Reapply "[AMDGPU] Make `getAssumedAddrSpace` return AS1 for pointer kernel arguments (#137488)"

This reverts commit 3c6211c183885afb5d89259a53c4f4f46a6bf399.

2 files changed, 26 insertions, 12 deletions

diff --git a/llvm/test/Transforms/InferAddressSpaces/AMDGPU/basic.ll b/llvm/test/Transforms/InferAddressSpaces/AMDGPU/basic.ll
index 99fe986cf637..60bb38f863e8 100644
--- a/llvm/test/Transforms/InferAddressSpaces/AMDGPU/basic.ll
+++ b/llvm/test/Transforms/InferAddressSpaces/AMDGPU/basic.ll
@@ -66,7 +66,9 @@ define amdgpu_kernel void @store_global_from_flat(ptr %generic_scalar) #0 {
 define amdgpu_kernel void @store_group_from_flat(ptr %generic_scalar) #0 {
 ; CHECK-LABEL: define amdgpu_kernel void @store_group_from_flat(
 ; CHECK-SAME: ptr [[GENERIC_SCALAR:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    [[_TMP0:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(3)
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(1)
+; CHECK-NEXT:    [[TMP2:%.*]] = addrspacecast ptr addrspace(1) [[TMP1]] to ptr
+; CHECK-NEXT:    [[_TMP0:%.*]] = addrspacecast ptr [[TMP2]] to ptr addrspace(3)
 ; CHECK-NEXT:    store float 0.000000e+00, ptr addrspace(3) [[_TMP0]], align 4
 ; CHECK-NEXT:    ret void
 ;
@@ -78,7 +80,9 @@ define amdgpu_kernel void @store_group_from_flat(ptr %generic_scalar) #0 {
 define amdgpu_kernel void @store_private_from_flat(ptr %generic_scalar) #0 {
 ; CHECK-LABEL: define amdgpu_kernel void @store_private_from_flat(
 ; CHECK-SAME: ptr [[GENERIC_SCALAR:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    [[_TMP0:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(5)
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[GENERIC_SCALAR]] to ptr addrspace(1)
+; CHECK-NEXT:    [[TMP2:%.*]] = addrspacecast ptr addrspace(1) [[TMP1]] to ptr
+; CHECK-NEXT:    [[_TMP0:%.*]] = addrspacecast ptr [[TMP2]] to ptr addrspace(5)
 ; CHECK-NEXT:    store float 0.000000e+00, ptr addrspace(5) [[_TMP0]], align 4
 ; CHECK-NEXT:    ret void
 ;
@@ -136,8 +140,10 @@ define amdgpu_kernel void @load_store_private(ptr addrspace(5) nocapture %input,
 define amdgpu_kernel void @load_store_flat(ptr nocapture %input, ptr nocapture %output) #0 {
 ; CHECK-LABEL: define amdgpu_kernel void @load_store_flat(
 ; CHECK-SAME: ptr captures(none) [[INPUT:%.*]], ptr captures(none) [[OUTPUT:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    [[VAL:%.*]] = load i32, ptr [[INPUT]], align 4
-; CHECK-NEXT:    store i32 [[VAL]], ptr [[OUTPUT]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[INPUT]] to ptr addrspace(1)
+; CHECK-NEXT:    [[TMP2:%.*]] = addrspacecast ptr [[OUTPUT]] to ptr addrspace(1)
+; CHECK-NEXT:    [[VAL:%.*]] = load i32, ptr addrspace(1) [[TMP1]], align 4
+; CHECK-NEXT:    store i32 [[VAL]], ptr addrspace(1) [[TMP2]], align 4
 ; CHECK-NEXT:    ret void
 ;
   %val = load i32, ptr %input, align 4
diff --git a/llvm/test/Transforms/InferAddressSpaces/AMDGPU/mem-intrinsics.ll b/llvm/test/Transforms/InferAddressSpaces/AMDGPU/mem-intrinsics.ll
index 57453d63d7e8..1c317786d1c2 100644
--- a/llvm/test/Transforms/InferAddressSpaces/AMDGPU/mem-intrinsics.ll
+++ b/llvm/test/Transforms/InferAddressSpaces/AMDGPU/mem-intrinsics.ll
@@ -48,7 +48,8 @@ define amdgpu_kernel void @memset_global_to_flat_no_md(ptr addrspace(1) %global.
 define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group(ptr %dest, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {
 ; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group(
 ; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p3.i64(ptr align 4 [[DEST]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false), !tbaa [[TBAA0]], !alias.scope [[META3]], !noalias [[META6]]
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)
+; 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 [[TBAA0]], !alias.scope [[META3]], !noalias [[META6]]
 ; CHECK-NEXT:    ret void
 ;
   %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr
@@ -59,7 +60,8 @@ define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group(ptr %dest,
 define amdgpu_kernel void @memcpy_inline_flat_to_flat_replace_src_with_group(ptr %dest, ptr addrspace(3) %src.group.ptr) #0 {
 ; CHECK-LABEL: define amdgpu_kernel void @memcpy_inline_flat_to_flat_replace_src_with_group(
 ; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    call void @llvm.memcpy.inline.p0.p3.i64(ptr align 4 [[DEST]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 42, i1 false), !tbaa [[TBAA0]], !alias.scope [[META3]], !noalias [[META6]]
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)
+; 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 [[TBAA0]], !alias.scope [[META3]], !noalias [[META6]]
 ; CHECK-NEXT:    ret void
 ;
   %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr
@@ -70,7 +72,8 @@ define amdgpu_kernel void @memcpy_inline_flat_to_flat_replace_src_with_group(ptr
 define amdgpu_kernel void @memcpy_flat_to_flat_replace_dest_with_group(ptr addrspace(3) %dest.group.ptr, ptr %src.ptr, i64 %size) #0 {
 ; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_dest_with_group(
 ; CHECK-SAME: ptr addrspace(3) [[DEST_GROUP_PTR:%.*]], ptr [[SRC_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    call void @llvm.memcpy.p3.p0.i64(ptr addrspace(3) align 4 [[DEST_GROUP_PTR]], ptr align 4 [[SRC_PTR]], i64 [[SIZE]], i1 false), !tbaa [[TBAA0]], !alias.scope [[META3]], !noalias [[META6]]
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[SRC_PTR]] to ptr addrspace(1)
+; 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 [[TBAA0]], !alias.scope [[META3]], !noalias [[META6]]
 ; CHECK-NEXT:    ret void
 ;
   %cast.dest = addrspacecast ptr addrspace(3) %dest.group.ptr to ptr
@@ -116,7 +119,8 @@ define amdgpu_kernel void @memcpy_group_to_flat_replace_dest_global(ptr addrspac
 define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_tbaa_struct(ptr %dest, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {
 ; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_tbaa_struct(
 ; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p3.i64(ptr align 4 [[DEST]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false), !tbaa.struct [[TBAA_STRUCT8:![0-9]+]]
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)
+; 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]+]]
 ; CHECK-NEXT:    ret void
 ;
   %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr
@@ -127,7 +131,8 @@ define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_tbaa_struc
 define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_no_md(ptr %dest, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {
 ; CHECK-LABEL: define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_no_md(
 ; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p3.i64(ptr align 4 [[DEST]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false)
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)
+; 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)
 ; CHECK-NEXT:    ret void
 ;
   %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr
@@ -138,8 +143,10 @@ define amdgpu_kernel void @memcpy_flat_to_flat_replace_src_with_group_no_md(ptr
 define 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 {
 ; CHECK-LABEL: define amdgpu_kernel void @multiple_memcpy_flat_to_flat_replace_src_with_group_no_md(
 ; CHECK-SAME: ptr [[DEST0:%.*]], ptr [[DEST1:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p3.i64(ptr align 4 [[DEST0]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false)
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p3.i64(ptr align 4 [[DEST1]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false)
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[DEST0]] to ptr addrspace(1)
+; CHECK-NEXT:    [[TMP2:%.*]] = addrspacecast ptr [[DEST1]] to ptr addrspace(1)
+; 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)
+; 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)
 ; CHECK-NEXT:    ret void
 ;
   %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr
@@ -162,7 +169,8 @@ define amdgpu_kernel void @memcpy_group_flat_to_flat_self(ptr addrspace(3) %grou
 define amdgpu_kernel void @memmove_flat_to_flat_replace_src_with_group(ptr %dest, ptr addrspace(3) %src.group.ptr, i64 %size) #0 {
 ; CHECK-LABEL: define amdgpu_kernel void @memmove_flat_to_flat_replace_src_with_group(
 ; CHECK-SAME: ptr [[DEST:%.*]], ptr addrspace(3) [[SRC_GROUP_PTR:%.*]], i64 [[SIZE:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    call void @llvm.memmove.p0.p3.i64(ptr align 4 [[DEST]], ptr addrspace(3) align 4 [[SRC_GROUP_PTR]], i64 [[SIZE]], i1 false), !tbaa [[TBAA0]], !alias.scope [[META3]], !noalias [[META6]]
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[DEST]] to ptr addrspace(1)
+; 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 [[TBAA0]], !alias.scope [[META3]], !noalias [[META6]]
 ; CHECK-NEXT:    ret void
 ;
   %cast.src = addrspacecast ptr addrspace(3) %src.group.ptr to ptr