AMDGPU: Implement addrspacecast
llvm-svn: 267452
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp
index b4b872e..c441a6d 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp
@@ -24,6 +24,8 @@
class AMDGPUAnnotateKernelFeatures : public ModulePass {
private:
+ static bool hasAddrSpaceCast(const Function &F);
+
void addAttrToCallers(Function *Intrin, StringRef AttrName);
bool addAttrsForIntrinsics(Module &M, ArrayRef<StringRef[2]>);
@@ -48,12 +50,29 @@
char &llvm::AMDGPUAnnotateKernelFeaturesID = AMDGPUAnnotateKernelFeatures::ID;
+INITIALIZE_PASS(AMDGPUAnnotateKernelFeatures, DEBUG_TYPE,
+ "Add AMDGPU function attributes", false, false)
-INITIALIZE_PASS_BEGIN(AMDGPUAnnotateKernelFeatures, DEBUG_TYPE,
- "Add AMDGPU function attributes", false, false)
-INITIALIZE_PASS_END(AMDGPUAnnotateKernelFeatures, DEBUG_TYPE,
- "Add AMDGPU function attributes", false, false)
+static bool castRequiresQueuePtr(const AddrSpaceCastInst *ASC) {
+ unsigned SrcAS = ASC->getSrcAddressSpace();
+ // The queue ptr is only needed when casting to flat, not from it.
+ return SrcAS == AMDGPUAS::LOCAL_ADDRESS || SrcAS == AMDGPUAS::PRIVATE_ADDRESS;
+}
+
+// Return true if an addrspacecast is used that requires the queue ptr.
+bool AMDGPUAnnotateKernelFeatures::hasAddrSpaceCast(const Function &F) {
+ for (const BasicBlock &BB : F) {
+ for (const Instruction &I : BB) {
+ if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(&I)) {
+ if (castRequiresQueuePtr(ASC))
+ return true;
+ }
+ }
+ }
+
+ return false;
+}
void AMDGPUAnnotateKernelFeatures::addAttrToCallers(Function *Intrin,
StringRef AttrName) {
@@ -117,9 +136,18 @@
// always initialized.
bool Changed = addAttrsForIntrinsics(M, IntrinsicToAttr);
- if (TT.getOS() == Triple::AMDHSA)
+ if (TT.getOS() == Triple::AMDHSA) {
Changed |= addAttrsForIntrinsics(M, HSAIntrinsicToAttr);
+ for (Function &F : M) {
+ if (F.hasFnAttribute("amdgpu-queue-ptr"))
+ continue;
+
+ if (hasAddrSpaceCast(F))
+ F.addFnAttr("amdgpu-queue-ptr");
+ }
+ }
+
return Changed;
}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index de500d8..0e3f7ed 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -147,7 +147,6 @@
bool SelectSMRDBufferImm(SDValue Addr, SDValue &Offset) const;
bool SelectSMRDBufferImm32(SDValue Addr, SDValue &Offset) const;
bool SelectSMRDBufferSgpr(SDValue Addr, SDValue &Offset) const;
- SDNode *SelectAddrSpaceCast(SDNode *N);
bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
bool SelectVOP3NoMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
@@ -526,8 +525,6 @@
Lowering.legalizeTargetIndependentNode(N, *CurDAG);
break;
}
- case ISD::ADDRSPACECAST:
- return SelectAddrSpaceCast(N);
case ISD::AND:
case ISD::SRL:
case ISD::SRA:
@@ -1332,69 +1329,6 @@
!isa<ConstantSDNode>(Offset);
}
-// FIXME: This is incorrect and only enough to be able to compile.
-SDNode *AMDGPUDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
- AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
- SDLoc DL(N);
-
- const MachineFunction &MF = CurDAG->getMachineFunction();
- DiagnosticInfoUnsupported NotImplemented(
- *MF.getFunction(), "addrspacecast not implemented", DL.getDebugLoc());
- CurDAG->getContext()->diagnose(NotImplemented);
-
- assert(Subtarget->hasFlatAddressSpace() &&
- "addrspacecast only supported with flat address space!");
-
- assert((ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS ||
- ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) &&
- "Can only cast to / from flat address space!");
-
- // The flat instructions read the address as the index of the VGPR holding the
- // address, so casting should just be reinterpreting the base VGPR, so just
- // insert trunc / bitcast / zext.
-
- SDValue Src = ASC->getOperand(0);
- EVT DestVT = ASC->getValueType(0);
- EVT SrcVT = Src.getValueType();
-
- unsigned SrcSize = SrcVT.getSizeInBits();
- unsigned DestSize = DestVT.getSizeInBits();
-
- if (SrcSize > DestSize) {
- assert(SrcSize == 64 && DestSize == 32);
- return CurDAG->getMachineNode(
- TargetOpcode::EXTRACT_SUBREG,
- DL,
- DestVT,
- Src,
- CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32));
- }
-
- if (DestSize > SrcSize) {
- assert(SrcSize == 32 && DestSize == 64);
-
- // FIXME: This is probably wrong, we should never be defining
- // a register class with both VGPRs and SGPRs
- SDValue RC = CurDAG->getTargetConstant(AMDGPU::VS_64RegClassID, DL,
- MVT::i32);
-
- const SDValue Ops[] = {
- RC,
- Src,
- CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
- SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32,
- CurDAG->getConstant(0, DL, MVT::i32)), 0),
- CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32)
- };
-
- return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
- DL, N->getValueType(0), Ops);
- }
-
- assert(SrcSize == 64 && DestSize == 64);
- return CurDAG->getNode(ISD::BITCAST, DL, DestVT, Src).getNode();
-}
-
SDNode *AMDGPUDAGToDAGISel::getS_BFE(unsigned Opcode, SDLoc DL, SDValue Val,
uint32_t Offset, uint32_t Width) {
// Transformation function, pack the offset and width of a BFE into
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index a5d1ec2..2674efd 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -278,6 +278,11 @@
setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i32, Expand);
setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i64, Expand);
+ if (Subtarget->hasFlatAddressSpace()) {
+ setOperationAction(ISD::ADDRSPACECAST, MVT::i32, Custom);
+ setOperationAction(ISD::ADDRSPACECAST, MVT::i64, Custom);
+ }
+
setTargetDAGCombine(ISD::FADD);
setTargetDAGCombine(ISD::FSUB);
setTargetDAGCombine(ISD::FMINNUM);
@@ -1232,6 +1237,7 @@
case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG);
+ case ISD::ADDRSPACECAST: return lowerADDRSPACECAST(Op, DAG);
}
return SDValue();
}
@@ -1390,6 +1396,84 @@
return Chain;
}
+SDValue SITargetLowering::getSegmentAperture(unsigned AS,
+ SelectionDAG &DAG) const {
+ SDLoc SL;
+ MachineFunction &MF = DAG.getMachineFunction();
+ SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
+ SDValue QueuePtr = CreateLiveInRegister(
+ DAG, &AMDGPU::SReg_64RegClass, Info->getQueuePtrUserSGPR(), MVT::i64);
+
+ // Offset into amd_queue_t for group_segment_aperture_base_hi /
+ // private_segment_aperture_base_hi.
+ uint32_t StructOffset = (AS == AMDGPUAS::LOCAL_ADDRESS) ? 0x40 : 0x44;
+
+ SDValue Ptr = DAG.getNode(ISD::ADD, SL, MVT::i64, QueuePtr,
+ DAG.getConstant(StructOffset, SL, MVT::i64));
+
+ // TODO: Use custom target PseudoSourceValue.
+ // TODO: We should use the value from the IR intrinsic call, but it might not
+ // be available and how do we get it?
+ Value *V = UndefValue::get(PointerType::get(Type::getInt8Ty(*DAG.getContext()),
+ AMDGPUAS::CONSTANT_ADDRESS));
+
+ MachinePointerInfo PtrInfo(V, StructOffset);
+ return DAG.getLoad(MVT::i32, SL, QueuePtr.getValue(1), Ptr,
+ PtrInfo, false,
+ false, true,
+ MinAlign(64, StructOffset));
+}
+
+SDValue SITargetLowering::lowerADDRSPACECAST(SDValue Op,
+ SelectionDAG &DAG) const {
+ SDLoc SL(Op);
+ const AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(Op);
+
+ SDValue Src = ASC->getOperand(0);
+
+ // FIXME: Really support non-0 null pointers.
+ SDValue SegmentNullPtr = DAG.getConstant(-1, SL, MVT::i32);
+ SDValue FlatNullPtr = DAG.getConstant(0, SL, MVT::i64);
+
+ // flat -> local/private
+ if (ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS) {
+ if (ASC->getDestAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
+ ASC->getDestAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
+ SDValue NonNull = DAG.getSetCC(SL, MVT::i1, Src, FlatNullPtr, ISD::SETNE);
+ SDValue Ptr = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, Src);
+
+ return DAG.getNode(ISD::SELECT, SL, MVT::i32,
+ NonNull, Ptr, SegmentNullPtr);
+ }
+ }
+
+ // local/private -> flat
+ if (ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) {
+ if (ASC->getSrcAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
+ ASC->getSrcAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
+ SDValue NonNull
+ = DAG.getSetCC(SL, MVT::i1, Src, SegmentNullPtr, ISD::SETNE);
+
+ SDValue Aperture = getSegmentAperture(ASC->getSrcAddressSpace(), DAG);
+ SDValue CvtPtr
+ = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Src, Aperture);
+
+ return DAG.getNode(ISD::SELECT, SL, MVT::i64, NonNull,
+ DAG.getNode(ISD::BITCAST, SL, MVT::i64, CvtPtr),
+ FlatNullPtr);
+ }
+ }
+
+ // global <-> flat are no-ops and never emitted.
+
+ const MachineFunction &MF = DAG.getMachineFunction();
+ DiagnosticInfoUnsupported InvalidAddrSpaceCast(
+ *MF.getFunction(), "invalid addrspacecast", SL.getDebugLoc());
+ DAG.getContext()->diagnose(InvalidAddrSpaceCast);
+
+ return DAG.getUNDEF(ASC->getValueType(0));
+}
+
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 4e7131a..e7de080 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -45,6 +45,9 @@
SDValue LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
+ SDValue getSegmentAperture(unsigned AS, SelectionDAG &DAG) const;
+ SDValue lowerADDRSPACECAST(SDValue Op, SelectionDAG &DAG) const;
+
void adjustWritemask(MachineSDNode *&N, SelectionDAG &DAG) const;
SDValue performUCharToFloatCombine(SDNode *N,
diff --git a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h
index acbd276..aa59602 100644
--- a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h
@@ -270,6 +270,10 @@
ScratchWaveOffsetReg = Reg;
}
+ unsigned getQueuePtrUserSGPR() const {
+ return QueuePtrUserSGPR;
+ }
+
bool hasSpilledSGPRs() const {
return HasSpilledSGPRs;
}