AMDGPU/SI: Make sure not to fold offsets into local address space globals

[• tstellarAMD]

Summary

Offset folding only works if you are emitting relocations, and we don't
emit relocations for local address space globals.

[nhaustov]

Looks good.

lib/Target/AMDGPU/SIISelLowering.h

public:
   TargetLoweringBase::LegalizeTypeAction
   getPreferredVectorAction(EVT VT) const override;
 
   bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
                                         Type *Ty) const override;
 
   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,
                                const SDLoc &DL, SelectionDAG &DAG,
                                SmallVectorImpl<SDValue> &InVals) const override;
 
   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
                       const SmallVectorImpl<ISD::OutputArg> &Outs,

lib/Target/AMDGPU/SIISelLowering.cpp

SDValue SITargetLowering::lowerADDRSPACECAST(SDValue Op,
   const MachineFunction &MF = DAG.getMachineFunction();
   DiagnosticInfoUnsupported InvalidAddrSpaceCast(
     *MF.getFunction(), "invalid addrspacecast", SL.getDebugLoc());
   DAG.getContext()->diagnose(InvalidAddrSpaceCast);
 
   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 {
   GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op);
 
   if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
       GSD->getAddressSpace() != AMDGPUAS::GLOBAL_ADDRESS)
     return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG);

test/CodeGen/AMDGPU/gv-offset-folding.ll

+; 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
+}