folded Reload13869; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13870; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:12213871; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1220 ; 4-byte Folded Reload13872; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13873; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:12313874; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Reload13875; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13876; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:12113877; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Reload13878; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13879; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11713880; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Reload13881; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13882; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11113883; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Reload13884; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13885; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11213886; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1176 ; 4-byte Folded Reload13887; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13888; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11013889; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Reload13890; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13891; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11613892; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1192 ; 4-byte Folded Reload13893; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13894; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11413895; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1196 ; 4-byte Folded Reload13896; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13897; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11513898; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1184 ; 4-byte Folded Reload13899; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13900; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:11313901; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Reload13902; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13903; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10913904; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Reload13905; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13906; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:37613907; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1156 ; 4-byte Folded Reload13908; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13909; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:38013910; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1132 ; 4-byte Folded Reload13911; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13912; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:37213913; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Reload13914; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13915; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:36813916; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1164 ; 4-byte Folded Reload13917; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13918; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10313919; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Reload13920; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13921; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10413922; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Reload13923; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13924; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10213925; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Reload13926; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13927; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10813928; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1136 ; 4-byte Folded Reload13929; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13930; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10613931; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1140 ; 4-byte Folded Reload13932; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13933; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10713934; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Reload13935; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13936; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10513937; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Reload13938; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13939; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10113940; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1104 ; 4-byte Folded Reload13941; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13942; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9513943; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1100 ; 4-byte Folded Reload13944; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13945; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9613946; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1096 ; 4-byte Folded Reload13947; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13948; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9413949; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Reload13950; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13951; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:10013952; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1112 ; 4-byte Folded Reload13953; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13954; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9813955; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1116 ; 4-byte Folded Reload13956; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13957; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9913958; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1108 ; 4-byte Folded Reload13959; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13960; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9713961; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Reload13962; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13963; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9313964; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Reload13965; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13966; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:32813967; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Reload13968; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13969; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:33213970; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Reload13971; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13972; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:32413973; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Reload13974; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13975; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:32013976; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1076 ; 4-byte Folded Reload13977; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13978; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8713979; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1072 ; 4-byte Folded Reload13980; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13981; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8813982; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1068 ; 4-byte Folded Reload13983; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13984; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8613985; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1060 ; 4-byte Folded Reload13986; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13987; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9213988; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Reload13989; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13990; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9013991; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1056 ; 4-byte Folded Reload13992; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13993; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:9113994; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Reload13995; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13996; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8913997; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Reload13998; ALIGNED-NEXT:    s_waitcnt vmcnt(0)13999; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8514000; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1024 ; 4-byte Folded Reload14001; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14002; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7914003; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Reload14004; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14005; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8014006; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1016 ; 4-byte Folded Reload14007; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14008; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7814009; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1040 ; 4-byte Folded Reload14010; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14011; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8414012; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1032 ; 4-byte Folded Reload14013; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14014; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8214015; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1036 ; 4-byte Folded Reload14016; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14017; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8314018; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1028 ; 4-byte Folded Reload14019; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14020; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:8114021; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1012 ; 4-byte Folded Reload14022; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14023; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7714024; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Reload14025; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14026; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:34414027; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:996 ; 4-byte Folded Reload14028; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14029; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:34814030; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:984 ; 4-byte Folded Reload14031; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14032; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:34014033; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload14034; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14035; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:33614036; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1004 ; 4-byte Folded Reload14037; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14038; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7114039; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1000 ; 4-byte Folded Reload14040; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14041; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7214042; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:992 ; 4-byte Folded Reload14043; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14044; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7014045; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:980 ; 4-byte Folded Reload14046; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14047; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7614048; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload14049; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14050; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7414051; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Reload14052; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14053; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7514054; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload14055; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14056; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:7314057; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:988 ; 4-byte Folded Reload14058; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14059; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:6914060; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload14061; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14062; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:6314063; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload14064; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14065; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:6414066; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload14067; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14068; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:6214069; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload14070; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14071; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:6814072; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload14073; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14074; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:6614075; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload14076; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14077; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:6714078; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload14079; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14080; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:6514081; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload14082; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14083; ALIGNED-NEXT:    flat_store_byte v[121:122], v0 offset:6414084; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload14085; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14086; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:42414087; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload14088; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14089; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:42814090; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload14091; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14092; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:42014093; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload14094; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14095; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:41614096; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload14097; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14098; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5814099; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload14100; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14101; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5514102; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload14103; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14104; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5614105; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload14106; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14107; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5414108; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload14109; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14110; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:6014111; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload14112; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14113; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5914114; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload14115; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14116; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5714117; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload14118; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14119; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5314120; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload14121; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14122; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5014123; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload14124; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14125; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4714126; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload14127; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14128; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4814129; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload14130; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14131; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4614132; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload14133; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14134; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5214135; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload14136; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14137; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:5114138; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload14139; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14140; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4914141; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload14142; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14143; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4514144; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload14145; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14146; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:44414147; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload14148; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14149; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:44014150; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload14151; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14152; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:43614153; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload14154; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14155; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:43214156; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload14157; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14158; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4014159; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload14160; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14161; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3914162; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload14163; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14164; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3814165; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload14166; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14167; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3714168; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload14169; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14170; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4414171; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload14172; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14173; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4314174; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload14175; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14176; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4214177; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload14178; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14179; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4114180; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload14181; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14182; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3214183; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload14184; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14185; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3114186; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload14187; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14188; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3014189; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload14190; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14191; ALIGNED-NEXT:    flat_store_byte v[121:122], v0 offset:3214192; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload14193; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14194; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3614195; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload14196; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14197; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3514198; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload14199; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14200; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3414201; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload14202; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14203; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3314204; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload14205; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14206; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:39214207; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload14208; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14209; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:39614210; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload14211; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14212; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:38814213; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:38414214; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload14215; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14216; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2314217; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload14218; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14219; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2414220; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload14221; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14222; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2214223; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload14224; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14225; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2814226; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload14227; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14228; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2614229; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload14230; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14231; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2714232; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload14233; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14234; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2514235; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload14236; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14237; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2114238; ALIGNED-NEXT:    flat_store_byte v[5:6], v124 offset:1514239; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1428 ; 4-byte Folded Reload14240; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14241; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1614242; ALIGNED-NEXT:    flat_store_byte v[5:6], v1 offset:1414243; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload14244; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14245; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2014246; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload14247; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14248; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1814249; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload14250; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14251; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1914252; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload14253; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14254; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1714255; ALIGNED-NEXT:    flat_store_byte v[121:122], v125 offset:1614256; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Reload14257; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14258; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:40814259; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1396 ; 4-byte Folded Reload14260; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14261; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:41214262; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Reload14263; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14264; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:40414265; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1320 ; 4-byte Folded Reload14266; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14267; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:40014268; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1380 ; 4-byte Folded Reload14269; ALIGNED-NEXT:    flat_store_byte v[5:6], v3 offset:714270; ALIGNED-NEXT:    flat_store_byte v[5:6], v4 offset:814271; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14272; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1014273; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Reload14274; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14275; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:614276; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1376 ; 4-byte Folded Reload14277; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14278; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1214279; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1372 ; 4-byte Folded Reload14280; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14281; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:1114282; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1360 ; 4-byte Folded Reload14283; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14284; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:914285; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1388 ; 4-byte Folded Reload14286; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14287; ALIGNED-NEXT:    flat_store_byte v[121:122], v0 offset:814288; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1300 ; 4-byte Folded Reload14289; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14290; ALIGNED-NEXT:    flat_store_byte v[121:122], v0 offset:214291; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1292 ; 4-byte Folded Reload14292; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14293; ALIGNED-NEXT:    flat_store_byte v[5:6], v014294; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1288 ; 4-byte Folded Reload14295; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14296; ALIGNED-NEXT:    flat_store_byte v[121:122], v0 offset:114297; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1340 ; 4-byte Folded Reload14298; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14299; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:414300; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1308 ; 4-byte Folded Reload14301; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14302; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:214303; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1324 ; 4-byte Folded Reload14304; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14305; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:314306; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1296 ; 4-byte Folded Reload14307; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14308; ALIGNED-NEXT:    flat_store_byte v[121:122], v0 offset:414309; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1268 ; 4-byte Folded Reload14310; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14311; ALIGNED-NEXT:    flat_store_byte v[121:122], v014312; ALIGNED-NEXT:    s_cbranch_scc1 .LBB9_114313; ALIGNED-NEXT:  .LBB9_2: ; %Flow1014314; ALIGNED-NEXT:    s_andn2_saveexec_b32 s8, s614315; ALIGNED-NEXT:    s_cbranch_execz .LBB9_514316; ALIGNED-NEXT:  ; %bb.3: ; %memmove_bwd_loop.preheader14317; ALIGNED-NEXT:    v_add_nc_u32_e32 v6, 0x700, v214318; ALIGNED-NEXT:    s_movk_i32 s6, 0xff0014319; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x70014320; ALIGNED-NEXT:    s_mov_b32 s7, -114321; ALIGNED-NEXT:  .LBB9_4: ; %memmove_bwd_loop14322; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=114323; ALIGNED-NEXT:    s_clause 0x3e14324; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:2014325; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:2114326; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:2214327; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:2314328; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:2414329; ALIGNED-NEXT:    buffer_load_ubyte v10, v6, s[0:3], 0 offen offset:2514330; ALIGNED-NEXT:    buffer_load_ubyte v127, v6, s[0:3], 0 offen offset:1914331; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:2814332; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:2914333; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:3014334; ALIGNED-NEXT:    buffer_load_ubyte v9, v6, s[0:3], 0 offen offset:3114335; ALIGNED-NEXT:    buffer_load_ubyte v12, v6, s[0:3], 0 offen offset:2614336; ALIGNED-NEXT:    buffer_load_ubyte v14, v6, s[0:3], 0 offen offset:3214337; ALIGNED-NEXT:    buffer_load_ubyte v15, v6, s[0:3], 0 offen offset:3314338; ALIGNED-NEXT:    buffer_load_ubyte v11, v6, s[0:3], 0 offen offset:2714339; ALIGNED-NEXT:    buffer_load_ubyte v19, v6, s[0:3], 0 offen offset:3414340; ALIGNED-NEXT:    buffer_load_ubyte v20, v6, s[0:3], 0 offen offset:3514341; ALIGNED-NEXT:    buffer_load_ubyte v13, v6, s[0:3], 0 offen offset:3614342; ALIGNED-NEXT:    buffer_load_ubyte v16, v6, s[0:3], 0 offen offset:3714343; ALIGNED-NEXT:    buffer_load_ubyte v17, v6, s[0:3], 0 offen offset:3814344; ALIGNED-NEXT:    buffer_load_ubyte v18, v6, s[0:3], 0 offen offset:3914345; ALIGNED-NEXT:    buffer_load_ubyte v22, v6, s[0:3], 0 offen offset:4014346; ALIGNED-NEXT:    buffer_load_ubyte v26, v6, s[0:3], 0 offen offset:4114347; ALIGNED-NEXT:    buffer_load_ubyte v27, v6, s[0:3], 0 offen offset:4214348; ALIGNED-NEXT:    buffer_load_ubyte v28, v6, s[0:3], 0 offen offset:4314349; ALIGNED-NEXT:    buffer_load_ubyte v21, v6, s[0:3], 0 offen offset:4414350; ALIGNED-NEXT:    buffer_load_ubyte v23, v6, s[0:3], 0 offen offset:4514351; ALIGNED-NEXT:    buffer_load_ubyte v24, v6, s[0:3], 0 offen offset:4614352; ALIGNED-NEXT:    buffer_load_ubyte v25, v6, s[0:3], 0 offen offset:4714353; ALIGNED-NEXT:    buffer_load_ubyte v30, v6, s[0:3], 0 offen offset:4814354; ALIGNED-NEXT:    buffer_load_ubyte v31, v6, s[0:3], 0 offen offset:4914355; ALIGNED-NEXT:    buffer_load_ubyte v33, v6, s[0:3], 0 offen offset:5014356; ALIGNED-NEXT:    buffer_load_ubyte v34, v6, s[0:3], 0 offen offset:5114357; ALIGNED-NEXT:    buffer_load_ubyte v32, v6, s[0:3], 0 offen offset:5214358; ALIGNED-NEXT:    buffer_load_ubyte v37, v6, s[0:3], 0 offen offset:5314359; ALIGNED-NEXT:    buffer_load_ubyte v35, v6, s[0:3], 0 offen offset:5414360; ALIGNED-NEXT:    buffer_load_ubyte v36, v6, s[0:3], 0 offen offset:5514361; ALIGNED-NEXT:    buffer_load_ubyte v38, v6, s[0:3], 0 offen offset:5614362; ALIGNED-NEXT:    buffer_load_ubyte v51, v6, s[0:3], 0 offen offset:5714363; ALIGNED-NEXT:    buffer_load_ubyte v39, v6, s[0:3], 0 offen offset:6014364; ALIGNED-NEXT:    buffer_load_ubyte v50, v6, s[0:3], 0 offen offset:6114365; ALIGNED-NEXT:    buffer_load_ubyte v48, v6, s[0:3], 0 offen offset:6214366; ALIGNED-NEXT:    buffer_load_ubyte v49, v6, s[0:3], 0 offen offset:6314367; ALIGNED-NEXT:    buffer_load_ubyte v52, v6, s[0:3], 0 offen offset:5814368; ALIGNED-NEXT:    buffer_load_ubyte v29, v6, s[0:3], 0 offen offset:6414369; ALIGNED-NEXT:    buffer_load_ubyte v55, v6, s[0:3], 0 offen offset:6514370; ALIGNED-NEXT:    buffer_load_ubyte v53, v6, s[0:3], 0 offen offset:5914371; ALIGNED-NEXT:    buffer_load_ubyte v66, v6, s[0:3], 0 offen offset:6614372; ALIGNED-NEXT:    buffer_load_ubyte v67, v6, s[0:3], 0 offen offset:6714373; ALIGNED-NEXT:    buffer_load_ubyte v54, v6, s[0:3], 0 offen offset:6814374; ALIGNED-NEXT:    buffer_load_ubyte v64, v6, s[0:3], 0 offen offset:6914375; ALIGNED-NEXT:    buffer_load_ubyte v65, v6, s[0:3], 0 offen offset:7014376; ALIGNED-NEXT:    buffer_load_ubyte v68, v6, s[0:3], 0 offen offset:7114377; ALIGNED-NEXT:    buffer_load_ubyte v69, v6, s[0:3], 0 offen offset:7614378; ALIGNED-NEXT:    buffer_load_ubyte v81, v6, s[0:3], 0 offen offset:7514379; ALIGNED-NEXT:    buffer_load_ubyte v70, v6, s[0:3], 0 offen offset:7714380; ALIGNED-NEXT:    buffer_load_ubyte v71, v6, s[0:3], 0 offen offset:7814381; ALIGNED-NEXT:    buffer_load_ubyte v80, v6, s[0:3], 0 offen offset:7914382; ALIGNED-NEXT:    buffer_load_ubyte v123, v6, s[0:3], 0 offen offset:15214383; ALIGNED-NEXT:    buffer_load_ubyte v122, v6, s[0:3], 0 offen offset:15314384; ALIGNED-NEXT:    buffer_load_ubyte v111, v6, s[0:3], 0 offen offset:15414385; ALIGNED-NEXT:    buffer_load_ubyte v108, v6, s[0:3], 0 offen offset:16014386; ALIGNED-NEXT:    buffer_load_ubyte v105, v6, s[0:3], 0 offen offset:16114387; ALIGNED-NEXT:    s_clause 0x3414388; ALIGNED-NEXT:    buffer_load_ubyte v94, v6, s[0:3], 0 offen offset:16214389; ALIGNED-NEXT:    buffer_load_ubyte v92, v6, s[0:3], 0 offen offset:16314390; ALIGNED-NEXT:    buffer_load_ubyte v106, v6, s[0:3], 0 offen offset:16414391; ALIGNED-NEXT:    buffer_load_ubyte v95, v6, s[0:3], 0 offen offset:16514392; ALIGNED-NEXT:    buffer_load_ubyte v104, v6, s[0:3], 0 offen offset:16614393; ALIGNED-NEXT:    buffer_load_ubyte v91, v6, s[0:3], 0 offen offset:16714394; ALIGNED-NEXT:    buffer_load_ubyte v78, v6, s[0:3], 0 offen offset:17214395; ALIGNED-NEXT:    buffer_load_ubyte v77, v6, s[0:3], 0 offen offset:17314396; ALIGNED-NEXT:    buffer_load_ubyte v76, v6, s[0:3], 0 offen offset:17414397; ALIGNED-NEXT:    buffer_load_ubyte v75, v6, s[0:3], 0 offen offset:17514398; ALIGNED-NEXT:    buffer_load_ubyte v74, v6, s[0:3], 0 offen offset:17114399; ALIGNED-NEXT:    buffer_load_ubyte v73, v6, s[0:3], 0 offen offset:16814400; ALIGNED-NEXT:    buffer_load_ubyte v63, v6, s[0:3], 0 offen offset:16914401; ALIGNED-NEXT:    buffer_load_ubyte v62, v6, s[0:3], 0 offen offset:17014402; ALIGNED-NEXT:    buffer_load_ubyte v61, v6, s[0:3], 0 offen offset:17614403; ALIGNED-NEXT:    buffer_load_ubyte v59, v6, s[0:3], 0 offen offset:17714404; ALIGNED-NEXT:    buffer_load_ubyte v47, v6, s[0:3], 0 offen offset:17814405; ALIGNED-NEXT:    buffer_load_ubyte v56, v6, s[0:3], 0 offen offset:17914406; ALIGNED-NEXT:    buffer_load_ubyte v60, v6, s[0:3], 0 offen offset:18014407; ALIGNED-NEXT:    buffer_load_ubyte v57, v6, s[0:3], 0 offen offset:18114408; ALIGNED-NEXT:    buffer_load_ubyte v58, v6, s[0:3], 0 offen offset:18214409; ALIGNED-NEXT:    buffer_load_ubyte v46, v6, s[0:3], 0 offen offset:18314410; ALIGNED-NEXT:    buffer_load_ubyte v44, v6, s[0:3], 0 offen offset:18814411; ALIGNED-NEXT:    buffer_load_ubyte v43, v6, s[0:3], 0 offen offset:18914412; ALIGNED-NEXT:    buffer_load_ubyte v42, v6, s[0:3], 0 offen offset:19014413; ALIGNED-NEXT:    buffer_load_ubyte v41, v6, s[0:3], 0 offen offset:19114414; ALIGNED-NEXT:    buffer_load_ubyte v40, v6, s[0:3], 0 offen offset:18714415; ALIGNED-NEXT:    buffer_load_ubyte v119, v6, s[0:3], 0 offen offset:18414416; ALIGNED-NEXT:    buffer_load_ubyte v118, v6, s[0:3], 0 offen offset:18514417; ALIGNED-NEXT:    buffer_load_ubyte v116, v6, s[0:3], 0 offen offset:18614418; ALIGNED-NEXT:    buffer_load_ubyte v115, v6, s[0:3], 0 offen offset:19214419; ALIGNED-NEXT:    buffer_load_ubyte v112, v6, s[0:3], 0 offen offset:19314420; ALIGNED-NEXT:    buffer_load_ubyte v101, v6, s[0:3], 0 offen offset:19414421; ALIGNED-NEXT:    buffer_load_ubyte v100, v6, s[0:3], 0 offen offset:19514422; ALIGNED-NEXT:    buffer_load_ubyte v113, v6, s[0:3], 0 offen offset:19614423; ALIGNED-NEXT:    buffer_load_ubyte v103, v6, s[0:3], 0 offen offset:19714424; ALIGNED-NEXT:    buffer_load_ubyte v102, v6, s[0:3], 0 offen offset:19814425; ALIGNED-NEXT:    buffer_load_ubyte v99, v6, s[0:3], 0 offen offset:19914426; ALIGNED-NEXT:    buffer_load_ubyte v97, v6, s[0:3], 0 offen offset:20414427; ALIGNED-NEXT:    buffer_load_ubyte v87, v6, s[0:3], 0 offen offset:20514428; ALIGNED-NEXT:    buffer_load_ubyte v96, v6, s[0:3], 0 offen offset:20614429; ALIGNED-NEXT:    buffer_load_ubyte v86, v6, s[0:3], 0 offen offset:20714430; ALIGNED-NEXT:    buffer_load_ubyte v85, v6, s[0:3], 0 offen offset:20314431; ALIGNED-NEXT:    buffer_load_ubyte v84, v6, s[0:3], 0 offen offset:20014432; ALIGNED-NEXT:    buffer_load_ubyte v83, v6, s[0:3], 0 offen offset:20114433; ALIGNED-NEXT:    buffer_load_ubyte v82, v6, s[0:3], 0 offen offset:20214434; ALIGNED-NEXT:    buffer_load_ubyte v124, v6, s[0:3], 0 offen offset:15514435; ALIGNED-NEXT:    buffer_load_ubyte v120, v6, s[0:3], 0 offen offset:214436; ALIGNED-NEXT:    buffer_load_ubyte v89, v6, s[0:3], 0 offen offset:314437; ALIGNED-NEXT:    buffer_load_ubyte v93, v6, s[0:3], 0 offen offset:414438; ALIGNED-NEXT:    buffer_load_ubyte v109, v6, s[0:3], 0 offen offset:514439; ALIGNED-NEXT:    buffer_load_ubyte v110, v6, s[0:3], 0 offen offset:614440; ALIGNED-NEXT:    buffer_load_ubyte v121, v6, s[0:3], 0 offen offset:714441; ALIGNED-NEXT:    s_waitcnt vmcnt(62)14442; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill14443; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill14444; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill14445; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill14446; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill14447; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill14448; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014449; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill14450; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill14451; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill14452; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill14453; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill14454; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill14455; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v214456; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill14457; ALIGNED-NEXT:    v_lshl_or_b32 v2, v7, 8, v414458; ALIGNED-NEXT:    v_lshl_or_b32 v3, v9, 8, v814459; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill14460; ALIGNED-NEXT:    v_lshl_or_b32 v4, v10, 8, v514461; ALIGNED-NEXT:    v_lshl_or_b32 v5, v11, 8, v1214462; ALIGNED-NEXT:    v_lshl_or_b32 v7, v15, 8, v1414463; ALIGNED-NEXT:    v_lshl_or_b32 v8, v20, 8, v1914464; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v1314465; ALIGNED-NEXT:    v_lshl_or_b32 v10, v18, 8, v1714466; ALIGNED-NEXT:    v_lshl_or_b32 v11, v26, 8, v2214467; ALIGNED-NEXT:    v_lshl_or_b32 v12, v28, 8, v2714468; ALIGNED-NEXT:    v_lshl_or_b32 v13, v23, 8, v2114469; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014470; ALIGNED-NEXT:    v_lshl_or_b32 v14, v25, 8, v2414471; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v214472; ALIGNED-NEXT:    v_lshl_or_b32 v2, v5, 16, v414473; ALIGNED-NEXT:    v_lshl_or_b32 v3, v8, 16, v714474; ALIGNED-NEXT:    v_lshl_or_b32 v4, v10, 16, v914475; ALIGNED-NEXT:    v_lshl_or_b32 v5, v12, 16, v1114476; ALIGNED-NEXT:    v_lshl_or_b32 v7, v14, 16, v1314477; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill14478; ALIGNED-NEXT:    v_lshl_or_b32 v15, v31, 8, v3014479; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill14480; ALIGNED-NEXT:    v_lshl_or_b32 v0, v34, 8, v3314481; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill14482; ALIGNED-NEXT:    v_lshl_or_b32 v1, v37, 8, v3214483; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill14484; ALIGNED-NEXT:    v_lshl_or_b32 v2, v36, 8, v3514485; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill14486; ALIGNED-NEXT:    v_lshl_or_b32 v3, v50, 8, v3914487; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill14488; ALIGNED-NEXT:    v_lshl_or_b32 v4, v49, 8, v4814489; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill14490; ALIGNED-NEXT:    v_lshl_or_b32 v5, v51, 8, v3814491; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill14492; ALIGNED-NEXT:    v_lshl_or_b32 v7, v53, 8, v5214493; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 16, v1514494; ALIGNED-NEXT:    v_lshl_or_b32 v1, v2, 16, v114495; ALIGNED-NEXT:    v_lshl_or_b32 v2, v4, 16, v314496; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:8514497; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 16, v514498; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill14499; ALIGNED-NEXT:    v_lshl_or_b32 v0, v55, 8, v2914500; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill14501; ALIGNED-NEXT:    v_lshl_or_b32 v1, v67, 8, v6614502; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill14503; ALIGNED-NEXT:    v_lshl_or_b32 v2, v64, 8, v5414504; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill14505; ALIGNED-NEXT:    v_lshl_or_b32 v3, v68, 8, v6514506; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014507; ALIGNED-NEXT:    s_clause 0x214508; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:8614509; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:8214510; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:8714511; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v214512; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill14513; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:8314514; ALIGNED-NEXT:    s_waitcnt vmcnt(62)14515; ALIGNED-NEXT:    v_lshl_or_b32 v0, v70, 8, v6914516; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:7414517; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:988 ; 4-byte Folded Spill14518; ALIGNED-NEXT:    v_lshl_or_b32 v1, v80, 8, v7114519; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill14520; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill14521; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill14522; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill14523; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014524; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:7314525; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill14526; ALIGNED-NEXT:    s_clause 0x414527; ALIGNED-NEXT:    buffer_load_ubyte v80, v6, s[0:3], 0 offen offset:21214528; ALIGNED-NEXT:    buffer_load_ubyte v67, v6, s[0:3], 0 offen offset:21314529; ALIGNED-NEXT:    buffer_load_ubyte v70, v6, s[0:3], 0 offen offset:21414530; ALIGNED-NEXT:    buffer_load_ubyte v64, v6, s[0:3], 0 offen offset:21514531; ALIGNED-NEXT:    buffer_load_ubyte v65, v6, s[0:3], 0 offen offset:21114532; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill14533; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:992 ; 4-byte Folded Spill14534; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:7214535; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill14536; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill14537; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill14538; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill14539; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill14540; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill14541; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill14542; ALIGNED-NEXT:    s_clause 0x714543; ALIGNED-NEXT:    buffer_load_ubyte v71, v6, s[0:3], 0 offen offset:21614544; ALIGNED-NEXT:    buffer_load_ubyte v66, v6, s[0:3], 0 offen offset:21714545; ALIGNED-NEXT:    buffer_load_ubyte v53, v6, s[0:3], 0 offen offset:21814546; ALIGNED-NEXT:    buffer_load_ubyte v52, v6, s[0:3], 0 offen offset:21914547; ALIGNED-NEXT:    buffer_load_ubyte v69, v6, s[0:3], 0 offen offset:22014548; ALIGNED-NEXT:    buffer_load_ubyte v55, v6, s[0:3], 0 offen offset:22114549; ALIGNED-NEXT:    buffer_load_ubyte v54, v6, s[0:3], 0 offen offset:22214550; ALIGNED-NEXT:    buffer_load_ubyte v51, v6, s[0:3], 0 offen offset:22314551; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill14552; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill14553; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill14554; ALIGNED-NEXT:    s_clause 0x214555; ALIGNED-NEXT:    buffer_load_ubyte v49, v6, s[0:3], 0 offen offset:20814556; ALIGNED-NEXT:    buffer_load_ubyte v38, v6, s[0:3], 0 offen offset:20914557; ALIGNED-NEXT:    buffer_load_ubyte v39, v6, s[0:3], 0 offen offset:21014558; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill14559; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill14560; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill14561; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill14562; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill14563; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill14564; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill14565; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill14566; ALIGNED-NEXT:    s_clause 0x714567; ALIGNED-NEXT:    buffer_load_ubyte v37, v6, s[0:3], 0 offen offset:22414568; ALIGNED-NEXT:    buffer_load_ubyte v35, v6, s[0:3], 0 offen offset:22514569; ALIGNED-NEXT:    buffer_load_ubyte v31, v6, s[0:3], 0 offen offset:22614570; ALIGNED-NEXT:    buffer_load_ubyte v32, v6, s[0:3], 0 offen offset:22714571; ALIGNED-NEXT:    buffer_load_ubyte v36, v6, s[0:3], 0 offen offset:22814572; ALIGNED-NEXT:    buffer_load_ubyte v33, v6, s[0:3], 0 offen offset:22914573; ALIGNED-NEXT:    buffer_load_ubyte v34, v6, s[0:3], 0 offen offset:23014574; ALIGNED-NEXT:    buffer_load_ubyte v30, v6, s[0:3], 0 offen offset:23114575; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill14576; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill14577; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill14578; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill14579; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill14580; ALIGNED-NEXT:    s_clause 0x414581; ALIGNED-NEXT:    buffer_load_ubyte v29, v6, s[0:3], 0 offen offset:23614582; ALIGNED-NEXT:    buffer_load_ubyte v27, v6, s[0:3], 0 offen offset:23714583; ALIGNED-NEXT:    buffer_load_ubyte v28, v6, s[0:3], 0 offen offset:23814584; ALIGNED-NEXT:    buffer_load_ubyte v26, v6, s[0:3], 0 offen offset:23914585; ALIGNED-NEXT:    buffer_load_ubyte v25, v6, s[0:3], 0 offen offset:23514586; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill14587; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill14588; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill14589; ALIGNED-NEXT:    s_clause 0x214590; ALIGNED-NEXT:    buffer_load_ubyte v24, v6, s[0:3], 0 offen offset:23214591; ALIGNED-NEXT:    buffer_load_ubyte v23, v6, s[0:3], 0 offen offset:23314592; ALIGNED-NEXT:    buffer_load_ubyte v22, v6, s[0:3], 0 offen offset:23414593; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill14594; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill14595; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill14596; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill14597; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill14598; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill14599; ALIGNED-NEXT:    s_clause 0xc14600; ALIGNED-NEXT:    buffer_load_ubyte v21, v6, s[0:3], 0 offen offset:24014601; ALIGNED-NEXT:    buffer_load_ubyte v19, v6, s[0:3], 0 offen offset:24114602; ALIGNED-NEXT:    buffer_load_ubyte v15, v6, s[0:3], 0 offen offset:24214603; ALIGNED-NEXT:    buffer_load_ubyte v16, v6, s[0:3], 0 offen offset:24314604; ALIGNED-NEXT:    buffer_load_ubyte v20, v6, s[0:3], 0 offen offset:24414605; ALIGNED-NEXT:    buffer_load_ubyte v17, v6, s[0:3], 0 offen offset:24514606; ALIGNED-NEXT:    buffer_load_ubyte v18, v6, s[0:3], 0 offen offset:24614607; ALIGNED-NEXT:    buffer_load_ubyte v14, v6, s[0:3], 0 offen offset:24714608; ALIGNED-NEXT:    buffer_load_ubyte v13, v6, s[0:3], 0 offen offset:25214609; ALIGNED-NEXT:    buffer_load_ubyte v11, v6, s[0:3], 0 offen offset:25314610; ALIGNED-NEXT:    buffer_load_ubyte v12, v6, s[0:3], 0 offen offset:25414611; ALIGNED-NEXT:    buffer_load_ubyte v10, v6, s[0:3], 0 offen offset:25514612; ALIGNED-NEXT:    buffer_load_ubyte v9, v6, s[0:3], 0 offen offset:25114613; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:1000 ; 4-byte Folded Spill14614; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill14615; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill14616; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill14617; ALIGNED-NEXT:    s_waitcnt vmcnt(57)14618; ALIGNED-NEXT:    buffer_store_dword v89, off, s[0:3], s32 offset:1364 ; 4-byte Folded Spill14619; ALIGNED-NEXT:    v_lshl_or_b32 v89, v89, 8, v12014620; ALIGNED-NEXT:    buffer_store_dword v120, off, s[0:3], s32 offset:1376 ; 4-byte Folded Spill14621; ALIGNED-NEXT:    s_clause 0x114622; ALIGNED-NEXT:    buffer_load_ubyte v126, v6, s[0:3], 0 offen offset:1514623; ALIGNED-NEXT:    buffer_load_ubyte v120, v6, s[0:3], 0 offen offset:1114624; ALIGNED-NEXT:    s_waitcnt vmcnt(57)14625; ALIGNED-NEXT:    buffer_store_dword v109, off, s[0:3], s32 offset:1384 ; 4-byte Folded Spill14626; ALIGNED-NEXT:    s_waitcnt vmcnt(56)14627; ALIGNED-NEXT:    buffer_store_dword v110, off, s[0:3], s32 offset:1388 ; 4-byte Folded Spill14628; ALIGNED-NEXT:    s_waitcnt vmcnt(55)14629; ALIGNED-NEXT:    buffer_store_dword v121, off, s[0:3], s32 offset:1392 ; 4-byte Folded Spill14630; ALIGNED-NEXT:    buffer_store_dword v93, off, s[0:3], s32 offset:1372 ; 4-byte Folded Spill14631; ALIGNED-NEXT:    s_waitcnt vmcnt(54)14632; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1032 ; 4-byte Folded Spill14633; ALIGNED-NEXT:    s_waitcnt vmcnt(53)14634; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1036 ; 4-byte Folded Spill14635; ALIGNED-NEXT:    s_waitcnt vmcnt(52)14636; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1028 ; 4-byte Folded Spill14637; ALIGNED-NEXT:    s_waitcnt vmcnt(51)14638; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1040 ; 4-byte Folded Spill14639; ALIGNED-NEXT:    s_waitcnt vmcnt(50)14640; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1020 ; 4-byte Folded Spill14641; ALIGNED-NEXT:    s_waitcnt vmcnt(49)14642; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1004 ; 4-byte Folded Spill14643; ALIGNED-NEXT:    s_waitcnt vmcnt(48)14644; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:996 ; 4-byte Folded Spill14645; ALIGNED-NEXT:    s_waitcnt vmcnt(42)14646; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill14647; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014648; ALIGNED-NEXT:    v_lshl_or_b32 v1, v81, 8, v214649; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:8414650; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014651; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:8114652; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Spill14653; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:8014654; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14655; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1024 ; 4-byte Folded Spill14656; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14657; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1016 ; 4-byte Folded Spill14658; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14659; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1012 ; 4-byte Folded Spill14660; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014661; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v714662; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:9814663; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014664; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v514665; ALIGNED-NEXT:    s_clause 0x114666; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:10214667; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:10314668; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Spill14669; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 8, v214670; ALIGNED-NEXT:    s_clause 0x114671; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:9414672; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:9514673; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014674; ALIGNED-NEXT:    s_clause 0x114675; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:9314676; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:9114677; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Spill14678; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:9214679; ALIGNED-NEXT:    s_waitcnt vmcnt(6)14680; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1116 ; 4-byte Folded Spill14681; ALIGNED-NEXT:    s_waitcnt vmcnt(5)14682; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1120 ; 4-byte Folded Spill14683; ALIGNED-NEXT:    s_waitcnt vmcnt(4)14684; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1060 ; 4-byte Folded Spill14685; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14686; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1064 ; 4-byte Folded Spill14687; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14688; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1056 ; 4-byte Folded Spill14689; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14690; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Spill14691; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014692; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v214693; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:9014694; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1080 ; 4-byte Folded Spill14695; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:10114696; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014697; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:8914698; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1072 ; 4-byte Folded Spill14699; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:8814700; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14701; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1084 ; 4-byte Folded Spill14702; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14703; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1112 ; 4-byte Folded Spill14704; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14705; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1076 ; 4-byte Folded Spill14706; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14707; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1068 ; 4-byte Folded Spill14708; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014709; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v214710; ALIGNED-NEXT:    s_clause 0x114711; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:9914712; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:10014713; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1108 ; 4-byte Folded Spill14714; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014715; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:9714716; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Spill14717; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:9614718; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14719; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1100 ; 4-byte Folded Spill14720; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14721; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1104 ; 4-byte Folded Spill14722; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14723; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1096 ; 4-byte Folded Spill14724; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14725; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Spill14726; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014727; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v714728; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:11414729; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014730; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v514731; ALIGNED-NEXT:    s_clause 0x114732; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:11814733; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:11914734; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Spill14735; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 8, v214736; ALIGNED-NEXT:    s_clause 0x114737; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:11014738; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:11114739; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014740; ALIGNED-NEXT:    s_clause 0x114741; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:10914742; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:10714743; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1132 ; 4-byte Folded Spill14744; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:10814745; ALIGNED-NEXT:    s_waitcnt vmcnt(6)14746; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1196 ; 4-byte Folded Spill14747; ALIGNED-NEXT:    s_waitcnt vmcnt(5)14748; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1200 ; 4-byte Folded Spill14749; ALIGNED-NEXT:    s_waitcnt vmcnt(4)14750; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1140 ; 4-byte Folded Spill14751; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14752; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1144 ; 4-byte Folded Spill14753; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14754; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1136 ; 4-byte Folded Spill14755; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14756; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Spill14757; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014758; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v214759; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:10614760; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1160 ; 4-byte Folded Spill14761; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:11714762; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014763; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:10514764; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Spill14765; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:10414766; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14767; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1164 ; 4-byte Folded Spill14768; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14769; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1192 ; 4-byte Folded Spill14770; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14771; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1156 ; 4-byte Folded Spill14772; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14773; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Spill14774; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014775; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v214776; ALIGNED-NEXT:    s_clause 0x114777; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:11514778; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:11614779; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1188 ; 4-byte Folded Spill14780; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014781; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:11314782; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Spill14783; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:11214784; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14785; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1180 ; 4-byte Folded Spill14786; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14787; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1184 ; 4-byte Folded Spill14788; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14789; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1176 ; 4-byte Folded Spill14790; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14791; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Spill14792; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014793; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v714794; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:13014795; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014796; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v514797; ALIGNED-NEXT:    s_clause 0x114798; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:13414799; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:13514800; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Spill14801; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 8, v214802; ALIGNED-NEXT:    s_clause 0x114803; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:12614804; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:12714805; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014806; ALIGNED-NEXT:    s_clause 0x114807; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:12514808; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:12314809; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1212 ; 4-byte Folded Spill14810; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:12414811; ALIGNED-NEXT:    s_waitcnt vmcnt(6)14812; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1276 ; 4-byte Folded Spill14813; ALIGNED-NEXT:    s_waitcnt vmcnt(5)14814; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1280 ; 4-byte Folded Spill14815; ALIGNED-NEXT:    s_waitcnt vmcnt(4)14816; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1220 ; 4-byte Folded Spill14817; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14818; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1224 ; 4-byte Folded Spill14819; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14820; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1216 ; 4-byte Folded Spill14821; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14822; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Spill14823; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014824; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v214825; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:12214826; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1240 ; 4-byte Folded Spill14827; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:13314828; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014829; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:12114830; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Spill14831; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:12014832; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14833; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1244 ; 4-byte Folded Spill14834; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14835; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1272 ; 4-byte Folded Spill14836; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14837; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1236 ; 4-byte Folded Spill14838; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14839; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Spill14840; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014841; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v214842; ALIGNED-NEXT:    s_clause 0x114843; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:13114844; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:13214845; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1268 ; 4-byte Folded Spill14846; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014847; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:12914848; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Spill14849; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:12814850; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14851; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1260 ; 4-byte Folded Spill14852; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14853; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1264 ; 4-byte Folded Spill14854; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14855; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1256 ; 4-byte Folded Spill14856; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14857; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1252 ; 4-byte Folded Spill14858; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014859; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v714860; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:14614861; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014862; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v514863; ALIGNED-NEXT:    s_clause 0x114864; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:15014865; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:15114866; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1284 ; 4-byte Folded Spill14867; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 8, v214868; ALIGNED-NEXT:    s_clause 0x114869; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:14214870; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:14314871; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014872; ALIGNED-NEXT:    s_clause 0x114873; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:14114874; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:13914875; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1292 ; 4-byte Folded Spill14876; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:14014877; ALIGNED-NEXT:    s_waitcnt vmcnt(6)14878; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1368 ; 4-byte Folded Spill14879; ALIGNED-NEXT:    s_waitcnt vmcnt(5)14880; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1380 ; 4-byte Folded Spill14881; ALIGNED-NEXT:    s_waitcnt vmcnt(4)14882; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1300 ; 4-byte Folded Spill14883; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14884; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1304 ; 4-byte Folded Spill14885; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14886; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1296 ; 4-byte Folded Spill14887; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14888; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1288 ; 4-byte Folded Spill14889; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014890; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v214891; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:13814892; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1320 ; 4-byte Folded Spill14893; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:14914894; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014895; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:13714896; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Spill14897; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:13614898; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14899; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1324 ; 4-byte Folded Spill14900; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14901; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1360 ; 4-byte Folded Spill14902; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14903; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1316 ; 4-byte Folded Spill14904; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14905; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1308 ; 4-byte Folded Spill14906; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014907; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v214908; ALIGNED-NEXT:    s_clause 0x114909; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:14714910; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:14814911; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1356 ; 4-byte Folded Spill14912; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014913; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:14514914; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1328 ; 4-byte Folded Spill14915; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:14414916; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14917; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1344 ; 4-byte Folded Spill14918; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14919; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1352 ; 4-byte Folded Spill14920; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14921; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1340 ; 4-byte Folded Spill14922; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14923; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1332 ; 4-byte Folded Spill14924; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014925; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v714926; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:15914927; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014928; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v514929; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 8, v1214930; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1396 ; 4-byte Folded Spill14931; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 8, v214932; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:15814933; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v1814934; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014935; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:15714936; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Spill14937; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:15614938; ALIGNED-NEXT:    s_waitcnt vmcnt(3)14939; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1424 ; 4-byte Folded Spill14940; ALIGNED-NEXT:    s_waitcnt vmcnt(2)14941; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1416 ; 4-byte Folded Spill14942; ALIGNED-NEXT:    s_waitcnt vmcnt(1)14943; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1412 ; 4-byte Folded Spill14944; ALIGNED-NEXT:    s_waitcnt vmcnt(0)14945; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Spill14946; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v014947; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v214948; ALIGNED-NEXT:    v_lshl_or_b32 v2, v100, 8, v10114949; ALIGNED-NEXT:    v_lshl_or_b32 v3, v86, 8, v9614950; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014951; ALIGNED-NEXT:    v_lshl_or_b32 v1, v124, 8, v11114952; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1436 ; 4-byte Folded Spill14953; ALIGNED-NEXT:    v_lshl_or_b32 v0, v122, 8, v12314954; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014955; ALIGNED-NEXT:    v_lshl_or_b32 v1, v92, 8, v9414956; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1440 ; 4-byte Folded Spill14957; ALIGNED-NEXT:    v_lshl_or_b32 v0, v105, 8, v10814958; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014959; ALIGNED-NEXT:    v_lshl_or_b32 v1, v91, 8, v10414960; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Spill14961; ALIGNED-NEXT:    v_lshl_or_b32 v0, v95, 8, v10614962; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014963; ALIGNED-NEXT:    v_lshl_or_b32 v1, v75, 8, v7614964; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1460 ; 4-byte Folded Spill14965; ALIGNED-NEXT:    v_lshl_or_b32 v0, v77, 8, v7814966; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014967; ALIGNED-NEXT:    v_lshl_or_b32 v1, v74, 8, v6214968; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1464 ; 4-byte Folded Spill14969; ALIGNED-NEXT:    v_lshl_or_b32 v0, v63, 8, v7314970; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014971; ALIGNED-NEXT:    v_lshl_or_b32 v1, v56, 8, v4714972; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1468 ; 4-byte Folded Spill14973; ALIGNED-NEXT:    v_lshl_or_b32 v0, v59, 8, v6114974; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014975; ALIGNED-NEXT:    v_lshl_or_b32 v1, v46, 8, v5814976; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1472 ; 4-byte Folded Spill14977; ALIGNED-NEXT:    v_lshl_or_b32 v0, v57, 8, v6014978; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014979; ALIGNED-NEXT:    v_lshl_or_b32 v1, v41, 8, v4214980; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1476 ; 4-byte Folded Spill14981; ALIGNED-NEXT:    v_lshl_or_b32 v0, v43, 8, v4414982; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014983; ALIGNED-NEXT:    v_lshl_or_b32 v1, v40, 8, v11614984; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1480 ; 4-byte Folded Spill14985; ALIGNED-NEXT:    v_lshl_or_b32 v0, v118, 8, v11914986; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v014987; ALIGNED-NEXT:    s_clause 0x214988; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:24814989; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:24914990; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:25014991; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Spill14992; ALIGNED-NEXT:    v_lshl_or_b32 v0, v112, 8, v11514993; ALIGNED-NEXT:    v_lshl_or_b32 v107, v2, 16, v014994; ALIGNED-NEXT:    v_lshl_or_b32 v0, v103, 8, v11314995; ALIGNED-NEXT:    v_lshl_or_b32 v2, v99, 8, v10214996; ALIGNED-NEXT:    v_lshl_or_b32 v90, v2, 16, v014997; ALIGNED-NEXT:    v_lshl_or_b32 v2, v87, 8, v9714998; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen14999; ALIGNED-NEXT:    v_lshl_or_b32 v79, v3, 16, v215000; ALIGNED-NEXT:    v_lshl_or_b32 v2, v83, 8, v8415001; ALIGNED-NEXT:    v_lshl_or_b32 v3, v85, 8, v8215002; ALIGNED-NEXT:    v_lshl_or_b32 v72, v3, 16, v215003; ALIGNED-NEXT:    v_lshl_or_b32 v2, v67, 8, v8015004; ALIGNED-NEXT:    v_lshl_or_b32 v3, v64, 8, v7015005; ALIGNED-NEXT:    v_lshl_or_b32 v45, v3, 16, v215006; ALIGNED-NEXT:    v_lshl_or_b32 v2, v66, 8, v7115007; ALIGNED-NEXT:    v_lshl_or_b32 v3, v52, 8, v5315008; ALIGNED-NEXT:    v_lshl_or_b32 v117, v3, 16, v215009; ALIGNED-NEXT:    v_lshl_or_b32 v2, v55, 8, v6915010; ALIGNED-NEXT:    v_lshl_or_b32 v3, v51, 8, v5415011; ALIGNED-NEXT:    v_lshl_or_b32 v114, v3, 16, v215012; ALIGNED-NEXT:    v_lshl_or_b32 v2, v38, 8, v4915013; ALIGNED-NEXT:    v_lshl_or_b32 v3, v65, 8, v3915014; ALIGNED-NEXT:    v_lshl_or_b32 v98, v3, 16, v215015; ALIGNED-NEXT:    v_lshl_or_b32 v2, v35, 8, v3715016; ALIGNED-NEXT:    v_lshl_or_b32 v3, v32, 8, v3115017; ALIGNED-NEXT:    v_lshl_or_b32 v81, v3, 16, v215018; ALIGNED-NEXT:    v_lshl_or_b32 v2, v33, 8, v3615019; ALIGNED-NEXT:    v_lshl_or_b32 v3, v30, 8, v3415020; ALIGNED-NEXT:    v_lshl_or_b32 v68, v3, 16, v215021; ALIGNED-NEXT:    v_lshl_or_b32 v2, v27, 8, v2915022; ALIGNED-NEXT:    v_lshl_or_b32 v3, v26, 8, v2815023; ALIGNED-NEXT:    v_lshl_or_b32 v50, v3, 16, v215024; ALIGNED-NEXT:    v_lshl_or_b32 v2, v23, 8, v2415025; ALIGNED-NEXT:    v_lshl_or_b32 v3, v25, 8, v2215026; ALIGNED-NEXT:    v_lshl_or_b32 v48, v3, 16, v215027; ALIGNED-NEXT:    v_lshl_or_b32 v2, v19, 8, v2115028; ALIGNED-NEXT:    v_lshl_or_b32 v3, v16, 8, v1515029; ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v215030; ALIGNED-NEXT:    v_lshl_or_b32 v3, v17, 8, v2015031; ALIGNED-NEXT:    v_lshl_or_b32 v3, v4, 16, v315032; ALIGNED-NEXT:    v_lshl_or_b32 v4, v11, 8, v1315033; ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 16, v415034; ALIGNED-NEXT:    s_waitcnt vmcnt(2)15035; ALIGNED-NEXT:    v_lshl_or_b32 v5, v7, 8, v815036; ALIGNED-NEXT:    s_waitcnt vmcnt(1)15037; ALIGNED-NEXT:    v_lshl_or_b32 v88, v9, 8, v115038; ALIGNED-NEXT:    v_lshl_or_b32 v5, v88, 16, v515039; ALIGNED-NEXT:    buffer_load_ubyte v88, v6, s[0:3], 0 offen offset:115040; ALIGNED-NEXT:    s_waitcnt vmcnt(1)15041; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1336 ; 4-byte Folded Spill15042; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15043; ALIGNED-NEXT:    buffer_store_dword v88, off, s[0:3], s32 offset:1348 ; 4-byte Folded Spill15044; ALIGNED-NEXT:    v_lshl_or_b32 v88, v88, 8, v015045; ALIGNED-NEXT:    v_lshl_or_b32 v0, v89, 16, v8815046; ALIGNED-NEXT:    v_lshl_or_b32 v88, v109, 8, v9315047; ALIGNED-NEXT:    v_lshl_or_b32 v89, v121, 8, v11015048; ALIGNED-NEXT:    s_clause 0x215049; ALIGNED-NEXT:    buffer_load_ubyte v121, v6, s[0:3], 0 offen offset:815050; ALIGNED-NEXT:    buffer_load_ubyte v110, v6, s[0:3], 0 offen offset:915051; ALIGNED-NEXT:    buffer_load_ubyte v109, v6, s[0:3], 0 offen offset:1015052; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Spill15053; ALIGNED-NEXT:    v_lshl_or_b32 v0, v89, 16, v8815054; ALIGNED-NEXT:    s_clause 0x115055; ALIGNED-NEXT:    buffer_load_ubyte v125, v6, s[0:3], 0 offen offset:1315056; ALIGNED-NEXT:    buffer_load_ubyte v89, v6, s[0:3], 0 offen offset:1415057; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Spill15058; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:1215059; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:1452 ; 4-byte Folded Spill15060; ALIGNED-NEXT:    s_waitcnt vmcnt(1)15061; ALIGNED-NEXT:    buffer_store_dword v89, off, s[0:3], s32 offset:1432 ; 4-byte Folded Spill15062; ALIGNED-NEXT:    v_lshl_or_b32 v89, v126, 8, v8915063; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15064; ALIGNED-NEXT:    v_lshl_or_b32 v88, v125, 8, v015065; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1428 ; 4-byte Folded Spill15066; ALIGNED-NEXT:    v_lshl_or_b32 v0, v89, 16, v8815067; ALIGNED-NEXT:    v_lshl_or_b32 v88, v110, 8, v12115068; ALIGNED-NEXT:    v_lshl_or_b32 v89, v120, 8, v10915069; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Spill15070; ALIGNED-NEXT:    v_lshl_or_b32 v0, v89, 16, v8815071; ALIGNED-NEXT:    s_clause 0x215072; ALIGNED-NEXT:    buffer_load_ubyte v88, v6, s[0:3], 0 offen offset:1815073; ALIGNED-NEXT:    buffer_load_ubyte v93, v6, s[0:3], 0 offen offset:1615074; ALIGNED-NEXT:    buffer_load_ubyte v89, v6, s[0:3], 0 offen offset:1715075; ALIGNED-NEXT:    v_add_nc_u32_e32 v6, 0xffffff00, v615076; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1448 ; 4-byte Folded Spill15077; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:48815078; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:49215079; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:48415080; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:48015081; ALIGNED-NEXT:    s_clause 0x1 ; 8-byte Folded Reload15082; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:70415083; ALIGNED-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:70815084; ALIGNED-NEXT:    s_waitcnt vmcnt(4)15085; ALIGNED-NEXT:    v_lshl_or_b32 v0, v127, 8, v8815086; ALIGNED-NEXT:    s_waitcnt vmcnt(2)15087; ALIGNED-NEXT:    v_lshl_or_b32 v127, v89, 8, v9315088; ALIGNED-NEXT:    s_waitcnt vmcnt(1)15089; ALIGNED-NEXT:    v_add_co_u32 v2, vcc_lo, v2, s415090; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15091; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v3, null, s5, v3, vcc_lo15092; ALIGNED-NEXT:    v_lshl_or_b32 v127, v0, 16, v12715093; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, 315094; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v3, vcc_lo15095; ALIGNED-NEXT:    flat_store_byte v[4:5], v1 offset:24715096; ALIGNED-NEXT:    flat_store_byte v[4:5], v9 offset:24815097; ALIGNED-NEXT:    flat_store_byte v[4:5], v7 offset:24615098; ALIGNED-NEXT:    flat_store_byte v[4:5], v10 offset:25215099; ALIGNED-NEXT:    flat_store_byte v[4:5], v11 offset:25015100; ALIGNED-NEXT:    flat_store_byte v[4:5], v12 offset:25115101; ALIGNED-NEXT:    flat_store_byte v[4:5], v13 offset:24915102; ALIGNED-NEXT:    flat_store_byte v[4:5], v8 offset:24515103; ALIGNED-NEXT:    flat_store_byte v[4:5], v15 offset:23915104; ALIGNED-NEXT:    flat_store_byte v[4:5], v16 offset:24015105; ALIGNED-NEXT:    flat_store_byte v[4:5], v19 offset:23815106; ALIGNED-NEXT:    flat_store_byte v[4:5], v14 offset:24415107; ALIGNED-NEXT:    flat_store_byte v[4:5], v17 offset:24215108; ALIGNED-NEXT:    flat_store_byte v[4:5], v18 offset:24315109; ALIGNED-NEXT:    flat_store_byte v[4:5], v20 offset:24115110; ALIGNED-NEXT:    flat_store_byte v[4:5], v21 offset:23715111; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:50415112; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:50815113; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:50015114; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:49615115; ALIGNED-NEXT:    flat_store_byte v[4:5], v22 offset:23115116; ALIGNED-NEXT:    flat_store_byte v[4:5], v25 offset:23215117; ALIGNED-NEXT:    flat_store_byte v[4:5], v23 offset:23015118; ALIGNED-NEXT:    flat_store_byte v[4:5], v26 offset:23615119; ALIGNED-NEXT:    flat_store_byte v[4:5], v27 offset:23415120; ALIGNED-NEXT:    flat_store_byte v[4:5], v28 offset:23515121; ALIGNED-NEXT:    flat_store_byte v[4:5], v29 offset:23315122; ALIGNED-NEXT:    flat_store_byte v[4:5], v24 offset:22915123; ALIGNED-NEXT:    flat_store_byte v[4:5], v31 offset:22315124; ALIGNED-NEXT:    flat_store_byte v[4:5], v32 offset:22415125; ALIGNED-NEXT:    flat_store_byte v[4:5], v35 offset:22215126; ALIGNED-NEXT:    flat_store_byte v[4:5], v30 offset:22815127; ALIGNED-NEXT:    flat_store_byte v[4:5], v33 offset:22615128; ALIGNED-NEXT:    flat_store_byte v[4:5], v34 offset:22715129; ALIGNED-NEXT:    flat_store_byte v[4:5], v36 offset:22515130; ALIGNED-NEXT:    flat_store_byte v[4:5], v37 offset:22115131; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:44815132; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:46015133; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:45615134; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:45215135; ALIGNED-NEXT:    flat_store_byte v[4:5], v67 offset:21015136; ALIGNED-NEXT:    flat_store_byte v[4:5], v64 offset:21215137; ALIGNED-NEXT:    flat_store_byte v[4:5], v38 offset:20615138; ALIGNED-NEXT:    flat_store_byte v[4:5], v65 offset:20815139; ALIGNED-NEXT:    flat_store_byte v[4:5], v39 offset:20715140; ALIGNED-NEXT:    flat_store_byte v[4:5], v70 offset:21115141; ALIGNED-NEXT:    flat_store_byte v[4:5], v80 offset:20915142; ALIGNED-NEXT:    flat_store_byte v[4:5], v53 offset:21515143; ALIGNED-NEXT:    flat_store_byte v[4:5], v52 offset:21615144; ALIGNED-NEXT:    flat_store_byte v[4:5], v66 offset:21415145; ALIGNED-NEXT:    flat_store_byte v[4:5], v51 offset:22015146; ALIGNED-NEXT:    flat_store_byte v[4:5], v55 offset:21815147; ALIGNED-NEXT:    flat_store_byte v[4:5], v54 offset:21915148; ALIGNED-NEXT:    flat_store_byte v[4:5], v69 offset:21715149; ALIGNED-NEXT:    flat_store_byte v[4:5], v71 offset:21315150; ALIGNED-NEXT:    flat_store_byte v[4:5], v49 offset:20515151; ALIGNED-NEXT:    buffer_store_dword v72, off, s[0:3], s32 offset:47215152; ALIGNED-NEXT:    buffer_store_dword v79, off, s[0:3], s32 offset:47615153; ALIGNED-NEXT:    buffer_store_dword v90, off, s[0:3], s32 offset:46815154; ALIGNED-NEXT:    buffer_store_dword v107, off, s[0:3], s32 offset:46415155; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Reload15156; ALIGNED-NEXT:    flat_store_byte v[4:5], v82 offset:19915157; ALIGNED-NEXT:    flat_store_byte v[4:5], v85 offset:20015158; ALIGNED-NEXT:    flat_store_byte v[4:5], v83 offset:19815159; ALIGNED-NEXT:    flat_store_byte v[4:5], v86 offset:20415160; ALIGNED-NEXT:    flat_store_byte v[4:5], v87 offset:20215161; ALIGNED-NEXT:    flat_store_byte v[4:5], v96 offset:20315162; ALIGNED-NEXT:    flat_store_byte v[4:5], v97 offset:20115163; ALIGNED-NEXT:    flat_store_byte v[4:5], v84 offset:19715164; ALIGNED-NEXT:    flat_store_byte v[4:5], v101 offset:19115165; ALIGNED-NEXT:    flat_store_byte v[4:5], v100 offset:19215166; ALIGNED-NEXT:    flat_store_byte v[4:5], v112 offset:19015167; ALIGNED-NEXT:    flat_store_byte v[4:5], v99 offset:19615168; ALIGNED-NEXT:    flat_store_byte v[4:5], v103 offset:19415169; ALIGNED-NEXT:    flat_store_byte v[4:5], v102 offset:19515170; ALIGNED-NEXT:    flat_store_byte v[4:5], v113 offset:19315171; ALIGNED-NEXT:    flat_store_byte v[4:5], v115 offset:18915172; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff0015173; ALIGNED-NEXT:    s_addc_u32 s5, s5, -115174; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]15175; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15176; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:55215177; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1480 ; 4-byte Folded Reload15178; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15179; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:55615180; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1476 ; 4-byte Folded Reload15181; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15182; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:54815183; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1472 ; 4-byte Folded Reload15184; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15185; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:54415186; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1468 ; 4-byte Folded Reload15187; ALIGNED-NEXT:    flat_store_byte v[4:5], v116 offset:18315188; ALIGNED-NEXT:    flat_store_byte v[4:5], v40 offset:18415189; ALIGNED-NEXT:    flat_store_byte v[4:5], v118 offset:18215190; ALIGNED-NEXT:    flat_store_byte v[4:5], v41 offset:18815191; ALIGNED-NEXT:    flat_store_byte v[4:5], v43 offset:18615192; ALIGNED-NEXT:    flat_store_byte v[4:5], v42 offset:18715193; ALIGNED-NEXT:    flat_store_byte v[4:5], v44 offset:18515194; ALIGNED-NEXT:    flat_store_byte v[4:5], v119 offset:18115195; ALIGNED-NEXT:    flat_store_byte v[4:5], v47 offset:17515196; ALIGNED-NEXT:    flat_store_byte v[4:5], v56 offset:17615197; ALIGNED-NEXT:    flat_store_byte v[4:5], v59 offset:17415198; ALIGNED-NEXT:    flat_store_byte v[4:5], v46 offset:18015199; ALIGNED-NEXT:    flat_store_byte v[4:5], v57 offset:17815200; ALIGNED-NEXT:    flat_store_byte v[4:5], v58 offset:17915201; ALIGNED-NEXT:    flat_store_byte v[4:5], v60 offset:17715202; ALIGNED-NEXT:    flat_store_byte v[4:5], v61 offset:17315203; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15204; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:56815205; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1464 ; 4-byte Folded Reload15206; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15207; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:57215208; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1460 ; 4-byte Folded Reload15209; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15210; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:56415211; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Reload15212; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15213; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:56015214; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1440 ; 4-byte Folded Reload15215; ALIGNED-NEXT:    flat_store_byte v[4:5], v62 offset:16715216; ALIGNED-NEXT:    flat_store_byte v[4:5], v74 offset:16815217; ALIGNED-NEXT:    flat_store_byte v[4:5], v63 offset:16615218; ALIGNED-NEXT:    flat_store_byte v[4:5], v75 offset:17215219; ALIGNED-NEXT:    flat_store_byte v[4:5], v77 offset:17015220; ALIGNED-NEXT:    flat_store_byte v[4:5], v76 offset:17115221; ALIGNED-NEXT:    flat_store_byte v[4:5], v78 offset:16915222; ALIGNED-NEXT:    flat_store_byte v[4:5], v73 offset:16515223; ALIGNED-NEXT:    flat_store_byte v[4:5], v94 offset:15915224; ALIGNED-NEXT:    flat_store_byte v[4:5], v92 offset:16015225; ALIGNED-NEXT:    flat_store_byte v[4:5], v105 offset:15815226; ALIGNED-NEXT:    flat_store_byte v[4:5], v91 offset:16415227; ALIGNED-NEXT:    flat_store_byte v[4:5], v95 offset:16215228; ALIGNED-NEXT:    flat_store_byte v[4:5], v104 offset:16315229; ALIGNED-NEXT:    flat_store_byte v[4:5], v106 offset:16115230; ALIGNED-NEXT:    flat_store_byte v[4:5], v108 offset:15715231; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15232; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:52015233; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1436 ; 4-byte Folded Reload15234; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15235; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:52415236; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Reload15237; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15238; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:51615239; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1396 ; 4-byte Folded Reload15240; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15241; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:51215242; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1424 ; 4-byte Folded Reload15243; ALIGNED-NEXT:    flat_store_byte v[4:5], v111 offset:15115244; ALIGNED-NEXT:    flat_store_byte v[4:5], v124 offset:15215245; ALIGNED-NEXT:    flat_store_byte v[4:5], v122 offset:15015246; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15247; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:15615248; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1412 ; 4-byte Folded Reload15249; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15250; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:15415251; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1416 ; 4-byte Folded Reload15252; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15253; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:15515254; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Reload15255; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15256; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:15315257; ALIGNED-NEXT:    flat_store_byte v[4:5], v123 offset:14915258; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1356 ; 4-byte Folded Reload15259; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15260; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:14315261; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1344 ; 4-byte Folded Reload15262; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15263; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:14415264; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1340 ; 4-byte Folded Reload15265; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15266; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:14215267; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1380 ; 4-byte Folded Reload15268; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15269; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:14815270; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1360 ; 4-byte Folded Reload15271; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15272; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:14615273; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1368 ; 4-byte Folded Reload15274; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15275; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:14715276; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1352 ; 4-byte Folded Reload15277; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15278; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:14515279; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1332 ; 4-byte Folded Reload15280; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15281; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:14115282; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1328 ; 4-byte Folded Reload15283; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15284; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:53615285; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Reload15286; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15287; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:54015288; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1292 ; 4-byte Folded Reload15289; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15290; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:53215291; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1284 ; 4-byte Folded Reload15292; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15293; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:52815294; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1324 ; 4-byte Folded Reload15295; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15296; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13515297; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1320 ; 4-byte Folded Reload15298; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15299; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13615300; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1316 ; 4-byte Folded Reload15301; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15302; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13415303; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1304 ; 4-byte Folded Reload15304; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15305; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:14015306; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1296 ; 4-byte Folded Reload15307; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15308; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13815309; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1300 ; 4-byte Folded Reload15310; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15311; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13915312; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1288 ; 4-byte Folded Reload15313; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15314; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13715315; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1308 ; 4-byte Folded Reload15316; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15317; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13315318; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1268 ; 4-byte Folded Reload15319; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15320; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:12715321; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1260 ; 4-byte Folded Reload15322; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15323; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:12815324; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1256 ; 4-byte Folded Reload15325; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15326; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:12615327; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1280 ; 4-byte Folded Reload15328; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15329; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13215330; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1272 ; 4-byte Folded Reload15331; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15332; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13015333; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1276 ; 4-byte Folded Reload15334; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15335; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:13115336; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1264 ; 4-byte Folded Reload15337; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15338; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:12915339; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1252 ; 4-byte Folded Reload15340; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15341; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:12815342; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Reload15343; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15344; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:61615345; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Reload15346; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15347; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:62015348; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1212 ; 4-byte Folded Reload15349; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15350; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:61215351; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Reload15352; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15353; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60815354; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1244 ; 4-byte Folded Reload15355; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15356; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11915357; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1240 ; 4-byte Folded Reload15358; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15359; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:12015360; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1236 ; 4-byte Folded Reload15361; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15362; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11815363; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1224 ; 4-byte Folded Reload15364; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15365; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:12415366; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1216 ; 4-byte Folded Reload15367; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15368; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:12215369; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1220 ; 4-byte Folded Reload15370; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15371; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:12315372; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Reload15373; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15374; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:12115375; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Reload15376; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15377; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11715378; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Reload15379; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15380; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11115381; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Reload15382; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15383; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11215384; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1176 ; 4-byte Folded Reload15385; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15386; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11015387; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Reload15388; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15389; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11615390; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1192 ; 4-byte Folded Reload15391; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15392; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11415393; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1196 ; 4-byte Folded Reload15394; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15395; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11515396; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1184 ; 4-byte Folded Reload15397; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15398; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:11315399; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Reload15400; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15401; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10915402; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Reload15403; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15404; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:63215405; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Reload15406; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15407; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:63615408; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1132 ; 4-byte Folded Reload15409; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15410; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:62815411; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Reload15412; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15413; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:62415414; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1164 ; 4-byte Folded Reload15415; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15416; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10315417; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Reload15418; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15419; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10415420; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1156 ; 4-byte Folded Reload15421; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15422; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10215423; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Reload15424; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15425; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10815426; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1136 ; 4-byte Folded Reload15427; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15428; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10615429; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1140 ; 4-byte Folded Reload15430; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15431; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10715432; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Reload15433; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15434; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10515435; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Reload15436; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15437; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10115438; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1108 ; 4-byte Folded Reload15439; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15440; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9515441; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1100 ; 4-byte Folded Reload15442; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15443; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9615444; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1096 ; 4-byte Folded Reload15445; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15446; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9415447; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Reload15448; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15449; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:10015450; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1112 ; 4-byte Folded Reload15451; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15452; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9815453; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1116 ; 4-byte Folded Reload15454; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15455; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9915456; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1104 ; 4-byte Folded Reload15457; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15458; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9715459; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Reload15460; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15461; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9315462; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Reload15463; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15464; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:58415465; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1072 ; 4-byte Folded Reload15466; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15467; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:58815468; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Reload15469; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15470; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:58015471; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Reload15472; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15473; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:57615474; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Reload15475; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15476; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8715477; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Reload15478; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15479; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8815480; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1076 ; 4-byte Folded Reload15481; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15482; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8615483; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Reload15484; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15485; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9215486; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1056 ; 4-byte Folded Reload15487; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15488; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9015489; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1060 ; 4-byte Folded Reload15490; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15491; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:9115492; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Reload15493; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15494; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8915495; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1068 ; 4-byte Folded Reload15496; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15497; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8515498; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1028 ; 4-byte Folded Reload15499; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15500; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7915501; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Reload15502; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15503; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8015504; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1016 ; 4-byte Folded Reload15505; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15506; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7815507; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1040 ; 4-byte Folded Reload15508; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15509; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8415510; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1032 ; 4-byte Folded Reload15511; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15512; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8215513; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1036 ; 4-byte Folded Reload15514; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15515; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8315516; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1024 ; 4-byte Folded Reload15517; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15518; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:8115519; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1012 ; 4-byte Folded Reload15520; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15521; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7715522; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Reload15523; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15524; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60015525; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:992 ; 4-byte Folded Reload15526; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15527; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60415528; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:988 ; 4-byte Folded Reload15529; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15530; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:59615531; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Reload15532; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15533; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:59215534; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1004 ; 4-byte Folded Reload15535; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15536; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7115537; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1000 ; 4-byte Folded Reload15538; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15539; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7215540; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:996 ; 4-byte Folded Reload15541; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15542; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7015543; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:980 ; 4-byte Folded Reload15544; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15545; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7615546; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload15547; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15548; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7415549; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload15550; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15551; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7515552; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload15553; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15554; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:7315555; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:984 ; 4-byte Folded Reload15556; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15557; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:6915558; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload15559; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15560; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:6315561; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload15562; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15563; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:6415564; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload15565; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15566; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:6215567; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload15568; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15569; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:6815570; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload15571; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15572; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:6615573; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload15574; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15575; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:6715576; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload15577; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15578; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:6515579; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload15580; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15581; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:6415582; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload15583; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15584; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68015585; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload15586; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15587; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68415588; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload15589; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15590; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:67615591; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload15592; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15593; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:67215594; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload15595; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15596; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5815597; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload15598; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15599; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5515600; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload15601; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15602; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5615603; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload15604; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15605; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5415606; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload15607; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15608; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:6015609; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload15610; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15611; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5915612; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload15613; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15614; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5715615; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload15616; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15617; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5315618; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload15619; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15620; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5015621; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload15622; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15623; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4715624; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload15625; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15626; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4815627; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload15628; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15629; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4615630; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload15631; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15632; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5215633; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload15634; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15635; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:5115636; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload15637; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15638; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4915639; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload15640; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15641; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4515642; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload15643; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15644; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:70015645; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload15646; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15647; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:69615648; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload15649; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15650; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:69215651; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload15652; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15653; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68815654; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload15655; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15656; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4015657; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload15658; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15659; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3915660; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload15661; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15662; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3815663; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload15664; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15665; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3715666; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload15667; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15668; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4415669; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload15670; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15671; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4315672; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload15673; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15674; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4215675; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload15676; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15677; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:4115678; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload15679; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15680; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3215681; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload15682; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15683; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3115684; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload15685; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15686; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3015687; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload15688; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15689; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:3215690; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload15691; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15692; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3615693; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload15694; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15695; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3515696; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload15697; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15698; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3415699; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload15700; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15701; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:3315702; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload15703; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15704; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:64815705; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload15706; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15707; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:65215708; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload15709; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15710; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:64415711; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:64015712; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload15713; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15714; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:2315715; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload15716; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15717; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:2415718; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload15719; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15720; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:2215721; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload15722; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15723; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:2815724; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload15725; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15726; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:2615727; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload15728; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15729; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:2715730; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload15731; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15732; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:2515733; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload15734; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15735; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:2115736; ALIGNED-NEXT:    flat_store_byte v[4:5], v88 offset:1515737; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1452 ; 4-byte Folded Reload15738; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15739; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:1615740; ALIGNED-NEXT:    flat_store_byte v[4:5], v89 offset:1415741; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload15742; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15743; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:2015744; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload15745; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15746; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:1815747; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload15748; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15749; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:1915750; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload15751; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15752; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:1715753; ALIGNED-NEXT:    flat_store_byte v[2:3], v93 offset:1615754; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1448 ; 4-byte Folded Reload15755; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15756; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:66415757; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Reload15758; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15759; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:66815760; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Reload15761; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15762; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:66015763; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Reload15764; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15765; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:65615766; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1432 ; 4-byte Folded Reload15767; ALIGNED-NEXT:    flat_store_byte v[4:5], v109 offset:715768; ALIGNED-NEXT:    flat_store_byte v[4:5], v120 offset:815769; ALIGNED-NEXT:    flat_store_byte v[4:5], v125 offset:1015770; ALIGNED-NEXT:    flat_store_byte v[4:5], v110 offset:615771; ALIGNED-NEXT:    flat_store_byte v[4:5], v126 offset:1215772; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15773; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:1115774; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1428 ; 4-byte Folded Reload15775; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15776; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:915777; ALIGNED-NEXT:    flat_store_byte v[2:3], v121 offset:815778; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1376 ; 4-byte Folded Reload15779; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15780; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:215781; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1364 ; 4-byte Folded Reload15782; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15783; ALIGNED-NEXT:    flat_store_byte v[4:5], v015784; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Reload15785; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15786; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:115787; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Reload15788; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15789; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:415790; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1384 ; 4-byte Folded Reload15791; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15792; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:215793; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1388 ; 4-byte Folded Reload15794; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15795; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:315796; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1372 ; 4-byte Folded Reload15797; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15798; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:415799; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1336 ; 4-byte Folded Reload15800; ALIGNED-NEXT:    s_waitcnt vmcnt(0)15801; ALIGNED-NEXT:    flat_store_byte v[2:3], v015802; ALIGNED-NEXT:    s_cbranch_scc0 .LBB9_415803; ALIGNED-NEXT:  .LBB9_5: ; %Flow1115804; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s815805; ALIGNED-NEXT:    s_clause 0x2f ; 192-byte Folded Reload15806; ALIGNED-NEXT:    buffer_load_dword v127, off, s[0:3], s3215807; ALIGNED-NEXT:    buffer_load_dword v126, off, s[0:3], s32 offset:415808; ALIGNED-NEXT:    buffer_load_dword v125, off, s[0:3], s32 offset:815809; ALIGNED-NEXT:    buffer_load_dword v124, off, s[0:3], s32 offset:1215810; ALIGNED-NEXT:    buffer_load_dword v123, off, s[0:3], s32 offset:1615811; ALIGNED-NEXT:    buffer_load_dword v122, off, s[0:3], s32 offset:2015812; ALIGNED-NEXT:    buffer_load_dword v121, off, s[0:3], s32 offset:2415813; ALIGNED-NEXT:    buffer_load_dword v120, off, s[0:3], s32 offset:2815814; ALIGNED-NEXT:    buffer_load_dword v111, off, s[0:3], s32 offset:3215815; ALIGNED-NEXT:    buffer_load_dword v110, off, s[0:3], s32 offset:3615816; ALIGNED-NEXT:    buffer_load_dword v109, off, s[0:3], s32 offset:4015817; ALIGNED-NEXT:    buffer_load_dword v108, off, s[0:3], s32 offset:4415818; ALIGNED-NEXT:    buffer_load_dword v107, off, s[0:3], s32 offset:4815819; ALIGNED-NEXT:    buffer_load_dword v106, off, s[0:3], s32 offset:5215820; ALIGNED-NEXT:    buffer_load_dword v105, off, s[0:3], s32 offset:5615821; ALIGNED-NEXT:    buffer_load_dword v104, off, s[0:3], s32 offset:6015822; ALIGNED-NEXT:    buffer_load_dword v95, off, s[0:3], s32 offset:6415823; ALIGNED-NEXT:    buffer_load_dword v94, off, s[0:3], s32 offset:6815824; ALIGNED-NEXT:    buffer_load_dword v93, off, s[0:3], s32 offset:7215825; ALIGNED-NEXT:    buffer_load_dword v92, off, s[0:3], s32 offset:7615826; ALIGNED-NEXT:    buffer_load_dword v91, off, s[0:3], s32 offset:8015827; ALIGNED-NEXT:    buffer_load_dword v90, off, s[0:3], s32 offset:8415828; ALIGNED-NEXT:    buffer_load_dword v89, off, s[0:3], s32 offset:8815829; ALIGNED-NEXT:    buffer_load_dword v88, off, s[0:3], s32 offset:9215830; ALIGNED-NEXT:    buffer_load_dword v79, off, s[0:3], s32 offset:9615831; ALIGNED-NEXT:    buffer_load_dword v78, off, s[0:3], s32 offset:10015832; ALIGNED-NEXT:    buffer_load_dword v77, off, s[0:3], s32 offset:10415833; ALIGNED-NEXT:    buffer_load_dword v76, off, s[0:3], s32 offset:10815834; ALIGNED-NEXT:    buffer_load_dword v75, off, s[0:3], s32 offset:11215835; ALIGNED-NEXT:    buffer_load_dword v74, off, s[0:3], s32 offset:11615836; ALIGNED-NEXT:    buffer_load_dword v73, off, s[0:3], s32 offset:12015837; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s32 offset:12415838; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12815839; ALIGNED-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:13215840; ALIGNED-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:13615841; ALIGNED-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:14015842; ALIGNED-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:14415843; ALIGNED-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:14815844; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:15215845; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:15615846; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:16015847; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:16415848; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:16815849; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:17215850; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:17615851; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:18015852; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:18415853; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:18815854; ALIGNED-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)15855; ALIGNED-NEXT:    s_setpc_b64 s[30:31]15856;15857; UNROLL3-LABEL: memmove_p0_p5_sz2048:15858; UNROLL3:       ; %bb.0: ; %entry15859; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)15860; UNROLL3-NEXT:    v_cmp_ne_u64_e32 vcc_lo, 0, v[0:1]15861; UNROLL3-NEXT:    s_mov_b64 s[4:5], 015862; UNROLL3-NEXT:    s_mov_b32 s6, exec_lo15863; UNROLL3-NEXT:    v_cndmask_b32_e32 v3, -1, v0, vcc_lo15864; UNROLL3-NEXT:    v_cmpx_ge_u32_e64 v2, v315865; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s615866; UNROLL3-NEXT:    s_cbranch_execz .LBB9_415867; UNROLL3-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader15868; UNROLL3-NEXT:    v_mov_b32_e32 v3, v215869; UNROLL3-NEXT:    s_inst_prefetch 0x115870; UNROLL3-NEXT:    .p2align 615871; UNROLL3-NEXT:  .LBB9_2: ; %memmove_fwd_loop15872; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=115873; UNROLL3-NEXT:    s_clause 0xb15874; UNROLL3-NEXT:    buffer_load_dword v4, v3, s[0:3], 0 offen15875; UNROLL3-NEXT:    buffer_load_dword v5, v3, s[0:3], 0 offen offset:415876; UNROLL3-NEXT:    buffer_load_dword v6, v3, s[0:3], 0 offen offset:815877; UNROLL3-NEXT:    buffer_load_dword v7, v3, s[0:3], 0 offen offset:1215878; UNROLL3-NEXT:    buffer_load_dword v8, v3, s[0:3], 0 offen offset:1615879; UNROLL3-NEXT:    buffer_load_dword v9, v3, s[0:3], 0 offen offset:2015880; UNROLL3-NEXT:    buffer_load_dword v10, v3, s[0:3], 0 offen offset:2415881; UNROLL3-NEXT:    buffer_load_dword v11, v3, s[0:3], 0 offen offset:2815882; UNROLL3-NEXT:    buffer_load_dword v12, v3, s[0:3], 0 offen offset:3215883; UNROLL3-NEXT:    buffer_load_dword v13, v3, s[0:3], 0 offen offset:3615884; UNROLL3-NEXT:    buffer_load_dword v14, v3, s[0:3], 0 offen offset:4015885; UNROLL3-NEXT:    buffer_load_dword v15, v3, s[0:3], 0 offen offset:4415886; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s415887; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo15888; UNROLL3-NEXT:    s_add_u32 s4, s4, 4815889; UNROLL3-NEXT:    v_add_nc_u32_e32 v3, 48, v315890; UNROLL3-NEXT:    s_addc_u32 s5, s5, 015891; UNROLL3-NEXT:    s_waitcnt vmcnt(4)15892; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11] offset:1615893; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7]15894; UNROLL3-NEXT:    s_waitcnt vmcnt(0)15895; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:3215896; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0x7e015897; UNROLL3-NEXT:    s_cbranch_scc1 .LBB9_215898; UNROLL3-NEXT:  ; %bb.3: ; %memmove_fwd_residual15899; UNROLL3-NEXT:    s_inst_prefetch 0x215900; UNROLL3-NEXT:    s_clause 0x315901; UNROLL3-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:201615902; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:202015903; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:202415904; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:202815905; UNROLL3-NEXT:    s_waitcnt vmcnt(0)15906; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:201615907; UNROLL3-NEXT:    s_clause 0x315908; UNROLL3-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:203215909; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:203615910; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:204015911; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:204415912; UNROLL3-NEXT:    ; implicit-def: $vgpr215913; UNROLL3-NEXT:    s_waitcnt vmcnt(0)15914; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:203215915; UNROLL3-NEXT:    ; implicit-def: $vgpr0_vgpr115916; UNROLL3-NEXT:  .LBB9_4: ; %Flow815917; UNROLL3-NEXT:    s_andn2_saveexec_b32 s8, s615918; UNROLL3-NEXT:    s_cbranch_execz .LBB9_715919; UNROLL3-NEXT:  ; %bb.5: ; %memmove_bwd_residual15920; UNROLL3-NEXT:    s_clause 0x315921; UNROLL3-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:203215922; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:203615923; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:204015924; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:204415925; UNROLL3-NEXT:    s_movk_i32 s6, 0xffd015926; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7b015927; UNROLL3-NEXT:    s_mov_b32 s7, -115928; UNROLL3-NEXT:    s_waitcnt vmcnt(0)15929; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:203215930; UNROLL3-NEXT:    s_clause 0x315931; UNROLL3-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:201615932; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:202015933; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:202415934; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:202815935; UNROLL3-NEXT:    v_add_nc_u32_e32 v2, 0x7b0, v215936; UNROLL3-NEXT:    s_waitcnt vmcnt(0)15937; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:201615938; UNROLL3-NEXT:    s_inst_prefetch 0x115939; UNROLL3-NEXT:    .p2align 615940; UNROLL3-NEXT:  .LBB9_6: ; %memmove_bwd_loop15941; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=115942; UNROLL3-NEXT:    s_clause 0xb15943; UNROLL3-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen15944; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:415945; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:815946; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:1215947; UNROLL3-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:1615948; UNROLL3-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:2015949; UNROLL3-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:2415950; UNROLL3-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:2815951; UNROLL3-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:3215952; UNROLL3-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:3615953; UNROLL3-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:4015954; UNROLL3-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:4415955; UNROLL3-NEXT:    v_add_co_u32 v15, vcc_lo, v0, s415956; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v16, null, s5, v1, vcc_lo15957; UNROLL3-NEXT:    v_subrev_nc_u32_e32 v2, 48, v215958; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd015959; UNROLL3-NEXT:    s_addc_u32 s5, s5, -115960; UNROLL3-NEXT:    s_waitcnt vmcnt(4)15961; UNROLL3-NEXT:    flat_store_dwordx4 v[15:16], v[7:10] offset:1615962; UNROLL3-NEXT:    flat_store_dwordx4 v[15:16], v[3:6]15963; UNROLL3-NEXT:    s_waitcnt vmcnt(0)15964; UNROLL3-NEXT:    flat_store_dwordx4 v[15:16], v[11:14] offset:3215965; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]15966; UNROLL3-NEXT:    s_cbranch_scc0 .LBB9_615967; UNROLL3-NEXT:  .LBB9_7: ; %Flow915968; UNROLL3-NEXT:    s_inst_prefetch 0x215969; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s815970; UNROLL3-NEXT:    s_waitcnt lgkmcnt(0)15971; UNROLL3-NEXT:    s_setpc_b64 s[30:31]15972entry:15973  tail call void @llvm.memmove.p0.p5.i64(ptr addrspace(0) noundef nonnull align 1 %dst, ptr addrspace(5) noundef nonnull align 1 %src, i64 2048, i1 false)15974  ret void15975}15976 15977 15978declare void @llvm.memcpy.p0.p0.i64(ptr addrspace(0) noalias nocapture writeonly, ptr addrspace(0) noalias nocapture readonly, i64, i1 immarg) #215979declare void @llvm.memcpy.p1.p1.i64(ptr addrspace(1) noalias nocapture writeonly, ptr addrspace(1) noalias nocapture readonly, i64, i1 immarg) #215980declare void @llvm.memcpy.p0.p4.i64(ptr addrspace(0) noalias nocapture writeonly, ptr addrspace(4) noalias nocapture readonly, i64, i1 immarg) #215981declare void @llvm.memcpy.p5.p5.i64(ptr addrspace(5) noalias nocapture writeonly, ptr addrspace(5) noalias nocapture readonly, i64, i1 immarg) #215982 15983declare void @llvm.memcpy.p0.p5.i64(ptr addrspace(0) noalias nocapture writeonly, ptr addrspace(5) noalias nocapture readonly, i64, i1 immarg) #215984 15985declare void @llvm.memmove.p0.p0.i64(ptr addrspace(0) nocapture writeonly, ptr addrspace(0) nocapture readonly, i64, i1 immarg) #215986declare void @llvm.memmove.p1.p1.i64(ptr addrspace(1) nocapture writeonly, ptr addrspace(1) nocapture readonly, i64, i1 immarg) #215987declare void @llvm.memmove.p0.p4.i64(ptr addrspace(0) nocapture writeonly, ptr addrspace(4) nocapture readonly, i64, i1 immarg) #215988declare void @llvm.memmove.p5.p5.i64(ptr addrspace(5) nocapture writeonly, ptr addrspace(5) nocapture readonly, i64, i1 immarg) #215989 15990declare void @llvm.memmove.p0.p5.i64(ptr addrspace(0) nocapture writeonly, ptr addrspace(5) nocapture readonly, i64, i1 immarg) #215991 15992attributes #2 = { nocallback nofree nounwind willreturn memory(argmem: readwrite) }15993