AMDGPU/SI: Don't emit multi-dword flat memory ops when they might access scratch
Summary:
A single flat memory operations that might access the scratch buffer
can only access MaxPrivateElementSize bytes.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, tony-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D25788
llvm-svn: 285198
diff --git a/llvm/test/CodeGen/AMDGPU/flat-address-space.ll b/llvm/test/CodeGen/AMDGPU/flat-address-space.ll
index 1b6d8dc..0cfe688 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-address-space.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-address-space.ll
@@ -1,5 +1,6 @@
-; RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=bonaire < %s | FileCheck %s
-; RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga < %s | FileCheck %s
+; RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=bonaire < %s | FileCheck %s
+; RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga < %s | FileCheck %s
+; RUN: llc -O0 -mtriple=amdgcn-amd-amdhsa -mcpu=fiji < %s | FileCheck -check-prefixes=CHECK,HSA %s
; Disable optimizations in case there are optimizations added that
; specialize away generic pointer accesses.
@@ -149,6 +150,28 @@
ret void
}
+; CHECK-LABEL: flat_scratch_multidword_load:
+; HSA: flat_load_dword
+; HSA: flat_load_dword
+; FIXME: These tests are broken for os = mesa3d, becasue it doesn't initialize flat_scr
+define void @flat_scratch_multidword_load() {
+ %scratch = alloca <2 x i32>
+ %fptr = addrspacecast <2 x i32>* %scratch to <2 x i32> addrspace(4)*
+ %ld = load volatile <2 x i32>, <2 x i32> addrspace(4)* %fptr
+ ret void
+}
+
+; CHECK-LABEL: flat_scratch_multidword_store:
+; HSA: flat_store_dword
+; HSA: flat_store_dword
+; FIXME: These tests are broken for os = mesa3d, becasue it doesn't initialize flat_scr
+define void @flat_scratch_multidword_store() {
+ %scratch = alloca <2 x i32>
+ %fptr = addrspacecast <2 x i32>* %scratch to <2 x i32> addrspace(4)*
+ store volatile <2 x i32> zeroinitializer, <2 x i32> addrspace(4)* %fptr
+ ret void
+}
+
attributes #0 = { nounwind }
attributes #1 = { nounwind convergent }
attributes #3 = { nounwind readnone }