1; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck %s 2; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck %s 3 4; Tests for indirect addressing on SI, which is implemented using dynamic 5; indexing of vectors. 6 7; CHECK-LABEL: {{^}}extract_w_offset: 8; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 4.0 9; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 0x40400000 10; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 2.0 11; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 1.0 12; CHECK: s_mov_b32 m0 13; CHECK-NEXT: v_movrels_b32_e32 14define void @extract_w_offset(float addrspace(1)* %out, i32 %in) { 15entry: 16 %idx = add i32 %in, 1 17 %elt = extractelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, i32 %idx 18 store float %elt, float addrspace(1)* %out 19 ret void 20} 21 22; XXX: Could do v_or_b32 directly 23; CHECK-LABEL: {{^}}extract_w_offset_salu_use_vector: 24; CHECK-DAG: s_or_b32 25; CHECK-DAG: s_or_b32 26; CHECK-DAG: s_or_b32 27; CHECK-DAG: s_or_b32 28; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}} 29; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}} 30; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}} 31; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}} 32; CHECK: s_mov_b32 m0 33; CHECK-NEXT: v_movrels_b32_e32 34define void @extract_w_offset_salu_use_vector(i32 addrspace(1)* %out, i32 %in, <4 x i32> %or.val) { 35entry: 36 %idx = add i32 %in, 1 37 %vec = or <4 x i32> %or.val, <i32 1, i32 2, i32 3, i32 4> 38 %elt = extractelement <4 x i32> %vec, i32 %idx 39 store i32 %elt, i32 addrspace(1)* %out 40 ret void 41} 42 43; CHECK-LABEL: {{^}}extract_wo_offset: 44; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 4.0 45; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 0x40400000 46; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 2.0 47; CHECK-DAG: v_mov_b32_e32 v{{[0-9]+}}, 1.0 48; CHECK: s_mov_b32 m0 49; CHECK-NEXT: v_movrels_b32_e32 50define void @extract_wo_offset(float addrspace(1)* %out, i32 %in) { 51entry: 52 %elt = extractelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, i32 %in 53 store float %elt, float addrspace(1)* %out 54 ret void 55} 56 57; CHECK-LABEL: {{^}}extract_neg_offset_sgpr: 58; The offset depends on the register that holds the first element of the vector. 59; CHECK: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}} 60; CHECK: v_movrels_b32_e32 v{{[0-9]}}, v0 61define void @extract_neg_offset_sgpr(i32 addrspace(1)* %out, i32 %offset) { 62entry: 63 %index = add i32 %offset, -512 64 %value = extractelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 %index 65 store i32 %value, i32 addrspace(1)* %out 66 ret void 67} 68 69; CHECK-LABEL: {{^}}extract_neg_offset_sgpr_loaded: 70; The offset depends on the register that holds the first element of the vector. 71; CHECK: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}} 72; CHECK: v_movrels_b32_e32 v{{[0-9]}}, v0 73define void @extract_neg_offset_sgpr_loaded(i32 addrspace(1)* %out, <4 x i32> %vec0, <4 x i32> %vec1, i32 %offset) { 74entry: 75 %index = add i32 %offset, -512 76 %or = or <4 x i32> %vec0, %vec1 77 %value = extractelement <4 x i32> %or, i32 %index 78 store i32 %value, i32 addrspace(1)* %out 79 ret void 80} 81 82; CHECK-LABEL: {{^}}extract_neg_offset_vgpr: 83; The offset depends on the register that holds the first element of the vector. 84; CHECK: v_readfirstlane_b32 85; CHECK: s_add_i32 m0, m0, 0xfffffe{{[0-9a-z]+}} 86; CHECK-NEXT: v_movrels_b32_e32 v{{[0-9]}}, v0 87; CHECK: s_cbranch_execnz 88define void @extract_neg_offset_vgpr(i32 addrspace(1)* %out) { 89entry: 90 %id = call i32 @llvm.amdgcn.workitem.id.x() #1 91 %index = add i32 %id, -512 92 %value = extractelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 %index 93 store i32 %value, i32 addrspace(1)* %out 94 ret void 95} 96 97; CHECK-LABEL: {{^}}extract_undef_offset_sgpr: 98define void @extract_undef_offset_sgpr(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in) { 99entry: 100 %ld = load volatile <4 x i32>, <4 x i32> addrspace(1)* %in 101 %value = extractelement <4 x i32> %ld, i32 undef 102 store i32 %value, i32 addrspace(1)* %out 103 ret void 104} 105 106; CHECK-LABEL: {{^}}insert_undef_offset_sgpr_vector_src: 107; CHECK: buffer_load_dwordx4 108; CHECK: s_mov_b32 m0, 109; CHECK-NEXT: v_movreld_b32 110define void @insert_undef_offset_sgpr_vector_src(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) { 111entry: 112 %ld = load <4 x i32>, <4 x i32> addrspace(1)* %in 113 %value = insertelement <4 x i32> %ld, i32 5, i32 undef 114 store <4 x i32> %value, <4 x i32> addrspace(1)* %out 115 ret void 116} 117 118; CHECK-LABEL: {{^}}insert_w_offset: 119; CHECK: s_mov_b32 m0 120; CHECK-NEXT: v_movreld_b32_e32 121define void @insert_w_offset(float addrspace(1)* %out, i32 %in) { 122entry: 123 %0 = add i32 %in, 1 124 %1 = insertelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, float 5.0, i32 %0 125 %2 = extractelement <4 x float> %1, i32 2 126 store float %2, float addrspace(1)* %out 127 ret void 128} 129 130; CHECK-LABEL: {{^}}insert_wo_offset: 131; CHECK: s_mov_b32 m0 132; CHECK-NEXT: v_movreld_b32_e32 133define void @insert_wo_offset(float addrspace(1)* %out, i32 %in) { 134entry: 135 %0 = insertelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, float 5.0, i32 %in 136 %1 = extractelement <4 x float> %0, i32 2 137 store float %1, float addrspace(1)* %out 138 ret void 139} 140 141; CHECK-LABEL: {{^}}insert_neg_offset_sgpr: 142; The offset depends on the register that holds the first element of the vector. 143; CHECK: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}} 144; CHECK: v_movreld_b32_e32 v0, v{{[0-9]}} 145define void @insert_neg_offset_sgpr(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out, i32 %offset) { 146entry: 147 %index = add i32 %offset, -512 148 %value = insertelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 5, i32 %index 149 store <4 x i32> %value, <4 x i32> addrspace(1)* %out 150 ret void 151} 152 153; The vector indexed into is originally loaded into an SGPR rather 154; than built with a reg_sequence 155 156; CHECK-LABEL: {{^}}insert_neg_offset_sgpr_loadreg: 157; The offset depends on the register that holds the first element of the vector. 158; CHECK: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}} 159; CHECK: v_movreld_b32_e32 v0, v{{[0-9]}} 160define void @insert_neg_offset_sgpr_loadreg(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out, <4 x i32> %vec, i32 %offset) { 161entry: 162 %index = add i32 %offset, -512 163 %value = insertelement <4 x i32> %vec, i32 5, i32 %index 164 store <4 x i32> %value, <4 x i32> addrspace(1)* %out 165 ret void 166} 167 168; CHECK-LABEL: {{^}}insert_neg_offset_vgpr: 169; The offset depends on the register that holds the first element of the vector. 170; CHECK: v_readfirstlane_b32 171; CHECK: s_add_i32 m0, m0, 0xfffffe{{[0-9a-z]+}} 172; CHECK-NEXT: v_movreld_b32_e32 v0, v{{[0-9]}} 173; CHECK: s_cbranch_execnz 174define void @insert_neg_offset_vgpr(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out) { 175entry: 176 %id = call i32 @llvm.amdgcn.workitem.id.x() #1 177 %index = add i32 %id, -512 178 %value = insertelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 5, i32 %index 179 store <4 x i32> %value, <4 x i32> addrspace(1)* %out 180 ret void 181} 182 183; CHECK-LABEL: {{^}}insert_neg_inline_offset_vgpr: 184; The offset depends on the register that holds the first element of the vector. 185; CHECK: v_readfirstlane_b32 186; CHECK: s_add_i32 m0, m0, -{{[0-9]+}} 187; CHECK-NEXT: v_movreld_b32_e32 v0, v{{[0-9]}} 188; CHECK: s_cbranch_execnz 189define void @insert_neg_inline_offset_vgpr(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out) { 190entry: 191 %id = call i32 @llvm.amdgcn.workitem.id.x() #1 192 %index = add i32 %id, -16 193 %value = insertelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 5, i32 %index 194 store <4 x i32> %value, <4 x i32> addrspace(1)* %out 195 ret void 196} 197 198; When the block is split to insert the loop, make sure any other 199; places that need to be expanded in the same block are also handled. 200 201; CHECK-LABEL: {{^}}extract_vgpr_offset_multiple_in_block: 202 203; CHECK-DAG: {{buffer|flat}}_load_dword [[IDX0:v[0-9]+]] 204; CHECK-DAG: s_mov_b32 [[S_ELT0:s[0-9]+]], 7 205; CHECK-DAG: s_mov_b32 [[S_ELT1:s[0-9]+]], 9 206; CHECK-DAG: v_mov_b32_e32 [[VEC_ELT0:v[0-9]+]], [[S_ELT0]] 207; CHECK-DAG: v_mov_b32_e32 [[VEC_ELT1:v[0-9]+]], [[S_ELT1]] 208; CHECK: s_waitcnt vmcnt(0) 209 210; CHECK: s_mov_b64 [[MASK:s\[[0-9]+:[0-9]+\]]], exec 211 212; CHECK: [[LOOP0:BB[0-9]+_[0-9]+]]: 213; CHECK: v_readfirstlane_b32 vcc_lo, [[IDX0]] 214; CHECK: s_mov_b32 m0, vcc_lo 215; CHECK: v_cmp_eq_u32_e32 vcc, m0, [[IDX0]] 216; CHECK: s_and_saveexec_b64 vcc, vcc 217; CHECK-NEXT: v_movrels_b32_e32 [[MOVREL0:v[0-9]+]], [[VEC_ELT0]] 218; CHECK-NEXT: s_xor_b64 exec, exec, vcc 219; CHECK: s_cbranch_execnz [[LOOP0]] 220 221; FIXME: Redundant copy 222; CHECK: s_mov_b64 exec, [[MASK]] 223; CHECK: s_mov_b64 [[MASK2:s\[[0-9]+:[0-9]+\]]], exec 224 225; CHECK: [[LOOP1:BB[0-9]+_[0-9]+]]: 226; CHECK: v_readfirstlane_b32 vcc_lo, [[IDX0]] 227; CHECK: s_mov_b32 m0, vcc_lo 228; CHECK: v_cmp_eq_u32_e32 vcc, m0, [[IDX0]] 229; CHECK: s_and_saveexec_b64 vcc, vcc 230; CHECK-NEXT: v_movrels_b32_e32 [[MOVREL1:v[0-9]+]], [[VEC_ELT1]] 231; CHECK-NEXT: s_xor_b64 exec, exec, vcc 232; CHECK: s_cbranch_execnz [[LOOP1]] 233 234; CHECK: buffer_store_dword [[MOVREL0]] 235; CHECK: buffer_store_dword [[MOVREL1]] 236define void @extract_vgpr_offset_multiple_in_block(i32 addrspace(1)* %out0, i32 addrspace(1)* %out1, i32 addrspace(1)* %in) #0 { 237entry: 238 %id = call i32 @llvm.amdgcn.workitem.id.x() #1 239 %id.ext = zext i32 %id to i64 240 %gep = getelementptr inbounds i32, i32 addrspace(1)* %in, i64 %id.ext 241 %idx0 = load volatile i32, i32 addrspace(1)* %gep 242 %idx1 = add i32 %idx0, 1 243 %val0 = extractelement <4 x i32> <i32 7, i32 9, i32 11, i32 13>, i32 %idx0 244 %live.out.reg = call i32 asm sideeffect "s_mov_b32 $0, 17", "={SGPR4}" () 245 %val1 = extractelement <4 x i32> <i32 7, i32 9, i32 11, i32 13>, i32 %idx1 246 store volatile i32 %val0, i32 addrspace(1)* %out0 247 store volatile i32 %val1, i32 addrspace(1)* %out0 248 %cmp = icmp eq i32 %id, 0 249 br i1 %cmp, label %bb1, label %bb2 250 251bb1: 252 store volatile i32 %live.out.reg, i32 addrspace(1)* undef 253 br label %bb2 254 255bb2: 256 ret void 257} 258 259; CHECK-LABEL: {{^}}insert_vgpr_offset_multiple_in_block: 260; CHECK-DAG: s_load_dwordx4 s{{\[}}[[S_ELT0:[0-9]+]]:[[S_ELT3:[0-9]+]]{{\]}} 261; CHECK-DAG: {{buffer|flat}}_load_dword [[IDX0:v[0-9]+]] 262; CHECK-DAG: v_mov_b32_e32 [[VEC_ELT0:v[0-9]+]], s[[S_ELT0]] 263; CHECK-DAG: v_mov_b32 [[INS0:v[0-9]+]], 62 264; CHECK-DAG: s_waitcnt vmcnt(0) 265 266; CHECK: s_mov_b64 [[MASK:s\[[0-9]+:[0-9]+\]]], exec 267 268; CHECK: [[LOOP0:BB[0-9]+_[0-9]+]]: 269; CHECK: v_readfirstlane_b32 vcc_lo, [[IDX0]] 270; CHECK: s_mov_b32 m0, vcc_lo 271; CHECK: v_cmp_eq_u32_e32 vcc, m0, [[IDX0]] 272; CHECK: s_and_saveexec_b64 vcc, vcc 273; CHECK-NEXT: v_movreld_b32_e32 v[[MOVREL0:[0-9]+]], [[INS0]] 274; CHECK-NEXT: s_xor_b64 exec, exec, vcc 275; CHECK: s_cbranch_execnz [[LOOP0]] 276 277; FIXME: Redundant copy 278; CHECK: s_mov_b64 exec, [[MASK]] 279; CHECK: v_mov_b32_e32 [[INS1:v[0-9]+]], 63 280; CHECK: s_mov_b64 [[MASK]], exec 281 282; CHECK: [[LOOP1:BB[0-9]+_[0-9]+]]: 283; CHECK: v_readfirstlane_b32 vcc_lo, [[IDX0]] 284; CHECK: s_mov_b32 m0, vcc_lo 285; CHECK: v_cmp_eq_u32_e32 vcc, m0, [[IDX0]] 286; CHECK: s_and_saveexec_b64 vcc, vcc 287; CHECK-NEXT: v_movreld_b32_e32 v[[MOVREL1:[0-9]+]], [[INS1]] 288; CHECK-NEXT: s_xor_b64 exec, exec, vcc 289; CHECK: s_cbranch_execnz [[LOOP1]] 290 291; CHECK: buffer_store_dwordx4 v{{\[}}[[MOVREL0]]: 292 293; CHECK: buffer_store_dword [[INS0]] 294define void @insert_vgpr_offset_multiple_in_block(<4 x i32> addrspace(1)* %out0, <4 x i32> addrspace(1)* %out1, i32 addrspace(1)* %in, <4 x i32> %vec0) #0 { 295entry: 296 %id = call i32 @llvm.amdgcn.workitem.id.x() #1 297 %id.ext = zext i32 %id to i64 298 %gep = getelementptr inbounds i32, i32 addrspace(1)* %in, i64 %id.ext 299 %idx0 = load volatile i32, i32 addrspace(1)* %gep 300 %idx1 = add i32 %idx0, 1 301 %live.out.val = call i32 asm sideeffect "v_mov_b32 $0, 62", "=v"() 302 %vec1 = insertelement <4 x i32> %vec0, i32 %live.out.val, i32 %idx0 303 %vec2 = insertelement <4 x i32> %vec1, i32 63, i32 %idx1 304 store volatile <4 x i32> %vec2, <4 x i32> addrspace(1)* %out0 305 %cmp = icmp eq i32 %id, 0 306 br i1 %cmp, label %bb1, label %bb2 307 308bb1: 309 store volatile i32 %live.out.val, i32 addrspace(1)* undef 310 br label %bb2 311 312bb2: 313 ret void 314} 315 316; CHECK-LABEL: {{^}}extract_adjacent_blocks: 317; CHECK: s_load_dword [[ARG:s[0-9]+]] 318; CHECK: s_cmp_lg_i32 319; CHECK: s_cbranch_scc0 [[BB4:BB[0-9]+_[0-9]+]] 320 321; CHECK: buffer_load_dwordx4 322; CHECK: s_mov_b32 m0, 323; CHECK: v_movrels_b32_e32 324; CHECK: s_branch [[ENDBB:BB[0-9]+_[0-9]+]] 325 326; CHECK: [[BB4]]: 327; CHECK: buffer_load_dwordx4 328; CHECK: s_mov_b32 m0, 329; CHECK: v_movrels_b32_e32 330 331; CHECK: [[ENDBB]]: 332; CHECK: buffer_store_dword 333; CHECK: s_endpgm 334define void @extract_adjacent_blocks(i32 %arg) #0 { 335bb: 336 %tmp = icmp eq i32 %arg, 0 337 br i1 %tmp, label %bb1, label %bb4 338 339bb1: 340 %tmp2 = load volatile <4 x float>, <4 x float> addrspace(1)* undef 341 %tmp3 = extractelement <4 x float> %tmp2, i32 undef 342 br label %bb7 343 344bb4: 345 %tmp5 = load volatile <4 x float>, <4 x float> addrspace(1)* undef 346 %tmp6 = extractelement <4 x float> %tmp5, i32 undef 347 br label %bb7 348 349bb7: 350 %tmp8 = phi float [ %tmp3, %bb1 ], [ %tmp6, %bb4 ] 351 store volatile float %tmp8, float addrspace(1)* undef 352 ret void 353} 354 355; CHECK-LABEL: {{^}}insert_adjacent_blocks: 356; CHECK: s_load_dword [[ARG:s[0-9]+]] 357; CHECK: s_cmp_lg_i32 358; CHECK: s_cbranch_scc0 [[BB4:BB[0-9]+_[0-9]+]] 359 360; CHECK: buffer_load_dwordx4 361; CHECK: s_mov_b32 m0, 362; CHECK: v_movreld_b32_e32 363; CHECK: s_branch [[ENDBB:BB[0-9]+_[0-9]+]] 364 365; CHECK: [[BB4]]: 366; CHECK: buffer_load_dwordx4 367; CHECK: s_mov_b32 m0, 368; CHECK: v_movreld_b32_e32 369 370; CHECK: [[ENDBB]]: 371; CHECK: buffer_store_dword 372; CHECK: s_endpgm 373define void @insert_adjacent_blocks(i32 %arg, float %val0) #0 { 374bb: 375 %tmp = icmp eq i32 %arg, 0 376 br i1 %tmp, label %bb1, label %bb4 377 378bb1: ; preds = %bb 379 %tmp2 = load volatile <4 x float>, <4 x float> addrspace(1)* undef 380 %tmp3 = insertelement <4 x float> %tmp2, float %val0, i32 undef 381 br label %bb7 382 383bb4: ; preds = %bb 384 %tmp5 = load volatile <4 x float>, <4 x float> addrspace(1)* undef 385 %tmp6 = insertelement <4 x float> %tmp5, float %val0, i32 undef 386 br label %bb7 387 388bb7: ; preds = %bb4, %bb1 389 %tmp8 = phi <4 x float> [ %tmp3, %bb1 ], [ %tmp6, %bb4 ] 390 store volatile <4 x float> %tmp8, <4 x float> addrspace(1)* undef 391 ret void 392} 393 394; FIXME: Should be able to fold zero input to movreld to inline imm? 395 396; CHECK-LABEL: {{^}}multi_same_block: 397; CHECK: s_load_dword [[ARG:s[0-9]+]] 398; CHECK-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}} 399; CHECK-DAG: s_add_i32 m0, [[ARG]], -16 400; CHECK: v_movreld_b32_e32 v{{[0-9]+}}, [[ZERO]] 401 402; CHECK: s_add_i32 m0, [[ARG]], -14 403; CHECK: v_movreld_b32_e32 v{{[0-9]+}}, v{{[0-9]+}} 404 405; CHECK: s_mov_b32 m0, -1 406; CHECK: ds_write_b32 407; CHECK: ds_write_b32 408; CHECK: s_endpgm 409define void @multi_same_block(i32 %arg) #0 { 410bb: 411 %tmp1 = add i32 %arg, -16 412 %tmp2 = insertelement <6 x float> <float 1.700000e+01, float 1.800000e+01, float 1.900000e+01, float 2.000000e+01, float 2.100000e+01, float 2.200000e+01>, float 0.000000e+00, i32 %tmp1 413 %tmp3 = add i32 %arg, -16 414 %tmp4 = insertelement <6 x float> <float 0x40311999A0000000, float 0x40321999A0000000, float 0x40331999A0000000, float 0x40341999A0000000, float 0x40351999A0000000, float 0x40361999A0000000>, float 0x3FB99999A0000000, i32 %tmp3 415 %tmp5 = bitcast <6 x float> %tmp2 to <6 x i32> 416 %tmp6 = extractelement <6 x i32> %tmp5, i32 1 417 %tmp7 = bitcast <6 x float> %tmp4 to <6 x i32> 418 %tmp8 = extractelement <6 x i32> %tmp7, i32 5 419 store volatile i32 %tmp6, i32 addrspace(3)* undef, align 4 420 store volatile i32 %tmp8, i32 addrspace(3)* undef, align 4 421 ret void 422} 423 424; offset puts outside of superegister bounaries, so clamp to 1st element. 425; CHECK-LABEL: {{^}}extract_largest_inbounds_offset: 426; CHECK: buffer_load_dwordx4 v{{\[}}[[LO_ELT:[0-9]+]]:[[HI_ELT:[0-9]+]]{{\]}} 427; CHECK: s_load_dword [[IDX:s[0-9]+]] 428; CHECK: s_mov_b32 m0, [[IDX]] 429; CHECK-NEXT: v_movrels_b32_e32 [[EXTRACT:v[0-9]+]], v[[HI_ELT]] 430; CHECK: buffer_store_dword [[EXTRACT]] 431define void @extract_largest_inbounds_offset(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %idx) { 432entry: 433 %ld = load volatile <4 x i32>, <4 x i32> addrspace(1)* %in 434 %offset = add i32 %idx, 3 435 %value = extractelement <4 x i32> %ld, i32 %offset 436 store i32 %value, i32 addrspace(1)* %out 437 ret void 438} 439 440; CHECK-LABL: {{^}}extract_out_of_bounds_offset: 441; CHECK: buffer_load_dwordx4 v{{\[}}[[LO_ELT:[0-9]+]]:[[HI_ELT:[0-9]+]]{{\]}} 442; CHECK: s_load_dword [[IDX:s[0-9]+]] 443; CHECK: s_add_i32 m0, [[IDX]], 4 444; CHECK-NEXT: v_movrels_b32_e32 [[EXTRACT:v[0-9]+]], v[[LO_ELT]] 445; CHECK: buffer_store_dword [[EXTRACT]] 446define void @extract_out_of_bounds_offset(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %idx) { 447entry: 448 %ld = load volatile <4 x i32>, <4 x i32> addrspace(1)* %in 449 %offset = add i32 %idx, 4 450 %value = extractelement <4 x i32> %ld, i32 %offset 451 store i32 %value, i32 addrspace(1)* %out 452 ret void 453} 454 455; Test that the or is folded into the base address register instead of 456; added to m0 457 458; GCN-LABEL: {{^}}extractelement_v4i32_or_index: 459; GCN: s_load_dword [[IDX_IN:s[0-9]+]] 460; GCN: s_lshl_b32 [[IDX_SHL:s[0-9]+]], [[IDX_IN]] 461; GCN-NOT: [[IDX_SHL]] 462; GCN: s_mov_b32 m0, [[IDX_SHL]] 463; GCN: v_movreld_b32_e32 v{{[0-9]+}}, v{{[0-9]+}} 464define void @extractelement_v4i32_or_index(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %idx.in) { 465entry: 466 %ld = load volatile <4 x i32>, <4 x i32> addrspace(1)* %in 467 %idx.shl = shl i32 %idx.in, 2 468 %idx = or i32 %idx.shl, 1 469 %value = extractelement <4 x i32> %ld, i32 %idx 470 store i32 %value, i32 addrspace(1)* %out 471 ret void 472} 473 474; GCN-LABEL: {{^}}insertelement_v4f32_or_index: 475; GCN: s_load_dword [[IDX_IN:s[0-9]+]] 476; GCN: s_lshl_b32 [[IDX_SHL:s[0-9]+]], [[IDX_IN]] 477; GCN-NOT: [[IDX_SHL]] 478; GCN: s_mov_b32 m0, [[IDX_SHL]] 479; GCN: v_movreld_b32_e32 v{{[0-9]+}}, v{{[0-9]+}} 480define void @insertelement_v4f32_or_index(<4 x float> addrspace(1)* %out, <4 x float> %a, i32 %idx.in) nounwind { 481 %idx.shl = shl i32 %idx.in, 2 482 %idx = or i32 %idx.shl, 1 483 %vecins = insertelement <4 x float> %a, float 5.000000e+00, i32 %idx 484 store <4 x float> %vecins, <4 x float> addrspace(1)* %out, align 16 485 ret void 486} 487 488declare i32 @llvm.amdgcn.workitem.id.x() #1 489 490attributes #0 = { nounwind } 491attributes #1 = { nounwind readnone } 492