[PATCH] D45993: AMDGPU/SI: Don't promote alloca to vector for AddrSpaceCast instruction.

[Changpeng Fang via Phabricator via llvm-commits]

cfang updated this revision to Diff 144933.
cfang added a comment.

Update the test:

1. correct a few wrong checking;
2. Remove the -amdgiz from the triple since it is no longer necessary;
3. Add "-data-layout=A5" to explicitly specify the data layout for address space 5 for alloca
  - and to remove the "target:" line for the same data layout purpose.


https://reviews.llvm.org/D45993

Files:
  lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
  test/CodeGen/AMDGPU/vector-alloca-addrspacecast.ll


Index: test/CodeGen/AMDGPU/vector-alloca-addrspacecast.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/vector-alloca-addrspacecast.ll
@@ -0,0 +1,27 @@
+; RUN: opt -S -mtriple=amdgcn-- -data-layout=A5 -amdgpu-promote-alloca -sroa -instcombine < %s | FileCheck -check-prefix=OPT %s
+
+; Should give up promoting alloca to vector with an addrspacecast.
+
+; OPT-LABEL: @vector_addrspacecast(
+; OPT: alloca [3 x i32]
+; OPT: store i32 0, i32 addrspace(5)* %a0, align 4
+; OPT: store i32 1, i32 addrspace(5)* %a1, align 4
+; OPT: store i32 2, i32 addrspace(5)* %a2, align 4
+; OPT: %tmp = getelementptr [3 x i32], [3 x i32] addrspace(5)* %alloca, i64 0, i64 %index
+; OPT: %ac = addrspacecast i32 addrspace(5)* %tmp to i32*
+; OPT: %data = load i32, i32* %ac, align 4
+define amdgpu_kernel void @vector_addrspacecast(i32 addrspace(1)* %out, i64 %index) {
+entry:
+  %alloca = alloca [3 x i32], addrspace(5)
+  %a0 = getelementptr [3 x i32], [3 x i32] addrspace(5)* %alloca, i32 0, i32 0
+  %a1 = getelementptr [3 x i32], [3 x i32] addrspace(5)* %alloca, i32 0, i32 1
+  %a2 = getelementptr [3 x i32], [3 x i32] addrspace(5)* %alloca, i32 0, i32 2
+  store i32 0, i32 addrspace(5)* %a0
+  store i32 1, i32 addrspace(5)* %a1
+  store i32 2, i32 addrspace(5)* %a2
+  %tmp = getelementptr [3 x i32], [3 x i32] addrspace(5)* %alloca, i64 0, i64 %index
+  %ac = addrspacecast i32 addrspace(5)* %tmp to i32 *
+  %data = load i32, i32 * %ac
+  store i32 %data, i32 addrspace(1)* %out
+  ret void
+}
Index: lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
+++ lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
@@ -328,7 +328,6 @@
     return isa<GetElementPtrInst>(LI->getPointerOperand()) && !LI->isVolatile();
   }
   case Instruction::BitCast:
-  case Instruction::AddrSpaceCast:
     return true;
   case Instruction::Store: {
     // Must be the stored pointer operand, not a stored value, plus


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D45993.144933.patch
Type: text/x-patch
Size: 2047 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180502/5d2b9d47/attachment.bin>