1; RUN: llc -march=amdgcn -mcpu=hawaii -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=CI %s 2; RUN: llc -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI %s 3; RUN: llc -march=amdgcn -mcpu=gfx900 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI %s 4 5; GCN-LABEL: {{^}}i1_func_void: 6; GCN: buffer_load_ubyte v0, off 7; GCN-NEXT: s_waitcnt 8; GCN-NEXT: s_setpc_b64 9define i1 @i1_func_void() #0 { 10 %val = load i1, i1 addrspace(1)* undef 11 ret i1 %val 12} 13 14; FIXME: Missing and? 15; GCN-LABEL: {{^}}i1_zeroext_func_void: 16; GCN: buffer_load_ubyte v0, off 17; GCN-NEXT: s_waitcnt vmcnt(0) 18; GCN-NEXT: s_setpc_b64 19define zeroext i1 @i1_zeroext_func_void() #0 { 20 %val = load i1, i1 addrspace(1)* undef 21 ret i1 %val 22} 23 24; GCN-LABEL: {{^}}i1_signext_func_void: 25; GCN: buffer_load_ubyte v0, off 26; GCN-NEXT: s_waitcnt vmcnt(0) 27; GCN-NEXT: v_bfe_i32 v0, v0, 0, 1{{$}} 28; GCN-NEXT: s_setpc_b64 29define signext i1 @i1_signext_func_void() #0 { 30 %val = load i1, i1 addrspace(1)* undef 31 ret i1 %val 32} 33 34; GCN-LABEL: {{^}}i8_func_void: 35; GCN: buffer_load_ubyte v0, off 36; GCN-NEXT: s_waitcnt vmcnt(0) 37; GCN-NEXT: s_setpc_b64 38define i8 @i8_func_void() #0 { 39 %val = load i8, i8 addrspace(1)* undef 40 ret i8 %val 41} 42 43; GCN-LABEL: {{^}}i8_zeroext_func_void: 44; GCN: buffer_load_ubyte v0, off 45; GCN-NEXT: s_waitcnt vmcnt(0) 46; GCN-NEXT: s_setpc_b64 47define zeroext i8 @i8_zeroext_func_void() #0 { 48 %val = load i8, i8 addrspace(1)* undef 49 ret i8 %val 50} 51 52; GCN-LABEL: {{^}}i8_signext_func_void: 53; GCN: buffer_load_sbyte v0, off 54; GCN-NEXT: s_waitcnt vmcnt(0) 55; GCN-NEXT: s_setpc_b64 56define signext i8 @i8_signext_func_void() #0 { 57 %val = load i8, i8 addrspace(1)* undef 58 ret i8 %val 59} 60 61; GCN-LABEL: {{^}}i16_func_void: 62; GCN: buffer_load_ushort v0, off 63; GCN-NEXT: s_waitcnt vmcnt(0) 64; GCN-NEXT: s_setpc_b64 65define i16 @i16_func_void() #0 { 66 %val = load i16, i16 addrspace(1)* undef 67 ret i16 %val 68} 69 70; GCN-LABEL: {{^}}i16_zeroext_func_void: 71; GCN: buffer_load_ushort v0, off 72; GCN-NEXT: s_waitcnt vmcnt(0) 73; GCN-NEXT: s_setpc_b64 74define zeroext i16 @i16_zeroext_func_void() #0 { 75 %val = load i16, i16 addrspace(1)* undef 76 ret i16 %val 77} 78 79; GCN-LABEL: {{^}}i16_signext_func_void: 80; GCN: buffer_load_sshort v0, off 81; GCN-NEXT: s_waitcnt vmcnt(0) 82; GCN-NEXT: s_setpc_b64 83define signext i16 @i16_signext_func_void() #0 { 84 %val = load i16, i16 addrspace(1)* undef 85 ret i16 %val 86} 87 88; GCN-LABEL: {{^}}i32_func_void: 89; GCN: buffer_load_dword v0, off 90; GCN-NEXT: s_waitcnt vmcnt(0) 91; GCN-NEXT: s_setpc_b64 92define i32 @i32_func_void() #0 { 93 %val = load i32, i32 addrspace(1)* undef 94 ret i32 %val 95} 96 97; GCN-LABEL: {{^}}i64_func_void: 98; GCN: buffer_load_dwordx2 v[0:1], off 99; GCN-NEXT: s_waitcnt vmcnt(0) 100; GCN-NEXT: s_setpc_b64 101define i64 @i64_func_void() #0 { 102 %val = load i64, i64 addrspace(1)* undef 103 ret i64 %val 104} 105 106; GCN-LABEL: {{^}}f32_func_void: 107; GCN: buffer_load_dword v0, off, s[8:11], 0 108; GCN-NEXT: s_waitcnt vmcnt(0) 109; GCN-NEXT: s_setpc_b64 110define float @f32_func_void() #0 { 111 %val = load float, float addrspace(1)* undef 112 ret float %val 113} 114 115; GCN-LABEL: {{^}}f64_func_void: 116; GCN: buffer_load_dwordx2 v[0:1], off 117; GCN-NEXT: s_waitcnt vmcnt(0) 118; GCN-NEXT: s_setpc_b64 119define double @f64_func_void() #0 { 120 %val = load double, double addrspace(1)* undef 121 ret double %val 122} 123 124; GCN-LABEL: {{^}}v2i32_func_void: 125; GCN: buffer_load_dwordx2 v[0:1], off 126; GCN-NEXT: s_waitcnt vmcnt(0) 127; GCN-NEXT: s_setpc_b64 128define <2 x i32> @v2i32_func_void() #0 { 129 %val = load <2 x i32>, <2 x i32> addrspace(1)* undef 130 ret <2 x i32> %val 131} 132 133; GCN-LABEL: {{^}}v3i32_func_void: 134; GCN: buffer_load_dwordx4 v[0:3], off 135; GCN-NEXT: s_waitcnt vmcnt(0) 136; GCN-NEXT: s_setpc_b64 137define <3 x i32> @v3i32_func_void() #0 { 138 %val = load <3 x i32>, <3 x i32> addrspace(1)* undef 139 ret <3 x i32> %val 140} 141 142; GCN-LABEL: {{^}}v4i32_func_void: 143; GCN: buffer_load_dwordx4 v[0:3], off 144; GCN-NEXT: s_waitcnt vmcnt(0) 145; GCN-NEXT: s_setpc_b64 146define <4 x i32> @v4i32_func_void() #0 { 147 %val = load <4 x i32>, <4 x i32> addrspace(1)* undef 148 ret <4 x i32> %val 149} 150 151; GCN-LABEL: {{^}}v5i32_func_void: 152; GCN-DAG: buffer_load_dword v4, off 153; GCN-DAG: buffer_load_dwordx4 v[0:3], off 154; GCN: s_waitcnt vmcnt(0) 155; GCN-NEXT: s_setpc_b64 156define <5 x i32> @v5i32_func_void() #0 { 157 %val = load volatile <5 x i32>, <5 x i32> addrspace(1)* undef 158 ret <5 x i32> %val 159} 160 161; GCN-LABEL: {{^}}v8i32_func_void: 162; GCN-DAG: buffer_load_dwordx4 v[0:3], off 163; GCN-DAG: buffer_load_dwordx4 v[4:7], off 164; GCN: s_waitcnt vmcnt(0) 165; GCN-NEXT: s_setpc_b64 166define <8 x i32> @v8i32_func_void() #0 { 167 %ptr = load volatile <8 x i32> addrspace(1)*, <8 x i32> addrspace(1)* addrspace(2)* undef 168 %val = load <8 x i32>, <8 x i32> addrspace(1)* %ptr 169 ret <8 x i32> %val 170} 171 172; GCN-LABEL: {{^}}v16i32_func_void: 173; GCN-DAG: buffer_load_dwordx4 v[0:3], off 174; GCN-DAG: buffer_load_dwordx4 v[4:7], off 175; GCN-DAG: buffer_load_dwordx4 v[8:11], off 176; GCN-DAG: buffer_load_dwordx4 v[12:15], off 177; GCN: s_waitcnt vmcnt(0) 178; GCN-NEXT: s_setpc_b64 179define <16 x i32> @v16i32_func_void() #0 { 180 %ptr = load volatile <16 x i32> addrspace(1)*, <16 x i32> addrspace(1)* addrspace(2)* undef 181 %val = load <16 x i32>, <16 x i32> addrspace(1)* %ptr 182 ret <16 x i32> %val 183} 184 185; GCN-LABEL: {{^}}v32i32_func_void: 186; GCN-DAG: buffer_load_dwordx4 v[0:3], off 187; GCN-DAG: buffer_load_dwordx4 v[4:7], off 188; GCN-DAG: buffer_load_dwordx4 v[8:11], off 189; GCN-DAG: buffer_load_dwordx4 v[12:15], off 190; GCN-DAG: buffer_load_dwordx4 v[16:19], off 191; GCN-DAG: buffer_load_dwordx4 v[20:23], off 192; GCN-DAG: buffer_load_dwordx4 v[24:27], off 193; GCN-DAG: buffer_load_dwordx4 v[28:31], off 194; GCN: s_waitcnt vmcnt(0) 195; GCN-NEXT: s_setpc_b64 196define <32 x i32> @v32i32_func_void() #0 { 197 %ptr = load volatile <32 x i32> addrspace(1)*, <32 x i32> addrspace(1)* addrspace(2)* undef 198 %val = load <32 x i32>, <32 x i32> addrspace(1)* %ptr 199 ret <32 x i32> %val 200} 201 202; GCN-LABEL: {{^}}v2i64_func_void: 203; GCN: buffer_load_dwordx4 v[0:3], off 204; GCN-NEXT: s_waitcnt vmcnt(0) 205; GCN-NEXT: s_setpc_b64 206define <2 x i64> @v2i64_func_void() #0 { 207 %val = load <2 x i64>, <2 x i64> addrspace(1)* undef 208 ret <2 x i64> %val 209} 210 211; GCN-LABEL: {{^}}v3i64_func_void: 212; GCN-DAG: buffer_load_dwordx4 v[0:3], off 213; GCN-DAG: buffer_load_dwordx4 v[4:7], off 214; GCN: s_waitcnt vmcnt(0) 215; GCN-NEXT: s_setpc_b64 216define <3 x i64> @v3i64_func_void() #0 { 217 %ptr = load volatile <3 x i64> addrspace(1)*, <3 x i64> addrspace(1)* addrspace(2)* undef 218 %val = load <3 x i64>, <3 x i64> addrspace(1)* %ptr 219 ret <3 x i64> %val 220} 221 222; GCN-LABEL: {{^}}v4i64_func_void: 223; GCN: buffer_load_dwordx4 v[0:3], off 224; GCN: buffer_load_dwordx4 v[4:7], off 225; GCN-NEXT: s_waitcnt vmcnt(0) 226; GCN-NEXT: s_setpc_b64 227define <4 x i64> @v4i64_func_void() #0 { 228 %ptr = load volatile <4 x i64> addrspace(1)*, <4 x i64> addrspace(1)* addrspace(2)* undef 229 %val = load <4 x i64>, <4 x i64> addrspace(1)* %ptr 230 ret <4 x i64> %val 231} 232 233; GCN-LABEL: {{^}}v5i64_func_void: 234; GCN-DAG: buffer_load_dwordx4 v[0:3], off 235; GCN-DAG: buffer_load_dwordx4 v[4:7], off 236; GCN-DAG: buffer_load_dwordx4 v[8:11], off 237; GCN: s_waitcnt vmcnt(0) 238; GCN-NEXT: s_setpc_b64 239define <5 x i64> @v5i64_func_void() #0 { 240 %ptr = load volatile <5 x i64> addrspace(1)*, <5 x i64> addrspace(1)* addrspace(2)* undef 241 %val = load <5 x i64>, <5 x i64> addrspace(1)* %ptr 242 ret <5 x i64> %val 243} 244 245; GCN-LABEL: {{^}}v8i64_func_void: 246; GCN-DAG: buffer_load_dwordx4 v[0:3], off 247; GCN-DAG: buffer_load_dwordx4 v[4:7], off 248; GCN-DAG: buffer_load_dwordx4 v[8:11], off 249; GCN-DAG: buffer_load_dwordx4 v[12:15], off 250; GCN: s_waitcnt vmcnt(0) 251; GCN-NEXT: s_setpc_b64 252define <8 x i64> @v8i64_func_void() #0 { 253 %ptr = load volatile <8 x i64> addrspace(1)*, <8 x i64> addrspace(1)* addrspace(2)* undef 254 %val = load <8 x i64>, <8 x i64> addrspace(1)* %ptr 255 ret <8 x i64> %val 256} 257 258; GCN-LABEL: {{^}}v16i64_func_void: 259; GCN-DAG: buffer_load_dwordx4 v[0:3], off 260; GCN-DAG: buffer_load_dwordx4 v[4:7], off 261; GCN-DAG: buffer_load_dwordx4 v[8:11], off 262; GCN-DAG: buffer_load_dwordx4 v[12:15], off 263; GCN-DAG: buffer_load_dwordx4 v[16:19], off 264; GCN-DAG: buffer_load_dwordx4 v[20:23], off 265; GCN-DAG: buffer_load_dwordx4 v[24:27], off 266; GCN-DAG: buffer_load_dwordx4 v[28:31], off 267; GCN: s_waitcnt vmcnt(0) 268; GCN-NEXT: s_setpc_b64 269define <16 x i64> @v16i64_func_void() #0 { 270 %ptr = load volatile <16 x i64> addrspace(1)*, <16 x i64> addrspace(1)* addrspace(2)* undef 271 %val = load <16 x i64>, <16 x i64> addrspace(1)* %ptr 272 ret <16 x i64> %val 273} 274 275; GCN-LABEL: {{^}}v2i16_func_void: 276; GFX9: buffer_load_dword v0, off 277; GFX9-NEXT: s_waitcnt vmcnt(0) 278; GFX9-NEXT: s_setpc_b64 279define <2 x i16> @v2i16_func_void() #0 { 280 %val = load <2 x i16>, <2 x i16> addrspace(1)* undef 281 ret <2 x i16> %val 282} 283 284; GCN-LABEL: {{^}}v3i16_func_void: 285; GFX9: buffer_load_dwordx2 v[0:1], off 286; GFX9-NEXT: s_waitcnt vmcnt(0) 287; GFX9-NEXT: s_setpc_b64 288define <3 x i16> @v3i16_func_void() #0 { 289 %val = load <3 x i16>, <3 x i16> addrspace(1)* undef 290 ret <3 x i16> %val 291} 292 293; GCN-LABEL: {{^}}v4i16_func_void: 294; GFX9: buffer_load_dwordx2 v[0:1], off 295; GFX9-NEXT: s_waitcnt vmcnt(0) 296; GFX9-NEXT: s_setpc_b64 297define <4 x i16> @v4i16_func_void() #0 { 298 %val = load <4 x i16>, <4 x i16> addrspace(1)* undef 299 ret <4 x i16> %val 300} 301 302; FIXME: Should not scalarize 303; GCN-LABEL: {{^}}v5i16_func_void: 304; GFX9: buffer_load_dwordx2 v[0:1] 305; GFX9: buffer_load_ushort v4 306; GFX9: v_lshrrev_b32_e32 v3, 16, v1 307; GFX9: v_mov_b32_e32 v2, v1 308; GFX9: v_lshrrev_b32_e32 v3, 16, v0 309; GCN: s_setpc_b64 310define <5 x i16> @v5i16_func_void() #0 { 311 %ptr = load volatile <5 x i16> addrspace(1)*, <5 x i16> addrspace(1)* addrspace(2)* undef 312 %val = load <5 x i16>, <5 x i16> addrspace(1)* %ptr 313 ret <5 x i16> %val 314} 315 316; GCN-LABEL: {{^}}v8i16_func_void: 317; GFX9-DAG: buffer_load_dwordx4 v[0:3], off 318; GFX9: s_waitcnt vmcnt(0) 319; GFX9-NEXT: s_setpc_b64 320define <8 x i16> @v8i16_func_void() #0 { 321 %ptr = load volatile <8 x i16> addrspace(1)*, <8 x i16> addrspace(1)* addrspace(2)* undef 322 %val = load <8 x i16>, <8 x i16> addrspace(1)* %ptr 323 ret <8 x i16> %val 324} 325 326; GCN-LABEL: {{^}}v16i16_func_void: 327; GFX9: buffer_load_dwordx4 v[0:3], off 328; GFX9: buffer_load_dwordx4 v[4:7], off 329; GFX9: s_waitcnt vmcnt(0) 330; GFX9-NEXT: s_setpc_b64 331define <16 x i16> @v16i16_func_void() #0 { 332 %ptr = load volatile <16 x i16> addrspace(1)*, <16 x i16> addrspace(1)* addrspace(2)* undef 333 %val = load <16 x i16>, <16 x i16> addrspace(1)* %ptr 334 ret <16 x i16> %val 335} 336 337; FIXME: Should pack 338; GCN-LABEL: {{^}}v16i8_func_void: 339; GCN-DAG: v12 340; GCN-DAG: v13 341; GCN-DAG: v14 342; GCN-DAG: v15 343define <16 x i8> @v16i8_func_void() #0 { 344 %ptr = load volatile <16 x i8> addrspace(1)*, <16 x i8> addrspace(1)* addrspace(2)* undef 345 %val = load <16 x i8>, <16 x i8> addrspace(1)* %ptr 346 ret <16 x i8> %val 347} 348 349; FIXME: Should pack 350; GCN-LABEL: {{^}}v4i8_func_void: 351; GCN: buffer_load_dword v0 352; GCN-DAG: v_lshrrev_b32_e32 v2, 16, v0 353; GCN-DAG: v_lshrrev_b32_e32 v3, 24, v0 354; CI-DAG: v_bfe_u32 v1, v0, 8, 8 355; VI-DAG: v_lshrrev_b16_e32 v1, 8, v0 356; GCN: s_setpc_b64 357define <4 x i8> @v4i8_func_void() #0 { 358 %ptr = load volatile <4 x i8> addrspace(1)*, <4 x i8> addrspace(1)* addrspace(2)* undef 359 %val = load <4 x i8>, <4 x i8> addrspace(1)* %ptr 360 ret <4 x i8> %val 361} 362 363; GCN-LABEL: {{^}}struct_i8_i32_func_void: 364; GCN-DAG: buffer_load_dword v1 365; GCN-DAG: buffer_load_ubyte v0 366; GCN: s_waitcnt vmcnt(0) 367; GCN-NEXT: s_setpc_b64 368define {i8, i32} @struct_i8_i32_func_void() #0 { 369 %val = load { i8, i32 }, { i8, i32 } addrspace(1)* undef 370 ret { i8, i32 } %val 371} 372 373; GCN-LABEL: {{^}}void_func_sret_struct_i8_i32: 374; GCN: buffer_load_ubyte [[VAL0:v[0-9]+]] 375; GCN: buffer_load_dword [[VAL1:v[0-9]+]] 376; GCN: buffer_store_byte [[VAL0]], v0, s[0:3], s4 offen{{$}} 377; GCN: buffer_store_dword [[VAL1]], v0, s[0:3], s4 offen offset:4{{$}} 378define void @void_func_sret_struct_i8_i32({ i8, i32 }* sret %arg0) #0 { 379 %val0 = load volatile i8, i8 addrspace(1)* undef 380 %val1 = load volatile i32, i32 addrspace(1)* undef 381 %gep0 = getelementptr inbounds { i8, i32 }, { i8, i32 }* %arg0, i32 0, i32 0 382 %gep1 = getelementptr inbounds { i8, i32 }, { i8, i32 }* %arg0, i32 0, i32 1 383 store i8 %val0, i8* %gep0 384 store i32 %val1, i32* %gep1 385 ret void 386} 387 388; FIXME: Should be able to fold offsets in all of these. Call lowering 389; introduces an extra CopyToReg/CopyFromReg obscuring the AssertZext 390; inserted. Not using it introduces the spills. 391 392; GCN-LABEL: {{^}}v33i32_func_void: 393; GCN: buffer_store_dword v32, off, s[0:3], s5 offset:12 ; 4-byte Folded Spill 394; GCN: buffer_store_dword v33, off, s[0:3], s5 offset:8 ; 4-byte Folded Spill 395; GCN: buffer_store_dword v34, off, s[0:3], s5 offset:4 ; 4-byte Folded Spill 396 397; GCN-DAG: buffer_store_dword v{{[0-9]+}}, v0, s[0:3], s4 offen{{$}} 398 399; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_4:v[0-9]+]], vcc, 4, v0 400; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_4]], s[0:3], s4 offen{{$}} 401 402; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_8:v[0-9]+]], vcc, 8, v0 403; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_8]], s[0:3], s4 offen{{$}} 404 405; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_12:v[0-9]+]], vcc, 12, v0 406; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_12]], s[0:3], s4 offen{{$}} 407 408; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_16:v[0-9]+]], vcc, 16, v0 409; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_16]], s[0:3], s4 offen{{$}} 410 411; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_20:v[0-9]+]], vcc, 20, v0 412; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_20]], s[0:3], s4 offen{{$}} 413 414; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_24:v[0-9]+]], vcc, 24, v0 415; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_24]], s[0:3], s4 offen{{$}} 416 417; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_28:v[0-9]+]], vcc, 28, v0 418; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_28]], s[0:3], s4 offen{{$}} 419 420; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_32:v[0-9]+]], vcc, 32, v0 421; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_32]], s[0:3], s4 offen{{$}} 422 423; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_36:v[0-9]+]], vcc, 36, v0 424; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_36]], s[0:3], s4 offen{{$}} 425 426; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_40:v[0-9]+]], vcc, 40, v0 427; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_40]], s[0:3], s4 offen{{$}} 428 429; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_44:v[0-9]+]], vcc, 44, v0 430; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_44]], s[0:3], s4 offen{{$}} 431 432; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_48:v[0-9]+]], vcc, 48, v0 433; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_48]], s[0:3], s4 offen{{$}} 434 435; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_52:v[0-9]+]], vcc, 52, v0 436; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_52]], s[0:3], s4 offen{{$}} 437 438; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_56:v[0-9]+]], vcc, 56, v0 439; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_56]], s[0:3], s4 offen{{$}} 440 441; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_60:v[0-9]+]], vcc, 60, v0 442; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_60]], s[0:3], s4 offen{{$}} 443 444; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_64:v[0-9]+]], vcc, 64, v0 445; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_64]], s[0:3], s4 offen{{$}} 446 447; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_68:v[0-9]+]], vcc, 0x44, v0 448; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_68]], s[0:3], s4 offen{{$}} 449 450; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_72:v[0-9]+]], vcc, 0x48, v0 451; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_72]], s[0:3], s4 offen{{$}} 452 453; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_76:v[0-9]+]], vcc, 0x4c, v0 454; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_76]], s[0:3], s4 offen{{$}} 455 456; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_80:v[0-9]+]], vcc, 0x50, v0 457; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_80]], s[0:3], s4 offen{{$}} 458 459; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_84:v[0-9]+]], vcc, 0x54, v0 460; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_84]], s[0:3], s4 offen{{$}} 461 462; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_88:v[0-9]+]], vcc, 0x58, v0 463; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_88]], s[0:3], s4 offen{{$}} 464 465; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_92:v[0-9]+]], vcc, 0x5c, v0 466; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_92]], s[0:3], s4 offen{{$}} 467 468; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_96:v[0-9]+]], vcc, 0x60, v0 469; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_96]], s[0:3], s4 offen{{$}} 470 471; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_100:v[0-9]+]], vcc, 0x64, v0 472; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_100]], s[0:3], s4 offen{{$}} 473 474; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_104:v[0-9]+]], vcc, 0x68, v0 475; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_104]], s[0:3], s4 offen{{$}} 476 477; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_108:v[0-9]+]], vcc, 0x6c, v0 478; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_108]], s[0:3], s4 offen{{$}} 479 480; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_112:v[0-9]+]], vcc, 0x70, v0 481; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_112]], s[0:3], s4 offen{{$}} 482 483; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_116:v[0-9]+]], vcc, 0x74, v0 484; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_116]], s[0:3], s4 offen{{$}} 485 486; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_120:v[0-9]+]], vcc, 0x78, v0 487; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_120]], s[0:3], s4 offen{{$}} 488 489; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_124:v[0-9]+]], vcc, 0x7c, v0 490; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_124]], s[0:3], s4 offen{{$}} 491 492; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_128:v[0-9]+]], vcc, 0x80, v0 493; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_128]], s[0:3], s4 offen{{$}} 494 495; GCN: buffer_load_dword v34 496; GCN: buffer_load_dword v33 497; GCN: buffer_load_dword v32 498; GCN: s_waitcnt vmcnt(0) 499; GCN-NEXT: s_setpc_b64 500define <33 x i32> @v33i32_func_void() #0 { 501 %ptr = load volatile <33 x i32> addrspace(1)*, <33 x i32> addrspace(1)* addrspace(2)* undef 502 %val = load <33 x i32>, <33 x i32> addrspace(1)* %ptr 503 ret <33 x i32> %val 504} 505 506; GCN-LABEL: {{^}}struct_v32i32_i32_func_void: 507; GCN: buffer_store_dword v32, off, s[0:3], s5 offset:12 ; 4-byte Folded Spill 508; GCN: buffer_store_dword v33, off, s[0:3], s5 offset:8 ; 4-byte Folded Spill 509; GCN: buffer_store_dword v34, off, s[0:3], s5 offset:4 ; 4-byte Folded Spill 510 511; GCN-DAG: buffer_store_dword v{{[0-9]+}}, v0, s[0:3], s4 offen{{$}} 512 513; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_4:v[0-9]+]], vcc, 4, v0 514; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_4]], s[0:3], s4 offen{{$}} 515 516; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_8:v[0-9]+]], vcc, 8, v0 517; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_8]], s[0:3], s4 offen{{$}} 518 519; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_12:v[0-9]+]], vcc, 12, v0 520; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_12]], s[0:3], s4 offen{{$}} 521 522; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_16:v[0-9]+]], vcc, 16, v0 523; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_16]], s[0:3], s4 offen{{$}} 524 525; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_20:v[0-9]+]], vcc, 20, v0 526; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_20]], s[0:3], s4 offen{{$}} 527 528; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_24:v[0-9]+]], vcc, 24, v0 529; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_24]], s[0:3], s4 offen{{$}} 530 531; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_28:v[0-9]+]], vcc, 28, v0 532; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_28]], s[0:3], s4 offen{{$}} 533 534; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_32:v[0-9]+]], vcc, 32, v0 535; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_32]], s[0:3], s4 offen{{$}} 536 537; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_36:v[0-9]+]], vcc, 36, v0 538; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_36]], s[0:3], s4 offen{{$}} 539 540; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_40:v[0-9]+]], vcc, 40, v0 541; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_40]], s[0:3], s4 offen{{$}} 542 543; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_44:v[0-9]+]], vcc, 44, v0 544; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_44]], s[0:3], s4 offen{{$}} 545 546; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_48:v[0-9]+]], vcc, 48, v0 547; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_48]], s[0:3], s4 offen{{$}} 548 549; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_52:v[0-9]+]], vcc, 52, v0 550; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_52]], s[0:3], s4 offen{{$}} 551 552; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_56:v[0-9]+]], vcc, 56, v0 553; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_56]], s[0:3], s4 offen{{$}} 554 555; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_60:v[0-9]+]], vcc, 60, v0 556; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_60]], s[0:3], s4 offen{{$}} 557 558; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_64:v[0-9]+]], vcc, 64, v0 559; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_64]], s[0:3], s4 offen{{$}} 560 561; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_68:v[0-9]+]], vcc, 0x44, v0 562; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_68]], s[0:3], s4 offen{{$}} 563 564; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_72:v[0-9]+]], vcc, 0x48, v0 565; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_72]], s[0:3], s4 offen{{$}} 566 567; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_76:v[0-9]+]], vcc, 0x4c, v0 568; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_76]], s[0:3], s4 offen{{$}} 569 570; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_80:v[0-9]+]], vcc, 0x50, v0 571; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_80]], s[0:3], s4 offen{{$}} 572 573; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_84:v[0-9]+]], vcc, 0x54, v0 574; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_84]], s[0:3], s4 offen{{$}} 575 576; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_88:v[0-9]+]], vcc, 0x58, v0 577; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_88]], s[0:3], s4 offen{{$}} 578 579; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_92:v[0-9]+]], vcc, 0x5c, v0 580; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_92]], s[0:3], s4 offen{{$}} 581 582; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_96:v[0-9]+]], vcc, 0x60, v0 583; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_96]], s[0:3], s4 offen{{$}} 584 585; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_100:v[0-9]+]], vcc, 0x64, v0 586; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_100]], s[0:3], s4 offen{{$}} 587 588; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_104:v[0-9]+]], vcc, 0x68, v0 589; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_104]], s[0:3], s4 offen{{$}} 590 591; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_108:v[0-9]+]], vcc, 0x6c, v0 592; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_108]], s[0:3], s4 offen{{$}} 593 594; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_112:v[0-9]+]], vcc, 0x70, v0 595; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_112]], s[0:3], s4 offen{{$}} 596 597; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_116:v[0-9]+]], vcc, 0x74, v0 598; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_116]], s[0:3], s4 offen{{$}} 599 600; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_120:v[0-9]+]], vcc, 0x78, v0 601; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_120]], s[0:3], s4 offen{{$}} 602 603; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_124:v[0-9]+]], vcc, 0x7c, v0 604; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_124]], s[0:3], s4 offen{{$}} 605 606; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_128:v[0-9]+]], vcc, 0x80, v0 607; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_128]], s[0:3], s4 offen{{$}} 608 609; GCN: buffer_load_dword v34 610; GCN: buffer_load_dword v33 611; GCN: buffer_load_dword v32 612; GCN: s_waitcnt vmcnt(0) 613; GCN-NEXT: s_setpc_b64 614define { <32 x i32>, i32 } @struct_v32i32_i32_func_void() #0 { 615 %ptr = load volatile { <32 x i32>, i32 } addrspace(1)*, { <32 x i32>, i32 } addrspace(1)* addrspace(2)* undef 616 %val = load { <32 x i32>, i32 }, { <32 x i32>, i32 } addrspace(1)* %ptr 617 ret { <32 x i32>, i32 }%val 618} 619 620; GCN-LABEL: {{^}}struct_i32_v32i32_func_void: 621; GCN: buffer_store_dword v32, off, s[0:3], s5 offset:8 ; 4-byte Folded Spill 622; GCN: buffer_store_dword v33, off, s[0:3], s5 offset:4 ; 4-byte Folded Spill 623 624; GCN-DAG: buffer_store_dword v{{[0-9]+}}, v0, s[0:3], s4 offen{{$}} 625 626; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_128:v[0-9]+]], vcc, 0x80, v0 627; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_128]], s[0:3], s4 offen{{$}} 628 629 630; GCN-DAG: v_add_{{[_coiu]*}}32_e32 [[ADD_256:v[0-9]+]], vcc, 0xfc, v0 631; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ADD_256]], s[0:3], s4 offen{{$}} 632 633; GCN: buffer_load_dword v33 634; GCN: buffer_load_dword v32 635; GCN: s_waitcnt vmcnt(0) 636; GCN-NEXT: s_setpc_b64 637define { i32, <32 x i32> } @struct_i32_v32i32_func_void() #0 { 638 %ptr = load volatile { i32, <32 x i32> } addrspace(1)*, { i32, <32 x i32> } addrspace(1)* addrspace(2)* undef 639 %val = load { i32, <32 x i32> }, { i32, <32 x i32> } addrspace(1)* %ptr 640 ret { i32, <32 x i32> }%val 641} 642 643attributes #0 = { nounwind } 644