3 files changed, 31 insertions, 0 deletions
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 60fe8c8bf54..3142e4c8f42 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1422,6 +1422,14 @@ SDValue SITargetLowering::lowerADDRSPACECAST(SDValue Op,
   return DAG.getUNDEF(ASC->getValueType(0));
 }
 
+bool
+SITargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
+  if (GA->getAddressSpace() != AMDGPUAS::GLOBAL_ADDRESS)
+    return false;
+
+  return TargetLowering::isOffsetFoldingLegal(GA);
+}
+
 SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
                                              SDValue Op,
                                              SelectionDAG &DAG) const {
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
index 2f013198970..032372b7b17 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -105,6 +105,8 @@ public:
 
   bool isTypeDesirableForOp(unsigned Op, EVT VT) const override;
 
+  bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
+
   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
                                bool isVarArg,
                                const SmallVectorImpl<ISD::InputArg> &Ins,
diff --git a/llvm/test/CodeGen/AMDGPU/gv-offset-folding.ll b/llvm/test/CodeGen/AMDGPU/gv-offset-folding.ll
new file mode 100644
index 00000000000..c75fdb35dd0
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/gv-offset-folding.ll
@@ -0,0 +1,21 @@
+; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji -relocation-model=static < %s | FileCheck %s
+
+@lds = external addrspace(3) global [4 x i32]
+
+; Function Attrs: nounwind
+
+; Offset folding is an optimization done for global variables with relocations,
+; which allows you to store the offset in the r_addend of the relocation entry.
+; The offset is apllied to the variables address at link time, which eliminates
+; the need to emit shader instructions to do this calculation.
+; We don't use relocations for local memory, so we should never fold offsets
+; for local memory globals.
+
+; CHECK-LABEL: lds_no_offset:
+; CHECK ds_write_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:4
+define void @lds_no_offset() {
+entry:
+  %ptr = getelementptr [4 x i32], [4 x i32] addrspace(3)* @lds, i32 0, i32 1
+  store i32 0, i32 addrspace(3)* %ptr
+  ret void
+}