266 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2;RUN: llc < %s -mtriple=amdgcn -mcpu=verde | FileCheck -check-prefixes=PREGFX10 %s3;RUN: llc < %s -mtriple=amdgcn -mcpu=tonga | FileCheck -check-prefixes=PREGFX10 %s4;RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 | FileCheck -check-prefixes=GFX10 %s5;RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 | FileCheck -check-prefixes=GFX11 %s6 7define amdgpu_ps void @tbuffer_store(ptr addrspace(8) inreg, <4 x float>, <4 x float>, <4 x float>) {8; PREGFX10-LABEL: tbuffer_store:9; PREGFX10: ; %bb.0: ; %main_body10; PREGFX10-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_DATA_FORMAT_16_16_16_16,BUF_NUM_FORMAT_USCALED]11; PREGFX10-NEXT: tbuffer_store_format_xyzw v[4:7], off, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_SSCALED] glc12; PREGFX10-NEXT: tbuffer_store_format_xyzw v[8:11], off, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32_32,BUF_NUM_FORMAT_UINT] slc13; PREGFX10-NEXT: tbuffer_store_format_xyzw v[8:11], off, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32_32,BUF_NUM_FORMAT_UINT] glc14; PREGFX10-NEXT: s_endpgm15;16; GFX10-LABEL: tbuffer_store:17; GFX10: ; %bb.0: ; %main_body18; GFX10-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_FMT_10_10_10_2_UNORM]19; GFX10-NEXT: tbuffer_store_format_xyzw v[4:7], off, s[0:3], 0 format:[BUF_FMT_8_8_8_8_SINT] glc20; GFX10-NEXT: tbuffer_store_format_xyzw v[8:11], off, s[0:3], 0 format:78 slc21; GFX10-NEXT: tbuffer_store_format_xyzw v[8:11], off, s[0:3], 0 format:78 glc dlc22; GFX10-NEXT: s_endpgm23;24; GFX11-LABEL: tbuffer_store:25; GFX11: ; %bb.0: ; %main_body26; GFX11-NEXT: s_clause 0x327; GFX11-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_FMT_8_8_8_8_USCALED]28; GFX11-NEXT: tbuffer_store_format_xyzw v[4:7], off, s[0:3], 0 format:[BUF_FMT_32_32_32_32_UINT] glc29; GFX11-NEXT: tbuffer_store_format_xyzw v[8:11], off, s[0:3], 0 format:78 slc30; GFX11-NEXT: tbuffer_store_format_xyzw v[8:11], off, s[0:3], 0 format:78 glc dlc31; GFX11-NEXT: s_endpgm32main_body:33 %in1 = bitcast <4 x float> %1 to <4 x i32>34 %in2 = bitcast <4 x float> %2 to <4 x i32>35 %in3 = bitcast <4 x float> %3 to <4 x i32>36 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4i32(<4 x i32> %in1, ptr addrspace(8) %0, i32 0, i32 0, i32 44, i32 0)37 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4i32(<4 x i32> %in2, ptr addrspace(8) %0, i32 0, i32 0, i32 61, i32 1)38 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4i32(<4 x i32> %in3, ptr addrspace(8) %0, i32 0, i32 0, i32 78, i32 2)39 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4f32(<4 x float> %3, ptr addrspace(8) %0, i32 0, i32 0, i32 78, i32 5)40 ret void41}42 43define amdgpu_ps void @tbuffer_store_immoffs(ptr addrspace(8) inreg, <4 x float>) {44; PREGFX10-LABEL: tbuffer_store_immoffs:45; PREGFX10: ; %bb.0: ; %main_body46; PREGFX10-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_DATA_FORMAT_16_16,BUF_NUM_FORMAT_FLOAT] offset:4247; PREGFX10-NEXT: s_endpgm48;49; GFX10-LABEL: tbuffer_store_immoffs:50; GFX10: ; %bb.0: ; %main_body51; GFX10-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], 0 format:117 offset:4252; GFX10-NEXT: s_endpgm53;54; GFX11-LABEL: tbuffer_store_immoffs:55; GFX11: ; %bb.0: ; %main_body56; GFX11-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], 0 format:117 offset:4257; GFX11-NEXT: s_endpgm58main_body:59 %in1 = bitcast <4 x float> %1 to <4 x i32>60 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4i32(<4 x i32> %in1, ptr addrspace(8) %0, i32 42, i32 0, i32 117, i32 0)61 ret void62}63 64define amdgpu_ps void @tbuffer_store_scalar_and_imm_offs(ptr addrspace(8) inreg, <4 x float> %vdata, i32 inreg %soffset) {65; PREGFX10-LABEL: tbuffer_store_scalar_and_imm_offs:66; PREGFX10: ; %bb.0: ; %main_body67; PREGFX10-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], s4 format:[BUF_DATA_FORMAT_16_16,BUF_NUM_FORMAT_FLOAT] offset:4268; PREGFX10-NEXT: s_endpgm69;70; GFX10-LABEL: tbuffer_store_scalar_and_imm_offs:71; GFX10: ; %bb.0: ; %main_body72; GFX10-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], s4 format:117 offset:4273; GFX10-NEXT: s_endpgm74;75; GFX11-LABEL: tbuffer_store_scalar_and_imm_offs:76; GFX11: ; %bb.0: ; %main_body77; GFX11-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], s4 format:117 offset:4278; GFX11-NEXT: s_endpgm79main_body:80 %in1 = bitcast <4 x float> %vdata to <4 x i32>81 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4i32(<4 x i32> %in1, ptr addrspace(8) %0, i32 42, i32 %soffset, i32 117, i32 0)82 ret void83}84 85define amdgpu_ps void @buffer_store_ofs(ptr addrspace(8) inreg, <4 x float> %vdata, i32 %voffset) {86; PREGFX10-LABEL: buffer_store_ofs:87; PREGFX10: ; %bb.0: ; %main_body88; PREGFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_DATA_FORMAT_8_8,BUF_NUM_FORMAT_FLOAT] offen89; PREGFX10-NEXT: s_endpgm90;91; GFX10-LABEL: buffer_store_ofs:92; GFX10: ; %bb.0: ; %main_body93; GFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:115 offen94; GFX10-NEXT: s_endpgm95;96; GFX11-LABEL: buffer_store_ofs:97; GFX11: ; %bb.0: ; %main_body98; GFX11-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:115 offen99; GFX11-NEXT: s_endpgm100main_body:101 %in1 = bitcast <4 x float> %vdata to <4 x i32>102 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4i32(<4 x i32> %in1, ptr addrspace(8) %0, i32 %voffset, i32 0, i32 115, i32 0)103 ret void104}105 106define amdgpu_ps void @buffer_store_x1(ptr addrspace(8) inreg %rsrc, float %data) {107; PREGFX10-LABEL: buffer_store_x1:108; PREGFX10: ; %bb.0: ; %main_body109; PREGFX10-NEXT: tbuffer_store_format_x v0, off, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_FLOAT]110; PREGFX10-NEXT: s_endpgm111;112; GFX10-LABEL: buffer_store_x1:113; GFX10: ; %bb.0: ; %main_body114; GFX10-NEXT: tbuffer_store_format_x v0, off, s[0:3], 0 format:125115; GFX10-NEXT: s_endpgm116;117; GFX11-LABEL: buffer_store_x1:118; GFX11: ; %bb.0: ; %main_body119; GFX11-NEXT: tbuffer_store_format_x v0, off, s[0:3], 0 format:125120; GFX11-NEXT: s_endpgm121main_body:122 %data.i = bitcast float %data to i32123 call void @llvm.amdgcn.raw.ptr.tbuffer.store.i32(i32 %data.i, ptr addrspace(8) %rsrc, i32 0, i32 0, i32 125, i32 0)124 ret void125}126 127define amdgpu_ps void @buffer_store_x2(ptr addrspace(8) inreg %rsrc, <2 x float> %data) {128; PREGFX10-LABEL: buffer_store_x2:129; PREGFX10: ; %bb.0: ; %main_body130; PREGFX10-NEXT: tbuffer_store_format_xy v[0:1], off, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED]131; PREGFX10-NEXT: s_endpgm132;133; GFX10-LABEL: buffer_store_x2:134; GFX10: ; %bb.0: ; %main_body135; GFX10-NEXT: tbuffer_store_format_xy v[0:1], off, s[0:3], 0 format:[BUF_FMT_10_11_11_SSCALED]136; GFX10-NEXT: s_endpgm137;138; GFX11-LABEL: buffer_store_x2:139; GFX11: ; %bb.0: ; %main_body140; GFX11-NEXT: tbuffer_store_format_xy v[0:1], off, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM]141; GFX11-NEXT: s_endpgm142main_body:143 %data.i = bitcast <2 x float> %data to <2 x i32>144 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v2i32(<2 x i32> %data.i, ptr addrspace(8) %rsrc, i32 0, i32 0, i32 33, i32 0)145 ret void146}147 148define amdgpu_ps void @buffer_store_voffset_large_12bit(ptr addrspace(8) inreg %rsrc, <4 x float> %data) {149; PREGFX10-LABEL: buffer_store_voffset_large_12bit:150; PREGFX10: ; %bb.0: ; %main_body151; PREGFX10-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offset:4092152; PREGFX10-NEXT: s_endpgm153;154; GFX10-LABEL: buffer_store_voffset_large_12bit:155; GFX10: ; %bb.0: ; %main_body156; GFX10-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offset:4092157; GFX10-NEXT: s_endpgm158;159; GFX11-LABEL: buffer_store_voffset_large_12bit:160; GFX11: ; %bb.0: ; %main_body161; GFX11-NEXT: tbuffer_store_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offset:4092162; GFX11-NEXT: s_endpgm163main_body:164 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4f32(<4 x float> %data, ptr addrspace(8) %rsrc, i32 4092, i32 0, i32 63, i32 0)165 ret void166}167 168define amdgpu_ps void @buffer_store_voffset_large_13bit(ptr addrspace(8) inreg %rsrc, <4 x float> %data) {169; PREGFX10-LABEL: buffer_store_voffset_large_13bit:170; PREGFX10: ; %bb.0: ; %main_body171; PREGFX10-NEXT: v_mov_b32_e32 v4, 0x1000172; PREGFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offen offset:4092173; PREGFX10-NEXT: s_endpgm174;175; GFX10-LABEL: buffer_store_voffset_large_13bit:176; GFX10: ; %bb.0: ; %main_body177; GFX10-NEXT: v_mov_b32_e32 v4, 0x1000178; GFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offen offset:4092179; GFX10-NEXT: s_endpgm180;181; GFX11-LABEL: buffer_store_voffset_large_13bit:182; GFX11: ; %bb.0: ; %main_body183; GFX11-NEXT: v_mov_b32_e32 v4, 0x1000184; GFX11-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offen offset:4092185; GFX11-NEXT: s_endpgm186main_body:187 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4f32(<4 x float> %data, ptr addrspace(8) %rsrc, i32 8188, i32 0, i32 63, i32 0)188 ret void189}190 191define amdgpu_ps void @buffer_store_voffset_large_16bit(ptr addrspace(8) inreg %rsrc, <4 x float> %data) {192; PREGFX10-LABEL: buffer_store_voffset_large_16bit:193; PREGFX10: ; %bb.0: ; %main_body194; PREGFX10-NEXT: v_mov_b32_e32 v4, 0xf000195; PREGFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offen offset:4092196; PREGFX10-NEXT: s_endpgm197;198; GFX10-LABEL: buffer_store_voffset_large_16bit:199; GFX10: ; %bb.0: ; %main_body200; GFX10-NEXT: v_mov_b32_e32 v4, 0xf000201; GFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offen offset:4092202; GFX10-NEXT: s_endpgm203;204; GFX11-LABEL: buffer_store_voffset_large_16bit:205; GFX11: ; %bb.0: ; %main_body206; GFX11-NEXT: v_mov_b32_e32 v4, 0xf000207; GFX11-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offen offset:4092208; GFX11-NEXT: s_endpgm209main_body:210 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4f32(<4 x float> %data, ptr addrspace(8) %rsrc, i32 65532, i32 0, i32 63, i32 0)211 ret void212}213 214define amdgpu_ps void @buffer_store_voffset_large_23bit(ptr addrspace(8) inreg %rsrc, <4 x float> %data) {215; PREGFX10-LABEL: buffer_store_voffset_large_23bit:216; PREGFX10: ; %bb.0: ; %main_body217; PREGFX10-NEXT: v_mov_b32_e32 v4, 0x7ff000218; PREGFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offen offset:4092219; PREGFX10-NEXT: s_endpgm220;221; GFX10-LABEL: buffer_store_voffset_large_23bit:222; GFX10: ; %bb.0: ; %main_body223; GFX10-NEXT: v_mov_b32_e32 v4, 0x7ff000224; GFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offen offset:4092225; GFX10-NEXT: s_endpgm226;227; GFX11-LABEL: buffer_store_voffset_large_23bit:228; GFX11: ; %bb.0: ; %main_body229; GFX11-NEXT: v_mov_b32_e32 v4, 0x7ff000230; GFX11-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offen offset:4092231; GFX11-NEXT: s_endpgm232main_body:233 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4f32(<4 x float> %data, ptr addrspace(8) %rsrc, i32 8388604, i32 0, i32 63, i32 0)234 ret void235}236 237define amdgpu_ps void @buffer_store_voffset_large_24bit(ptr addrspace(8) inreg %rsrc, <4 x float> %data) {238; PREGFX10-LABEL: buffer_store_voffset_large_24bit:239; PREGFX10: ; %bb.0: ; %main_body240; PREGFX10-NEXT: v_mov_b32_e32 v4, 0xfff000241; PREGFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offen offset:4092242; PREGFX10-NEXT: s_endpgm243;244; GFX10-LABEL: buffer_store_voffset_large_24bit:245; GFX10: ; %bb.0: ; %main_body246; GFX10-NEXT: v_mov_b32_e32 v4, 0xfff000247; GFX10-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offen offset:4092248; GFX10-NEXT: s_endpgm249;250; GFX11-LABEL: buffer_store_voffset_large_24bit:251; GFX11: ; %bb.0: ; %main_body252; GFX11-NEXT: v_mov_b32_e32 v4, 0xfff000253; GFX11-NEXT: tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offen offset:4092254; GFX11-NEXT: s_endpgm255main_body:256 call void @llvm.amdgcn.raw.ptr.tbuffer.store.v4f32(<4 x float> %data, ptr addrspace(8) %rsrc, i32 16777212, i32 0, i32 63, i32 0)257 ret void258}259 260declare void @llvm.amdgcn.raw.ptr.tbuffer.store.i32(i32, ptr addrspace(8), i32, i32, i32, i32) #0261declare void @llvm.amdgcn.raw.ptr.tbuffer.store.v2i32(<2 x i32>, ptr addrspace(8), i32, i32, i32, i32) #0262declare void @llvm.amdgcn.raw.ptr.tbuffer.store.v4i32(<4 x i32>, ptr addrspace(8), i32, i32, i32, i32) #0263declare void @llvm.amdgcn.raw.ptr.tbuffer.store.v4f32(<4 x float>, ptr addrspace(8), i32, i32, i32, i32) #0264attributes #0 = { nounwind }265attributes #1 = { nounwind readonly }266