267 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --prefix-filecheck-ir-name _ --version 52; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes=infer-address-spaces %s | FileCheck %s3 4; Trivial optimization of generic addressing5 6define float @load_global_from_flat(ptr %generic_scalar) #0 {7; CHECK-LABEL: define float @load_global_from_flat(8; CHECK-SAME: ptr [[GENERIC_SCALAR:%.*]]) #[[ATTR0:[0-9]+]] {9; CHECK-NEXT: [[_TMP0:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(1)10; CHECK-NEXT: [[_TMP1:%.*]] = load float, ptr addrspace(1) [[_TMP0]], align 411; CHECK-NEXT: ret float [[_TMP1]]12;13 %tmp0 = addrspacecast ptr %generic_scalar to ptr addrspace(1)14 %tmp1 = load float, ptr addrspace(1) %tmp015 ret float %tmp116}17 18define float @load_constant_from_flat(ptr %generic_scalar) #0 {19; CHECK-LABEL: define float @load_constant_from_flat(20; CHECK-SAME: ptr [[GENERIC_SCALAR:%.*]]) #[[ATTR0]] {21; CHECK-NEXT: [[_TMP0:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(4)22; CHECK-NEXT: [[_TMP1:%.*]] = load float, ptr addrspace(4) [[_TMP0]], align 423; CHECK-NEXT: ret float [[_TMP1]]24;25 %tmp0 = addrspacecast ptr %generic_scalar to ptr addrspace(4)26 %tmp1 = load float, ptr addrspace(4) %tmp027 ret float %tmp128}29 30define float @load_group_from_flat(ptr %generic_scalar) #0 {31; CHECK-LABEL: define float @load_group_from_flat(32; CHECK-SAME: ptr [[GENERIC_SCALAR:%.*]]) #[[ATTR0]] {33; CHECK-NEXT: [[_TMP0:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(3)34; CHECK-NEXT: [[_TMP1:%.*]] = load float, ptr addrspace(3) [[_TMP0]], align 435; CHECK-NEXT: ret float [[_TMP1]]36;37 %tmp0 = addrspacecast ptr %generic_scalar to ptr addrspace(3)38 %tmp1 = load float, ptr addrspace(3) %tmp039 ret float %tmp140}41 42define float @load_private_from_flat(ptr %generic_scalar) #0 {43; CHECK-LABEL: define float @load_private_from_flat(44; CHECK-SAME: ptr [[GENERIC_SCALAR:%.*]]) #[[ATTR0]] {45; CHECK-NEXT: [[_TMP0:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(5)46; CHECK-NEXT: [[_TMP1:%.*]] = load float, ptr addrspace(5) [[_TMP0]], align 447; CHECK-NEXT: ret float [[_TMP1]]48;49 %tmp0 = addrspacecast ptr %generic_scalar to ptr addrspace(5)50 %tmp1 = load float, ptr addrspace(5) %tmp051 ret float %tmp152}53 54define amdgpu_kernel void @store_global_from_flat(ptr %generic_scalar) #0 {55; CHECK-LABEL: define amdgpu_kernel void @store_global_from_flat(56; CHECK-SAME: ptr [[GENERIC_SCALAR:%.*]]) #[[ATTR0]] {57; CHECK-NEXT: [[_TMP0:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(1)58; CHECK-NEXT: store float 0.000000e+00, ptr addrspace(1) [[_TMP0]], align 459; CHECK-NEXT: ret void60;61 %tmp0 = addrspacecast ptr %generic_scalar to ptr addrspace(1)62 store float 0.0, ptr addrspace(1) %tmp063 ret void64}65 66define amdgpu_kernel void @store_group_from_flat(ptr %generic_scalar) #0 {67; CHECK-LABEL: define amdgpu_kernel void @store_group_from_flat(68; CHECK-SAME: ptr [[GENERIC_SCALAR:%.*]]) #[[ATTR0]] {69; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(1)70; CHECK-NEXT: [[TMP2:%.*]] = addrspacecast ptr addrspace(1) [[TMP1]] to ptr71; CHECK-NEXT: [[_TMP0:%.*]] = addrspacecast ptr [[TMP2]] to ptr addrspace(3)72; CHECK-NEXT: store float 0.000000e+00, ptr addrspace(3) [[_TMP0]], align 473; CHECK-NEXT: ret void74;75 %tmp0 = addrspacecast ptr %generic_scalar to ptr addrspace(3)76 store float 0.0, ptr addrspace(3) %tmp077 ret void78}79 80define amdgpu_kernel void @store_private_from_flat(ptr %generic_scalar) #0 {81; CHECK-LABEL: define amdgpu_kernel void @store_private_from_flat(82; CHECK-SAME: ptr [[GENERIC_SCALAR:%.*]]) #[[ATTR0]] {83; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(1)84; CHECK-NEXT: [[TMP2:%.*]] = addrspacecast ptr addrspace(1) [[TMP1]] to ptr85; CHECK-NEXT: [[_TMP0:%.*]] = addrspacecast ptr [[TMP2]] to ptr addrspace(5)86; CHECK-NEXT: store float 0.000000e+00, ptr addrspace(5) [[_TMP0]], align 487; CHECK-NEXT: ret void88;89 %tmp0 = addrspacecast ptr %generic_scalar to ptr addrspace(5)90 store float 0.0, ptr addrspace(5) %tmp091 ret void92}93 94; optimized to global load/store.95define amdgpu_kernel void @load_store_global(ptr addrspace(1) nocapture %input, ptr addrspace(1) nocapture %output) #0 {96; CHECK-LABEL: define amdgpu_kernel void @load_store_global(97; CHECK-SAME: ptr addrspace(1) captures(none) [[INPUT:%.*]], ptr addrspace(1) captures(none) [[OUTPUT:%.*]]) #[[ATTR0]] {98; CHECK-NEXT: [[VAL:%.*]] = load i32, ptr addrspace(1) [[INPUT]], align 499; CHECK-NEXT: store i32 [[VAL]], ptr addrspace(1) [[OUTPUT]], align 4100; CHECK-NEXT: ret void101;102 %tmp0 = addrspacecast ptr addrspace(1) %input to ptr103 %tmp1 = addrspacecast ptr addrspace(1) %output to ptr104 %val = load i32, ptr %tmp0, align 4105 store i32 %val, ptr %tmp1, align 4106 ret void107}108 109; Optimized to group load/store.110define amdgpu_kernel void @load_store_group(ptr addrspace(3) nocapture %input, ptr addrspace(3) nocapture %output) #0 {111; CHECK-LABEL: define amdgpu_kernel void @load_store_group(112; CHECK-SAME: ptr addrspace(3) captures(none) [[INPUT:%.*]], ptr addrspace(3) captures(none) [[OUTPUT:%.*]]) #[[ATTR0]] {113; CHECK-NEXT: [[VAL:%.*]] = load i32, ptr addrspace(3) [[INPUT]], align 4114; CHECK-NEXT: store i32 [[VAL]], ptr addrspace(3) [[OUTPUT]], align 4115; CHECK-NEXT: ret void116;117 %tmp0 = addrspacecast ptr addrspace(3) %input to ptr118 %tmp1 = addrspacecast ptr addrspace(3) %output to ptr119 %val = load i32, ptr %tmp0, align 4120 store i32 %val, ptr %tmp1, align 4121 ret void122}123 124; Optimized to private load/store.125define amdgpu_kernel void @load_store_private(ptr addrspace(5) nocapture %input, ptr addrspace(5) nocapture %output) #0 {126; CHECK-LABEL: define amdgpu_kernel void @load_store_private(127; CHECK-SAME: ptr addrspace(5) captures(none) [[INPUT:%.*]], ptr addrspace(5) captures(none) [[OUTPUT:%.*]]) #[[ATTR0]] {128; CHECK-NEXT: [[VAL:%.*]] = load i32, ptr addrspace(5) [[INPUT]], align 4129; CHECK-NEXT: store i32 [[VAL]], ptr addrspace(5) [[OUTPUT]], align 4130; CHECK-NEXT: ret void131;132 %tmp0 = addrspacecast ptr addrspace(5) %input to ptr133 %tmp1 = addrspacecast ptr addrspace(5) %output to ptr134 %val = load i32, ptr %tmp0, align 4135 store i32 %val, ptr %tmp1, align 4136 ret void137}138 139; No optimization. flat load/store.140define amdgpu_kernel void @load_store_flat(ptr nocapture %input, ptr nocapture %output) #0 {141; CHECK-LABEL: define amdgpu_kernel void @load_store_flat(142; CHECK-SAME: ptr captures(none) [[INPUT:%.*]], ptr captures(none) [[OUTPUT:%.*]]) #[[ATTR0]] {143; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[INPUT]] to ptr addrspace(1)144; CHECK-NEXT: [[TMP2:%.*]] = addrspacecast ptr [[OUTPUT]] to ptr addrspace(1)145; CHECK-NEXT: [[VAL:%.*]] = load i32, ptr addrspace(1) [[TMP1]], align 4146; CHECK-NEXT: store i32 [[VAL]], ptr addrspace(1) [[TMP2]], align 4147; CHECK-NEXT: ret void148;149 %val = load i32, ptr %input, align 4150 store i32 %val, ptr %output, align 4151 ret void152}153 154define amdgpu_kernel void @store_addrspacecast_ptr_value(ptr addrspace(1) nocapture %input, ptr addrspace(1) nocapture %output) #0 {155; CHECK-LABEL: define amdgpu_kernel void @store_addrspacecast_ptr_value(156; CHECK-SAME: ptr addrspace(1) captures(none) [[INPUT:%.*]], ptr addrspace(1) captures(none) [[OUTPUT:%.*]]) #[[ATTR0]] {157; CHECK-NEXT: [[CAST:%.*]] = addrspacecast ptr addrspace(1) [[INPUT]] to ptr158; CHECK-NEXT: store ptr [[CAST]], ptr addrspace(1) [[OUTPUT]], align 4159; CHECK-NEXT: ret void160;161 %cast = addrspacecast ptr addrspace(1) %input to ptr162 store ptr %cast, ptr addrspace(1) %output, align 4163 ret void164}165 166define i32 @atomicrmw_add_global_to_flat(ptr addrspace(1) %global.ptr, i32 %y) #0 {167; CHECK-LABEL: define i32 @atomicrmw_add_global_to_flat(168; CHECK-SAME: ptr addrspace(1) [[GLOBAL_PTR:%.*]], i32 [[Y:%.*]]) #[[ATTR0]] {169; CHECK-NEXT: [[RET:%.*]] = atomicrmw add ptr addrspace(1) [[GLOBAL_PTR]], i32 [[Y]] seq_cst, align 4170; CHECK-NEXT: ret i32 [[RET]]171;172 %cast = addrspacecast ptr addrspace(1) %global.ptr to ptr173 %ret = atomicrmw add ptr %cast, i32 %y seq_cst174 ret i32 %ret175}176 177define i32 @atomicrmw_add_group_to_flat(ptr addrspace(3) %group.ptr, i32 %y) #0 {178; CHECK-LABEL: define i32 @atomicrmw_add_group_to_flat(179; CHECK-SAME: ptr addrspace(3) [[GROUP_PTR:%.*]], i32 [[Y:%.*]]) #[[ATTR0]] {180; CHECK-NEXT: [[RET:%.*]] = atomicrmw add ptr addrspace(3) [[GROUP_PTR]], i32 [[Y]] seq_cst, align 4181; CHECK-NEXT: ret i32 [[RET]]182;183 %cast = addrspacecast ptr addrspace(3) %group.ptr to ptr184 %ret = atomicrmw add ptr %cast, i32 %y seq_cst185 ret i32 %ret186}187 188define { i32, i1 } @cmpxchg_global_to_flat(ptr addrspace(1) %global.ptr, i32 %cmp, i32 %val) #0 {189; CHECK-LABEL: define { i32, i1 } @cmpxchg_global_to_flat(190; CHECK-SAME: ptr addrspace(1) [[GLOBAL_PTR:%.*]], i32 [[CMP:%.*]], i32 [[VAL:%.*]]) #[[ATTR0]] {191; CHECK-NEXT: [[RET:%.*]] = cmpxchg ptr addrspace(1) [[GLOBAL_PTR]], i32 [[CMP]], i32 [[VAL]] seq_cst monotonic, align 4192; CHECK-NEXT: ret { i32, i1 } [[RET]]193;194 %cast = addrspacecast ptr addrspace(1) %global.ptr to ptr195 %ret = cmpxchg ptr %cast, i32 %cmp, i32 %val seq_cst monotonic196 ret { i32, i1 } %ret197}198 199define { i32, i1 } @cmpxchg_group_to_flat(ptr addrspace(3) %group.ptr, i32 %cmp, i32 %val) #0 {200; CHECK-LABEL: define { i32, i1 } @cmpxchg_group_to_flat(201; CHECK-SAME: ptr addrspace(3) [[GROUP_PTR:%.*]], i32 [[CMP:%.*]], i32 [[VAL:%.*]]) #[[ATTR0]] {202; CHECK-NEXT: [[RET:%.*]] = cmpxchg ptr addrspace(3) [[GROUP_PTR]], i32 [[CMP]], i32 [[VAL]] seq_cst monotonic, align 4203; CHECK-NEXT: ret { i32, i1 } [[RET]]204;205 %cast = addrspacecast ptr addrspace(3) %group.ptr to ptr206 %ret = cmpxchg ptr %cast, i32 %cmp, i32 %val seq_cst monotonic207 ret { i32, i1 } %ret208}209 210; Not pointer operand211define { ptr, i1 } @cmpxchg_group_to_flat_wrong_operand(ptr addrspace(3) %cas.ptr, ptr addrspace(3) %cmp.ptr, ptr %val) #0 {212; CHECK-LABEL: define { ptr, i1 } @cmpxchg_group_to_flat_wrong_operand(213; CHECK-SAME: ptr addrspace(3) [[CAS_PTR:%.*]], ptr addrspace(3) [[CMP_PTR:%.*]], ptr [[VAL:%.*]]) #[[ATTR0]] {214; CHECK-NEXT: [[CAST_CMP:%.*]] = addrspacecast ptr addrspace(3) [[CMP_PTR]] to ptr215; CHECK-NEXT: [[RET:%.*]] = cmpxchg ptr addrspace(3) [[CAS_PTR]], ptr [[CAST_CMP]], ptr [[VAL]] seq_cst monotonic, align 8216; CHECK-NEXT: ret { ptr, i1 } [[RET]]217;218 %cast.cmp = addrspacecast ptr addrspace(3) %cmp.ptr to ptr219 %ret = cmpxchg ptr addrspace(3) %cas.ptr, ptr %cast.cmp, ptr %val seq_cst monotonic220 ret { ptr, i1 } %ret221}222 223; Null pointer in local addr space224define void @local_nullptr(ptr addrspace(1) nocapture %results, ptr addrspace(3) %a) {225; CHECK-LABEL: define void @local_nullptr(226; CHECK-SAME: ptr addrspace(1) captures(none) [[RESULTS:%.*]], ptr addrspace(3) [[A:%.*]]) {227; CHECK-NEXT: [[ENTRY:.*:]]228; CHECK-NEXT: [[TOBOOL:%.*]] = icmp ne ptr addrspace(3) [[A]], addrspacecast (ptr addrspace(5) null to ptr addrspace(3))229; CHECK-NEXT: [[CONV:%.*]] = zext i1 [[TOBOOL]] to i32230; CHECK-NEXT: store i32 [[CONV]], ptr addrspace(1) [[RESULTS]], align 4231; CHECK-NEXT: ret void232;233entry:234 %tobool = icmp ne ptr addrspace(3) %a, addrspacecast (ptr addrspace(5) null to ptr addrspace(3))235 %conv = zext i1 %tobool to i32236 store i32 %conv, ptr addrspace(1) %results, align 4237 ret void238}239 240define i32 @atomicrmw_add_global_to_flat_preserve_amdgpu_md(ptr addrspace(1) %global.ptr, i32 %y) #0 {241; CHECK-LABEL: define i32 @atomicrmw_add_global_to_flat_preserve_amdgpu_md(242; CHECK-SAME: ptr addrspace(1) [[GLOBAL_PTR:%.*]], i32 [[Y:%.*]]) #[[ATTR0]] {243; CHECK-NEXT: [[RET:%.*]] = atomicrmw add ptr addrspace(1) [[GLOBAL_PTR]], i32 [[Y]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]], !amdgpu.no.remote.memory [[META0]]244; CHECK-NEXT: ret i32 [[RET]]245;246 %cast = addrspacecast ptr addrspace(1) %global.ptr to ptr247 %ret = atomicrmw add ptr %cast, i32 %y seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0248 ret i32 %ret249}250 251; Make sure there's no assert252define ptr @try_infer_getelementptr_constant_null() {253; CHECK-LABEL: define ptr @try_infer_getelementptr_constant_null() {254; CHECK-NEXT: [[CE:%.*]] = getelementptr i8, ptr getelementptr inbounds (i8, ptr null, i64 8), i64 0255; CHECK-NEXT: ret ptr [[CE]]256;257 %ce = getelementptr i8, ptr getelementptr inbounds (i8, ptr null, i64 8), i64 0258 ret ptr %ce259}260 261attributes #0 = { nounwind }262 263!0 = !{}264;.265; CHECK: [[META0]] = !{}266;.267