1; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs -mattr=+load-store-opt < %s | FileCheck -strict-whitespace -check-prefixes=GCN,CI %s 2; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs -mattr=+load-store-opt < %s | FileCheck -strict-whitespace -check-prefixes=GCN,GFX9 %s 3 4@lds = addrspace(3) global [512 x float] undef, align 4 5@lds.f64 = addrspace(3) global [512 x double] undef, align 8 6 7 8; GCN-LABEL: {{^}}simple_write2_one_val_f32: 9; CI-DAG: s_mov_b32 m0 10; GFX9-NOT: m0 11 12; GCN-DAG: {{buffer|global}}_load_dword [[VAL:v[0-9]+]] 13; GCN-DAG: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 2, v{{[0-9]+}} 14; GCN: ds_write2_b32 [[VPTR]], [[VAL]], [[VAL]] offset1:8 15; GCN: s_endpgm 16define amdgpu_kernel void @simple_write2_one_val_f32(float addrspace(1)* %C, float addrspace(1)* %in) #0 { 17 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 18 %in.gep = getelementptr float, float addrspace(1)* %in, i32 %x.i 19 %val = load float, float addrspace(1)* %in.gep, align 4 20 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i 21 store float %val, float addrspace(3)* %arrayidx0, align 4 22 %add.x = add nsw i32 %x.i, 8 23 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x 24 store float %val, float addrspace(3)* %arrayidx1, align 4 25 ret void 26} 27 28; GCN-LABEL: {{^}}simple_write2_two_val_f32: 29; CI-DAG: s_mov_b32 m0 30; GFX9-NOT: m0 31 32; CI-DAG: buffer_load_dword [[VAL0:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}} 33; CI-DAG: buffer_load_dword [[VAL1:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4 34 35; GFX9-DAG: global_load_dword [[VAL0:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, off{{$}} 36; GFX9-DAG: global_load_dword [[VAL1:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, off offset:4 37 38; GCN-DAG: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 2, v{{[0-9]+}} 39; GCN: ds_write2_b32 [[VPTR]], [[VAL0]], [[VAL1]] offset1:8 40; GCN: s_endpgm 41define amdgpu_kernel void @simple_write2_two_val_f32(float addrspace(1)* %C, float addrspace(1)* %in) #0 { 42 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 43 %in.gep.0 = getelementptr float, float addrspace(1)* %in, i32 %x.i 44 %in.gep.1 = getelementptr float, float addrspace(1)* %in.gep.0, i32 1 45 %val0 = load volatile float, float addrspace(1)* %in.gep.0, align 4 46 %val1 = load volatile float, float addrspace(1)* %in.gep.1, align 4 47 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i 48 store float %val0, float addrspace(3)* %arrayidx0, align 4 49 %add.x = add nsw i32 %x.i, 8 50 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x 51 store float %val1, float addrspace(3)* %arrayidx1, align 4 52 ret void 53} 54 55; GCN-LABEL: @simple_write2_two_val_f32_volatile_0 56; CI-DAG: s_mov_b32 m0 57; GFX9-NOT: m0 58 59; GCN-NOT: ds_write2_b32 60; GCN: ds_write_b32 {{v[0-9]+}}, {{v[0-9]+}} 61; GCN: ds_write_b32 {{v[0-9]+}}, {{v[0-9]+}} offset:32 62; GCN: s_endpgm 63define amdgpu_kernel void @simple_write2_two_val_f32_volatile_0(float addrspace(1)* %C, float addrspace(1)* %in0, float addrspace(1)* %in1) #0 { 64 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 65 %in0.gep = getelementptr float, float addrspace(1)* %in0, i32 %x.i 66 %in1.gep = getelementptr float, float addrspace(1)* %in1, i32 %x.i 67 %val0 = load volatile float, float addrspace(1)* %in0.gep, align 4 68 %val1 = load volatile float, float addrspace(1)* %in1.gep, align 4 69 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i 70 store volatile float %val0, float addrspace(3)* %arrayidx0, align 4 71 %add.x = add nsw i32 %x.i, 8 72 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x 73 store float %val1, float addrspace(3)* %arrayidx1, align 4 74 ret void 75} 76 77; GCN-LABEL: @simple_write2_two_val_f32_volatile_1 78; CI-DAG: s_mov_b32 m0 79; GFX9-NOT: m0 80 81; GCN-NOT: ds_write2_b32 82; GCN: ds_write_b32 {{v[0-9]+}}, {{v[0-9]+}} 83; GCN: ds_write_b32 {{v[0-9]+}}, {{v[0-9]+}} offset:32 84; GCN: s_endpgm 85define amdgpu_kernel void @simple_write2_two_val_f32_volatile_1(float addrspace(1)* %C, float addrspace(1)* %in0, float addrspace(1)* %in1) #0 { 86 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 87 %in0.gep = getelementptr float, float addrspace(1)* %in0, i32 %x.i 88 %in1.gep = getelementptr float, float addrspace(1)* %in1, i32 %x.i 89 %val0 = load volatile float, float addrspace(1)* %in0.gep, align 4 90 %val1 = load volatile float, float addrspace(1)* %in1.gep, align 4 91 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i 92 store float %val0, float addrspace(3)* %arrayidx0, align 4 93 %add.x = add nsw i32 %x.i, 8 94 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x 95 store volatile float %val1, float addrspace(3)* %arrayidx1, align 4 96 ret void 97} 98 99; 2 data subregisters from different super registers. 100; GCN-LABEL: {{^}}simple_write2_two_val_subreg2_mixed_f32: 101; GFX9-NOT: m0 102 103; CI: buffer_load_dwordx2 v{{\[}}[[VAL0:[0-9]+]]:{{[0-9]+\]}} 104; CI: buffer_load_dwordx2 v{{\[[0-9]+}}:[[VAL1:[0-9]+]]{{\]}} 105; CI-DAG: s_mov_b32 m0 106 107; GCN-DAG: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 2, v{{[0-9]+}} 108 109; GFX9: global_load_dwordx2 v{{\[}}[[VAL0:[0-9]+]]:{{[0-9]+\]}} 110; GFX9: global_load_dwordx2 v{{\[[0-9]+}}:[[VAL1:[0-9]+]]{{\]}} 111; GCN: ds_write2_b32 [[VPTR]], v[[VAL0]], v[[VAL1]] offset1:8 112; GCN: s_endpgm 113define amdgpu_kernel void @simple_write2_two_val_subreg2_mixed_f32(float addrspace(1)* %C, <2 x float> addrspace(1)* %in) #0 { 114 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 115 %in.gep.0 = getelementptr <2 x float>, <2 x float> addrspace(1)* %in, i32 %x.i 116 %in.gep.1 = getelementptr <2 x float>, <2 x float> addrspace(1)* %in.gep.0, i32 1 117 %val0 = load volatile <2 x float>, <2 x float> addrspace(1)* %in.gep.0, align 8 118 %val1 = load volatile <2 x float>, <2 x float> addrspace(1)* %in.gep.1, align 8 119 %val0.0 = extractelement <2 x float> %val0, i32 0 120 %val1.1 = extractelement <2 x float> %val1, i32 1 121 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i 122 store float %val0.0, float addrspace(3)* %arrayidx0, align 4 123 %add.x = add nsw i32 %x.i, 8 124 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x 125 store float %val1.1, float addrspace(3)* %arrayidx1, align 4 126 ret void 127} 128 129; GCN-LABEL: @simple_write2_two_val_subreg2_f32 130; CI-DAG: s_mov_b32 m0 131; GFX9-NOT: m0 132 133; GCN-DAG: {{buffer|global}}_load_dwordx2 v{{\[}}[[VAL0:[0-9]+]]:[[VAL1:[0-9]+]]{{\]}} 134; GCN-DAG: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 2, v{{[0-9]+}} 135; GCN: ds_write2_b32 [[VPTR]], v[[VAL0]], v[[VAL1]] offset1:8 136; GCN: s_endpgm 137define amdgpu_kernel void @simple_write2_two_val_subreg2_f32(float addrspace(1)* %C, <2 x float> addrspace(1)* %in) #0 { 138 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 139 %in.gep = getelementptr <2 x float>, <2 x float> addrspace(1)* %in, i32 %x.i 140 %val = load <2 x float>, <2 x float> addrspace(1)* %in.gep, align 8 141 %val0 = extractelement <2 x float> %val, i32 0 142 %val1 = extractelement <2 x float> %val, i32 1 143 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i 144 store float %val0, float addrspace(3)* %arrayidx0, align 4 145 %add.x = add nsw i32 %x.i, 8 146 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x 147 store float %val1, float addrspace(3)* %arrayidx1, align 4 148 ret void 149} 150 151; GCN-LABEL: @simple_write2_two_val_subreg4_f32 152; CI-DAG: s_mov_b32 m0 153; GFX9-NOT: m0 154 155; GCN-DAG: {{buffer|global}}_load_dwordx4 v{{\[}}[[VAL0:[0-9]+]]:[[VAL1:[0-9]+]]{{\]}} 156; GCN-DAG: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 2, v{{[0-9]+}} 157; GCN: ds_write2_b32 [[VPTR]], v[[VAL0]], v[[VAL1]] offset1:8 158; GCN: s_endpgm 159define amdgpu_kernel void @simple_write2_two_val_subreg4_f32(float addrspace(1)* %C, <4 x float> addrspace(1)* %in) #0 { 160 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 161 %in.gep = getelementptr <4 x float>, <4 x float> addrspace(1)* %in, i32 %x.i 162 %val = load <4 x float>, <4 x float> addrspace(1)* %in.gep, align 16 163 %val0 = extractelement <4 x float> %val, i32 0 164 %val1 = extractelement <4 x float> %val, i32 3 165 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i 166 store float %val0, float addrspace(3)* %arrayidx0, align 4 167 %add.x = add nsw i32 %x.i, 8 168 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x 169 store float %val1, float addrspace(3)* %arrayidx1, align 4 170 ret void 171} 172 173; GCN-LABEL: @simple_write2_two_val_max_offset_f32 174; CI-DAG: s_mov_b32 m0 175; GFX9-NOT: m0 176 177; CI-DAG: buffer_load_dword [[VAL0:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}} 178; CI-DAG: buffer_load_dword [[VAL1:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4 179 180; GFX9-DAG: global_load_dword [[VAL0:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, off{{$}} 181; GFX9-DAG: global_load_dword [[VAL1:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, off offset:4 182 183; GCN-DAG: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 2, v{{[0-9]+}} 184; GCN: ds_write2_b32 [[VPTR]], [[VAL0]], [[VAL1]] offset1:255 185; GCN: s_endpgm 186define amdgpu_kernel void @simple_write2_two_val_max_offset_f32(float addrspace(1)* %C, float addrspace(1)* %in) #0 { 187 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 188 %in.gep.0 = getelementptr float, float addrspace(1)* %in, i32 %x.i 189 %in.gep.1 = getelementptr float, float addrspace(1)* %in.gep.0, i32 1 190 %val0 = load volatile float, float addrspace(1)* %in.gep.0, align 4 191 %val1 = load volatile float, float addrspace(1)* %in.gep.1, align 4 192 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i 193 store float %val0, float addrspace(3)* %arrayidx0, align 4 194 %add.x = add nsw i32 %x.i, 255 195 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x 196 store float %val1, float addrspace(3)* %arrayidx1, align 4 197 ret void 198} 199 200; GCN-LABEL: @simple_write2_two_val_too_far_f32 201; CI-DAG: s_mov_b32 m0 202; GFX9-NOT: m0 203 204; GCN: ds_write_b32 v{{[0-9]+}}, v{{[0-9]+}} 205; GCN: ds_write_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:1028 206; GCN: s_endpgm 207define amdgpu_kernel void @simple_write2_two_val_too_far_f32(float addrspace(1)* %C, float addrspace(1)* %in0, float addrspace(1)* %in1) #0 { 208 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 209 %in0.gep = getelementptr float, float addrspace(1)* %in0, i32 %x.i 210 %in1.gep = getelementptr float, float addrspace(1)* %in1, i32 %x.i 211 %val0 = load float, float addrspace(1)* %in0.gep, align 4 212 %val1 = load float, float addrspace(1)* %in1.gep, align 4 213 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i 214 store float %val0, float addrspace(3)* %arrayidx0, align 4 215 %add.x = add nsw i32 %x.i, 257 216 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x 217 store float %val1, float addrspace(3)* %arrayidx1, align 4 218 ret void 219} 220 221; GCN-LABEL: @simple_write2_two_val_f32_x2 222; CI-DAG: s_mov_b32 m0 223; GFX9-NOT: m0 224 225; GCN: ds_write2_b32 [[BASEADDR:v[0-9]+]], [[VAL0:v[0-9]+]], [[VAL1:v[0-9]+]] offset1:8 226; GCN: ds_write2_b32 [[BASEADDR:v[0-9]+]], [[VAL0]], [[VAL1]] offset0:11 offset1:27 227; GCN: s_endpgm 228define amdgpu_kernel void @simple_write2_two_val_f32_x2(float addrspace(1)* %C, float addrspace(1)* %in0, float addrspace(1)* %in1) #0 { 229 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1 230 %in0.gep = getelementptr float, float addrspace(1)* %in0, i32 %tid.x 231 %in1.gep = getelementptr float, float addrspace(1)* %in1, i32 %tid.x 232 %val0 = load float, float addrspace(1)* %in0.gep, align 4 233 %val1 = load float, float addrspace(1)* %in1.gep, align 4 234 235 %idx.0 = add nsw i32 %tid.x, 0 236 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0 237 store float %val0, float addrspace(3)* %arrayidx0, align 4 238 239 %idx.1 = add nsw i32 %tid.x, 8 240 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1 241 store float %val1, float addrspace(3)* %arrayidx1, align 4 242 243 %idx.2 = add nsw i32 %tid.x, 11 244 %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2 245 store float %val0, float addrspace(3)* %arrayidx2, align 4 246 247 %idx.3 = add nsw i32 %tid.x, 27 248 %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3 249 store float %val1, float addrspace(3)* %arrayidx3, align 4 250 251 ret void 252} 253 254; GCN-LABEL: @simple_write2_two_val_f32_x2_nonzero_base 255; CI-DAG: s_mov_b32 m0 256; GFX9-NOT: m0 257 258; GCN: ds_write2_b32 [[BASEADDR:v[0-9]+]], [[VAL0:v[0-9]+]], [[VAL1:v[0-9]+]] offset0:3 offset1:8 259; GCN: ds_write2_b32 [[BASEADDR:v[0-9]+]], [[VAL0]], [[VAL1]] offset0:11 offset1:27 260; GCN: s_endpgm 261define amdgpu_kernel void @simple_write2_two_val_f32_x2_nonzero_base(float addrspace(1)* %C, float addrspace(1)* %in0, float addrspace(1)* %in1) #0 { 262 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1 263 %in0.gep = getelementptr float, float addrspace(1)* %in0, i32 %tid.x 264 %in1.gep = getelementptr float, float addrspace(1)* %in1, i32 %tid.x 265 %val0 = load float, float addrspace(1)* %in0.gep, align 4 266 %val1 = load float, float addrspace(1)* %in1.gep, align 4 267 268 %idx.0 = add nsw i32 %tid.x, 3 269 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0 270 store float %val0, float addrspace(3)* %arrayidx0, align 4 271 272 %idx.1 = add nsw i32 %tid.x, 8 273 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1 274 store float %val1, float addrspace(3)* %arrayidx1, align 4 275 276 %idx.2 = add nsw i32 %tid.x, 11 277 %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2 278 store float %val0, float addrspace(3)* %arrayidx2, align 4 279 280 %idx.3 = add nsw i32 %tid.x, 27 281 %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3 282 store float %val1, float addrspace(3)* %arrayidx3, align 4 283 284 ret void 285} 286 287; GCN-LABEL: @write2_ptr_subreg_arg_two_val_f32 288; CI-DAG: s_mov_b32 m0 289; GFX9-NOT: m0 290 291; GCN-NOT: ds_write2_b32 292; GCN: ds_write_b32 293; GCN: ds_write_b32 294; GCN: s_endpgm 295define amdgpu_kernel void @write2_ptr_subreg_arg_two_val_f32(float addrspace(1)* %C, float addrspace(1)* %in0, float addrspace(1)* %in1, <2 x float addrspace(3)*> %lds.ptr) #0 { 296 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 297 %in0.gep = getelementptr float, float addrspace(1)* %in0, i32 %x.i 298 %in1.gep = getelementptr float, float addrspace(1)* %in1, i32 %x.i 299 %val0 = load float, float addrspace(1)* %in0.gep, align 4 300 %val1 = load float, float addrspace(1)* %in1.gep, align 4 301 302 %index.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0 303 %index.1 = insertelement <2 x i32> %index.0, i32 8, i32 0 304 %gep = getelementptr inbounds float, <2 x float addrspace(3)*> %lds.ptr, <2 x i32> %index.1 305 %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0 306 %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1 307 308 ; Apply an additional offset after the vector that will be more obviously folded. 309 %gep.1.offset = getelementptr float, float addrspace(3)* %gep.1, i32 8 310 store float %val0, float addrspace(3)* %gep.0, align 4 311 312 %add.x = add nsw i32 %x.i, 8 313 store float %val1, float addrspace(3)* %gep.1.offset, align 4 314 ret void 315} 316 317; GCN-LABEL: @simple_write2_one_val_f64 318; CI-DAG: s_mov_b32 m0 319; GFX9-NOT: m0 320 321; GCN-DAG: {{buffer|global}}_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]], 322; GCN-DAG: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 3, v{{[0-9]+}} 323; GCN: ds_write2_b64 [[VPTR]], [[VAL]], [[VAL]] offset1:8 324; GCN: s_endpgm 325define amdgpu_kernel void @simple_write2_one_val_f64(double addrspace(1)* %C, double addrspace(1)* %in) #0 { 326 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 327 %in.gep = getelementptr double, double addrspace(1)* %in, i32 %x.i 328 %val = load double, double addrspace(1)* %in.gep, align 8 329 %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i 330 store double %val, double addrspace(3)* %arrayidx0, align 8 331 %add.x = add nsw i32 %x.i, 8 332 %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x 333 store double %val, double addrspace(3)* %arrayidx1, align 8 334 ret void 335} 336 337; GCN-LABEL: @misaligned_simple_write2_one_val_f64 338; CI-DAG: s_mov_b32 m0 339; GFX9-NOT: m0 340 341; GCN-DAG: {{buffer|global}}_load_dwordx2 v{{\[}}[[VAL0:[0-9]+]]:[[VAL1:[0-9]+]]{{\]}} 342; GCN-DAG: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 3, v{{[0-9]+}} 343; GCN: ds_write2_b32 [[VPTR]], v[[VAL0]], v[[VAL1]] offset1:1 344; GCN: ds_write2_b32 [[VPTR]], v[[VAL0]], v[[VAL1]] offset0:14 offset1:15 345; GCN: s_endpgm 346define amdgpu_kernel void @misaligned_simple_write2_one_val_f64(double addrspace(1)* %C, double addrspace(1)* %in, double addrspace(3)* %lds) #0 { 347 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 348 %in.gep = getelementptr double, double addrspace(1)* %in, i32 %x.i 349 %val = load double, double addrspace(1)* %in.gep, align 8 350 %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %x.i 351 store double %val, double addrspace(3)* %arrayidx0, align 4 352 %add.x = add nsw i32 %x.i, 7 353 %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x 354 store double %val, double addrspace(3)* %arrayidx1, align 4 355 ret void 356} 357 358; GCN-LABEL: @simple_write2_two_val_f64 359; CI-DAG: s_mov_b32 m0 360; GFX9-NOT: m0 361 362; CI-DAG: buffer_load_dwordx2 [[VAL0:v\[[0-9]+:[0-9]+\]]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}} 363; CI-DAG: buffer_load_dwordx2 [[VAL1:v\[[0-9]+:[0-9]+\]]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8 364 365; GFX9-DAG: global_load_dwordx2 [[VAL0:v\[[0-9]+:[0-9]+\]]], {{v\[[0-9]+:[0-9]+\]}}, off{{$}} 366; GFX9-DAG: global_load_dwordx2 [[VAL1:v\[[0-9]+:[0-9]+\]]], {{v\[[0-9]+:[0-9]+\]}}, off offset:8 367 368 369; GCN-DAG: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 3, v{{[0-9]+}} 370; GCN: ds_write2_b64 [[VPTR]], [[VAL0]], [[VAL1]] offset1:8 371; GCN: s_endpgm 372define amdgpu_kernel void @simple_write2_two_val_f64(double addrspace(1)* %C, double addrspace(1)* %in) #0 { 373 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 374 %in.gep.0 = getelementptr double, double addrspace(1)* %in, i32 %x.i 375 %in.gep.1 = getelementptr double, double addrspace(1)* %in.gep.0, i32 1 376 %val0 = load volatile double, double addrspace(1)* %in.gep.0, align 8 377 %val1 = load volatile double, double addrspace(1)* %in.gep.1, align 8 378 %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i 379 store double %val0, double addrspace(3)* %arrayidx0, align 8 380 %add.x = add nsw i32 %x.i, 8 381 %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x 382 store double %val1, double addrspace(3)* %arrayidx1, align 8 383 ret void 384} 385 386@foo = addrspace(3) global [4 x i32] undef, align 4 387 388; GCN-LABEL: @store_constant_adjacent_offsets 389; CI-DAG: s_mov_b32 m0 390; GFX9-NOT: m0 391 392; GCN-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}} 393; GCN: ds_write2_b32 [[ZERO]], v{{[0-9]+}}, v{{[0-9]+}} offset1:1 394define amdgpu_kernel void @store_constant_adjacent_offsets() { 395 store i32 123, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4 396 store i32 123, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 1), align 4 397 ret void 398} 399 400; GCN-LABEL: @store_constant_disjoint_offsets 401; CI-DAG: s_mov_b32 m0 402; GFX9-NOT: m0 403 404; GCN-DAG: v_mov_b32_e32 [[VAL:v[0-9]+]], 0x7b{{$}} 405; GCN-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}} 406; GCN: ds_write2_b32 [[ZERO]], [[VAL]], [[VAL]] offset1:2 407define amdgpu_kernel void @store_constant_disjoint_offsets() { 408 store i32 123, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4 409 store i32 123, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 2), align 4 410 ret void 411} 412 413@bar = addrspace(3) global [4 x i64] undef, align 4 414 415; GCN-LABEL: @store_misaligned64_constant_offsets 416; CI-DAG: s_mov_b32 m0 417; GFX9-NOT: m0 418 419; GCN-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}} 420; GCN-DAG: ds_write2_b32 [[ZERO]], v{{[0-9]+}}, v{{[0-9]+}} offset1:1 421; GCN-DAG: ds_write2_b32 [[ZERO]], v{{[0-9]+}}, v{{[0-9]+}} offset0:2 offset1:3 422; GCN: s_endpgm 423define amdgpu_kernel void @store_misaligned64_constant_offsets() { 424 store i64 123, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 0), align 4 425 store i64 123, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 1), align 4 426 ret void 427} 428 429@bar.large = addrspace(3) global [4096 x i64] undef, align 4 430 431; GCN-LABEL: @store_misaligned64_constant_large_offsets 432; CI-DAG: s_mov_b32 m0 433; GFX9-NOT: m0 434 435; GCN-DAG: v_mov_b32_e32 [[BASE0:v[0-9]+]], 0x7ff8{{$}} 436; GCN-DAG: v_mov_b32_e32 [[BASE1:v[0-9]+]], 0x4000{{$}} 437; GCN-DAG: ds_write2_b32 [[BASE0]], v{{[0-9]+}}, v{{[0-9]+}} offset1:1 438; GCN-DAG: ds_write2_b32 [[BASE1]], v{{[0-9]+}}, v{{[0-9]+}} offset1:1 439; GCN: s_endpgm 440define amdgpu_kernel void @store_misaligned64_constant_large_offsets() { 441 store i64 123, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 2048), align 4 442 store i64 123, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 4095), align 4 443 ret void 444} 445 446@sgemm.lA = internal unnamed_addr addrspace(3) global [264 x float] undef, align 4 447@sgemm.lB = internal unnamed_addr addrspace(3) global [776 x float] undef, align 4 448 449define amdgpu_kernel void @write2_sgemm_sequence(float addrspace(1)* %C, i32 %lda, i32 %ldb, float addrspace(1)* %in) #0 { 450 %x.i = tail call i32 @llvm.amdgcn.workgroup.id.x() #1 451 %y.i = tail call i32 @llvm.amdgcn.workitem.id.y() #1 452 %val = load float, float addrspace(1)* %in 453 %arrayidx44 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %x.i 454 store float %val, float addrspace(3)* %arrayidx44, align 4 455 %add47 = add nsw i32 %x.i, 1 456 %arrayidx48 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add47 457 store float %val, float addrspace(3)* %arrayidx48, align 4 458 %add51 = add nsw i32 %x.i, 16 459 %arrayidx52 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add51 460 store float %val, float addrspace(3)* %arrayidx52, align 4 461 %add55 = add nsw i32 %x.i, 17 462 %arrayidx56 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add55 463 store float %val, float addrspace(3)* %arrayidx56, align 4 464 %arrayidx60 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %y.i 465 store float %val, float addrspace(3)* %arrayidx60, align 4 466 %add63 = add nsw i32 %y.i, 1 467 %arrayidx64 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add63 468 store float %val, float addrspace(3)* %arrayidx64, align 4 469 %add67 = add nsw i32 %y.i, 32 470 %arrayidx68 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add67 471 store float %val, float addrspace(3)* %arrayidx68, align 4 472 %add71 = add nsw i32 %y.i, 33 473 %arrayidx72 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add71 474 store float %val, float addrspace(3)* %arrayidx72, align 4 475 %add75 = add nsw i32 %y.i, 64 476 %arrayidx76 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add75 477 store float %val, float addrspace(3)* %arrayidx76, align 4 478 %add79 = add nsw i32 %y.i, 65 479 %arrayidx80 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add79 480 store float %val, float addrspace(3)* %arrayidx80, align 4 481 ret void 482} 483 484; GCN-LABEL: {{^}}simple_write2_v4f32_superreg_align4: 485; CI: s_mov_b32 m0 486; GFX9-NOT: m0 487 488; GCN: ds_write2_b32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}} offset0:2 offset1:3{{$}} 489; GCN: ds_write2_b32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}} offset1:1{{$}} 490define amdgpu_kernel void @simple_write2_v4f32_superreg_align4(<4 x float> addrspace(3)* %out, <4 x float> addrspace(1)* %in) #0 { 491 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1 492 %in.gep = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %in 493 %val0 = load <4 x float>, <4 x float> addrspace(1)* %in.gep, align 4 494 %out.gep = getelementptr inbounds <4 x float>, <4 x float> addrspace(3)* %out, i32 %x.i 495 store <4 x float> %val0, <4 x float> addrspace(3)* %out.gep, align 4 496 ret void 497} 498 499declare i32 @llvm.amdgcn.workgroup.id.x() #1 500declare i32 @llvm.amdgcn.workgroup.id.y() #1 501declare i32 @llvm.amdgcn.workitem.id.x() #1 502declare i32 @llvm.amdgcn.workitem.id.y() #1 503 504attributes #0 = { nounwind } 505attributes #1 = { nounwind readnone } 506attributes #2 = { convergent nounwind } 507