1; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=SI %s 2; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=CI %s 3; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=CI --check-prefix=GCN-HSA %s 4 5declare i32 @llvm.amdgcn.workitem.id.x() #0 6declare i32 @llvm.amdgcn.workitem.id.y() #0 7 8; In this test both the pointer and the offset operands to the 9; BUFFER_LOAD instructions end up being stored in vgprs. This 10; requires us to add the pointer and offset together, store the 11; result in the offset operand (vaddr), and then store 0 in an 12; sgpr register pair and use that for the pointer operand 13; (low 64-bits of srsrc). 14 15; GCN-LABEL: {{^}}mubuf: 16 17; Make sure we aren't using VGPRs for the source operand of s_mov_b64 18; GCN-NOT: s_mov_b64 s[{{[0-9]+:[0-9]+}}], v 19 20; Make sure we aren't using VGPR's for the srsrc operand of BUFFER_LOAD_* 21; instructions 22; GCN-NOHSA: buffer_load_ubyte v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], 0 addr64 23; GCN-NOHSA: buffer_load_ubyte v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], 0 addr64 24; GCN-HSA: flat_load_ubyte v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}} 25; GCN-HSA: flat_load_ubyte v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}} 26 27define void @mubuf(i32 addrspace(1)* %out, i8 addrspace(1)* %in) #1 { 28entry: 29 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 30 %tmp1 = call i32 @llvm.amdgcn.workitem.id.y() 31 %tmp2 = sext i32 %tmp to i64 32 %tmp3 = sext i32 %tmp1 to i64 33 br label %loop 34 35loop: ; preds = %loop, %entry 36 %tmp4 = phi i64 [ 0, %entry ], [ %tmp5, %loop ] 37 %tmp5 = add i64 %tmp2, %tmp4 38 %tmp6 = getelementptr i8, i8 addrspace(1)* %in, i64 %tmp5 39 %tmp7 = load i8, i8 addrspace(1)* %tmp6, align 1 40 %tmp8 = or i64 %tmp5, 1 41 %tmp9 = getelementptr i8, i8 addrspace(1)* %in, i64 %tmp8 42 %tmp10 = load i8, i8 addrspace(1)* %tmp9, align 1 43 %tmp11 = add i8 %tmp7, %tmp10 44 %tmp12 = sext i8 %tmp11 to i32 45 store i32 %tmp12, i32 addrspace(1)* %out 46 %tmp13 = icmp slt i64 %tmp5, 10 47 br i1 %tmp13, label %loop, label %done 48 49done: ; preds = %loop 50 ret void 51} 52 53; Test moving an SMRD instruction to the VALU 54; FIXME: movs can be moved before nop to reduce count 55 56; GCN-LABEL: {{^}}smrd_valu: 57; SI: s_movk_i32 [[OFFSET:s[0-9]+]], 0x2ee0 58; GCN: v_readfirstlane_b32 s[[PTR_LO:[0-9]+]], v{{[0-9]+}} 59; GCN: v_readfirstlane_b32 s[[PTR_HI:[0-9]+]], v{{[0-9]+}} 60; SI: s_nop 3 61; SI: s_load_dword [[OUT:s[0-9]+]], s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, [[OFFSET]] 62; SI: s_mov_b32 63 64; CI: s_load_dword [[OUT:s[0-9]+]], s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0xbb8 65; GCN: v_mov_b32_e32 [[V_OUT:v[0-9]+]], [[OUT]] 66; GCN-NOHSA: buffer_store_dword [[V_OUT]] 67; GCN-HSA: flat_store_dword {{.*}}, [[V_OUT]] 68define void @smrd_valu(i32 addrspace(2)* addrspace(1)* %in, i32 %a, i32 %b, i32 addrspace(1)* %out) #1 { 69entry: 70 %tmp = icmp ne i32 %a, 0 71 br i1 %tmp, label %if, label %else 72 73if: ; preds = %entry 74 %tmp1 = load i32 addrspace(2)*, i32 addrspace(2)* addrspace(1)* %in 75 br label %endif 76 77else: ; preds = %entry 78 %tmp2 = getelementptr i32 addrspace(2)*, i32 addrspace(2)* addrspace(1)* %in 79 %tmp3 = load i32 addrspace(2)*, i32 addrspace(2)* addrspace(1)* %tmp2 80 br label %endif 81 82endif: ; preds = %else, %if 83 %tmp4 = phi i32 addrspace(2)* [ %tmp1, %if ], [ %tmp3, %else ] 84 %tmp5 = getelementptr i32, i32 addrspace(2)* %tmp4, i32 3000 85 %tmp6 = load i32, i32 addrspace(2)* %tmp5 86 store i32 %tmp6, i32 addrspace(1)* %out 87 ret void 88} 89 90; Test moving an SMRD with an immediate offset to the VALU 91 92; GCN-LABEL: {{^}}smrd_valu2: 93; GCN-NOHSA-NOT: v_add 94; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], 0 addr64 offset:16{{$}} 95; GCN-HSA: flat_load_dword v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}] 96define void @smrd_valu2(i32 addrspace(1)* %out, [8 x i32] addrspace(2)* %in) #1 { 97entry: 98 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 99 %tmp1 = add i32 %tmp, 4 100 %tmp2 = getelementptr [8 x i32], [8 x i32] addrspace(2)* %in, i32 %tmp, i32 4 101 %tmp3 = load i32, i32 addrspace(2)* %tmp2 102 store i32 %tmp3, i32 addrspace(1)* %out 103 ret void 104} 105 106; Use a big offset that will use the SMRD literal offset on CI 107; GCN-LABEL: {{^}}smrd_valu_ci_offset: 108; GCN-NOHSA-NOT: v_add 109; GCN-NOHSA: s_movk_i32 [[OFFSET:s[0-9]+]], 0x4e20{{$}} 110; GCN-NOHSA-NOT: v_add 111; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET]] addr64{{$}} 112; GCN-NOHSA: v_add_i32_e32 113; GCN-NOHSA: buffer_store_dword 114; GCN-HSA: flat_load_dword v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}] 115; GCN-HSA: flat_store_dword v[{{[0-9]+:[0-9]+}}], v{{[0-9]+}} 116define void @smrd_valu_ci_offset(i32 addrspace(1)* %out, i32 addrspace(2)* %in, i32 %c) #1 { 117entry: 118 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 119 %tmp2 = getelementptr i32, i32 addrspace(2)* %in, i32 %tmp 120 %tmp3 = getelementptr i32, i32 addrspace(2)* %tmp2, i32 5000 121 %tmp4 = load i32, i32 addrspace(2)* %tmp3 122 %tmp5 = add i32 %tmp4, %c 123 store i32 %tmp5, i32 addrspace(1)* %out 124 ret void 125} 126 127; GCN-LABEL: {{^}}smrd_valu_ci_offset_x2: 128; GCN-NOHSA-NOT: v_add 129; GCN-NOHSA: s_mov_b32 [[OFFSET:s[0-9]+]], 0x9c40{{$}} 130; GCN-NOHSA-NOT: v_add 131; GCN-NOHSA: buffer_load_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET]] addr64{{$}} 132; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 133; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 134; GCN-NOHSA: buffer_store_dwordx2 135; GCN-HSA: flat_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}] 136define void @smrd_valu_ci_offset_x2(i64 addrspace(1)* %out, i64 addrspace(2)* %in, i64 %c) #1 { 137entry: 138 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 139 %tmp2 = getelementptr i64, i64 addrspace(2)* %in, i32 %tmp 140 %tmp3 = getelementptr i64, i64 addrspace(2)* %tmp2, i32 5000 141 %tmp4 = load i64, i64 addrspace(2)* %tmp3 142 %tmp5 = or i64 %tmp4, %c 143 store i64 %tmp5, i64 addrspace(1)* %out 144 ret void 145} 146 147; GCN-LABEL: {{^}}smrd_valu_ci_offset_x4: 148; GCN-NOHSA-NOT: v_add 149; GCN-NOHSA: s_movk_i32 [[OFFSET:s[0-9]+]], 0x4d20{{$}} 150; GCN-NOHSA-NOT: v_add 151; GCN-NOHSA: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET]] addr64{{$}} 152; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 153; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 154; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 155; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 156; GCN-NOHSA: buffer_store_dwordx4 157; GCN-HSA: flat_load_dwordx4 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}] 158define void @smrd_valu_ci_offset_x4(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(2)* %in, <4 x i32> %c) #1 { 159entry: 160 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 161 %tmp2 = getelementptr <4 x i32>, <4 x i32> addrspace(2)* %in, i32 %tmp 162 %tmp3 = getelementptr <4 x i32>, <4 x i32> addrspace(2)* %tmp2, i32 1234 163 %tmp4 = load <4 x i32>, <4 x i32> addrspace(2)* %tmp3 164 %tmp5 = or <4 x i32> %tmp4, %c 165 store <4 x i32> %tmp5, <4 x i32> addrspace(1)* %out 166 ret void 167} 168 169; Original scalar load uses SGPR offset on SI and 32-bit literal on 170; CI. 171 172; GCN-LABEL: {{^}}smrd_valu_ci_offset_x8: 173; GCN-NOHSA-NOT: v_add 174; GCN-NOHSA: s_mov_b32 [[OFFSET0:s[0-9]+]], 0x9a40{{$}} 175; GCN-NOHSA-NOT: v_add 176; GCN-NOHSA: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET0]] addr64{{$}} 177; GCN-NOHSA-NOT: v_add 178; GCN-NOHSA: s_mov_b32 [[OFFSET1:s[0-9]+]], 0x9a50{{$}} 179; GCN-NOHSA-NOT: v_add 180; GCN-NOHSA: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET1]] addr64{{$}} 181 182; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 183; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 184; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 185; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 186; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 187; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 188; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 189; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 190; GCN-NOHSA: buffer_store_dwordx4 191; GCN-NOHSA: buffer_store_dwordx4 192; GCN-HSA: flat_load_dwordx4 193; GCN-HSA: flat_load_dwordx4 194define void @smrd_valu_ci_offset_x8(<8 x i32> addrspace(1)* %out, <8 x i32> addrspace(2)* %in, <8 x i32> %c) #1 { 195entry: 196 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 197 %tmp2 = getelementptr <8 x i32>, <8 x i32> addrspace(2)* %in, i32 %tmp 198 %tmp3 = getelementptr <8 x i32>, <8 x i32> addrspace(2)* %tmp2, i32 1234 199 %tmp4 = load <8 x i32>, <8 x i32> addrspace(2)* %tmp3 200 %tmp5 = or <8 x i32> %tmp4, %c 201 store <8 x i32> %tmp5, <8 x i32> addrspace(1)* %out 202 ret void 203} 204 205; GCN-LABEL: {{^}}smrd_valu_ci_offset_x16: 206 207; GCN-NOHSA-DAG: s_mov_b32 [[OFFSET0:s[0-9]+]], 0x13480{{$}} 208; GCN-NOHSA-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET0]] addr64{{$}} 209; GCN-NOHSA-DAG: s_mov_b32 [[OFFSET1:s[0-9]+]], 0x13490{{$}} 210; GCN-NOHSA-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET1]] addr64{{$}} 211; GCN-NOHSA-DAG: s_mov_b32 [[OFFSET2:s[0-9]+]], 0x134a0{{$}} 212; GCN-NOHSA-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET2]] addr64{{$}} 213; GCN-NOHSA-DAG: s_mov_b32 [[OFFSET3:s[0-9]+]], 0x134b0{{$}} 214; GCN-NOHSA-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET3]] addr64{{$}} 215 216; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 217; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 218; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 219; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 220; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 221; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 222; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 223; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}} 224; GCN-NOHSA: buffer_store_dwordx4 225; GCN-NOHSA: buffer_store_dwordx4 226; GCN-NOHSA: buffer_store_dwordx4 227; GCN-NOHSA: buffer_store_dwordx4 228 229; GCN-HSA: flat_load_dwordx4 230; GCN-HSA: flat_load_dwordx4 231; GCN-HSA: flat_load_dwordx4 232; GCN-HSA: flat_load_dwordx4 233 234; GCN: s_endpgm 235define void @smrd_valu_ci_offset_x16(<16 x i32> addrspace(1)* %out, <16 x i32> addrspace(2)* %in, <16 x i32> %c) #1 { 236entry: 237 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 238 %tmp2 = getelementptr <16 x i32>, <16 x i32> addrspace(2)* %in, i32 %tmp 239 %tmp3 = getelementptr <16 x i32>, <16 x i32> addrspace(2)* %tmp2, i32 1234 240 %tmp4 = load <16 x i32>, <16 x i32> addrspace(2)* %tmp3 241 %tmp5 = or <16 x i32> %tmp4, %c 242 store <16 x i32> %tmp5, <16 x i32> addrspace(1)* %out 243 ret void 244} 245 246; GCN-LABEL: {{^}}smrd_valu2_salu_user: 247; GCN-NOHSA: buffer_load_dword [[MOVED:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}} 248; GCN-HSA: flat_load_dword [[MOVED:v[0-9]+]], v[{{[0-9+:[0-9]+}}] 249; GCN: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, s{{[0-9]+}}, [[MOVED]] 250; GCN-NOHSA: buffer_store_dword [[ADD]] 251; GCN-HSA: flat_store_dword {{.*}}, [[ADD]] 252define void @smrd_valu2_salu_user(i32 addrspace(1)* %out, [8 x i32] addrspace(2)* %in, i32 %a) #1 { 253entry: 254 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 255 %tmp1 = add i32 %tmp, 4 256 %tmp2 = getelementptr [8 x i32], [8 x i32] addrspace(2)* %in, i32 %tmp, i32 4 257 %tmp3 = load i32, i32 addrspace(2)* %tmp2 258 %tmp4 = add i32 %tmp3, %a 259 store i32 %tmp4, i32 addrspace(1)* %out 260 ret void 261} 262 263; GCN-LABEL: {{^}}smrd_valu2_max_smrd_offset: 264; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:1020{{$}} 265; GCN-HSA flat_load_dword v{{[0-9]}}, v{{[0-9]+:[0-9]+}} 266define void @smrd_valu2_max_smrd_offset(i32 addrspace(1)* %out, [1024 x i32] addrspace(2)* %in) #1 { 267entry: 268 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 269 %tmp1 = add i32 %tmp, 4 270 %tmp2 = getelementptr [1024 x i32], [1024 x i32] addrspace(2)* %in, i32 %tmp, i32 255 271 %tmp3 = load i32, i32 addrspace(2)* %tmp2 272 store i32 %tmp3, i32 addrspace(1)* %out 273 ret void 274} 275 276; GCN-LABEL: {{^}}smrd_valu2_mubuf_offset: 277; GCN-NOHSA-NOT: v_add 278; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:1024{{$}} 279; GCN-HSA: flat_load_dword v{{[0-9]}}, v[{{[0-9]+:[0-9]+}}] 280define void @smrd_valu2_mubuf_offset(i32 addrspace(1)* %out, [1024 x i32] addrspace(2)* %in) #1 { 281entry: 282 %tmp = call i32 @llvm.amdgcn.workitem.id.x() 283 %tmp1 = add i32 %tmp, 4 284 %tmp2 = getelementptr [1024 x i32], [1024 x i32] addrspace(2)* %in, i32 %tmp, i32 256 285 %tmp3 = load i32, i32 addrspace(2)* %tmp2 286 store i32 %tmp3, i32 addrspace(1)* %out 287 ret void 288} 289 290; GCN-LABEL: {{^}}s_load_imm_v8i32: 291; GCN-NOHSA: buffer_load_dwordx4 292; GCN-NOHSA: buffer_load_dwordx4 293; GCN-HSA: flat_load_dwordx4 294; GCN-HSA: flat_load_dwordx4 295define void @s_load_imm_v8i32(<8 x i32> addrspace(1)* %out, i32 addrspace(2)* nocapture readonly %in) #1 { 296entry: 297 %tmp0 = tail call i32 @llvm.amdgcn.workitem.id.x() 298 %tmp1 = getelementptr inbounds i32, i32 addrspace(2)* %in, i32 %tmp0 299 %tmp2 = bitcast i32 addrspace(2)* %tmp1 to <8 x i32> addrspace(2)* 300 %tmp3 = load <8 x i32>, <8 x i32> addrspace(2)* %tmp2, align 4 301 store <8 x i32> %tmp3, <8 x i32> addrspace(1)* %out, align 32 302 ret void 303} 304 305; GCN-LABEL: {{^}}s_load_imm_v8i32_salu_user: 306; GCN-NOHSA: buffer_load_dwordx4 307; GCN-NOHSA: buffer_load_dwordx4 308; GCN-NOHSA: v_add_i32_e32 309; GCN-NOHSA: v_add_i32_e32 310; GCN-NOHSA: v_add_i32_e32 311; GCN-NOHSA: v_add_i32_e32 312; GCN-NOHSA: v_add_i32_e32 313; GCN-NOHSA: v_add_i32_e32 314; GCN-NOHSA: v_add_i32_e32 315; GCN-NOHSA: buffer_store_dword 316; GCN-HSA: flat_load_dwordx4 317; GCN-HSA: flat_load_dwordx4 318define void @s_load_imm_v8i32_salu_user(i32 addrspace(1)* %out, i32 addrspace(2)* nocapture readonly %in) #1 { 319entry: 320 %tmp0 = tail call i32 @llvm.amdgcn.workitem.id.x() 321 %tmp1 = getelementptr inbounds i32, i32 addrspace(2)* %in, i32 %tmp0 322 %tmp2 = bitcast i32 addrspace(2)* %tmp1 to <8 x i32> addrspace(2)* 323 %tmp3 = load <8 x i32>, <8 x i32> addrspace(2)* %tmp2, align 4 324 325 %elt0 = extractelement <8 x i32> %tmp3, i32 0 326 %elt1 = extractelement <8 x i32> %tmp3, i32 1 327 %elt2 = extractelement <8 x i32> %tmp3, i32 2 328 %elt3 = extractelement <8 x i32> %tmp3, i32 3 329 %elt4 = extractelement <8 x i32> %tmp3, i32 4 330 %elt5 = extractelement <8 x i32> %tmp3, i32 5 331 %elt6 = extractelement <8 x i32> %tmp3, i32 6 332 %elt7 = extractelement <8 x i32> %tmp3, i32 7 333 334 %add0 = add i32 %elt0, %elt1 335 %add1 = add i32 %add0, %elt2 336 %add2 = add i32 %add1, %elt3 337 %add3 = add i32 %add2, %elt4 338 %add4 = add i32 %add3, %elt5 339 %add5 = add i32 %add4, %elt6 340 %add6 = add i32 %add5, %elt7 341 342 store i32 %add6, i32 addrspace(1)* %out 343 ret void 344} 345 346; GCN-LABEL: {{^}}s_load_imm_v16i32: 347; GCN-NOHSA: buffer_load_dwordx4 348; GCN-NOHSA: buffer_load_dwordx4 349; GCN-NOHSA: buffer_load_dwordx4 350; GCN-NOHSA: buffer_load_dwordx4 351; GCN-HSA: flat_load_dwordx4 352; GCN-HSA: flat_load_dwordx4 353; GCN-HSA: flat_load_dwordx4 354; GCN-HSA: flat_load_dwordx4 355define void @s_load_imm_v16i32(<16 x i32> addrspace(1)* %out, i32 addrspace(2)* nocapture readonly %in) #1 { 356entry: 357 %tmp0 = tail call i32 @llvm.amdgcn.workitem.id.x() 358 %tmp1 = getelementptr inbounds i32, i32 addrspace(2)* %in, i32 %tmp0 359 %tmp2 = bitcast i32 addrspace(2)* %tmp1 to <16 x i32> addrspace(2)* 360 %tmp3 = load <16 x i32>, <16 x i32> addrspace(2)* %tmp2, align 4 361 store <16 x i32> %tmp3, <16 x i32> addrspace(1)* %out, align 32 362 ret void 363} 364 365; GCN-LABEL: {{^}}s_load_imm_v16i32_salu_user: 366; GCN-NOHSA: buffer_load_dwordx4 367; GCN-NOHSA: buffer_load_dwordx4 368; GCN-NOHSA: buffer_load_dwordx4 369; GCN-NOHSA: buffer_load_dwordx4 370; GCN-NOHSA: v_add_i32_e32 371; GCN-NOHSA: v_add_i32_e32 372; GCN-NOHSA: v_add_i32_e32 373; GCN-NOHSA: v_add_i32_e32 374; GCN-NOHSA: v_add_i32_e32 375; GCN-NOHSA: v_add_i32_e32 376; GCN-NOHSA: v_add_i32_e32 377; GCN-NOHSA: v_add_i32_e32 378; GCN-NOHSA: v_add_i32_e32 379; GCN-NOHSA: v_add_i32_e32 380; GCN-NOHSA: v_add_i32_e32 381; GCN-NOHSA: v_add_i32_e32 382; GCN-NOHSA: v_add_i32_e32 383; GCN-NOHSA: v_add_i32_e32 384; GCN-NOHSA: v_add_i32_e32 385; GCN-NOHSA: buffer_store_dword 386; GCN-HSA: flat_load_dwordx4 387; GCN-HSA: flat_load_dwordx4 388; GCN-HSA: flat_load_dwordx4 389; GCN-HSA: flat_load_dwordx4 390define void @s_load_imm_v16i32_salu_user(i32 addrspace(1)* %out, i32 addrspace(2)* nocapture readonly %in) #1 { 391entry: 392 %tmp0 = tail call i32 @llvm.amdgcn.workitem.id.x() 393 %tmp1 = getelementptr inbounds i32, i32 addrspace(2)* %in, i32 %tmp0 394 %tmp2 = bitcast i32 addrspace(2)* %tmp1 to <16 x i32> addrspace(2)* 395 %tmp3 = load <16 x i32>, <16 x i32> addrspace(2)* %tmp2, align 4 396 397 %elt0 = extractelement <16 x i32> %tmp3, i32 0 398 %elt1 = extractelement <16 x i32> %tmp3, i32 1 399 %elt2 = extractelement <16 x i32> %tmp3, i32 2 400 %elt3 = extractelement <16 x i32> %tmp3, i32 3 401 %elt4 = extractelement <16 x i32> %tmp3, i32 4 402 %elt5 = extractelement <16 x i32> %tmp3, i32 5 403 %elt6 = extractelement <16 x i32> %tmp3, i32 6 404 %elt7 = extractelement <16 x i32> %tmp3, i32 7 405 %elt8 = extractelement <16 x i32> %tmp3, i32 8 406 %elt9 = extractelement <16 x i32> %tmp3, i32 9 407 %elt10 = extractelement <16 x i32> %tmp3, i32 10 408 %elt11 = extractelement <16 x i32> %tmp3, i32 11 409 %elt12 = extractelement <16 x i32> %tmp3, i32 12 410 %elt13 = extractelement <16 x i32> %tmp3, i32 13 411 %elt14 = extractelement <16 x i32> %tmp3, i32 14 412 %elt15 = extractelement <16 x i32> %tmp3, i32 15 413 414 %add0 = add i32 %elt0, %elt1 415 %add1 = add i32 %add0, %elt2 416 %add2 = add i32 %add1, %elt3 417 %add3 = add i32 %add2, %elt4 418 %add4 = add i32 %add3, %elt5 419 %add5 = add i32 %add4, %elt6 420 %add6 = add i32 %add5, %elt7 421 %add7 = add i32 %add6, %elt8 422 %add8 = add i32 %add7, %elt9 423 %add9 = add i32 %add8, %elt10 424 %add10 = add i32 %add9, %elt11 425 %add11 = add i32 %add10, %elt12 426 %add12 = add i32 %add11, %elt13 427 %add13 = add i32 %add12, %elt14 428 %add14 = add i32 %add13, %elt15 429 430 store i32 %add14, i32 addrspace(1)* %out 431 ret void 432} 433 434; Make sure we legalize vopc operands after moving an sopc to the value. 435 436; {{^}}sopc_vopc_legalize_bug: 437; GCN: s_load_dword [[SGPR:s[0-9]+]] 438; GCN: v_cmp_le_u32_e32 vcc, [[SGPR]], v{{[0-9]+}} 439; GCN: s_and_b64 vcc, exec, vcc 440; GCN: s_cbranch_vccnz [[EXIT:[A-Z0-9_]+]] 441; GCN: v_mov_b32_e32 [[ONE:v[0-9]+]], 1 442; GCN-NOHSA: buffer_store_dword [[ONE]] 443; GCN-HSA: flat_store_dword v[{{[0-9]+:[0-9]+}}], [[ONE]] 444; GCN; {{^}}[[EXIT]]: 445; GCN: s_endpgm 446define void @sopc_vopc_legalize_bug(i32 %cond, i32 addrspace(1)* %out, i32 addrspace(1)* %in) { 447bb3: ; preds = %bb2 448 %tmp0 = bitcast i32 %cond to float 449 %tmp1 = fadd float %tmp0, 2.500000e-01 450 %tmp2 = bitcast float %tmp1 to i32 451 %tmp3 = icmp ult i32 %tmp2, %cond 452 br i1 %tmp3, label %bb6, label %bb7 453 454bb6: 455 store i32 1, i32 addrspace(1)* %out 456 br label %bb7 457 458bb7: ; preds = %bb3 459 ret void 460} 461 462attributes #0 = { nounwind readnone } 463attributes #1 = { nounwind } 464