249 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 62; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes=infer-address-spaces %s | FileCheck %s3 4define amdgpu_kernel void @memset_group_to_flat(ptr addrspace(3) %group.ptr, i32 %y) #0 {5; CHECK-LABEL: define amdgpu_kernel void @memset_group_to_flat(6; CHECK-SAME: ptr addrspace(3) [[GROUP_PTR:%.*]], i32 [[Y:%.*]]) #[[ATTR0:[0-9]+]] {7; CHECK-NEXT: call void @llvm.memset.p3.i64(ptr addrspace(3) align 4 [[GROUP_PTR]], i8 4, i64 32, i1 false), !tbaa [[A_TBAA0:![0-9]+]], !alias.scope [[META3:![0-9]+]], !noalias [[META6:![0-9]+]]8; CHECK-NEXT: ret void9;10 %cast = addrspacecast ptr addrspace(3) %group.ptr to ptr11 call void @llvm.memset.p0.i64(ptr align 4 %cast, i8 4, i64 32, i1 false), !tbaa !0, !alias.scope !3, !noalias !612 ret void13}14 15define amdgpu_kernel void @memset_global_to_flat(ptr addrspace(1) %global.ptr, i32 %y) #0 {16; CHECK-LABEL: define amdgpu_kernel void @memset_global_to_flat(17; CHECK-SAME: ptr addrspace(1) [[GLOBAL_PTR:%.*]], i32 [[Y:%.*]]) #[[ATTR0]] {18; CHECK-NEXT: call void @llvm.memset.p1.i64(ptr addrspace(1) align 4 [[GLOBAL_PTR]], i8 4, i64 32, i1 false), !tbaa [[A_TBAA0]], !alias.scope [[META3]], !noalias [[META6]]19; CHECK-NEXT: ret void20;21 %cast = addrspacecast ptr addrspace(1) %global.ptr to ptr22 call void @llvm.memset.p0.i64(ptr align 4 %cast, i8 4, i64 32, i1 false), !tbaa !0, !alias.scope !3, !noalias !623 ret void24}25 26define amdgpu_kernel void @memset_group_to_flat_no_md(ptr addrspace(3) %group.ptr, i64 %size) #0 {27; CHECK-LABEL: define amdgpu_kernel void @memset_group_to_flat_no_md(28; CHECK-SAME: ptr addrspace(3) [[GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {29; CHECK-NEXT: call void @llvm.memset.p3.i64(ptr addrspace(3) align 4 [[GROUP_PTR]], i8 4, i64 [[SIZE]], i1 false)30; CHECK-NEXT: ret void31;32 %cast = addrspacecast ptr addrspace(3) %group.ptr to ptr33 call void @llvm.memset.p0.i64(ptr align 4 %cast, i8 4, i64 %size, i1 false)34 ret void35}36 37define amdgpu_kernel void @memset_global_to_flat_no_md(ptr addrspace(1) %global.ptr, i64 %size) #0 {38; CHECK-LABEL: define amdgpu_kernel void @memset_global_to_flat_no_md(39; CHECK-SAME: ptr addrspace(1) [[GLOBAL_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {40; CHECK-NEXT: call void @llvm.memset.p1.i64(ptr addrspace(1) align 4 [[GLOBAL_PTR]], i8 4, i64 [[SIZE]], i1 false)41; CHECK-NEXT: ret void42;43 %cast = addrspacecast ptr addrspace(1) %global.ptr to ptr44 call void @llvm.memset.p0.i64(ptr align 4 %cast, i8 4, i64 %size, i1 false)45 ret void46}47 48define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group(ptr %dest, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {49; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group(50; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {51; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)52; CHECK-NEXT: call void @llvm.memcpy.p1.p3.i64(ptr addrspace(1) align 4 [[TMP1]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false), !tbaa [[A_TBAA0]], !alias.scope [[META3]], !noalias [[META6]]53; CHECK-NEXT: ret void54;55 %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr56 call void @llvm.memcpy.p0.p0.i64(ptr align 4 %dest, ptr align 4 %cast.src, i64 %size, i1 false), !tbaa !0, !alias.scope !3, !noalias !657 ret void58}59 60define amdgpu_kernel void @memcpy_inline_flat_to_flat_replace_src_with_group(ptr %dest, ptr addrspace(3) %src.group.ptr) #0 {61; CHECK-LABEL: define amdgpu_kernel void @memcpy_inline_flat_to_flat_replace_src_with_group(62; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]]) #[[ATTR0]] {63; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)64; CHECK-NEXT: call void @llvm.memcpy.inline.p1.p3.i64(ptr addrspace(1) align 4 [[TMP1]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 42, i1 false), !tbaa [[A_TBAA0]], !alias.scope [[META3]], !noalias [[META6]]65; CHECK-NEXT: ret void66;67 %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr68 call void @llvm.memcpy.inline.p0.p0.i64(ptr align 4 %dest, ptr align 4 %cast.src, i64 42, i1 false), !tbaa !0, !alias.scope !3, !noalias !669 ret void70}71 72define amdgpu_kernel void @memcpy_flat_to_flat_replace_dest_with_group(ptr addrspace(3) %dest.group.ptr, ptr %src.ptr, i64 %size) #0 {73; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_dest_with_group(74; CHECK-SAME: ptr addrspace(3) [[DEST_GROUP_PTR:%.*]], ptr [[SRC_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {75; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[SRC_PTR]] to ptr addrspace(1)76; CHECK-NEXT: call void @llvm.memcpy.p3.p1.i64(ptr addrspace(3) align 4 [[DEST_GROUP_PTR]], ptr addrspace(1) align 4 [[TMP1]], i64 [[SIZE]], i1 false), !tbaa [[A_TBAA0]], !alias.scope [[META3]], !noalias [[META6]]77; CHECK-NEXT: ret void78;79 %cast.dest = addrspacecast ptr addrspace(3) %dest.group.ptr to ptr80 call void @llvm.memcpy.p0.p0.i64(ptr align 4 %cast.dest, ptr align 4 %src.ptr, i64 %size, i1 false), !tbaa !0, !alias.scope !3, !noalias !681 ret void82}83 84define amdgpu_kernel void @memcpy_flat_to_flat_replace_dest_src_with_group(ptr addrspace(3) %dest.group.ptr, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {85; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_dest_src_with_group(86; CHECK-SAME: ptr addrspace(3) [[DEST_GROUP_PTR:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {87; CHECK-NEXT: call void @llvm.memcpy.p3.p3.i64(ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false), !tbaa [[A_TBAA0]], !alias.scope [[META3]], !noalias [[META6]]88; CHECK-NEXT: ret void89;90 %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr91 %cast.dest = addrspacecast ptr addrspace(3) %src.group.ptr to ptr92 call void @llvm.memcpy.p0.p0.i64(ptr align 4 %cast.dest, ptr align 4 %cast.src, i64 %size, i1 false), !tbaa !0, !alias.scope !3, !noalias !693 ret void94}95 96define amdgpu_kernel void @memcpy_flat_to_flat_replace_dest_group_src_global(ptr addrspace(3) %dest.group.ptr, ptr addrspace(1) %src.global.ptr, i64 %size) #0 {97; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_dest_group_src_global(98; CHECK-SAME: ptr addrspace(3) [[DEST_GROUP_PTR:%.*]], ptr addrspace(1) [[SRC_GLOBAL_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {99; CHECK-NEXT: call void @llvm.memcpy.p3.p1.i64(ptr addrspace(3) align 4 [[DEST_GROUP_PTR]], ptr addrspace(1) align 4 [[SRC_GLOBAL_PTR]], i64 [[SIZE]], i1 false), !tbaa [[A_TBAA0]], !alias.scope [[META3]], !noalias [[META6]]100; CHECK-NEXT: ret void101;102 %cast.src = addrspacecast ptr addrspace(1) %src.global.ptr to ptr103 %cast.dest = addrspacecast ptr addrspace(3) %dest.group.ptr to ptr104 call void @llvm.memcpy.p0.p0.i64(ptr align 4 %cast.dest, ptr align 4 %cast.src, i64 %size, i1 false), !tbaa !0, !alias.scope !3, !noalias !6105 ret void106}107 108define amdgpu_kernel void @memcpy_group_to_flat_replace_dest_global(ptr addrspace(1) %dest.global.ptr, ptr addrspace(3) %src.group.ptr, i32 %size) #0 {109; CHECK-LABEL: define amdgpu_kernel void @memcpy_group_to_flat_replace_dest_global(110; CHECK-SAME: ptr addrspace(1) [[DEST_GLOBAL_PTR:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i32 [[SIZE:%.*]]) #[[ATTR0]] {111; CHECK-NEXT: call void @llvm.memcpy.p1.p3.i32(ptr addrspace(1) align 4 [[DEST_GLOBAL_PTR]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i32 [[SIZE]], i1 false), !tbaa [[A_TBAA0]], !alias.scope [[META3]], !noalias [[META6]]112; CHECK-NEXT: ret void113;114 %cast.dest = addrspacecast ptr addrspace(1) %dest.global.ptr to ptr115 call void @llvm.memcpy.p0.p3.i32(ptr align 4 %cast.dest, ptr addrspace(3) align 4 %src.group.ptr, i32 %size, i1 false), !tbaa !0, !alias.scope !3, !noalias !6116 ret void117}118 119define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_tbaa_struct(ptr %dest, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {120; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_tbaa_struct(121; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {122; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)123; CHECK-NEXT: call void @llvm.memcpy.p1.p3.i64(ptr addrspace(1) align 4 [[TMP1]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false), !tbaa.struct [[TBAA_STRUCT8:![0-9]+]]124; CHECK-NEXT: ret void125;126 %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr127 call void @llvm.memcpy.p0.p0.i64(ptr align 4 %dest, ptr align 4 %cast.src, i64 %size, i1 false), !tbaa.struct !8128 ret void129}130 131define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_no_md(ptr %dest, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {132; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_no_md(133; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {134; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)135; CHECK-NEXT: call void @llvm.memcpy.p1.p3.i64(ptr addrspace(1) align 4 [[TMP1]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false)136; CHECK-NEXT: ret void137;138 %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr139 call void @llvm.memcpy.p0.p0.i64(ptr align 4 %dest, ptr align 4 %cast.src, i64 %size, i1 false)140 ret void141}142 143define amdgpu_kernel void @multiple_memcpy_flat_to_flat_replace_src_with_group_no_md(ptr %dest0, ptr %dest1, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {144; CHECK-LABEL: define amdgpu_kernel void @multiple_memcpy_flat_to_flat_replace_src_with_group_no_md(145; CHECK-SAME: ptr [[DEST0:%.*]], ptr [[DEST1:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {146; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[DEST0]] to ptr addrspace(1)147; CHECK-NEXT: [[TMP2:%.*]] = addrspacecast ptr [[DEST1]] to ptr addrspace(1)148; CHECK-NEXT: call void @llvm.memcpy.p1.p3.i64(ptr addrspace(1) align 4 [[TMP1]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false)149; CHECK-NEXT: call void @llvm.memcpy.p1.p3.i64(ptr addrspace(1) align 4 [[TMP2]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false)150; CHECK-NEXT: ret void151;152 %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr153 call void @llvm.memcpy.p0.p0.i64(ptr align 4 %dest0, ptr align 4 %cast.src, i64 %size, i1 false)154 call void @llvm.memcpy.p0.p0.i64(ptr align 4 %dest1, ptr align 4 %cast.src, i64 %size, i1 false)155 ret void156}157 158; Check for iterator problems if the pointer has 2 uses in the same call159define amdgpu_kernel void @memcpy_group_flat_to_flat_self(ptr addrspace(3) %group.ptr) #0 {160; CHECK-LABEL: define amdgpu_kernel void @memcpy_group_flat_to_flat_self(161; CHECK-SAME: ptr addrspace(3) [[GROUP_PTR:%.*]]) #[[ATTR0]] {162; CHECK-NEXT: call void @llvm.memcpy.p3.p3.i64(ptr addrspace(3) align 4 [[GROUP_PTR]], ptr addrspace(3) align 4 [[GROUP_PTR]], i64 32, i1 false), !tbaa [[A_TBAA0]], !alias.scope [[META3]], !noalias [[META6]]163; CHECK-NEXT: ret void164;165 %cast = addrspacecast ptr addrspace(3) %group.ptr to ptr166 call void @llvm.memcpy.p0.p0.i64(ptr align 4 %cast, ptr align 4 %cast, i64 32, i1 false), !tbaa !0, !alias.scope !3, !noalias !6167 ret void168}169define amdgpu_kernel void @memmove_flat_to_flat_replace_src_with_group(ptr %dest, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {170; CHECK-LABEL: define amdgpu_kernel void @memmove_flat_to_flat_replace_src_with_group(171; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {172; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)173; CHECK-NEXT: call void @llvm.memmove.p1.p3.i64(ptr addrspace(1) align 4 [[TMP1]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false), !tbaa [[A_TBAA0]], !alias.scope [[META3]], !noalias [[META6]]174; CHECK-NEXT: ret void175;176 %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr177 call void @llvm.memmove.p0.p0.i64(ptr align 4 %dest, ptr align 4 %cast.src, i64 %size, i1 false), !tbaa !0, !alias.scope !3, !noalias !6178 ret void179}180 181define amdgpu_kernel void @load_to_lds_global_as_flat(ptr addrspace(1) %global.ptr, ptr addrspace(3) %group.ptr) #0 {182; CHECK-LABEL: define amdgpu_kernel void @load_to_lds_global_as_flat(183; CHECK-SAME: ptr addrspace(1) [[GLOBAL_PTR:%.*]], ptr addrspace(3) [[GROUP_PTR:%.*]]) #[[ATTR0]] {184; CHECK-NEXT: call void @llvm.amdgcn.load.to.lds.p1(ptr addrspace(1) [[GLOBAL_PTR]], ptr addrspace(3) [[GROUP_PTR]], i32 4, i32 0, i32 0)185; CHECK-NEXT: ret void186;187 %cast = addrspacecast ptr addrspace(1) %global.ptr to ptr188 call void @llvm.amdgcn.load.to.lds.p0(ptr %cast, ptr addrspace(3) %group.ptr, i32 4, i32 0, i32 0)189 ret void190}191 192define amdgpu_kernel void @load_to_lds_fat_pointer_as_flat(ptr addrspace(7) %buffer.fat.ptr, ptr addrspace(3) %group.ptr) #0 {193; CHECK-LABEL: define amdgpu_kernel void @load_to_lds_fat_pointer_as_flat(194; CHECK-SAME: ptr addrspace(7) [[BUFFER_FAT_PTR:%.*]], ptr addrspace(3) [[GROUP_PTR:%.*]]) #[[ATTR0]] {195; CHECK-NEXT: call void @llvm.amdgcn.load.to.lds.p7(ptr addrspace(7) [[BUFFER_FAT_PTR]], ptr addrspace(3) [[GROUP_PTR]], i32 4, i32 0, i32 0)196; CHECK-NEXT: ret void197;198 %cast = addrspacecast ptr addrspace(7) %buffer.fat.ptr to ptr199 call void @llvm.amdgcn.load.to.lds.p0(ptr %cast, ptr addrspace(3) %group.ptr, i32 4, i32 0, i32 0)200 ret void201}202 203define amdgpu_kernel void @make_buffer_rsrc_global_as_flat(ptr addrspace(1) %global, i64 %extent) {204 ;; NOTE: flags value not representative of real input205; CHECK-LABEL: define amdgpu_kernel void @make_buffer_rsrc_global_as_flat(206; CHECK-SAME: ptr addrspace(1) [[GLOBAL:%.*]], i64 [[EXTENT:%.*]]) {207; CHECK-NEXT: [[BUFFER_FAT_PTR:%.*]] = call ptr addrspace(7) @llvm.amdgcn.make.buffer.rsrc.p7.p1(ptr addrspace(1) [[GLOBAL]], i16 0, i64 [[EXTENT]], i32 0)208; CHECK-NEXT: store i64 [[EXTENT]], ptr addrspace(7) [[BUFFER_FAT_PTR]], align 8209; CHECK-NEXT: ret void210;211 %cast = addrspacecast ptr addrspace(1) %global to ptr212 %buffer.fat.ptr = call ptr addrspace(7) @llvm.amdgcn.make.buffer.rsrc.p7.p0(ptr %cast, i16 0, i64 %extent, i32 0)213 store i64 %extent, ptr addrspace(7) %buffer.fat.ptr214 ret void215}216 217declare void @llvm.memset.p0.i64(ptr nocapture writeonly, i8, i64, i1) #1218declare void @llvm.memcpy.p0.p0.i64(ptr nocapture writeonly, ptr nocapture readonly, i64, i1) #1219declare void @llvm.memcpy.inline.p0.p0.i64(ptr nocapture writeonly, ptr nocapture readonly, i64, i1) #1220declare void @llvm.memcpy.p0.p3.i32(ptr nocapture writeonly, ptr addrspace(3) nocapture readonly, i32, i1) #1221declare void @llvm.memmove.p0.p0.i64(ptr nocapture writeonly, ptr nocapture readonly, i64, i1) #1222declare void @llvm.amdgcn.load.to.lds.p0(ptr nocapture readonly, ptr addrspace(3) nocapture writeonly, i32 immarg, i32 immarg, i32 immarg) #1223 224declare ptr addrspace(7) @llvm.amdgcn.make.buffer.rsrc.p7.p0(ptr readnone, i16, i64, i32) #0225 226attributes #0 = { nounwind }227attributes #1 = { argmemonly nounwind }228 229!0 = !{!1, !1, i64 0}230!1 = !{!"A", !2}231!2 = !{!"tbaa root"}232!3 = !{!4}233!4 = distinct !{!4, !5, !"some scope 1"}234!5 = distinct !{!5, !"some domain"}235!6 = !{!7}236!7 = distinct !{!7, !5, !"some scope 2"}237!8 = !{i64 0, i64 8, null}238;.239; CHECK: [[A_TBAA0]] = !{[[META1:![0-9]+]], [[META1]], i64 0}240; CHECK: [[META1]] = !{!"A", [[META2:![0-9]+]]}241; CHECK: [[META2]] = !{!"tbaa root"}242; CHECK: [[META3]] = !{[[META4:![0-9]+]]}243; CHECK: [[META4]] = distinct !{[[META4]], [[META5:![0-9]+]], !"some scope 1"}244; CHECK: [[META5]] = distinct !{[[META5]], !"some domain"}245; CHECK: [[META6]] = !{[[META7:![0-9]+]]}246; CHECK: [[META7]] = distinct !{[[META7]], [[META5]], !"some scope 2"}247; CHECK: [[TBAA_STRUCT8]] = !{i64 0, i64 8, null}248;.249