AMDGPU: Implement isCheapAddrSpaceCast

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@288523 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp
index ac5fddfd42d5826cdfac22df6543c6fed1948c9d..ef61fc409f860a45bb8c849c6e513c823a6b80e2 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -600,8 +600,8 @@ EVT SITargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
 
 static bool isFlatGlobalAddrSpace(unsigned AS) {
   return AS == AMDGPUAS::GLOBAL_ADDRESS ||
-    AS == AMDGPUAS::FLAT_ADDRESS ||
-    AS == AMDGPUAS::CONSTANT_ADDRESS;
+         AS == AMDGPUAS::FLAT_ADDRESS ||
+         AS == AMDGPUAS::CONSTANT_ADDRESS;
 }
 
 bool SITargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
@@ -609,6 +609,16 @@ bool SITargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
   return isFlatGlobalAddrSpace(SrcAS) && isFlatGlobalAddrSpace(DestAS);
 }
 
+bool SITargetLowering::isCheapAddrSpaceCast(unsigned SrcAS,
+                                            unsigned DestAS) const {
+  // Flat -> private/local is a simple truncate.
+  // Flat -> global is no-op
+  if (SrcAS == AMDGPUAS::FLAT_ADDRESS)
+    return true;
+
+  return isNoopAddrSpaceCast(SrcAS, DestAS);
+}
+
 bool SITargetLowering::isMemOpUniform(const SDNode *N) const {
   const MemSDNode *MemNode = cast<MemSDNode>(N);
   const Value *Ptr = MemNode->getMemOperand()->getValue();

diff --git a/lib/Target/AMDGPU/SIISelLowering.h b/lib/Target/AMDGPU/SIISelLowering.h
index 2e14f499a85f050e283d6fd8c5de055f103e26dc..03846fd54732e46144a078826bcc0e560e9eb8c0 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.h
+++ b/lib/Target/AMDGPU/SIISelLowering.h
@@ -130,6 +130,7 @@ public:
 
   bool isMemOpUniform(const SDNode *N) const;
   bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
+  bool isCheapAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
 
   TargetLoweringBase::LegalizeTypeAction
   getPreferredVectorAction(EVT VT) const override;

diff --git a/test/Transforms/CodeGenPrepare/AMDGPU/sink-addrspacecast.ll b/test/Transforms/CodeGenPrepare/AMDGPU/sink-addrspacecast.ll
new file mode 100644 (file)
index 0000000..adeba26
--- /dev/null
+++ b/test/Transforms/CodeGenPrepare/AMDGPU/sink-addrspacecast.ll
@@ -0,0 +1,121 @@
+; RUN: opt -S -codegenprepare -mtriple=amdgcn--amdhsa < %s | FileCheck %s
+
+; CHECK-LABEL: @no_sink_local_to_flat(
+; CHECK: addrspacecast
+; CHECK: br
+; CHECK-NOT: addrspacecast
+define i64 @no_sink_local_to_flat(i1 %pred, i64 addrspace(3)* %ptr) {
+  %ptr_cast = addrspacecast i64 addrspace(3)* %ptr to i64 addrspace(4)*
+  br i1 %pred, label %l1, label %l2
+
+l1:
+  %v1 = load i64, i64 addrspace(3)* %ptr
+  ret i64 %v1
+
+l2:
+  %v2 = load i64, i64 addrspace(4)* %ptr_cast
+  ret i64 %v2
+}
+
+; CHECK-LABEL: @no_sink_private_to_flat(
+; CHECK: addrspacecast
+; CHECK: br
+; CHECK-NOT: addrspacecast
+define i64 @no_sink_private_to_flat(i1 %pred, i64* %ptr) {
+  %ptr_cast = addrspacecast i64* %ptr to i64 addrspace(4)*
+  br i1 %pred, label %l1, label %l2
+
+l1:
+  %v1 = load i64, i64* %ptr
+  ret i64 %v1
+
+l2:
+  %v2 = load i64, i64 addrspace(4)* %ptr_cast
+  ret i64 %v2
+}
+
+
+; CHECK-LABEL: @sink_global_to_flat(
+; CHECK-NOT: addrspacecast
+; CHECK: br
+; CHECK: addrspacecast
+define i64 @sink_global_to_flat(i1 %pred, i64 addrspace(1)* %ptr) {
+  %ptr_cast = addrspacecast i64 addrspace(1)* %ptr to i64 addrspace(4)*
+  br i1 %pred, label %l1, label %l2
+
+l1:
+  %v1 = load i64, i64 addrspace(1)* %ptr
+  ret i64 %v1
+
+l2:
+  %v2 = load i64, i64 addrspace(4)* %ptr_cast
+  ret i64 %v2
+}
+
+; CHECK-LABEL: @sink_flat_to_global(
+; CHECK-NOT: addrspacecast
+; CHECK: br
+; CHECK: addrspacecast
+define i64 @sink_flat_to_global(i1 %pred, i64 addrspace(4)* %ptr) {
+  %ptr_cast = addrspacecast i64 addrspace(4)* %ptr to i64 addrspace(1)*
+  br i1 %pred, label %l1, label %l2
+
+l1:
+  %v1 = load i64, i64 addrspace(4)* %ptr
+  ret i64 %v1
+
+l2:
+  %v2 = load i64, i64 addrspace(1)* %ptr_cast
+  ret i64 %v2
+}
+
+; CHECK-LABEL: @sink_flat_to_constant(
+; CHECK-NOT: addrspacecast
+; CHECK: br
+; CHECK: addrspacecast
+define i64 @sink_flat_to_constant(i1 %pred, i64 addrspace(4)* %ptr) {
+  %ptr_cast = addrspacecast i64 addrspace(4)* %ptr to i64 addrspace(2)*
+  br i1 %pred, label %l1, label %l2
+
+l1:
+  %v1 = load i64, i64 addrspace(4)* %ptr
+  ret i64 %v1
+
+l2:
+  %v2 = load i64, i64 addrspace(2)* %ptr_cast
+  ret i64 %v2
+}
+
+; CHECK-LABEL: @sink_flat_to_local(
+; CHECK-NOT: addrspacecast
+; CHECK: br
+; CHECK: addrspacecast
+define i64 @sink_flat_to_local(i1 %pred, i64 addrspace(4)* %ptr) {
+  %ptr_cast = addrspacecast i64 addrspace(4)* %ptr to i64 addrspace(3)*
+  br i1 %pred, label %l1, label %l2
+
+l1:
+  %v1 = load i64, i64 addrspace(4)* %ptr
+  ret i64 %v1
+
+l2:
+  %v2 = load i64, i64 addrspace(3)* %ptr_cast
+  ret i64 %v2
+}
+
+; CHECK-LABEL: @sink_flat_to_private(
+; CHECK-NOT: addrspacecast
+; CHECK: br
+; CHECK: addrspacecast
+define i64 @sink_flat_to_private(i1 %pred, i64 addrspace(4)* %ptr) {
+  %ptr_cast = addrspacecast i64 addrspace(4)* %ptr to i64*
+  br i1 %pred, label %l1, label %l2
+
+l1:
+  %v1 = load i64, i64 addrspace(4)* %ptr
+  ret i64 %v1
+
+l2:
+  %v2 = load i64, i64* %ptr_cast
+  ret i64 %v2
+}