330 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=amdgcn -mcpu=gfx90a < %s | FileCheck -check-prefix=GCN %s3 4define amdgpu_kernel void @test_iglp_opt() #0 {5; GCN-LABEL: test_iglp_opt:6; GCN: ; %bb.0: ; %entry7; GCN-NEXT: ; iglp_opt mask(0x00000000)8; GCN-NEXT: s_endpgm9entry:10 call void @llvm.amdgcn.iglp.opt(i32 0) #111 ret void12}13 14define amdgpu_kernel void @test_iglp_opt_mfma_gemm(ptr addrspace(3) noalias %in, ptr addrspace(3) noalias %out) #0 {15; GCN-LABEL: test_iglp_opt_mfma_gemm:16; GCN: ; %bb.0: ; %entry17; GCN-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2418; GCN-NEXT: v_lshlrev_b32_e32 v0, 7, v019; GCN-NEXT: v_and_b32_e32 v0, 0x1ff80, v020; GCN-NEXT: v_mov_b32_e32 v3, 2.021; GCN-NEXT: ; iglp_opt mask(0x00000000)22; GCN-NEXT: s_waitcnt lgkmcnt(0)23; GCN-NEXT: v_add_u32_e32 v1, s0, v024; GCN-NEXT: v_add_u32_e32 v2, 0x6000, v125; GCN-NEXT: ds_read_b128 a[28:31], v2 offset:5745626; GCN-NEXT: ds_read_b128 a[24:27], v2 offset:5744027; GCN-NEXT: ds_read_b128 a[20:23], v2 offset:5742428; GCN-NEXT: ds_read_b128 a[16:19], v2 offset:5740829; GCN-NEXT: ds_read_b128 a[0:3], v2 offset:5734430; GCN-NEXT: ds_read_b128 a[4:7], v2 offset:5736031; GCN-NEXT: ds_read_b128 a[8:11], v2 offset:5737632; GCN-NEXT: ds_read_b128 a[12:15], v2 offset:5739233; GCN-NEXT: v_mov_b32_e32 v2, 1.034; GCN-NEXT: ds_read_b128 a[60:63], v1 offset:4926435; GCN-NEXT: ds_read_b128 a[56:59], v1 offset:4924836; GCN-NEXT: ds_read_b128 a[52:55], v1 offset:4923237; GCN-NEXT: ds_read_b128 a[48:51], v1 offset:4921638; GCN-NEXT: ds_read_b128 a[44:47], v1 offset:4920039; GCN-NEXT: ds_read_b128 a[40:43], v1 offset:4918440; GCN-NEXT: ds_read_b128 a[36:39], v1 offset:4916841; GCN-NEXT: ds_read_b128 a[32:35], v1 offset:4915242; GCN-NEXT: s_waitcnt lgkmcnt(8)43; GCN-NEXT: v_mfma_f32_32x32x1f32 a[0:31], v2, v3, a[0:31]44; GCN-NEXT: ds_read_b128 a[156:159], v1 offset:11245; GCN-NEXT: ds_read_b128 a[152:155], v1 offset:9646; GCN-NEXT: ds_read_b128 a[68:71], v1 offset:2459247; GCN-NEXT: ds_read_b128 a[64:67], v1 offset:2457648; GCN-NEXT: v_add_u32_e32 v0, s1, v049; GCN-NEXT: s_waitcnt lgkmcnt(4)50; GCN-NEXT: v_mfma_f32_32x32x1f32 a[32:63], v2, v3, a[32:63]51; GCN-NEXT: ds_read_b128 a[148:151], v1 offset:8052; GCN-NEXT: ds_read_b128 a[144:147], v1 offset:6453; GCN-NEXT: ds_read_b128 a[128:131], v154; GCN-NEXT: ds_read_b128 a[132:135], v1 offset:1655; GCN-NEXT: ds_read_b128 a[136:139], v1 offset:3256; GCN-NEXT: ds_read_b128 a[140:143], v1 offset:4857; GCN-NEXT: s_waitcnt lgkmcnt(0)58; GCN-NEXT: v_mfma_f32_32x32x1f32 a[128:159], v2, v3, a[128:159]59; GCN-NEXT: ds_read_b128 a[124:127], v1 offset:830460; GCN-NEXT: ds_read_b128 a[120:123], v1 offset:828861; GCN-NEXT: ds_read_b128 a[116:119], v1 offset:827262; GCN-NEXT: ds_read_b128 a[112:115], v1 offset:825663; GCN-NEXT: ds_read_b128 a[108:111], v1 offset:824064; GCN-NEXT: ds_read_b128 a[104:107], v1 offset:822465; GCN-NEXT: ds_read_b128 a[100:103], v1 offset:820866; GCN-NEXT: ds_read_b128 a[96:99], v1 offset:819267; GCN-NEXT: s_waitcnt lgkmcnt(0)68; GCN-NEXT: v_mfma_f32_32x32x1f32 a[96:127], v2, v3, a[96:127]69; GCN-NEXT: ds_read_b128 a[92:95], v1 offset:2468870; GCN-NEXT: ds_read_b128 a[88:91], v1 offset:2467271; GCN-NEXT: ds_read_b128 a[84:87], v1 offset:2465672; GCN-NEXT: ds_read_b128 a[80:83], v1 offset:2464073; GCN-NEXT: ds_read_b128 a[76:79], v1 offset:2462474; GCN-NEXT: ds_read_b128 a[72:75], v1 offset:2460875; GCN-NEXT: s_nop 276; GCN-NEXT: ds_write_b128 v0, a[156:159] offset:11277; GCN-NEXT: ds_write_b128 v0, a[152:155] offset:9678; GCN-NEXT: ds_write_b128 v0, a[148:151] offset:8079; GCN-NEXT: ds_write_b128 v0, a[144:147] offset:6480; GCN-NEXT: ds_write_b128 v0, a[140:143] offset:4881; GCN-NEXT: ds_write_b128 v0, a[136:139] offset:3282; GCN-NEXT: ds_write_b128 v0, a[132:135] offset:1683; GCN-NEXT: ds_write_b128 v0, a[128:131]84; GCN-NEXT: v_mov_b32_e32 v0, s185; GCN-NEXT: s_waitcnt lgkmcnt(8)86; GCN-NEXT: v_mfma_f32_32x32x1f32 a[64:95], v2, v3, a[64:95]87; GCN-NEXT: ds_write_b128 v0, a[56:59] offset:2467288; GCN-NEXT: ds_write_b128 v0, a[60:63] offset:2468889; GCN-NEXT: ds_write_b128 v0, a[48:51] offset:2464090; GCN-NEXT: ds_write_b128 v0, a[120:123] offset:828891; GCN-NEXT: ds_write_b128 v0, a[124:127] offset:830492; GCN-NEXT: ds_write_b128 v0, a[112:115] offset:825693; GCN-NEXT: ds_write_b128 v0, a[116:119] offset:827294; GCN-NEXT: ds_write_b128 v0, a[104:107] offset:822495; GCN-NEXT: ds_write_b128 v0, a[108:111] offset:824096; GCN-NEXT: ds_write_b128 v0, a[96:99] offset:819297; GCN-NEXT: ds_write_b128 v0, a[100:103] offset:820898; GCN-NEXT: ds_write_b128 v0, a[52:55] offset:2465699; GCN-NEXT: ds_write_b128 v0, a[40:43] offset:24608100; GCN-NEXT: ds_write_b128 v0, a[44:47] offset:24624101; GCN-NEXT: ds_write_b128 v0, a[32:35] offset:24576102; GCN-NEXT: ds_write_b128 v0, a[36:39] offset:24592103; GCN-NEXT: ds_write_b128 v0, a[24:27] offset:32864104; GCN-NEXT: ds_write_b128 v0, a[28:31] offset:32880105; GCN-NEXT: ds_write_b128 v0, a[16:19] offset:32832106; GCN-NEXT: ds_write_b128 v0, a[88:91] offset:16480107; GCN-NEXT: ds_write_b128 v0, a[92:95] offset:16496108; GCN-NEXT: ds_write_b128 v0, a[80:83] offset:16448109; GCN-NEXT: ds_write_b128 v0, a[84:87] offset:16464110; GCN-NEXT: ds_write_b128 v0, a[72:75] offset:16416111; GCN-NEXT: ds_write_b128 v0, a[76:79] offset:16432112; GCN-NEXT: ds_write_b128 v0, a[64:67] offset:16384113; GCN-NEXT: ds_write_b128 v0, a[68:71] offset:16400114; GCN-NEXT: ds_write_b128 v0, a[20:23] offset:32848115; GCN-NEXT: ds_write_b128 v0, a[8:11] offset:32800116; GCN-NEXT: ds_write_b128 v0, a[12:15] offset:32816117; GCN-NEXT: ds_write_b128 v0, a[0:3] offset:32768118; GCN-NEXT: ds_write_b128 v0, a[4:7] offset:32784119; GCN-NEXT: s_endpgm120entry:121 call void @llvm.amdgcn.iglp.opt(i32 0)122 %idx = call i32 @llvm.amdgcn.workitem.id.x()123 %load.0.addr = getelementptr <32 x float>, ptr addrspace(3) %in, i32 %idx124 %load.0 = load <32 x float>, ptr addrspace(3) %load.0.addr125 %load.1.addr = getelementptr <32 x float>, ptr addrspace(3) %load.0.addr, i32 64126 %load.1 = load <32 x float>, ptr addrspace(3) %load.1.addr127 %load.2.addr = getelementptr <32 x float>, ptr addrspace(3) %load.1.addr, i32 128128 %load.2 = load <32 x float>, ptr addrspace(3) %load.2.addr129 %load.3.addr = getelementptr <32 x float>, ptr addrspace(3) %load.2.addr, i32 192130 %load.3 = load <32 x float>, ptr addrspace(3) %load.3.addr131 %load.4.addr = getelementptr <32 x float>, ptr addrspace(3) %load.3.addr, i32 256132 %load.4 = load <32 x float>, ptr addrspace(3) %load.4.addr133 %mai.0 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.0, i32 0, i32 0, i32 0)134 %mai.1 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.1, i32 0, i32 0, i32 0)135 %mai.2 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.2, i32 0, i32 0, i32 0)136 %mai.3 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.3, i32 0, i32 0, i32 0)137 %mai.4 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.4, i32 0, i32 0, i32 0)138 %store.0.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 %idx139 store <32 x float> %mai.0, ptr addrspace(3) %store.0.addr140 %store.1.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 64141 store <32 x float> %mai.1, ptr addrspace(3) %store.1.addr142 %store.2.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 128143 store <32 x float> %mai.2, ptr addrspace(3) %store.2.addr144 %store.3.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 192145 store <32 x float> %mai.3, ptr addrspace(3) %store.3.addr146 %store.4.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 256147 store <32 x float> %mai.4, ptr addrspace(3) %store.4.addr148 ret void149}150 151 152define amdgpu_kernel void @test_iglp_opt_rev_mfma_gemm(ptr addrspace(3) noalias %in, ptr addrspace(3) noalias %out) #0 {153; GCN-LABEL: test_iglp_opt_rev_mfma_gemm:154; GCN: ; %bb.0: ; %entry155; GCN-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24156; GCN-NEXT: v_lshlrev_b32_e32 v0, 7, v0157; GCN-NEXT: v_and_b32_e32 v0, 0x1ff80, v0158; GCN-NEXT: v_mov_b32_e32 v2, 1.0159; GCN-NEXT: v_mov_b32_e32 v1, 2.0160; GCN-NEXT: s_waitcnt lgkmcnt(0)161; GCN-NEXT: v_add_u32_e32 v3, s0, v0162; GCN-NEXT: ds_read_b128 a[28:31], v3 offset:112163; GCN-NEXT: ds_read_b128 a[24:27], v3 offset:96164; GCN-NEXT: ds_read_b128 a[20:23], v3 offset:80165; GCN-NEXT: ds_read_b128 a[16:19], v3 offset:64166; GCN-NEXT: ds_read_b128 a[0:3], v3167; GCN-NEXT: ds_read_b128 a[4:7], v3 offset:16168; GCN-NEXT: ds_read_b128 a[8:11], v3 offset:32169; GCN-NEXT: ds_read_b128 a[12:15], v3 offset:48170; GCN-NEXT: s_waitcnt lgkmcnt(0)171; GCN-NEXT: v_mfma_f32_32x32x1f32 a[0:31], v2, v1, a[0:31]172; GCN-NEXT: ds_read_b128 a[156:159], v3 offset:8304173; GCN-NEXT: ds_read_b128 a[152:155], v3 offset:8288174; GCN-NEXT: ds_read_b128 a[148:151], v3 offset:8272175; GCN-NEXT: ds_read_b128 a[144:147], v3 offset:8256176; GCN-NEXT: ds_read_b128 a[140:143], v3 offset:8240177; GCN-NEXT: ds_read_b128 a[136:139], v3 offset:8224178; GCN-NEXT: ds_read_b128 a[132:135], v3 offset:8208179; GCN-NEXT: ds_read_b128 a[128:131], v3 offset:8192180; GCN-NEXT: v_add_u32_e32 v4, 0x6000, v3181; GCN-NEXT: v_add_u32_e32 v0, s1, v0182; GCN-NEXT: ; iglp_opt mask(0x00000001)183; GCN-NEXT: s_waitcnt lgkmcnt(0)184; GCN-NEXT: v_mfma_f32_32x32x1f32 a[128:159], v2, v1, a[128:159]185; GCN-NEXT: ds_read_b128 a[124:127], v3 offset:24688186; GCN-NEXT: ds_read_b128 a[120:123], v3 offset:24672187; GCN-NEXT: ds_read_b128 a[116:119], v3 offset:24656188; GCN-NEXT: ds_read_b128 a[112:115], v3 offset:24640189; GCN-NEXT: ds_read_b128 a[108:111], v3 offset:24624190; GCN-NEXT: ds_read_b128 a[104:107], v3 offset:24608191; GCN-NEXT: ds_read_b128 a[100:103], v3 offset:24592192; GCN-NEXT: ds_read_b128 a[96:99], v3 offset:24576193; GCN-NEXT: s_waitcnt lgkmcnt(0)194; GCN-NEXT: v_mfma_f32_32x32x1f32 a[96:127], v2, v1, a[96:127]195; GCN-NEXT: ds_read_b128 a[92:95], v3 offset:49264196; GCN-NEXT: ds_read_b128 a[88:91], v3 offset:49248197; GCN-NEXT: ds_read_b128 a[84:87], v3 offset:49232198; GCN-NEXT: ds_read_b128 a[80:83], v3 offset:49216199; GCN-NEXT: ds_read_b128 a[76:79], v3 offset:49200200; GCN-NEXT: ds_read_b128 a[72:75], v3 offset:49184201; GCN-NEXT: ds_read_b128 a[68:71], v3 offset:49168202; GCN-NEXT: ds_read_b128 a[64:67], v3 offset:49152203; GCN-NEXT: s_waitcnt lgkmcnt(0)204; GCN-NEXT: v_mfma_f32_32x32x1f32 a[64:95], v2, v1, a[64:95]205; GCN-NEXT: ds_read_b128 a[60:63], v4 offset:57456206; GCN-NEXT: ds_read_b128 a[56:59], v4 offset:57440207; GCN-NEXT: ds_read_b128 a[52:55], v4 offset:57424208; GCN-NEXT: ds_read_b128 a[48:51], v4 offset:57408209; GCN-NEXT: ds_read_b128 a[32:35], v4 offset:57344210; GCN-NEXT: ds_read_b128 a[36:39], v4 offset:57360211; GCN-NEXT: ds_read_b128 a[40:43], v4 offset:57376212; GCN-NEXT: ds_read_b128 a[44:47], v4 offset:57392213; GCN-NEXT: s_waitcnt lgkmcnt(0)214; GCN-NEXT: v_mfma_f32_32x32x1f32 a[32:63], v2, v1, a[32:63]215; GCN-NEXT: ds_write_b128 v0, a[28:31] offset:112216; GCN-NEXT: ds_write_b128 v0, a[24:27] offset:96217; GCN-NEXT: ds_write_b128 v0, a[20:23] offset:80218; GCN-NEXT: ds_write_b128 v0, a[16:19] offset:64219; GCN-NEXT: ds_write_b128 v0, a[12:15] offset:48220; GCN-NEXT: ds_write_b128 v0, a[8:11] offset:32221; GCN-NEXT: ds_write_b128 v0, a[4:7] offset:16222; GCN-NEXT: ds_write_b128 v0, a[0:3]223; GCN-NEXT: v_mov_b32_e32 v0, s1224; GCN-NEXT: ds_write_b128 v0, a[152:155] offset:8288225; GCN-NEXT: ds_write_b128 v0, a[156:159] offset:8304226; GCN-NEXT: ds_write_b128 v0, a[144:147] offset:8256227; GCN-NEXT: ds_write_b128 v0, a[148:151] offset:8272228; GCN-NEXT: ds_write_b128 v0, a[136:139] offset:8224229; GCN-NEXT: ds_write_b128 v0, a[140:143] offset:8240230; GCN-NEXT: ds_write_b128 v0, a[128:131] offset:8192231; GCN-NEXT: ds_write_b128 v0, a[132:135] offset:8208232; GCN-NEXT: ds_write_b128 v0, a[120:123] offset:16480233; GCN-NEXT: ds_write_b128 v0, a[124:127] offset:16496234; GCN-NEXT: ds_write_b128 v0, a[112:115] offset:16448235; GCN-NEXT: ds_write_b128 v0, a[116:119] offset:16464236; GCN-NEXT: ds_write_b128 v0, a[104:107] offset:16416237; GCN-NEXT: ds_write_b128 v0, a[108:111] offset:16432238; GCN-NEXT: ds_write_b128 v0, a[96:99] offset:16384239; GCN-NEXT: ds_write_b128 v0, a[100:103] offset:16400240; GCN-NEXT: ds_write_b128 v0, a[88:91] offset:24672241; GCN-NEXT: ds_write_b128 v0, a[92:95] offset:24688242; GCN-NEXT: ds_write_b128 v0, a[80:83] offset:24640243; GCN-NEXT: ds_write_b128 v0, a[84:87] offset:24656244; GCN-NEXT: ds_write_b128 v0, a[72:75] offset:24608245; GCN-NEXT: ds_write_b128 v0, a[76:79] offset:24624246; GCN-NEXT: ds_write_b128 v0, a[64:67] offset:24576247; GCN-NEXT: ds_write_b128 v0, a[68:71] offset:24592248; GCN-NEXT: ds_write_b128 v0, a[56:59] offset:32864249; GCN-NEXT: ds_write_b128 v0, a[60:63] offset:32880250; GCN-NEXT: ds_write_b128 v0, a[48:51] offset:32832251; GCN-NEXT: ds_write_b128 v0, a[52:55] offset:32848252; GCN-NEXT: ds_write_b128 v0, a[40:43] offset:32800253; GCN-NEXT: ds_write_b128 v0, a[44:47] offset:32816254; GCN-NEXT: ds_write_b128 v0, a[32:35] offset:32768255; GCN-NEXT: ds_write_b128 v0, a[36:39] offset:32784256; GCN-NEXT: s_endpgm257entry:258 call void @llvm.amdgcn.iglp.opt(i32 1)259 %idx = call i32 @llvm.amdgcn.workitem.id.x()260 %load.0.addr = getelementptr <32 x float>, ptr addrspace(3) %in, i32 %idx261 %load.0 = load <32 x float>, ptr addrspace(3) %load.0.addr262 %load.1.addr = getelementptr <32 x float>, ptr addrspace(3) %load.0.addr, i32 64263 %load.1 = load <32 x float>, ptr addrspace(3) %load.1.addr264 %load.2.addr = getelementptr <32 x float>, ptr addrspace(3) %load.1.addr, i32 128265 %load.2 = load <32 x float>, ptr addrspace(3) %load.2.addr266 %load.3.addr = getelementptr <32 x float>, ptr addrspace(3) %load.2.addr, i32 192267 %load.3 = load <32 x float>, ptr addrspace(3) %load.3.addr268 %load.4.addr = getelementptr <32 x float>, ptr addrspace(3) %load.3.addr, i32 256269 %load.4 = load <32 x float>, ptr addrspace(3) %load.4.addr270 %mai.0 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.0, i32 0, i32 0, i32 0)271 %mai.1 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.1, i32 0, i32 0, i32 0)272 %mai.2 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.2, i32 0, i32 0, i32 0)273 %mai.3 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.3, i32 0, i32 0, i32 0)274 %mai.4 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %load.4, i32 0, i32 0, i32 0)275 %store.0.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 %idx276 store <32 x float> %mai.0, ptr addrspace(3) %store.0.addr277 %store.1.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 64278 store <32 x float> %mai.1, ptr addrspace(3) %store.1.addr279 %store.2.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 128280 store <32 x float> %mai.2, ptr addrspace(3) %store.2.addr281 %store.3.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 192282 store <32 x float> %mai.3, ptr addrspace(3) %store.3.addr283 %store.4.addr = getelementptr <32 x float>, ptr addrspace(3) %out, i32 256284 store <32 x float> %mai.4, ptr addrspace(3) %store.4.addr285 ret void286}287 288define amdgpu_kernel void @test_iglp_opt_asm_sideeffect(ptr addrspace(3) noalias %in, ptr addrspace(3) noalias %out) #0 {289; GCN-LABEL: test_iglp_opt_asm_sideeffect:290; GCN: ; %bb.0: ; %entry291; GCN-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24292; GCN-NEXT: v_lshlrev_b32_e32 v0, 2, v0293; GCN-NEXT: v_and_b32_e32 v0, 0xffc, v0294; GCN-NEXT: ; iglp_opt mask(0x00000000)295; GCN-NEXT: s_waitcnt lgkmcnt(0)296; GCN-NEXT: v_add_u32_e32 v1, s0, v0297; GCN-NEXT: ds_read_b32 v2, v1298; GCN-NEXT: v_add_u32_e32 v0, s1, v0299; GCN-NEXT: v_mov_b32_e32 v1, s0300; GCN-NEXT: s_waitcnt lgkmcnt(0)301; GCN-NEXT: ds_write_b32 v0, v2302; GCN-NEXT: ;;#ASMSTART303; GCN-NEXT: ;;#ASMEND304; GCN-NEXT: ds_read_b32 v1, v1 offset:256305; GCN-NEXT: v_mov_b32_e32 v0, s1306; GCN-NEXT: s_waitcnt lgkmcnt(0)307; GCN-NEXT: ds_write_b32 v0, v1 offset:256308; GCN-NEXT: s_endpgm309entry:310 %idx = call i32 @llvm.amdgcn.workitem.id.x()311 %load.0.addr = getelementptr float, ptr addrspace(3) %in, i32 %idx312 %load.0 = load float, ptr addrspace(3) %load.0.addr313 %store.0.addr = getelementptr float, ptr addrspace(3) %out, i32 %idx314 store float %load.0, ptr addrspace(3) %store.0.addr315 call void asm sideeffect "", ""() #1316 call void @llvm.amdgcn.iglp.opt(i32 0) #1317 %load.1.addr = getelementptr float, ptr addrspace(3) %in, i32 64318 %load.1 = load float, ptr addrspace(3) %load.1.addr319 %store.1.addr = getelementptr float, ptr addrspace(3) %out, i32 64320 store float %load.1, ptr addrspace(3) %store.1.addr321 ret void322}323 324declare void @llvm.amdgcn.iglp.opt(i32) #1325declare i32 @llvm.amdgcn.workitem.id.x() #1326declare <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float, float, <32 x float>, i32, i32, i32) #1327 328attributes #0 = { nounwind "amdgpu-flat-work-group-size"="1,256" }329attributes #1 = { convergent nounwind }330