[InferAddressSpaces] Enhance the handling of cosntexpr.

Summary:
- Constant expressions may not be added in strict postorder as the
  forward instruction scan order. Thus, for a constant express (CE0), if
  its operand (CE1) is used in an previous instruction, they are not in
  postorder. However, different from
  `cloneInstructionWithNewAddressSpace`,
  `cloneConstantExprWithNewAddressSpace` doesn't bookkeep uninferred
  instructions for later resolving. That results in failure of inferring
  constant address.
- This patch adds the support to infer constant expression operand
  recursively, since there won't be loop, if that operand is another
  constant expression.

Reviewers: arsenm

Subscribers: jholewinski, jvesely, wdng, nhaehnle, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D61760

llvm-svn: 360431

2 files changed, 13 insertions, 1 deletions

diff --git a/llvm/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll b/llvm/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll
index f9b788f07fd..e5010104e88 100644
--- a/llvm/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll
+++ b/llvm/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll
@@ -71,3 +71,15 @@ define void @repeated_constexpr_gep_addrspacecast(i64 %idx0, i64 %idx1) {
 
   ret void
 }
+
+; CHECK-LABEL: @unorder_constexpr_gep_bitcast(
+; CHECK-NEXT: %x0 = load i32, i32 addrspace(3)* bitcast ([648 x double] addrspace(3)* @lds to i32 addrspace(3)*), align 4
+; CHECK-NEXT: %x1 = load i32, i32 addrspace(3)* getelementptr (i32, i32 addrspace(3)* bitcast ([648 x double] addrspace(3)* @lds to i32 addrspace(3)*), i32 1), align 4
+define void @unorder_constexpr_gep_bitcast() {
+  %x0 = load i32, i32* bitcast ([648 x double]* addrspacecast ([648 x double] addrspace(3)* @lds to [648 x double]*) to i32*), align 4
+  %x1 = load i32, i32* getelementptr (i32, i32* bitcast ([648 x double]* addrspacecast ([648 x double] addrspace(3)* @lds to [648 x double]*) to i32*), i32 1), align 4
+  call void @use(i32 %x0, i32 %x1)
+  ret void
+}
+
+declare void @use(i32, i32)
diff --git a/llvm/test/Transforms/InferAddressSpaces/NVPTX/clone_constexpr.ll b/llvm/test/Transforms/InferAddressSpaces/NVPTX/clone_constexpr.ll
index 1b324062057..53dff196d32 100644
--- a/llvm/test/Transforms/InferAddressSpaces/NVPTX/clone_constexpr.ll
+++ b/llvm/test/Transforms/InferAddressSpaces/NVPTX/clone_constexpr.ll
@@ -11,7 +11,7 @@ $g1 = comdat any
 ; CHECK:  %idxprom.i = zext i32 %x0 to i64
 ; CHECK:  %arrayidx.i = getelementptr %struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 %idxprom.i
 ; CHECK:  tail call void @f1(i32* %arrayidx.i, i32 undef) #0
-; CHECK:  %x1 = load i32, i32* getelementptr (%struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 0), align 4
+; CHECK:  %x1 = load i32, i32 addrspace(3)* getelementptr inbounds (%struct.S, %struct.S addrspace(3)* @g1, i64 0, i32 0, i64 0), align 4
 ; CHECK:  %L.sroa.0.0.insert.ext.i = zext i32 %x1 to i64
 ; CHECK:  tail call void @f2(i64* null, i64 %L.sroa.0.0.insert.ext.i) #0
 ; CHECK:  ret void