360 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -global-isel -mtriple=amdgcn-- -amdgpu-memcpy-loop-unroll=2 -mem-intrinsic-expand-size=35 %s -o - | FileCheck -check-prefix=LOOP %s3; RUN: llc -global-isel -mtriple=amdgcn-- -amdgpu-memcpy-loop-unroll=2 -mem-intrinsic-expand-size=37 %s -o - | FileCheck -check-prefix=UNROLL %s4 5declare void @llvm.memcpy.p1.p1.i32(ptr addrspace(1), ptr addrspace(1), i32, i1 immarg)6 7define amdgpu_cs void @memcpy_p1i8(ptr addrspace(1) %dst, ptr addrspace(1) %src) {8; LOOP-LABEL: memcpy_p1i8:9; LOOP: ; %bb.0:10; LOOP-NEXT: s_mov_b64 s[0:1], 011; LOOP-NEXT: s_mov_b32 s2, 012; LOOP-NEXT: s_mov_b32 s3, 0xf00013; LOOP-NEXT: v_mov_b32_e32 v5, s114; LOOP-NEXT: v_mov_b32_e32 v4, s015; LOOP-NEXT: .LBB0_1: ; %load-store-loop16; LOOP-NEXT: ; =>This Inner Loop Header: Depth=117; LOOP-NEXT: v_add_i32_e32 v6, vcc, v2, v418; LOOP-NEXT: v_addc_u32_e32 v7, vcc, v3, v5, vcc19; LOOP-NEXT: buffer_load_ubyte v26, v[6:7], s[0:3], 0 addr6420; LOOP-NEXT: s_waitcnt expcnt(5)21; LOOP-NEXT: buffer_load_ubyte v29, v[6:7], s[0:3], 0 addr64 offset:122; LOOP-NEXT: s_waitcnt expcnt(2)23; LOOP-NEXT: buffer_load_ubyte v31, v[6:7], s[0:3], 0 addr64 offset:224; LOOP-NEXT: buffer_load_ubyte v32, v[6:7], s[0:3], 0 addr64 offset:325; LOOP-NEXT: buffer_load_ubyte v36, v[6:7], s[0:3], 0 addr64 offset:426; LOOP-NEXT: buffer_load_ubyte v37, v[6:7], s[0:3], 0 addr64 offset:527; LOOP-NEXT: buffer_load_ubyte v38, v[6:7], s[0:3], 0 addr64 offset:628; LOOP-NEXT: buffer_load_ubyte v39, v[6:7], s[0:3], 0 addr64 offset:729; LOOP-NEXT: buffer_load_ubyte v8, v[6:7], s[0:3], 0 addr64 offset:830; LOOP-NEXT: buffer_load_ubyte v11, v[6:7], s[0:3], 0 addr64 offset:931; LOOP-NEXT: buffer_load_ubyte v12, v[6:7], s[0:3], 0 addr64 offset:1032; LOOP-NEXT: s_waitcnt expcnt(0)33; LOOP-NEXT: buffer_load_ubyte v13, v[6:7], s[0:3], 0 addr64 offset:1134; LOOP-NEXT: buffer_load_ubyte v9, v[6:7], s[0:3], 0 addr64 offset:1235; LOOP-NEXT: buffer_load_ubyte v15, v[6:7], s[0:3], 0 addr64 offset:1336; LOOP-NEXT: buffer_load_ubyte v16, v[6:7], s[0:3], 0 addr64 offset:1437; LOOP-NEXT: buffer_load_ubyte v17, v[6:7], s[0:3], 0 addr64 offset:1538; LOOP-NEXT: buffer_load_ubyte v10, v[6:7], s[0:3], 0 addr64 offset:1639; LOOP-NEXT: buffer_load_ubyte v19, v[6:7], s[0:3], 0 addr64 offset:1740; LOOP-NEXT: buffer_load_ubyte v20, v[6:7], s[0:3], 0 addr64 offset:1841; LOOP-NEXT: buffer_load_ubyte v21, v[6:7], s[0:3], 0 addr64 offset:1942; LOOP-NEXT: buffer_load_ubyte v14, v[6:7], s[0:3], 0 addr64 offset:2043; LOOP-NEXT: buffer_load_ubyte v23, v[6:7], s[0:3], 0 addr64 offset:2144; LOOP-NEXT: buffer_load_ubyte v24, v[6:7], s[0:3], 0 addr64 offset:2245; LOOP-NEXT: buffer_load_ubyte v25, v[6:7], s[0:3], 0 addr64 offset:2346; LOOP-NEXT: buffer_load_ubyte v18, v[6:7], s[0:3], 0 addr64 offset:2447; LOOP-NEXT: buffer_load_ubyte v27, v[6:7], s[0:3], 0 addr64 offset:2548; LOOP-NEXT: buffer_load_ubyte v28, v[6:7], s[0:3], 0 addr64 offset:2649; LOOP-NEXT: buffer_load_ubyte v30, v[6:7], s[0:3], 0 addr64 offset:2750; LOOP-NEXT: buffer_load_ubyte v22, v[6:7], s[0:3], 0 addr64 offset:2851; LOOP-NEXT: buffer_load_ubyte v33, v[6:7], s[0:3], 0 addr64 offset:2952; LOOP-NEXT: buffer_load_ubyte v34, v[6:7], s[0:3], 0 addr64 offset:3053; LOOP-NEXT: buffer_load_ubyte v35, v[6:7], s[0:3], 0 addr64 offset:3154; LOOP-NEXT: s_waitcnt vmcnt(14)55; LOOP-NEXT: v_lshlrev_b32_e32 v6, 8, v2956; LOOP-NEXT: v_or_b32_e32 v26, v6, v2657; LOOP-NEXT: v_lshlrev_b32_e32 v6, 24, v3258; LOOP-NEXT: v_lshlrev_b32_e32 v7, 16, v3159; LOOP-NEXT: v_or_b32_e32 v29, v6, v760; LOOP-NEXT: v_lshlrev_b32_e32 v6, 8, v3761; LOOP-NEXT: v_lshlrev_b32_e32 v7, 24, v3962; LOOP-NEXT: v_lshlrev_b32_e32 v32, 16, v3863; LOOP-NEXT: v_or_b32_e32 v31, v6, v3664; LOOP-NEXT: v_or_b32_e32 v32, v7, v3265; LOOP-NEXT: v_add_i32_e32 v6, vcc, v0, v466; LOOP-NEXT: v_addc_u32_e32 v7, vcc, v1, v5, vcc67; LOOP-NEXT: v_add_i32_e32 v4, vcc, 32, v468; LOOP-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc69; LOOP-NEXT: v_cmp_gt_u32_e32 vcc, 32, v470; LOOP-NEXT: v_lshlrev_b32_e32 v11, 8, v1171; LOOP-NEXT: v_lshlrev_b32_e32 v13, 24, v1372; LOOP-NEXT: v_lshlrev_b32_e32 v12, 16, v1273; LOOP-NEXT: v_lshlrev_b32_e32 v15, 8, v1574; LOOP-NEXT: v_lshlrev_b32_e32 v17, 24, v1775; LOOP-NEXT: v_lshlrev_b32_e32 v16, 16, v1676; LOOP-NEXT: v_lshlrev_b32_e32 v19, 8, v1977; LOOP-NEXT: s_waitcnt vmcnt(12)78; LOOP-NEXT: v_lshlrev_b32_e32 v21, 24, v2179; LOOP-NEXT: v_lshlrev_b32_e32 v20, 16, v2080; LOOP-NEXT: s_waitcnt vmcnt(10)81; LOOP-NEXT: v_lshlrev_b32_e32 v23, 8, v2382; LOOP-NEXT: s_waitcnt vmcnt(8)83; LOOP-NEXT: v_lshlrev_b32_e32 v25, 24, v2584; LOOP-NEXT: v_lshlrev_b32_e32 v24, 16, v2485; LOOP-NEXT: s_waitcnt vmcnt(6)86; LOOP-NEXT: v_lshlrev_b32_e32 v27, 8, v2787; LOOP-NEXT: s_waitcnt vmcnt(4)88; LOOP-NEXT: v_lshlrev_b32_e32 v30, 24, v3089; LOOP-NEXT: v_lshlrev_b32_e32 v28, 16, v2890; LOOP-NEXT: s_waitcnt vmcnt(2)91; LOOP-NEXT: v_lshlrev_b32_e32 v33, 8, v3392; LOOP-NEXT: s_waitcnt vmcnt(0)93; LOOP-NEXT: v_lshlrev_b32_e32 v35, 24, v3594; LOOP-NEXT: v_lshlrev_b32_e32 v34, 16, v3495; LOOP-NEXT: v_or_b32_e32 v8, v11, v896; LOOP-NEXT: v_or_b32_e32 v11, v13, v1297; LOOP-NEXT: v_or_b32_e32 v9, v15, v998; LOOP-NEXT: v_or_b32_e32 v12, v17, v1699; LOOP-NEXT: v_or_b32_e32 v10, v19, v10100; LOOP-NEXT: v_or_b32_e32 v13, v21, v20101; LOOP-NEXT: v_or_b32_e32 v14, v23, v14102; LOOP-NEXT: v_or_b32_e32 v15, v25, v24103; LOOP-NEXT: v_or_b32_e32 v16, v27, v18104; LOOP-NEXT: v_or_b32_e32 v17, v30, v28105; LOOP-NEXT: v_or_b32_e32 v18, v33, v22106; LOOP-NEXT: v_or_b32_e32 v19, v35, v34107; LOOP-NEXT: v_or_b32_e32 v20, v29, v26108; LOOP-NEXT: v_or_b32_e32 v21, v32, v31109; LOOP-NEXT: v_or_b32_e32 v8, v11, v8110; LOOP-NEXT: v_or_b32_e32 v9, v12, v9111; LOOP-NEXT: v_or_b32_e32 v10, v13, v10112; LOOP-NEXT: v_or_b32_e32 v11, v15, v14113; LOOP-NEXT: v_or_b32_e32 v12, v17, v16114; LOOP-NEXT: v_or_b32_e32 v13, v19, v18115; LOOP-NEXT: v_lshrrev_b32_e32 v14, 16, v20116; LOOP-NEXT: v_bfe_u32 v15, v20, 8, 8117; LOOP-NEXT: buffer_store_byte v20, v[6:7], s[0:3], 0 addr64118; LOOP-NEXT: v_lshrrev_b32_e32 v16, 24, v20119; LOOP-NEXT: v_lshrrev_b32_e32 v17, 16, v21120; LOOP-NEXT: v_bfe_u32 v18, v21, 8, 8121; LOOP-NEXT: buffer_store_byte v21, v[6:7], s[0:3], 0 addr64 offset:4122; LOOP-NEXT: v_lshrrev_b32_e32 v19, 24, v21123; LOOP-NEXT: s_waitcnt expcnt(1)124; LOOP-NEXT: v_lshrrev_b32_e32 v20, 16, v8125; LOOP-NEXT: s_waitcnt expcnt(0)126; LOOP-NEXT: v_bfe_u32 v21, v8, 8, 8127; LOOP-NEXT: buffer_store_byte v8, v[6:7], s[0:3], 0 addr64 offset:8128; LOOP-NEXT: s_waitcnt expcnt(0)129; LOOP-NEXT: v_lshrrev_b32_e32 v8, 24, v8130; LOOP-NEXT: v_lshrrev_b32_e32 v22, 16, v9131; LOOP-NEXT: v_bfe_u32 v23, v9, 8, 8132; LOOP-NEXT: buffer_store_byte v9, v[6:7], s[0:3], 0 addr64 offset:12133; LOOP-NEXT: s_waitcnt expcnt(0)134; LOOP-NEXT: v_lshrrev_b32_e32 v9, 24, v9135; LOOP-NEXT: v_lshrrev_b32_e32 v24, 16, v10136; LOOP-NEXT: v_bfe_u32 v25, v10, 8, 8137; LOOP-NEXT: buffer_store_byte v10, v[6:7], s[0:3], 0 addr64 offset:16138; LOOP-NEXT: s_waitcnt expcnt(0)139; LOOP-NEXT: v_lshrrev_b32_e32 v10, 24, v10140; LOOP-NEXT: v_lshrrev_b32_e32 v26, 16, v11141; LOOP-NEXT: v_bfe_u32 v27, v11, 8, 8142; LOOP-NEXT: buffer_store_byte v11, v[6:7], s[0:3], 0 addr64 offset:20143; LOOP-NEXT: s_waitcnt expcnt(0)144; LOOP-NEXT: v_lshrrev_b32_e32 v11, 24, v11145; LOOP-NEXT: v_lshrrev_b32_e32 v28, 16, v12146; LOOP-NEXT: v_bfe_u32 v29, v12, 8, 8147; LOOP-NEXT: buffer_store_byte v12, v[6:7], s[0:3], 0 addr64 offset:24148; LOOP-NEXT: s_waitcnt expcnt(0)149; LOOP-NEXT: v_lshrrev_b32_e32 v12, 24, v12150; LOOP-NEXT: v_lshrrev_b32_e32 v30, 16, v13151; LOOP-NEXT: v_bfe_u32 v31, v13, 8, 8152; LOOP-NEXT: buffer_store_byte v13, v[6:7], s[0:3], 0 addr64 offset:28153; LOOP-NEXT: s_waitcnt expcnt(0)154; LOOP-NEXT: v_lshrrev_b32_e32 v13, 24, v13155; LOOP-NEXT: buffer_store_byte v15, v[6:7], s[0:3], 0 addr64 offset:1156; LOOP-NEXT: buffer_store_byte v14, v[6:7], s[0:3], 0 addr64 offset:2157; LOOP-NEXT: buffer_store_byte v16, v[6:7], s[0:3], 0 addr64 offset:3158; LOOP-NEXT: buffer_store_byte v18, v[6:7], s[0:3], 0 addr64 offset:5159; LOOP-NEXT: buffer_store_byte v17, v[6:7], s[0:3], 0 addr64 offset:6160; LOOP-NEXT: buffer_store_byte v19, v[6:7], s[0:3], 0 addr64 offset:7161; LOOP-NEXT: buffer_store_byte v21, v[6:7], s[0:3], 0 addr64 offset:9162; LOOP-NEXT: buffer_store_byte v20, v[6:7], s[0:3], 0 addr64 offset:10163; LOOP-NEXT: buffer_store_byte v8, v[6:7], s[0:3], 0 addr64 offset:11164; LOOP-NEXT: buffer_store_byte v23, v[6:7], s[0:3], 0 addr64 offset:13165; LOOP-NEXT: buffer_store_byte v22, v[6:7], s[0:3], 0 addr64 offset:14166; LOOP-NEXT: buffer_store_byte v9, v[6:7], s[0:3], 0 addr64 offset:15167; LOOP-NEXT: buffer_store_byte v25, v[6:7], s[0:3], 0 addr64 offset:17168; LOOP-NEXT: buffer_store_byte v24, v[6:7], s[0:3], 0 addr64 offset:18169; LOOP-NEXT: buffer_store_byte v10, v[6:7], s[0:3], 0 addr64 offset:19170; LOOP-NEXT: buffer_store_byte v27, v[6:7], s[0:3], 0 addr64 offset:21171; LOOP-NEXT: buffer_store_byte v26, v[6:7], s[0:3], 0 addr64 offset:22172; LOOP-NEXT: buffer_store_byte v11, v[6:7], s[0:3], 0 addr64 offset:23173; LOOP-NEXT: buffer_store_byte v29, v[6:7], s[0:3], 0 addr64 offset:25174; LOOP-NEXT: buffer_store_byte v28, v[6:7], s[0:3], 0 addr64 offset:26175; LOOP-NEXT: buffer_store_byte v12, v[6:7], s[0:3], 0 addr64 offset:27176; LOOP-NEXT: buffer_store_byte v31, v[6:7], s[0:3], 0 addr64 offset:29177; LOOP-NEXT: buffer_store_byte v30, v[6:7], s[0:3], 0 addr64 offset:30178; LOOP-NEXT: buffer_store_byte v13, v[6:7], s[0:3], 0 addr64 offset:31179; LOOP-NEXT: s_cbranch_vccnz .LBB0_1180; LOOP-NEXT: ; %bb.2: ; %memcpy-split181; LOOP-NEXT: s_mov_b32 s2, 0182; LOOP-NEXT: s_mov_b32 s3, 0xf000183; LOOP-NEXT: s_mov_b64 s[0:1], 0184; LOOP-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:33185; LOOP-NEXT: buffer_load_ubyte v5, v[2:3], s[0:3], 0 addr64 offset:35186; LOOP-NEXT: buffer_load_ubyte v6, v[2:3], s[0:3], 0 addr64 offset:34187; LOOP-NEXT: buffer_load_ubyte v2, v[2:3], s[0:3], 0 addr64 offset:32188; LOOP-NEXT: s_waitcnt vmcnt(3)189; LOOP-NEXT: v_lshlrev_b32_e32 v3, 8, v4190; LOOP-NEXT: s_waitcnt vmcnt(2)191; LOOP-NEXT: v_lshlrev_b32_e32 v4, 24, v5192; LOOP-NEXT: s_waitcnt vmcnt(1)193; LOOP-NEXT: v_lshlrev_b32_e32 v5, 16, v6194; LOOP-NEXT: s_waitcnt vmcnt(0)195; LOOP-NEXT: v_or_b32_e32 v2, v3, v2196; LOOP-NEXT: v_or_b32_e32 v3, v4, v5197; LOOP-NEXT: v_or_b32_e32 v2, v3, v2198; LOOP-NEXT: v_lshrrev_b32_e32 v3, 16, v2199; LOOP-NEXT: v_bfe_u32 v4, v2, 8, 8200; LOOP-NEXT: buffer_store_byte v2, v[0:1], s[0:3], 0 addr64 offset:32201; LOOP-NEXT: s_waitcnt expcnt(0)202; LOOP-NEXT: v_lshrrev_b32_e32 v2, 24, v2203; LOOP-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:33204; LOOP-NEXT: buffer_store_byte v3, v[0:1], s[0:3], 0 addr64 offset:34205; LOOP-NEXT: buffer_store_byte v2, v[0:1], s[0:3], 0 addr64 offset:35206; LOOP-NEXT: s_endpgm207;208; UNROLL-LABEL: memcpy_p1i8:209; UNROLL: ; %bb.0:210; UNROLL-NEXT: s_mov_b32 s2, 0211; UNROLL-NEXT: s_mov_b32 s3, 0xf000212; UNROLL-NEXT: s_mov_b64 s[0:1], 0213; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64214; UNROLL-NEXT: s_waitcnt vmcnt(0)215; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64216; UNROLL-NEXT: s_waitcnt expcnt(0)217; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:1218; UNROLL-NEXT: s_waitcnt vmcnt(0)219; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:1220; UNROLL-NEXT: s_waitcnt expcnt(0)221; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:2222; UNROLL-NEXT: s_waitcnt vmcnt(0)223; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:2224; UNROLL-NEXT: s_waitcnt expcnt(0)225; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:3226; UNROLL-NEXT: s_waitcnt vmcnt(0)227; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:3228; UNROLL-NEXT: s_waitcnt expcnt(0)229; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:4230; UNROLL-NEXT: s_waitcnt vmcnt(0)231; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:4232; UNROLL-NEXT: s_waitcnt expcnt(0)233; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:5234; UNROLL-NEXT: s_waitcnt vmcnt(0)235; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:5236; UNROLL-NEXT: s_waitcnt expcnt(0)237; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:6238; UNROLL-NEXT: s_waitcnt vmcnt(0)239; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:6240; UNROLL-NEXT: s_waitcnt expcnt(0)241; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:7242; UNROLL-NEXT: s_waitcnt vmcnt(0)243; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:7244; UNROLL-NEXT: s_waitcnt expcnt(0)245; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:8246; UNROLL-NEXT: s_waitcnt vmcnt(0)247; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:8248; UNROLL-NEXT: s_waitcnt expcnt(0)249; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:9250; UNROLL-NEXT: s_waitcnt vmcnt(0)251; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:9252; UNROLL-NEXT: s_waitcnt expcnt(0)253; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:10254; UNROLL-NEXT: s_waitcnt vmcnt(0)255; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:10256; UNROLL-NEXT: s_waitcnt expcnt(0)257; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:11258; UNROLL-NEXT: s_waitcnt vmcnt(0)259; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:11260; UNROLL-NEXT: s_waitcnt expcnt(0)261; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:12262; UNROLL-NEXT: s_waitcnt vmcnt(0)263; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:12264; UNROLL-NEXT: s_waitcnt expcnt(0)265; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:13266; UNROLL-NEXT: s_waitcnt vmcnt(0)267; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:13268; UNROLL-NEXT: s_waitcnt expcnt(0)269; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:14270; UNROLL-NEXT: s_waitcnt vmcnt(0)271; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:14272; UNROLL-NEXT: s_waitcnt expcnt(0)273; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:15274; UNROLL-NEXT: s_waitcnt vmcnt(0)275; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:15276; UNROLL-NEXT: s_waitcnt expcnt(0)277; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:16278; UNROLL-NEXT: s_waitcnt vmcnt(0)279; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:16280; UNROLL-NEXT: s_waitcnt expcnt(0)281; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:17282; UNROLL-NEXT: s_waitcnt vmcnt(0)283; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:17284; UNROLL-NEXT: s_waitcnt expcnt(0)285; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:18286; UNROLL-NEXT: s_waitcnt vmcnt(0)287; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:18288; UNROLL-NEXT: s_waitcnt expcnt(0)289; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:19290; UNROLL-NEXT: s_waitcnt vmcnt(0)291; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:19292; UNROLL-NEXT: s_waitcnt expcnt(0)293; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:20294; UNROLL-NEXT: s_waitcnt vmcnt(0)295; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:20296; UNROLL-NEXT: s_waitcnt expcnt(0)297; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:21298; UNROLL-NEXT: s_waitcnt vmcnt(0)299; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:21300; UNROLL-NEXT: s_waitcnt expcnt(0)301; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:22302; UNROLL-NEXT: s_waitcnt vmcnt(0)303; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:22304; UNROLL-NEXT: s_waitcnt expcnt(0)305; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:23306; UNROLL-NEXT: s_waitcnt vmcnt(0)307; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:23308; UNROLL-NEXT: s_waitcnt expcnt(0)309; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:24310; UNROLL-NEXT: s_waitcnt vmcnt(0)311; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:24312; UNROLL-NEXT: s_waitcnt expcnt(0)313; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:25314; UNROLL-NEXT: s_waitcnt vmcnt(0)315; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:25316; UNROLL-NEXT: s_waitcnt expcnt(0)317; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:26318; UNROLL-NEXT: s_waitcnt vmcnt(0)319; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:26320; UNROLL-NEXT: s_waitcnt expcnt(0)321; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:27322; UNROLL-NEXT: s_waitcnt vmcnt(0)323; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:27324; UNROLL-NEXT: s_waitcnt expcnt(0)325; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:28326; UNROLL-NEXT: s_waitcnt vmcnt(0)327; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:28328; UNROLL-NEXT: s_waitcnt expcnt(0)329; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:29330; UNROLL-NEXT: s_waitcnt vmcnt(0)331; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:29332; UNROLL-NEXT: s_waitcnt expcnt(0)333; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:30334; UNROLL-NEXT: s_waitcnt vmcnt(0)335; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:30336; UNROLL-NEXT: s_waitcnt expcnt(0)337; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:31338; UNROLL-NEXT: s_waitcnt vmcnt(0)339; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:31340; UNROLL-NEXT: s_waitcnt expcnt(0)341; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:32342; UNROLL-NEXT: s_waitcnt vmcnt(0)343; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:32344; UNROLL-NEXT: s_waitcnt expcnt(0)345; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:33346; UNROLL-NEXT: s_waitcnt vmcnt(0)347; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:33348; UNROLL-NEXT: s_waitcnt expcnt(0)349; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:34350; UNROLL-NEXT: s_waitcnt vmcnt(0)351; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:34352; UNROLL-NEXT: buffer_load_ubyte v2, v[2:3], s[0:3], 0 addr64 offset:35353; UNROLL-NEXT: s_waitcnt vmcnt(0)354; UNROLL-NEXT: buffer_store_byte v2, v[0:1], s[0:3], 0 addr64 offset:35355; UNROLL-NEXT: s_endpgm356 call void @llvm.memcpy.p1.p1.i32(ptr addrspace(1) %dst, ptr addrspace(1) %src, i32 36, i1 false)357 ret void358}359 360