1; RUN: llc -march=amdgcn -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri -mattr=-code-object-v3,-promote-alloca -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=HSA -check-prefix=CI %s 2; RUN: llc -march=amdgcn -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=-code-object-v3,-promote-alloca -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=HSA -check-prefix=GFX9 %s 3 4; HSA-LABEL: {{^}}use_group_to_flat_addrspacecast: 5; HSA: enable_sgpr_private_segment_buffer = 1 6; HSA: enable_sgpr_dispatch_ptr = 0 7; CI: enable_sgpr_queue_ptr = 1 8; GFX9: enable_sgpr_queue_ptr = 0 9 10; CI-DAG: s_load_dword [[PTR:s[0-9]+]], s[6:7], 0x0{{$}} 11; CI-DAG: s_load_dword [[APERTURE:s[0-9]+]], s[4:5], 0x10{{$}} 12; CI-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[APERTURE]] 13; CI-DAG: v_cmp_ne_u32_e64 vcc, [[PTR]], -1 14; CI-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc 15; CI-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]] 16; CI-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]] 17 18; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7 19; GFX9-DAG: s_load_dword [[PTR:s[0-9]+]], s[4:5], 0x0{{$}} 20; GFX9-DAG: s_getreg_b32 [[SSRC_SHARED:s[0-9]+]], hwreg(HW_REG_SH_MEM_BASES, 16, 16) 21; GFX9-DAG: s_lshl_b32 [[SSRC_SHARED_BASE:s[0-9]+]], [[SSRC_SHARED]], 16 22; GFX9-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[SSRC_SHARED_BASE]] 23 24; GFX9-XXX: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], src_shared_base 25; GFX9: v_cmp_ne_u32_e64 vcc, [[PTR]], -1 26; GFX9: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc 27; GFX9-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]] 28; GFX9-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]] 29 30; HSA: flat_store_dword v{{\[}}[[LO]]:[[HI]]{{\]}}, [[K]] 31 32; At most 2 digits. Make sure src_shared_base is not counted as a high 33; number SGPR. 34 35; CI: NumSgprs: {{[0-9][0-9]+}} 36; GFX9: NumSgprs: {{[0-9]+}} 37define amdgpu_kernel void @use_group_to_flat_addrspacecast(i32 addrspace(3)* %ptr) #0 { 38 %stof = addrspacecast i32 addrspace(3)* %ptr to i32* 39 store volatile i32 7, i32* %stof 40 ret void 41} 42 43; Test handling inside a non-kernel 44; HSA-LABEL: {{^}}use_group_to_flat_addrspacecast_func: 45; CI-DAG: s_load_dword [[APERTURE:s[0-9]+]], s[4:5], 0x10{{$}} 46; CI-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[APERTURE]] 47; CI-DAG: v_cmp_ne_u32_e32 vcc, -1, v0 48; CI-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc 49; CI-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, v0 50 51; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7 52; GFX9-DAG: s_getreg_b32 [[SSRC_SHARED:s[0-9]+]], hwreg(HW_REG_SH_MEM_BASES, 16, 16) 53; GFX9-DAG: s_lshl_b32 [[SSRC_SHARED_BASE:s[0-9]+]], [[SSRC_SHARED]], 16 54; GFX9-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[SSRC_SHARED_BASE]] 55 56; GFX9-XXX: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], src_shared_base 57; GFX9-DAG: v_cmp_ne_u32_e32 vcc, -1, v0 58; GFX9-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, v0, vcc 59; GFX9-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc 60 61; HSA: flat_store_dword v{{\[}}[[LO]]:[[HI]]{{\]}}, [[K]] 62define void @use_group_to_flat_addrspacecast_func(i32 addrspace(3)* %ptr) #0 { 63 %stof = addrspacecast i32 addrspace(3)* %ptr to i32* 64 store volatile i32 7, i32* %stof 65 ret void 66} 67 68; HSA-LABEL: {{^}}use_private_to_flat_addrspacecast: 69; HSA: enable_sgpr_private_segment_buffer = 1 70; HSA: enable_sgpr_dispatch_ptr = 0 71; CI: enable_sgpr_queue_ptr = 1 72; GFX9: enable_sgpr_queue_ptr = 0 73 74; CI-DAG: s_load_dword [[PTR:s[0-9]+]], s[6:7], 0x0{{$}} 75; CI-DAG: s_load_dword [[APERTURE:s[0-9]+]], s[4:5], 0x11{{$}} 76; CI-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[APERTURE]] 77 78; CI-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7 79; CI-DAG: v_cmp_ne_u32_e64 vcc, [[PTR]], -1 80; CI-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc 81; CI-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]] 82; CI-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]] 83 84; GFX9-DAG: s_load_dword [[PTR:s[0-9]+]], s[4:5], 0x0{{$}} 85; GFX9-DAG: s_getreg_b32 [[SSRC_PRIVATE:s[0-9]+]], hwreg(HW_REG_SH_MEM_BASES, 0, 16) 86; GFX9-DAG: s_lshl_b32 [[SSRC_PRIVATE_BASE:s[0-9]+]], [[SSRC_PRIVATE]], 16 87; GFX9-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[SSRC_PRIVATE_BASE]] 88 89; GFX9-XXX: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], src_private_base 90 91; GFX9-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7 92; GFX9: v_cmp_ne_u32_e64 vcc, [[PTR]], -1 93; GFX9: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc 94; GFX9: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]] 95; GFX9-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]] 96 97; HSA: flat_store_dword v{{\[}}[[LO]]:[[HI]]{{\]}}, [[K]] 98 99; CI: NumSgprs: {{[0-9][0-9]+}} 100; GFX9: NumSgprs: {{[0-9]+}} 101define amdgpu_kernel void @use_private_to_flat_addrspacecast(i32 addrspace(5)* %ptr) #0 { 102 %stof = addrspacecast i32 addrspace(5)* %ptr to i32* 103 store volatile i32 7, i32* %stof 104 ret void 105} 106 107; no-op 108; HSA-LABEL: {{^}}use_global_to_flat_addrspacecast: 109; HSA: enable_sgpr_queue_ptr = 0 110 111; HSA: s_load_dwordx2 s{{\[}}[[PTRLO:[0-9]+]]:[[PTRHI:[0-9]+]]{{\]}} 112; HSA-DAG: v_mov_b32_e32 v[[VPTRLO:[0-9]+]], s[[PTRLO]] 113; HSA-DAG: v_mov_b32_e32 v[[VPTRHI:[0-9]+]], s[[PTRHI]] 114; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7 115; HSA: flat_store_dword v{{\[}}[[VPTRLO]]:[[VPTRHI]]{{\]}}, [[K]] 116define amdgpu_kernel void @use_global_to_flat_addrspacecast(i32 addrspace(1)* %ptr) #0 { 117 %stof = addrspacecast i32 addrspace(1)* %ptr to i32* 118 store volatile i32 7, i32* %stof 119 ret void 120} 121 122; no-op 123; HSA-LABEl: {{^}}use_constant_to_flat_addrspacecast: 124; HSA: s_load_dwordx2 s{{\[}}[[PTRLO:[0-9]+]]:[[PTRHI:[0-9]+]]{{\]}} 125; HSA-DAG: v_mov_b32_e32 v[[VPTRLO:[0-9]+]], s[[PTRLO]] 126; HSA-DAG: v_mov_b32_e32 v[[VPTRHI:[0-9]+]], s[[PTRHI]] 127; HSA: flat_load_dword v{{[0-9]+}}, v{{\[}}[[VPTRLO]]:[[VPTRHI]]{{\]}} 128define amdgpu_kernel void @use_constant_to_flat_addrspacecast(i32 addrspace(4)* %ptr) #0 { 129 %stof = addrspacecast i32 addrspace(4)* %ptr to i32* 130 %ld = load volatile i32, i32* %stof 131 ret void 132} 133 134; HSA-LABEl: {{^}}use_constant_to_global_addrspacecast: 135; HSA: s_load_dwordx2 s{{\[}}[[PTRLO:[0-9]+]]:[[PTRHI:[0-9]+]]{{\]}} 136; HSA-DAG: v_mov_b32_e32 v[[VPTRLO:[0-9]+]], s[[PTRLO]] 137; HSA-DAG: v_mov_b32_e32 v[[VPTRHI:[0-9]+]], s[[PTRHI]] 138; HSA: {{flat|global}}_load_dword v{{[0-9]+}}, v{{\[}}[[VPTRLO]]:[[VPTRHI]]{{\]}} 139define amdgpu_kernel void @use_constant_to_global_addrspacecast(i32 addrspace(4)* %ptr) #0 { 140 %stof = addrspacecast i32 addrspace(4)* %ptr to i32 addrspace(1)* 141 %ld = load volatile i32, i32 addrspace(1)* %stof 142 ret void 143} 144 145; HSA-LABEL: {{^}}use_flat_to_group_addrspacecast: 146; HSA: enable_sgpr_private_segment_buffer = 1 147; HSA: enable_sgpr_dispatch_ptr = 0 148; HSA: enable_sgpr_queue_ptr = 0 149 150; HSA: s_load_dwordx2 s{{\[}}[[PTR_LO:[0-9]+]]:[[PTR_HI:[0-9]+]]{{\]}} 151; HSA-DAG: v_cmp_ne_u64_e64 vcc, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0{{$}} 152; HSA-DAG: v_mov_b32_e32 v[[VPTR_LO:[0-9]+]], s[[PTR_LO]] 153; HSA-DAG: v_cndmask_b32_e32 [[CASTPTR:v[0-9]+]], -1, v[[VPTR_LO]] 154; HSA-DAG: v_mov_b32_e32 v[[K:[0-9]+]], 0{{$}} 155; HSA: ds_write_b32 [[CASTPTR]], v[[K]] 156define amdgpu_kernel void @use_flat_to_group_addrspacecast(i32* %ptr) #0 { 157 %ftos = addrspacecast i32* %ptr to i32 addrspace(3)* 158 store volatile i32 0, i32 addrspace(3)* %ftos 159 ret void 160} 161 162; HSA-LABEL: {{^}}use_flat_to_private_addrspacecast: 163; HSA: enable_sgpr_private_segment_buffer = 1 164; HSA: enable_sgpr_dispatch_ptr = 0 165; HSA: enable_sgpr_queue_ptr = 0 166 167; HSA: s_load_dwordx2 s{{\[}}[[PTR_LO:[0-9]+]]:[[PTR_HI:[0-9]+]]{{\]}} 168; HSA-DAG: v_cmp_ne_u64_e64 vcc, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0{{$}} 169; HSA-DAG: v_mov_b32_e32 v[[VPTR_LO:[0-9]+]], s[[PTR_LO]] 170; HSA-DAG: v_cndmask_b32_e32 [[CASTPTR:v[0-9]+]], -1, v[[VPTR_LO]] 171; HSA-DAG: v_mov_b32_e32 v[[K:[0-9]+]], 0{{$}} 172; HSA: buffer_store_dword v[[K]], [[CASTPTR]], s{{\[[0-9]+:[0-9]+\]}}, 0 offen{{$}} 173define amdgpu_kernel void @use_flat_to_private_addrspacecast(i32* %ptr) #0 { 174 %ftos = addrspacecast i32* %ptr to i32 addrspace(5)* 175 store volatile i32 0, i32 addrspace(5)* %ftos 176 ret void 177} 178 179; HSA-LABEL: {{^}}use_flat_to_global_addrspacecast: 180; HSA: enable_sgpr_queue_ptr = 0 181 182; HSA: s_load_dwordx2 s{{\[}}[[PTRLO:[0-9]+]]:[[PTRHI:[0-9]+]]{{\]}}, s[4:5], 0x0 183; HSA-DAG: v_mov_b32_e32 v[[VPTRLO:[0-9]+]], s[[PTRLO]] 184; HSA-DAG: v_mov_b32_e32 v[[VPTRHI:[0-9]+]], s[[PTRHI]] 185; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 0 186; HSA: {{flat|global}}_store_dword v{{\[}}[[VPTRLO]]:[[VPTRHI]]{{\]}}, [[K]] 187define amdgpu_kernel void @use_flat_to_global_addrspacecast(i32* %ptr) #0 { 188 %ftos = addrspacecast i32* %ptr to i32 addrspace(1)* 189 store volatile i32 0, i32 addrspace(1)* %ftos 190 ret void 191} 192 193; HSA-LABEL: {{^}}use_flat_to_constant_addrspacecast: 194; HSA: enable_sgpr_queue_ptr = 0 195 196; HSA: s_load_dwordx2 s{{\[}}[[PTRLO:[0-9]+]]:[[PTRHI:[0-9]+]]{{\]}}, s[4:5], 0x0 197; HSA: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTRLO]]:[[PTRHI]]{{\]}}, 0x0 198define amdgpu_kernel void @use_flat_to_constant_addrspacecast(i32* %ptr) #0 { 199 %ftos = addrspacecast i32* %ptr to i32 addrspace(4)* 200 load volatile i32, i32 addrspace(4)* %ftos 201 ret void 202} 203 204; HSA-LABEL: {{^}}cast_0_group_to_flat_addrspacecast: 205; CI: s_load_dword [[APERTURE:s[0-9]+]], s[4:5], 0x10 206; CI-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], [[APERTURE]] 207; GFX9-DAG: s_getreg_b32 [[SSRC_SHARED:s[0-9]+]], hwreg(HW_REG_SH_MEM_BASES, 16, 16) 208; GFX9-DAG: s_lshl_b32 [[SSRC_SHARED_BASE:s[0-9]+]], [[SSRC_SHARED]], 16 209; GFX9-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], [[SSRC_SHARED_BASE]] 210 211; GFX9-XXX: v_mov_b32_e32 v[[HI:[0-9]+]], src_shared_base 212 213; HSA-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 214; HSA-DAG: v_mov_b32_e32 v[[K:[0-9]+]], 7{{$}} 215; HSA: {{flat|global}}_store_dword v{{\[}}[[LO]]:[[HI]]{{\]}}, v[[K]] 216define amdgpu_kernel void @cast_0_group_to_flat_addrspacecast() #0 { 217 %cast = addrspacecast i32 addrspace(3)* null to i32* 218 store volatile i32 7, i32* %cast 219 ret void 220} 221 222; HSA-LABEL: {{^}}cast_0_flat_to_group_addrspacecast: 223; HSA-DAG: v_mov_b32_e32 [[PTR:v[0-9]+]], -1{{$}} 224; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7{{$}} 225; HSA: ds_write_b32 [[PTR]], [[K]] 226define amdgpu_kernel void @cast_0_flat_to_group_addrspacecast() #0 { 227 %cast = addrspacecast i32* null to i32 addrspace(3)* 228 store volatile i32 7, i32 addrspace(3)* %cast 229 ret void 230} 231 232; HSA-LABEL: {{^}}cast_neg1_group_to_flat_addrspacecast: 233; HSA: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 234; HSA: v_mov_b32_e32 v[[K:[0-9]+]], 7{{$}} 235; HSA: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}} 236; HSA: {{flat|global}}_store_dword v{{\[}}[[LO]]:[[HI]]{{\]}}, v[[K]] 237define amdgpu_kernel void @cast_neg1_group_to_flat_addrspacecast() #0 { 238 %cast = addrspacecast i32 addrspace(3)* inttoptr (i32 -1 to i32 addrspace(3)*) to i32* 239 store volatile i32 7, i32* %cast 240 ret void 241} 242 243; HSA-LABEL: {{^}}cast_neg1_flat_to_group_addrspacecast: 244; HSA-DAG: v_mov_b32_e32 [[PTR:v[0-9]+]], -1{{$}} 245; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7{{$}} 246; HSA: ds_write_b32 [[PTR]], [[K]] 247define amdgpu_kernel void @cast_neg1_flat_to_group_addrspacecast() #0 { 248 %cast = addrspacecast i32* inttoptr (i64 -1 to i32*) to i32 addrspace(3)* 249 store volatile i32 7, i32 addrspace(3)* %cast 250 ret void 251} 252 253; FIXME: Shouldn't need to enable queue ptr 254; HSA-LABEL: {{^}}cast_0_private_to_flat_addrspacecast: 255; CI: s_load_dword [[APERTURE:s[0-9]+]], s[4:5], 0x11 256; CI-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], [[APERTURE]] 257; GFX9-DAG: s_getreg_b32 [[SSRC_SHARED:s[0-9]+]], hwreg(HW_REG_SH_MEM_BASES, 0, 16) 258; GFX9-DAG: s_lshl_b32 [[SSRC_SHARED_BASE:s[0-9]+]], [[SSRC_SHARED]], 16 259; GFX9-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], [[SSRC_SHARED_BASE]] 260 261; GFX9-XXX: v_mov_b32_e32 v[[HI:[0-9]+]], src_shared_base 262 263; HSA-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 264; HSA-DAG: v_mov_b32_e32 v[[K:[0-9]+]], 7{{$}} 265; HSA: {{flat|global}}_store_dword v{{\[}}[[LO]]:[[HI]]{{\]}}, v[[K]] 266define amdgpu_kernel void @cast_0_private_to_flat_addrspacecast() #0 { 267 %cast = addrspacecast i32 addrspace(5)* null to i32* 268 store volatile i32 7, i32* %cast 269 ret void 270} 271 272; HSA-LABEL: {{^}}cast_0_flat_to_private_addrspacecast: 273; HSA-DAG: v_mov_b32_e32 [[PTR:v[0-9]+]], -1{{$}} 274; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7{{$}} 275; HSA: buffer_store_dword [[K]], [[PTR]], s{{\[[0-9]+:[0-9]+\]}}, 0 276define amdgpu_kernel void @cast_0_flat_to_private_addrspacecast() #0 { 277 %cast = addrspacecast i32* null to i32 addrspace(5)* 278 store volatile i32 7, i32 addrspace(5)* %cast 279 ret void 280} 281 282 283; HSA-LABEL: {{^}}cast_neg1_private_to_flat_addrspacecast: 284; CI: enable_sgpr_queue_ptr = 1 285; GFX9: enable_sgpr_queue_ptr = 0 286 287; HSA-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 288; HSA-DAG: v_mov_b32_e32 v[[K:[0-9]+]], 7{{$}} 289; HSA: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}} 290; HSA: {{flat|global}}_store_dword v{{\[}}[[LO]]:[[HI]]{{\]}}, v[[K]] 291define amdgpu_kernel void @cast_neg1_private_to_flat_addrspacecast() #0 { 292 %cast = addrspacecast i32 addrspace(5)* inttoptr (i32 -1 to i32 addrspace(5)*) to i32* 293 store volatile i32 7, i32* %cast 294 ret void 295} 296 297; HSA-LABEL: {{^}}cast_neg1_flat_to_private_addrspacecast: 298; HSA-DAG: v_mov_b32_e32 [[PTR:v[0-9]+]], -1{{$}} 299; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7{{$}} 300; HSA: buffer_store_dword [[K]], [[PTR]], s{{\[[0-9]+:[0-9]+\]}}, 0 301define amdgpu_kernel void @cast_neg1_flat_to_private_addrspacecast() #0 { 302 %cast = addrspacecast i32* inttoptr (i64 -1 to i32*) to i32 addrspace(5)* 303 store volatile i32 7, i32 addrspace(5)* %cast 304 ret void 305} 306 307 308; Disable optimizations in case there are optimizations added that 309; specialize away generic pointer accesses. 310 311; HSA-LABEL: {{^}}branch_use_flat_i32: 312; HSA: {{flat|global}}_store_dword {{v\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} 313; HSA: s_endpgm 314define amdgpu_kernel void @branch_use_flat_i32(i32 addrspace(1)* noalias %out, i32 addrspace(1)* %gptr, i32 addrspace(3)* %lptr, i32 %x, i32 %c) #0 { 315entry: 316 %cmp = icmp ne i32 %c, 0 317 br i1 %cmp, label %local, label %global 318 319local: 320 %flat_local = addrspacecast i32 addrspace(3)* %lptr to i32* 321 br label %end 322 323global: 324 %flat_global = addrspacecast i32 addrspace(1)* %gptr to i32* 325 br label %end 326 327end: 328 %fptr = phi i32* [ %flat_local, %local ], [ %flat_global, %global ] 329 store volatile i32 %x, i32* %fptr, align 4 330; %val = load i32, i32* %fptr, align 4 331; store i32 %val, i32 addrspace(1)* %out, align 4 332 ret void 333} 334 335; Check for prologue initializing special SGPRs pointing to scratch. 336; HSA-LABEL: {{^}}store_flat_scratch: 337; CI-DAG: s_mov_b32 flat_scratch_lo, s9 338; CI-DAG: s_add_u32 [[ADD:s[0-9]+]], s8, s11 339; CI: s_lshr_b32 flat_scratch_hi, [[ADD]], 8 340 341; GFX9: s_add_u32 flat_scratch_lo, s6, s9 342; GFX9: s_addc_u32 flat_scratch_hi, s7, 0 343 344; HSA: {{flat|global}}_store_dword 345; HSA: s_barrier 346; HSA: {{flat|global}}_load_dword 347define amdgpu_kernel void @store_flat_scratch(i32 addrspace(1)* noalias %out, i32) #0 { 348 %alloca = alloca i32, i32 9, align 4, addrspace(5) 349 %x = call i32 @llvm.amdgcn.workitem.id.x() #2 350 %pptr = getelementptr i32, i32 addrspace(5)* %alloca, i32 %x 351 %fptr = addrspacecast i32 addrspace(5)* %pptr to i32* 352 store volatile i32 %x, i32* %fptr 353 ; Dummy call 354 call void @llvm.amdgcn.s.barrier() #1 355 %reload = load volatile i32, i32* %fptr, align 4 356 store volatile i32 %reload, i32 addrspace(1)* %out, align 4 357 ret void 358} 359 360; HSA-LABEL: {{^}}use_constant_to_constant32_addrspacecast 361; GFX9: s_load_dwordx2 [[PTRPTR:s\[[0-9]+:[0-9]+\]]], s[4:5], 0x0{{$}} 362; GFX9: s_load_dword [[OFFSET:s[0-9]+]], s[4:5], 0x8{{$}} 363; GFX9: s_load_dwordx2 s{{\[}}[[PTR_LO:[0-9]+]]:[[PTR_HI:[0-9]+]]{{\]}}, [[PTRPTR]], 0x0{{$}} 364; GFX9: s_mov_b32 s[[PTR_HI]], 0{{$}} 365; GFX9: s_add_i32 s[[PTR_LO]], s[[PTR_LO]], [[OFFSET]] 366; GFX9: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x0{{$}} 367define amdgpu_kernel void @use_constant_to_constant32_addrspacecast(i8 addrspace(4)* addrspace(4)* %ptr.ptr, i32 %offset) #0 { 368 %ptr = load volatile i8 addrspace(4)*, i8 addrspace(4)* addrspace(4)* %ptr.ptr 369 %addrspacecast = addrspacecast i8 addrspace(4)* %ptr to i8 addrspace(6)* 370 %gep = getelementptr i8, i8 addrspace(6)* %addrspacecast, i32 %offset 371 %ptr.cast = bitcast i8 addrspace(6)* %gep to i32 addrspace(6)* 372 %load = load volatile i32, i32 addrspace(6)* %ptr.cast, align 4 373 ret void 374} 375 376; HSA-LABEL: {{^}}use_global_to_constant32_addrspacecast 377; GFX9: s_load_dwordx2 [[PTRPTR:s\[[0-9]+:[0-9]+\]]], s[4:5], 0x0{{$}} 378; GFX9: s_load_dword [[OFFSET:s[0-9]+]], s[4:5], 0x8{{$}} 379; GFX9: s_load_dwordx2 s{{\[}}[[PTR_LO:[0-9]+]]:[[PTR_HI:[0-9]+]]{{\]}}, [[PTRPTR]], 0x0{{$}} 380; GFX9: s_mov_b32 s[[PTR_HI]], 0{{$}} 381; GFX9: s_add_i32 s[[PTR_LO]], s[[PTR_LO]], [[OFFSET]] 382; GFX9: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x0{{$}} 383define amdgpu_kernel void @use_global_to_constant32_addrspacecast(i8 addrspace(1)* addrspace(4)* %ptr.ptr, i32 %offset) #0 { 384 %ptr = load volatile i8 addrspace(1)*, i8 addrspace(1)* addrspace(4)* %ptr.ptr 385 %addrspacecast = addrspacecast i8 addrspace(1)* %ptr to i8 addrspace(6)* 386 %gep = getelementptr i8, i8 addrspace(6)* %addrspacecast, i32 %offset 387 %ptr.cast = bitcast i8 addrspace(6)* %gep to i32 addrspace(6)* 388 %load = load volatile i32, i32 addrspace(6)* %ptr.cast, align 4 389 ret void 390} 391 392declare void @llvm.amdgcn.s.barrier() #1 393declare i32 @llvm.amdgcn.workitem.id.x() #2 394 395attributes #0 = { nounwind } 396attributes #1 = { nounwind convergent } 397attributes #2 = { nounwind readnone } 398