//===-- TargetLowering.cpp - Implement the TargetLowering class -----------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This implements the TargetLowering class. // //===----------------------------------------------------------------------===// #include "llvm/CodeGen/TargetLowering.h" #include "llvm/ADT/STLExtras.h" #include "llvm/Analysis/VectorUtils.h" #include "llvm/CodeGen/CallingConvLower.h" #include "llvm/CodeGen/CodeGenCommonISel.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineJumpTableInfo.h" #include "llvm/CodeGen/MachineModuleInfoImpls.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/SelectionDAG.h" #include "llvm/CodeGen/TargetRegisterInfo.h" #include "llvm/IR/DataLayout.h" #include "llvm/IR/DerivedTypes.h" #include "llvm/IR/GlobalVariable.h" #include "llvm/IR/LLVMContext.h" #include "llvm/MC/MCAsmInfo.h" #include "llvm/MC/MCExpr.h" #include "llvm/Support/DivisionByConstantInfo.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/KnownBits.h" #include "llvm/Support/MathExtras.h" #include "llvm/Target/TargetMachine.h" #include using namespace llvm; /// NOTE: The TargetMachine owns TLOF. TargetLowering::TargetLowering(const TargetMachine &tm) : TargetLoweringBase(tm) {} const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { return nullptr; } bool TargetLowering::isPositionIndependent() const { return getTargetMachine().isPositionIndependent(); } /// Check whether a given call node is in tail position within its function. If /// so, it sets Chain to the input chain of the tail call. bool TargetLowering::isInTailCallPosition(SelectionDAG &DAG, SDNode *Node, SDValue &Chain) const { const Function &F = DAG.getMachineFunction().getFunction(); // First, check if tail calls have been disabled in this function. if (F.getFnAttribute("disable-tail-calls").getValueAsBool()) return false; // Conservatively require the attributes of the call to match those of // the return. Ignore following attributes because they don't affect the // call sequence. AttrBuilder CallerAttrs(F.getContext(), F.getAttributes().getRetAttrs()); for (const auto &Attr : {Attribute::Alignment, Attribute::Dereferenceable, Attribute::DereferenceableOrNull, Attribute::NoAlias, Attribute::NonNull, Attribute::NoUndef, Attribute::Range}) CallerAttrs.removeAttribute(Attr); if (CallerAttrs.hasAttributes()) return false; // It's not safe to eliminate the sign / zero extension of the return value. if (CallerAttrs.contains(Attribute::ZExt) || CallerAttrs.contains(Attribute::SExt)) return false; // Check if the only use is a function return node. return isUsedByReturnOnly(Node, Chain); } bool TargetLowering::parametersInCSRMatch(const MachineRegisterInfo &MRI, const uint32_t *CallerPreservedMask, const SmallVectorImpl &ArgLocs, const SmallVectorImpl &OutVals) const { for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) { const CCValAssign &ArgLoc = ArgLocs[I]; if (!ArgLoc.isRegLoc()) continue; MCRegister Reg = ArgLoc.getLocReg(); // Only look at callee saved registers. if (MachineOperand::clobbersPhysReg(CallerPreservedMask, Reg)) continue; // Check that we pass the value used for the caller. // (We look for a CopyFromReg reading a virtual register that is used // for the function live-in value of register Reg) SDValue Value = OutVals[I]; if (Value->getOpcode() == ISD::AssertZext) Value = Value.getOperand(0); if (Value->getOpcode() != ISD::CopyFromReg) return false; Register ArgReg = cast(Value->getOperand(1))->getReg(); if (MRI.getLiveInPhysReg(ArgReg) != Reg) return false; } return true; } /// Set CallLoweringInfo attribute flags based on a call instruction /// and called function attributes. void TargetLoweringBase::ArgListEntry::setAttributes(const CallBase *Call, unsigned ArgIdx) { IsSExt = Call->paramHasAttr(ArgIdx, Attribute::SExt); IsZExt = Call->paramHasAttr(ArgIdx, Attribute::ZExt); IsInReg = Call->paramHasAttr(ArgIdx, Attribute::InReg); IsSRet = Call->paramHasAttr(ArgIdx, Attribute::StructRet); IsNest = Call->paramHasAttr(ArgIdx, Attribute::Nest); IsByVal = Call->paramHasAttr(ArgIdx, Attribute::ByVal); IsPreallocated = Call->paramHasAttr(ArgIdx, Attribute::Preallocated); IsInAlloca = Call->paramHasAttr(ArgIdx, Attribute::InAlloca); IsReturned = Call->paramHasAttr(ArgIdx, Attribute::Returned); IsSwiftSelf = Call->paramHasAttr(ArgIdx, Attribute::SwiftSelf); IsSwiftAsync = Call->paramHasAttr(ArgIdx, Attribute::SwiftAsync); IsSwiftError = Call->paramHasAttr(ArgIdx, Attribute::SwiftError); Alignment = Call->getParamStackAlign(ArgIdx); IndirectType = nullptr; assert(IsByVal + IsPreallocated + IsInAlloca + IsSRet <= 1 && "multiple ABI attributes?"); if (IsByVal) { IndirectType = Call->getParamByValType(ArgIdx); if (!Alignment) Alignment = Call->getParamAlign(ArgIdx); } if (IsPreallocated) IndirectType = Call->getParamPreallocatedType(ArgIdx); if (IsInAlloca) IndirectType = Call->getParamInAllocaType(ArgIdx); if (IsSRet) IndirectType = Call->getParamStructRetType(ArgIdx); } /// Generate a libcall taking the given operands as arguments and returning a /// result of type RetVT. std::pair TargetLowering::makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC, EVT RetVT, ArrayRef Ops, MakeLibCallOptions CallOptions, const SDLoc &dl, SDValue InChain) const { if (!InChain) InChain = DAG.getEntryNode(); TargetLowering::ArgListTy Args; Args.reserve(Ops.size()); TargetLowering::ArgListEntry Entry; for (unsigned i = 0; i < Ops.size(); ++i) { SDValue NewOp = Ops[i]; Entry.Node = NewOp; Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext()); Entry.IsSExt = shouldSignExtendTypeInLibCall(NewOp.getValueType(), CallOptions.IsSExt); Entry.IsZExt = !Entry.IsSExt; if (CallOptions.IsSoften && !shouldExtendTypeInLibCall(CallOptions.OpsVTBeforeSoften[i])) { Entry.IsSExt = Entry.IsZExt = false; } Args.push_back(Entry); } if (LC == RTLIB::UNKNOWN_LIBCALL) report_fatal_error("Unsupported library call operation!"); SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC), getPointerTy(DAG.getDataLayout())); Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext()); TargetLowering::CallLoweringInfo CLI(DAG); bool signExtend = shouldSignExtendTypeInLibCall(RetVT, CallOptions.IsSExt); bool zeroExtend = !signExtend; if (CallOptions.IsSoften && !shouldExtendTypeInLibCall(CallOptions.RetVTBeforeSoften)) { signExtend = zeroExtend = false; } CLI.setDebugLoc(dl) .setChain(InChain) .setLibCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args)) .setNoReturn(CallOptions.DoesNotReturn) .setDiscardResult(!CallOptions.IsReturnValueUsed) .setIsPostTypeLegalization(CallOptions.IsPostTypeLegalization) .setSExtResult(signExtend) .setZExtResult(zeroExtend); return LowerCallTo(CLI); } bool TargetLowering::findOptimalMemOpLowering( std::vector &MemOps, unsigned Limit, const MemOp &Op, unsigned DstAS, unsigned SrcAS, const AttributeList &FuncAttributes) const { if (Limit != ~unsigned(0) && Op.isMemcpyWithFixedDstAlign() && Op.getSrcAlign() < Op.getDstAlign()) return false; EVT VT = getOptimalMemOpType(Op, FuncAttributes); if (VT == MVT::Other) { // Use the largest integer type whose alignment constraints are satisfied. // We only need to check DstAlign here as SrcAlign is always greater or // equal to DstAlign (or zero). VT = MVT::LAST_INTEGER_VALUETYPE; if (Op.isFixedDstAlign()) while (Op.getDstAlign() < (VT.getSizeInBits() / 8) && !allowsMisalignedMemoryAccesses(VT, DstAS, Op.getDstAlign())) VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1); assert(VT.isInteger()); // Find the largest legal integer type. MVT LVT = MVT::LAST_INTEGER_VALUETYPE; while (!isTypeLegal(LVT)) LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1); assert(LVT.isInteger()); // If the type we've chosen is larger than the largest legal integer type // then use that instead. if (VT.bitsGT(LVT)) VT = LVT; } unsigned NumMemOps = 0; uint64_t Size = Op.size(); while (Size) { unsigned VTSize = VT.getSizeInBits() / 8; while (VTSize > Size) { // For now, only use non-vector load / store's for the left-over pieces. EVT NewVT = VT; unsigned NewVTSize; bool Found = false; if (VT.isVector() || VT.isFloatingPoint()) { NewVT = (VT.getSizeInBits() > 64) ? MVT::i64 : MVT::i32; if (isOperationLegalOrCustom(ISD::STORE, NewVT) && isSafeMemOpType(NewVT.getSimpleVT())) Found = true; else if (NewVT == MVT::i64 && isOperationLegalOrCustom(ISD::STORE, MVT::f64) && isSafeMemOpType(MVT::f64)) { // i64 is usually not legal on 32-bit targets, but f64 may be. NewVT = MVT::f64; Found = true; } } if (!Found) { do { NewVT = (MVT::SimpleValueType)(NewVT.getSimpleVT().SimpleTy - 1); if (NewVT == MVT::i8) break; } while (!isSafeMemOpType(NewVT.getSimpleVT())); } NewVTSize = NewVT.getSizeInBits() / 8; // If the new VT cannot cover all of the remaining bits, then consider // issuing a (or a pair of) unaligned and overlapping load / store. unsigned Fast; if (NumMemOps && Op.allowOverlap() && NewVTSize < Size && allowsMisalignedMemoryAccesses( VT, DstAS, Op.isFixedDstAlign() ? Op.getDstAlign() : Align(1), MachineMemOperand::MONone, &Fast) && Fast) VTSize = Size; else { VT = NewVT; VTSize = NewVTSize; } } if (++NumMemOps > Limit) return false; MemOps.push_back(VT); Size -= VTSize; } return true; } /// Soften the operands of a comparison. This code is shared among BR_CC, /// SELECT_CC, and SETCC handlers. void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT, SDValue &NewLHS, SDValue &NewRHS, ISD::CondCode &CCCode, const SDLoc &dl, const SDValue OldLHS, const SDValue OldRHS) const { SDValue Chain; return softenSetCCOperands(DAG, VT, NewLHS, NewRHS, CCCode, dl, OldLHS, OldRHS, Chain); } void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT, SDValue &NewLHS, SDValue &NewRHS, ISD::CondCode &CCCode, const SDLoc &dl, const SDValue OldLHS, const SDValue OldRHS, SDValue &Chain, bool IsSignaling) const { // FIXME: Currently we cannot really respect all IEEE predicates due to libgcc // not supporting it. We can update this code when libgcc provides such // functions. assert((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128 || VT == MVT::ppcf128) && "Unsupported setcc type!"); // Expand into one or more soft-fp libcall(s). RTLIB::Libcall LC1 = RTLIB::UNKNOWN_LIBCALL, LC2 = RTLIB::UNKNOWN_LIBCALL; bool ShouldInvertCC = false; switch (CCCode) { case ISD::SETEQ: case ISD::SETOEQ: LC1 = (VT == MVT::f32) ? RTLIB::OEQ_F32 : (VT == MVT::f64) ? RTLIB::OEQ_F64 : (VT == MVT::f128) ? RTLIB::OEQ_F128 : RTLIB::OEQ_PPCF128; break; case ISD::SETNE: case ISD::SETUNE: LC1 = (VT == MVT::f32) ? RTLIB::UNE_F32 : (VT == MVT::f64) ? RTLIB::UNE_F64 : (VT == MVT::f128) ? RTLIB::UNE_F128 : RTLIB::UNE_PPCF128; break; case ISD::SETGE: case ISD::SETOGE: LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 : (VT == MVT::f64) ? RTLIB::OGE_F64 : (VT == MVT::f128) ? RTLIB::OGE_F128 : RTLIB::OGE_PPCF128; break; case ISD::SETLT: case ISD::SETOLT: LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 : (VT == MVT::f64) ? RTLIB::OLT_F64 : (VT == MVT::f128) ? RTLIB::OLT_F128 : RTLIB::OLT_PPCF128; break; case ISD::SETLE: case ISD::SETOLE: LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 : (VT == MVT::f64) ? RTLIB::OLE_F64 : (VT == MVT::f128) ? RTLIB::OLE_F128 : RTLIB::OLE_PPCF128; break; case ISD::SETGT: case ISD::SETOGT: LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 : (VT == MVT::f64) ? RTLIB::OGT_F64 : (VT == MVT::f128) ? RTLIB::OGT_F128 : RTLIB::OGT_PPCF128; break; case ISD::SETO: ShouldInvertCC = true; [[fallthrough]]; case ISD::SETUO: LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 : (VT == MVT::f64) ? RTLIB::UO_F64 : (VT == MVT::f128) ? RTLIB::UO_F128 : RTLIB::UO_PPCF128; break; case ISD::SETONE: // SETONE = O && UNE ShouldInvertCC = true; [[fallthrough]]; case ISD::SETUEQ: LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 : (VT == MVT::f64) ? RTLIB::UO_F64 : (VT == MVT::f128) ? RTLIB::UO_F128 : RTLIB::UO_PPCF128; LC2 = (VT == MVT::f32) ? RTLIB::OEQ_F32 : (VT == MVT::f64) ? RTLIB::OEQ_F64 : (VT == MVT::f128) ? RTLIB::OEQ_F128 : RTLIB::OEQ_PPCF128; break; default: // Invert CC for unordered comparisons ShouldInvertCC = true; switch (CCCode) { case ISD::SETULT: LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 : (VT == MVT::f64) ? RTLIB::OGE_F64 : (VT == MVT::f128) ? RTLIB::OGE_F128 : RTLIB::OGE_PPCF128; break; case ISD::SETULE: LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 : (VT == MVT::f64) ? RTLIB::OGT_F64 : (VT == MVT::f128) ? RTLIB::OGT_F128 : RTLIB::OGT_PPCF128; break; case ISD::SETUGT: LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 : (VT == MVT::f64) ? RTLIB::OLE_F64 : (VT == MVT::f128) ? RTLIB::OLE_F128 : RTLIB::OLE_PPCF128; break; case ISD::SETUGE: LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 : (VT == MVT::f64) ? RTLIB::OLT_F64 : (VT == MVT::f128) ? RTLIB::OLT_F128 : RTLIB::OLT_PPCF128; break; default: llvm_unreachable("Do not know how to soften this setcc!"); } } // Use the target specific return value for comparison lib calls. EVT RetVT = getCmpLibcallReturnType(); SDValue Ops[2] = {NewLHS, NewRHS}; TargetLowering::MakeLibCallOptions CallOptions; EVT OpsVT[2] = { OldLHS.getValueType(), OldRHS.getValueType() }; CallOptions.setTypeListBeforeSoften(OpsVT, RetVT, true); auto Call = makeLibCall(DAG, LC1, RetVT, Ops, CallOptions, dl, Chain); NewLHS = Call.first; NewRHS = DAG.getConstant(0, dl, RetVT); CCCode = getCmpLibcallCC(LC1); if (ShouldInvertCC) { assert(RetVT.isInteger()); CCCode = getSetCCInverse(CCCode, RetVT); } if (LC2 == RTLIB::UNKNOWN_LIBCALL) { // Update Chain. Chain = Call.second; } else { EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), RetVT); SDValue Tmp = DAG.getSetCC(dl, SetCCVT, NewLHS, NewRHS, CCCode); auto Call2 = makeLibCall(DAG, LC2, RetVT, Ops, CallOptions, dl, Chain); CCCode = getCmpLibcallCC(LC2); if (ShouldInvertCC) CCCode = getSetCCInverse(CCCode, RetVT); NewLHS = DAG.getSetCC(dl, SetCCVT, Call2.first, NewRHS, CCCode); if (Chain) Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Call.second, Call2.second); NewLHS = DAG.getNode(ShouldInvertCC ? ISD::AND : ISD::OR, dl, Tmp.getValueType(), Tmp, NewLHS); NewRHS = SDValue(); } } /// Return the entry encoding for a jump table in the current function. The /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum. unsigned TargetLowering::getJumpTableEncoding() const { // In non-pic modes, just use the address of a block. if (!isPositionIndependent()) return MachineJumpTableInfo::EK_BlockAddress; // In PIC mode, if the target supports a GPRel32 directive, use it. if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != nullptr) return MachineJumpTableInfo::EK_GPRel32BlockAddress; // Otherwise, use a label difference. return MachineJumpTableInfo::EK_LabelDifference32; } SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, SelectionDAG &DAG) const { // If our PIC model is GP relative, use the global offset table as the base. unsigned JTEncoding = getJumpTableEncoding(); if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) || (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress)) return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(DAG.getDataLayout())); return Table; } /// This returns the relocation base for the given PIC jumptable, the same as /// getPICJumpTableRelocBase, but as an MCExpr. const MCExpr * TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,MCContext &Ctx) const{ // The normal PIC reloc base is the label at the start of the jump table. return MCSymbolRefExpr::create(MF->getJTISymbol(JTI, Ctx), Ctx); } SDValue TargetLowering::expandIndirectJTBranch(const SDLoc &dl, SDValue Value, SDValue Addr, int JTI, SelectionDAG &DAG) const { SDValue Chain = Value; // Jump table debug info is only needed if CodeView is enabled. if (DAG.getTarget().getTargetTriple().isOSBinFormatCOFF()) { Chain = DAG.getJumpTableDebugInfo(JTI, Chain, dl); } return DAG.getNode(ISD::BRIND, dl, MVT::Other, Chain, Addr); } bool TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { const TargetMachine &TM = getTargetMachine(); const GlobalValue *GV = GA->getGlobal(); // If the address is not even local to this DSO we will have to load it from // a got and then add the offset. if (!TM.shouldAssumeDSOLocal(GV)) return false; // If the code is position independent we will have to add a base register. if (isPositionIndependent()) return false; // Otherwise we can do it. return true; } //===----------------------------------------------------------------------===// // Optimization Methods //===----------------------------------------------------------------------===// /// If the specified instruction has a constant integer operand and there are /// bits set in that constant that are not demanded, then clear those bits and /// return true. bool TargetLowering::ShrinkDemandedConstant(SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, TargetLoweringOpt &TLO) const { SDLoc DL(Op); unsigned Opcode = Op.getOpcode(); // Early-out if we've ended up calling an undemanded node, leave this to // constant folding. if (DemandedBits.isZero() || DemandedElts.isZero()) return false; // Do target-specific constant optimization. if (targetShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO)) return TLO.New.getNode(); // FIXME: ISD::SELECT, ISD::SELECT_CC switch (Opcode) { default: break; case ISD::XOR: case ISD::AND: case ISD::OR: { auto *Op1C = dyn_cast(Op.getOperand(1)); if (!Op1C || Op1C->isOpaque()) return false; // If this is a 'not' op, don't touch it because that's a canonical form. const APInt &C = Op1C->getAPIntValue(); if (Opcode == ISD::XOR && DemandedBits.isSubsetOf(C)) return false; if (!C.isSubsetOf(DemandedBits)) { EVT VT = Op.getValueType(); SDValue NewC = TLO.DAG.getConstant(DemandedBits & C, DL, VT); SDValue NewOp = TLO.DAG.getNode(Opcode, DL, VT, Op.getOperand(0), NewC, Op->getFlags()); return TLO.CombineTo(Op, NewOp); } break; } } return false; } bool TargetLowering::ShrinkDemandedConstant(SDValue Op, const APInt &DemandedBits, TargetLoweringOpt &TLO) const { EVT VT = Op.getValueType(); APInt DemandedElts = VT.isVector() ? APInt::getAllOnes(VT.getVectorNumElements()) : APInt(1, 1); return ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO); } /// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free. /// This uses isTruncateFree/isZExtFree and ANY_EXTEND for the widening cast, /// but it could be generalized for targets with other types of implicit /// widening casts. bool TargetLowering::ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &DemandedBits, TargetLoweringOpt &TLO) const { assert(Op.getNumOperands() == 2 && "ShrinkDemandedOp only supports binary operators!"); assert(Op.getNode()->getNumValues() == 1 && "ShrinkDemandedOp only supports nodes with one result!"); EVT VT = Op.getValueType(); SelectionDAG &DAG = TLO.DAG; SDLoc dl(Op); // Early return, as this function cannot handle vector types. if (VT.isVector()) return false; assert(Op.getOperand(0).getValueType().getScalarSizeInBits() == BitWidth && Op.getOperand(1).getValueType().getScalarSizeInBits() == BitWidth && "ShrinkDemandedOp only supports operands that have the same size!"); // Don't do this if the node has another user, which may require the // full value. if (!Op.getNode()->hasOneUse()) return false; // Search for the smallest integer type with free casts to and from // Op's type. For expedience, just check power-of-2 integer types. const TargetLowering &TLI = DAG.getTargetLoweringInfo(); unsigned DemandedSize = DemandedBits.getActiveBits(); for (unsigned SmallVTBits = llvm::bit_ceil(DemandedSize); SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits); if (TLI.isTruncateFree(VT, SmallVT) && TLI.isZExtFree(SmallVT, VT)) { // We found a type with free casts. SDValue X = DAG.getNode( Op.getOpcode(), dl, SmallVT, DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(0)), DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(1))); assert(DemandedSize <= SmallVTBits && "Narrowed below demanded bits?"); SDValue Z = DAG.getNode(ISD::ANY_EXTEND, dl, VT, X); return TLO.CombineTo(Op, Z); } } return false; } bool TargetLowering::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, DAGCombinerInfo &DCI) const { SelectionDAG &DAG = DCI.DAG; TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), !DCI.isBeforeLegalizeOps()); KnownBits Known; bool Simplified = SimplifyDemandedBits(Op, DemandedBits, Known, TLO); if (Simplified) { DCI.AddToWorklist(Op.getNode()); DCI.CommitTargetLoweringOpt(TLO); } return Simplified; } bool TargetLowering::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, DAGCombinerInfo &DCI) const { SelectionDAG &DAG = DCI.DAG; TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), !DCI.isBeforeLegalizeOps()); KnownBits Known; bool Simplified = SimplifyDemandedBits(Op, DemandedBits, DemandedElts, Known, TLO); if (Simplified) { DCI.AddToWorklist(Op.getNode()); DCI.CommitTargetLoweringOpt(TLO); } return Simplified; } bool TargetLowering::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, KnownBits &Known, TargetLoweringOpt &TLO, unsigned Depth, bool AssumeSingleUse) const { EVT VT = Op.getValueType(); // Since the number of lanes in a scalable vector is unknown at compile time, // we track one bit which is implicitly broadcast to all lanes. This means // that all lanes in a scalable vector are considered demanded. APInt DemandedElts = VT.isFixedLengthVector() ? APInt::getAllOnes(VT.getVectorNumElements()) : APInt(1, 1); return SimplifyDemandedBits(Op, DemandedBits, DemandedElts, Known, TLO, Depth, AssumeSingleUse); } // TODO: Under what circumstances can we create nodes? Constant folding? SDValue TargetLowering::SimplifyMultipleUseDemandedBits( SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, SelectionDAG &DAG, unsigned Depth) const { EVT VT = Op.getValueType(); // Limit search depth. if (Depth >= SelectionDAG::MaxRecursionDepth) return SDValue(); // Ignore UNDEFs. if (Op.isUndef()) return SDValue(); // Not demanding any bits/elts from Op. if (DemandedBits == 0 || DemandedElts == 0) return DAG.getUNDEF(VT); bool IsLE = DAG.getDataLayout().isLittleEndian(); unsigned NumElts = DemandedElts.getBitWidth(); unsigned BitWidth = DemandedBits.getBitWidth(); KnownBits LHSKnown, RHSKnown; switch (Op.getOpcode()) { case ISD::BITCAST: { if (VT.isScalableVector()) return SDValue(); SDValue Src = peekThroughBitcasts(Op.getOperand(0)); EVT SrcVT = Src.getValueType(); EVT DstVT = Op.getValueType(); if (SrcVT == DstVT) return Src; unsigned NumSrcEltBits = SrcVT.getScalarSizeInBits(); unsigned NumDstEltBits = DstVT.getScalarSizeInBits(); if (NumSrcEltBits == NumDstEltBits) if (SDValue V = SimplifyMultipleUseDemandedBits( Src, DemandedBits, DemandedElts, DAG, Depth + 1)) return DAG.getBitcast(DstVT, V); if (SrcVT.isVector() && (NumDstEltBits % NumSrcEltBits) == 0) { unsigned Scale = NumDstEltBits / NumSrcEltBits; unsigned NumSrcElts = SrcVT.getVectorNumElements(); APInt DemandedSrcBits = APInt::getZero(NumSrcEltBits); APInt DemandedSrcElts = APInt::getZero(NumSrcElts); for (unsigned i = 0; i != Scale; ++i) { unsigned EltOffset = IsLE ? i : (Scale - 1 - i); unsigned BitOffset = EltOffset * NumSrcEltBits; APInt Sub = DemandedBits.extractBits(NumSrcEltBits, BitOffset); if (!Sub.isZero()) { DemandedSrcBits |= Sub; for (unsigned j = 0; j != NumElts; ++j) if (DemandedElts[j]) DemandedSrcElts.setBit((j * Scale) + i); } } if (SDValue V = SimplifyMultipleUseDemandedBits( Src, DemandedSrcBits, DemandedSrcElts, DAG, Depth + 1)) return DAG.getBitcast(DstVT, V); } // TODO - bigendian once we have test coverage. if (IsLE && (NumSrcEltBits % NumDstEltBits) == 0) { unsigned Scale = NumSrcEltBits / NumDstEltBits; unsigned NumSrcElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1; APInt DemandedSrcBits = APInt::getZero(NumSrcEltBits); APInt DemandedSrcElts = APInt::getZero(NumSrcElts); for (unsigned i = 0; i != NumElts; ++i) if (DemandedElts[i]) { unsigned Offset = (i % Scale) * NumDstEltBits; DemandedSrcBits.insertBits(DemandedBits, Offset); DemandedSrcElts.setBit(i / Scale); } if (SDValue V = SimplifyMultipleUseDemandedBits( Src, DemandedSrcBits, DemandedSrcElts, DAG, Depth + 1)) return DAG.getBitcast(DstVT, V); } break; } case ISD::FREEZE: { SDValue N0 = Op.getOperand(0); if (DAG.isGuaranteedNotToBeUndefOrPoison(N0, DemandedElts, /*PoisonOnly=*/false)) return N0; break; } case ISD::AND: { LHSKnown = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); RHSKnown = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); // If all of the demanded bits are known 1 on one side, return the other. // These bits cannot contribute to the result of the 'and' in this // context. if (DemandedBits.isSubsetOf(LHSKnown.Zero | RHSKnown.One)) return Op.getOperand(0); if (DemandedBits.isSubsetOf(RHSKnown.Zero | LHSKnown.One)) return Op.getOperand(1); break; } case ISD::OR: { LHSKnown = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); RHSKnown = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); // If all of the demanded bits are known zero on one side, return the // other. These bits cannot contribute to the result of the 'or' in this // context. if (DemandedBits.isSubsetOf(LHSKnown.One | RHSKnown.Zero)) return Op.getOperand(0); if (DemandedBits.isSubsetOf(RHSKnown.One | LHSKnown.Zero)) return Op.getOperand(1); break; } case ISD::XOR: { LHSKnown = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); RHSKnown = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); // If all of the demanded bits are known zero on one side, return the // other. if (DemandedBits.isSubsetOf(RHSKnown.Zero)) return Op.getOperand(0); if (DemandedBits.isSubsetOf(LHSKnown.Zero)) return Op.getOperand(1); break; } case ISD::SHL: { // If we are only demanding sign bits then we can use the shift source // directly. if (std::optional MaxSA = DAG.getValidMaximumShiftAmount(Op, DemandedElts, Depth + 1)) { SDValue Op0 = Op.getOperand(0); unsigned ShAmt = *MaxSA; unsigned NumSignBits = DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1); unsigned UpperDemandedBits = BitWidth - DemandedBits.countr_zero(); if (NumSignBits > ShAmt && (NumSignBits - ShAmt) >= (UpperDemandedBits)) return Op0; } break; } case ISD::SETCC: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); ISD::CondCode CC = cast(Op.getOperand(2))->get(); // If (1) we only need the sign-bit, (2) the setcc operands are the same // width as the setcc result, and (3) the result of a setcc conforms to 0 or // -1, we may be able to bypass the setcc. if (DemandedBits.isSignMask() && Op0.getScalarValueSizeInBits() == BitWidth && getBooleanContents(Op0.getValueType()) == BooleanContent::ZeroOrNegativeOneBooleanContent) { // If we're testing X < 0, then this compare isn't needed - just use X! // FIXME: We're limiting to integer types here, but this should also work // if we don't care about FP signed-zero. The use of SETLT with FP means // that we don't care about NaNs. if (CC == ISD::SETLT && Op1.getValueType().isInteger() && (isNullConstant(Op1) || ISD::isBuildVectorAllZeros(Op1.getNode()))) return Op0; } break; } case ISD::SIGN_EXTEND_INREG: { // If none of the extended bits are demanded, eliminate the sextinreg. SDValue Op0 = Op.getOperand(0); EVT ExVT = cast(Op.getOperand(1))->getVT(); unsigned ExBits = ExVT.getScalarSizeInBits(); if (DemandedBits.getActiveBits() <= ExBits && shouldRemoveRedundantExtend(Op)) return Op0; // If the input is already sign extended, just drop the extension. unsigned NumSignBits = DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1); if (NumSignBits >= (BitWidth - ExBits + 1)) return Op0; break; } case ISD::ANY_EXTEND_VECTOR_INREG: case ISD::SIGN_EXTEND_VECTOR_INREG: case ISD::ZERO_EXTEND_VECTOR_INREG: { if (VT.isScalableVector()) return SDValue(); // If we only want the lowest element and none of extended bits, then we can // return the bitcasted source vector. SDValue Src = Op.getOperand(0); EVT SrcVT = Src.getValueType(); EVT DstVT = Op.getValueType(); if (IsLE && DemandedElts == 1 && DstVT.getSizeInBits() == SrcVT.getSizeInBits() && DemandedBits.getActiveBits() <= SrcVT.getScalarSizeInBits()) { return DAG.getBitcast(DstVT, Src); } break; } case ISD::INSERT_VECTOR_ELT: { if (VT.isScalableVector()) return SDValue(); // If we don't demand the inserted element, return the base vector. SDValue Vec = Op.getOperand(0); auto *CIdx = dyn_cast(Op.getOperand(2)); EVT VecVT = Vec.getValueType(); if (CIdx && CIdx->getAPIntValue().ult(VecVT.getVectorNumElements()) && !DemandedElts[CIdx->getZExtValue()]) return Vec; break; } case ISD::INSERT_SUBVECTOR: { if (VT.isScalableVector()) return SDValue(); SDValue Vec = Op.getOperand(0); SDValue Sub = Op.getOperand(1); uint64_t Idx = Op.getConstantOperandVal(2); unsigned NumSubElts = Sub.getValueType().getVectorNumElements(); APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx); // If we don't demand the inserted subvector, return the base vector. if (DemandedSubElts == 0) return Vec; break; } case ISD::VECTOR_SHUFFLE: { assert(!VT.isScalableVector()); ArrayRef ShuffleMask = cast(Op)->getMask(); // If all the demanded elts are from one operand and are inline, // then we can use the operand directly. bool AllUndef = true, IdentityLHS = true, IdentityRHS = true; for (unsigned i = 0; i != NumElts; ++i) { int M = ShuffleMask[i]; if (M < 0 || !DemandedElts[i]) continue; AllUndef = false; IdentityLHS &= (M == (int)i); IdentityRHS &= ((M - NumElts) == i); } if (AllUndef) return DAG.getUNDEF(Op.getValueType()); if (IdentityLHS) return Op.getOperand(0); if (IdentityRHS) return Op.getOperand(1); break; } default: // TODO: Probably okay to remove after audit; here to reduce change size // in initial enablement patch for scalable vectors if (VT.isScalableVector()) return SDValue(); if (Op.getOpcode() >= ISD::BUILTIN_OP_END) if (SDValue V = SimplifyMultipleUseDemandedBitsForTargetNode( Op, DemandedBits, DemandedElts, DAG, Depth)) return V; break; } return SDValue(); } SDValue TargetLowering::SimplifyMultipleUseDemandedBits( SDValue Op, const APInt &DemandedBits, SelectionDAG &DAG, unsigned Depth) const { EVT VT = Op.getValueType(); // Since the number of lanes in a scalable vector is unknown at compile time, // we track one bit which is implicitly broadcast to all lanes. This means // that all lanes in a scalable vector are considered demanded. APInt DemandedElts = VT.isFixedLengthVector() ? APInt::getAllOnes(VT.getVectorNumElements()) : APInt(1, 1); return SimplifyMultipleUseDemandedBits(Op, DemandedBits, DemandedElts, DAG, Depth); } SDValue TargetLowering::SimplifyMultipleUseDemandedVectorElts( SDValue Op, const APInt &DemandedElts, SelectionDAG &DAG, unsigned Depth) const { APInt DemandedBits = APInt::getAllOnes(Op.getScalarValueSizeInBits()); return SimplifyMultipleUseDemandedBits(Op, DemandedBits, DemandedElts, DAG, Depth); } // Attempt to form ext(avgfloor(A, B)) from shr(add(ext(A), ext(B)), 1). // or to form ext(avgceil(A, B)) from shr(add(ext(A), ext(B), 1), 1). static SDValue combineShiftToAVG(SDValue Op, TargetLowering::TargetLoweringOpt &TLO, const TargetLowering &TLI, const APInt &DemandedBits, const APInt &DemandedElts, unsigned Depth) { assert((Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) && "SRL or SRA node is required here!"); // Is the right shift using an immediate value of 1? ConstantSDNode *N1C = isConstOrConstSplat(Op.getOperand(1), DemandedElts); if (!N1C || !N1C->isOne()) return SDValue(); // We are looking for an avgfloor // add(ext, ext) // or one of these as a avgceil // add(add(ext, ext), 1) // add(add(ext, 1), ext) // add(ext, add(ext, 1)) SDValue Add = Op.getOperand(0); if (Add.getOpcode() != ISD::ADD) return SDValue(); SDValue ExtOpA = Add.getOperand(0); SDValue ExtOpB = Add.getOperand(1); SDValue Add2; auto MatchOperands = [&](SDValue Op1, SDValue Op2, SDValue Op3, SDValue A) { ConstantSDNode *ConstOp; if ((ConstOp = isConstOrConstSplat(Op2, DemandedElts)) && ConstOp->isOne()) { ExtOpA = Op1; ExtOpB = Op3; Add2 = A; return true; } if ((ConstOp = isConstOrConstSplat(Op3, DemandedElts)) && ConstOp->isOne()) { ExtOpA = Op1; ExtOpB = Op2; Add2 = A; return true; } return false; }; bool IsCeil = (ExtOpA.getOpcode() == ISD::ADD && MatchOperands(ExtOpA.getOperand(0), ExtOpA.getOperand(1), ExtOpB, ExtOpA)) || (ExtOpB.getOpcode() == ISD::ADD && MatchOperands(ExtOpB.getOperand(0), ExtOpB.getOperand(1), ExtOpA, ExtOpB)); // If the shift is signed (sra): // - Needs >= 2 sign bit for both operands. // - Needs >= 2 zero bits. // If the shift is unsigned (srl): // - Needs >= 1 zero bit for both operands. // - Needs 1 demanded bit zero and >= 2 sign bits. SelectionDAG &DAG = TLO.DAG; unsigned ShiftOpc = Op.getOpcode(); bool IsSigned = false; unsigned KnownBits; unsigned NumSignedA = DAG.ComputeNumSignBits(ExtOpA, DemandedElts, Depth); unsigned NumSignedB = DAG.ComputeNumSignBits(ExtOpB, DemandedElts, Depth); unsigned NumSigned = std::min(NumSignedA, NumSignedB) - 1; unsigned NumZeroA = DAG.computeKnownBits(ExtOpA, DemandedElts, Depth).countMinLeadingZeros(); unsigned NumZeroB = DAG.computeKnownBits(ExtOpB, DemandedElts, Depth).countMinLeadingZeros(); unsigned NumZero = std::min(NumZeroA, NumZeroB); switch (ShiftOpc) { default: llvm_unreachable("Unexpected ShiftOpc in combineShiftToAVG"); case ISD::SRA: { if (NumZero >= 2 && NumSigned < NumZero) { IsSigned = false; KnownBits = NumZero; break; } if (NumSigned >= 1) { IsSigned = true; KnownBits = NumSigned; break; } return SDValue(); } case ISD::SRL: { if (NumZero >= 1 && NumSigned < NumZero) { IsSigned = false; KnownBits = NumZero; break; } if (NumSigned >= 1 && DemandedBits.isSignBitClear()) { IsSigned = true; KnownBits = NumSigned; break; } return SDValue(); } } unsigned AVGOpc = IsCeil ? (IsSigned ? ISD::AVGCEILS : ISD::AVGCEILU) : (IsSigned ? ISD::AVGFLOORS : ISD::AVGFLOORU); // Find the smallest power-2 type that is legal for this vector size and // operation, given the original type size and the number of known sign/zero // bits. EVT VT = Op.getValueType(); unsigned MinWidth = std::max(VT.getScalarSizeInBits() - KnownBits, 8); EVT NVT = EVT::getIntegerVT(*DAG.getContext(), llvm::bit_ceil(MinWidth)); if (NVT.getScalarSizeInBits() > VT.getScalarSizeInBits()) return SDValue(); if (VT.isVector()) NVT = EVT::getVectorVT(*DAG.getContext(), NVT, VT.getVectorElementCount()); if (TLO.LegalTypes() && !TLI.isOperationLegal(AVGOpc, NVT)) { // If we could not transform, and (both) adds are nuw/nsw, we can use the // larger type size to do the transform. if (TLO.LegalOperations() && !TLI.isOperationLegal(AVGOpc, VT)) return SDValue(); if (DAG.willNotOverflowAdd(IsSigned, Add.getOperand(0), Add.getOperand(1)) && (!Add2 || DAG.willNotOverflowAdd(IsSigned, Add2.getOperand(0), Add2.getOperand(1)))) NVT = VT; else return SDValue(); } // Don't create a AVGFLOOR node with a scalar constant unless its legal as // this is likely to stop other folds (reassociation, value tracking etc.) if (!IsCeil && !TLI.isOperationLegal(AVGOpc, NVT) && (isa(ExtOpA) || isa(ExtOpB))) return SDValue(); SDLoc DL(Op); SDValue ResultAVG = DAG.getNode(AVGOpc, DL, NVT, DAG.getExtOrTrunc(IsSigned, ExtOpA, DL, NVT), DAG.getExtOrTrunc(IsSigned, ExtOpB, DL, NVT)); return DAG.getExtOrTrunc(IsSigned, ResultAVG, DL, VT); } /// Look at Op. At this point, we know that only the OriginalDemandedBits of the /// result of Op are ever used downstream. If we can use this information to /// simplify Op, create a new simplified DAG node and return true, returning the /// original and new nodes in Old and New. Otherwise, analyze the expression and /// return a mask of Known bits for the expression (used to simplify the /// caller). The Known bits may only be accurate for those bits in the /// OriginalDemandedBits and OriginalDemandedElts. bool TargetLowering::SimplifyDemandedBits( SDValue Op, const APInt &OriginalDemandedBits, const APInt &OriginalDemandedElts, KnownBits &Known, TargetLoweringOpt &TLO, unsigned Depth, bool AssumeSingleUse) const { unsigned BitWidth = OriginalDemandedBits.getBitWidth(); assert(Op.getScalarValueSizeInBits() == BitWidth && "Mask size mismatches value type size!"); // Don't know anything. Known = KnownBits(BitWidth); EVT VT = Op.getValueType(); bool IsLE = TLO.DAG.getDataLayout().isLittleEndian(); unsigned NumElts = OriginalDemandedElts.getBitWidth(); assert((!VT.isFixedLengthVector() || NumElts == VT.getVectorNumElements()) && "Unexpected vector size"); APInt DemandedBits = OriginalDemandedBits; APInt DemandedElts = OriginalDemandedElts; SDLoc dl(Op); // Undef operand. if (Op.isUndef()) return false; // We can't simplify target constants. if (Op.getOpcode() == ISD::TargetConstant) return false; if (Op.getOpcode() == ISD::Constant) { // We know all of the bits for a constant! Known = KnownBits::makeConstant(Op->getAsAPIntVal()); return false; } if (Op.getOpcode() == ISD::ConstantFP) { // We know all of the bits for a floating point constant! Known = KnownBits::makeConstant( cast(Op)->getValueAPF().bitcastToAPInt()); return false; } // Other users may use these bits. bool HasMultiUse = false; if (!AssumeSingleUse && !Op.getNode()->hasOneUse()) { if (Depth >= SelectionDAG::MaxRecursionDepth) { // Limit search depth. return false; } // Allow multiple uses, just set the DemandedBits/Elts to all bits. DemandedBits = APInt::getAllOnes(BitWidth); DemandedElts = APInt::getAllOnes(NumElts); HasMultiUse = true; } else if (OriginalDemandedBits == 0 || OriginalDemandedElts == 0) { // Not demanding any bits/elts from Op. return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); } else if (Depth >= SelectionDAG::MaxRecursionDepth) { // Limit search depth. return false; } KnownBits Known2; switch (Op.getOpcode()) { case ISD::SCALAR_TO_VECTOR: { if (VT.isScalableVector()) return false; if (!DemandedElts[0]) return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); KnownBits SrcKnown; SDValue Src = Op.getOperand(0); unsigned SrcBitWidth = Src.getScalarValueSizeInBits(); APInt SrcDemandedBits = DemandedBits.zext(SrcBitWidth); if (SimplifyDemandedBits(Src, SrcDemandedBits, SrcKnown, TLO, Depth + 1)) return true; // Upper elements are undef, so only get the knownbits if we just demand // the bottom element. if (DemandedElts == 1) Known = SrcKnown.anyextOrTrunc(BitWidth); break; } case ISD::BUILD_VECTOR: // Collect the known bits that are shared by every demanded element. // TODO: Call SimplifyDemandedBits for non-constant demanded elements. Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); return false; // Don't fall through, will infinitely loop. case ISD::SPLAT_VECTOR: { SDValue Scl = Op.getOperand(0); APInt DemandedSclBits = DemandedBits.zextOrTrunc(Scl.getValueSizeInBits()); KnownBits KnownScl; if (SimplifyDemandedBits(Scl, DemandedSclBits, KnownScl, TLO, Depth + 1)) return true; // Implicitly truncate the bits to match the official semantics of // SPLAT_VECTOR. Known = KnownScl.trunc(BitWidth); break; } case ISD::LOAD: { auto *LD = cast(Op); if (getTargetConstantFromLoad(LD)) { Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); return false; // Don't fall through, will infinitely loop. } if (ISD::isZEXTLoad(Op.getNode()) && Op.getResNo() == 0) { // If this is a ZEXTLoad and we are looking at the loaded value. EVT MemVT = LD->getMemoryVT(); unsigned MemBits = MemVT.getScalarSizeInBits(); Known.Zero.setBitsFrom(MemBits); return false; // Don't fall through, will infinitely loop. } break; } case ISD::INSERT_VECTOR_ELT: { if (VT.isScalableVector()) return false; SDValue Vec = Op.getOperand(0); SDValue Scl = Op.getOperand(1); auto *CIdx = dyn_cast(Op.getOperand(2)); EVT VecVT = Vec.getValueType(); // If index isn't constant, assume we need all vector elements AND the // inserted element. APInt DemandedVecElts(DemandedElts); if (CIdx && CIdx->getAPIntValue().ult(VecVT.getVectorNumElements())) { unsigned Idx = CIdx->getZExtValue(); DemandedVecElts.clearBit(Idx); // Inserted element is not required. if (!DemandedElts[Idx]) return TLO.CombineTo(Op, Vec); } KnownBits KnownScl; unsigned NumSclBits = Scl.getScalarValueSizeInBits(); APInt DemandedSclBits = DemandedBits.zextOrTrunc(NumSclBits); if (SimplifyDemandedBits(Scl, DemandedSclBits, KnownScl, TLO, Depth + 1)) return true; Known = KnownScl.anyextOrTrunc(BitWidth); KnownBits KnownVec; if (SimplifyDemandedBits(Vec, DemandedBits, DemandedVecElts, KnownVec, TLO, Depth + 1)) return true; if (!!DemandedVecElts) Known = Known.intersectWith(KnownVec); return false; } case ISD::INSERT_SUBVECTOR: { if (VT.isScalableVector()) return false; // Demand any elements from the subvector and the remainder from the src its // inserted into. SDValue Src = Op.getOperand(0); SDValue Sub = Op.getOperand(1); uint64_t Idx = Op.getConstantOperandVal(2); unsigned NumSubElts = Sub.getValueType().getVectorNumElements(); APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx); APInt DemandedSrcElts = DemandedElts; DemandedSrcElts.insertBits(APInt::getZero(NumSubElts), Idx); KnownBits KnownSub, KnownSrc; if (SimplifyDemandedBits(Sub, DemandedBits, DemandedSubElts, KnownSub, TLO, Depth + 1)) return true; if (SimplifyDemandedBits(Src, DemandedBits, DemandedSrcElts, KnownSrc, TLO, Depth + 1)) return true; Known.Zero.setAllBits(); Known.One.setAllBits(); if (!!DemandedSubElts) Known = Known.intersectWith(KnownSub); if (!!DemandedSrcElts) Known = Known.intersectWith(KnownSrc); // Attempt to avoid multi-use src if we don't need anything from it. if (!DemandedBits.isAllOnes() || !DemandedSubElts.isAllOnes() || !DemandedSrcElts.isAllOnes()) { SDValue NewSub = SimplifyMultipleUseDemandedBits( Sub, DemandedBits, DemandedSubElts, TLO.DAG, Depth + 1); SDValue NewSrc = SimplifyMultipleUseDemandedBits( Src, DemandedBits, DemandedSrcElts, TLO.DAG, Depth + 1); if (NewSub || NewSrc) { NewSub = NewSub ? NewSub : Sub; NewSrc = NewSrc ? NewSrc : Src; SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, NewSrc, NewSub, Op.getOperand(2)); return TLO.CombineTo(Op, NewOp); } } break; } case ISD::EXTRACT_SUBVECTOR: { if (VT.isScalableVector()) return false; // Offset the demanded elts by the subvector index. SDValue Src = Op.getOperand(0); if (Src.getValueType().isScalableVector()) break; uint64_t Idx = Op.getConstantOperandVal(1); unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx); if (SimplifyDemandedBits(Src, DemandedBits, DemandedSrcElts, Known, TLO, Depth + 1)) return true; // Attempt to avoid multi-use src if we don't need anything from it. if (!DemandedBits.isAllOnes() || !DemandedSrcElts.isAllOnes()) { SDValue DemandedSrc = SimplifyMultipleUseDemandedBits( Src, DemandedBits, DemandedSrcElts, TLO.DAG, Depth + 1); if (DemandedSrc) { SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, DemandedSrc, Op.getOperand(1)); return TLO.CombineTo(Op, NewOp); } } break; } case ISD::CONCAT_VECTORS: { if (VT.isScalableVector()) return false; Known.Zero.setAllBits(); Known.One.setAllBits(); EVT SubVT = Op.getOperand(0).getValueType(); unsigned NumSubVecs = Op.getNumOperands(); unsigned NumSubElts = SubVT.getVectorNumElements(); for (unsigned i = 0; i != NumSubVecs; ++i) { APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, i * NumSubElts); if (SimplifyDemandedBits(Op.getOperand(i), DemandedBits, DemandedSubElts, Known2, TLO, Depth + 1)) return true; // Known bits are shared by every demanded subvector element. if (!!DemandedSubElts) Known = Known.intersectWith(Known2); } break; } case ISD::VECTOR_SHUFFLE: { assert(!VT.isScalableVector()); ArrayRef ShuffleMask = cast(Op)->getMask(); // Collect demanded elements from shuffle operands.. APInt DemandedLHS, DemandedRHS; if (!getShuffleDemandedElts(NumElts, ShuffleMask, DemandedElts, DemandedLHS, DemandedRHS)) break; if (!!DemandedLHS || !!DemandedRHS) { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); Known.Zero.setAllBits(); Known.One.setAllBits(); if (!!DemandedLHS) { if (SimplifyDemandedBits(Op0, DemandedBits, DemandedLHS, Known2, TLO, Depth + 1)) return true; Known = Known.intersectWith(Known2); } if (!!DemandedRHS) { if (SimplifyDemandedBits(Op1, DemandedBits, DemandedRHS, Known2, TLO, Depth + 1)) return true; Known = Known.intersectWith(Known2); } // Attempt to avoid multi-use ops if we don't need anything from them. SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( Op0, DemandedBits, DemandedLHS, TLO.DAG, Depth + 1); SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( Op1, DemandedBits, DemandedRHS, TLO.DAG, Depth + 1); if (DemandedOp0 || DemandedOp1) { Op0 = DemandedOp0 ? DemandedOp0 : Op0; Op1 = DemandedOp1 ? DemandedOp1 : Op1; SDValue NewOp = TLO.DAG.getVectorShuffle(VT, dl, Op0, Op1, ShuffleMask); return TLO.CombineTo(Op, NewOp); } } break; } case ISD::AND: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); // If the RHS is a constant, check to see if the LHS would be zero without // using the bits from the RHS. Below, we use knowledge about the RHS to // simplify the LHS, here we're using information from the LHS to simplify // the RHS. if (ConstantSDNode *RHSC = isConstOrConstSplat(Op1, DemandedElts)) { // Do not increment Depth here; that can cause an infinite loop. KnownBits LHSKnown = TLO.DAG.computeKnownBits(Op0, DemandedElts, Depth); // If the LHS already has zeros where RHSC does, this 'and' is dead. if ((LHSKnown.Zero & DemandedBits) == (~RHSC->getAPIntValue() & DemandedBits)) return TLO.CombineTo(Op, Op0); // If any of the set bits in the RHS are known zero on the LHS, shrink // the constant. if (ShrinkDemandedConstant(Op, ~LHSKnown.Zero & DemandedBits, DemandedElts, TLO)) return true; // Bitwise-not (xor X, -1) is a special case: we don't usually shrink its // constant, but if this 'and' is only clearing bits that were just set by // the xor, then this 'and' can be eliminated by shrinking the mask of // the xor. For example, for a 32-bit X: // and (xor (srl X, 31), -1), 1 --> xor (srl X, 31), 1 if (isBitwiseNot(Op0) && Op0.hasOneUse() && LHSKnown.One == ~RHSC->getAPIntValue()) { SDValue Xor = TLO.DAG.getNode(ISD::XOR, dl, VT, Op0.getOperand(0), Op1); return TLO.CombineTo(Op, Xor); } } // AND(INSERT_SUBVECTOR(C,X,I),M) -> INSERT_SUBVECTOR(AND(C,M),X,I) // iff 'C' is Undef/Constant and AND(X,M) == X (for DemandedBits). if (Op0.getOpcode() == ISD::INSERT_SUBVECTOR && !VT.isScalableVector() && (Op0.getOperand(0).isUndef() || ISD::isBuildVectorOfConstantSDNodes(Op0.getOperand(0).getNode())) && Op0->hasOneUse()) { unsigned NumSubElts = Op0.getOperand(1).getValueType().getVectorNumElements(); unsigned SubIdx = Op0.getConstantOperandVal(2); APInt DemandedSub = APInt::getBitsSet(NumElts, SubIdx, SubIdx + NumSubElts); KnownBits KnownSubMask = TLO.DAG.computeKnownBits(Op1, DemandedSub & DemandedElts, Depth + 1); if (DemandedBits.isSubsetOf(KnownSubMask.One)) { SDValue NewAnd = TLO.DAG.getNode(ISD::AND, dl, VT, Op0.getOperand(0), Op1); SDValue NewInsert = TLO.DAG.getNode(ISD::INSERT_SUBVECTOR, dl, VT, NewAnd, Op0.getOperand(1), Op0.getOperand(2)); return TLO.CombineTo(Op, NewInsert); } } if (SimplifyDemandedBits(Op1, DemandedBits, DemandedElts, Known, TLO, Depth + 1)) return true; if (SimplifyDemandedBits(Op0, ~Known.Zero & DemandedBits, DemandedElts, Known2, TLO, Depth + 1)) return true; // If all of the demanded bits are known one on one side, return the other. // These bits cannot contribute to the result of the 'and'. if (DemandedBits.isSubsetOf(Known2.Zero | Known.One)) return TLO.CombineTo(Op, Op0); if (DemandedBits.isSubsetOf(Known.Zero | Known2.One)) return TLO.CombineTo(Op, Op1); // If all of the demanded bits in the inputs are known zeros, return zero. if (DemandedBits.isSubsetOf(Known.Zero | Known2.Zero)) return TLO.CombineTo(Op, TLO.DAG.getConstant(0, dl, VT)); // If the RHS is a constant, see if we can simplify it. if (ShrinkDemandedConstant(Op, ~Known2.Zero & DemandedBits, DemandedElts, TLO)) return true; // If the operation can be done in a smaller type, do so. if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) return true; // Attempt to avoid multi-use ops if we don't need anything from them. if (!DemandedBits.isAllOnes() || !DemandedElts.isAllOnes()) { SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( Op0, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( Op1, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); if (DemandedOp0 || DemandedOp1) { Op0 = DemandedOp0 ? DemandedOp0 : Op0; Op1 = DemandedOp1 ? DemandedOp1 : Op1; SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1); return TLO.CombineTo(Op, NewOp); } } Known &= Known2; break; } case ISD::OR: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); SDNodeFlags Flags = Op.getNode()->getFlags(); if (SimplifyDemandedBits(Op1, DemandedBits, DemandedElts, Known, TLO, Depth + 1)) { if (Flags.hasDisjoint()) { Flags.setDisjoint(false); Op->setFlags(Flags); } return true; } if (SimplifyDemandedBits(Op0, ~Known.One & DemandedBits, DemandedElts, Known2, TLO, Depth + 1)) { if (Flags.hasDisjoint()) { Flags.setDisjoint(false); Op->setFlags(Flags); } return true; } // If all of the demanded bits are known zero on one side, return the other. // These bits cannot contribute to the result of the 'or'. if (DemandedBits.isSubsetOf(Known2.One | Known.Zero)) return TLO.CombineTo(Op, Op0); if (DemandedBits.isSubsetOf(Known.One | Known2.Zero)) return TLO.CombineTo(Op, Op1); // If the RHS is a constant, see if we can simplify it. if (ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO)) return true; // If the operation can be done in a smaller type, do so. if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) return true; // Attempt to avoid multi-use ops if we don't need anything from them. if (!DemandedBits.isAllOnes() || !DemandedElts.isAllOnes()) { SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( Op0, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( Op1, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); if (DemandedOp0 || DemandedOp1) { Op0 = DemandedOp0 ? DemandedOp0 : Op0; Op1 = DemandedOp1 ? DemandedOp1 : Op1; SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1); return TLO.CombineTo(Op, NewOp); } } // (or (and X, C1), (and (or X, Y), C2)) -> (or (and X, C1|C2), (and Y, C2)) // TODO: Use SimplifyMultipleUseDemandedBits to peek through masks. if (Op0.getOpcode() == ISD::AND && Op1.getOpcode() == ISD::AND && Op0->hasOneUse() && Op1->hasOneUse()) { // Attempt to match all commutations - m_c_Or would've been useful! for (int I = 0; I != 2; ++I) { SDValue X = Op.getOperand(I).getOperand(0); SDValue C1 = Op.getOperand(I).getOperand(1); SDValue Alt = Op.getOperand(1 - I).getOperand(0); SDValue C2 = Op.getOperand(1 - I).getOperand(1); if (Alt.getOpcode() == ISD::OR) { for (int J = 0; J != 2; ++J) { if (X == Alt.getOperand(J)) { SDValue Y = Alt.getOperand(1 - J); if (SDValue C12 = TLO.DAG.FoldConstantArithmetic(ISD::OR, dl, VT, {C1, C2})) { SDValue MaskX = TLO.DAG.getNode(ISD::AND, dl, VT, X, C12); SDValue MaskY = TLO.DAG.getNode(ISD::AND, dl, VT, Y, C2); return TLO.CombineTo( Op, TLO.DAG.getNode(ISD::OR, dl, VT, MaskX, MaskY)); } } } } } } Known |= Known2; break; } case ISD::XOR: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); if (SimplifyDemandedBits(Op1, DemandedBits, DemandedElts, Known, TLO, Depth + 1)) return true; if (SimplifyDemandedBits(Op0, DemandedBits, DemandedElts, Known2, TLO, Depth + 1)) return true; // If all of the demanded bits are known zero on one side, return the other. // These bits cannot contribute to the result of the 'xor'. if (DemandedBits.isSubsetOf(Known.Zero)) return TLO.CombineTo(Op, Op0); if (DemandedBits.isSubsetOf(Known2.Zero)) return TLO.CombineTo(Op, Op1); // If the operation can be done in a smaller type, do so. if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) return true; // If all of the unknown bits are known to be zero on one side or the other // turn this into an *inclusive* or. // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 if (DemandedBits.isSubsetOf(Known.Zero | Known2.Zero)) return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, VT, Op0, Op1)); ConstantSDNode *C = isConstOrConstSplat(Op1, DemandedElts); if (C) { // If one side is a constant, and all of the set bits in the constant are // also known set on the other side, turn this into an AND, as we know // the bits will be cleared. // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 // NB: it is okay if more bits are known than are requested if (C->getAPIntValue() == Known2.One) { SDValue ANDC = TLO.DAG.getConstant(~C->getAPIntValue() & DemandedBits, dl, VT); return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, Op0, ANDC)); } // If the RHS is a constant, see if we can change it. Don't alter a -1 // constant because that's a 'not' op, and that is better for combining // and codegen. if (!C->isAllOnes() && DemandedBits.isSubsetOf(C->getAPIntValue())) { // We're flipping all demanded bits. Flip the undemanded bits too. SDValue New = TLO.DAG.getNOT(dl, Op0, VT); return TLO.CombineTo(Op, New); } unsigned Op0Opcode = Op0.getOpcode(); if ((Op0Opcode == ISD::SRL || Op0Opcode == ISD::SHL) && Op0.hasOneUse()) { if (ConstantSDNode *ShiftC = isConstOrConstSplat(Op0.getOperand(1), DemandedElts)) { // Don't crash on an oversized shift. We can not guarantee that a // bogus shift has been simplified to undef. if (ShiftC->getAPIntValue().ult(BitWidth)) { uint64_t ShiftAmt = ShiftC->getZExtValue(); APInt Ones = APInt::getAllOnes(BitWidth); Ones = Op0Opcode == ISD::SHL ? Ones.shl(ShiftAmt) : Ones.lshr(ShiftAmt); const TargetLowering &TLI = TLO.DAG.getTargetLoweringInfo(); if ((DemandedBits & C->getAPIntValue()) == (DemandedBits & Ones) && TLI.isDesirableToCommuteXorWithShift(Op.getNode())) { // If the xor constant is a demanded mask, do a 'not' before the // shift: // xor (X << ShiftC), XorC --> (not X) << ShiftC // xor (X >> ShiftC), XorC --> (not X) >> ShiftC SDValue Not = TLO.DAG.getNOT(dl, Op0.getOperand(0), VT); return TLO.CombineTo(Op, TLO.DAG.getNode(Op0Opcode, dl, VT, Not, Op0.getOperand(1))); } } } } } // If we can't turn this into a 'not', try to shrink the constant. if (!C || !C->isAllOnes()) if (ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO)) return true; // Attempt to avoid multi-use ops if we don't need anything from them. if (!DemandedBits.isAllOnes() || !DemandedElts.isAllOnes()) { SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( Op0, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( Op1, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); if (DemandedOp0 || DemandedOp1) { Op0 = DemandedOp0 ? DemandedOp0 : Op0; Op1 = DemandedOp1 ? DemandedOp1 : Op1; SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1); return TLO.CombineTo(Op, NewOp); } } Known ^= Known2; break; } case ISD::SELECT: if (SimplifyDemandedBits(Op.getOperand(2), DemandedBits, DemandedElts, Known, TLO, Depth + 1)) return true; if (SimplifyDemandedBits(Op.getOperand(1), DemandedBits, DemandedElts, Known2, TLO, Depth + 1)) return true; // If the operands are constants, see if we can simplify them. if (ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO)) return true; // Only known if known in both the LHS and RHS. Known = Known.intersectWith(Known2); break; case ISD::VSELECT: if (SimplifyDemandedBits(Op.getOperand(2), DemandedBits, DemandedElts, Known, TLO, Depth + 1)) return true; if (SimplifyDemandedBits(Op.getOperand(1), DemandedBits, DemandedElts, Known2, TLO, Depth + 1)) return true; // Only known if known in both the LHS and RHS. Known = Known.intersectWith(Known2); break; case ISD::SELECT_CC: if (SimplifyDemandedBits(Op.getOperand(3), DemandedBits, DemandedElts, Known, TLO, Depth + 1)) return true; if (SimplifyDemandedBits(Op.getOperand(2), DemandedBits, DemandedElts, Known2, TLO, Depth + 1)) return true; // If the operands are constants, see if we can simplify them. if (ShrinkDemandedConstant(Op, DemandedBits, DemandedElts, TLO)) return true; // Only known if known in both the LHS and RHS. Known = Known.intersectWith(Known2); break; case ISD::SETCC: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); ISD::CondCode CC = cast(Op.getOperand(2))->get(); // If (1) we only need the sign-bit, (2) the setcc operands are the same // width as the setcc result, and (3) the result of a setcc conforms to 0 or // -1, we may be able to bypass the setcc. if (DemandedBits.isSignMask() && Op0.getScalarValueSizeInBits() == BitWidth && getBooleanContents(Op0.getValueType()) == BooleanContent::ZeroOrNegativeOneBooleanContent) { // If we're testing X < 0, then this compare isn't needed - just use X! // FIXME: We're limiting to integer types here, but this should also work // if we don't care about FP signed-zero. The use of SETLT with FP means // that we don't care about NaNs. if (CC == ISD::SETLT && Op1.getValueType().isInteger() && (isNullConstant(Op1) || ISD::isBuildVectorAllZeros(Op1.getNode()))) return TLO.CombineTo(Op, Op0); // TODO: Should we check for other forms of sign-bit comparisons? // Examples: X <= -1, X >= 0 } if (getBooleanContents(Op0.getValueType()) == TargetLowering::ZeroOrOneBooleanContent && BitWidth > 1) Known.Zero.setBitsFrom(1); break; } case ISD::SHL: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); EVT ShiftVT = Op1.getValueType(); if (std::optional KnownSA = TLO.DAG.getValidShiftAmount(Op, DemandedElts, Depth + 1)) { unsigned ShAmt = *KnownSA; if (ShAmt == 0) return TLO.CombineTo(Op, Op0); // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a // single shift. We can do this if the bottom bits (which are shifted // out) are never demanded. // TODO - support non-uniform vector amounts. if (Op0.getOpcode() == ISD::SRL) { if (!DemandedBits.intersects(APInt::getLowBitsSet(BitWidth, ShAmt))) { if (std::optional InnerSA = TLO.DAG.getValidShiftAmount(Op0, DemandedElts, Depth + 2)) { unsigned C1 = *InnerSA; unsigned Opc = ISD::SHL; int Diff = ShAmt - C1; if (Diff < 0) { Diff = -Diff; Opc = ISD::SRL; } SDValue NewSA = TLO.DAG.getConstant(Diff, dl, ShiftVT); return TLO.CombineTo( Op, TLO.DAG.getNode(Opc, dl, VT, Op0.getOperand(0), NewSA)); } } } // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits // are not demanded. This will likely allow the anyext to be folded away. // TODO - support non-uniform vector amounts. if (Op0.getOpcode() == ISD::ANY_EXTEND) { SDValue InnerOp = Op0.getOperand(0); EVT InnerVT = InnerOp.getValueType(); unsigned InnerBits = InnerVT.getScalarSizeInBits(); if (ShAmt < InnerBits && DemandedBits.getActiveBits() <= InnerBits && isTypeDesirableForOp(ISD::SHL, InnerVT)) { SDValue NarrowShl = TLO.DAG.getNode( ISD::SHL, dl, InnerVT, InnerOp, TLO.DAG.getShiftAmountConstant(ShAmt, InnerVT, dl)); return TLO.CombineTo( Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT, NarrowShl)); } // Repeat the SHL optimization above in cases where an extension // intervenes: (shl (anyext (shr x, c1)), c2) to // (shl (anyext x), c2-c1). This requires that the bottom c1 bits // aren't demanded (as above) and that the shifted upper c1 bits of // x aren't demanded. // TODO - support non-uniform vector amounts. if (InnerOp.getOpcode() == ISD::SRL && Op0.hasOneUse() && InnerOp.hasOneUse()) { if (std::optional SA2 = TLO.DAG.getValidShiftAmount( InnerOp, DemandedElts, Depth + 2)) { unsigned InnerShAmt = *SA2; if (InnerShAmt < ShAmt && InnerShAmt < InnerBits && DemandedBits.getActiveBits() <= (InnerBits - InnerShAmt + ShAmt) && DemandedBits.countr_zero() >= ShAmt) { SDValue NewSA = TLO.DAG.getConstant(ShAmt - InnerShAmt, dl, ShiftVT); SDValue NewExt = TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT, InnerOp.getOperand(0)); return TLO.CombineTo( Op, TLO.DAG.getNode(ISD::SHL, dl, VT, NewExt, NewSA)); } } } } APInt InDemandedMask = DemandedBits.lshr(ShAmt); if (SimplifyDemandedBits(Op0, InDemandedMask, DemandedElts, Known, TLO, Depth + 1)) { SDNodeFlags Flags = Op.getNode()->getFlags(); if (Flags.hasNoSignedWrap() || Flags.hasNoUnsignedWrap()) { // Disable the nsw and nuw flags. We can no longer guarantee that we // won't wrap after simplification. Flags.setNoSignedWrap(false); Flags.setNoUnsignedWrap(false); Op->setFlags(Flags); } return true; } Known.Zero <<= ShAmt; Known.One <<= ShAmt; // low bits known zero. Known.Zero.setLowBits(ShAmt); // Attempt to avoid multi-use ops if we don't need anything from them. if (!InDemandedMask.isAllOnes() || !DemandedElts.isAllOnes()) { SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( Op0, InDemandedMask, DemandedElts, TLO.DAG, Depth + 1); if (DemandedOp0) { SDValue NewOp = TLO.DAG.getNode(ISD::SHL, dl, VT, DemandedOp0, Op1); return TLO.CombineTo(Op, NewOp); } } // TODO: Can we merge this fold with the one below? // Try shrinking the operation as long as the shift amount will still be // in range. if (ShAmt < DemandedBits.getActiveBits() && !VT.isVector() && Op.getNode()->hasOneUse()) { // Search for the smallest integer type with free casts to and from // Op's type. For expedience, just check power-of-2 integer types. unsigned DemandedSize = DemandedBits.getActiveBits(); for (unsigned SmallVTBits = llvm::bit_ceil(DemandedSize); SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { EVT SmallVT = EVT::getIntegerVT(*TLO.DAG.getContext(), SmallVTBits); if (isNarrowingProfitable(VT, SmallVT) && isTypeDesirableForOp(ISD::SHL, SmallVT) && isTruncateFree(VT, SmallVT) && isZExtFree(SmallVT, VT) && (!TLO.LegalOperations() || isOperationLegal(ISD::SHL, SmallVT))) { assert(DemandedSize <= SmallVTBits && "Narrowed below demanded bits?"); // We found a type with free casts. SDValue NarrowShl = TLO.DAG.getNode( ISD::SHL, dl, SmallVT, TLO.DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(0)), TLO.DAG.getShiftAmountConstant(ShAmt, SmallVT, dl)); return TLO.CombineTo( Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT, NarrowShl)); } } } // Narrow shift to lower half - similar to ShrinkDemandedOp. // (shl i64:x, K) -> (i64 zero_extend (shl (i32 (trunc i64:x)), K)) // Only do this if we demand the upper half so the knownbits are correct. unsigned HalfWidth = BitWidth / 2; if ((BitWidth % 2) == 0 && !VT.isVector() && ShAmt < HalfWidth && DemandedBits.countLeadingOnes() >= HalfWidth) { EVT HalfVT = EVT::getIntegerVT(*TLO.DAG.getContext(), HalfWidth); if (isNarrowingProfitable(VT, HalfVT) && isTypeDesirableForOp(ISD::SHL, HalfVT) && isTruncateFree(VT, HalfVT) && isZExtFree(HalfVT, VT) && (!TLO.LegalOperations() || isOperationLegal(ISD::SHL, HalfVT))) { // If we're demanding the upper bits at all, we must ensure // that the upper bits of the shift result are known to be zero, // which is equivalent to the narrow shift being NUW. if (bool IsNUW = (Known.countMinLeadingZeros() >= HalfWidth)) { bool IsNSW = Known.countMinSignBits() > HalfWidth; SDNodeFlags Flags; Flags.setNoSignedWrap(IsNSW); Flags.setNoUnsignedWrap(IsNUW); SDValue NewOp = TLO.DAG.getNode(ISD::TRUNCATE, dl, HalfVT, Op0); SDValue NewShiftAmt = TLO.DAG.getShiftAmountConstant(ShAmt, HalfVT, dl); SDValue NewShift = TLO.DAG.getNode(ISD::SHL, dl, HalfVT, NewOp, NewShiftAmt, Flags); SDValue NewExt = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, VT, NewShift); return TLO.CombineTo(Op, NewExt); } } } } else { // This is a variable shift, so we can't shift the demand mask by a known // amount. But if we are not demanding high bits, then we are not // demanding those bits from the pre-shifted operand either. if (unsigned CTLZ = DemandedBits.countl_zero()) { APInt DemandedFromOp(APInt::getLowBitsSet(BitWidth, BitWidth - CTLZ)); if (SimplifyDemandedBits(Op0, DemandedFromOp, DemandedElts, Known, TLO, Depth + 1)) { SDNodeFlags Flags = Op.getNode()->getFlags(); if (Flags.hasNoSignedWrap() || Flags.hasNoUnsignedWrap()) { // Disable the nsw and nuw flags. We can no longer guarantee that we // won't wrap after simplification. Flags.setNoSignedWrap(false); Flags.setNoUnsignedWrap(false); Op->setFlags(Flags); } return true; } Known.resetAll(); } } // If we are only demanding sign bits then we can use the shift source // directly. if (std::optional MaxSA = TLO.DAG.getValidMaximumShiftAmount(Op, DemandedElts, Depth + 1)) { unsigned ShAmt = *MaxSA; unsigned NumSignBits = TLO.DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1); unsigned UpperDemandedBits = BitWidth - DemandedBits.countr_zero(); if (NumSignBits > ShAmt && (NumSignBits - ShAmt) >= (UpperDemandedBits)) return TLO.CombineTo(Op, Op0); } break; } case ISD::SRL: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); EVT ShiftVT = Op1.getValueType(); if (std::optional KnownSA = TLO.DAG.getValidShiftAmount(Op, DemandedElts, Depth + 1)) { unsigned ShAmt = *KnownSA; if (ShAmt == 0) return TLO.CombineTo(Op, Op0); // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a // single shift. We can do this if the top bits (which are shifted out) // are never demanded. // TODO - support non-uniform vector amounts. if (Op0.getOpcode() == ISD::SHL) { if (!DemandedBits.intersects(APInt::getHighBitsSet(BitWidth, ShAmt))) { if (std::optional InnerSA = TLO.DAG.getValidShiftAmount(Op0, DemandedElts, Depth + 2)) { unsigned C1 = *InnerSA; unsigned Opc = ISD::SRL; int Diff = ShAmt - C1; if (Diff < 0) { Diff = -Diff; Opc = ISD::SHL; } SDValue NewSA = TLO.DAG.getConstant(Diff, dl, ShiftVT); return TLO.CombineTo( Op, TLO.DAG.getNode(Opc, dl, VT, Op0.getOperand(0), NewSA)); } } } APInt InDemandedMask = (DemandedBits << ShAmt); // If the shift is exact, then it does demand the low bits (and knows that // they are zero). if (Op->getFlags().hasExact()) InDemandedMask.setLowBits(ShAmt); // Narrow shift to lower half - similar to ShrinkDemandedOp. // (srl i64:x, K) -> (i64 zero_extend (srl (i32 (trunc i64:x)), K)) if ((BitWidth % 2) == 0 && !VT.isVector()) { APInt HiBits = APInt::getHighBitsSet(BitWidth, BitWidth / 2); EVT HalfVT = EVT::getIntegerVT(*TLO.DAG.getContext(), BitWidth / 2); if (isNarrowingProfitable(VT, HalfVT) && isTypeDesirableForOp(ISD::SRL, HalfVT) && isTruncateFree(VT, HalfVT) && isZExtFree(HalfVT, VT) && (!TLO.LegalOperations() || isOperationLegal(ISD::SRL, HalfVT)) && ((InDemandedMask.countLeadingZeros() >= (BitWidth / 2)) || TLO.DAG.MaskedValueIsZero(Op0, HiBits))) { SDValue NewOp = TLO.DAG.getNode(ISD::TRUNCATE, dl, HalfVT, Op0); SDValue NewShiftAmt = TLO.DAG.getShiftAmountConstant(ShAmt, HalfVT, dl); SDValue NewShift = TLO.DAG.getNode(ISD::SRL, dl, HalfVT, NewOp, NewShiftAmt); return TLO.CombineTo( Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, VT, NewShift)); } } // Compute the new bits that are at the top now. if (SimplifyDemandedBits(Op0, InDemandedMask, DemandedElts, Known, TLO, Depth + 1)) return true; Known.Zero.lshrInPlace(ShAmt); Known.One.lshrInPlace(ShAmt); // High bits known zero. Known.Zero.setHighBits(ShAmt); // Attempt to avoid multi-use ops if we don't need anything from them. if (!InDemandedMask.isAllOnes() || !DemandedElts.isAllOnes()) { SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( Op0, InDemandedMask, DemandedElts, TLO.DAG, Depth + 1); if (DemandedOp0) { SDValue NewOp = TLO.DAG.getNode(ISD::SRL, dl, VT, DemandedOp0, Op1); return TLO.CombineTo(Op, NewOp); } } } else { // Use generic knownbits computation as it has support for non-uniform // shift amounts. Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); } // Try to match AVG patterns (after shift simplification). if (SDValue AVG = combineShiftToAVG(Op, TLO, *this, DemandedBits, DemandedElts, Depth + 1)) return TLO.CombineTo(Op, AVG); break; } case ISD::SRA: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); EVT ShiftVT = Op1.getValueType(); // If we only want bits that already match the signbit then we don't need // to shift. unsigned NumHiDemandedBits = BitWidth - DemandedBits.countr_zero(); if (TLO.DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1) >= NumHiDemandedBits) return TLO.CombineTo(Op, Op0); // If this is an arithmetic shift right and only the low-bit is set, we can // always convert this into a logical shr, even if the shift amount is // variable. The low bit of the shift cannot be an input sign bit unless // the shift amount is >= the size of the datatype, which is undefined. if (DemandedBits.isOne()) return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, Op1)); if (std::optional KnownSA = TLO.DAG.getValidShiftAmount(Op, DemandedElts, Depth + 1)) { unsigned ShAmt = *KnownSA; if (ShAmt == 0) return TLO.CombineTo(Op, Op0); // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target // supports sext_inreg. if (Op0.getOpcode() == ISD::SHL) { if (std::optional InnerSA = TLO.DAG.getValidShiftAmount(Op0, DemandedElts, Depth + 2)) { unsigned LowBits = BitWidth - ShAmt; EVT ExtVT = EVT::getIntegerVT(*TLO.DAG.getContext(), LowBits); if (VT.isVector()) ExtVT = EVT::getVectorVT(*TLO.DAG.getContext(), ExtVT, VT.getVectorElementCount()); if (*InnerSA == ShAmt) { if (!TLO.LegalOperations() || getOperationAction(ISD::SIGN_EXTEND_INREG, ExtVT) == Legal) return TLO.CombineTo( Op, TLO.DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, VT, Op0.getOperand(0), TLO.DAG.getValueType(ExtVT))); // Even if we can't convert to sext_inreg, we might be able to // remove this shift pair if the input is already sign extended. unsigned NumSignBits = TLO.DAG.ComputeNumSignBits(Op0.getOperand(0), DemandedElts); if (NumSignBits > ShAmt) return TLO.CombineTo(Op, Op0.getOperand(0)); } } } APInt InDemandedMask = (DemandedBits << ShAmt); // If the shift is exact, then it does demand the low bits (and knows that // they are zero). if (Op->getFlags().hasExact()) InDemandedMask.setLowBits(ShAmt); // If any of the demanded bits are produced by the sign extension, we also // demand the input sign bit. if (DemandedBits.countl_zero() < ShAmt) InDemandedMask.setSignBit(); if (SimplifyDemandedBits(Op0, InDemandedMask, DemandedElts, Known, TLO, Depth + 1)) return true; Known.Zero.lshrInPlace(ShAmt); Known.One.lshrInPlace(ShAmt); // If the input sign bit is known to be zero, or if none of the top bits // are demanded, turn this into an unsigned shift right. if (Known.Zero[BitWidth - ShAmt - 1] || DemandedBits.countl_zero() >= ShAmt) { SDNodeFlags Flags; Flags.setExact(Op->getFlags().hasExact()); return TLO.CombineTo( Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, Op1, Flags)); } int Log2 = DemandedBits.exactLogBase2(); if (Log2 >= 0) { // The bit must come from the sign. SDValue NewSA = TLO.DAG.getConstant(BitWidth - 1 - Log2, dl, ShiftVT); return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, NewSA)); } if (Known.One[BitWidth - ShAmt - 1]) // New bits are known one. Known.One.setHighBits(ShAmt); // Attempt to avoid multi-use ops if we don't need anything from them. if (!InDemandedMask.isAllOnes() || !DemandedElts.isAllOnes()) { SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( Op0, InDemandedMask, DemandedElts, TLO.DAG, Depth + 1); if (DemandedOp0) { SDValue NewOp = TLO.DAG.getNode(ISD::SRA, dl, VT, DemandedOp0, Op1); return TLO.CombineTo(Op, NewOp); } } } // Try to match AVG patterns (after shift simplification). if (SDValue AVG = combineShiftToAVG(Op, TLO, *this, DemandedBits, DemandedElts, Depth + 1)) return TLO.CombineTo(Op, AVG); break; } case ISD::FSHL: case ISD::FSHR: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); SDValue Op2 = Op.getOperand(2); bool IsFSHL = (Op.getOpcode() == ISD::FSHL); if (ConstantSDNode *SA = isConstOrConstSplat(Op2, DemandedElts)) { unsigned Amt = SA->getAPIntValue().urem(BitWidth); // For fshl, 0-shift returns the 1st arg. // For fshr, 0-shift returns the 2nd arg. if (Amt == 0) { if (SimplifyDemandedBits(IsFSHL ? Op0 : Op1, DemandedBits, DemandedElts, Known, TLO, Depth + 1)) return true; break; } // fshl: (Op0 << Amt) | (Op1 >> (BW - Amt)) // fshr: (Op0 << (BW - Amt)) | (Op1 >> Amt) APInt Demanded0 = DemandedBits.lshr(IsFSHL ? Amt : (BitWidth - Amt)); APInt Demanded1 = DemandedBits << (IsFSHL ? (BitWidth - Amt) : Amt); if (SimplifyDemandedBits(Op0, Demanded0, DemandedElts, Known2, TLO, Depth + 1)) return true; if (SimplifyDemandedBits(Op1, Demanded1, DemandedElts, Known, TLO, Depth + 1)) return true; Known2.One <<= (IsFSHL ? Amt : (BitWidth - Amt)); Known2.Zero <<= (IsFSHL ? Amt : (BitWidth - Amt)); Known.One.lshrInPlace(IsFSHL ? (BitWidth - Amt) : Amt); Known.Zero.lshrInPlace(IsFSHL ? (BitWidth - Amt) : Amt); Known = Known.unionWith(Known2); // Attempt to avoid multi-use ops if we don't need anything from them. if (!Demanded0.isAllOnes() || !Demanded1.isAllOnes() || !DemandedElts.isAllOnes()) { SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( Op0, Demanded0, DemandedElts, TLO.DAG, Depth + 1); SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( Op1, Demanded1, DemandedElts, TLO.DAG, Depth + 1); if (DemandedOp0 || DemandedOp1) { DemandedOp0 = DemandedOp0 ? DemandedOp0 : Op0; DemandedOp1 = DemandedOp1 ? DemandedOp1 : Op1; SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, DemandedOp0, DemandedOp1, Op2); return TLO.CombineTo(Op, NewOp); } } } // For pow-2 bitwidths we only demand the bottom modulo amt bits. if (isPowerOf2_32(BitWidth)) { APInt DemandedAmtBits(Op2.getScalarValueSizeInBits(), BitWidth - 1); if (SimplifyDemandedBits(Op2, DemandedAmtBits, DemandedElts, Known2, TLO, Depth + 1)) return true; } break; } case ISD::ROTL: case ISD::ROTR: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); bool IsROTL = (Op.getOpcode() == ISD::ROTL); // If we're rotating an 0/-1 value, then it stays an 0/-1 value. if (BitWidth == TLO.DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1)) return TLO.CombineTo(Op, Op0); if (ConstantSDNode *SA = isConstOrConstSplat(Op1, DemandedElts)) { unsigned Amt = SA->getAPIntValue().urem(BitWidth); unsigned RevAmt = BitWidth - Amt; // rotl: (Op0 << Amt) | (Op0 >> (BW - Amt)) // rotr: (Op0 << (BW - Amt)) | (Op0 >> Amt) APInt Demanded0 = DemandedBits.rotr(IsROTL ? Amt : RevAmt); if (SimplifyDemandedBits(Op0, Demanded0, DemandedElts, Known2, TLO, Depth + 1)) return true; // rot*(x, 0) --> x if (Amt == 0) return TLO.CombineTo(Op, Op0); // See if we don't demand either half of the rotated bits. if ((!TLO.LegalOperations() || isOperationLegal(ISD::SHL, VT)) && DemandedBits.countr_zero() >= (IsROTL ? Amt : RevAmt)) { Op1 = TLO.DAG.getConstant(IsROTL ? Amt : RevAmt, dl, Op1.getValueType()); return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT, Op0, Op1)); } if ((!TLO.LegalOperations() || isOperationLegal(ISD::SRL, VT)) && DemandedBits.countl_zero() >= (IsROTL ? RevAmt : Amt)) { Op1 = TLO.DAG.getConstant(IsROTL ? RevAmt : Amt, dl, Op1.getValueType()); return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, Op1)); } } // For pow-2 bitwidths we only demand the bottom modulo amt bits. if (isPowerOf2_32(BitWidth)) { APInt DemandedAmtBits(Op1.getScalarValueSizeInBits(), BitWidth - 1); if (SimplifyDemandedBits(Op1, DemandedAmtBits, DemandedElts, Known2, TLO, Depth + 1)) return true; } break; } case ISD::SMIN: case ISD::SMAX: case ISD::UMIN: case ISD::UMAX: { unsigned Opc = Op.getOpcode(); SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); // If we're only demanding signbits, then we can simplify to OR/AND node. unsigned BitOp = (Opc == ISD::SMIN || Opc == ISD::UMAX) ? ISD::OR : ISD::AND; unsigned NumSignBits = std::min(TLO.DAG.ComputeNumSignBits(Op0, DemandedElts, Depth + 1), TLO.DAG.ComputeNumSignBits(Op1, DemandedElts, Depth + 1)); unsigned NumDemandedUpperBits = BitWidth - DemandedBits.countr_zero(); if (NumSignBits >= NumDemandedUpperBits) return TLO.CombineTo(Op, TLO.DAG.getNode(BitOp, SDLoc(Op), VT, Op0, Op1)); // Check if one arg is always less/greater than (or equal) to the other arg. KnownBits Known0 = TLO.DAG.computeKnownBits(Op0, DemandedElts, Depth + 1); KnownBits Known1 = TLO.DAG.computeKnownBits(Op1, DemandedElts, Depth + 1); switch (Opc) { case ISD::SMIN: if (std::optional IsSLE = KnownBits::sle(Known0, Known1)) return TLO.CombineTo(Op, *IsSLE ? Op0 : Op1); if (std::optional IsSLT = KnownBits::slt(Known0, Known1)) return TLO.CombineTo(Op, *IsSLT ? Op0 : Op1); Known = KnownBits::smin(Known0, Known1); break; case ISD::SMAX: if (std::optional IsSGE = KnownBits::sge(Known0, Known1)) return TLO.CombineTo(Op, *IsSGE ? Op0 : Op1); if (std::optional IsSGT = KnownBits::sgt(Known0, Known1)) return TLO.CombineTo(Op, *IsSGT ? Op0 : Op1); Known = KnownBits::smax(Known0, Known1); break; case ISD::UMIN: if (std::optional IsULE = KnownBits::ule(Known0, Known1)) return TLO.CombineTo(Op, *IsULE ? Op0 : Op1); if (std::optional IsULT = KnownBits::ult(Known0, Known1)) return TLO.CombineTo(Op, *IsULT ? Op0 : Op1); Known = KnownBits::umin(Known0, Known1); break; case ISD::UMAX: if (std::optional IsUGE = KnownBits::uge(Known0, Known1)) return TLO.CombineTo(Op, *IsUGE ? Op0 : Op1); if (std::optional IsUGT = KnownBits::ugt(Known0, Known1)) return TLO.CombineTo(Op, *IsUGT ? Op0 : Op1); Known = KnownBits::umax(Known0, Known1); break; } break; } case ISD::BITREVERSE: { SDValue Src = Op.getOperand(0); APInt DemandedSrcBits = DemandedBits.reverseBits(); if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedElts, Known2, TLO, Depth + 1)) return true; Known.One = Known2.One.reverseBits(); Known.Zero = Known2.Zero.reverseBits(); break; } case ISD::BSWAP: { SDValue Src = Op.getOperand(0); // If the only bits demanded come from one byte of the bswap result, // just shift the input byte into position to eliminate the bswap. unsigned NLZ = DemandedBits.countl_zero(); unsigned NTZ = DemandedBits.countr_zero(); // Round NTZ down to the next byte. If we have 11 trailing zeros, then // we need all the bits down to bit 8. Likewise, round NLZ. If we // have 14 leading zeros, round to 8. NLZ = alignDown(NLZ, 8); NTZ = alignDown(NTZ, 8); // If we need exactly one byte, we can do this transformation. if (BitWidth - NLZ - NTZ == 8) { // Replace this with either a left or right shift to get the byte into // the right place. unsigned ShiftOpcode = NLZ > NTZ ? ISD::SRL : ISD::SHL; if (!TLO.LegalOperations() || isOperationLegal(ShiftOpcode, VT)) { unsigned ShiftAmount = NLZ > NTZ ? NLZ - NTZ : NTZ - NLZ; SDValue ShAmt = TLO.DAG.getShiftAmountConstant(ShiftAmount, VT, dl); SDValue NewOp = TLO.DAG.getNode(ShiftOpcode, dl, VT, Src, ShAmt); return TLO.CombineTo(Op, NewOp); } } APInt DemandedSrcBits = DemandedBits.byteSwap(); if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedElts, Known2, TLO, Depth + 1)) return true; Known.One = Known2.One.byteSwap(); Known.Zero = Known2.Zero.byteSwap(); break; } case ISD::CTPOP: { // If only 1 bit is demanded, replace with PARITY as long as we're before // op legalization. // FIXME: Limit to scalars for now. if (DemandedBits.isOne() && !TLO.LegalOps && !VT.isVector()) return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::PARITY, dl, VT, Op.getOperand(0))); Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); break; } case ISD::SIGN_EXTEND_INREG: { SDValue Op0 = Op.getOperand(0); EVT ExVT = cast(Op.getOperand(1))->getVT(); unsigned ExVTBits = ExVT.getScalarSizeInBits(); // If we only care about the highest bit, don't bother shifting right. if (DemandedBits.isSignMask()) { unsigned MinSignedBits = TLO.DAG.ComputeMaxSignificantBits(Op0, DemandedElts, Depth + 1); bool AlreadySignExtended = ExVTBits >= MinSignedBits; // However if the input is already sign extended we expect the sign // extension to be dropped altogether later and do not simplify. if (!AlreadySignExtended) { // Compute the correct shift amount type, which must be getShiftAmountTy // for scalar types after legalization. SDValue ShiftAmt = TLO.DAG.getShiftAmountConstant(BitWidth - ExVTBits, VT, dl); return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT, Op0, ShiftAmt)); } } // If none of the extended bits are demanded, eliminate the sextinreg. if (DemandedBits.getActiveBits() <= ExVTBits) return TLO.CombineTo(Op, Op0); APInt InputDemandedBits = DemandedBits.getLoBits(ExVTBits); // Since the sign extended bits are demanded, we know that the sign // bit is demanded. InputDemandedBits.setBit(ExVTBits - 1); if (SimplifyDemandedBits(Op0, InputDemandedBits, DemandedElts, Known, TLO, Depth + 1)) return true; // If the sign bit of the input is known set or clear, then we know the // top bits of the result. // If the input sign bit is known zero, convert this into a zero extension. if (Known.Zero[ExVTBits - 1]) return TLO.CombineTo(Op, TLO.DAG.getZeroExtendInReg(Op0, dl, ExVT)); APInt Mask = APInt::getLowBitsSet(BitWidth, ExVTBits); if (Known.One[ExVTBits - 1]) { // Input sign bit known set Known.One.setBitsFrom(ExVTBits); Known.Zero &= Mask; } else { // Input sign bit unknown Known.Zero &= Mask; Known.One &= Mask; } break; } case ISD::BUILD_PAIR: { EVT HalfVT = Op.getOperand(0).getValueType(); unsigned HalfBitWidth = HalfVT.getScalarSizeInBits(); APInt MaskLo = DemandedBits.getLoBits(HalfBitWidth).trunc(HalfBitWidth); APInt MaskHi = DemandedBits.getHiBits(HalfBitWidth).trunc(HalfBitWidth); KnownBits KnownLo, KnownHi; if (SimplifyDemandedBits(Op.getOperand(0), MaskLo, KnownLo, TLO, Depth + 1)) return true; if (SimplifyDemandedBits(Op.getOperand(1), MaskHi, KnownHi, TLO, Depth + 1)) return true; Known = KnownHi.concat(KnownLo); break; } case ISD::ZERO_EXTEND_VECTOR_INREG: if (VT.isScalableVector()) return false; [[fallthrough]]; case ISD::ZERO_EXTEND: { SDValue Src = Op.getOperand(0); EVT SrcVT = Src.getValueType(); unsigned InBits = SrcVT.getScalarSizeInBits(); unsigned InElts = SrcVT.isFixedLengthVector() ? SrcVT.getVectorNumElements() : 1; bool IsVecInReg = Op.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG; // If none of the top bits are demanded, convert this into an any_extend. if (DemandedBits.getActiveBits() <= InBits) { // If we only need the non-extended bits of the bottom element // then we can just bitcast to the result. if (IsLE && IsVecInReg && DemandedElts == 1 && VT.getSizeInBits() == SrcVT.getSizeInBits()) return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src)); unsigned Opc = IsVecInReg ? ISD::ANY_EXTEND_VECTOR_INREG : ISD::ANY_EXTEND; if (!TLO.LegalOperations() || isOperationLegal(Opc, VT)) return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src)); } SDNodeFlags Flags = Op->getFlags(); APInt InDemandedBits = DemandedBits.trunc(InBits); APInt InDemandedElts = DemandedElts.zext(InElts); if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO, Depth + 1)) { if (Flags.hasNonNeg()) { Flags.setNonNeg(false); Op->setFlags(Flags); } return true; } assert(Known.getBitWidth() == InBits && "Src width has changed?"); Known = Known.zext(BitWidth); // Attempt to avoid multi-use ops if we don't need anything from them. if (SDValue NewSrc = SimplifyMultipleUseDemandedBits( Src, InDemandedBits, InDemandedElts, TLO.DAG, Depth + 1)) return TLO.CombineTo(Op, TLO.DAG.getNode(Op.getOpcode(), dl, VT, NewSrc)); break; } case ISD::SIGN_EXTEND_VECTOR_INREG: if (VT.isScalableVector()) return false; [[fallthrough]]; case ISD::SIGN_EXTEND: { SDValue Src = Op.getOperand(0); EVT SrcVT = Src.getValueType(); unsigned InBits = SrcVT.getScalarSizeInBits(); unsigned InElts = SrcVT.isFixedLengthVector() ? SrcVT.getVectorNumElements() : 1; bool IsVecInReg = Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG; APInt InDemandedElts = DemandedElts.zext(InElts); APInt InDemandedBits = DemandedBits.trunc(InBits); // Since some of the sign extended bits are demanded, we know that the sign // bit is demanded. InDemandedBits.setBit(InBits - 1); // If none of the top bits are demanded, convert this into an any_extend. if (DemandedBits.getActiveBits() <= InBits) { // If we only need the non-extended bits of the bottom element // then we can just bitcast to the result. if (IsLE && IsVecInReg && DemandedElts == 1 && VT.getSizeInBits() == SrcVT.getSizeInBits()) return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src)); // Don't lose an all signbits 0/-1 splat on targets with 0/-1 booleans. if (getBooleanContents(VT) != ZeroOrNegativeOneBooleanContent || TLO.DAG.ComputeNumSignBits(Src, InDemandedElts, Depth + 1) != InBits) { unsigned Opc = IsVecInReg ? ISD::ANY_EXTEND_VECTOR_INREG : ISD::ANY_EXTEND; if (!TLO.LegalOperations() || isOperationLegal(Opc, VT)) return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src)); } } if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO, Depth + 1)) return true; assert(Known.getBitWidth() == InBits && "Src width has changed?"); // If the sign bit is known one, the top bits match. Known = Known.sext(BitWidth); // If the sign bit is known zero, convert this to a zero extend. if (Known.isNonNegative()) { unsigned Opc = IsVecInReg ? ISD::ZERO_EXTEND_VECTOR_INREG : ISD::ZERO_EXTEND; if (!TLO.LegalOperations() || isOperationLegal(Opc, VT)) { SDNodeFlags Flags; if (!IsVecInReg) Flags.setNonNeg(true); return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src, Flags)); } } // Attempt to avoid multi-use ops if we don't need anything from them. if (SDValue NewSrc = SimplifyMultipleUseDemandedBits( Src, InDemandedBits, InDemandedElts, TLO.DAG, Depth + 1)) return TLO.CombineTo(Op, TLO.DAG.getNode(Op.getOpcode(), dl, VT, NewSrc)); break; } case ISD::ANY_EXTEND_VECTOR_INREG: if (VT.isScalableVector()) return false; [[fallthrough]]; case ISD::ANY_EXTEND: { SDValue Src = Op.getOperand(0); EVT SrcVT = Src.getValueType(); unsigned InBits = SrcVT.getScalarSizeInBits(); unsigned InElts = SrcVT.isFixedLengthVector() ? SrcVT.getVectorNumElements() : 1; bool IsVecInReg = Op.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG; // If we only need the bottom element then we can just bitcast. // TODO: Handle ANY_EXTEND? if (IsLE && IsVecInReg && DemandedElts == 1 && VT.getSizeInBits() == SrcVT.getSizeInBits()) return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src)); APInt InDemandedBits = DemandedBits.trunc(InBits); APInt InDemandedElts = DemandedElts.zext(InElts); if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO, Depth + 1)) return true; assert(Known.getBitWidth() == InBits && "Src width has changed?"); Known = Known.anyext(BitWidth); // Attempt to avoid multi-use ops if we don't need anything from them. if (SDValue NewSrc = SimplifyMultipleUseDemandedBits( Src, InDemandedBits, InDemandedElts, TLO.DAG, Depth + 1)) return TLO.CombineTo(Op, TLO.DAG.getNode(Op.getOpcode(), dl, VT, NewSrc)); break; } case ISD::TRUNCATE: { SDValue Src = Op.getOperand(0); // Simplify the input, using demanded bit information, and compute the known // zero/one bits live out. unsigned OperandBitWidth = Src.getScalarValueSizeInBits(); APInt TruncMask = DemandedBits.zext(OperandBitWidth); if (SimplifyDemandedBits(Src, TruncMask, DemandedElts, Known, TLO, Depth + 1)) return true; Known = Known.trunc(BitWidth); // Attempt to avoid multi-use ops if we don't need anything from them. if (SDValue NewSrc = SimplifyMultipleUseDemandedBits( Src, TruncMask, DemandedElts, TLO.DAG, Depth + 1)) return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::TRUNCATE, dl, VT, NewSrc)); // If the input is only used by this truncate, see if we can shrink it based // on the known demanded bits. switch (Src.getOpcode()) { default: break; case ISD::SRL: // Shrink SRL by a constant if none of the high bits shifted in are // demanded. if (TLO.LegalTypes() && !isTypeDesirableForOp(ISD::SRL, VT)) // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is // undesirable. break; if (Src.getNode()->hasOneUse()) { if (isTruncateFree(Src, VT) && !isTruncateFree(Src.getValueType(), VT)) { // If truncate is only free at trunc(srl), do not turn it into // srl(trunc). The check is done by first check the truncate is free // at Src's opcode(srl), then check the truncate is not done by // referencing sub-register. In test, if both trunc(srl) and // srl(trunc)'s trunc are free, srl(trunc) performs better. If only // trunc(srl)'s trunc is free, trunc(srl) is better. break; } std::optional ShAmtC = TLO.DAG.getValidShiftAmount(Src, DemandedElts, Depth + 2); if (!ShAmtC || *ShAmtC >= BitWidth) break; uint64_t ShVal = *ShAmtC; APInt HighBits = APInt::getHighBitsSet(OperandBitWidth, OperandBitWidth - BitWidth); HighBits.lshrInPlace(ShVal); HighBits = HighBits.trunc(BitWidth); if (!(HighBits & DemandedBits)) { // None of the shifted in bits are needed. Add a truncate of the // shift input, then shift it. SDValue NewShAmt = TLO.DAG.getShiftAmountConstant(ShVal, VT, dl); SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, VT, Src.getOperand(0)); return TLO.CombineTo( Op, TLO.DAG.getNode(ISD::SRL, dl, VT, NewTrunc, NewShAmt)); } } break; } break; } case ISD::AssertZext: { // AssertZext demands all of the high bits, plus any of the low bits // demanded by its users. EVT ZVT = cast(Op.getOperand(1))->getVT(); APInt InMask = APInt::getLowBitsSet(BitWidth, ZVT.getSizeInBits()); if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | DemandedBits, Known, TLO, Depth + 1)) return true; Known.Zero |= ~InMask; Known.One &= (~Known.Zero); break; } case ISD::EXTRACT_VECTOR_ELT: { SDValue Src = Op.getOperand(0); SDValue Idx = Op.getOperand(1); ElementCount SrcEltCnt = Src.getValueType().getVectorElementCount(); unsigned EltBitWidth = Src.getScalarValueSizeInBits(); if (SrcEltCnt.isScalable()) return false; // Demand the bits from every vector element without a constant index. unsigned NumSrcElts = SrcEltCnt.getFixedValue(); APInt DemandedSrcElts = APInt::getAllOnes(NumSrcElts); if (auto *CIdx = dyn_cast(Idx)) if (CIdx->getAPIntValue().ult(NumSrcElts)) DemandedSrcElts = APInt::getOneBitSet(NumSrcElts, CIdx->getZExtValue()); // If BitWidth > EltBitWidth the value is anyext:ed. So we do not know // anything about the extended bits. APInt DemandedSrcBits = DemandedBits; if (BitWidth > EltBitWidth) DemandedSrcBits = DemandedSrcBits.trunc(EltBitWidth); if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts, Known2, TLO, Depth + 1)) return true; // Attempt to avoid multi-use ops if we don't need anything from them. if (!DemandedSrcBits.isAllOnes() || !DemandedSrcElts.isAllOnes()) { if (SDValue DemandedSrc = SimplifyMultipleUseDemandedBits( Src, DemandedSrcBits, DemandedSrcElts, TLO.DAG, Depth + 1)) { SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, DemandedSrc, Idx); return TLO.CombineTo(Op, NewOp); } } Known = Known2; if (BitWidth > EltBitWidth) Known = Known.anyext(BitWidth); break; } case ISD::BITCAST: { if (VT.isScalableVector()) return false; SDValue Src = Op.getOperand(0); EVT SrcVT = Src.getValueType(); unsigned NumSrcEltBits = SrcVT.getScalarSizeInBits(); // If this is an FP->Int bitcast and if the sign bit is the only // thing demanded, turn this into a FGETSIGN. if (!TLO.LegalOperations() && !VT.isVector() && !SrcVT.isVector() && DemandedBits == APInt::getSignMask(Op.getValueSizeInBits()) && SrcVT.isFloatingPoint()) { bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, VT); bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32); if ((OpVTLegal || i32Legal) && VT.isSimple() && SrcVT != MVT::f16 && SrcVT != MVT::f128) { // Cannot eliminate/lower SHL for f128 yet. EVT Ty = OpVTLegal ? VT : MVT::i32; // Make a FGETSIGN + SHL to move the sign bit into the appropriate // place. We expect the SHL to be eliminated by other optimizations. SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Src); unsigned OpVTSizeInBits = Op.getValueSizeInBits(); if (!OpVTLegal && OpVTSizeInBits > 32) Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Sign); unsigned ShVal = Op.getValueSizeInBits() - 1; SDValue ShAmt = TLO.DAG.getConstant(ShVal, dl, VT); return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT, Sign, ShAmt)); } } // Bitcast from a vector using SimplifyDemanded Bits/VectorElts. // Demand the elt/bit if any of the original elts/bits are demanded. if (SrcVT.isVector() && (BitWidth % NumSrcEltBits) == 0) { unsigned Scale = BitWidth / NumSrcEltBits; unsigned NumSrcElts = SrcVT.getVectorNumElements(); APInt DemandedSrcBits = APInt::getZero(NumSrcEltBits); APInt DemandedSrcElts = APInt::getZero(NumSrcElts); for (unsigned i = 0; i != Scale; ++i) { unsigned EltOffset = IsLE ? i : (Scale - 1 - i); unsigned BitOffset = EltOffset * NumSrcEltBits; APInt Sub = DemandedBits.extractBits(NumSrcEltBits, BitOffset); if (!Sub.isZero()) { DemandedSrcBits |= Sub; for (unsigned j = 0; j != NumElts; ++j) if (DemandedElts[j]) DemandedSrcElts.setBit((j * Scale) + i); } } APInt KnownSrcUndef, KnownSrcZero; if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownSrcUndef, KnownSrcZero, TLO, Depth + 1)) return true; KnownBits KnownSrcBits; if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts, KnownSrcBits, TLO, Depth + 1)) return true; } else if (IsLE && (NumSrcEltBits % BitWidth) == 0) { // TODO - bigendian once we have test coverage. unsigned Scale = NumSrcEltBits / BitWidth; unsigned NumSrcElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1; APInt DemandedSrcBits = APInt::getZero(NumSrcEltBits); APInt DemandedSrcElts = APInt::getZero(NumSrcElts); for (unsigned i = 0; i != NumElts; ++i) if (DemandedElts[i]) { unsigned Offset = (i % Scale) * BitWidth; DemandedSrcBits.insertBits(DemandedBits, Offset); DemandedSrcElts.setBit(i / Scale); } if (SrcVT.isVector()) { APInt KnownSrcUndef, KnownSrcZero; if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownSrcUndef, KnownSrcZero, TLO, Depth + 1)) return true; } KnownBits KnownSrcBits; if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts, KnownSrcBits, TLO, Depth + 1)) return true; // Attempt to avoid multi-use ops if we don't need anything from them. if (!DemandedSrcBits.isAllOnes() || !DemandedSrcElts.isAllOnes()) { if (SDValue DemandedSrc = SimplifyMultipleUseDemandedBits( Src, DemandedSrcBits, DemandedSrcElts, TLO.DAG, Depth + 1)) { SDValue NewOp = TLO.DAG.getBitcast(VT, DemandedSrc); return TLO.CombineTo(Op, NewOp); } } } // If this is a bitcast, let computeKnownBits handle it. Only do this on a // recursive call where Known may be useful to the caller. if (Depth > 0) { Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); return false; } break; } case ISD::MUL: if (DemandedBits.isPowerOf2()) { // The LSB of X*Y is set only if (X & 1) == 1 and (Y & 1) == 1. // If we demand exactly one bit N and we have "X * (C' << N)" where C' is // odd (has LSB set), then the left-shifted low bit of X is the answer. unsigned CTZ = DemandedBits.countr_zero(); ConstantSDNode *C = isConstOrConstSplat(Op.getOperand(1), DemandedElts); if (C && C->getAPIntValue().countr_zero() == CTZ) { SDValue AmtC = TLO.DAG.getShiftAmountConstant(CTZ, VT, dl); SDValue Shl = TLO.DAG.getNode(ISD::SHL, dl, VT, Op.getOperand(0), AmtC); return TLO.CombineTo(Op, Shl); } } // For a squared value "X * X", the bottom 2 bits are 0 and X[0] because: // X * X is odd iff X is odd. // 'Quadratic Reciprocity': X * X -> 0 for bit[1] if (Op.getOperand(0) == Op.getOperand(1) && DemandedBits.ult(4)) { SDValue One = TLO.DAG.getConstant(1, dl, VT); SDValue And1 = TLO.DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), One); return TLO.CombineTo(Op, And1); } [[fallthrough]]; case ISD::ADD: case ISD::SUB: { // Add, Sub, and Mul don't demand any bits in positions beyond that // of the highest bit demanded of them. SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1); SDNodeFlags Flags = Op.getNode()->getFlags(); unsigned DemandedBitsLZ = DemandedBits.countl_zero(); APInt LoMask = APInt::getLowBitsSet(BitWidth, BitWidth - DemandedBitsLZ); KnownBits KnownOp0, KnownOp1; auto GetDemandedBitsLHSMask = [&](APInt Demanded, const KnownBits &KnownRHS) { if (Op.getOpcode() == ISD::MUL) Demanded.clearHighBits(KnownRHS.countMinTrailingZeros()); return Demanded; }; if (SimplifyDemandedBits(Op1, LoMask, DemandedElts, KnownOp1, TLO, Depth + 1) || SimplifyDemandedBits(Op0, GetDemandedBitsLHSMask(LoMask, KnownOp1), DemandedElts, KnownOp0, TLO, Depth + 1) || // See if the operation should be performed at a smaller bit width. ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) { if (Flags.hasNoSignedWrap() || Flags.hasNoUnsignedWrap()) { // Disable the nsw and nuw flags. We can no longer guarantee that we // won't wrap after simplification. Flags.setNoSignedWrap(false); Flags.setNoUnsignedWrap(false); Op->setFlags(Flags); } return true; } // neg x with only low bit demanded is simply x. if (Op.getOpcode() == ISD::SUB && DemandedBits.isOne() && isNullConstant(Op0)) return TLO.CombineTo(Op, Op1); // Attempt to avoid multi-use ops if we don't need anything from them. if (!LoMask.isAllOnes() || !DemandedElts.isAllOnes()) { SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( Op0, LoMask, DemandedElts, TLO.DAG, Depth + 1); SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( Op1, LoMask, DemandedElts, TLO.DAG, Depth + 1); if (DemandedOp0 || DemandedOp1) { Flags.setNoSignedWrap(false); Flags.setNoUnsignedWrap(false); Op0 = DemandedOp0 ? DemandedOp0 : Op0; Op1 = DemandedOp1 ? DemandedOp1 : Op1; SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1, Flags); return TLO.CombineTo(Op, NewOp); } } // If we have a constant operand, we may be able to turn it into -1 if we // do not demand the high bits. This can make the constant smaller to // encode, allow more general folding, or match specialized instruction // patterns (eg, 'blsr' on x86). Don't bother changing 1 to -1 because that // is probably not useful (and could be detrimental). ConstantSDNode *C = isConstOrConstSplat(Op1); APInt HighMask = APInt::getHighBitsSet(BitWidth, DemandedBitsLZ); if (C && !C->isAllOnes() && !C->isOne() && (C->getAPIntValue() | HighMask).isAllOnes()) { SDValue Neg1 = TLO.DAG.getAllOnesConstant(dl, VT); // Disable the nsw and nuw flags. We can no longer guarantee that we // won't wrap after simplification. Flags.setNoSignedWrap(false); Flags.setNoUnsignedWrap(false); SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Neg1, Flags); return TLO.CombineTo(Op, NewOp); } // Match a multiply with a disguised negated-power-of-2 and convert to a // an equivalent shift-left amount. // Example: (X * MulC) + Op1 --> Op1 - (X << log2(-MulC)) auto getShiftLeftAmt = [&HighMask](SDValue Mul) -> unsigned { if (Mul.getOpcode() != ISD::MUL || !Mul.hasOneUse()) return 0; // Don't touch opaque constants. Also, ignore zero and power-of-2 // multiplies. Those will get folded later. ConstantSDNode *MulC = isConstOrConstSplat(Mul.getOperand(1)); if (MulC && !MulC->isOpaque() && !MulC->isZero() && !MulC->getAPIntValue().isPowerOf2()) { APInt UnmaskedC = MulC->getAPIntValue() | HighMask; if (UnmaskedC.isNegatedPowerOf2()) return (-UnmaskedC).logBase2(); } return 0; }; auto foldMul = [&](ISD::NodeType NT, SDValue X, SDValue Y, unsigned ShlAmt) { SDValue ShlAmtC = TLO.DAG.getShiftAmountConstant(ShlAmt, VT, dl); SDValue Shl = TLO.DAG.getNode(ISD::SHL, dl, VT, X, ShlAmtC); SDValue Res = TLO.DAG.getNode(NT, dl, VT, Y, Shl); return TLO.CombineTo(Op, Res); }; if (isOperationLegalOrCustom(ISD::SHL, VT)) { if (Op.getOpcode() == ISD::ADD) { // (X * MulC) + Op1 --> Op1 - (X << log2(-MulC)) if (unsigned ShAmt = getShiftLeftAmt(Op0)) return foldMul(ISD::SUB, Op0.getOperand(0), Op1, ShAmt); // Op0 + (X * MulC) --> Op0 - (X << log2(-MulC)) if (unsigned ShAmt = getShiftLeftAmt(Op1)) return foldMul(ISD::SUB, Op1.getOperand(0), Op0, ShAmt); } if (Op.getOpcode() == ISD::SUB) { // Op0 - (X * MulC) --> Op0 + (X << log2(-MulC)) if (unsigned ShAmt = getShiftLeftAmt(Op1)) return foldMul(ISD::ADD, Op1.getOperand(0), Op0, ShAmt); } } if (Op.getOpcode() == ISD::MUL) { Known = KnownBits::mul(KnownOp0, KnownOp1); } else { // Op.getOpcode() is either ISD::ADD or ISD::SUB. Known = KnownBits::computeForAddSub( Op.getOpcode() == ISD::ADD, Flags.hasNoSignedWrap(), Flags.hasNoUnsignedWrap(), KnownOp0, KnownOp1); } break; } default: // We also ask the target about intrinsics (which could be specific to it). if (Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN) { // TODO: Probably okay to remove after audit; here to reduce change size // in initial enablement patch for scalable vectors if (Op.getValueType().isScalableVector()) break; if (SimplifyDemandedBitsForTargetNode(Op, DemandedBits, DemandedElts, Known, TLO, Depth)) return true; break; } // Just use computeKnownBits to compute output bits. Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); break; } // If we know the value of all of the demanded bits, return this as a // constant. if (!isTargetCanonicalConstantNode(Op) && DemandedBits.isSubsetOf(Known.Zero | Known.One)) { // Avoid folding to a constant if any OpaqueConstant is involved. const SDNode *N = Op.getNode(); for (SDNode *Op : llvm::make_range(SDNodeIterator::begin(N), SDNodeIterator::end(N))) { if (auto *C = dyn_cast(Op)) if (C->isOpaque()) return false; } if (VT.isInteger()) return TLO.CombineTo(Op, TLO.DAG.getConstant(Known.One, dl, VT)); if (VT.isFloatingPoint()) return TLO.CombineTo( Op, TLO.DAG.getConstantFP( APFloat(TLO.DAG.EVTToAPFloatSemantics(VT), Known.One), dl, VT)); } // A multi use 'all demanded elts' simplify failed to find any knownbits. // Try again just for the original demanded elts. // Ensure we do this AFTER constant folding above. if (HasMultiUse && Known.isUnknown() && !OriginalDemandedElts.isAllOnes()) Known = TLO.DAG.computeKnownBits(Op, OriginalDemandedElts, Depth); return false; } bool TargetLowering::SimplifyDemandedVectorElts(SDValue Op, const APInt &DemandedElts, DAGCombinerInfo &DCI) const { SelectionDAG &DAG = DCI.DAG; TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), !DCI.isBeforeLegalizeOps()); APInt KnownUndef, KnownZero; bool Simplified = SimplifyDemandedVectorElts(Op, DemandedElts, KnownUndef, KnownZero, TLO); if (Simplified) { DCI.AddToWorklist(Op.getNode()); DCI.CommitTargetLoweringOpt(TLO); } return Simplified; } /// Given a vector binary operation and known undefined elements for each input /// operand, compute whether each element of the output is undefined. static APInt getKnownUndefForVectorBinop(SDValue BO, SelectionDAG &DAG, const APInt &UndefOp0, const APInt &UndefOp1) { EVT VT = BO.getValueType(); assert(DAG.getTargetLoweringInfo().isBinOp(BO.getOpcode()) && VT.isVector() && "Vector binop only"); EVT EltVT = VT.getVectorElementType(); unsigned NumElts = VT.isFixedLengthVector() ? VT.getVectorNumElements() : 1; assert(UndefOp0.getBitWidth() == NumElts && UndefOp1.getBitWidth() == NumElts && "Bad type for undef analysis"); auto getUndefOrConstantElt = [&](SDValue V, unsigned Index, const APInt &UndefVals) { if (UndefVals[Index]) return DAG.getUNDEF(EltVT); if (auto *BV = dyn_cast(V)) { // Try hard to make sure that the getNode() call is not creating temporary // nodes. Ignore opaque integers because they do not constant fold. SDValue Elt = BV->getOperand(Index); auto *C = dyn_cast(Elt); if (isa(Elt) || Elt.isUndef() || (C && !C->isOpaque())) return Elt; } return SDValue(); }; APInt KnownUndef = APInt::getZero(NumElts); for (unsigned i = 0; i != NumElts; ++i) { // If both inputs for this element are either constant or undef and match // the element type, compute the constant/undef result for this element of // the vector. // TODO: Ideally we would use FoldConstantArithmetic() here, but that does // not handle FP constants. The code within getNode() should be refactored // to avoid the danger of creating a bogus temporary node here. SDValue C0 = getUndefOrConstantElt(BO.getOperand(0), i, UndefOp0); SDValue C1 = getUndefOrConstantElt(BO.getOperand(1), i, UndefOp1); if (C0 && C1 && C0.getValueType() == EltVT && C1.getValueType() == EltVT) if (DAG.getNode(BO.getOpcode(), SDLoc(BO), EltVT, C0, C1).isUndef()) KnownUndef.setBit(i); } return KnownUndef; } bool TargetLowering::SimplifyDemandedVectorElts( SDValue Op, const APInt &OriginalDemandedElts, APInt &KnownUndef, APInt &KnownZero, TargetLoweringOpt &TLO, unsigned Depth, bool AssumeSingleUse) const { EVT VT = Op.getValueType(); unsigned Opcode = Op.getOpcode(); APInt DemandedElts = OriginalDemandedElts; unsigned NumElts = DemandedElts.getBitWidth(); assert(VT.isVector() && "Expected vector op"); KnownUndef = KnownZero = APInt::getZero(NumElts); const TargetLowering &TLI = TLO.DAG.getTargetLoweringInfo(); if (!TLI.shouldSimplifyDemandedVectorElts(Op, TLO)) return false; // TODO: For now we assume we know nothing about scalable vectors. if (VT.isScalableVector()) return false; assert(VT.getVectorNumElements() == NumElts && "Mask size mismatches value type element count!"); // Undef operand. if (Op.isUndef()) { KnownUndef.setAllBits(); return false; } // If Op has other users, assume that all elements are needed. if (!AssumeSingleUse && !Op.getNode()->hasOneUse()) DemandedElts.setAllBits(); // Not demanding any elements from Op. if (DemandedElts == 0) { KnownUndef.setAllBits(); return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); } // Limit search depth. if (Depth >= SelectionDAG::MaxRecursionDepth) return false; SDLoc DL(Op); unsigned EltSizeInBits = VT.getScalarSizeInBits(); bool IsLE = TLO.DAG.getDataLayout().isLittleEndian(); // Helper for demanding the specified elements and all the bits of both binary // operands. auto SimplifyDemandedVectorEltsBinOp = [&](SDValue Op0, SDValue Op1) { SDValue NewOp0 = SimplifyMultipleUseDemandedVectorElts(Op0, DemandedElts, TLO.DAG, Depth + 1); SDValue NewOp1 = SimplifyMultipleUseDemandedVectorElts(Op1, DemandedElts, TLO.DAG, Depth + 1); if (NewOp0 || NewOp1) { SDValue NewOp = TLO.DAG.getNode(Opcode, SDLoc(Op), VT, NewOp0 ? NewOp0 : Op0, NewOp1 ? NewOp1 : Op1, Op->getFlags()); return TLO.CombineTo(Op, NewOp); } return false; }; switch (Opcode) { case ISD::SCALAR_TO_VECTOR: { if (!DemandedElts[0]) { KnownUndef.setAllBits(); return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); } SDValue ScalarSrc = Op.getOperand(0); if (ScalarSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { SDValue Src = ScalarSrc.getOperand(0); SDValue Idx = ScalarSrc.getOperand(1); EVT SrcVT = Src.getValueType(); ElementCount SrcEltCnt = SrcVT.getVectorElementCount(); if (SrcEltCnt.isScalable()) return false; unsigned NumSrcElts = SrcEltCnt.getFixedValue(); if (isNullConstant(Idx)) { APInt SrcDemandedElts = APInt::getOneBitSet(NumSrcElts, 0); APInt SrcUndef = KnownUndef.zextOrTrunc(NumSrcElts); APInt SrcZero = KnownZero.zextOrTrunc(NumSrcElts); if (SimplifyDemandedVectorElts(Src, SrcDemandedElts, SrcUndef, SrcZero, TLO, Depth + 1)) return true; } } KnownUndef.setHighBits(NumElts - 1); break; } case ISD::BITCAST: { SDValue Src = Op.getOperand(0); EVT SrcVT = Src.getValueType(); // We only handle vectors here. // TODO - investigate calling SimplifyDemandedBits/ComputeKnownBits? if (!SrcVT.isVector()) break; // Fast handling of 'identity' bitcasts. unsigned NumSrcElts = SrcVT.getVectorNumElements(); if (NumSrcElts == NumElts) return SimplifyDemandedVectorElts(Src, DemandedElts, KnownUndef, KnownZero, TLO, Depth + 1); APInt SrcDemandedElts, SrcZero, SrcUndef; // Bitcast from 'large element' src vector to 'small element' vector, we // must demand a source element if any DemandedElt maps to it. if ((NumElts % NumSrcElts) == 0) { unsigned Scale = NumElts / NumSrcElts; SrcDemandedElts = APIntOps::ScaleBitMask(DemandedElts, NumSrcElts); if (SimplifyDemandedVectorElts(Src, SrcDemandedElts, SrcUndef, SrcZero, TLO, Depth + 1)) return true; // Try calling SimplifyDemandedBits, converting demanded elts to the bits // of the large element. // TODO - bigendian once we have test coverage. if (IsLE) { unsigned SrcEltSizeInBits = SrcVT.getScalarSizeInBits(); APInt SrcDemandedBits = APInt::getZero(SrcEltSizeInBits); for (unsigned i = 0; i != NumElts; ++i) if (DemandedElts[i]) { unsigned Ofs = (i % Scale) * EltSizeInBits; SrcDemandedBits.setBits(Ofs, Ofs + EltSizeInBits); } KnownBits Known; if (SimplifyDemandedBits(Src, SrcDemandedBits, SrcDemandedElts, Known, TLO, Depth + 1)) return true; // The bitcast has split each wide element into a number of // narrow subelements. We have just computed the Known bits // for wide elements. See if element splitting results in // some subelements being zero. Only for demanded elements! for (unsigned SubElt = 0; SubElt != Scale; ++SubElt) { if (!Known.Zero.extractBits(EltSizeInBits, SubElt * EltSizeInBits) .isAllOnes()) continue; for (unsigned SrcElt = 0; SrcElt != NumSrcElts; ++SrcElt) { unsigned Elt = Scale * SrcElt + SubElt; if (DemandedElts[Elt]) KnownZero.setBit(Elt); } } } // If the src element is zero/undef then all the output elements will be - // only demanded elements are guaranteed to be correct. for (unsigned i = 0; i != NumSrcElts; ++i) { if (SrcDemandedElts[i]) { if (SrcZero[i]) KnownZero.setBits(i * Scale, (i + 1) * Scale); if (SrcUndef[i]) KnownUndef.setBits(i * Scale, (i + 1) * Scale); } } } // Bitcast from 'small element' src vector to 'large element' vector, we // demand all smaller source elements covered by the larger demanded element // of this vector. if ((NumSrcElts % NumElts) == 0) { unsigned Scale = NumSrcElts / NumElts; SrcDemandedElts = APIntOps::ScaleBitMask(DemandedElts, NumSrcElts); if (SimplifyDemandedVectorElts(Src, SrcDemandedElts, SrcUndef, SrcZero, TLO, Depth + 1)) return true; // If all the src elements covering an output element are zero/undef, then // the output element will be as well, assuming it was demanded. for (unsigned i = 0; i != NumElts; ++i) { if (DemandedElts[i]) { if (SrcZero.extractBits(Scale, i * Scale).isAllOnes()) KnownZero.setBit(i); if (SrcUndef.extractBits(Scale, i * Scale).isAllOnes()) KnownUndef.setBit(i); } } } break; } case ISD::FREEZE: { SDValue N0 = Op.getOperand(0); if (TLO.DAG.isGuaranteedNotToBeUndefOrPoison(N0, DemandedElts, /*PoisonOnly=*/false)) return TLO.CombineTo(Op, N0); // TODO: Replace this with the general fold from DAGCombiner::visitFREEZE // freeze(op(x, ...)) -> op(freeze(x), ...). if (N0.getOpcode() == ISD::SCALAR_TO_VECTOR && DemandedElts == 1) return TLO.CombineTo( Op, TLO.DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, TLO.DAG.getFreeze(N0.getOperand(0)))); break; } case ISD::BUILD_VECTOR: { // Check all elements and simplify any unused elements with UNDEF. if (!DemandedElts.isAllOnes()) { // Don't simplify BROADCASTS. if (llvm::any_of(Op->op_values(), [&](SDValue Elt) { return Op.getOperand(0) != Elt; })) { SmallVector Ops(Op->op_begin(), Op->op_end()); bool Updated = false; for (unsigned i = 0; i != NumElts; ++i) { if (!DemandedElts[i] && !Ops[i].isUndef()) { Ops[i] = TLO.DAG.getUNDEF(Ops[0].getValueType()); KnownUndef.setBit(i); Updated = true; } } if (Updated) return TLO.CombineTo(Op, TLO.DAG.getBuildVector(VT, DL, Ops)); } } for (unsigned i = 0; i != NumElts; ++i) { SDValue SrcOp = Op.getOperand(i); if (SrcOp.isUndef()) { KnownUndef.setBit(i); } else if (EltSizeInBits == SrcOp.getScalarValueSizeInBits() && (isNullConstant(SrcOp) || isNullFPConstant(SrcOp))) { KnownZero.setBit(i); } } break; } case ISD::CONCAT_VECTORS: { EVT SubVT = Op.getOperand(0).getValueType(); unsigned NumSubVecs = Op.getNumOperands(); unsigned NumSubElts = SubVT.getVectorNumElements(); for (unsigned i = 0; i != NumSubVecs; ++i) { SDValue SubOp = Op.getOperand(i); APInt SubElts = DemandedElts.extractBits(NumSubElts, i * NumSubElts); APInt SubUndef, SubZero; if (SimplifyDemandedVectorElts(SubOp, SubElts, SubUndef, SubZero, TLO, Depth + 1)) return true; KnownUndef.insertBits(SubUndef, i * NumSubElts); KnownZero.insertBits(SubZero, i * NumSubElts); } // Attempt to avoid multi-use ops if we don't need anything from them. if (!DemandedElts.isAllOnes()) { bool FoundNewSub = false; SmallVector DemandedSubOps; for (unsigned i = 0; i != NumSubVecs; ++i) { SDValue SubOp = Op.getOperand(i); APInt SubElts = DemandedElts.extractBits(NumSubElts, i * NumSubElts); SDValue NewSubOp = SimplifyMultipleUseDemandedVectorElts( SubOp, SubElts, TLO.DAG, Depth + 1); DemandedSubOps.push_back(NewSubOp ? NewSubOp : SubOp); FoundNewSub = NewSubOp ? true : FoundNewSub; } if (FoundNewSub) { SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, DemandedSubOps); return TLO.CombineTo(Op, NewOp); } } break; } case ISD::INSERT_SUBVECTOR: { // Demand any elements from the subvector and the remainder from the src its // inserted into. SDValue Src = Op.getOperand(0); SDValue Sub = Op.getOperand(1); uint64_t Idx = Op.getConstantOperandVal(2); unsigned NumSubElts = Sub.getValueType().getVectorNumElements(); APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx); APInt DemandedSrcElts = DemandedElts; DemandedSrcElts.insertBits(APInt::getZero(NumSubElts), Idx); APInt SubUndef, SubZero; if (SimplifyDemandedVectorElts(Sub, DemandedSubElts, SubUndef, SubZero, TLO, Depth + 1)) return true; // If none of the src operand elements are demanded, replace it with undef. if (!DemandedSrcElts && !Src.isUndef()) return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, TLO.DAG.getUNDEF(VT), Sub, Op.getOperand(2))); if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownUndef, KnownZero, TLO, Depth + 1)) return true; KnownUndef.insertBits(SubUndef, Idx); KnownZero.insertBits(SubZero, Idx); // Attempt to avoid multi-use ops if we don't need anything from them. if (!DemandedSrcElts.isAllOnes() || !DemandedSubElts.isAllOnes()) { SDValue NewSrc = SimplifyMultipleUseDemandedVectorElts( Src, DemandedSrcElts, TLO.DAG, Depth + 1); SDValue NewSub = SimplifyMultipleUseDemandedVectorElts( Sub, DemandedSubElts, TLO.DAG, Depth + 1); if (NewSrc || NewSub) { NewSrc = NewSrc ? NewSrc : Src; NewSub = NewSub ? NewSub : Sub; SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, NewSrc, NewSub, Op.getOperand(2)); return TLO.CombineTo(Op, NewOp); } } break; } case ISD::EXTRACT_SUBVECTOR: { // Offset the demanded elts by the subvector index. SDValue Src = Op.getOperand(0); if (Src.getValueType().isScalableVector()) break; uint64_t Idx = Op.getConstantOperandVal(1); unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx); APInt SrcUndef, SrcZero; if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, SrcUndef, SrcZero, TLO, Depth + 1)) return true; KnownUndef = SrcUndef.extractBits(NumElts, Idx); KnownZero = SrcZero.extractBits(NumElts, Idx); // Attempt to avoid multi-use ops if we don't need anything from them. if (!DemandedElts.isAllOnes()) { SDValue NewSrc = SimplifyMultipleUseDemandedVectorElts( Src, DemandedSrcElts, TLO.DAG, Depth + 1); if (NewSrc) { SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, NewSrc, Op.getOperand(1)); return TLO.CombineTo(Op, NewOp); } } break; } case ISD::INSERT_VECTOR_ELT: { SDValue Vec = Op.getOperand(0); SDValue Scl = Op.getOperand(1); auto *CIdx = dyn_cast(Op.getOperand(2)); // For a legal, constant insertion index, if we don't need this insertion // then strip it, else remove it from the demanded elts. if (CIdx && CIdx->getAPIntValue().ult(NumElts)) { unsigned Idx = CIdx->getZExtValue(); if (!DemandedElts[Idx]) return TLO.CombineTo(Op, Vec); APInt DemandedVecElts(DemandedElts); DemandedVecElts.clearBit(Idx); if (SimplifyDemandedVectorElts(Vec, DemandedVecElts, KnownUndef, KnownZero, TLO, Depth + 1)) return true; KnownUndef.setBitVal(Idx, Scl.isUndef()); KnownZero.setBitVal(Idx, isNullConstant(Scl) || isNullFPConstant(Scl)); break; } APInt VecUndef, VecZero; if (SimplifyDemandedVectorElts(Vec, DemandedElts, VecUndef, VecZero, TLO, Depth + 1)) return true; // Without knowing the insertion index we can't set KnownUndef/KnownZero. break; } case ISD::VSELECT: { SDValue Sel = Op.getOperand(0); SDValue LHS = Op.getOperand(1); SDValue RHS = Op.getOperand(2); // Try to transform the select condition based on the current demanded // elements. APInt UndefSel, ZeroSel; if (SimplifyDemandedVectorElts(Sel, DemandedElts, UndefSel, ZeroSel, TLO, Depth + 1)) return true; // See if we can simplify either vselect operand. APInt DemandedLHS(DemandedElts); APInt DemandedRHS(DemandedElts); APInt UndefLHS, ZeroLHS; APInt UndefRHS, ZeroRHS; if (SimplifyDemandedVectorElts(LHS, DemandedLHS, UndefLHS, ZeroLHS, TLO, Depth + 1)) return true; if (SimplifyDemandedVectorElts(RHS, DemandedRHS, UndefRHS, ZeroRHS, TLO, Depth + 1)) return true; KnownUndef = UndefLHS & UndefRHS; KnownZero = ZeroLHS & ZeroRHS; // If we know that the selected element is always zero, we don't need the // select value element. APInt DemandedSel = DemandedElts & ~KnownZero; if (DemandedSel != DemandedElts) if (SimplifyDemandedVectorElts(Sel, DemandedSel, UndefSel, ZeroSel, TLO, Depth + 1)) return true; break; } case ISD::VECTOR_SHUFFLE: { SDValue LHS = Op.getOperand(0); SDValue RHS = Op.getOperand(1); ArrayRef ShuffleMask = cast(Op)->getMask(); // Collect demanded elements from shuffle operands.. APInt DemandedLHS(NumElts, 0); APInt DemandedRHS(NumElts, 0); for (unsigned i = 0; i != NumElts; ++i) { int M = ShuffleMask[i]; if (M < 0 || !DemandedElts[i]) continue; assert(0 <= M && M < (int)(2 * NumElts) && "Shuffle index out of range"); if (M < (int)NumElts) DemandedLHS.setBit(M); else DemandedRHS.setBit(M - NumElts); } // See if we can simplify either shuffle operand. APInt UndefLHS, ZeroLHS; APInt UndefRHS, ZeroRHS; if (SimplifyDemandedVectorElts(LHS, DemandedLHS, UndefLHS, ZeroLHS, TLO, Depth + 1)) return true; if (SimplifyDemandedVectorElts(RHS, DemandedRHS, UndefRHS, ZeroRHS, TLO, Depth + 1)) return true; // Simplify mask using undef elements from LHS/RHS. bool Updated = false; bool IdentityLHS = true, IdentityRHS = true; SmallVector NewMask(ShuffleMask); for (unsigned i = 0; i != NumElts; ++i) { int &M = NewMask[i]; if (M < 0) continue; if (!DemandedElts[i] || (M < (int)NumElts && UndefLHS[M]) || (M >= (int)NumElts && UndefRHS[M - NumElts])) { Updated = true; M = -1; } IdentityLHS &= (M < 0) || (M == (int)i); IdentityRHS &= (M < 0) || ((M - NumElts) == i); } // Update legal shuffle masks based on demanded elements if it won't reduce // to Identity which can cause premature removal of the shuffle mask. if (Updated && !IdentityLHS && !IdentityRHS && !TLO.LegalOps) { SDValue LegalShuffle = buildLegalVectorShuffle(VT, DL, LHS, RHS, NewMask, TLO.DAG); if (LegalShuffle) return TLO.CombineTo(Op, LegalShuffle); } // Propagate undef/zero elements from LHS/RHS. for (unsigned i = 0; i != NumElts; ++i) { int M = ShuffleMask[i]; if (M < 0) { KnownUndef.setBit(i); } else if (M < (int)NumElts) { if (UndefLHS[M]) KnownUndef.setBit(i); if (ZeroLHS[M]) KnownZero.setBit(i); } else { if (UndefRHS[M - NumElts]) KnownUndef.setBit(i); if (ZeroRHS[M - NumElts]) KnownZero.setBit(i); } } break; } case ISD::ANY_EXTEND_VECTOR_INREG: case ISD::SIGN_EXTEND_VECTOR_INREG: case ISD::ZERO_EXTEND_VECTOR_INREG: { APInt SrcUndef, SrcZero; SDValue Src = Op.getOperand(0); unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts); if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, SrcUndef, SrcZero, TLO, Depth + 1)) return true; KnownZero = SrcZero.zextOrTrunc(NumElts); KnownUndef = SrcUndef.zextOrTrunc(NumElts); if (IsLE && Op.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG && Op.getValueSizeInBits() == Src.getValueSizeInBits() && DemandedSrcElts == 1) { // aext - if we just need the bottom element then we can bitcast. return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src)); } if (Op.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG) { // zext(undef) upper bits are guaranteed to be zero. if (DemandedElts.isSubsetOf(KnownUndef)) return TLO.CombineTo(Op, TLO.DAG.getConstant(0, SDLoc(Op), VT)); KnownUndef.clearAllBits(); // zext - if we just need the bottom element then we can mask: // zext(and(x,c)) -> and(x,c') iff the zext is the only user of the and. if (IsLE && DemandedSrcElts == 1 && Src.getOpcode() == ISD::AND && Op->isOnlyUserOf(Src.getNode()) && Op.getValueSizeInBits() == Src.getValueSizeInBits()) { SDLoc DL(Op); EVT SrcVT = Src.getValueType(); EVT SrcSVT = SrcVT.getScalarType(); SmallVector MaskElts; MaskElts.push_back(TLO.DAG.getAllOnesConstant(DL, SrcSVT)); MaskElts.append(NumSrcElts - 1, TLO.DAG.getConstant(0, DL, SrcSVT)); SDValue Mask = TLO.DAG.getBuildVector(SrcVT, DL, MaskElts); if (SDValue Fold = TLO.DAG.FoldConstantArithmetic( ISD::AND, DL, SrcVT, {Src.getOperand(1), Mask})) { Fold = TLO.DAG.getNode(ISD::AND, DL, SrcVT, Src.getOperand(0), Fold); return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Fold)); } } } break; } // TODO: There are more binop opcodes that could be handled here - MIN, // MAX, saturated math, etc. case ISD::ADD: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); if (Op0 == Op1 && Op->isOnlyUserOf(Op0.getNode())) { APInt UndefLHS, ZeroLHS; if (SimplifyDemandedVectorElts(Op0, DemandedElts, UndefLHS, ZeroLHS, TLO, Depth + 1, /*AssumeSingleUse*/ true)) return true; } [[fallthrough]]; } case ISD::AVGCEILS: case ISD::AVGCEILU: case ISD::AVGFLOORS: case ISD::AVGFLOORU: case ISD::OR: case ISD::XOR: case ISD::SUB: case ISD::FADD: case ISD::FSUB: case ISD::FMUL: case ISD::FDIV: case ISD::FREM: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); APInt UndefRHS, ZeroRHS; if (SimplifyDemandedVectorElts(Op1, DemandedElts, UndefRHS, ZeroRHS, TLO, Depth + 1)) return true; APInt UndefLHS, ZeroLHS; if (SimplifyDemandedVectorElts(Op0, DemandedElts, UndefLHS, ZeroLHS, TLO, Depth + 1)) return true; KnownZero = ZeroLHS & ZeroRHS; KnownUndef = getKnownUndefForVectorBinop(Op, TLO.DAG, UndefLHS, UndefRHS); // Attempt to avoid multi-use ops if we don't need anything from them. // TODO - use KnownUndef to relax the demandedelts? if (!DemandedElts.isAllOnes()) if (SimplifyDemandedVectorEltsBinOp(Op0, Op1)) return true; break; } case ISD::SHL: case ISD::SRL: case ISD::SRA: case ISD::ROTL: case ISD::ROTR: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); APInt UndefRHS, ZeroRHS; if (SimplifyDemandedVectorElts(Op1, DemandedElts, UndefRHS, ZeroRHS, TLO, Depth + 1)) return true; APInt UndefLHS, ZeroLHS; if (SimplifyDemandedVectorElts(Op0, DemandedElts, UndefLHS, ZeroLHS, TLO, Depth + 1)) return true; KnownZero = ZeroLHS; KnownUndef = UndefLHS & UndefRHS; // TODO: use getKnownUndefForVectorBinop? // Attempt to avoid multi-use ops if we don't need anything from them. // TODO - use KnownUndef to relax the demandedelts? if (!DemandedElts.isAllOnes()) if (SimplifyDemandedVectorEltsBinOp(Op0, Op1)) return true; break; } case ISD::MUL: case ISD::MULHU: case ISD::MULHS: case ISD::AND: { SDValue Op0 = Op.getOperand(0); SDValue Op1 = Op.getOperand(1); APInt SrcUndef, SrcZero; if (SimplifyDemandedVectorElts(Op1, DemandedElts, SrcUndef, SrcZero, TLO, Depth + 1)) return true; // If we know that a demanded element was zero in Op1 we don't need to // demand it in Op0 - its guaranteed to be zero. APInt DemandedElts0 = DemandedElts & ~SrcZero; if (SimplifyDemandedVectorElts(Op0, DemandedElts0, KnownUndef, KnownZero, TLO, Depth + 1)) return true; KnownUndef &= DemandedElts0; KnownZero &= DemandedElts0; // If every element pair has a zero/undef then just fold to zero. // fold (and x, undef) -> 0 / (and x, 0) -> 0 // fold (mul x, undef) -> 0 / (mul x, 0) -> 0 if (DemandedElts.isSubsetOf(SrcZero | KnownZero | SrcUndef | KnownUndef)) return TLO.CombineTo(Op, TLO.DAG.getConstant(0, SDLoc(Op), VT)); // If either side has a zero element, then the result element is zero, even // if the other is an UNDEF. // TODO: Extend getKnownUndefForVectorBinop to also deal with known zeros // and then handle 'and' nodes with the rest of the binop opcodes. KnownZero |= SrcZero; KnownUndef &= SrcUndef; KnownUndef &= ~KnownZero; // Attempt to avoid multi-use ops if we don't need anything from them. if (!DemandedElts.isAllOnes()) if (SimplifyDemandedVectorEltsBinOp(Op0, Op1)) return true; break; } case ISD::TRUNCATE: case ISD::SIGN_EXTEND: case ISD::ZERO_EXTEND: if (SimplifyDemandedVectorElts(Op.getOperand(0), DemandedElts, KnownUndef, KnownZero, TLO, Depth + 1)) return true; if (Op.getOpcode() == ISD::ZERO_EXTEND) { // zext(undef) upper bits are guaranteed to be zero. if (DemandedElts.isSubsetOf(KnownUndef)) return TLO.CombineTo(Op, TLO.DAG.getConstant(0, SDLoc(Op), VT)); KnownUndef.clearAllBits(); } break; default: { if (Op.getOpcode() >= ISD::BUILTIN_OP_END) { if (SimplifyDemandedVectorEltsForTargetNode(Op, DemandedElts, KnownUndef, KnownZero, TLO, Depth)) return true; } else { KnownBits Known; APInt DemandedBits = APInt::getAllOnes(EltSizeInBits); if (SimplifyDemandedBits(Op, DemandedBits, OriginalDemandedElts, Known, TLO, Depth, AssumeSingleUse)) return true; } break; } } assert((KnownUndef & KnownZero) == 0 && "Elements flagged as undef AND zero"); // Constant fold all undef cases. // TODO: Handle zero cases as well. if (DemandedElts.isSubsetOf(KnownUndef)) return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); return false; } /// Determine which of the bits specified in Mask are known to be either zero or /// one and return them in the Known. void TargetLowering::computeKnownBitsForTargetNode(const SDValue Op, KnownBits &Known, const APInt &DemandedElts, const SelectionDAG &DAG, unsigned Depth) const { assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!"); Known.resetAll(); } void TargetLowering::computeKnownBitsForTargetInstr( GISelKnownBits &Analysis, Register R, KnownBits &Known, const APInt &DemandedElts, const MachineRegisterInfo &MRI, unsigned Depth) const { Known.resetAll(); } void TargetLowering::computeKnownBitsForFrameIndex( const int FrameIdx, KnownBits &Known, const MachineFunction &MF) const { // The low bits are known zero if the pointer is aligned. Known.Zero.setLowBits(Log2(MF.getFrameInfo().getObjectAlign(FrameIdx))); } Align TargetLowering::computeKnownAlignForTargetInstr( GISelKnownBits &Analysis, Register R, const MachineRegisterInfo &MRI, unsigned Depth) const { return Align(1); } /// This method can be implemented by targets that want to expose additional /// information about sign bits to the DAG Combiner. unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, const APInt &, const SelectionDAG &, unsigned Depth) const { assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use ComputeNumSignBits if you don't know whether Op" " is a target node!"); return 1; } unsigned TargetLowering::computeNumSignBitsForTargetInstr( GISelKnownBits &Analysis, Register R, const APInt &DemandedElts, const MachineRegisterInfo &MRI, unsigned Depth) const { return 1; } bool TargetLowering::SimplifyDemandedVectorEltsForTargetNode( SDValue Op, const APInt &DemandedElts, APInt &KnownUndef, APInt &KnownZero, TargetLoweringOpt &TLO, unsigned Depth) const { assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedVectorElts if you don't know whether Op" " is a target node!"); return false; } bool TargetLowering::SimplifyDemandedBitsForTargetNode( SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, KnownBits &Known, TargetLoweringOpt &TLO, unsigned Depth) const { assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedBits if you don't know whether Op" " is a target node!"); computeKnownBitsForTargetNode(Op, Known, DemandedElts, TLO.DAG, Depth); return false; } SDValue TargetLowering::SimplifyMultipleUseDemandedBitsForTargetNode( SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, SelectionDAG &DAG, unsigned Depth) const { assert( (Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyMultipleUseDemandedBits if you don't know whether Op" " is a target node!"); return SDValue(); } SDValue TargetLowering::buildLegalVectorShuffle(EVT VT, const SDLoc &DL, SDValue N0, SDValue N1, MutableArrayRef Mask, SelectionDAG &DAG) const { bool LegalMask = isShuffleMaskLegal(Mask, VT); if (!LegalMask) { std::swap(N0, N1); ShuffleVectorSDNode::commuteMask(Mask); LegalMask = isShuffleMaskLegal(Mask, VT); } if (!LegalMask) return SDValue(); return DAG.getVectorShuffle(VT, DL, N0, N1, Mask); } const Constant *TargetLowering::getTargetConstantFromLoad(LoadSDNode*) const { return nullptr; } bool TargetLowering::isGuaranteedNotToBeUndefOrPoisonForTargetNode( SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG, bool PoisonOnly, unsigned Depth) const { assert( (Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use isGuaranteedNotToBeUndefOrPoison if you don't know whether Op" " is a target node!"); // If Op can't create undef/poison and none of its operands are undef/poison // then Op is never undef/poison. return !canCreateUndefOrPoisonForTargetNode(Op, DemandedElts, DAG, PoisonOnly, /*ConsiderFlags*/ true, Depth) && all_of(Op->ops(), [&](SDValue V) { return DAG.isGuaranteedNotToBeUndefOrPoison(V, PoisonOnly, Depth + 1); }); } bool TargetLowering::canCreateUndefOrPoisonForTargetNode( SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG, bool PoisonOnly, bool ConsiderFlags, unsigned Depth) const { assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use canCreateUndefOrPoison if you don't know whether Op" " is a target node!"); // Be conservative and return true. return true; } bool TargetLowering::isKnownNeverNaNForTargetNode(SDValue Op, const SelectionDAG &DAG, bool SNaN, unsigned Depth) const { assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use isKnownNeverNaN if you don't know whether Op" " is a target node!"); return false; } bool TargetLowering::isSplatValueForTargetNode(SDValue Op, const APInt &DemandedElts, APInt &UndefElts, const SelectionDAG &DAG, unsigned Depth) const { assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use isSplatValue if you don't know whether Op" " is a target node!"); return false; } // FIXME: Ideally, this would use ISD::isConstantSplatVector(), but that must // work with truncating build vectors and vectors with elements of less than // 8 bits. bool TargetLowering::isConstTrueVal(SDValue N) const { if (!N) return false; unsigned EltWidth; APInt CVal; if (ConstantSDNode *CN = isConstOrConstSplat(N, /*AllowUndefs=*/false, /*AllowTruncation=*/true)) { CVal = CN->getAPIntValue(); EltWidth = N.getValueType().getScalarSizeInBits(); } else return false; // If this is a truncating splat, truncate the splat value. // Otherwise, we may fail to match the expected values below. if (EltWidth < CVal.getBitWidth()) CVal = CVal.trunc(EltWidth); switch (getBooleanContents(N.getValueType())) { case UndefinedBooleanContent: return CVal[0]; case ZeroOrOneBooleanContent: return CVal.isOne(); case ZeroOrNegativeOneBooleanContent: return CVal.isAllOnes(); } llvm_unreachable("Invalid boolean contents"); } bool TargetLowering::isConstFalseVal(SDValue N) const { if (!N) return false; const ConstantSDNode *CN = dyn_cast(N); if (!CN) { const BuildVectorSDNode *BV = dyn_cast(N); if (!BV) return false; // Only interested in constant splats, we don't care about undef // elements in identifying boolean constants and getConstantSplatNode // returns NULL if all ops are undef; CN = BV->getConstantSplatNode(); if (!CN) return false; } if (getBooleanContents(N->getValueType(0)) == UndefinedBooleanContent) return !CN->getAPIntValue()[0]; return CN->isZero(); } bool TargetLowering::isExtendedTrueVal(const ConstantSDNode *N, EVT VT, bool SExt) const { if (VT == MVT::i1) return N->isOne(); TargetLowering::BooleanContent Cnt = getBooleanContents(VT); switch (Cnt) { case TargetLowering::ZeroOrOneBooleanContent: // An extended value of 1 is always true, unless its original type is i1, // in which case it will be sign extended to -1. return (N->isOne() && !SExt) || (SExt && (N->getValueType(0) != MVT::i1)); case TargetLowering::UndefinedBooleanContent: case TargetLowering::ZeroOrNegativeOneBooleanContent: return N->isAllOnes() && SExt; } llvm_unreachable("Unexpected enumeration."); } /// This helper function of SimplifySetCC tries to optimize the comparison when /// either operand of the SetCC node is a bitwise-and instruction. SDValue TargetLowering::foldSetCCWithAnd(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, const SDLoc &DL, DAGCombinerInfo &DCI) const { if (N1.getOpcode() == ISD::AND && N0.getOpcode() != ISD::AND) std::swap(N0, N1); SelectionDAG &DAG = DCI.DAG; EVT OpVT = N0.getValueType(); if (N0.getOpcode() != ISD::AND || !OpVT.isInteger() || (Cond != ISD::SETEQ && Cond != ISD::SETNE)) return SDValue(); // (X & Y) != 0 --> zextOrTrunc(X & Y) // iff everything but LSB is known zero: if (Cond == ISD::SETNE && isNullConstant(N1) && (getBooleanContents(OpVT) == TargetLowering::UndefinedBooleanContent || getBooleanContents(OpVT) == TargetLowering::ZeroOrOneBooleanContent)) { unsigned NumEltBits = OpVT.getScalarSizeInBits(); APInt UpperBits = APInt::getHighBitsSet(NumEltBits, NumEltBits - 1); if (DAG.MaskedValueIsZero(N0, UpperBits)) return DAG.getBoolExtOrTrunc(N0, DL, VT, OpVT); } // Try to eliminate a power-of-2 mask constant by converting to a signbit // test in a narrow type that we can truncate to with no cost. Examples: // (i32 X & 32768) == 0 --> (trunc X to i16) >= 0 // (i32 X & 32768) != 0 --> (trunc X to i16) < 0 // TODO: This conservatively checks for type legality on the source and // destination types. That may inhibit optimizations, but it also // allows setcc->shift transforms that may be more beneficial. auto *AndC = dyn_cast(N0.getOperand(1)); if (AndC && isNullConstant(N1) && AndC->getAPIntValue().isPowerOf2() && isTypeLegal(OpVT) && N0.hasOneUse()) { EVT NarrowVT = EVT::getIntegerVT(*DAG.getContext(), AndC->getAPIntValue().getActiveBits()); if (isTruncateFree(OpVT, NarrowVT) && isTypeLegal(NarrowVT)) { SDValue Trunc = DAG.getZExtOrTrunc(N0.getOperand(0), DL, NarrowVT); SDValue Zero = DAG.getConstant(0, DL, NarrowVT); return DAG.getSetCC(DL, VT, Trunc, Zero, Cond == ISD::SETEQ ? ISD::SETGE : ISD::SETLT); } } // Match these patterns in any of their permutations: // (X & Y) == Y // (X & Y) != Y SDValue X, Y; if (N0.getOperand(0) == N1) { X = N0.getOperand(1); Y = N0.getOperand(0); } else if (N0.getOperand(1) == N1) { X = N0.getOperand(0); Y = N0.getOperand(1); } else { return SDValue(); } // TODO: We should invert (X & Y) eq/ne 0 -> (X & Y) ne/eq Y if // `isXAndYEqZeroPreferableToXAndYEqY` is false. This is a bit difficult as // its liable to create and infinite loop. SDValue Zero = DAG.getConstant(0, DL, OpVT); if (isXAndYEqZeroPreferableToXAndYEqY(Cond, OpVT) && DAG.isKnownToBeAPowerOfTwo(Y)) { // Simplify X & Y == Y to X & Y != 0 if Y has exactly one bit set. // Note that where Y is variable and is known to have at most one bit set // (for example, if it is Z & 1) we cannot do this; the expressions are not // equivalent when Y == 0. assert(OpVT.isInteger()); Cond = ISD::getSetCCInverse(Cond, OpVT); if (DCI.isBeforeLegalizeOps() || isCondCodeLegal(Cond, N0.getSimpleValueType())) return DAG.getSetCC(DL, VT, N0, Zero, Cond); } else if (N0.hasOneUse() && hasAndNotCompare(Y)) { // If the target supports an 'and-not' or 'and-complement' logic operation, // try to use that to make a comparison operation more efficient. // But don't do this transform if the mask is a single bit because there are // more efficient ways to deal with that case (for example, 'bt' on x86 or // 'rlwinm' on PPC). // Bail out if the compare operand that we want to turn into a zero is // already a zero (otherwise, infinite loop). if (isNullConstant(Y)) return SDValue(); // Transform this into: ~X & Y == 0. SDValue NotX = DAG.getNOT(SDLoc(X), X, OpVT); SDValue NewAnd = DAG.getNode(ISD::AND, SDLoc(N0), OpVT, NotX, Y); return DAG.getSetCC(DL, VT, NewAnd, Zero, Cond); } return SDValue(); } /// There are multiple IR patterns that could be checking whether certain /// truncation of a signed number would be lossy or not. The pattern which is /// best at IR level, may not lower optimally. Thus, we want to unfold it. /// We are looking for the following pattern: (KeptBits is a constant) /// (add %x, (1 << (KeptBits-1))) srccond (1 << KeptBits) /// KeptBits won't be bitwidth(x), that will be constant-folded to true/false. /// KeptBits also can't be 1, that would have been folded to %x dstcond 0 /// We will unfold it into the natural trunc+sext pattern: /// ((%x << C) a>> C) dstcond %x /// Where C = bitwidth(x) - KeptBits and C u< bitwidth(x) SDValue TargetLowering::optimizeSetCCOfSignedTruncationCheck( EVT SCCVT, SDValue N0, SDValue N1, ISD::CondCode Cond, DAGCombinerInfo &DCI, const SDLoc &DL) const { // We must be comparing with a constant. ConstantSDNode *C1; if (!(C1 = dyn_cast(N1))) return SDValue(); // N0 should be: add %x, (1 << (KeptBits-1)) if (N0->getOpcode() != ISD::ADD) return SDValue(); // And we must be 'add'ing a constant. ConstantSDNode *C01; if (!(C01 = dyn_cast(N0->getOperand(1)))) return SDValue(); SDValue X = N0->getOperand(0); EVT XVT = X.getValueType(); // Validate constants ... APInt I1 = C1->getAPIntValue(); ISD::CondCode NewCond; if (Cond == ISD::CondCode::SETULT) { NewCond = ISD::CondCode::SETEQ; } else if (Cond == ISD::CondCode::SETULE) { NewCond = ISD::CondCode::SETEQ; // But need to 'canonicalize' the constant. I1 += 1; } else if (Cond == ISD::CondCode::SETUGT) { NewCond = ISD::CondCode::SETNE; // But need to 'canonicalize' the constant. I1 += 1; } else if (Cond == ISD::CondCode::SETUGE) { NewCond = ISD::CondCode::SETNE; } else return SDValue(); APInt I01 = C01->getAPIntValue(); auto checkConstants = [&I1, &I01]() -> bool { // Both of them must be power-of-two, and the constant from setcc is bigger. return I1.ugt(I01) && I1.isPowerOf2() && I01.isPowerOf2(); }; if (checkConstants()) { // Great, e.g. got icmp ult i16 (add i16 %x, 128), 256 } else { // What if we invert constants? (and the target predicate) I1.negate(); I01.negate(); assert(XVT.isInteger()); NewCond = getSetCCInverse(NewCond, XVT); if (!checkConstants()) return SDValue(); // Great, e.g. got icmp uge i16 (add i16 %x, -128), -256 } // They are power-of-two, so which bit is set? const unsigned KeptBits = I1.logBase2(); const unsigned KeptBitsMinusOne = I01.logBase2(); // Magic! if (KeptBits != (KeptBitsMinusOne + 1)) return SDValue(); assert(KeptBits > 0 && KeptBits < XVT.getSizeInBits() && "unreachable"); // We don't want to do this in every single case. SelectionDAG &DAG = DCI.DAG; if (!DAG.getTargetLoweringInfo().shouldTransformSignedTruncationCheck( XVT, KeptBits)) return SDValue(); // Unfold into: sext_inreg(%x) cond %x // Where 'cond' will be either 'eq' or 'ne'. SDValue SExtInReg = DAG.getNode( ISD::SIGN_EXTEND_INREG, DL, XVT, X, DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), KeptBits))); return DAG.getSetCC(DL, SCCVT, SExtInReg, X, NewCond); } // (X & (C l>>/<< Y)) ==/!= 0 --> ((X <> Y) & C) ==/!= 0 SDValue TargetLowering::optimizeSetCCByHoistingAndByConstFromLogicalShift( EVT SCCVT, SDValue N0, SDValue N1C, ISD::CondCode Cond, DAGCombinerInfo &DCI, const SDLoc &DL) const { assert(isConstOrConstSplat(N1C) && isConstOrConstSplat(N1C)->isZero() && "Should be a comparison with 0."); assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Valid only for [in]equality comparisons."); unsigned NewShiftOpcode; SDValue X, C, Y; SelectionDAG &DAG = DCI.DAG; const TargetLowering &TLI = DAG.getTargetLoweringInfo(); // Look for '(C l>>/<< Y)'. auto Match = [&NewShiftOpcode, &X, &C, &Y, &TLI, &DAG](SDValue V) { // The shift should be one-use. if (!V.hasOneUse()) return false; unsigned OldShiftOpcode = V.getOpcode(); switch (OldShiftOpcode) { case ISD::SHL: NewShiftOpcode = ISD::SRL; break; case ISD::SRL: NewShiftOpcode = ISD::SHL; break; default: return false; // must be a logical shift. } // We should be shifting a constant. // FIXME: best to use isConstantOrConstantVector(). C = V.getOperand(0); ConstantSDNode *CC = isConstOrConstSplat(C, /*AllowUndefs=*/true, /*AllowTruncation=*/true); if (!CC) return false; Y = V.getOperand(1); ConstantSDNode *XC = isConstOrConstSplat(X, /*AllowUndefs=*/true, /*AllowTruncation=*/true); return TLI.shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd( X, XC, CC, Y, OldShiftOpcode, NewShiftOpcode, DAG); }; // LHS of comparison should be an one-use 'and'. if (N0.getOpcode() != ISD::AND || !N0.hasOneUse()) return SDValue(); X = N0.getOperand(0); SDValue Mask = N0.getOperand(1); // 'and' is commutative! if (!Match(Mask)) { std::swap(X, Mask); if (!Match(Mask)) return SDValue(); } EVT VT = X.getValueType(); // Produce: // ((X 'OppositeShiftOpcode' Y) & C) Cond 0 SDValue T0 = DAG.getNode(NewShiftOpcode, DL, VT, X, Y); SDValue T1 = DAG.getNode(ISD::AND, DL, VT, T0, C); SDValue T2 = DAG.getSetCC(DL, SCCVT, T1, N1C, Cond); return T2; } /// Try to fold an equality comparison with a {add/sub/xor} binary operation as /// the 1st operand (N0). Callers are expected to swap the N0/N1 parameters to /// handle the commuted versions of these patterns. SDValue TargetLowering::foldSetCCWithBinOp(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, const SDLoc &DL, DAGCombinerInfo &DCI) const { unsigned BOpcode = N0.getOpcode(); assert((BOpcode == ISD::ADD || BOpcode == ISD::SUB || BOpcode == ISD::XOR) && "Unexpected binop"); assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Unexpected condcode"); // (X + Y) == X --> Y == 0 // (X - Y) == X --> Y == 0 // (X ^ Y) == X --> Y == 0 SelectionDAG &DAG = DCI.DAG; EVT OpVT = N0.getValueType(); SDValue X = N0.getOperand(0); SDValue Y = N0.getOperand(1); if (X == N1) return DAG.getSetCC(DL, VT, Y, DAG.getConstant(0, DL, OpVT), Cond); if (Y != N1) return SDValue(); // (X + Y) == Y --> X == 0 // (X ^ Y) == Y --> X == 0 if (BOpcode == ISD::ADD || BOpcode == ISD::XOR) return DAG.getSetCC(DL, VT, X, DAG.getConstant(0, DL, OpVT), Cond); // The shift would not be valid if the operands are boolean (i1). if (!N0.hasOneUse() || OpVT.getScalarSizeInBits() == 1) return SDValue(); // (X - Y) == Y --> X == Y << 1 SDValue One = DAG.getShiftAmountConstant(1, OpVT, DL); SDValue YShl1 = DAG.getNode(ISD::SHL, DL, N1.getValueType(), Y, One); if (!DCI.isCalledByLegalizer()) DCI.AddToWorklist(YShl1.getNode()); return DAG.getSetCC(DL, VT, X, YShl1, Cond); } static SDValue simplifySetCCWithCTPOP(const TargetLowering &TLI, EVT VT, SDValue N0, const APInt &C1, ISD::CondCode Cond, const SDLoc &dl, SelectionDAG &DAG) { // Look through truncs that don't change the value of a ctpop. // FIXME: Add vector support? Need to be careful with setcc result type below. SDValue CTPOP = N0; if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() && !VT.isVector() && N0.getScalarValueSizeInBits() > Log2_32(N0.getOperand(0).getScalarValueSizeInBits())) CTPOP = N0.getOperand(0); if (CTPOP.getOpcode() != ISD::CTPOP || !CTPOP.hasOneUse()) return SDValue(); EVT CTVT = CTPOP.getValueType(); SDValue CTOp = CTPOP.getOperand(0); // Expand a power-of-2-or-zero comparison based on ctpop: // (ctpop x) u< 2 -> (x & x-1) == 0 // (ctpop x) u> 1 -> (x & x-1) != 0 if (Cond == ISD::SETULT || Cond == ISD::SETUGT) { // Keep the CTPOP if it is a cheap vector op. if (CTVT.isVector() && TLI.isCtpopFast(CTVT)) return SDValue(); unsigned CostLimit = TLI.getCustomCtpopCost(CTVT, Cond); if (C1.ugt(CostLimit + (Cond == ISD::SETULT))) return SDValue(); if (C1 == 0 && (Cond == ISD::SETULT)) return SDValue(); // This is handled elsewhere. unsigned Passes = C1.getLimitedValue() - (Cond == ISD::SETULT); SDValue NegOne = DAG.getAllOnesConstant(dl, CTVT); SDValue Result = CTOp; for (unsigned i = 0; i < Passes; i++) { SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, Result, NegOne); Result = DAG.getNode(ISD::AND, dl, CTVT, Result, Add); } ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE; return DAG.getSetCC(dl, VT, Result, DAG.getConstant(0, dl, CTVT), CC); } // Expand a power-of-2 comparison based on ctpop if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && C1 == 1) { // Keep the CTPOP if it is cheap. if (TLI.isCtpopFast(CTVT)) return SDValue(); SDValue Zero = DAG.getConstant(0, dl, CTVT); SDValue NegOne = DAG.getAllOnesConstant(dl, CTVT); assert(CTVT.isInteger()); SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, CTOp, NegOne); // Its not uncommon for known-never-zero X to exist in (ctpop X) eq/ne 1, so // check before emitting a potentially unnecessary op. if (DAG.isKnownNeverZero(CTOp)) { // (ctpop x) == 1 --> (x & x-1) == 0 // (ctpop x) != 1 --> (x & x-1) != 0 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Add); SDValue RHS = DAG.getSetCC(dl, VT, And, Zero, Cond); return RHS; } // (ctpop x) == 1 --> (x ^ x-1) > x-1 // (ctpop x) != 1 --> (x ^ x-1) <= x-1 SDValue Xor = DAG.getNode(ISD::XOR, dl, CTVT, CTOp, Add); ISD::CondCode CmpCond = Cond == ISD::SETEQ ? ISD::SETUGT : ISD::SETULE; return DAG.getSetCC(dl, VT, Xor, Add, CmpCond); } return SDValue(); } static SDValue foldSetCCWithRotate(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, const SDLoc &dl, SelectionDAG &DAG) { if (Cond != ISD::SETEQ && Cond != ISD::SETNE) return SDValue(); auto *C1 = isConstOrConstSplat(N1, /* AllowUndefs */ true); if (!C1 || !(C1->isZero() || C1->isAllOnes())) return SDValue(); auto getRotateSource = [](SDValue X) { if (X.getOpcode() == ISD::ROTL || X.getOpcode() == ISD::ROTR) return X.getOperand(0); return SDValue(); }; // Peek through a rotated value compared against 0 or -1: // (rot X, Y) == 0/-1 --> X == 0/-1 // (rot X, Y) != 0/-1 --> X != 0/-1 if (SDValue R = getRotateSource(N0)) return DAG.getSetCC(dl, VT, R, N1, Cond); // Peek through an 'or' of a rotated value compared against 0: // or (rot X, Y), Z ==/!= 0 --> (or X, Z) ==/!= 0 // or Z, (rot X, Y) ==/!= 0 --> (or X, Z) ==/!= 0 // // TODO: Add the 'and' with -1 sibling. // TODO: Recurse through a series of 'or' ops to find the rotate. EVT OpVT = N0.getValueType(); if (N0.hasOneUse() && N0.getOpcode() == ISD::OR && C1->isZero()) { if (SDValue R = getRotateSource(N0.getOperand(0))) { SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, R, N0.getOperand(1)); return DAG.getSetCC(dl, VT, NewOr, N1, Cond); } if (SDValue R = getRotateSource(N0.getOperand(1))) { SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, R, N0.getOperand(0)); return DAG.getSetCC(dl, VT, NewOr, N1, Cond); } } return SDValue(); } static SDValue foldSetCCWithFunnelShift(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, const SDLoc &dl, SelectionDAG &DAG) { // If we are testing for all-bits-clear, we might be able to do that with // less shifting since bit-order does not matter. if (Cond != ISD::SETEQ && Cond != ISD::SETNE) return SDValue(); auto *C1 = isConstOrConstSplat(N1, /* AllowUndefs */ true); if (!C1 || !C1->isZero()) return SDValue(); if (!N0.hasOneUse() || (N0.getOpcode() != ISD::FSHL && N0.getOpcode() != ISD::FSHR)) return SDValue(); unsigned BitWidth = N0.getScalarValueSizeInBits(); auto *ShAmtC = isConstOrConstSplat(N0.getOperand(2)); if (!ShAmtC || ShAmtC->getAPIntValue().uge(BitWidth)) return SDValue(); // Canonicalize fshr as fshl to reduce pattern-matching. unsigned ShAmt = ShAmtC->getZExtValue(); if (N0.getOpcode() == ISD::FSHR) ShAmt = BitWidth - ShAmt; // Match an 'or' with a specific operand 'Other' in either commuted variant. SDValue X, Y; auto matchOr = [&X, &Y](SDValue Or, SDValue Other) { if (Or.getOpcode() != ISD::OR || !Or.hasOneUse()) return false; if (Or.getOperand(0) == Other) { X = Or.getOperand(0); Y = Or.getOperand(1); return true; } if (Or.getOperand(1) == Other) { X = Or.getOperand(1); Y = Or.getOperand(0); return true; } return false; }; EVT OpVT = N0.getValueType(); EVT ShAmtVT = N0.getOperand(2).getValueType(); SDValue F0 = N0.getOperand(0); SDValue F1 = N0.getOperand(1); if (matchOr(F0, F1)) { // fshl (or X, Y), X, C ==/!= 0 --> or (shl Y, C), X ==/!= 0 SDValue NewShAmt = DAG.getConstant(ShAmt, dl, ShAmtVT); SDValue Shift = DAG.getNode(ISD::SHL, dl, OpVT, Y, NewShAmt); SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, Shift, X); return DAG.getSetCC(dl, VT, NewOr, N1, Cond); } if (matchOr(F1, F0)) { // fshl X, (or X, Y), C ==/!= 0 --> or (srl Y, BW-C), X ==/!= 0 SDValue NewShAmt = DAG.getConstant(BitWidth - ShAmt, dl, ShAmtVT); SDValue Shift = DAG.getNode(ISD::SRL, dl, OpVT, Y, NewShAmt); SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, Shift, X); return DAG.getSetCC(dl, VT, NewOr, N1, Cond); } return SDValue(); } /// Try to simplify a setcc built with the specified operands and cc. If it is /// unable to simplify it, return a null SDValue. SDValue TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, bool foldBooleans, DAGCombinerInfo &DCI, const SDLoc &dl) const { SelectionDAG &DAG = DCI.DAG; const DataLayout &Layout = DAG.getDataLayout(); EVT OpVT = N0.getValueType(); AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); // Constant fold or commute setcc. if (SDValue Fold = DAG.FoldSetCC(VT, N0, N1, Cond, dl)) return Fold; bool N0ConstOrSplat = isConstOrConstSplat(N0, /*AllowUndefs*/ false, /*AllowTruncate*/ true); bool N1ConstOrSplat = isConstOrConstSplat(N1, /*AllowUndefs*/ false, /*AllowTruncate*/ true); // Canonicalize toward having the constant on the RHS. // TODO: Handle non-splat vector constants. All undef causes trouble. // FIXME: We can't yet fold constant scalable vector splats, so avoid an // infinite loop here when we encounter one. ISD::CondCode SwappedCC = ISD::getSetCCSwappedOperands(Cond); if (N0ConstOrSplat && !N1ConstOrSplat && (DCI.isBeforeLegalizeOps() || isCondCodeLegal(SwappedCC, N0.getSimpleValueType()))) return DAG.getSetCC(dl, VT, N1, N0, SwappedCC); // If we have a subtract with the same 2 non-constant operands as this setcc // -- but in reverse order -- then try to commute the operands of this setcc // to match. A matching pair of setcc (cmp) and sub may be combined into 1 // instruction on some targets. if (!N0ConstOrSplat && !N1ConstOrSplat && (DCI.isBeforeLegalizeOps() || isCondCodeLegal(SwappedCC, N0.getSimpleValueType())) && DAG.doesNodeExist(ISD::SUB, DAG.getVTList(OpVT), {N1, N0}) && !DAG.doesNodeExist(ISD::SUB, DAG.getVTList(OpVT), {N0, N1})) return DAG.getSetCC(dl, VT, N1, N0, SwappedCC); if (SDValue V = foldSetCCWithRotate(VT, N0, N1, Cond, dl, DAG)) return V; if (SDValue V = foldSetCCWithFunnelShift(VT, N0, N1, Cond, dl, DAG)) return V; if (auto *N1C = isConstOrConstSplat(N1)) { const APInt &C1 = N1C->getAPIntValue(); // Optimize some CTPOP cases. if (SDValue V = simplifySetCCWithCTPOP(*this, VT, N0, C1, Cond, dl, DAG)) return V; // For equality to 0 of a no-wrap multiply, decompose and test each op: // X * Y == 0 --> (X == 0) || (Y == 0) // X * Y != 0 --> (X != 0) && (Y != 0) // TODO: This bails out if minsize is set, but if the target doesn't have a // single instruction multiply for this type, it would likely be // smaller to decompose. if (C1.isZero() && (Cond == ISD::SETEQ || Cond == ISD::SETNE) && N0.getOpcode() == ISD::MUL && N0.hasOneUse() && (N0->getFlags().hasNoUnsignedWrap() || N0->getFlags().hasNoSignedWrap()) && !Attr.hasFnAttr(Attribute::MinSize)) { SDValue IsXZero = DAG.getSetCC(dl, VT, N0.getOperand(0), N1, Cond); SDValue IsYZero = DAG.getSetCC(dl, VT, N0.getOperand(1), N1, Cond); unsigned LogicOp = Cond == ISD::SETEQ ? ISD::OR : ISD::AND; return DAG.getNode(LogicOp, dl, VT, IsXZero, IsYZero); } // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an // equality comparison, then we're just comparing whether X itself is // zero. if (N0.getOpcode() == ISD::SRL && (C1.isZero() || C1.isOne()) && N0.getOperand(0).getOpcode() == ISD::CTLZ && llvm::has_single_bit(N0.getScalarValueSizeInBits())) { if (ConstantSDNode *ShAmt = isConstOrConstSplat(N0.getOperand(1))) { if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && ShAmt->getAPIntValue() == Log2_32(N0.getScalarValueSizeInBits())) { if ((C1 == 0) == (Cond == ISD::SETEQ)) { // (srl (ctlz x), 5) == 0 -> X != 0 // (srl (ctlz x), 5) != 1 -> X != 0 Cond = ISD::SETNE; } else { // (srl (ctlz x), 5) != 0 -> X == 0 // (srl (ctlz x), 5) == 1 -> X == 0 Cond = ISD::SETEQ; } SDValue Zero = DAG.getConstant(0, dl, N0.getValueType()); return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), Zero, Cond); } } } } // FIXME: Support vectors. if (auto *N1C = dyn_cast(N1.getNode())) { const APInt &C1 = N1C->getAPIntValue(); // (zext x) == C --> x == (trunc C) // (sext x) == C --> x == (trunc C) if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && DCI.isBeforeLegalize() && N0->hasOneUse()) { unsigned MinBits = N0.getValueSizeInBits(); SDValue PreExt; bool Signed = false; if (N0->getOpcode() == ISD::ZERO_EXTEND) { // ZExt MinBits = N0->getOperand(0).getValueSizeInBits(); PreExt = N0->getOperand(0); } else if (N0->getOpcode() == ISD::AND) { // DAGCombine turns costly ZExts into ANDs if (auto *C = dyn_cast(N0->getOperand(1))) if ((C->getAPIntValue()+1).isPowerOf2()) { MinBits = C->getAPIntValue().countr_one(); PreExt = N0->getOperand(0); } } else if (N0->getOpcode() == ISD::SIGN_EXTEND) { // SExt MinBits = N0->getOperand(0).getValueSizeInBits(); PreExt = N0->getOperand(0); Signed = true; } else if (auto *LN0 = dyn_cast(N0)) { // ZEXTLOAD / SEXTLOAD if (LN0->getExtensionType() == ISD::ZEXTLOAD) { MinBits = LN0->getMemoryVT().getSizeInBits(); PreExt = N0; } else if (LN0->getExtensionType() == ISD::SEXTLOAD) { Signed = true; MinBits = LN0->getMemoryVT().getSizeInBits(); PreExt = N0; } } // Figure out how many bits we need to preserve this constant. unsigned ReqdBits = Signed ? C1.getSignificantBits() : C1.getActiveBits(); // Make sure we're not losing bits from the constant. if (MinBits > 0 && MinBits < C1.getBitWidth() && MinBits >= ReqdBits) { EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits); if (isTypeDesirableForOp(ISD::SETCC, MinVT)) { // Will get folded away. SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreExt); if (MinBits == 1 && C1 == 1) // Invert the condition. return DAG.getSetCC(dl, VT, Trunc, DAG.getConstant(0, dl, MVT::i1), Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); SDValue C = DAG.getConstant(C1.trunc(MinBits), dl, MinVT); return DAG.getSetCC(dl, VT, Trunc, C, Cond); } // If truncating the setcc operands is not desirable, we can still // simplify the expression in some cases: // setcc ([sz]ext (setcc x, y, cc)), 0, setne) -> setcc (x, y, cc) // setcc ([sz]ext (setcc x, y, cc)), 0, seteq) -> setcc (x, y, inv(cc)) // setcc (zext (setcc x, y, cc)), 1, setne) -> setcc (x, y, inv(cc)) // setcc (zext (setcc x, y, cc)), 1, seteq) -> setcc (x, y, cc) // setcc (sext (setcc x, y, cc)), -1, setne) -> setcc (x, y, inv(cc)) // setcc (sext (setcc x, y, cc)), -1, seteq) -> setcc (x, y, cc) SDValue TopSetCC = N0->getOperand(0); unsigned N0Opc = N0->getOpcode(); bool SExt = (N0Opc == ISD::SIGN_EXTEND); if (TopSetCC.getValueType() == MVT::i1 && VT == MVT::i1 && TopSetCC.getOpcode() == ISD::SETCC && (N0Opc == ISD::ZERO_EXTEND || N0Opc == ISD::SIGN_EXTEND) && (isConstFalseVal(N1) || isExtendedTrueVal(N1C, N0->getValueType(0), SExt))) { bool Inverse = (N1C->isZero() && Cond == ISD::SETEQ) || (!N1C->isZero() && Cond == ISD::SETNE); if (!Inverse) return TopSetCC; ISD::CondCode InvCond = ISD::getSetCCInverse( cast(TopSetCC.getOperand(2))->get(), TopSetCC.getOperand(0).getValueType()); return DAG.getSetCC(dl, VT, TopSetCC.getOperand(0), TopSetCC.getOperand(1), InvCond); } } } // If the LHS is '(and load, const)', the RHS is 0, the test is for // equality or unsigned, and all 1 bits of the const are in the same // partial word, see if we can shorten the load. if (DCI.isBeforeLegalize() && !ISD::isSignedIntSetCC(Cond) && N0.getOpcode() == ISD::AND && C1 == 0 && N0.getNode()->hasOneUse() && isa(N0.getOperand(0)) && N0.getOperand(0).getNode()->hasOneUse() && isa(N0.getOperand(1))) { auto *Lod = cast(N0.getOperand(0)); APInt bestMask; unsigned bestWidth = 0, bestOffset = 0; if (Lod->isSimple() && Lod->isUnindexed() && (Lod->getMemoryVT().isByteSized() || isPaddedAtMostSignificantBitsWhenStored(Lod->getMemoryVT()))) { unsigned memWidth = Lod->getMemoryVT().getStoreSizeInBits(); unsigned origWidth = N0.getValueSizeInBits(); unsigned maskWidth = origWidth; // We can narrow (e.g.) 16-bit extending loads on 32-bit target to // 8 bits, but have to be careful... if (Lod->getExtensionType() != ISD::NON_EXTLOAD) origWidth = Lod->getMemoryVT().getSizeInBits(); const APInt &Mask = N0.getConstantOperandAPInt(1); // Only consider power-of-2 widths (and at least one byte) as candiates // for the narrowed load. for (unsigned width = 8; width < origWidth; width *= 2) { EVT newVT = EVT::getIntegerVT(*DAG.getContext(), width); if (!shouldReduceLoadWidth(Lod, ISD::NON_EXTLOAD, newVT)) continue; APInt newMask = APInt::getLowBitsSet(maskWidth, width); // Avoid accessing any padding here for now (we could use memWidth // instead of origWidth here otherwise). unsigned maxOffset = origWidth - width; for (unsigned offset = 0; offset <= maxOffset; offset += 8) { if (Mask.isSubsetOf(newMask)) { unsigned ptrOffset = Layout.isLittleEndian() ? offset : memWidth - width - offset; unsigned IsFast = 0; Align NewAlign = commonAlignment(Lod->getAlign(), ptrOffset / 8); if (allowsMemoryAccess( *DAG.getContext(), Layout, newVT, Lod->getAddressSpace(), NewAlign, Lod->getMemOperand()->getFlags(), &IsFast) && IsFast) { bestOffset = ptrOffset / 8; bestMask = Mask.lshr(offset); bestWidth = width; break; } } newMask <<= 8; } if (bestWidth) break; } } if (bestWidth) { EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth); SDValue Ptr = Lod->getBasePtr(); if (bestOffset != 0) Ptr = DAG.getObjectPtrOffset(dl, Ptr, TypeSize::getFixed(bestOffset)); SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, Lod->getPointerInfo().getWithOffset(bestOffset), Lod->getOriginalAlign()); SDValue And = DAG.getNode(ISD::AND, dl, newVT, NewLoad, DAG.getConstant(bestMask.trunc(bestWidth), dl, newVT)); return DAG.getSetCC(dl, VT, And, DAG.getConstant(0LL, dl, newVT), Cond); } } // If the LHS is a ZERO_EXTEND, perform the comparison on the input. if (N0.getOpcode() == ISD::ZERO_EXTEND) { unsigned InSize = N0.getOperand(0).getValueSizeInBits(); // If the comparison constant has bits in the upper part, the // zero-extended value could never match. if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), C1.getBitWidth() - InSize))) { switch (Cond) { case ISD::SETUGT: case ISD::SETUGE: case ISD::SETEQ: return DAG.getConstant(0, dl, VT); case ISD::SETULT: case ISD::SETULE: case ISD::SETNE: return DAG.getConstant(1, dl, VT); case ISD::SETGT: case ISD::SETGE: // True if the sign bit of C1 is set. return DAG.getConstant(C1.isNegative(), dl, VT); case ISD::SETLT: case ISD::SETLE: // True if the sign bit of C1 isn't set. return DAG.getConstant(C1.isNonNegative(), dl, VT); default: break; } } // Otherwise, we can perform the comparison with the low bits. switch (Cond) { case ISD::SETEQ: case ISD::SETNE: case ISD::SETUGT: case ISD::SETUGE: case ISD::SETULT: case ISD::SETULE: { EVT newVT = N0.getOperand(0).getValueType(); if (DCI.isBeforeLegalizeOps() || (isOperationLegal(ISD::SETCC, newVT) && isCondCodeLegal(Cond, newVT.getSimpleVT()))) { EVT NewSetCCVT = getSetCCResultType(Layout, *DAG.getContext(), newVT); SDValue NewConst = DAG.getConstant(C1.trunc(InSize), dl, newVT); SDValue NewSetCC = DAG.getSetCC(dl, NewSetCCVT, N0.getOperand(0), NewConst, Cond); return DAG.getBoolExtOrTrunc(NewSetCC, dl, VT, N0.getValueType()); } break; } default: break; // todo, be more careful with signed comparisons } } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && (Cond == ISD::SETEQ || Cond == ISD::SETNE) && !isSExtCheaperThanZExt(cast(N0.getOperand(1))->getVT(), OpVT)) { EVT ExtSrcTy = cast(N0.getOperand(1))->getVT(); unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); EVT ExtDstTy = N0.getValueType(); unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); // If the constant doesn't fit into the number of bits for the source of // the sign extension, it is impossible for both sides to be equal. if (C1.getSignificantBits() > ExtSrcTyBits) return DAG.getBoolConstant(Cond == ISD::SETNE, dl, VT, OpVT); assert(ExtDstTy == N0.getOperand(0).getValueType() && ExtDstTy != ExtSrcTy && "Unexpected types!"); APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); SDValue ZextOp = DAG.getNode(ISD::AND, dl, ExtDstTy, N0.getOperand(0), DAG.getConstant(Imm, dl, ExtDstTy)); if (!DCI.isCalledByLegalizer()) DCI.AddToWorklist(ZextOp.getNode()); // Otherwise, make this a use of a zext. return DAG.getSetCC(dl, VT, ZextOp, DAG.getConstant(C1 & Imm, dl, ExtDstTy), Cond); } else if ((N1C->isZero() || N1C->isOne()) && (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { // SETCC (X), [0|1], [EQ|NE] -> X if X is known 0/1. i1 types are // excluded as they are handled below whilst checking for foldBooleans. if ((N0.getOpcode() == ISD::SETCC || VT.getScalarType() != MVT::i1) && isTypeLegal(VT) && VT.bitsLE(N0.getValueType()) && (N0.getValueType() == MVT::i1 || getBooleanContents(N0.getValueType()) == ZeroOrOneBooleanContent) && DAG.MaskedValueIsZero( N0, APInt::getBitsSetFrom(N0.getValueSizeInBits(), 1))) { bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (!N1C->isOne()); if (TrueWhenTrue) return DAG.getNode(ISD::TRUNCATE, dl, VT, N0); // Invert the condition. if (N0.getOpcode() == ISD::SETCC) { ISD::CondCode CC = cast(N0.getOperand(2))->get(); CC = ISD::getSetCCInverse(CC, N0.getOperand(0).getValueType()); if (DCI.isBeforeLegalizeOps() || isCondCodeLegal(CC, N0.getOperand(0).getSimpleValueType())) return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); } } if ((N0.getOpcode() == ISD::XOR || (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::XOR && N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && isOneConstant(N0.getOperand(1))) { // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We // can only do this if the top bits are known zero. unsigned BitWidth = N0.getValueSizeInBits(); if (DAG.MaskedValueIsZero(N0, APInt::getHighBitsSet(BitWidth, BitWidth-1))) { // Okay, get the un-inverted input value. SDValue Val; if (N0.getOpcode() == ISD::XOR) { Val = N0.getOperand(0); } else { assert(N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::XOR); // ((X^1)&1)^1 -> X & 1 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), N0.getOperand(0).getOperand(0), N0.getOperand(1)); } return DAG.getSetCC(dl, VT, Val, N1, Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); } } else if (N1C->isOne()) { SDValue Op0 = N0; if (Op0.getOpcode() == ISD::TRUNCATE) Op0 = Op0.getOperand(0); if ((Op0.getOpcode() == ISD::XOR) && Op0.getOperand(0).getOpcode() == ISD::SETCC && Op0.getOperand(1).getOpcode() == ISD::SETCC) { SDValue XorLHS = Op0.getOperand(0); SDValue XorRHS = Op0.getOperand(1); // Ensure that the input setccs return an i1 type or 0/1 value. if (Op0.getValueType() == MVT::i1 || (getBooleanContents(XorLHS.getOperand(0).getValueType()) == ZeroOrOneBooleanContent && getBooleanContents(XorRHS.getOperand(0).getValueType()) == ZeroOrOneBooleanContent)) { // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc) Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ; return DAG.getSetCC(dl, VT, XorLHS, XorRHS, Cond); } } if (Op0.getOpcode() == ISD::AND && isOneConstant(Op0.getOperand(1))) { // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0. if (Op0.getValueType().bitsGT(VT)) Op0 = DAG.getNode(ISD::AND, dl, VT, DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)), DAG.getConstant(1, dl, VT)); else if (Op0.getValueType().bitsLT(VT)) Op0 = DAG.getNode(ISD::AND, dl, VT, DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)), DAG.getConstant(1, dl, VT)); return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, Op0.getValueType()), Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); } if (Op0.getOpcode() == ISD::AssertZext && cast(Op0.getOperand(1))->getVT() == MVT::i1) return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, Op0.getValueType()), Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); } } // Given: // icmp eq/ne (urem %x, %y), 0 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem': // icmp eq/ne %x, 0 if (N0.getOpcode() == ISD::UREM && N1C->isZero() && (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { KnownBits XKnown = DAG.computeKnownBits(N0.getOperand(0)); KnownBits YKnown = DAG.computeKnownBits(N0.getOperand(1)); if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2) return DAG.getSetCC(dl, VT, N0.getOperand(0), N1, Cond); } // Fold set_cc seteq (ashr X, BW-1), -1 -> set_cc setlt X, 0 // and set_cc setne (ashr X, BW-1), -1 -> set_cc setge X, 0 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && N0.getOpcode() == ISD::SRA && isa(N0.getOperand(1)) && N0.getConstantOperandAPInt(1) == OpVT.getScalarSizeInBits() - 1 && N1C && N1C->isAllOnes()) { return DAG.getSetCC(dl, VT, N0.getOperand(0), DAG.getConstant(0, dl, OpVT), Cond == ISD::SETEQ ? ISD::SETLT : ISD::SETGE); } if (SDValue V = optimizeSetCCOfSignedTruncationCheck(VT, N0, N1, Cond, DCI, dl)) return V; } // These simplifications apply to splat vectors as well. // TODO: Handle more splat vector cases. if (auto *N1C = isConstOrConstSplat(N1)) { const APInt &C1 = N1C->getAPIntValue(); APInt MinVal, MaxVal; unsigned OperandBitSize = N1C->getValueType(0).getScalarSizeInBits(); if (ISD::isSignedIntSetCC(Cond)) { MinVal = APInt::getSignedMinValue(OperandBitSize); MaxVal = APInt::getSignedMaxValue(OperandBitSize); } else { MinVal = APInt::getMinValue(OperandBitSize); MaxVal = APInt::getMaxValue(OperandBitSize); } // Canonicalize GE/LE comparisons to use GT/LT comparisons. if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { // X >= MIN --> true if (C1 == MinVal) return DAG.getBoolConstant(true, dl, VT, OpVT); if (!VT.isVector()) { // TODO: Support this for vectors. // X >= C0 --> X > (C0 - 1) APInt C = C1 - 1; ISD::CondCode NewCC = (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT; if ((DCI.isBeforeLegalizeOps() || isCondCodeLegal(NewCC, VT.getSimpleVT())) && (!N1C->isOpaque() || (C.getBitWidth() <= 64 && isLegalICmpImmediate(C.getSExtValue())))) { return DAG.getSetCC(dl, VT, N0, DAG.getConstant(C, dl, N1.getValueType()), NewCC); } } } if (Cond == ISD::SETLE || Cond == ISD::SETULE) { // X <= MAX --> true if (C1 == MaxVal) return DAG.getBoolConstant(true, dl, VT, OpVT); // X <= C0 --> X < (C0 + 1) if (!VT.isVector()) { // TODO: Support this for vectors. APInt C = C1 + 1; ISD::CondCode NewCC = (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT; if ((DCI.isBeforeLegalizeOps() || isCondCodeLegal(NewCC, VT.getSimpleVT())) && (!N1C->isOpaque() || (C.getBitWidth() <= 64 && isLegalICmpImmediate(C.getSExtValue())))) { return DAG.getSetCC(dl, VT, N0, DAG.getConstant(C, dl, N1.getValueType()), NewCC); } } } if (Cond == ISD::SETLT || Cond == ISD::SETULT) { if (C1 == MinVal) return DAG.getBoolConstant(false, dl, VT, OpVT); // X < MIN --> false // TODO: Support this for vectors after legalize ops. if (!VT.isVector() || DCI.isBeforeLegalizeOps()) { // Canonicalize setlt X, Max --> setne X, Max if (C1 == MaxVal) return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); // If we have setult X, 1, turn it into seteq X, 0 if (C1 == MinVal+1) return DAG.getSetCC(dl, VT, N0, DAG.getConstant(MinVal, dl, N0.getValueType()), ISD::SETEQ); } } if (Cond == ISD::SETGT || Cond == ISD::SETUGT) { if (C1 == MaxVal) return DAG.getBoolConstant(false, dl, VT, OpVT); // X > MAX --> false // TODO: Support this for vectors after legalize ops. if (!VT.isVector() || DCI.isBeforeLegalizeOps()) { // Canonicalize setgt X, Min --> setne X, Min if (C1 == MinVal) return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); // If we have setugt X, Max-1, turn it into seteq X, Max if (C1 == MaxVal-1) return DAG.getSetCC(dl, VT, N0, DAG.getConstant(MaxVal, dl, N0.getValueType()), ISD::SETEQ); } } if (Cond == ISD::SETEQ || Cond == ISD::SETNE) { // (X & (C l>>/<< Y)) ==/!= 0 --> ((X <> Y) & C) ==/!= 0 if (C1.isZero()) if (SDValue CC = optimizeSetCCByHoistingAndByConstFromLogicalShift( VT, N0, N1, Cond, DCI, dl)) return CC; // For all/any comparisons, replace or(x,shl(y,bw/2)) with and/or(x,y). // For example, when high 32-bits of i64 X are known clear: // all bits clear: (X | (Y<<32)) == 0 --> (X | Y) == 0 // all bits set: (X | (Y<<32)) == -1 --> (X & Y) == -1 bool CmpZero = N1C->isZero(); bool CmpNegOne = N1C->isAllOnes(); if ((CmpZero || CmpNegOne) && N0.hasOneUse()) { // Match or(lo,shl(hi,bw/2)) pattern. auto IsConcat = [&](SDValue V, SDValue &Lo, SDValue &Hi) { unsigned EltBits = V.getScalarValueSizeInBits(); if (V.getOpcode() != ISD::OR || (EltBits % 2) != 0) return false; SDValue LHS = V.getOperand(0); SDValue RHS = V.getOperand(1); APInt HiBits = APInt::getHighBitsSet(EltBits, EltBits / 2); // Unshifted element must have zero upperbits. if (RHS.getOpcode() == ISD::SHL && isa(RHS.getOperand(1)) && RHS.getConstantOperandAPInt(1) == (EltBits / 2) && DAG.MaskedValueIsZero(LHS, HiBits)) { Lo = LHS; Hi = RHS.getOperand(0); return true; } if (LHS.getOpcode() == ISD::SHL && isa(LHS.getOperand(1)) && LHS.getConstantOperandAPInt(1) == (EltBits / 2) && DAG.MaskedValueIsZero(RHS, HiBits)) { Lo = RHS; Hi = LHS.getOperand(0); return true; } return false; }; auto MergeConcat = [&](SDValue Lo, SDValue Hi) { unsigned EltBits = N0.getScalarValueSizeInBits(); unsigned HalfBits = EltBits / 2; APInt HiBits = APInt::getHighBitsSet(EltBits, HalfBits); SDValue LoBits = DAG.getConstant(~HiBits, dl, OpVT); SDValue HiMask = DAG.getNode(ISD::AND, dl, OpVT, Hi, LoBits); SDValue NewN0 = DAG.getNode(CmpZero ? ISD::OR : ISD::AND, dl, OpVT, Lo, HiMask); SDValue NewN1 = CmpZero ? DAG.getConstant(0, dl, OpVT) : LoBits; return DAG.getSetCC(dl, VT, NewN0, NewN1, Cond); }; SDValue Lo, Hi; if (IsConcat(N0, Lo, Hi)) return MergeConcat(Lo, Hi); if (N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR) { SDValue Lo0, Lo1, Hi0, Hi1; if (IsConcat(N0.getOperand(0), Lo0, Hi0) && IsConcat(N0.getOperand(1), Lo1, Hi1)) { return MergeConcat(DAG.getNode(N0.getOpcode(), dl, OpVT, Lo0, Lo1), DAG.getNode(N0.getOpcode(), dl, OpVT, Hi0, Hi1)); } } } } // If we have "setcc X, C0", check to see if we can shrink the immediate // by changing cc. // TODO: Support this for vectors after legalize ops. if (!VT.isVector() || DCI.isBeforeLegalizeOps()) { // SETUGT X, SINTMAX -> SETLT X, 0 // SETUGE X, SINTMIN -> SETLT X, 0 if ((Cond == ISD::SETUGT && C1.isMaxSignedValue()) || (Cond == ISD::SETUGE && C1.isMinSignedValue())) return DAG.getSetCC(dl, VT, N0, DAG.getConstant(0, dl, N1.getValueType()), ISD::SETLT); // SETULT X, SINTMIN -> SETGT X, -1 // SETULE X, SINTMAX -> SETGT X, -1 if ((Cond == ISD::SETULT && C1.isMinSignedValue()) || (Cond == ISD::SETULE && C1.isMaxSignedValue())) return DAG.getSetCC(dl, VT, N0, DAG.getAllOnesConstant(dl, N1.getValueType()), ISD::SETGT); } } // Back to non-vector simplifications. // TODO: Can we do these for vector splats? if (auto *N1C = dyn_cast(N1.getNode())) { const TargetLowering &TLI = DAG.getTargetLoweringInfo(); const APInt &C1 = N1C->getAPIntValue(); EVT ShValTy = N0.getValueType(); // Fold bit comparisons when we can. This will result in an // incorrect value when boolean false is negative one, unless // the bitsize is 1 in which case the false value is the same // in practice regardless of the representation. if ((VT.getSizeInBits() == 1 || getBooleanContents(N0.getValueType()) == ZeroOrOneBooleanContent) && (Cond == ISD::SETEQ || Cond == ISD::SETNE) && (VT == ShValTy || (isTypeLegal(VT) && VT.bitsLE(ShValTy))) && N0.getOpcode() == ISD::AND) { if (auto *AndRHS = dyn_cast(N0.getOperand(1))) { if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 // Perform the xform if the AND RHS is a single bit. unsigned ShCt = AndRHS->getAPIntValue().logBase2(); if (AndRHS->getAPIntValue().isPowerOf2() && !TLI.shouldAvoidTransformToShift(ShValTy, ShCt)) { return DAG.getNode( ISD::TRUNCATE, dl, VT, DAG.getNode(ISD::SRL, dl, ShValTy, N0, DAG.getShiftAmountConstant(ShCt, ShValTy, dl))); } } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) { // (X & 8) == 8 --> (X & 8) >> 3 // Perform the xform if C1 is a single bit. unsigned ShCt = C1.logBase2(); if (C1.isPowerOf2() && !TLI.shouldAvoidTransformToShift(ShValTy, ShCt)) { return DAG.getNode( ISD::TRUNCATE, dl, VT, DAG.getNode(ISD::SRL, dl, ShValTy, N0, DAG.getShiftAmountConstant(ShCt, ShValTy, dl))); } } } } if (C1.getSignificantBits() <= 64 && !isLegalICmpImmediate(C1.getSExtValue())) { // (X & -256) == 256 -> (X >> 8) == 1 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && N0.getOpcode() == ISD::AND && N0.hasOneUse()) { if (auto *AndRHS = dyn_cast(N0.getOperand(1))) { const APInt &AndRHSC = AndRHS->getAPIntValue(); if (AndRHSC.isNegatedPowerOf2() && (AndRHSC & C1) == C1) { unsigned ShiftBits = AndRHSC.countr_zero(); if (!TLI.shouldAvoidTransformToShift(ShValTy, ShiftBits)) { SDValue Shift = DAG.getNode( ISD::SRL, dl, ShValTy, N0.getOperand(0), DAG.getShiftAmountConstant(ShiftBits, ShValTy, dl)); SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), dl, ShValTy); return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond); } } } } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE || Cond == ISD::SETULE || Cond == ISD::SETUGT) { bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT); // X < 0x100000000 -> (X >> 32) < 1 // X >= 0x100000000 -> (X >> 32) >= 1 // X <= 0x0ffffffff -> (X >> 32) < 1 // X > 0x0ffffffff -> (X >> 32) >= 1 unsigned ShiftBits; APInt NewC = C1; ISD::CondCode NewCond = Cond; if (AdjOne) { ShiftBits = C1.countr_one(); NewC = NewC + 1; NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; } else { ShiftBits = C1.countr_zero(); } NewC.lshrInPlace(ShiftBits); if (ShiftBits && NewC.getSignificantBits() <= 64 && isLegalICmpImmediate(NewC.getSExtValue()) && !TLI.shouldAvoidTransformToShift(ShValTy, ShiftBits)) { SDValue Shift = DAG.getNode(ISD::SRL, dl, ShValTy, N0, DAG.getShiftAmountConstant(ShiftBits, ShValTy, dl)); SDValue CmpRHS = DAG.getConstant(NewC, dl, ShValTy); return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond); } } } } if (!isa(N0) && isa(N1)) { auto *CFP = cast(N1); assert(!CFP->getValueAPF().isNaN() && "Unexpected NaN value"); // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the // constant if knowing that the operand is non-nan is enough. We prefer to // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to // materialize 0.0. if (Cond == ISD::SETO || Cond == ISD::SETUO) return DAG.getSetCC(dl, VT, N0, N0, Cond); // setcc (fneg x), C -> setcc swap(pred) x, -C if (N0.getOpcode() == ISD::FNEG) { ISD::CondCode SwapCond = ISD::getSetCCSwappedOperands(Cond); if (DCI.isBeforeLegalizeOps() || isCondCodeLegal(SwapCond, N0.getSimpleValueType())) { SDValue NegN1 = DAG.getNode(ISD::FNEG, dl, N0.getValueType(), N1); return DAG.getSetCC(dl, VT, N0.getOperand(0), NegN1, SwapCond); } } // setueq/setoeq X, (fabs Inf) -> is_fpclass X, fcInf if (isOperationLegalOrCustom(ISD::IS_FPCLASS, N0.getValueType()) && !isFPImmLegal(CFP->getValueAPF(), CFP->getValueType(0))) { bool IsFabs = N0.getOpcode() == ISD::FABS; SDValue Op = IsFabs ? N0.getOperand(0) : N0; if ((Cond == ISD::SETOEQ || Cond == ISD::SETUEQ) && CFP->isInfinity()) { FPClassTest Flag = CFP->isNegative() ? (IsFabs ? fcNone : fcNegInf) : (IsFabs ? fcInf : fcPosInf); if (Cond == ISD::SETUEQ) Flag |= fcNan; return DAG.getNode(ISD::IS_FPCLASS, dl, VT, Op, DAG.getTargetConstant(Flag, dl, MVT::i32)); } } // If the condition is not legal, see if we can find an equivalent one // which is legal. if (!isCondCodeLegal(Cond, N0.getSimpleValueType())) { // If the comparison was an awkward floating-point == or != and one of // the comparison operands is infinity or negative infinity, convert the // condition to a less-awkward <= or >=. if (CFP->getValueAPF().isInfinity()) { bool IsNegInf = CFP->getValueAPF().isNegative(); ISD::CondCode NewCond = ISD::SETCC_INVALID; switch (Cond) { case ISD::SETOEQ: NewCond = IsNegInf ? ISD::SETOLE : ISD::SETOGE; break; case ISD::SETUEQ: NewCond = IsNegInf ? ISD::SETULE : ISD::SETUGE; break; case ISD::SETUNE: NewCond = IsNegInf ? ISD::SETUGT : ISD::SETULT; break; case ISD::SETONE: NewCond = IsNegInf ? ISD::SETOGT : ISD::SETOLT; break; default: break; } if (NewCond != ISD::SETCC_INVALID && isCondCodeLegal(NewCond, N0.getSimpleValueType())) return DAG.getSetCC(dl, VT, N0, N1, NewCond); } } } if (N0 == N1) { // The sext(setcc()) => setcc() optimization relies on the appropriate // constant being emitted. assert(!N0.getValueType().isInteger() && "Integer types should be handled by FoldSetCC"); bool EqTrue = ISD::isTrueWhenEqual(Cond); unsigned UOF = ISD::getUnorderedFlavor(Cond); if (UOF == 2) // FP operators that are undefined on NaNs. return DAG.getBoolConstant(EqTrue, dl, VT, OpVT); if (UOF == unsigned(EqTrue)) return DAG.getBoolConstant(EqTrue, dl, VT, OpVT); // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO // if it is not already. ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; if (NewCond != Cond && (DCI.isBeforeLegalizeOps() || isCondCodeLegal(NewCond, N0.getSimpleValueType()))) return DAG.getSetCC(dl, VT, N0, N1, NewCond); } // ~X > ~Y --> Y > X // ~X < ~Y --> Y < X // ~X < C --> X > ~C // ~X > C --> X < ~C if ((isSignedIntSetCC(Cond) || isUnsignedIntSetCC(Cond)) && N0.getValueType().isInteger()) { if (isBitwiseNot(N0)) { if (isBitwiseNot(N1)) return DAG.getSetCC(dl, VT, N1.getOperand(0), N0.getOperand(0), Cond); if (DAG.isConstantIntBuildVectorOrConstantInt(N1) && !DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(0))) { SDValue Not = DAG.getNOT(dl, N1, OpVT); return DAG.getSetCC(dl, VT, Not, N0.getOperand(0), Cond); } } } if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && N0.getValueType().isInteger()) { if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || N0.getOpcode() == ISD::XOR) { // Simplify (X+Y) == (X+Z) --> Y == Z if (N0.getOpcode() == N1.getOpcode()) { if (N0.getOperand(0) == N1.getOperand(0)) return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); if (N0.getOperand(1) == N1.getOperand(1)) return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); if (isCommutativeBinOp(N0.getOpcode())) { // If X op Y == Y op X, try other combinations. if (N0.getOperand(0) == N1.getOperand(1)) return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), Cond); if (N0.getOperand(1) == N1.getOperand(0)) return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), Cond); } } // If RHS is a legal immediate value for a compare instruction, we need // to be careful about increasing register pressure needlessly. bool LegalRHSImm = false; if (auto *RHSC = dyn_cast(N1)) { if (auto *LHSR = dyn_cast(N0.getOperand(1))) { // Turn (X+C1) == C2 --> X == C2-C1 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) return DAG.getSetCC( dl, VT, N0.getOperand(0), DAG.getConstant(RHSC->getAPIntValue() - LHSR->getAPIntValue(), dl, N0.getValueType()), Cond); // Turn (X^C1) == C2 --> X == C1^C2 if (N0.getOpcode() == ISD::XOR && N0.getNode()->hasOneUse()) return DAG.getSetCC( dl, VT, N0.getOperand(0), DAG.getConstant(LHSR->getAPIntValue() ^ RHSC->getAPIntValue(), dl, N0.getValueType()), Cond); } // Turn (C1-X) == C2 --> X == C1-C2 if (auto *SUBC = dyn_cast(N0.getOperand(0))) if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) return DAG.getSetCC( dl, VT, N0.getOperand(1), DAG.getConstant(SUBC->getAPIntValue() - RHSC->getAPIntValue(), dl, N0.getValueType()), Cond); // Could RHSC fold directly into a compare? if (RHSC->getValueType(0).getSizeInBits() <= 64) LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue()); } // (X+Y) == X --> Y == 0 and similar folds. // Don't do this if X is an immediate that can fold into a cmp // instruction and X+Y has other uses. It could be an induction variable // chain, and the transform would increase register pressure. if (!LegalRHSImm || N0.hasOneUse()) if (SDValue V = foldSetCCWithBinOp(VT, N0, N1, Cond, dl, DCI)) return V; } if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::XOR) if (SDValue V = foldSetCCWithBinOp(VT, N1, N0, Cond, dl, DCI)) return V; if (SDValue V = foldSetCCWithAnd(VT, N0, N1, Cond, dl, DCI)) return V; } // Fold remainder of division by a constant. if ((N0.getOpcode() == ISD::UREM || N0.getOpcode() == ISD::SREM) && N0.hasOneUse() && (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { // When division is cheap or optimizing for minimum size, // fall through to DIVREM creation by skipping this fold. if (!isIntDivCheap(VT, Attr) && !Attr.hasFnAttr(Attribute::MinSize)) { if (N0.getOpcode() == ISD::UREM) { if (SDValue Folded = buildUREMEqFold(VT, N0, N1, Cond, DCI, dl)) return Folded; } else if (N0.getOpcode() == ISD::SREM) { if (SDValue Folded = buildSREMEqFold(VT, N0, N1, Cond, DCI, dl)) return Folded; } } } // Fold away ALL boolean setcc's. if (N0.getValueType().getScalarType() == MVT::i1 && foldBooleans) { SDValue Temp; switch (Cond) { default: llvm_unreachable("Unknown integer setcc!"); case ISD::SETEQ: // X == Y -> ~(X^Y) Temp = DAG.getNode(ISD::XOR, dl, OpVT, N0, N1); N0 = DAG.getNOT(dl, Temp, OpVT); if (!DCI.isCalledByLegalizer()) DCI.AddToWorklist(Temp.getNode()); break; case ISD::SETNE: // X != Y --> (X^Y) N0 = DAG.getNode(ISD::XOR, dl, OpVT, N0, N1); break; case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y case ISD::SETULT: // X X == 0 & Y == 1 --> ~X & Y Temp = DAG.getNOT(dl, N0, OpVT); N0 = DAG.getNode(ISD::AND, dl, OpVT, N1, Temp); if (!DCI.isCalledByLegalizer()) DCI.AddToWorklist(Temp.getNode()); break; case ISD::SETLT: // X X == 1 & Y == 0 --> ~Y & X case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X Temp = DAG.getNOT(dl, N1, OpVT); N0 = DAG.getNode(ISD::AND, dl, OpVT, N0, Temp); if (!DCI.isCalledByLegalizer()) DCI.AddToWorklist(Temp.getNode()); break; case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y Temp = DAG.getNOT(dl, N0, OpVT); N0 = DAG.getNode(ISD::OR, dl, OpVT, N1, Temp); if (!DCI.isCalledByLegalizer()) DCI.AddToWorklist(Temp.getNode()); break; case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X Temp = DAG.getNOT(dl, N1, OpVT); N0 = DAG.getNode(ISD::OR, dl, OpVT, N0, Temp); break; } if (VT.getScalarType() != MVT::i1) { if (!DCI.isCalledByLegalizer()) DCI.AddToWorklist(N0.getNode()); // FIXME: If running after legalize, we probably can't do this. ISD::NodeType ExtendCode = getExtendForContent(getBooleanContents(OpVT)); N0 = DAG.getNode(ExtendCode, dl, VT, N0); } return N0; } // Could not fold it. return SDValue(); } /// Returns true (and the GlobalValue and the offset) if the node is a /// GlobalAddress + offset. bool TargetLowering::isGAPlusOffset(SDNode *WN, const GlobalValue *&GA, int64_t &Offset) const { SDNode *N = unwrapAddress(SDValue(WN, 0)).getNode(); if (auto *GASD = dyn_cast(N)) { GA = GASD->getGlobal(); Offset += GASD->getOffset(); return true; } if (N->getOpcode() == ISD::ADD) { SDValue N1 = N->getOperand(0); SDValue N2 = N->getOperand(1); if (isGAPlusOffset(N1.getNode(), GA, Offset)) { if (auto *V = dyn_cast(N2)) { Offset += V->getSExtValue(); return true; } } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { if (auto *V = dyn_cast(N1)) { Offset += V->getSExtValue(); return true; } } } return false; } SDValue TargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { // Default implementation: no optimization. return SDValue(); } //===----------------------------------------------------------------------===// // Inline Assembler Implementation Methods //===----------------------------------------------------------------------===// TargetLowering::ConstraintType TargetLowering::getConstraintType(StringRef Constraint) const { unsigned S = Constraint.size(); if (S == 1) { switch (Constraint[0]) { default: break; case 'r': return C_RegisterClass; case 'm': // memory case 'o': // offsetable case 'V': // not offsetable return C_Memory; case 'p': // Address. return C_Address; case 'n': // Simple Integer case 'E': // Floating Point Constant case 'F': // Floating Point Constant return C_Immediate; case 'i': // Simple Integer or Relocatable Constant case 's': // Relocatable Constant case 'X': // Allow ANY value. case 'I': // Target registers. case 'J': case 'K': case 'L': case 'M': case 'N': case 'O': case 'P': case '<': case '>': return C_Other; } } if (S > 1 && Constraint[0] == '{' && Constraint[S - 1] == '}') { if (S == 8 && Constraint.substr(1, 6) == "memory") // "{memory}" return C_Memory; return C_Register; } return C_Unknown; } /// Try to replace an X constraint, which matches anything, with another that /// has more specific requirements based on the type of the corresponding /// operand. const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const { if (ConstraintVT.isInteger()) return "r"; if (ConstraintVT.isFloatingPoint()) return "f"; // works for many targets return nullptr; } SDValue TargetLowering::LowerAsmOutputForConstraint( SDValue &Chain, SDValue &Glue, const SDLoc &DL, const AsmOperandInfo &OpInfo, SelectionDAG &DAG) const { return SDValue(); } /// Lower the specified operand into the Ops vector. /// If it is invalid, don't add anything to Ops. void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, StringRef Constraint, std::vector &Ops, SelectionDAG &DAG) const { if (Constraint.size() > 1) return; char ConstraintLetter = Constraint[0]; switch (ConstraintLetter) { default: break; case 'X': // Allows any operand case 'i': // Simple Integer or Relocatable Constant case 'n': // Simple Integer case 's': { // Relocatable Constant ConstantSDNode *C; uint64_t Offset = 0; // Match (GA) or (C) or (GA+C) or (GA-C) or ((GA+C)+C) or (((GA+C)+C)+C), // etc., since getelementpointer is variadic. We can't use // SelectionDAG::FoldSymbolOffset because it expects the GA to be accessible // while in this case the GA may be furthest from the root node which is // likely an ISD::ADD. while (true) { if ((C = dyn_cast(Op)) && ConstraintLetter != 's') { // gcc prints these as sign extended. Sign extend value to 64 bits // now; without this it would get ZExt'd later in // ScheduleDAGSDNodes::EmitNode, which is very generic. bool IsBool = C->getConstantIntValue()->getBitWidth() == 1; BooleanContent BCont = getBooleanContents(MVT::i64); ISD::NodeType ExtOpc = IsBool ? getExtendForContent(BCont) : ISD::SIGN_EXTEND; int64_t ExtVal = ExtOpc == ISD::ZERO_EXTEND ? C->getZExtValue() : C->getSExtValue(); Ops.push_back( DAG.getTargetConstant(Offset + ExtVal, SDLoc(C), MVT::i64)); return; } if (ConstraintLetter != 'n') { if (const auto *GA = dyn_cast(Op)) { Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), SDLoc(Op), GA->getValueType(0), Offset + GA->getOffset())); return; } if (const auto *BA = dyn_cast(Op)) { Ops.push_back(DAG.getTargetBlockAddress( BA->getBlockAddress(), BA->getValueType(0), Offset + BA->getOffset(), BA->getTargetFlags())); return; } if (isa(Op)) { Ops.push_back(Op); return; } } const unsigned OpCode = Op.getOpcode(); if (OpCode == ISD::ADD || OpCode == ISD::SUB) { if ((C = dyn_cast(Op.getOperand(0)))) Op = Op.getOperand(1); // Subtraction is not commutative. else if (OpCode == ISD::ADD && (C = dyn_cast(Op.getOperand(1)))) Op = Op.getOperand(0); else return; Offset += (OpCode == ISD::ADD ? 1 : -1) * C->getSExtValue(); continue; } return; } break; } } } void TargetLowering::CollectTargetIntrinsicOperands( const CallInst &I, SmallVectorImpl &Ops, SelectionDAG &DAG) const { } std::pair TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *RI, StringRef Constraint, MVT VT) const { if (!Constraint.starts_with("{")) return std::make_pair(0u, static_cast(nullptr)); assert(*(Constraint.end() - 1) == '}' && "Not a brace enclosed constraint?"); // Remove the braces from around the name. StringRef RegName(Constraint.data() + 1, Constraint.size() - 2); std::pair R = std::make_pair(0u, static_cast(nullptr)); // Figure out which register class contains this reg. for (const TargetRegisterClass *RC : RI->regclasses()) { // If none of the value types for this register class are valid, we // can't use it. For example, 64-bit reg classes on 32-bit targets. if (!isLegalRC(*RI, *RC)) continue; for (const MCPhysReg &PR : *RC) { if (RegName.equals_insensitive(RI->getRegAsmName(PR))) { std::pair S = std::make_pair(PR, RC); // If this register class has the requested value type, return it, // otherwise keep searching and return the first class found // if no other is found which explicitly has the requested type. if (RI->isTypeLegalForClass(*RC, VT)) return S; if (!R.second) R = S; } } } return R; } //===----------------------------------------------------------------------===// // Constraint Selection. /// Return true of this is an input operand that is a matching constraint like /// "4". bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { assert(!ConstraintCode.empty() && "No known constraint!"); return isdigit(static_cast(ConstraintCode[0])); } /// If this is an input matching constraint, this method returns the output /// operand it matches. unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { assert(!ConstraintCode.empty() && "No known constraint!"); return atoi(ConstraintCode.c_str()); } /// Split up the constraint string from the inline assembly value into the /// specific constraints and their prefixes, and also tie in the associated /// operand values. /// If this returns an empty vector, and if the constraint string itself /// isn't empty, there was an error parsing. TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints(const DataLayout &DL, const TargetRegisterInfo *TRI, const CallBase &Call) const { /// Information about all of the constraints. AsmOperandInfoVector ConstraintOperands; const InlineAsm *IA = cast(Call.getCalledOperand()); unsigned maCount = 0; // Largest number of multiple alternative constraints. // Do a prepass over the constraints, canonicalizing them, and building up the // ConstraintOperands list. unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. unsigned ResNo = 0; // ResNo - The result number of the next output. unsigned LabelNo = 0; // LabelNo - CallBr indirect dest number. for (InlineAsm::ConstraintInfo &CI : IA->ParseConstraints()) { ConstraintOperands.emplace_back(std::move(CI)); AsmOperandInfo &OpInfo = ConstraintOperands.back(); // Update multiple alternative constraint count. if (OpInfo.multipleAlternatives.size() > maCount) maCount = OpInfo.multipleAlternatives.size(); OpInfo.ConstraintVT = MVT::Other; // Compute the value type for each operand. switch (OpInfo.Type) { case InlineAsm::isOutput: // Indirect outputs just consume an argument. if (OpInfo.isIndirect) { OpInfo.CallOperandVal = Call.getArgOperand(ArgNo); break; } // The return value of the call is this value. As such, there is no // corresponding argument. assert(!Call.getType()->isVoidTy() && "Bad inline asm!"); if (auto *STy = dyn_cast(Call.getType())) { OpInfo.ConstraintVT = getSimpleValueType(DL, STy->getElementType(ResNo)); } else { assert(ResNo == 0 && "Asm only has one result!"); OpInfo.ConstraintVT = getAsmOperandValueType(DL, Call.getType()).getSimpleVT(); } ++ResNo; break; case InlineAsm::isInput: OpInfo.CallOperandVal = Call.getArgOperand(ArgNo); break; case InlineAsm::isLabel: OpInfo.CallOperandVal = cast(&Call)->getIndirectDest(LabelNo); ++LabelNo; continue; case InlineAsm::isClobber: // Nothing to do. break; } if (OpInfo.CallOperandVal) { llvm::Type *OpTy = OpInfo.CallOperandVal->getType(); if (OpInfo.isIndirect) { OpTy = Call.getParamElementType(ArgNo); assert(OpTy && "Indirect operand must have elementtype attribute"); } // Look for vector wrapped in a struct. e.g. { <16 x i8> }. if (StructType *STy = dyn_cast(OpTy)) if (STy->getNumElements() == 1) OpTy = STy->getElementType(0); // If OpTy is not a single value, it may be a struct/union that we // can tile with integers. if (!OpTy->isSingleValueType() && OpTy->isSized()) { unsigned BitSize = DL.getTypeSizeInBits(OpTy); switch (BitSize) { default: break; case 1: case 8: case 16: case 32: case 64: case 128: OpTy = IntegerType::get(OpTy->getContext(), BitSize); break; } } EVT VT = getAsmOperandValueType(DL, OpTy, true); OpInfo.ConstraintVT = VT.isSimple() ? VT.getSimpleVT() : MVT::Other; ArgNo++; } } // If we have multiple alternative constraints, select the best alternative. if (!ConstraintOperands.empty()) { if (maCount) { unsigned bestMAIndex = 0; int bestWeight = -1; // weight: -1 = invalid match, and 0 = so-so match to 5 = good match. int weight = -1; unsigned maIndex; // Compute the sums of the weights for each alternative, keeping track // of the best (highest weight) one so far. for (maIndex = 0; maIndex < maCount; ++maIndex) { int weightSum = 0; for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); cIndex != eIndex; ++cIndex) { AsmOperandInfo &OpInfo = ConstraintOperands[cIndex]; if (OpInfo.Type == InlineAsm::isClobber) continue; // If this is an output operand with a matching input operand, // look up the matching input. If their types mismatch, e.g. one // is an integer, the other is floating point, or their sizes are // different, flag it as an maCantMatch. if (OpInfo.hasMatchingInput()) { AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; if (OpInfo.ConstraintVT != Input.ConstraintVT) { if ((OpInfo.ConstraintVT.isInteger() != Input.ConstraintVT.isInteger()) || (OpInfo.ConstraintVT.getSizeInBits() != Input.ConstraintVT.getSizeInBits())) { weightSum = -1; // Can't match. break; } } } weight = getMultipleConstraintMatchWeight(OpInfo, maIndex); if (weight == -1) { weightSum = -1; break; } weightSum += weight; } // Update best. if (weightSum > bestWeight) { bestWeight = weightSum; bestMAIndex = maIndex; } } // Now select chosen alternative in each constraint. for (AsmOperandInfo &cInfo : ConstraintOperands) if (cInfo.Type != InlineAsm::isClobber) cInfo.selectAlternative(bestMAIndex); } } // Check and hook up tied operands, choose constraint code to use. for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); cIndex != eIndex; ++cIndex) { AsmOperandInfo &OpInfo = ConstraintOperands[cIndex]; // If this is an output operand with a matching input operand, look up the // matching input. If their types mismatch, e.g. one is an integer, the // other is floating point, or their sizes are different, flag it as an // error. if (OpInfo.hasMatchingInput()) { AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; if (OpInfo.ConstraintVT != Input.ConstraintVT) { std::pair MatchRC = getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode, OpInfo.ConstraintVT); std::pair InputRC = getRegForInlineAsmConstraint(TRI, Input.ConstraintCode, Input.ConstraintVT); if ((OpInfo.ConstraintVT.isInteger() != Input.ConstraintVT.isInteger()) || (MatchRC.second != InputRC.second)) { report_fatal_error("Unsupported asm: input constraint" " with a matching output constraint of" " incompatible type!"); } } } } return ConstraintOperands; } /// Return a number indicating our preference for chosing a type of constraint /// over another, for the purpose of sorting them. Immediates are almost always /// preferrable (when they can be emitted). A higher return value means a /// stronger preference for one constraint type relative to another. /// FIXME: We should prefer registers over memory but doing so may lead to /// unrecoverable register exhaustion later. /// https://github.com/llvm/llvm-project/issues/20571 static unsigned getConstraintPiority(TargetLowering::ConstraintType CT) { switch (CT) { case TargetLowering::C_Immediate: case TargetLowering::C_Other: return 4; case TargetLowering::C_Memory: case TargetLowering::C_Address: return 3; case TargetLowering::C_RegisterClass: return 2; case TargetLowering::C_Register: return 1; case TargetLowering::C_Unknown: return 0; } llvm_unreachable("Invalid constraint type"); } /// Examine constraint type and operand type and determine a weight value. /// This object must already have been set up with the operand type /// and the current alternative constraint selected. TargetLowering::ConstraintWeight TargetLowering::getMultipleConstraintMatchWeight( AsmOperandInfo &info, int maIndex) const { InlineAsm::ConstraintCodeVector *rCodes; if (maIndex >= (int)info.multipleAlternatives.size()) rCodes = &info.Codes; else rCodes = &info.multipleAlternatives[maIndex].Codes; ConstraintWeight BestWeight = CW_Invalid; // Loop over the options, keeping track of the most general one. for (const std::string &rCode : *rCodes) { ConstraintWeight weight = getSingleConstraintMatchWeight(info, rCode.c_str()); if (weight > BestWeight) BestWeight = weight; } return BestWeight; } /// Examine constraint type and operand type and determine a weight value. /// This object must already have been set up with the operand type /// and the current alternative constraint selected. TargetLowering::ConstraintWeight TargetLowering::getSingleConstraintMatchWeight( AsmOperandInfo &info, const char *constraint) const { ConstraintWeight weight = CW_Invalid; Value *CallOperandVal = info.CallOperandVal; // If we don't have a value, we can't do a match, // but allow it at the lowest weight. if (!CallOperandVal) return CW_Default; // Look at the constraint type. switch (*constraint) { case 'i': // immediate integer. case 'n': // immediate integer with a known value. if (isa(CallOperandVal)) weight = CW_Constant; break; case 's': // non-explicit intregal immediate. if (isa(CallOperandVal)) weight = CW_Constant; break; case 'E': // immediate float if host format. case 'F': // immediate float. if (isa(CallOperandVal)) weight = CW_Constant; break; case '<': // memory operand with autodecrement. case '>': // memory operand with autoincrement. case 'm': // memory operand. case 'o': // offsettable memory operand case 'V': // non-offsettable memory operand weight = CW_Memory; break; case 'r': // general register. case 'g': // general register, memory operand or immediate integer. // note: Clang converts "g" to "imr". if (CallOperandVal->getType()->isIntegerTy()) weight = CW_Register; break; case 'X': // any operand. default: weight = CW_Default; break; } return weight; } /// If there are multiple different constraints that we could pick for this /// operand (e.g. "imr") try to pick the 'best' one. /// This is somewhat tricky: constraints (TargetLowering::ConstraintType) fall /// into seven classes: /// Register -> one specific register /// RegisterClass -> a group of regs /// Memory -> memory /// Address -> a symbolic memory reference /// Immediate -> immediate values /// Other -> magic values (such as "Flag Output Operands") /// Unknown -> something we don't recognize yet and can't handle /// Ideally, we would pick the most specific constraint possible: if we have /// something that fits into a register, we would pick it. The problem here /// is that if we have something that could either be in a register or in /// memory that use of the register could cause selection of *other* /// operands to fail: they might only succeed if we pick memory. Because of /// this the heuristic we use is: /// /// 1) If there is an 'other' constraint, and if the operand is valid for /// that constraint, use it. This makes us take advantage of 'i' /// constraints when available. /// 2) Otherwise, pick the most general constraint present. This prefers /// 'm' over 'r', for example. /// TargetLowering::ConstraintGroup TargetLowering::getConstraintPreferences( TargetLowering::AsmOperandInfo &OpInfo) const { ConstraintGroup Ret; Ret.reserve(OpInfo.Codes.size()); for (StringRef Code : OpInfo.Codes) { TargetLowering::ConstraintType CType = getConstraintType(Code); // Indirect 'other' or 'immediate' constraints are not allowed. if (OpInfo.isIndirect && !(CType == TargetLowering::C_Memory || CType == TargetLowering::C_Register || CType == TargetLowering::C_RegisterClass)) continue; // Things with matching constraints can only be registers, per gcc // documentation. This mainly affects "g" constraints. if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput()) continue; Ret.emplace_back(Code, CType); } std::stable_sort( Ret.begin(), Ret.end(), [](ConstraintPair a, ConstraintPair b) { return getConstraintPiority(a.second) > getConstraintPiority(b.second); }); return Ret; } /// If we have an immediate, see if we can lower it. Return true if we can, /// false otherwise. static bool lowerImmediateIfPossible(TargetLowering::ConstraintPair &P, SDValue Op, SelectionDAG *DAG, const TargetLowering &TLI) { assert((P.second == TargetLowering::C_Other || P.second == TargetLowering::C_Immediate) && "need immediate or other"); if (!Op.getNode()) return false; std::vector ResultOps; TLI.LowerAsmOperandForConstraint(Op, P.first, ResultOps, *DAG); return !ResultOps.empty(); } /// Determines the constraint code and constraint type to use for the specific /// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType. void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, SDValue Op, SelectionDAG *DAG) const { assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); // Single-letter constraints ('r') are very common. if (OpInfo.Codes.size() == 1) { OpInfo.ConstraintCode = OpInfo.Codes[0]; OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); } else { ConstraintGroup G = getConstraintPreferences(OpInfo); if (G.empty()) return; unsigned BestIdx = 0; for (const unsigned E = G.size(); BestIdx < E && (G[BestIdx].second == TargetLowering::C_Other || G[BestIdx].second == TargetLowering::C_Immediate); ++BestIdx) { if (lowerImmediateIfPossible(G[BestIdx], Op, DAG, *this)) break; // If we're out of constraints, just pick the first one. if (BestIdx + 1 == E) { BestIdx = 0; break; } } OpInfo.ConstraintCode = G[BestIdx].first; OpInfo.ConstraintType = G[BestIdx].second; } // 'X' matches anything. if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { // Constants are handled elsewhere. For Functions, the type here is the // type of the result, which is not what we want to look at; leave them // alone. Value *v = OpInfo.CallOperandVal; if (isa(v) || isa(v)) { return; } if (isa(v) || isa(v)) { OpInfo.ConstraintCode = "i"; return; } // Otherwise, try to resolve it to something we know about by looking at // the actual operand type. if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { OpInfo.ConstraintCode = Repl; OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); } } } /// Given an exact SDIV by a constant, create a multiplication /// with the multiplicative inverse of the constant. /// Ref: "Hacker's Delight" by Henry Warren, 2nd Edition, p. 242 static SDValue BuildExactSDIV(const TargetLowering &TLI, SDNode *N, const SDLoc &dl, SelectionDAG &DAG, SmallVectorImpl &Created) { SDValue Op0 = N->getOperand(0); SDValue Op1 = N->getOperand(1); EVT VT = N->getValueType(0); EVT SVT = VT.getScalarType(); EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); EVT ShSVT = ShVT.getScalarType(); bool UseSRA = false; SmallVector Shifts, Factors; auto BuildSDIVPattern = [&](ConstantSDNode *C) { if (C->isZero()) return false; APInt Divisor = C->getAPIntValue(); unsigned Shift = Divisor.countr_zero(); if (Shift) { Divisor.ashrInPlace(Shift); UseSRA = true; } APInt Factor = Divisor.multiplicativeInverse(); Shifts.push_back(DAG.getConstant(Shift, dl, ShSVT)); Factors.push_back(DAG.getConstant(Factor, dl, SVT)); return true; }; // Collect all magic values from the build vector. if (!ISD::matchUnaryPredicate(Op1, BuildSDIVPattern)) return SDValue(); SDValue Shift, Factor; if (Op1.getOpcode() == ISD::BUILD_VECTOR) { Shift = DAG.getBuildVector(ShVT, dl, Shifts); Factor = DAG.getBuildVector(VT, dl, Factors); } else if (Op1.getOpcode() == ISD::SPLAT_VECTOR) { assert(Shifts.size() == 1 && Factors.size() == 1 && "Expected matchUnaryPredicate to return one element for scalable " "vectors"); Shift = DAG.getSplatVector(ShVT, dl, Shifts[0]); Factor = DAG.getSplatVector(VT, dl, Factors[0]); } else { assert(isa(Op1) && "Expected a constant"); Shift = Shifts[0]; Factor = Factors[0]; } SDValue Res = Op0; if (UseSRA) { SDNodeFlags Flags; Flags.setExact(true); Res = DAG.getNode(ISD::SRA, dl, VT, Res, Shift, Flags); Created.push_back(Res.getNode()); } return DAG.getNode(ISD::MUL, dl, VT, Res, Factor); } /// Given an exact UDIV by a constant, create a multiplication /// with the multiplicative inverse of the constant. /// Ref: "Hacker's Delight" by Henry Warren, 2nd Edition, p. 242 static SDValue BuildExactUDIV(const TargetLowering &TLI, SDNode *N, const SDLoc &dl, SelectionDAG &DAG, SmallVectorImpl &Created) { EVT VT = N->getValueType(0); EVT SVT = VT.getScalarType(); EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); EVT ShSVT = ShVT.getScalarType(); bool UseSRL = false; SmallVector Shifts, Factors; auto BuildUDIVPattern = [&](ConstantSDNode *C) { if (C->isZero()) return false; APInt Divisor = C->getAPIntValue(); unsigned Shift = Divisor.countr_zero(); if (Shift) { Divisor.lshrInPlace(Shift); UseSRL = true; } // Calculate the multiplicative inverse modulo BW. APInt Factor = Divisor.multiplicativeInverse(); Shifts.push_back(DAG.getConstant(Shift, dl, ShSVT)); Factors.push_back(DAG.getConstant(Factor, dl, SVT)); return true; }; SDValue Op1 = N->getOperand(1); // Collect all magic values from the build vector. if (!ISD::matchUnaryPredicate(Op1, BuildUDIVPattern)) return SDValue(); SDValue Shift, Factor; if (Op1.getOpcode() == ISD::BUILD_VECTOR) { Shift = DAG.getBuildVector(ShVT, dl, Shifts); Factor = DAG.getBuildVector(VT, dl, Factors); } else if (Op1.getOpcode() == ISD::SPLAT_VECTOR) { assert(Shifts.size() == 1 && Factors.size() == 1 && "Expected matchUnaryPredicate to return one element for scalable " "vectors"); Shift = DAG.getSplatVector(ShVT, dl, Shifts[0]); Factor = DAG.getSplatVector(VT, dl, Factors[0]); } else { assert(isa(Op1) && "Expected a constant"); Shift = Shifts[0]; Factor = Factors[0]; } SDValue Res = N->getOperand(0); if (UseSRL) { SDNodeFlags Flags; Flags.setExact(true); Res = DAG.getNode(ISD::SRL, dl, VT, Res, Shift, Flags); Created.push_back(Res.getNode()); } return DAG.getNode(ISD::MUL, dl, VT, Res, Factor); } SDValue TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, SmallVectorImpl &Created) const { AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); if (TLI.isIntDivCheap(N->getValueType(0), Attr)) return SDValue(N, 0); // Lower SDIV as SDIV return SDValue(); } SDValue TargetLowering::BuildSREMPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, SmallVectorImpl &Created) const { AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); if (TLI.isIntDivCheap(N->getValueType(0), Attr)) return SDValue(N, 0); // Lower SREM as SREM return SDValue(); } /// Build sdiv by power-of-2 with conditional move instructions /// Ref: "Hacker's Delight" by Henry Warren 10-1 /// If conditional move/branch is preferred, we lower sdiv x, +/-2**k into: /// bgez x, label /// add x, x, 2**k-1 /// label: /// sra res, x, k /// neg res, res (when the divisor is negative) SDValue TargetLowering::buildSDIVPow2WithCMov( SDNode *N, const APInt &Divisor, SelectionDAG &DAG, SmallVectorImpl &Created) const { unsigned Lg2 = Divisor.countr_zero(); EVT VT = N->getValueType(0); SDLoc DL(N); SDValue N0 = N->getOperand(0); SDValue Zero = DAG.getConstant(0, DL, VT); APInt Lg2Mask = APInt::getLowBitsSet(VT.getSizeInBits(), Lg2); SDValue Pow2MinusOne = DAG.getConstant(Lg2Mask, DL, VT); // If N0 is negative, we need to add (Pow2 - 1) to it before shifting right. EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDValue Cmp = DAG.getSetCC(DL, CCVT, N0, Zero, ISD::SETLT); SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne); SDValue CMov = DAG.getNode(ISD::SELECT, DL, VT, Cmp, Add, N0); Created.push_back(Cmp.getNode()); Created.push_back(Add.getNode()); Created.push_back(CMov.getNode()); // Divide by pow2. SDValue SRA = DAG.getNode(ISD::SRA, DL, VT, CMov, DAG.getConstant(Lg2, DL, VT)); // If we're dividing by a positive value, we're done. Otherwise, we must // negate the result. if (Divisor.isNonNegative()) return SRA; Created.push_back(SRA.getNode()); return DAG.getNode(ISD::SUB, DL, VT, Zero, SRA); } /// Given an ISD::SDIV node expressing a divide by constant, /// return a DAG expression to select that will generate the same value by /// multiplying by a magic number. /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, SmallVectorImpl &Created) const { SDLoc dl(N); EVT VT = N->getValueType(0); EVT SVT = VT.getScalarType(); EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); EVT ShSVT = ShVT.getScalarType(); unsigned EltBits = VT.getScalarSizeInBits(); EVT MulVT; // Check to see if we can do this. // FIXME: We should be more aggressive here. if (!isTypeLegal(VT)) { // Limit this to simple scalars for now. if (VT.isVector() || !VT.isSimple()) return SDValue(); // If this type will be promoted to a large enough type with a legal // multiply operation, we can go ahead and do this transform. if (getTypeAction(VT.getSimpleVT()) != TypePromoteInteger) return SDValue(); MulVT = getTypeToTransformTo(*DAG.getContext(), VT); if (MulVT.getSizeInBits() < (2 * EltBits) || !isOperationLegal(ISD::MUL, MulVT)) return SDValue(); } // If the sdiv has an 'exact' bit we can use a simpler lowering. if (N->getFlags().hasExact()) return BuildExactSDIV(*this, N, dl, DAG, Created); SmallVector MagicFactors, Factors, Shifts, ShiftMasks; auto BuildSDIVPattern = [&](ConstantSDNode *C) { if (C->isZero()) return false; const APInt &Divisor = C->getAPIntValue(); SignedDivisionByConstantInfo magics = SignedDivisionByConstantInfo::get(Divisor); int NumeratorFactor = 0; int ShiftMask = -1; if (Divisor.isOne() || Divisor.isAllOnes()) { // If d is +1/-1, we just multiply the numerator by +1/-1. NumeratorFactor = Divisor.getSExtValue(); magics.Magic = 0; magics.ShiftAmount = 0; ShiftMask = 0; } else if (Divisor.isStrictlyPositive() && magics.Magic.isNegative()) { // If d > 0 and m < 0, add the numerator. NumeratorFactor = 1; } else if (Divisor.isNegative() && magics.Magic.isStrictlyPositive()) { // If d < 0 and m > 0, subtract the numerator. NumeratorFactor = -1; } MagicFactors.push_back(DAG.getConstant(magics.Magic, dl, SVT)); Factors.push_back(DAG.getConstant(NumeratorFactor, dl, SVT)); Shifts.push_back(DAG.getConstant(magics.ShiftAmount, dl, ShSVT)); ShiftMasks.push_back(DAG.getConstant(ShiftMask, dl, SVT)); return true; }; SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); // Collect the shifts / magic values from each element. if (!ISD::matchUnaryPredicate(N1, BuildSDIVPattern)) return SDValue(); SDValue MagicFactor, Factor, Shift, ShiftMask; if (N1.getOpcode() == ISD::BUILD_VECTOR) { MagicFactor = DAG.getBuildVector(VT, dl, MagicFactors); Factor = DAG.getBuildVector(VT, dl, Factors); Shift = DAG.getBuildVector(ShVT, dl, Shifts); ShiftMask = DAG.getBuildVector(VT, dl, ShiftMasks); } else if (N1.getOpcode() == ISD::SPLAT_VECTOR) { assert(MagicFactors.size() == 1 && Factors.size() == 1 && Shifts.size() == 1 && ShiftMasks.size() == 1 && "Expected matchUnaryPredicate to return one element for scalable " "vectors"); MagicFactor = DAG.getSplatVector(VT, dl, MagicFactors[0]); Factor = DAG.getSplatVector(VT, dl, Factors[0]); Shift = DAG.getSplatVector(ShVT, dl, Shifts[0]); ShiftMask = DAG.getSplatVector(VT, dl, ShiftMasks[0]); } else { assert(isa(N1) && "Expected a constant"); MagicFactor = MagicFactors[0]; Factor = Factors[0]; Shift = Shifts[0]; ShiftMask = ShiftMasks[0]; } // Multiply the numerator (operand 0) by the magic value. // FIXME: We should support doing a MUL in a wider type. auto GetMULHS = [&](SDValue X, SDValue Y) { // If the type isn't legal, use a wider mul of the type calculated // earlier. if (!isTypeLegal(VT)) { X = DAG.getNode(ISD::SIGN_EXTEND, dl, MulVT, X); Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MulVT, Y); Y = DAG.getNode(ISD::MUL, dl, MulVT, X, Y); Y = DAG.getNode(ISD::SRL, dl, MulVT, Y, DAG.getShiftAmountConstant(EltBits, MulVT, dl)); return DAG.getNode(ISD::TRUNCATE, dl, VT, Y); } if (isOperationLegalOrCustom(ISD::MULHS, VT, IsAfterLegalization)) return DAG.getNode(ISD::MULHS, dl, VT, X, Y); if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT, IsAfterLegalization)) { SDValue LoHi = DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), X, Y); return SDValue(LoHi.getNode(), 1); } // If type twice as wide legal, widen and use a mul plus a shift. unsigned Size = VT.getScalarSizeInBits(); EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), Size * 2); if (VT.isVector()) WideVT = EVT::getVectorVT(*DAG.getContext(), WideVT, VT.getVectorElementCount()); if (isOperationLegalOrCustom(ISD::MUL, WideVT)) { X = DAG.getNode(ISD::SIGN_EXTEND, dl, WideVT, X); Y = DAG.getNode(ISD::SIGN_EXTEND, dl, WideVT, Y); Y = DAG.getNode(ISD::MUL, dl, WideVT, X, Y); Y = DAG.getNode(ISD::SRL, dl, WideVT, Y, DAG.getShiftAmountConstant(EltBits, WideVT, dl)); return DAG.getNode(ISD::TRUNCATE, dl, VT, Y); } return SDValue(); }; SDValue Q = GetMULHS(N0, MagicFactor); if (!Q) return SDValue(); Created.push_back(Q.getNode()); // (Optionally) Add/subtract the numerator using Factor. Factor = DAG.getNode(ISD::MUL, dl, VT, N0, Factor); Created.push_back(Factor.getNode()); Q = DAG.getNode(ISD::ADD, dl, VT, Q, Factor); Created.push_back(Q.getNode()); // Shift right algebraic by shift value. Q = DAG.getNode(ISD::SRA, dl, VT, Q, Shift); Created.push_back(Q.getNode()); // Extract the sign bit, mask it and add it to the quotient. SDValue SignShift = DAG.getConstant(EltBits - 1, dl, ShVT); SDValue T = DAG.getNode(ISD::SRL, dl, VT, Q, SignShift); Created.push_back(T.getNode()); T = DAG.getNode(ISD::AND, dl, VT, T, ShiftMask); Created.push_back(T.getNode()); return DAG.getNode(ISD::ADD, dl, VT, Q, T); } /// Given an ISD::UDIV node expressing a divide by constant, /// return a DAG expression to select that will generate the same value by /// multiplying by a magic number. /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, SmallVectorImpl &Created) const { SDLoc dl(N); EVT VT = N->getValueType(0); EVT SVT = VT.getScalarType(); EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); EVT ShSVT = ShVT.getScalarType(); unsigned EltBits = VT.getScalarSizeInBits(); EVT MulVT; // Check to see if we can do this. // FIXME: We should be more aggressive here. if (!isTypeLegal(VT)) { // Limit this to simple scalars for now. if (VT.isVector() || !VT.isSimple()) return SDValue(); // If this type will be promoted to a large enough type with a legal // multiply operation, we can go ahead and do this transform. if (getTypeAction(VT.getSimpleVT()) != TypePromoteInteger) return SDValue(); MulVT = getTypeToTransformTo(*DAG.getContext(), VT); if (MulVT.getSizeInBits() < (2 * EltBits) || !isOperationLegal(ISD::MUL, MulVT)) return SDValue(); } // If the udiv has an 'exact' bit we can use a simpler lowering. if (N->getFlags().hasExact()) return BuildExactUDIV(*this, N, dl, DAG, Created); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); // Try to use leading zeros of the dividend to reduce the multiplier and // avoid expensive fixups. unsigned KnownLeadingZeros = DAG.computeKnownBits(N0).countMinLeadingZeros(); bool UseNPQ = false, UsePreShift = false, UsePostShift = false; SmallVector PreShifts, PostShifts, MagicFactors, NPQFactors; auto BuildUDIVPattern = [&](ConstantSDNode *C) { if (C->isZero()) return false; const APInt& Divisor = C->getAPIntValue(); SDValue PreShift, MagicFactor, NPQFactor, PostShift; // Magic algorithm doesn't work for division by 1. We need to emit a select // at the end. if (Divisor.isOne()) { PreShift = PostShift = DAG.getUNDEF(ShSVT); MagicFactor = NPQFactor = DAG.getUNDEF(SVT); } else { UnsignedDivisionByConstantInfo magics = UnsignedDivisionByConstantInfo::get( Divisor, std::min(KnownLeadingZeros, Divisor.countl_zero())); MagicFactor = DAG.getConstant(magics.Magic, dl, SVT); assert(magics.PreShift < Divisor.getBitWidth() && "We shouldn't generate an undefined shift!"); assert(magics.PostShift < Divisor.getBitWidth() && "We shouldn't generate an undefined shift!"); assert((!magics.IsAdd || magics.PreShift == 0) && "Unexpected pre-shift"); PreShift = DAG.getConstant(magics.PreShift, dl, ShSVT); PostShift = DAG.getConstant(magics.PostShift, dl, ShSVT); NPQFactor = DAG.getConstant( magics.IsAdd ? APInt::getOneBitSet(EltBits, EltBits - 1) : APInt::getZero(EltBits), dl, SVT); UseNPQ |= magics.IsAdd; UsePreShift |= magics.PreShift != 0; UsePostShift |= magics.PostShift != 0; } PreShifts.push_back(PreShift); MagicFactors.push_back(MagicFactor); NPQFactors.push_back(NPQFactor); PostShifts.push_back(PostShift); return true; }; // Collect the shifts/magic values from each element. if (!ISD::matchUnaryPredicate(N1, BuildUDIVPattern)) return SDValue(); SDValue PreShift, PostShift, MagicFactor, NPQFactor; if (N1.getOpcode() == ISD::BUILD_VECTOR) { PreShift = DAG.getBuildVector(ShVT, dl, PreShifts); MagicFactor = DAG.getBuildVector(VT, dl, MagicFactors); NPQFactor = DAG.getBuildVector(VT, dl, NPQFactors); PostShift = DAG.getBuildVector(ShVT, dl, PostShifts); } else if (N1.getOpcode() == ISD::SPLAT_VECTOR) { assert(PreShifts.size() == 1 && MagicFactors.size() == 1 && NPQFactors.size() == 1 && PostShifts.size() == 1 && "Expected matchUnaryPredicate to return one for scalable vectors"); PreShift = DAG.getSplatVector(ShVT, dl, PreShifts[0]); MagicFactor = DAG.getSplatVector(VT, dl, MagicFactors[0]); NPQFactor = DAG.getSplatVector(VT, dl, NPQFactors[0]); PostShift = DAG.getSplatVector(ShVT, dl, PostShifts[0]); } else { assert(isa(N1) && "Expected a constant"); PreShift = PreShifts[0]; MagicFactor = MagicFactors[0]; PostShift = PostShifts[0]; } SDValue Q = N0; if (UsePreShift) { Q = DAG.getNode(ISD::SRL, dl, VT, Q, PreShift); Created.push_back(Q.getNode()); } // FIXME: We should support doing a MUL in a wider type. auto GetMULHU = [&](SDValue X, SDValue Y) { // If the type isn't legal, use a wider mul of the type calculated // earlier. if (!isTypeLegal(VT)) { X = DAG.getNode(ISD::ZERO_EXTEND, dl, MulVT, X); Y = DAG.getNode(ISD::ZERO_EXTEND, dl, MulVT, Y); Y = DAG.getNode(ISD::MUL, dl, MulVT, X, Y); Y = DAG.getNode(ISD::SRL, dl, MulVT, Y, DAG.getShiftAmountConstant(EltBits, MulVT, dl)); return DAG.getNode(ISD::TRUNCATE, dl, VT, Y); } if (isOperationLegalOrCustom(ISD::MULHU, VT, IsAfterLegalization)) return DAG.getNode(ISD::MULHU, dl, VT, X, Y); if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT, IsAfterLegalization)) { SDValue LoHi = DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), X, Y); return SDValue(LoHi.getNode(), 1); } // If type twice as wide legal, widen and use a mul plus a shift. unsigned Size = VT.getScalarSizeInBits(); EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), Size * 2); if (VT.isVector()) WideVT = EVT::getVectorVT(*DAG.getContext(), WideVT, VT.getVectorElementCount()); if (isOperationLegalOrCustom(ISD::MUL, WideVT)) { X = DAG.getNode(ISD::ZERO_EXTEND, dl, WideVT, X); Y = DAG.getNode(ISD::ZERO_EXTEND, dl, WideVT, Y); Y = DAG.getNode(ISD::MUL, dl, WideVT, X, Y); Y = DAG.getNode(ISD::SRL, dl, WideVT, Y, DAG.getShiftAmountConstant(EltBits, WideVT, dl)); return DAG.getNode(ISD::TRUNCATE, dl, VT, Y); } return SDValue(); // No mulhu or equivalent }; // Multiply the numerator (operand 0) by the magic value. Q = GetMULHU(Q, MagicFactor); if (!Q) return SDValue(); Created.push_back(Q.getNode()); if (UseNPQ) { SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N0, Q); Created.push_back(NPQ.getNode()); // For vectors we might have a mix of non-NPQ/NPQ paths, so use // MULHU to act as a SRL-by-1 for NPQ, else multiply by zero. if (VT.isVector()) NPQ = GetMULHU(NPQ, NPQFactor); else NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, DAG.getConstant(1, dl, ShVT)); Created.push_back(NPQ.getNode()); Q = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); Created.push_back(Q.getNode()); } if (UsePostShift) { Q = DAG.getNode(ISD::SRL, dl, VT, Q, PostShift); Created.push_back(Q.getNode()); } EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDValue One = DAG.getConstant(1, dl, VT); SDValue IsOne = DAG.getSetCC(dl, SetCCVT, N1, One, ISD::SETEQ); return DAG.getSelect(dl, VT, IsOne, N0, Q); } /// If all values in Values that *don't* match the predicate are same 'splat' /// value, then replace all values with that splat value. /// Else, if AlternativeReplacement was provided, then replace all values that /// do match predicate with AlternativeReplacement value. static void turnVectorIntoSplatVector(MutableArrayRef Values, std::function Predicate, SDValue AlternativeReplacement = SDValue()) { SDValue Replacement; // Is there a value for which the Predicate does *NOT* match? What is it? auto SplatValue = llvm::find_if_not(Values, Predicate); if (SplatValue != Values.end()) { // Does Values consist only of SplatValue's and values matching Predicate? if (llvm::all_of(Values, [Predicate, SplatValue](SDValue Value) { return Value == *SplatValue || Predicate(Value); })) // Then we shall replace values matching predicate with SplatValue. Replacement = *SplatValue; } if (!Replacement) { // Oops, we did not find the "baseline" splat value. if (!AlternativeReplacement) return; // Nothing to do. // Let's replace with provided value then. Replacement = AlternativeReplacement; } std::replace_if(Values.begin(), Values.end(), Predicate, Replacement); } /// Given an ISD::UREM used only by an ISD::SETEQ or ISD::SETNE /// where the divisor is constant and the comparison target is zero, /// return a DAG expression that will generate the same comparison result /// using only multiplications, additions and shifts/rotations. /// Ref: "Hacker's Delight" 10-17. SDValue TargetLowering::buildUREMEqFold(EVT SETCCVT, SDValue REMNode, SDValue CompTargetNode, ISD::CondCode Cond, DAGCombinerInfo &DCI, const SDLoc &DL) const { SmallVector Built; if (SDValue Folded = prepareUREMEqFold(SETCCVT, REMNode, CompTargetNode, Cond, DCI, DL, Built)) { for (SDNode *N : Built) DCI.AddToWorklist(N); return Folded; } return SDValue(); } SDValue TargetLowering::prepareUREMEqFold(EVT SETCCVT, SDValue REMNode, SDValue CompTargetNode, ISD::CondCode Cond, DAGCombinerInfo &DCI, const SDLoc &DL, SmallVectorImpl &Created) const { // fold (seteq/ne (urem N, D), 0) -> (setule/ugt (rotr (mul N, P), K), Q) // - D must be constant, with D = D0 * 2^K where D0 is odd // - P is the multiplicative inverse of D0 modulo 2^W // - Q = floor(((2^W) - 1) / D) // where W is the width of the common type of N and D. assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Only applicable for (in)equality comparisons."); SelectionDAG &DAG = DCI.DAG; EVT VT = REMNode.getValueType(); EVT SVT = VT.getScalarType(); EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); EVT ShSVT = ShVT.getScalarType(); // If MUL is unavailable, we cannot proceed in any case. if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::MUL, VT)) return SDValue(); bool ComparingWithAllZeros = true; bool AllComparisonsWithNonZerosAreTautological = true; bool HadTautologicalLanes = false; bool AllLanesAreTautological = true; bool HadEvenDivisor = false; bool AllDivisorsArePowerOfTwo = true; bool HadTautologicalInvertedLanes = false; SmallVector PAmts, KAmts, QAmts, IAmts; auto BuildUREMPattern = [&](ConstantSDNode *CDiv, ConstantSDNode *CCmp) { // Division by 0 is UB. Leave it to be constant-folded elsewhere. if (CDiv->isZero()) return false; const APInt &D = CDiv->getAPIntValue(); const APInt &Cmp = CCmp->getAPIntValue(); ComparingWithAllZeros &= Cmp.isZero(); // x u% C1` is *always* less than C1. So given `x u% C1 == C2`, // if C2 is not less than C1, the comparison is always false. // But we will only be able to produce the comparison that will give the // opposive tautological answer. So this lane would need to be fixed up. bool TautologicalInvertedLane = D.ule(Cmp); HadTautologicalInvertedLanes |= TautologicalInvertedLane; // If all lanes are tautological (either all divisors are ones, or divisor // is not greater than the constant we are comparing with), // we will prefer to avoid the fold. bool TautologicalLane = D.isOne() || TautologicalInvertedLane; HadTautologicalLanes |= TautologicalLane; AllLanesAreTautological &= TautologicalLane; // If we are comparing with non-zero, we need'll need to subtract said // comparison value from the LHS. But there is no point in doing that if // every lane where we are comparing with non-zero is tautological.. if (!Cmp.isZero()) AllComparisonsWithNonZerosAreTautological &= TautologicalLane; // Decompose D into D0 * 2^K unsigned K = D.countr_zero(); assert((!D.isOne() || (K == 0)) && "For divisor '1' we won't rotate."); APInt D0 = D.lshr(K); // D is even if it has trailing zeros. HadEvenDivisor |= (K != 0); // D is a power-of-two if D0 is one. // If all divisors are power-of-two, we will prefer to avoid the fold. AllDivisorsArePowerOfTwo &= D0.isOne(); // P = inv(D0, 2^W) // 2^W requires W + 1 bits, so we have to extend and then truncate. unsigned W = D.getBitWidth(); APInt P = D0.multiplicativeInverse(); assert((D0 * P).isOne() && "Multiplicative inverse basic check failed."); // Q = floor((2^W - 1) u/ D) // R = ((2^W - 1) u% D) APInt Q, R; APInt::udivrem(APInt::getAllOnes(W), D, Q, R); // If we are comparing with zero, then that comparison constant is okay, // else it may need to be one less than that. if (Cmp.ugt(R)) Q -= 1; assert(APInt::getAllOnes(ShSVT.getSizeInBits()).ugt(K) && "We are expecting that K is always less than all-ones for ShSVT"); // If the lane is tautological the result can be constant-folded. if (TautologicalLane) { // Set P and K amount to a bogus values so we can try to splat them. P = 0; K = -1; // And ensure that comparison constant is tautological, // it will always compare true/false. Q = -1; } PAmts.push_back(DAG.getConstant(P, DL, SVT)); KAmts.push_back( DAG.getConstant(APInt(ShSVT.getSizeInBits(), K), DL, ShSVT)); QAmts.push_back(DAG.getConstant(Q, DL, SVT)); return true; }; SDValue N = REMNode.getOperand(0); SDValue D = REMNode.getOperand(1); // Collect the values from each element. if (!ISD::matchBinaryPredicate(D, CompTargetNode, BuildUREMPattern)) return SDValue(); // If all lanes are tautological, the result can be constant-folded. if (AllLanesAreTautological) return SDValue(); // If this is a urem by a powers-of-two, avoid the fold since it can be // best implemented as a bit test. if (AllDivisorsArePowerOfTwo) return SDValue(); SDValue PVal, KVal, QVal; if (D.getOpcode() == ISD::BUILD_VECTOR) { if (HadTautologicalLanes) { // Try to turn PAmts into a splat, since we don't care about the values // that are currently '0'. If we can't, just keep '0'`s. turnVectorIntoSplatVector(PAmts, isNullConstant); // Try to turn KAmts into a splat, since we don't care about the values // that are currently '-1'. If we can't, change them to '0'`s. turnVectorIntoSplatVector(KAmts, isAllOnesConstant, DAG.getConstant(0, DL, ShSVT)); } PVal = DAG.getBuildVector(VT, DL, PAmts); KVal = DAG.getBuildVector(ShVT, DL, KAmts); QVal = DAG.getBuildVector(VT, DL, QAmts); } else if (D.getOpcode() == ISD::SPLAT_VECTOR) { assert(PAmts.size() == 1 && KAmts.size() == 1 && QAmts.size() == 1 && "Expected matchBinaryPredicate to return one element for " "SPLAT_VECTORs"); PVal = DAG.getSplatVector(VT, DL, PAmts[0]); KVal = DAG.getSplatVector(ShVT, DL, KAmts[0]); QVal = DAG.getSplatVector(VT, DL, QAmts[0]); } else { PVal = PAmts[0]; KVal = KAmts[0]; QVal = QAmts[0]; } if (!ComparingWithAllZeros && !AllComparisonsWithNonZerosAreTautological) { if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::SUB, VT)) return SDValue(); // FIXME: Could/should use `ISD::ADD`? assert(CompTargetNode.getValueType() == N.getValueType() && "Expecting that the types on LHS and RHS of comparisons match."); N = DAG.getNode(ISD::SUB, DL, VT, N, CompTargetNode); } // (mul N, P) SDValue Op0 = DAG.getNode(ISD::MUL, DL, VT, N, PVal); Created.push_back(Op0.getNode()); // Rotate right only if any divisor was even. We avoid rotates for all-odd // divisors as a performance improvement, since rotating by 0 is a no-op. if (HadEvenDivisor) { // We need ROTR to do this. if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::ROTR, VT)) return SDValue(); // UREM: (rotr (mul N, P), K) Op0 = DAG.getNode(ISD::ROTR, DL, VT, Op0, KVal); Created.push_back(Op0.getNode()); } // UREM: (setule/setugt (rotr (mul N, P), K), Q) SDValue NewCC = DAG.getSetCC(DL, SETCCVT, Op0, QVal, ((Cond == ISD::SETEQ) ? ISD::SETULE : ISD::SETUGT)); if (!HadTautologicalInvertedLanes) return NewCC; // If any lanes previously compared always-false, the NewCC will give // always-true result for them, so we need to fixup those lanes. // Or the other way around for inequality predicate. assert(VT.isVector() && "Can/should only get here for vectors."); Created.push_back(NewCC.getNode()); // x u% C1` is *always* less than C1. So given `x u% C1 == C2`, // if C2 is not less than C1, the comparison is always false. // But we have produced the comparison that will give the // opposive tautological answer. So these lanes would need to be fixed up. SDValue TautologicalInvertedChannels = DAG.getSetCC(DL, SETCCVT, D, CompTargetNode, ISD::SETULE); Created.push_back(TautologicalInvertedChannels.getNode()); // NOTE: we avoid letting illegal types through even if we're before legalize // ops – legalization has a hard time producing good code for this. if (isOperationLegalOrCustom(ISD::VSELECT, SETCCVT)) { // If we have a vector select, let's replace the comparison results in the // affected lanes with the correct tautological result. SDValue Replacement = DAG.getBoolConstant(Cond == ISD::SETEQ ? false : true, DL, SETCCVT, SETCCVT); return DAG.getNode(ISD::VSELECT, DL, SETCCVT, TautologicalInvertedChannels, Replacement, NewCC); } // Else, we can just invert the comparison result in the appropriate lanes. // // NOTE: see the note above VSELECT above. if (isOperationLegalOrCustom(ISD::XOR, SETCCVT)) return DAG.getNode(ISD::XOR, DL, SETCCVT, NewCC, TautologicalInvertedChannels); return SDValue(); // Don't know how to lower. } /// Given an ISD::SREM used only by an ISD::SETEQ or ISD::SETNE /// where the divisor is constant and the comparison target is zero, /// return a DAG expression that will generate the same comparison result /// using only multiplications, additions and shifts/rotations. /// Ref: "Hacker's Delight" 10-17. SDValue TargetLowering::buildSREMEqFold(EVT SETCCVT, SDValue REMNode, SDValue CompTargetNode, ISD::CondCode Cond, DAGCombinerInfo &DCI, const SDLoc &DL) const { SmallVector Built; if (SDValue Folded = prepareSREMEqFold(SETCCVT, REMNode, CompTargetNode, Cond, DCI, DL, Built)) { assert(Built.size() <= 7 && "Max size prediction failed."); for (SDNode *N : Built) DCI.AddToWorklist(N); return Folded; } return SDValue(); } SDValue TargetLowering::prepareSREMEqFold(EVT SETCCVT, SDValue REMNode, SDValue CompTargetNode, ISD::CondCode Cond, DAGCombinerInfo &DCI, const SDLoc &DL, SmallVectorImpl &Created) const { // Derived from Hacker's Delight, 2nd Edition, by Hank Warren. Section 10-17. // Fold: // (seteq/ne (srem N, D), 0) // To: // (setule/ugt (rotr (add (mul N, P), A), K), Q) // // - D must be constant, with D = D0 * 2^K where D0 is odd // - P is the multiplicative inverse of D0 modulo 2^W // - A = bitwiseand(floor((2^(W - 1) - 1) / D0), (-(2^k))) // - Q = floor((2 * A) / (2^K)) // where W is the width of the common type of N and D. // // When D is a power of two (and thus D0 is 1), the normal // formula for A and Q don't apply, because the derivation // depends on D not dividing 2^(W-1), and thus theorem ZRS // does not apply. This specifically fails when N = INT_MIN. // // Instead, for power-of-two D, we use: // - A = 2^(W-1) // |-> Order-preserving map from [-2^(W-1), 2^(W-1) - 1] to [0,2^W - 1]) // - Q = 2^(W-K) - 1 // |-> Test that the top K bits are zero after rotation assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Only applicable for (in)equality comparisons."); SelectionDAG &DAG = DCI.DAG; EVT VT = REMNode.getValueType(); EVT SVT = VT.getScalarType(); EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); EVT ShSVT = ShVT.getScalarType(); // If we are after ops legalization, and MUL is unavailable, we can not // proceed. if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::MUL, VT)) return SDValue(); // TODO: Could support comparing with non-zero too. ConstantSDNode *CompTarget = isConstOrConstSplat(CompTargetNode); if (!CompTarget || !CompTarget->isZero()) return SDValue(); bool HadIntMinDivisor = false; bool HadOneDivisor = false; bool AllDivisorsAreOnes = true; bool HadEvenDivisor = false; bool NeedToApplyOffset = false; bool AllDivisorsArePowerOfTwo = true; SmallVector PAmts, AAmts, KAmts, QAmts; auto BuildSREMPattern = [&](ConstantSDNode *C) { // Division by 0 is UB. Leave it to be constant-folded elsewhere. if (C->isZero()) return false; // FIXME: we don't fold `rem %X, -C` to `rem %X, C` in DAGCombine. // WARNING: this fold is only valid for positive divisors! APInt D = C->getAPIntValue(); if (D.isNegative()) D.negate(); // `rem %X, -C` is equivalent to `rem %X, C` HadIntMinDivisor |= D.isMinSignedValue(); // If all divisors are ones, we will prefer to avoid the fold. HadOneDivisor |= D.isOne(); AllDivisorsAreOnes &= D.isOne(); // Decompose D into D0 * 2^K unsigned K = D.countr_zero(); assert((!D.isOne() || (K == 0)) && "For divisor '1' we won't rotate."); APInt D0 = D.lshr(K); if (!D.isMinSignedValue()) { // D is even if it has trailing zeros; unless it's INT_MIN, in which case // we don't care about this lane in this fold, we'll special-handle it. HadEvenDivisor |= (K != 0); } // D is a power-of-two if D0 is one. This includes INT_MIN. // If all divisors are power-of-two, we will prefer to avoid the fold. AllDivisorsArePowerOfTwo &= D0.isOne(); // P = inv(D0, 2^W) // 2^W requires W + 1 bits, so we have to extend and then truncate. unsigned W = D.getBitWidth(); APInt P = D0.multiplicativeInverse(); assert((D0 * P).isOne() && "Multiplicative inverse basic check failed."); // A = floor((2^(W - 1) - 1) / D0) & -2^K APInt A = APInt::getSignedMaxValue(W).udiv(D0); A.clearLowBits(K); if (!D.isMinSignedValue()) { // If divisor INT_MIN, then we don't care about this lane in this fold, // we'll special-handle it. NeedToApplyOffset |= A != 0; } // Q = floor((2 * A) / (2^K)) APInt Q = (2 * A).udiv(APInt::getOneBitSet(W, K)); assert(APInt::getAllOnes(SVT.getSizeInBits()).ugt(A) && "We are expecting that A is always less than all-ones for SVT"); assert(APInt::getAllOnes(ShSVT.getSizeInBits()).ugt(K) && "We are expecting that K is always less than all-ones for ShSVT"); // If D was a power of two, apply the alternate constant derivation. if (D0.isOne()) { // A = 2^(W-1) A = APInt::getSignedMinValue(W); // - Q = 2^(W-K) - 1 Q = APInt::getAllOnes(W - K).zext(W); } // If the divisor is 1 the result can be constant-folded. Likewise, we // don't care about INT_MIN lanes, those can be set to undef if appropriate. if (D.isOne()) { // Set P, A and K to a bogus values so we can try to splat them. P = 0; A = -1; K = -1; // x ?% 1 == 0 <--> true <--> x u<= -1 Q = -1; } PAmts.push_back(DAG.getConstant(P, DL, SVT)); AAmts.push_back(DAG.getConstant(A, DL, SVT)); KAmts.push_back( DAG.getConstant(APInt(ShSVT.getSizeInBits(), K), DL, ShSVT)); QAmts.push_back(DAG.getConstant(Q, DL, SVT)); return true; }; SDValue N = REMNode.getOperand(0); SDValue D = REMNode.getOperand(1); // Collect the values from each element. if (!ISD::matchUnaryPredicate(D, BuildSREMPattern)) return SDValue(); // If this is a srem by a one, avoid the fold since it can be constant-folded. if (AllDivisorsAreOnes) return SDValue(); // If this is a srem by a powers-of-two (including INT_MIN), avoid the fold // since it can be best implemented as a bit test. if (AllDivisorsArePowerOfTwo) return SDValue(); SDValue PVal, AVal, KVal, QVal; if (D.getOpcode() == ISD::BUILD_VECTOR) { if (HadOneDivisor) { // Try to turn PAmts into a splat, since we don't care about the values // that are currently '0'. If we can't, just keep '0'`s. turnVectorIntoSplatVector(PAmts, isNullConstant); // Try to turn AAmts into a splat, since we don't care about the // values that are currently '-1'. If we can't, change them to '0'`s. turnVectorIntoSplatVector(AAmts, isAllOnesConstant, DAG.getConstant(0, DL, SVT)); // Try to turn KAmts into a splat, since we don't care about the values // that are currently '-1'. If we can't, change them to '0'`s. turnVectorIntoSplatVector(KAmts, isAllOnesConstant, DAG.getConstant(0, DL, ShSVT)); } PVal = DAG.getBuildVector(VT, DL, PAmts); AVal = DAG.getBuildVector(VT, DL, AAmts); KVal = DAG.getBuildVector(ShVT, DL, KAmts); QVal = DAG.getBuildVector(VT, DL, QAmts); } else if (D.getOpcode() == ISD::SPLAT_VECTOR) { assert(PAmts.size() == 1 && AAmts.size() == 1 && KAmts.size() == 1 && QAmts.size() == 1 && "Expected matchUnaryPredicate to return one element for scalable " "vectors"); PVal = DAG.getSplatVector(VT, DL, PAmts[0]); AVal = DAG.getSplatVector(VT, DL, AAmts[0]); KVal = DAG.getSplatVector(ShVT, DL, KAmts[0]); QVal = DAG.getSplatVector(VT, DL, QAmts[0]); } else { assert(isa(D) && "Expected a constant"); PVal = PAmts[0]; AVal = AAmts[0]; KVal = KAmts[0]; QVal = QAmts[0]; } // (mul N, P) SDValue Op0 = DAG.getNode(ISD::MUL, DL, VT, N, PVal); Created.push_back(Op0.getNode()); if (NeedToApplyOffset) { // We need ADD to do this. if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::ADD, VT)) return SDValue(); // (add (mul N, P), A) Op0 = DAG.getNode(ISD::ADD, DL, VT, Op0, AVal); Created.push_back(Op0.getNode()); } // Rotate right only if any divisor was even. We avoid rotates for all-odd // divisors as a performance improvement, since rotating by 0 is a no-op. if (HadEvenDivisor) { // We need ROTR to do this. if (!DCI.isBeforeLegalizeOps() && !isOperationLegalOrCustom(ISD::ROTR, VT)) return SDValue(); // SREM: (rotr (add (mul N, P), A), K) Op0 = DAG.getNode(ISD::ROTR, DL, VT, Op0, KVal); Created.push_back(Op0.getNode()); } // SREM: (setule/setugt (rotr (add (mul N, P), A), K), Q) SDValue Fold = DAG.getSetCC(DL, SETCCVT, Op0, QVal, ((Cond == ISD::SETEQ) ? ISD::SETULE : ISD::SETUGT)); // If we didn't have lanes with INT_MIN divisor, then we're done. if (!HadIntMinDivisor) return Fold; // That fold is only valid for positive divisors. Which effectively means, // it is invalid for INT_MIN divisors. So if we have such a lane, // we must fix-up results for said lanes. assert(VT.isVector() && "Can/should only get here for vectors."); // NOTE: we avoid letting illegal types through even if we're before legalize // ops – legalization has a hard time producing good code for the code that // follows. if (!isOperationLegalOrCustom(ISD::SETCC, SETCCVT) || !isOperationLegalOrCustom(ISD::AND, VT) || !isCondCodeLegalOrCustom(Cond, VT.getSimpleVT()) || !isOperationLegalOrCustom(ISD::VSELECT, SETCCVT)) return SDValue(); Created.push_back(Fold.getNode()); SDValue IntMin = DAG.getConstant( APInt::getSignedMinValue(SVT.getScalarSizeInBits()), DL, VT); SDValue IntMax = DAG.getConstant( APInt::getSignedMaxValue(SVT.getScalarSizeInBits()), DL, VT); SDValue Zero = DAG.getConstant(APInt::getZero(SVT.getScalarSizeInBits()), DL, VT); // Which lanes had INT_MIN divisors? Divisor is constant, so const-folded. SDValue DivisorIsIntMin = DAG.getSetCC(DL, SETCCVT, D, IntMin, ISD::SETEQ); Created.push_back(DivisorIsIntMin.getNode()); // (N s% INT_MIN) ==/!= 0 <--> (N & INT_MAX) ==/!= 0 SDValue Masked = DAG.getNode(ISD::AND, DL, VT, N, IntMax); Created.push_back(Masked.getNode()); SDValue MaskedIsZero = DAG.getSetCC(DL, SETCCVT, Masked, Zero, Cond); Created.push_back(MaskedIsZero.getNode()); // To produce final result we need to blend 2 vectors: 'SetCC' and // 'MaskedIsZero'. If the divisor for channel was *NOT* INT_MIN, we pick // from 'Fold', else pick from 'MaskedIsZero'. Since 'DivisorIsIntMin' is // constant-folded, select can get lowered to a shuffle with constant mask. SDValue Blended = DAG.getNode(ISD::VSELECT, DL, SETCCVT, DivisorIsIntMin, MaskedIsZero, Fold); return Blended; } bool TargetLowering:: verifyReturnAddressArgumentIsConstant(SDValue Op, SelectionDAG &DAG) const { if (!isa(Op.getOperand(0))) { DAG.getContext()->emitError("argument to '__builtin_return_address' must " "be a constant integer"); return true; } return false; } SDValue TargetLowering::getSqrtInputTest(SDValue Op, SelectionDAG &DAG, const DenormalMode &Mode) const { SDLoc DL(Op); EVT VT = Op.getValueType(); EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDValue FPZero = DAG.getConstantFP(0.0, DL, VT); // This is specifically a check for the handling of denormal inputs, not the // result. if (Mode.Input == DenormalMode::PreserveSign || Mode.Input == DenormalMode::PositiveZero) { // Test = X == 0.0 return DAG.getSetCC(DL, CCVT, Op, FPZero, ISD::SETEQ); } // Testing it with denormal inputs to avoid wrong estimate. // // Test = fabs(X) < SmallestNormal const fltSemantics &FltSem = DAG.EVTToAPFloatSemantics(VT); APFloat SmallestNorm = APFloat::getSmallestNormalized(FltSem); SDValue NormC = DAG.getConstantFP(SmallestNorm, DL, VT); SDValue Fabs = DAG.getNode(ISD::FABS, DL, VT, Op); return DAG.getSetCC(DL, CCVT, Fabs, NormC, ISD::SETLT); } SDValue TargetLowering::getNegatedExpression(SDValue Op, SelectionDAG &DAG, bool LegalOps, bool OptForSize, NegatibleCost &Cost, unsigned Depth) const { // fneg is removable even if it has multiple uses. if (Op.getOpcode() == ISD::FNEG || Op.getOpcode() == ISD::VP_FNEG) { Cost = NegatibleCost::Cheaper; return Op.getOperand(0); } // Don't recurse exponentially. if (Depth > SelectionDAG::MaxRecursionDepth) return SDValue(); // Pre-increment recursion depth for use in recursive calls. ++Depth; const SDNodeFlags Flags = Op->getFlags(); const TargetOptions &Options = DAG.getTarget().Options; EVT VT = Op.getValueType(); unsigned Opcode = Op.getOpcode(); // Don't allow anything with multiple uses unless we know it is free. if (!Op.hasOneUse() && Opcode != ISD::ConstantFP) { bool IsFreeExtend = Opcode == ISD::FP_EXTEND && isFPExtFree(VT, Op.getOperand(0).getValueType()); if (!IsFreeExtend) return SDValue(); } auto RemoveDeadNode = [&](SDValue N) { if (N && N.getNode()->use_empty()) DAG.RemoveDeadNode(N.getNode()); }; SDLoc DL(Op); // Because getNegatedExpression can delete nodes we need a handle to keep // temporary nodes alive in case the recursion manages to create an identical // node. std::list Handles; switch (Opcode) { case ISD::ConstantFP: { // Don't invert constant FP values after legalization unless the target says // the negated constant is legal. bool IsOpLegal = isOperationLegal(ISD::ConstantFP, VT) || isFPImmLegal(neg(cast(Op)->getValueAPF()), VT, OptForSize); if (LegalOps && !IsOpLegal) break; APFloat V = cast(Op)->getValueAPF(); V.changeSign(); SDValue CFP = DAG.getConstantFP(V, DL, VT); // If we already have the use of the negated floating constant, it is free // to negate it even it has multiple uses. if (!Op.hasOneUse() && CFP.use_empty()) break; Cost = NegatibleCost::Neutral; return CFP; } case ISD::BUILD_VECTOR: { // Only permit BUILD_VECTOR of constants. if (llvm::any_of(Op->op_values(), [&](SDValue N) { return !N.isUndef() && !isa(N); })) break; bool IsOpLegal = (isOperationLegal(ISD::ConstantFP, VT) && isOperationLegal(ISD::BUILD_VECTOR, VT)) || llvm::all_of(Op->op_values(), [&](SDValue N) { return N.isUndef() || isFPImmLegal(neg(cast(N)->getValueAPF()), VT, OptForSize); }); if (LegalOps && !IsOpLegal) break; SmallVector Ops; for (SDValue C : Op->op_values()) { if (C.isUndef()) { Ops.push_back(C); continue; } APFloat V = cast(C)->getValueAPF(); V.changeSign(); Ops.push_back(DAG.getConstantFP(V, DL, C.getValueType())); } Cost = NegatibleCost::Neutral; return DAG.getBuildVector(VT, DL, Ops); } case ISD::FADD: { if (!Options.NoSignedZerosFPMath && !Flags.hasNoSignedZeros()) break; // After operation legalization, it might not be legal to create new FSUBs. if (LegalOps && !isOperationLegalOrCustom(ISD::FSUB, VT)) break; SDValue X = Op.getOperand(0), Y = Op.getOperand(1); // fold (fneg (fadd X, Y)) -> (fsub (fneg X), Y) NegatibleCost CostX = NegatibleCost::Expensive; SDValue NegX = getNegatedExpression(X, DAG, LegalOps, OptForSize, CostX, Depth); // Prevent this node from being deleted by the next call. if (NegX) Handles.emplace_back(NegX); // fold (fneg (fadd X, Y)) -> (fsub (fneg Y), X) NegatibleCost CostY = NegatibleCost::Expensive; SDValue NegY = getNegatedExpression(Y, DAG, LegalOps, OptForSize, CostY, Depth); // We're done with the handles. Handles.clear(); // Negate the X if its cost is less or equal than Y. if (NegX && (CostX <= CostY)) { Cost = CostX; SDValue N = DAG.getNode(ISD::FSUB, DL, VT, NegX, Y, Flags); if (NegY != N) RemoveDeadNode(NegY); return N; } // Negate the Y if it is not expensive. if (NegY) { Cost = CostY; SDValue N = DAG.getNode(ISD::FSUB, DL, VT, NegY, X, Flags); if (NegX != N) RemoveDeadNode(NegX); return N; } break; } case ISD::FSUB: { // We can't turn -(A-B) into B-A when we honor signed zeros. if (!Options.NoSignedZerosFPMath && !Flags.hasNoSignedZeros()) break; SDValue X = Op.getOperand(0), Y = Op.getOperand(1); // fold (fneg (fsub 0, Y)) -> Y if (ConstantFPSDNode *C = isConstOrConstSplatFP(X, /*AllowUndefs*/ true)) if (C->isZero()) { Cost = NegatibleCost::Cheaper; return Y; } // fold (fneg (fsub X, Y)) -> (fsub Y, X) Cost = NegatibleCost::Neutral; return DAG.getNode(ISD::FSUB, DL, VT, Y, X, Flags); } case ISD::FMUL: case ISD::FDIV: { SDValue X = Op.getOperand(0), Y = Op.getOperand(1); // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) NegatibleCost CostX = NegatibleCost::Expensive; SDValue NegX = getNegatedExpression(X, DAG, LegalOps, OptForSize, CostX, Depth); // Prevent this node from being deleted by the next call. if (NegX) Handles.emplace_back(NegX); // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y)) NegatibleCost CostY = NegatibleCost::Expensive; SDValue NegY = getNegatedExpression(Y, DAG, LegalOps, OptForSize, CostY, Depth); // We're done with the handles. Handles.clear(); // Negate the X if its cost is less or equal than Y. if (NegX && (CostX <= CostY)) { Cost = CostX; SDValue N = DAG.getNode(Opcode, DL, VT, NegX, Y, Flags); if (NegY != N) RemoveDeadNode(NegY); return N; } // Ignore X * 2.0 because that is expected to be canonicalized to X + X. if (auto *C = isConstOrConstSplatFP(Op.getOperand(1))) if (C->isExactlyValue(2.0) && Op.getOpcode() == ISD::FMUL) break; // Negate the Y if it is not expensive. if (NegY) { Cost = CostY; SDValue N = DAG.getNode(Opcode, DL, VT, X, NegY, Flags); if (NegX != N) RemoveDeadNode(NegX); return N; } break; } case ISD::FMA: case ISD::FMAD: { if (!Options.NoSignedZerosFPMath && !Flags.hasNoSignedZeros()) break; SDValue X = Op.getOperand(0), Y = Op.getOperand(1), Z = Op.getOperand(2); NegatibleCost CostZ = NegatibleCost::Expensive; SDValue NegZ = getNegatedExpression(Z, DAG, LegalOps, OptForSize, CostZ, Depth); // Give up if fail to negate the Z. if (!NegZ) break; // Prevent this node from being deleted by the next two calls. Handles.emplace_back(NegZ); // fold (fneg (fma X, Y, Z)) -> (fma (fneg X), Y, (fneg Z)) NegatibleCost CostX = NegatibleCost::Expensive; SDValue NegX = getNegatedExpression(X, DAG, LegalOps, OptForSize, CostX, Depth); // Prevent this node from being deleted by the next call. if (NegX) Handles.emplace_back(NegX); // fold (fneg (fma X, Y, Z)) -> (fma X, (fneg Y), (fneg Z)) NegatibleCost CostY = NegatibleCost::Expensive; SDValue NegY = getNegatedExpression(Y, DAG, LegalOps, OptForSize, CostY, Depth); // We're done with the handles. Handles.clear(); // Negate the X if its cost is less or equal than Y. if (NegX && (CostX <= CostY)) { Cost = std::min(CostX, CostZ); SDValue N = DAG.getNode(Opcode, DL, VT, NegX, Y, NegZ, Flags); if (NegY != N) RemoveDeadNode(NegY); return N; } // Negate the Y if it is not expensive. if (NegY) { Cost = std::min(CostY, CostZ); SDValue N = DAG.getNode(Opcode, DL, VT, X, NegY, NegZ, Flags); if (NegX != N) RemoveDeadNode(NegX); return N; } break; } case ISD::FP_EXTEND: case ISD::FSIN: if (SDValue NegV = getNegatedExpression(Op.getOperand(0), DAG, LegalOps, OptForSize, Cost, Depth)) return DAG.getNode(Opcode, DL, VT, NegV); break; case ISD::FP_ROUND: if (SDValue NegV = getNegatedExpression(Op.getOperand(0), DAG, LegalOps, OptForSize, Cost, Depth)) return DAG.getNode(ISD::FP_ROUND, DL, VT, NegV, Op.getOperand(1)); break; case ISD::SELECT: case ISD::VSELECT: { // fold (fneg (select C, LHS, RHS)) -> (select C, (fneg LHS), (fneg RHS)) // iff at least one cost is cheaper and the other is neutral/cheaper SDValue LHS = Op.getOperand(1); NegatibleCost CostLHS = NegatibleCost::Expensive; SDValue NegLHS = getNegatedExpression(LHS, DAG, LegalOps, OptForSize, CostLHS, Depth); if (!NegLHS || CostLHS > NegatibleCost::Neutral) { RemoveDeadNode(NegLHS); break; } // Prevent this node from being deleted by the next call. Handles.emplace_back(NegLHS); SDValue RHS = Op.getOperand(2); NegatibleCost CostRHS = NegatibleCost::Expensive; SDValue NegRHS = getNegatedExpression(RHS, DAG, LegalOps, OptForSize, CostRHS, Depth); // We're done with the handles. Handles.clear(); if (!NegRHS || CostRHS > NegatibleCost::Neutral || (CostLHS != NegatibleCost::Cheaper && CostRHS != NegatibleCost::Cheaper)) { RemoveDeadNode(NegLHS); RemoveDeadNode(NegRHS); break; } Cost = std::min(CostLHS, CostRHS); return DAG.getSelect(DL, VT, Op.getOperand(0), NegLHS, NegRHS); } } return SDValue(); } //===----------------------------------------------------------------------===// // Legalization Utilities //===----------------------------------------------------------------------===// bool TargetLowering::expandMUL_LOHI(unsigned Opcode, EVT VT, const SDLoc &dl, SDValue LHS, SDValue RHS, SmallVectorImpl &Result, EVT HiLoVT, SelectionDAG &DAG, MulExpansionKind Kind, SDValue LL, SDValue LH, SDValue RL, SDValue RH) const { assert(Opcode == ISD::MUL || Opcode == ISD::UMUL_LOHI || Opcode == ISD::SMUL_LOHI); bool HasMULHS = (Kind == MulExpansionKind::Always) || isOperationLegalOrCustom(ISD::MULHS, HiLoVT); bool HasMULHU = (Kind == MulExpansionKind::Always) || isOperationLegalOrCustom(ISD::MULHU, HiLoVT); bool HasSMUL_LOHI = (Kind == MulExpansionKind::Always) || isOperationLegalOrCustom(ISD::SMUL_LOHI, HiLoVT); bool HasUMUL_LOHI = (Kind == MulExpansionKind::Always) || isOperationLegalOrCustom(ISD::UMUL_LOHI, HiLoVT); if (!HasMULHU && !HasMULHS && !HasUMUL_LOHI && !HasSMUL_LOHI) return false; unsigned OuterBitSize = VT.getScalarSizeInBits(); unsigned InnerBitSize = HiLoVT.getScalarSizeInBits(); // LL, LH, RL, and RH must be either all NULL or all set to a value. assert((LL.getNode() && LH.getNode() && RL.getNode() && RH.getNode()) || (!LL.getNode() && !LH.getNode() && !RL.getNode() && !RH.getNode())); SDVTList VTs = DAG.getVTList(HiLoVT, HiLoVT); auto MakeMUL_LOHI = [&](SDValue L, SDValue R, SDValue &Lo, SDValue &Hi, bool Signed) -> bool { if ((Signed && HasSMUL_LOHI) || (!Signed && HasUMUL_LOHI)) { Lo = DAG.getNode(Signed ? ISD::SMUL_LOHI : ISD::UMUL_LOHI, dl, VTs, L, R); Hi = SDValue(Lo.getNode(), 1); return true; } if ((Signed && HasMULHS) || (!Signed && HasMULHU)) { Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, L, R); Hi = DAG.getNode(Signed ? ISD::MULHS : ISD::MULHU, dl, HiLoVT, L, R); return true; } return false; }; SDValue Lo, Hi; if (!LL.getNode() && !RL.getNode() && isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) { LL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, LHS); RL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, RHS); } if (!LL.getNode()) return false; APInt HighMask = APInt::getHighBitsSet(OuterBitSize, InnerBitSize); if (DAG.MaskedValueIsZero(LHS, HighMask) && DAG.MaskedValueIsZero(RHS, HighMask)) { // The inputs are both zero-extended. if (MakeMUL_LOHI(LL, RL, Lo, Hi, false)) { Result.push_back(Lo); Result.push_back(Hi); if (Opcode != ISD::MUL) { SDValue Zero = DAG.getConstant(0, dl, HiLoVT); Result.push_back(Zero); Result.push_back(Zero); } return true; } } if (!VT.isVector() && Opcode == ISD::MUL && DAG.ComputeMaxSignificantBits(LHS) <= InnerBitSize && DAG.ComputeMaxSignificantBits(RHS) <= InnerBitSize) { // The input values are both sign-extended. // TODO non-MUL case? if (MakeMUL_LOHI(LL, RL, Lo, Hi, true)) { Result.push_back(Lo); Result.push_back(Hi); return true; } } unsigned ShiftAmount = OuterBitSize - InnerBitSize; SDValue Shift = DAG.getShiftAmountConstant(ShiftAmount, VT, dl); if (!LH.getNode() && !RH.getNode() && isOperationLegalOrCustom(ISD::SRL, VT) && isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) { LH = DAG.getNode(ISD::SRL, dl, VT, LHS, Shift); LH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, LH); RH = DAG.getNode(ISD::SRL, dl, VT, RHS, Shift); RH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, RH); } if (!LH.getNode()) return false; if (!MakeMUL_LOHI(LL, RL, Lo, Hi, false)) return false; Result.push_back(Lo); if (Opcode == ISD::MUL) { RH = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RH); LH = DAG.getNode(ISD::MUL, dl, HiLoVT, LH, RL); Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, RH); Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, LH); Result.push_back(Hi); return true; } // Compute the full width result. auto Merge = [&](SDValue Lo, SDValue Hi) -> SDValue { Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo); Hi = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Hi); Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift); return DAG.getNode(ISD::OR, dl, VT, Lo, Hi); }; SDValue Next = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Hi); if (!MakeMUL_LOHI(LL, RH, Lo, Hi, false)) return false; // This is effectively the add part of a multiply-add of half-sized operands, // so it cannot overflow. Next = DAG.getNode(ISD::ADD, dl, VT, Next, Merge(Lo, Hi)); if (!MakeMUL_LOHI(LH, RL, Lo, Hi, false)) return false; SDValue Zero = DAG.getConstant(0, dl, HiLoVT); EVT BoolType = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); bool UseGlue = (isOperationLegalOrCustom(ISD::ADDC, VT) && isOperationLegalOrCustom(ISD::ADDE, VT)); if (UseGlue) Next = DAG.getNode(ISD::ADDC, dl, DAG.getVTList(VT, MVT::Glue), Next, Merge(Lo, Hi)); else Next = DAG.getNode(ISD::UADDO_CARRY, dl, DAG.getVTList(VT, BoolType), Next, Merge(Lo, Hi), DAG.getConstant(0, dl, BoolType)); SDValue Carry = Next.getValue(1); Result.push_back(DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, Next)); Next = DAG.getNode(ISD::SRL, dl, VT, Next, Shift); if (!MakeMUL_LOHI(LH, RH, Lo, Hi, Opcode == ISD::SMUL_LOHI)) return false; if (UseGlue) Hi = DAG.getNode(ISD::ADDE, dl, DAG.getVTList(HiLoVT, MVT::Glue), Hi, Zero, Carry); else Hi = DAG.getNode(ISD::UADDO_CARRY, dl, DAG.getVTList(HiLoVT, BoolType), Hi, Zero, Carry); Next = DAG.getNode(ISD::ADD, dl, VT, Next, Merge(Lo, Hi)); if (Opcode == ISD::SMUL_LOHI) { SDValue NextSub = DAG.getNode(ISD::SUB, dl, VT, Next, DAG.getNode(ISD::ZERO_EXTEND, dl, VT, RL)); Next = DAG.getSelectCC(dl, LH, Zero, NextSub, Next, ISD::SETLT); NextSub = DAG.getNode(ISD::SUB, dl, VT, Next, DAG.getNode(ISD::ZERO_EXTEND, dl, VT, LL)); Next = DAG.getSelectCC(dl, RH, Zero, NextSub, Next, ISD::SETLT); } Result.push_back(DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, Next)); Next = DAG.getNode(ISD::SRL, dl, VT, Next, Shift); Result.push_back(DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, Next)); return true; } bool TargetLowering::expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT, SelectionDAG &DAG, MulExpansionKind Kind, SDValue LL, SDValue LH, SDValue RL, SDValue RH) const { SmallVector Result; bool Ok = expandMUL_LOHI(N->getOpcode(), N->getValueType(0), SDLoc(N), N->getOperand(0), N->getOperand(1), Result, HiLoVT, DAG, Kind, LL, LH, RL, RH); if (Ok) { assert(Result.size() == 2); Lo = Result[0]; Hi = Result[1]; } return Ok; } // Optimize unsigned division or remainder by constants for types twice as large // as a legal VT. // // If (1 << (BitWidth / 2)) % Constant == 1, then the remainder // can be computed // as: // Sum += __builtin_uadd_overflow(Lo, High, &Sum); // Remainder = Sum % Constant // This is based on "Remainder by Summing Digits" from Hacker's Delight. // // For division, we can compute the remainder using the algorithm described // above, subtract it from the dividend to get an exact multiple of Constant. // Then multiply that exact multiply by the multiplicative inverse modulo // (1 << (BitWidth / 2)) to get the quotient. // If Constant is even, we can shift right the dividend and the divisor by the // number of trailing zeros in Constant before applying the remainder algorithm. // If we're after the quotient, we can subtract this value from the shifted // dividend and multiply by the multiplicative inverse of the shifted divisor. // If we want the remainder, we shift the value left by the number of trailing // zeros and add the bits that were shifted out of the dividend. bool TargetLowering::expandDIVREMByConstant(SDNode *N, SmallVectorImpl &Result, EVT HiLoVT, SelectionDAG &DAG, SDValue LL, SDValue LH) const { unsigned Opcode = N->getOpcode(); EVT VT = N->getValueType(0); // TODO: Support signed division/remainder. if (Opcode == ISD::SREM || Opcode == ISD::SDIV || Opcode == ISD::SDIVREM) return false; assert( (Opcode == ISD::UREM || Opcode == ISD::UDIV || Opcode == ISD::UDIVREM) && "Unexpected opcode"); auto *CN = dyn_cast(N->getOperand(1)); if (!CN) return false; APInt Divisor = CN->getAPIntValue(); unsigned BitWidth = Divisor.getBitWidth(); unsigned HBitWidth = BitWidth / 2; assert(VT.getScalarSizeInBits() == BitWidth && HiLoVT.getScalarSizeInBits() == HBitWidth && "Unexpected VTs"); // Divisor needs to less than (1 << HBitWidth). APInt HalfMaxPlus1 = APInt::getOneBitSet(BitWidth, HBitWidth); if (Divisor.uge(HalfMaxPlus1)) return false; // We depend on the UREM by constant optimization in DAGCombiner that requires // high multiply. if (!isOperationLegalOrCustom(ISD::MULHU, HiLoVT) && !isOperationLegalOrCustom(ISD::UMUL_LOHI, HiLoVT)) return false; // Don't expand if optimizing for size. if (DAG.shouldOptForSize()) return false; // Early out for 0 or 1 divisors. if (Divisor.ule(1)) return false; // If the divisor is even, shift it until it becomes odd. unsigned TrailingZeros = 0; if (!Divisor[0]) { TrailingZeros = Divisor.countr_zero(); Divisor.lshrInPlace(TrailingZeros); } SDLoc dl(N); SDValue Sum; SDValue PartialRem; // If (1 << HBitWidth) % divisor == 1, we can add the two halves together and // then add in the carry. // TODO: If we can't split it in half, we might be able to split into 3 or // more pieces using a smaller bit width. if (HalfMaxPlus1.urem(Divisor).isOne()) { assert(!LL == !LH && "Expected both input halves or no input halves!"); if (!LL) std::tie(LL, LH) = DAG.SplitScalar(N->getOperand(0), dl, HiLoVT, HiLoVT); // Shift the input by the number of TrailingZeros in the divisor. The // shifted out bits will be added to the remainder later. if (TrailingZeros) { // Save the shifted off bits if we need the remainder. if (Opcode != ISD::UDIV) { APInt Mask = APInt::getLowBitsSet(HBitWidth, TrailingZeros); PartialRem = DAG.getNode(ISD::AND, dl, HiLoVT, LL, DAG.getConstant(Mask, dl, HiLoVT)); } LL = DAG.getNode( ISD::OR, dl, HiLoVT, DAG.getNode(ISD::SRL, dl, HiLoVT, LL, DAG.getShiftAmountConstant(TrailingZeros, HiLoVT, dl)), DAG.getNode(ISD::SHL, dl, HiLoVT, LH, DAG.getShiftAmountConstant(HBitWidth - TrailingZeros, HiLoVT, dl))); LH = DAG.getNode(ISD::SRL, dl, HiLoVT, LH, DAG.getShiftAmountConstant(TrailingZeros, HiLoVT, dl)); } // Use uaddo_carry if we can, otherwise use a compare to detect overflow. EVT SetCCType = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), HiLoVT); if (isOperationLegalOrCustom(ISD::UADDO_CARRY, HiLoVT)) { SDVTList VTList = DAG.getVTList(HiLoVT, SetCCType); Sum = DAG.getNode(ISD::UADDO, dl, VTList, LL, LH); Sum = DAG.getNode(ISD::UADDO_CARRY, dl, VTList, Sum, DAG.getConstant(0, dl, HiLoVT), Sum.getValue(1)); } else { Sum = DAG.getNode(ISD::ADD, dl, HiLoVT, LL, LH); SDValue Carry = DAG.getSetCC(dl, SetCCType, Sum, LL, ISD::SETULT); // If the boolean for the target is 0 or 1, we can add the setcc result // directly. if (getBooleanContents(HiLoVT) == TargetLoweringBase::ZeroOrOneBooleanContent) Carry = DAG.getZExtOrTrunc(Carry, dl, HiLoVT); else Carry = DAG.getSelect(dl, HiLoVT, Carry, DAG.getConstant(1, dl, HiLoVT), DAG.getConstant(0, dl, HiLoVT)); Sum = DAG.getNode(ISD::ADD, dl, HiLoVT, Sum, Carry); } } // If we didn't find a sum, we can't do the expansion. if (!Sum) return false; // Perform a HiLoVT urem on the Sum using truncated divisor. SDValue RemL = DAG.getNode(ISD::UREM, dl, HiLoVT, Sum, DAG.getConstant(Divisor.trunc(HBitWidth), dl, HiLoVT)); SDValue RemH = DAG.getConstant(0, dl, HiLoVT); if (Opcode != ISD::UREM) { // Subtract the remainder from the shifted dividend. SDValue Dividend = DAG.getNode(ISD::BUILD_PAIR, dl, VT, LL, LH); SDValue Rem = DAG.getNode(ISD::BUILD_PAIR, dl, VT, RemL, RemH); Dividend = DAG.getNode(ISD::SUB, dl, VT, Dividend, Rem); // Multiply by the multiplicative inverse of the divisor modulo // (1 << BitWidth). APInt MulFactor = Divisor.multiplicativeInverse(); SDValue Quotient = DAG.getNode(ISD::MUL, dl, VT, Dividend, DAG.getConstant(MulFactor, dl, VT)); // Split the quotient into low and high parts. SDValue QuotL, QuotH; std::tie(QuotL, QuotH) = DAG.SplitScalar(Quotient, dl, HiLoVT, HiLoVT); Result.push_back(QuotL); Result.push_back(QuotH); } if (Opcode != ISD::UDIV) { // If we shifted the input, shift the remainder left and add the bits we // shifted off the input. if (TrailingZeros) { APInt Mask = APInt::getLowBitsSet(HBitWidth, TrailingZeros); RemL = DAG.getNode(ISD::SHL, dl, HiLoVT, RemL, DAG.getShiftAmountConstant(TrailingZeros, HiLoVT, dl)); RemL = DAG.getNode(ISD::ADD, dl, HiLoVT, RemL, PartialRem); } Result.push_back(RemL); Result.push_back(DAG.getConstant(0, dl, HiLoVT)); } return true; } // Check that (every element of) Z is undef or not an exact multiple of BW. static bool isNonZeroModBitWidthOrUndef(SDValue Z, unsigned BW) { return ISD::matchUnaryPredicate( Z, [=](ConstantSDNode *C) { return !C || C->getAPIntValue().urem(BW) != 0; }, true); } static SDValue expandVPFunnelShift(SDNode *Node, SelectionDAG &DAG) { EVT VT = Node->getValueType(0); SDValue ShX, ShY; SDValue ShAmt, InvShAmt; SDValue X = Node->getOperand(0); SDValue Y = Node->getOperand(1); SDValue Z = Node->getOperand(2); SDValue Mask = Node->getOperand(3); SDValue VL = Node->getOperand(4); unsigned BW = VT.getScalarSizeInBits(); bool IsFSHL = Node->getOpcode() == ISD::VP_FSHL; SDLoc DL(SDValue(Node, 0)); EVT ShVT = Z.getValueType(); if (isNonZeroModBitWidthOrUndef(Z, BW)) { // fshl: X << C | Y >> (BW - C) // fshr: X << (BW - C) | Y >> C // where C = Z % BW is not zero SDValue BitWidthC = DAG.getConstant(BW, DL, ShVT); ShAmt = DAG.getNode(ISD::VP_UREM, DL, ShVT, Z, BitWidthC, Mask, VL); InvShAmt = DAG.getNode(ISD::VP_SUB, DL, ShVT, BitWidthC, ShAmt, Mask, VL); ShX = DAG.getNode(ISD::VP_SHL, DL, VT, X, IsFSHL ? ShAmt : InvShAmt, Mask, VL); ShY = DAG.getNode(ISD::VP_SRL, DL, VT, Y, IsFSHL ? InvShAmt : ShAmt, Mask, VL); } else { // fshl: X << (Z % BW) | Y >> 1 >> (BW - 1 - (Z % BW)) // fshr: X << 1 << (BW - 1 - (Z % BW)) | Y >> (Z % BW) SDValue BitMask = DAG.getConstant(BW - 1, DL, ShVT); if (isPowerOf2_32(BW)) { // Z % BW -> Z & (BW - 1) ShAmt = DAG.getNode(ISD::VP_AND, DL, ShVT, Z, BitMask, Mask, VL); // (BW - 1) - (Z % BW) -> ~Z & (BW - 1) SDValue NotZ = DAG.getNode(ISD::VP_XOR, DL, ShVT, Z, DAG.getAllOnesConstant(DL, ShVT), Mask, VL); InvShAmt = DAG.getNode(ISD::VP_AND, DL, ShVT, NotZ, BitMask, Mask, VL); } else { SDValue BitWidthC = DAG.getConstant(BW, DL, ShVT); ShAmt = DAG.getNode(ISD::VP_UREM, DL, ShVT, Z, BitWidthC, Mask, VL); InvShAmt = DAG.getNode(ISD::VP_SUB, DL, ShVT, BitMask, ShAmt, Mask, VL); } SDValue One = DAG.getConstant(1, DL, ShVT); if (IsFSHL) { ShX = DAG.getNode(ISD::VP_SHL, DL, VT, X, ShAmt, Mask, VL); SDValue ShY1 = DAG.getNode(ISD::VP_SRL, DL, VT, Y, One, Mask, VL); ShY = DAG.getNode(ISD::VP_SRL, DL, VT, ShY1, InvShAmt, Mask, VL); } else { SDValue ShX1 = DAG.getNode(ISD::VP_SHL, DL, VT, X, One, Mask, VL); ShX = DAG.getNode(ISD::VP_SHL, DL, VT, ShX1, InvShAmt, Mask, VL); ShY = DAG.getNode(ISD::VP_SRL, DL, VT, Y, ShAmt, Mask, VL); } } return DAG.getNode(ISD::VP_OR, DL, VT, ShX, ShY, Mask, VL); } SDValue TargetLowering::expandFunnelShift(SDNode *Node, SelectionDAG &DAG) const { if (Node->isVPOpcode()) return expandVPFunnelShift(Node, DAG); EVT VT = Node->getValueType(0); if (VT.isVector() && (!isOperationLegalOrCustom(ISD::SHL, VT) || !isOperationLegalOrCustom(ISD::SRL, VT) || !isOperationLegalOrCustom(ISD::SUB, VT) || !isOperationLegalOrCustomOrPromote(ISD::OR, VT))) return SDValue(); SDValue X = Node->getOperand(0); SDValue Y = Node->getOperand(1); SDValue Z = Node->getOperand(2); unsigned BW = VT.getScalarSizeInBits(); bool IsFSHL = Node->getOpcode() == ISD::FSHL; SDLoc DL(SDValue(Node, 0)); EVT ShVT = Z.getValueType(); // If a funnel shift in the other direction is more supported, use it. unsigned RevOpcode = IsFSHL ? ISD::FSHR : ISD::FSHL; if (!isOperationLegalOrCustom(Node->getOpcode(), VT) && isOperationLegalOrCustom(RevOpcode, VT) && isPowerOf2_32(BW)) { if (isNonZeroModBitWidthOrUndef(Z, BW)) { // fshl X, Y, Z -> fshr X, Y, -Z // fshr X, Y, Z -> fshl X, Y, -Z SDValue Zero = DAG.getConstant(0, DL, ShVT); Z = DAG.getNode(ISD::SUB, DL, VT, Zero, Z); } else { // fshl X, Y, Z -> fshr (srl X, 1), (fshr X, Y, 1), ~Z // fshr X, Y, Z -> fshl (fshl X, Y, 1), (shl Y, 1), ~Z SDValue One = DAG.getConstant(1, DL, ShVT); if (IsFSHL) { Y = DAG.getNode(RevOpcode, DL, VT, X, Y, One); X = DAG.getNode(ISD::SRL, DL, VT, X, One); } else { X = DAG.getNode(RevOpcode, DL, VT, X, Y, One); Y = DAG.getNode(ISD::SHL, DL, VT, Y, One); } Z = DAG.getNOT(DL, Z, ShVT); } return DAG.getNode(RevOpcode, DL, VT, X, Y, Z); } SDValue ShX, ShY; SDValue ShAmt, InvShAmt; if (isNonZeroModBitWidthOrUndef(Z, BW)) { // fshl: X << C | Y >> (BW - C) // fshr: X << (BW - C) | Y >> C // where C = Z % BW is not zero SDValue BitWidthC = DAG.getConstant(BW, DL, ShVT); ShAmt = DAG.getNode(ISD::UREM, DL, ShVT, Z, BitWidthC); InvShAmt = DAG.getNode(ISD::SUB, DL, ShVT, BitWidthC, ShAmt); ShX = DAG.getNode(ISD::SHL, DL, VT, X, IsFSHL ? ShAmt : InvShAmt); ShY = DAG.getNode(ISD::SRL, DL, VT, Y, IsFSHL ? InvShAmt : ShAmt); } else { // fshl: X << (Z % BW) | Y >> 1 >> (BW - 1 - (Z % BW)) // fshr: X << 1 << (BW - 1 - (Z % BW)) | Y >> (Z % BW) SDValue Mask = DAG.getConstant(BW - 1, DL, ShVT); if (isPowerOf2_32(BW)) { // Z % BW -> Z & (BW - 1) ShAmt = DAG.getNode(ISD::AND, DL, ShVT, Z, Mask); // (BW - 1) - (Z % BW) -> ~Z & (BW - 1) InvShAmt = DAG.getNode(ISD::AND, DL, ShVT, DAG.getNOT(DL, Z, ShVT), Mask); } else { SDValue BitWidthC = DAG.getConstant(BW, DL, ShVT); ShAmt = DAG.getNode(ISD::UREM, DL, ShVT, Z, BitWidthC); InvShAmt = DAG.getNode(ISD::SUB, DL, ShVT, Mask, ShAmt); } SDValue One = DAG.getConstant(1, DL, ShVT); if (IsFSHL) { ShX = DAG.getNode(ISD::SHL, DL, VT, X, ShAmt); SDValue ShY1 = DAG.getNode(ISD::SRL, DL, VT, Y, One); ShY = DAG.getNode(ISD::SRL, DL, VT, ShY1, InvShAmt); } else { SDValue ShX1 = DAG.getNode(ISD::SHL, DL, VT, X, One); ShX = DAG.getNode(ISD::SHL, DL, VT, ShX1, InvShAmt); ShY = DAG.getNode(ISD::SRL, DL, VT, Y, ShAmt); } } return DAG.getNode(ISD::OR, DL, VT, ShX, ShY); } // TODO: Merge with expandFunnelShift. SDValue TargetLowering::expandROT(SDNode *Node, bool AllowVectorOps, SelectionDAG &DAG) const { EVT VT = Node->getValueType(0); unsigned EltSizeInBits = VT.getScalarSizeInBits(); bool IsLeft = Node->getOpcode() == ISD::ROTL; SDValue Op0 = Node->getOperand(0); SDValue Op1 = Node->getOperand(1); SDLoc DL(SDValue(Node, 0)); EVT ShVT = Op1.getValueType(); SDValue Zero = DAG.getConstant(0, DL, ShVT); // If a rotate in the other direction is more supported, use it. unsigned RevRot = IsLeft ? ISD::ROTR : ISD::ROTL; if (!isOperationLegalOrCustom(Node->getOpcode(), VT) && isOperationLegalOrCustom(RevRot, VT) && isPowerOf2_32(EltSizeInBits)) { SDValue Sub = DAG.getNode(ISD::SUB, DL, ShVT, Zero, Op1); return DAG.getNode(RevRot, DL, VT, Op0, Sub); } if (!AllowVectorOps && VT.isVector() && (!isOperationLegalOrCustom(ISD::SHL, VT) || !isOperationLegalOrCustom(ISD::SRL, VT) || !isOperationLegalOrCustom(ISD::SUB, VT) || !isOperationLegalOrCustomOrPromote(ISD::OR, VT) || !isOperationLegalOrCustomOrPromote(ISD::AND, VT))) return SDValue(); unsigned ShOpc = IsLeft ? ISD::SHL : ISD::SRL; unsigned HsOpc = IsLeft ? ISD::SRL : ISD::SHL; SDValue BitWidthMinusOneC = DAG.getConstant(EltSizeInBits - 1, DL, ShVT); SDValue ShVal; SDValue HsVal; if (isPowerOf2_32(EltSizeInBits)) { // (rotl x, c) -> x << (c & (w - 1)) | x >> (-c & (w - 1)) // (rotr x, c) -> x >> (c & (w - 1)) | x << (-c & (w - 1)) SDValue NegOp1 = DAG.getNode(ISD::SUB, DL, ShVT, Zero, Op1); SDValue ShAmt = DAG.getNode(ISD::AND, DL, ShVT, Op1, BitWidthMinusOneC); ShVal = DAG.getNode(ShOpc, DL, VT, Op0, ShAmt); SDValue HsAmt = DAG.getNode(ISD::AND, DL, ShVT, NegOp1, BitWidthMinusOneC); HsVal = DAG.getNode(HsOpc, DL, VT, Op0, HsAmt); } else { // (rotl x, c) -> x << (c % w) | x >> 1 >> (w - 1 - (c % w)) // (rotr x, c) -> x >> (c % w) | x << 1 << (w - 1 - (c % w)) SDValue BitWidthC = DAG.getConstant(EltSizeInBits, DL, ShVT); SDValue ShAmt = DAG.getNode(ISD::UREM, DL, ShVT, Op1, BitWidthC); ShVal = DAG.getNode(ShOpc, DL, VT, Op0, ShAmt); SDValue HsAmt = DAG.getNode(ISD::SUB, DL, ShVT, BitWidthMinusOneC, ShAmt); SDValue One = DAG.getConstant(1, DL, ShVT); HsVal = DAG.getNode(HsOpc, DL, VT, DAG.getNode(HsOpc, DL, VT, Op0, One), HsAmt); } return DAG.getNode(ISD::OR, DL, VT, ShVal, HsVal); } void TargetLowering::expandShiftParts(SDNode *Node, SDValue &Lo, SDValue &Hi, SelectionDAG &DAG) const { assert(Node->getNumOperands() == 3 && "Not a double-shift!"); EVT VT = Node->getValueType(0); unsigned VTBits = VT.getScalarSizeInBits(); assert(isPowerOf2_32(VTBits) && "Power-of-two integer type expected"); bool IsSHL = Node->getOpcode() == ISD::SHL_PARTS; bool IsSRA = Node->getOpcode() == ISD::SRA_PARTS; SDValue ShOpLo = Node->getOperand(0); SDValue ShOpHi = Node->getOperand(1); SDValue ShAmt = Node->getOperand(2); EVT ShAmtVT = ShAmt.getValueType(); EVT ShAmtCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), ShAmtVT); SDLoc dl(Node); // ISD::FSHL and ISD::FSHR have defined overflow behavior but ISD::SHL and // ISD::SRA/L nodes haven't. Insert an AND to be safe, it's usually optimized // away during isel. SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, ShAmtVT, ShAmt, DAG.getConstant(VTBits - 1, dl, ShAmtVT)); SDValue Tmp1 = IsSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi, DAG.getConstant(VTBits - 1, dl, ShAmtVT)) : DAG.getConstant(0, dl, VT); SDValue Tmp2, Tmp3; if (IsSHL) { Tmp2 = DAG.getNode(ISD::FSHL, dl, VT, ShOpHi, ShOpLo, ShAmt); Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt); } else { Tmp2 = DAG.getNode(ISD::FSHR, dl, VT, ShOpHi, ShOpLo, ShAmt); Tmp3 = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt); } // If the shift amount is larger or equal than the width of a part we don't // use the result from the FSHL/FSHR. Insert a test and select the appropriate // values for large shift amounts. SDValue AndNode = DAG.getNode(ISD::AND, dl, ShAmtVT, ShAmt, DAG.getConstant(VTBits, dl, ShAmtVT)); SDValue Cond = DAG.getSetCC(dl, ShAmtCCVT, AndNode, DAG.getConstant(0, dl, ShAmtVT), ISD::SETNE); if (IsSHL) { Hi = DAG.getNode(ISD::SELECT, dl, VT, Cond, Tmp3, Tmp2); Lo = DAG.getNode(ISD::SELECT, dl, VT, Cond, Tmp1, Tmp3); } else { Lo = DAG.getNode(ISD::SELECT, dl, VT, Cond, Tmp3, Tmp2); Hi = DAG.getNode(ISD::SELECT, dl, VT, Cond, Tmp1, Tmp3); } } bool TargetLowering::expandFP_TO_SINT(SDNode *Node, SDValue &Result, SelectionDAG &DAG) const { unsigned OpNo = Node->isStrictFPOpcode() ? 1 : 0; SDValue Src = Node->getOperand(OpNo); EVT SrcVT = Src.getValueType(); EVT DstVT = Node->getValueType(0); SDLoc dl(SDValue(Node, 0)); // FIXME: Only f32 to i64 conversions are supported. if (SrcVT != MVT::f32 || DstVT != MVT::i64) return false; if (Node->isStrictFPOpcode()) // When a NaN is converted to an integer a trap is allowed. We can't // use this expansion here because it would eliminate that trap. Other // traps are also allowed and cannot be eliminated. See // IEEE 754-2008 sec 5.8. return false; // Expand f32 -> i64 conversion // This algorithm comes from compiler-rt's implementation of fixsfdi: // https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/builtins/fixsfdi.c unsigned SrcEltBits = SrcVT.getScalarSizeInBits(); EVT IntVT = SrcVT.changeTypeToInteger(); EVT IntShVT = getShiftAmountTy(IntVT, DAG.getDataLayout()); SDValue ExponentMask = DAG.getConstant(0x7F800000, dl, IntVT); SDValue ExponentLoBit = DAG.getConstant(23, dl, IntVT); SDValue Bias = DAG.getConstant(127, dl, IntVT); SDValue SignMask = DAG.getConstant(APInt::getSignMask(SrcEltBits), dl, IntVT); SDValue SignLowBit = DAG.getConstant(SrcEltBits - 1, dl, IntVT); SDValue MantissaMask = DAG.getConstant(0x007FFFFF, dl, IntVT); SDValue Bits = DAG.getNode(ISD::BITCAST, dl, IntVT, Src); SDValue ExponentBits = DAG.getNode( ISD::SRL, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, ExponentMask), DAG.getZExtOrTrunc(ExponentLoBit, dl, IntShVT)); SDValue Exponent = DAG.getNode(ISD::SUB, dl, IntVT, ExponentBits, Bias); SDValue Sign = DAG.getNode(ISD::SRA, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, SignMask), DAG.getZExtOrTrunc(SignLowBit, dl, IntShVT)); Sign = DAG.getSExtOrTrunc(Sign, dl, DstVT); SDValue R = DAG.getNode(ISD::OR, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, MantissaMask), DAG.getConstant(0x00800000, dl, IntVT)); R = DAG.getZExtOrTrunc(R, dl, DstVT); R = DAG.getSelectCC( dl, Exponent, ExponentLoBit, DAG.getNode(ISD::SHL, dl, DstVT, R, DAG.getZExtOrTrunc( DAG.getNode(ISD::SUB, dl, IntVT, Exponent, ExponentLoBit), dl, IntShVT)), DAG.getNode(ISD::SRL, dl, DstVT, R, DAG.getZExtOrTrunc( DAG.getNode(ISD::SUB, dl, IntVT, ExponentLoBit, Exponent), dl, IntShVT)), ISD::SETGT); SDValue Ret = DAG.getNode(ISD::SUB, dl, DstVT, DAG.getNode(ISD::XOR, dl, DstVT, R, Sign), Sign); Result = DAG.getSelectCC(dl, Exponent, DAG.getConstant(0, dl, IntVT), DAG.getConstant(0, dl, DstVT), Ret, ISD::SETLT); return true; } bool TargetLowering::expandFP_TO_UINT(SDNode *Node, SDValue &Result, SDValue &Chain, SelectionDAG &DAG) const { SDLoc dl(SDValue(Node, 0)); unsigned OpNo = Node->isStrictFPOpcode() ? 1 : 0; SDValue Src = Node->getOperand(OpNo); EVT SrcVT = Src.getValueType(); EVT DstVT = Node->getValueType(0); EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT); EVT DstSetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), DstVT); // Only expand vector types if we have the appropriate vector bit operations. unsigned SIntOpcode = Node->isStrictFPOpcode() ? ISD::STRICT_FP_TO_SINT : ISD::FP_TO_SINT; if (DstVT.isVector() && (!isOperationLegalOrCustom(SIntOpcode, DstVT) || !isOperationLegalOrCustomOrPromote(ISD::XOR, SrcVT))) return false; // If the maximum float value is smaller then the signed integer range, // the destination signmask can't be represented by the float, so we can // just use FP_TO_SINT directly. const fltSemantics &APFSem = DAG.EVTToAPFloatSemantics(SrcVT); APFloat APF(APFSem, APInt::getZero(SrcVT.getScalarSizeInBits())); APInt SignMask = APInt::getSignMask(DstVT.getScalarSizeInBits()); if (APFloat::opOverflow & APF.convertFromAPInt(SignMask, false, APFloat::rmNearestTiesToEven)) { if (Node->isStrictFPOpcode()) { Result = DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, { DstVT, MVT::Other }, { Node->getOperand(0), Src }); Chain = Result.getValue(1); } else Result = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, Src); return true; } // Don't expand it if there isn't cheap fsub instruction. if (!isOperationLegalOrCustom( Node->isStrictFPOpcode() ? ISD::STRICT_FSUB : ISD::FSUB, SrcVT)) return false; SDValue Cst = DAG.getConstantFP(APF, dl, SrcVT); SDValue Sel; if (Node->isStrictFPOpcode()) { Sel = DAG.getSetCC(dl, SetCCVT, Src, Cst, ISD::SETLT, Node->getOperand(0), /*IsSignaling*/ true); Chain = Sel.getValue(1); } else { Sel = DAG.getSetCC(dl, SetCCVT, Src, Cst, ISD::SETLT); } bool Strict = Node->isStrictFPOpcode() || shouldUseStrictFP_TO_INT(SrcVT, DstVT, /*IsSigned*/ false); if (Strict) { // Expand based on maximum range of FP_TO_SINT, if the value exceeds the // signmask then offset (the result of which should be fully representable). // Sel = Src < 0x8000000000000000 // FltOfs = select Sel, 0, 0x8000000000000000 // IntOfs = select Sel, 0, 0x8000000000000000 // Result = fp_to_sint(Src - FltOfs) ^ IntOfs // TODO: Should any fast-math-flags be set for the FSUB? SDValue FltOfs = DAG.getSelect(dl, SrcVT, Sel, DAG.getConstantFP(0.0, dl, SrcVT), Cst); Sel = DAG.getBoolExtOrTrunc(Sel, dl, DstSetCCVT, DstVT); SDValue IntOfs = DAG.getSelect(dl, DstVT, Sel, DAG.getConstant(0, dl, DstVT), DAG.getConstant(SignMask, dl, DstVT)); SDValue SInt; if (Node->isStrictFPOpcode()) { SDValue Val = DAG.getNode(ISD::STRICT_FSUB, dl, { SrcVT, MVT::Other }, { Chain, Src, FltOfs }); SInt = DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, { DstVT, MVT::Other }, { Val.getValue(1), Val }); Chain = SInt.getValue(1); } else { SDValue Val = DAG.getNode(ISD::FSUB, dl, SrcVT, Src, FltOfs); SInt = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, Val); } Result = DAG.getNode(ISD::XOR, dl, DstVT, SInt, IntOfs); } else { // Expand based on maximum range of FP_TO_SINT: // True = fp_to_sint(Src) // False = 0x8000000000000000 + fp_to_sint(Src - 0x8000000000000000) // Result = select (Src < 0x8000000000000000), True, False SDValue True = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, Src); // TODO: Should any fast-math-flags be set for the FSUB? SDValue False = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, DAG.getNode(ISD::FSUB, dl, SrcVT, Src, Cst)); False = DAG.getNode(ISD::XOR, dl, DstVT, False, DAG.getConstant(SignMask, dl, DstVT)); Sel = DAG.getBoolExtOrTrunc(Sel, dl, DstSetCCVT, DstVT); Result = DAG.getSelect(dl, DstVT, Sel, True, False); } return true; } bool TargetLowering::expandUINT_TO_FP(SDNode *Node, SDValue &Result, SDValue &Chain, SelectionDAG &DAG) const { // This transform is not correct for converting 0 when rounding mode is set // to round toward negative infinity which will produce -0.0. So disable under // strictfp. if (Node->isStrictFPOpcode()) return false; SDValue Src = Node->getOperand(0); EVT SrcVT = Src.getValueType(); EVT DstVT = Node->getValueType(0); if (SrcVT.getScalarType() != MVT::i64 || DstVT.getScalarType() != MVT::f64) return false; // Only expand vector types if we have the appropriate vector bit operations. if (SrcVT.isVector() && (!isOperationLegalOrCustom(ISD::SRL, SrcVT) || !isOperationLegalOrCustom(ISD::FADD, DstVT) || !isOperationLegalOrCustom(ISD::FSUB, DstVT) || !isOperationLegalOrCustomOrPromote(ISD::OR, SrcVT) || !isOperationLegalOrCustomOrPromote(ISD::AND, SrcVT))) return false; SDLoc dl(SDValue(Node, 0)); EVT ShiftVT = getShiftAmountTy(SrcVT, DAG.getDataLayout()); // Implementation of unsigned i64 to f64 following the algorithm in // __floatundidf in compiler_rt. This implementation performs rounding // correctly in all rounding modes with the exception of converting 0 // when rounding toward negative infinity. In that case the fsub will produce // -0.0. This will be added to +0.0 and produce -0.0 which is incorrect. SDValue TwoP52 = DAG.getConstant(UINT64_C(0x4330000000000000), dl, SrcVT); SDValue TwoP84PlusTwoP52 = DAG.getConstantFP( llvm::bit_cast(UINT64_C(0x4530000000100000)), dl, DstVT); SDValue TwoP84 = DAG.getConstant(UINT64_C(0x4530000000000000), dl, SrcVT); SDValue LoMask = DAG.getConstant(UINT64_C(0x00000000FFFFFFFF), dl, SrcVT); SDValue HiShift = DAG.getConstant(32, dl, ShiftVT); SDValue Lo = DAG.getNode(ISD::AND, dl, SrcVT, Src, LoMask); SDValue Hi = DAG.getNode(ISD::SRL, dl, SrcVT, Src, HiShift); SDValue LoOr = DAG.getNode(ISD::OR, dl, SrcVT, Lo, TwoP52); SDValue HiOr = DAG.getNode(ISD::OR, dl, SrcVT, Hi, TwoP84); SDValue LoFlt = DAG.getBitcast(DstVT, LoOr); SDValue HiFlt = DAG.getBitcast(DstVT, HiOr); SDValue HiSub = DAG.getNode(ISD::FSUB, dl, DstVT, HiFlt, TwoP84PlusTwoP52); Result = DAG.getNode(ISD::FADD, dl, DstVT, LoFlt, HiSub); return true; } SDValue TargetLowering::createSelectForFMINNUM_FMAXNUM(SDNode *Node, SelectionDAG &DAG) const { unsigned Opcode = Node->getOpcode(); assert((Opcode == ISD::FMINNUM || Opcode == ISD::FMAXNUM || Opcode == ISD::STRICT_FMINNUM || Opcode == ISD::STRICT_FMAXNUM) && "Wrong opcode"); if (Node->getFlags().hasNoNaNs()) { ISD::CondCode Pred = Opcode == ISD::FMINNUM ? ISD::SETLT : ISD::SETGT; SDValue Op1 = Node->getOperand(0); SDValue Op2 = Node->getOperand(1); SDValue SelCC = DAG.getSelectCC(SDLoc(Node), Op1, Op2, Op1, Op2, Pred); // Copy FMF flags, but always set the no-signed-zeros flag // as this is implied by the FMINNUM/FMAXNUM semantics. SDNodeFlags Flags = Node->getFlags(); Flags.setNoSignedZeros(true); SelCC->setFlags(Flags); return SelCC; } return SDValue(); } SDValue TargetLowering::expandFMINNUM_FMAXNUM(SDNode *Node, SelectionDAG &DAG) const { SDLoc dl(Node); unsigned NewOp = Node->getOpcode() == ISD::FMINNUM ? ISD::FMINNUM_IEEE : ISD::FMAXNUM_IEEE; EVT VT = Node->getValueType(0); if (VT.isScalableVector()) report_fatal_error( "Expanding fminnum/fmaxnum for scalable vectors is undefined."); if (isOperationLegalOrCustom(NewOp, VT)) { SDValue Quiet0 = Node->getOperand(0); SDValue Quiet1 = Node->getOperand(1); if (!Node->getFlags().hasNoNaNs()) { // Insert canonicalizes if it's possible we need to quiet to get correct // sNaN behavior. if (!DAG.isKnownNeverSNaN(Quiet0)) { Quiet0 = DAG.getNode(ISD::FCANONICALIZE, dl, VT, Quiet0, Node->getFlags()); } if (!DAG.isKnownNeverSNaN(Quiet1)) { Quiet1 = DAG.getNode(ISD::FCANONICALIZE, dl, VT, Quiet1, Node->getFlags()); } } return DAG.getNode(NewOp, dl, VT, Quiet0, Quiet1, Node->getFlags()); } // If the target has FMINIMUM/FMAXIMUM but not FMINNUM/FMAXNUM use that // instead if there are no NaNs and there can't be an incompatible zero // compare: at least one operand isn't +/-0, or there are no signed-zeros. if ((Node->getFlags().hasNoNaNs() || (DAG.isKnownNeverNaN(Node->getOperand(0)) && DAG.isKnownNeverNaN(Node->getOperand(1)))) && (Node->getFlags().hasNoSignedZeros() || DAG.isKnownNeverZeroFloat(Node->getOperand(0)) || DAG.isKnownNeverZeroFloat(Node->getOperand(1)))) { unsigned IEEE2018Op = Node->getOpcode() == ISD::FMINNUM ? ISD::FMINIMUM : ISD::FMAXIMUM; if (isOperationLegalOrCustom(IEEE2018Op, VT)) return DAG.getNode(IEEE2018Op, dl, VT, Node->getOperand(0), Node->getOperand(1), Node->getFlags()); } if (SDValue SelCC = createSelectForFMINNUM_FMAXNUM(Node, DAG)) return SelCC; return SDValue(); } SDValue TargetLowering::expandFMINIMUM_FMAXIMUM(SDNode *N, SelectionDAG &DAG) const { SDLoc DL(N); SDValue LHS = N->getOperand(0); SDValue RHS = N->getOperand(1); unsigned Opc = N->getOpcode(); EVT VT = N->getValueType(0); EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); bool IsMax = Opc == ISD::FMAXIMUM; SDNodeFlags Flags = N->getFlags(); // First, implement comparison not propagating NaN. If no native fmin or fmax // available, use plain select with setcc instead. SDValue MinMax; unsigned CompOpcIeee = IsMax ? ISD::FMAXNUM_IEEE : ISD::FMINNUM_IEEE; unsigned CompOpc = IsMax ? ISD::FMAXNUM : ISD::FMINNUM; // FIXME: We should probably define fminnum/fmaxnum variants with correct // signed zero behavior. bool MinMaxMustRespectOrderedZero = false; if (isOperationLegalOrCustom(CompOpcIeee, VT)) { MinMax = DAG.getNode(CompOpcIeee, DL, VT, LHS, RHS, Flags); MinMaxMustRespectOrderedZero = true; } else if (isOperationLegalOrCustom(CompOpc, VT)) { MinMax = DAG.getNode(CompOpc, DL, VT, LHS, RHS, Flags); } else { if (VT.isVector() && !isOperationLegalOrCustom(ISD::VSELECT, VT)) return DAG.UnrollVectorOp(N); // NaN (if exists) will be propagated later, so orderness doesn't matter. SDValue Compare = DAG.getSetCC(DL, CCVT, LHS, RHS, IsMax ? ISD::SETGT : ISD::SETLT); MinMax = DAG.getSelect(DL, VT, Compare, LHS, RHS, Flags); } // Propagate any NaN of both operands if (!N->getFlags().hasNoNaNs() && (!DAG.isKnownNeverNaN(RHS) || !DAG.isKnownNeverNaN(LHS))) { ConstantFP *FPNaN = ConstantFP::get( *DAG.getContext(), APFloat::getNaN(DAG.EVTToAPFloatSemantics(VT))); MinMax = DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, LHS, RHS, ISD::SETUO), DAG.getConstantFP(*FPNaN, DL, VT), MinMax, Flags); } // fminimum/fmaximum requires -0.0 less than +0.0 if (!MinMaxMustRespectOrderedZero && !N->getFlags().hasNoSignedZeros() && !DAG.isKnownNeverZeroFloat(RHS) && !DAG.isKnownNeverZeroFloat(LHS)) { SDValue IsZero = DAG.getSetCC(DL, CCVT, MinMax, DAG.getConstantFP(0.0, DL, VT), ISD::SETEQ); SDValue TestZero = DAG.getTargetConstant(IsMax ? fcPosZero : fcNegZero, DL, MVT::i32); SDValue LCmp = DAG.getSelect( DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, LHS, TestZero), LHS, MinMax, Flags); SDValue RCmp = DAG.getSelect( DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, RHS, TestZero), RHS, LCmp, Flags); MinMax = DAG.getSelect(DL, VT, IsZero, RCmp, MinMax, Flags); } return MinMax; } /// Returns a true value if if this FPClassTest can be performed with an ordered /// fcmp to 0, and a false value if it's an unordered fcmp to 0. Returns /// std::nullopt if it cannot be performed as a compare with 0. static std::optional isFCmpEqualZero(FPClassTest Test, const fltSemantics &Semantics, const MachineFunction &MF) { FPClassTest OrderedMask = Test & ~fcNan; FPClassTest NanTest = Test & fcNan; bool IsOrdered = NanTest == fcNone; bool IsUnordered = NanTest == fcNan; // Skip cases that are testing for only a qnan or snan. if (!IsOrdered && !IsUnordered) return std::nullopt; if (OrderedMask == fcZero && MF.getDenormalMode(Semantics).Input == DenormalMode::IEEE) return IsOrdered; if (OrderedMask == (fcZero | fcSubnormal) && MF.getDenormalMode(Semantics).inputsAreZero()) return IsOrdered; return std::nullopt; } SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op, FPClassTest Test, SDNodeFlags Flags, const SDLoc &DL, SelectionDAG &DAG) const { EVT OperandVT = Op.getValueType(); assert(OperandVT.isFloatingPoint()); // Degenerated cases. if (Test == fcNone) return DAG.getBoolConstant(false, DL, ResultVT, OperandVT); if ((Test & fcAllFlags) == fcAllFlags) return DAG.getBoolConstant(true, DL, ResultVT, OperandVT); // PPC double double is a pair of doubles, of which the higher part determines // the value class. if (OperandVT == MVT::ppcf128) { Op = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::f64, Op, DAG.getConstant(1, DL, MVT::i32)); OperandVT = MVT::f64; } // Some checks may be represented as inversion of simpler check, for example // "inf|normal|subnormal|zero" => !"nan". bool IsInverted = false; if (FPClassTest InvertedCheck = invertFPClassTestIfSimpler(Test)) { IsInverted = true; Test = InvertedCheck; } // Floating-point type properties. EVT ScalarFloatVT = OperandVT.getScalarType(); const Type *FloatTy = ScalarFloatVT.getTypeForEVT(*DAG.getContext()); const llvm::fltSemantics &Semantics = FloatTy->getFltSemantics(); bool IsF80 = (ScalarFloatVT == MVT::f80); // Some checks can be implemented using float comparisons, if floating point // exceptions are ignored. if (Flags.hasNoFPExcept() && isOperationLegalOrCustom(ISD::SETCC, OperandVT.getScalarType())) { ISD::CondCode OrderedCmpOpcode = IsInverted ? ISD::SETUNE : ISD::SETOEQ; ISD::CondCode UnorderedCmpOpcode = IsInverted ? ISD::SETONE : ISD::SETUEQ; if (std::optional IsCmp0 = isFCmpEqualZero(Test, Semantics, DAG.getMachineFunction()); IsCmp0 && (isCondCodeLegalOrCustom( *IsCmp0 ? OrderedCmpOpcode : UnorderedCmpOpcode, OperandVT.getScalarType().getSimpleVT()))) { // If denormals could be implicitly treated as 0, this is not equivalent // to a compare with 0 since it will also be true for denormals. return DAG.getSetCC(DL, ResultVT, Op, DAG.getConstantFP(0.0, DL, OperandVT), *IsCmp0 ? OrderedCmpOpcode : UnorderedCmpOpcode); } if (Test == fcNan && isCondCodeLegalOrCustom(IsInverted ? ISD::SETO : ISD::SETUO, OperandVT.getScalarType().getSimpleVT())) { return DAG.getSetCC(DL, ResultVT, Op, Op, IsInverted ? ISD::SETO : ISD::SETUO); } if (Test == fcInf && isCondCodeLegalOrCustom(IsInverted ? ISD::SETUNE : ISD::SETOEQ, OperandVT.getScalarType().getSimpleVT()) && isOperationLegalOrCustom(ISD::FABS, OperandVT.getScalarType())) { // isinf(x) --> fabs(x) == inf SDValue Abs = DAG.getNode(ISD::FABS, DL, OperandVT, Op); SDValue Inf = DAG.getConstantFP(APFloat::getInf(Semantics), DL, OperandVT); return DAG.getSetCC(DL, ResultVT, Abs, Inf, IsInverted ? ISD::SETUNE : ISD::SETOEQ); } } // In the general case use integer operations. unsigned BitSize = OperandVT.getScalarSizeInBits(); EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), BitSize); if (OperandVT.isVector()) IntVT = EVT::getVectorVT(*DAG.getContext(), IntVT, OperandVT.getVectorElementCount()); SDValue OpAsInt = DAG.getBitcast(IntVT, Op); // Various masks. APInt SignBit = APInt::getSignMask(BitSize); APInt ValueMask = APInt::getSignedMaxValue(BitSize); // All bits but sign. APInt Inf = APFloat::getInf(Semantics).bitcastToAPInt(); // Exp and int bit. const unsigned ExplicitIntBitInF80 = 63; APInt ExpMask = Inf; if (IsF80) ExpMask.clearBit(ExplicitIntBitInF80); APInt AllOneMantissa = APFloat::getLargest(Semantics).bitcastToAPInt() & ~Inf; APInt QNaNBitMask = APInt::getOneBitSet(BitSize, AllOneMantissa.getActiveBits() - 1); APInt InvertionMask = APInt::getAllOnes(ResultVT.getScalarSizeInBits()); SDValue ValueMaskV = DAG.getConstant(ValueMask, DL, IntVT); SDValue SignBitV = DAG.getConstant(SignBit, DL, IntVT); SDValue ExpMaskV = DAG.getConstant(ExpMask, DL, IntVT); SDValue ZeroV = DAG.getConstant(0, DL, IntVT); SDValue InfV = DAG.getConstant(Inf, DL, IntVT); SDValue ResultInvertionMask = DAG.getConstant(InvertionMask, DL, ResultVT); SDValue Res; const auto appendResult = [&](SDValue PartialRes) { if (PartialRes) { if (Res) Res = DAG.getNode(ISD::OR, DL, ResultVT, Res, PartialRes); else Res = PartialRes; } }; SDValue IntBitIsSetV; // Explicit integer bit in f80 mantissa is set. const auto getIntBitIsSet = [&]() -> SDValue { if (!IntBitIsSetV) { APInt IntBitMask(BitSize, 0); IntBitMask.setBit(ExplicitIntBitInF80); SDValue IntBitMaskV = DAG.getConstant(IntBitMask, DL, IntVT); SDValue IntBitV = DAG.getNode(ISD::AND, DL, IntVT, OpAsInt, IntBitMaskV); IntBitIsSetV = DAG.getSetCC(DL, ResultVT, IntBitV, ZeroV, ISD::SETNE); } return IntBitIsSetV; }; // Split the value into sign bit and absolute value. SDValue AbsV = DAG.getNode(ISD::AND, DL, IntVT, OpAsInt, ValueMaskV); SDValue SignV = DAG.getSetCC(DL, ResultVT, OpAsInt, DAG.getConstant(0.0, DL, IntVT), ISD::SETLT); // Tests that involve more than one class should be processed first. SDValue PartialRes; if (IsF80) ; // Detect finite numbers of f80 by checking individual classes because // they have different settings of the explicit integer bit. else if ((Test & fcFinite) == fcFinite) { // finite(V) ==> abs(V) < exp_mask PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, ExpMaskV, ISD::SETLT); Test &= ~fcFinite; } else if ((Test & fcFinite) == fcPosFinite) { // finite(V) && V > 0 ==> V < exp_mask PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, ExpMaskV, ISD::SETULT); Test &= ~fcPosFinite; } else if ((Test & fcFinite) == fcNegFinite) { // finite(V) && V < 0 ==> abs(V) < exp_mask && signbit == 1 PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, ExpMaskV, ISD::SETLT); PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, SignV); Test &= ~fcNegFinite; } appendResult(PartialRes); if (FPClassTest PartialCheck = Test & (fcZero | fcSubnormal)) { // fcZero | fcSubnormal => test all exponent bits are 0 // TODO: Handle sign bit specific cases if (PartialCheck == (fcZero | fcSubnormal)) { SDValue ExpBits = DAG.getNode(ISD::AND, DL, IntVT, OpAsInt, ExpMaskV); SDValue ExpIsZero = DAG.getSetCC(DL, ResultVT, ExpBits, ZeroV, ISD::SETEQ); appendResult(ExpIsZero); Test &= ~PartialCheck & fcAllFlags; } } // Check for individual classes. if (unsigned PartialCheck = Test & fcZero) { if (PartialCheck == fcPosZero) PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, ZeroV, ISD::SETEQ); else if (PartialCheck == fcZero) PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, ZeroV, ISD::SETEQ); else // ISD::fcNegZero PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, SignBitV, ISD::SETEQ); appendResult(PartialRes); } if (unsigned PartialCheck = Test & fcSubnormal) { // issubnormal(V) ==> unsigned(abs(V) - 1) < (all mantissa bits set) // issubnormal(V) && V>0 ==> unsigned(V - 1) < (all mantissa bits set) SDValue V = (PartialCheck == fcPosSubnormal) ? OpAsInt : AbsV; SDValue MantissaV = DAG.getConstant(AllOneMantissa, DL, IntVT); SDValue VMinusOneV = DAG.getNode(ISD::SUB, DL, IntVT, V, DAG.getConstant(1, DL, IntVT)); PartialRes = DAG.getSetCC(DL, ResultVT, VMinusOneV, MantissaV, ISD::SETULT); if (PartialCheck == fcNegSubnormal) PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, SignV); appendResult(PartialRes); } if (unsigned PartialCheck = Test & fcInf) { if (PartialCheck == fcPosInf) PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, InfV, ISD::SETEQ); else if (PartialCheck == fcInf) PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, InfV, ISD::SETEQ); else { // ISD::fcNegInf APInt NegInf = APFloat::getInf(Semantics, true).bitcastToAPInt(); SDValue NegInfV = DAG.getConstant(NegInf, DL, IntVT); PartialRes = DAG.getSetCC(DL, ResultVT, OpAsInt, NegInfV, ISD::SETEQ); } appendResult(PartialRes); } if (unsigned PartialCheck = Test & fcNan) { APInt InfWithQnanBit = Inf | QNaNBitMask; SDValue InfWithQnanBitV = DAG.getConstant(InfWithQnanBit, DL, IntVT); if (PartialCheck == fcNan) { // isnan(V) ==> abs(V) > int(inf) PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, InfV, ISD::SETGT); if (IsF80) { // Recognize unsupported values as NaNs for compatibility with glibc. // In them (exp(V)==0) == int_bit. SDValue ExpBits = DAG.getNode(ISD::AND, DL, IntVT, AbsV, ExpMaskV); SDValue ExpIsZero = DAG.getSetCC(DL, ResultVT, ExpBits, ZeroV, ISD::SETEQ); SDValue IsPseudo = DAG.getSetCC(DL, ResultVT, getIntBitIsSet(), ExpIsZero, ISD::SETEQ); PartialRes = DAG.getNode(ISD::OR, DL, ResultVT, PartialRes, IsPseudo); } } else if (PartialCheck == fcQNan) { // isquiet(V) ==> abs(V) >= (unsigned(Inf) | quiet_bit) PartialRes = DAG.getSetCC(DL, ResultVT, AbsV, InfWithQnanBitV, ISD::SETGE); } else { // ISD::fcSNan // issignaling(V) ==> abs(V) > unsigned(Inf) && // abs(V) < (unsigned(Inf) | quiet_bit) SDValue IsNan = DAG.getSetCC(DL, ResultVT, AbsV, InfV, ISD::SETGT); SDValue IsNotQnan = DAG.getSetCC(DL, ResultVT, AbsV, InfWithQnanBitV, ISD::SETLT); PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, IsNan, IsNotQnan); } appendResult(PartialRes); } if (unsigned PartialCheck = Test & fcNormal) { // isnormal(V) ==> (0 < exp < max_exp) ==> (unsigned(exp-1) < (max_exp-1)) APInt ExpLSB = ExpMask & ~(ExpMask.shl(1)); SDValue ExpLSBV = DAG.getConstant(ExpLSB, DL, IntVT); SDValue ExpMinus1 = DAG.getNode(ISD::SUB, DL, IntVT, AbsV, ExpLSBV); APInt ExpLimit = ExpMask - ExpLSB; SDValue ExpLimitV = DAG.getConstant(ExpLimit, DL, IntVT); PartialRes = DAG.getSetCC(DL, ResultVT, ExpMinus1, ExpLimitV, ISD::SETULT); if (PartialCheck == fcNegNormal) PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, SignV); else if (PartialCheck == fcPosNormal) { SDValue PosSignV = DAG.getNode(ISD::XOR, DL, ResultVT, SignV, ResultInvertionMask); PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, PosSignV); } if (IsF80) PartialRes = DAG.getNode(ISD::AND, DL, ResultVT, PartialRes, getIntBitIsSet()); appendResult(PartialRes); } if (!Res) return DAG.getConstant(IsInverted, DL, ResultVT); if (IsInverted) Res = DAG.getNode(ISD::XOR, DL, ResultVT, Res, ResultInvertionMask); return Res; } // Only expand vector types if we have the appropriate vector bit operations. static bool canExpandVectorCTPOP(const TargetLowering &TLI, EVT VT) { assert(VT.isVector() && "Expected vector type"); unsigned Len = VT.getScalarSizeInBits(); return TLI.isOperationLegalOrCustom(ISD::ADD, VT) && TLI.isOperationLegalOrCustom(ISD::SUB, VT) && TLI.isOperationLegalOrCustom(ISD::SRL, VT) && (Len == 8 || TLI.isOperationLegalOrCustom(ISD::MUL, VT)) && TLI.isOperationLegalOrCustomOrPromote(ISD::AND, VT); } SDValue TargetLowering::expandCTPOP(SDNode *Node, SelectionDAG &DAG) const { SDLoc dl(Node); EVT VT = Node->getValueType(0); EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); SDValue Op = Node->getOperand(0); unsigned Len = VT.getScalarSizeInBits(); assert(VT.isInteger() && "CTPOP not implemented for this type."); // TODO: Add support for irregular type lengths. if (!(Len <= 128 && Len % 8 == 0)) return SDValue(); // Only expand vector types if we have the appropriate vector bit operations. if (VT.isVector() && !canExpandVectorCTPOP(*this, VT)) return SDValue(); // This is the "best" algorithm from // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), dl, VT); SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), dl, VT); SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), dl, VT); // v = v - ((v >> 1) & 0x55555555...) Op = DAG.getNode(ISD::SUB, dl, VT, Op, DAG.getNode(ISD::AND, dl, VT, DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(1, dl, ShVT)), Mask55)); // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...) Op = DAG.getNode(ISD::ADD, dl, VT, DAG.getNode(ISD::AND, dl, VT, Op, Mask33), DAG.getNode(ISD::AND, dl, VT, DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(2, dl, ShVT)), Mask33)); // v = (v + (v >> 4)) & 0x0F0F0F0F... Op = DAG.getNode(ISD::AND, dl, VT, DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(4, dl, ShVT))), Mask0F); if (Len <= 8) return Op; // Avoid the multiply if we only have 2 bytes to add. // TODO: Only doing this for scalars because vectors weren't as obviously // improved. if (Len == 16 && !VT.isVector()) { // v = (v + (v >> 8)) & 0x00FF; return DAG.getNode(ISD::AND, dl, VT, DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, ShVT))), DAG.getConstant(0xFF, dl, VT)); } // v = (v * 0x01010101...) >> (Len - 8) SDValue V; if (isOperationLegalOrCustomOrPromote( ISD::MUL, getTypeToTransformTo(*DAG.getContext(), VT))) { SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), dl, VT); V = DAG.getNode(ISD::MUL, dl, VT, Op, Mask01); } else { V = Op; for (unsigned Shift = 8; Shift < Len; Shift *= 2) { SDValue ShiftC = DAG.getShiftAmountConstant(Shift, VT, dl); V = DAG.getNode(ISD::ADD, dl, VT, V, DAG.getNode(ISD::SHL, dl, VT, V, ShiftC)); } } return DAG.getNode(ISD::SRL, dl, VT, V, DAG.getConstant(Len - 8, dl, ShVT)); } SDValue TargetLowering::expandVPCTPOP(SDNode *Node, SelectionDAG &DAG) const { SDLoc dl(Node); EVT VT = Node->getValueType(0); EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); SDValue Op = Node->getOperand(0); SDValue Mask = Node->getOperand(1); SDValue VL = Node->getOperand(2); unsigned Len = VT.getScalarSizeInBits(); assert(VT.isInteger() && "VP_CTPOP not implemented for this type."); // TODO: Add support for irregular type lengths. if (!(Len <= 128 && Len % 8 == 0)) return SDValue(); // This is same algorithm of expandCTPOP from // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), dl, VT); SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), dl, VT); SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), dl, VT); SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5; // v = v - ((v >> 1) & 0x55555555...) Tmp1 = DAG.getNode(ISD::VP_AND, dl, VT, DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(1, dl, ShVT), Mask, VL), Mask55, Mask, VL); Op = DAG.getNode(ISD::VP_SUB, dl, VT, Op, Tmp1, Mask, VL); // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...) Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Op, Mask33, Mask, VL); Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(2, dl, ShVT), Mask, VL), Mask33, Mask, VL); Op = DAG.getNode(ISD::VP_ADD, dl, VT, Tmp2, Tmp3, Mask, VL); // v = (v + (v >> 4)) & 0x0F0F0F0F... Tmp4 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(4, dl, ShVT), Mask, VL), Tmp5 = DAG.getNode(ISD::VP_ADD, dl, VT, Op, Tmp4, Mask, VL); Op = DAG.getNode(ISD::VP_AND, dl, VT, Tmp5, Mask0F, Mask, VL); if (Len <= 8) return Op; // v = (v * 0x01010101...) >> (Len - 8) SDValue V; if (isOperationLegalOrCustomOrPromote( ISD::VP_MUL, getTypeToTransformTo(*DAG.getContext(), VT))) { SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), dl, VT); V = DAG.getNode(ISD::VP_MUL, dl, VT, Op, Mask01, Mask, VL); } else { V = Op; for (unsigned Shift = 8; Shift < Len; Shift *= 2) { SDValue ShiftC = DAG.getShiftAmountConstant(Shift, VT, dl); V = DAG.getNode(ISD::VP_ADD, dl, VT, V, DAG.getNode(ISD::VP_SHL, dl, VT, V, ShiftC, Mask, VL), Mask, VL); } } return DAG.getNode(ISD::VP_SRL, dl, VT, V, DAG.getConstant(Len - 8, dl, ShVT), Mask, VL); } SDValue TargetLowering::expandCTLZ(SDNode *Node, SelectionDAG &DAG) const { SDLoc dl(Node); EVT VT = Node->getValueType(0); EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); SDValue Op = Node->getOperand(0); unsigned NumBitsPerElt = VT.getScalarSizeInBits(); // If the non-ZERO_UNDEF version is supported we can use that instead. if (Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF && isOperationLegalOrCustom(ISD::CTLZ, VT)) return DAG.getNode(ISD::CTLZ, dl, VT, Op); // If the ZERO_UNDEF version is supported use that and handle the zero case. if (isOperationLegalOrCustom(ISD::CTLZ_ZERO_UNDEF, VT)) { EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDValue CTLZ = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, dl, VT, Op); SDValue Zero = DAG.getConstant(0, dl, VT); SDValue SrcIsZero = DAG.getSetCC(dl, SetCCVT, Op, Zero, ISD::SETEQ); return DAG.getSelect(dl, VT, SrcIsZero, DAG.getConstant(NumBitsPerElt, dl, VT), CTLZ); } // Only expand vector types if we have the appropriate vector bit operations. // This includes the operations needed to expand CTPOP if it isn't supported. if (VT.isVector() && (!isPowerOf2_32(NumBitsPerElt) || (!isOperationLegalOrCustom(ISD::CTPOP, VT) && !canExpandVectorCTPOP(*this, VT)) || !isOperationLegalOrCustom(ISD::SRL, VT) || !isOperationLegalOrCustomOrPromote(ISD::OR, VT))) return SDValue(); // for now, we do this: // x = x | (x >> 1); // x = x | (x >> 2); // ... // x = x | (x >>16); // x = x | (x >>32); // for 64-bit input // return popcount(~x); // // Ref: "Hacker's Delight" by Henry Warren for (unsigned i = 0; (1U << i) < NumBitsPerElt; ++i) { SDValue Tmp = DAG.getConstant(1ULL << i, dl, ShVT); Op = DAG.getNode(ISD::OR, dl, VT, Op, DAG.getNode(ISD::SRL, dl, VT, Op, Tmp)); } Op = DAG.getNOT(dl, Op, VT); return DAG.getNode(ISD::CTPOP, dl, VT, Op); } SDValue TargetLowering::expandVPCTLZ(SDNode *Node, SelectionDAG &DAG) const { SDLoc dl(Node); EVT VT = Node->getValueType(0); EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); SDValue Op = Node->getOperand(0); SDValue Mask = Node->getOperand(1); SDValue VL = Node->getOperand(2); unsigned NumBitsPerElt = VT.getScalarSizeInBits(); // do this: // x = x | (x >> 1); // x = x | (x >> 2); // ... // x = x | (x >>16); // x = x | (x >>32); // for 64-bit input // return popcount(~x); for (unsigned i = 0; (1U << i) < NumBitsPerElt; ++i) { SDValue Tmp = DAG.getConstant(1ULL << i, dl, ShVT); Op = DAG.getNode(ISD::VP_OR, dl, VT, Op, DAG.getNode(ISD::VP_SRL, dl, VT, Op, Tmp, Mask, VL), Mask, VL); } Op = DAG.getNode(ISD::VP_XOR, dl, VT, Op, DAG.getConstant(-1, dl, VT), Mask, VL); return DAG.getNode(ISD::VP_CTPOP, dl, VT, Op, Mask, VL); } SDValue TargetLowering::CTTZTableLookup(SDNode *Node, SelectionDAG &DAG, const SDLoc &DL, EVT VT, SDValue Op, unsigned BitWidth) const { if (BitWidth != 32 && BitWidth != 64) return SDValue(); APInt DeBruijn = BitWidth == 32 ? APInt(32, 0x077CB531U) : APInt(64, 0x0218A392CD3D5DBFULL); const DataLayout &TD = DAG.getDataLayout(); MachinePointerInfo PtrInfo = MachinePointerInfo::getConstantPool(DAG.getMachineFunction()); unsigned ShiftAmt = BitWidth - Log2_32(BitWidth); SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Op); SDValue Lookup = DAG.getNode( ISD::SRL, DL, VT, DAG.getNode(ISD::MUL, DL, VT, DAG.getNode(ISD::AND, DL, VT, Op, Neg), DAG.getConstant(DeBruijn, DL, VT)), DAG.getConstant(ShiftAmt, DL, VT)); Lookup = DAG.getSExtOrTrunc(Lookup, DL, getPointerTy(TD)); SmallVector Table(BitWidth, 0); for (unsigned i = 0; i < BitWidth; i++) { APInt Shl = DeBruijn.shl(i); APInt Lshr = Shl.lshr(ShiftAmt); Table[Lshr.getZExtValue()] = i; } // Create a ConstantArray in Constant Pool auto *CA = ConstantDataArray::get(*DAG.getContext(), Table); SDValue CPIdx = DAG.getConstantPool(CA, getPointerTy(TD), TD.getPrefTypeAlign(CA->getType())); SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, VT, DAG.getEntryNode(), DAG.getMemBasePlusOffset(CPIdx, Lookup, DL), PtrInfo, MVT::i8); if (Node->getOpcode() == ISD::CTTZ_ZERO_UNDEF) return ExtLoad; EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDValue Zero = DAG.getConstant(0, DL, VT); SDValue SrcIsZero = DAG.getSetCC(DL, SetCCVT, Op, Zero, ISD::SETEQ); return DAG.getSelect(DL, VT, SrcIsZero, DAG.getConstant(BitWidth, DL, VT), ExtLoad); } SDValue TargetLowering::expandCTTZ(SDNode *Node, SelectionDAG &DAG) const { SDLoc dl(Node); EVT VT = Node->getValueType(0); SDValue Op = Node->getOperand(0); unsigned NumBitsPerElt = VT.getScalarSizeInBits(); // If the non-ZERO_UNDEF version is supported we can use that instead. if (Node->getOpcode() == ISD::CTTZ_ZERO_UNDEF && isOperationLegalOrCustom(ISD::CTTZ, VT)) return DAG.getNode(ISD::CTTZ, dl, VT, Op); // If the ZERO_UNDEF version is supported use that and handle the zero case. if (isOperationLegalOrCustom(ISD::CTTZ_ZERO_UNDEF, VT)) { EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDValue CTTZ = DAG.getNode(ISD::CTTZ_ZERO_UNDEF, dl, VT, Op); SDValue Zero = DAG.getConstant(0, dl, VT); SDValue SrcIsZero = DAG.getSetCC(dl, SetCCVT, Op, Zero, ISD::SETEQ); return DAG.getSelect(dl, VT, SrcIsZero, DAG.getConstant(NumBitsPerElt, dl, VT), CTTZ); } // Only expand vector types if we have the appropriate vector bit operations. // This includes the operations needed to expand CTPOP if it isn't supported. if (VT.isVector() && (!isPowerOf2_32(NumBitsPerElt) || (!isOperationLegalOrCustom(ISD::CTPOP, VT) && !isOperationLegalOrCustom(ISD::CTLZ, VT) && !canExpandVectorCTPOP(*this, VT)) || !isOperationLegalOrCustom(ISD::SUB, VT) || !isOperationLegalOrCustomOrPromote(ISD::AND, VT) || !isOperationLegalOrCustomOrPromote(ISD::XOR, VT))) return SDValue(); // Emit Table Lookup if ISD::CTLZ and ISD::CTPOP are not legal. if (!VT.isVector() && isOperationExpand(ISD::CTPOP, VT) && !isOperationLegal(ISD::CTLZ, VT)) if (SDValue V = CTTZTableLookup(Node, DAG, dl, VT, Op, NumBitsPerElt)) return V; // for now, we use: { return popcount(~x & (x - 1)); } // unless the target has ctlz but not ctpop, in which case we use: // { return 32 - nlz(~x & (x-1)); } // Ref: "Hacker's Delight" by Henry Warren SDValue Tmp = DAG.getNode( ISD::AND, dl, VT, DAG.getNOT(dl, Op, VT), DAG.getNode(ISD::SUB, dl, VT, Op, DAG.getConstant(1, dl, VT))); // If ISD::CTLZ is legal and CTPOP isn't, then do that instead. if (isOperationLegal(ISD::CTLZ, VT) && !isOperationLegal(ISD::CTPOP, VT)) { return DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(NumBitsPerElt, dl, VT), DAG.getNode(ISD::CTLZ, dl, VT, Tmp)); } return DAG.getNode(ISD::CTPOP, dl, VT, Tmp); } SDValue TargetLowering::expandVPCTTZ(SDNode *Node, SelectionDAG &DAG) const { SDValue Op = Node->getOperand(0); SDValue Mask = Node->getOperand(1); SDValue VL = Node->getOperand(2); SDLoc dl(Node); EVT VT = Node->getValueType(0); // Same as the vector part of expandCTTZ, use: popcount(~x & (x - 1)) SDValue Not = DAG.getNode(ISD::VP_XOR, dl, VT, Op, DAG.getConstant(-1, dl, VT), Mask, VL); SDValue MinusOne = DAG.getNode(ISD::VP_SUB, dl, VT, Op, DAG.getConstant(1, dl, VT), Mask, VL); SDValue Tmp = DAG.getNode(ISD::VP_AND, dl, VT, Not, MinusOne, Mask, VL); return DAG.getNode(ISD::VP_CTPOP, dl, VT, Tmp, Mask, VL); } SDValue TargetLowering::expandVPCTTZElements(SDNode *N, SelectionDAG &DAG) const { // %cond = to_bool_vec %source // %splat = splat /*val=*/VL // %tz = step_vector // %v = vp.select %cond, /*true=*/tz, /*false=*/%splat // %r = vp.reduce.umin %v SDLoc DL(N); SDValue Source = N->getOperand(0); SDValue Mask = N->getOperand(1); SDValue EVL = N->getOperand(2); EVT SrcVT = Source.getValueType(); EVT ResVT = N->getValueType(0); EVT ResVecVT = EVT::getVectorVT(*DAG.getContext(), ResVT, SrcVT.getVectorElementCount()); // Convert to boolean vector. if (SrcVT.getScalarType() != MVT::i1) { SDValue AllZero = DAG.getConstant(0, DL, SrcVT); SrcVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, SrcVT.getVectorElementCount()); Source = DAG.getNode(ISD::VP_SETCC, DL, SrcVT, Source, AllZero, DAG.getCondCode(ISD::SETNE), Mask, EVL); } SDValue ExtEVL = DAG.getZExtOrTrunc(EVL, DL, ResVT); SDValue Splat = DAG.getSplat(ResVecVT, DL, ExtEVL); SDValue StepVec = DAG.getStepVector(DL, ResVecVT); SDValue Select = DAG.getNode(ISD::VP_SELECT, DL, ResVecVT, Source, StepVec, Splat, EVL); return DAG.getNode(ISD::VP_REDUCE_UMIN, DL, ResVT, ExtEVL, Select, Mask, EVL); } SDValue TargetLowering::expandABS(SDNode *N, SelectionDAG &DAG, bool IsNegative) const { SDLoc dl(N); EVT VT = N->getValueType(0); SDValue Op = N->getOperand(0); // abs(x) -> smax(x,sub(0,x)) if (!IsNegative && isOperationLegal(ISD::SUB, VT) && isOperationLegal(ISD::SMAX, VT)) { SDValue Zero = DAG.getConstant(0, dl, VT); Op = DAG.getFreeze(Op); return DAG.getNode(ISD::SMAX, dl, VT, Op, DAG.getNode(ISD::SUB, dl, VT, Zero, Op)); } // abs(x) -> umin(x,sub(0,x)) if (!IsNegative && isOperationLegal(ISD::SUB, VT) && isOperationLegal(ISD::UMIN, VT)) { SDValue Zero = DAG.getConstant(0, dl, VT); Op = DAG.getFreeze(Op); return DAG.getNode(ISD::UMIN, dl, VT, Op, DAG.getNode(ISD::SUB, dl, VT, Zero, Op)); } // 0 - abs(x) -> smin(x, sub(0,x)) if (IsNegative && isOperationLegal(ISD::SUB, VT) && isOperationLegal(ISD::SMIN, VT)) { SDValue Zero = DAG.getConstant(0, dl, VT); Op = DAG.getFreeze(Op); return DAG.getNode(ISD::SMIN, dl, VT, Op, DAG.getNode(ISD::SUB, dl, VT, Zero, Op)); } // Only expand vector types if we have the appropriate vector operations. if (VT.isVector() && (!isOperationLegalOrCustom(ISD::SRA, VT) || (!IsNegative && !isOperationLegalOrCustom(ISD::ADD, VT)) || (IsNegative && !isOperationLegalOrCustom(ISD::SUB, VT)) || !isOperationLegalOrCustomOrPromote(ISD::XOR, VT))) return SDValue(); Op = DAG.getFreeze(Op); SDValue Shift = DAG.getNode( ISD::SRA, dl, VT, Op, DAG.getShiftAmountConstant(VT.getScalarSizeInBits() - 1, VT, dl)); SDValue Xor = DAG.getNode(ISD::XOR, dl, VT, Op, Shift); // abs(x) -> Y = sra (X, size(X)-1); sub (xor (X, Y), Y) if (!IsNegative) return DAG.getNode(ISD::SUB, dl, VT, Xor, Shift); // 0 - abs(x) -> Y = sra (X, size(X)-1); sub (Y, xor (X, Y)) return DAG.getNode(ISD::SUB, dl, VT, Shift, Xor); } SDValue TargetLowering::expandABD(SDNode *N, SelectionDAG &DAG) const { SDLoc dl(N); EVT VT = N->getValueType(0); SDValue LHS = DAG.getFreeze(N->getOperand(0)); SDValue RHS = DAG.getFreeze(N->getOperand(1)); bool IsSigned = N->getOpcode() == ISD::ABDS; // abds(lhs, rhs) -> sub(smax(lhs,rhs), smin(lhs,rhs)) // abdu(lhs, rhs) -> sub(umax(lhs,rhs), umin(lhs,rhs)) unsigned MaxOpc = IsSigned ? ISD::SMAX : ISD::UMAX; unsigned MinOpc = IsSigned ? ISD::SMIN : ISD::UMIN; if (isOperationLegal(MaxOpc, VT) && isOperationLegal(MinOpc, VT)) { SDValue Max = DAG.getNode(MaxOpc, dl, VT, LHS, RHS); SDValue Min = DAG.getNode(MinOpc, dl, VT, LHS, RHS); return DAG.getNode(ISD::SUB, dl, VT, Max, Min); } // abdu(lhs, rhs) -> or(usubsat(lhs,rhs), usubsat(rhs,lhs)) if (!IsSigned && isOperationLegal(ISD::USUBSAT, VT)) return DAG.getNode(ISD::OR, dl, VT, DAG.getNode(ISD::USUBSAT, dl, VT, LHS, RHS), DAG.getNode(ISD::USUBSAT, dl, VT, RHS, LHS)); EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); ISD::CondCode CC = IsSigned ? ISD::CondCode::SETGT : ISD::CondCode::SETUGT; SDValue Cmp = DAG.getSetCC(dl, CCVT, LHS, RHS, CC); // Branchless expansion iff cmp result is allbits: // abds(lhs, rhs) -> sub(sgt(lhs, rhs), xor(sgt(lhs, rhs), sub(lhs, rhs))) // abdu(lhs, rhs) -> sub(ugt(lhs, rhs), xor(ugt(lhs, rhs), sub(lhs, rhs))) if (CCVT == VT && getBooleanContents(VT) == ZeroOrNegativeOneBooleanContent) { SDValue Diff = DAG.getNode(ISD::SUB, dl, VT, LHS, RHS); SDValue Xor = DAG.getNode(ISD::XOR, dl, VT, Diff, Cmp); return DAG.getNode(ISD::SUB, dl, VT, Cmp, Xor); } // abds(lhs, rhs) -> select(sgt(lhs,rhs), sub(lhs,rhs), sub(rhs,lhs)) // abdu(lhs, rhs) -> select(ugt(lhs,rhs), sub(lhs,rhs), sub(rhs,lhs)) return DAG.getSelect(dl, VT, Cmp, DAG.getNode(ISD::SUB, dl, VT, LHS, RHS), DAG.getNode(ISD::SUB, dl, VT, RHS, LHS)); } SDValue TargetLowering::expandAVG(SDNode *N, SelectionDAG &DAG) const { SDLoc dl(N); EVT VT = N->getValueType(0); SDValue LHS = N->getOperand(0); SDValue RHS = N->getOperand(1); unsigned Opc = N->getOpcode(); bool IsFloor = Opc == ISD::AVGFLOORS || Opc == ISD::AVGFLOORU; bool IsSigned = Opc == ISD::AVGCEILS || Opc == ISD::AVGFLOORS; unsigned SumOpc = IsFloor ? ISD::ADD : ISD::SUB; unsigned SignOpc = IsFloor ? ISD::AND : ISD::OR; unsigned ShiftOpc = IsSigned ? ISD::SRA : ISD::SRL; unsigned ExtOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; assert((Opc == ISD::AVGFLOORS || Opc == ISD::AVGCEILS || Opc == ISD::AVGFLOORU || Opc == ISD::AVGCEILU) && "Unknown AVG node"); // If the operands are already extended, we can add+shift. bool IsExt = (IsSigned && DAG.ComputeNumSignBits(LHS) >= 2 && DAG.ComputeNumSignBits(RHS) >= 2) || (!IsSigned && DAG.computeKnownBits(LHS).countMinLeadingZeros() >= 1 && DAG.computeKnownBits(RHS).countMinLeadingZeros() >= 1); if (IsExt) { SDValue Sum = DAG.getNode(ISD::ADD, dl, VT, LHS, RHS); if (!IsFloor) Sum = DAG.getNode(ISD::ADD, dl, VT, Sum, DAG.getConstant(1, dl, VT)); return DAG.getNode(ShiftOpc, dl, VT, Sum, DAG.getShiftAmountConstant(1, VT, dl)); } // For scalars, see if we can efficiently extend/truncate to use add+shift. if (VT.isScalarInteger()) { unsigned BW = VT.getScalarSizeInBits(); EVT ExtVT = VT.getIntegerVT(*DAG.getContext(), 2 * BW); if (isTypeLegal(ExtVT) && isTruncateFree(ExtVT, VT)) { LHS = DAG.getNode(ExtOpc, dl, ExtVT, LHS); RHS = DAG.getNode(ExtOpc, dl, ExtVT, RHS); SDValue Avg = DAG.getNode(ISD::ADD, dl, ExtVT, LHS, RHS); if (!IsFloor) Avg = DAG.getNode(ISD::ADD, dl, ExtVT, Avg, DAG.getConstant(1, dl, ExtVT)); // Just use SRL as we will be truncating away the extended sign bits. Avg = DAG.getNode(ISD::SRL, dl, ExtVT, Avg, DAG.getShiftAmountConstant(1, ExtVT, dl)); return DAG.getNode(ISD::TRUNCATE, dl, VT, Avg); } } // avgceils(lhs, rhs) -> sub(or(lhs,rhs),ashr(xor(lhs,rhs),1)) // avgceilu(lhs, rhs) -> sub(or(lhs,rhs),lshr(xor(lhs,rhs),1)) // avgfloors(lhs, rhs) -> add(and(lhs,rhs),ashr(xor(lhs,rhs),1)) // avgflooru(lhs, rhs) -> add(and(lhs,rhs),lshr(xor(lhs,rhs),1)) LHS = DAG.getFreeze(LHS); RHS = DAG.getFreeze(RHS); SDValue Sign = DAG.getNode(SignOpc, dl, VT, LHS, RHS); SDValue Xor = DAG.getNode(ISD::XOR, dl, VT, LHS, RHS); SDValue Shift = DAG.getNode(ShiftOpc, dl, VT, Xor, DAG.getShiftAmountConstant(1, VT, dl)); return DAG.getNode(SumOpc, dl, VT, Sign, Shift); } SDValue TargetLowering::expandBSWAP(SDNode *N, SelectionDAG &DAG) const { SDLoc dl(N); EVT VT = N->getValueType(0); SDValue Op = N->getOperand(0); if (!VT.isSimple()) return SDValue(); EVT SHVT = getShiftAmountTy(VT, DAG.getDataLayout()); SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8; switch (VT.getSimpleVT().getScalarType().SimpleTy) { default: return SDValue(); case MVT::i16: // Use a rotate by 8. This can be further expanded if necessary. return DAG.getNode(ISD::ROTL, dl, VT, Op, DAG.getConstant(8, dl, SHVT)); case MVT::i32: Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT)); Tmp3 = DAG.getNode(ISD::AND, dl, VT, Op, DAG.getConstant(0xFF00, dl, VT)); Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(8, dl, SHVT)); Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT)); Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, dl, VT)); Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT)); Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3); Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1); return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2); case MVT::i64: Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, dl, SHVT)); Tmp7 = DAG.getNode(ISD::AND, dl, VT, Op, DAG.getConstant(255ULL<<8, dl, VT)); Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Tmp7, DAG.getConstant(40, dl, SHVT)); Tmp6 = DAG.getNode(ISD::AND, dl, VT, Op, DAG.getConstant(255ULL<<16, dl, VT)); Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Tmp6, DAG.getConstant(24, dl, SHVT)); Tmp5 = DAG.getNode(ISD::AND, dl, VT, Op, DAG.getConstant(255ULL<<24, dl, VT)); Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Tmp5, DAG.getConstant(8, dl, SHVT)); Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT)); Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, dl, VT)); Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT)); Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, dl, VT)); Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, dl, SHVT)); Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8, dl, VT)); Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, dl, SHVT)); Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7); Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5); Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3); Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1); Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6); Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2); return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4); } } SDValue TargetLowering::expandVPBSWAP(SDNode *N, SelectionDAG &DAG) const { SDLoc dl(N); EVT VT = N->getValueType(0); SDValue Op = N->getOperand(0); SDValue Mask = N->getOperand(1); SDValue EVL = N->getOperand(2); if (!VT.isSimple()) return SDValue(); EVT SHVT = getShiftAmountTy(VT, DAG.getDataLayout()); SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8; switch (VT.getSimpleVT().getScalarType().SimpleTy) { default: return SDValue(); case MVT::i16: Tmp1 = DAG.getNode(ISD::VP_SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT), Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT), Mask, EVL); return DAG.getNode(ISD::VP_OR, dl, VT, Tmp1, Tmp2, Mask, EVL); case MVT::i32: Tmp4 = DAG.getNode(ISD::VP_SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Op, DAG.getConstant(0xFF00, dl, VT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp3, DAG.getConstant(8, dl, SHVT), Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT), Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, dl, VT), Mask, EVL); Tmp1 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT), Mask, EVL); Tmp4 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp4, Tmp3, Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp1, Mask, EVL); return DAG.getNode(ISD::VP_OR, dl, VT, Tmp4, Tmp2, Mask, EVL); case MVT::i64: Tmp8 = DAG.getNode(ISD::VP_SHL, dl, VT, Op, DAG.getConstant(56, dl, SHVT), Mask, EVL); Tmp7 = DAG.getNode(ISD::VP_AND, dl, VT, Op, DAG.getConstant(255ULL << 8, dl, VT), Mask, EVL); Tmp7 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp7, DAG.getConstant(40, dl, SHVT), Mask, EVL); Tmp6 = DAG.getNode(ISD::VP_AND, dl, VT, Op, DAG.getConstant(255ULL << 16, dl, VT), Mask, EVL); Tmp6 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp6, DAG.getConstant(24, dl, SHVT), Mask, EVL); Tmp5 = DAG.getNode(ISD::VP_AND, dl, VT, Op, DAG.getConstant(255ULL << 24, dl, VT), Mask, EVL); Tmp5 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp5, DAG.getConstant(8, dl, SHVT), Mask, EVL); Tmp4 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT), Mask, EVL); Tmp4 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp4, DAG.getConstant(255ULL << 24, dl, VT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp3, DAG.getConstant(255ULL << 16, dl, VT), Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(40, dl, SHVT), Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2, DAG.getConstant(255ULL << 8, dl, VT), Mask, EVL); Tmp1 = DAG.getNode(ISD::VP_SRL, dl, VT, Op, DAG.getConstant(56, dl, SHVT), Mask, EVL); Tmp8 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp8, Tmp7, Mask, EVL); Tmp6 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp6, Tmp5, Mask, EVL); Tmp4 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp4, Tmp3, Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp1, Mask, EVL); Tmp8 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp8, Tmp6, Mask, EVL); Tmp4 = DAG.getNode(ISD::VP_OR, dl, VT, Tmp4, Tmp2, Mask, EVL); return DAG.getNode(ISD::VP_OR, dl, VT, Tmp8, Tmp4, Mask, EVL); } } SDValue TargetLowering::expandBITREVERSE(SDNode *N, SelectionDAG &DAG) const { SDLoc dl(N); EVT VT = N->getValueType(0); SDValue Op = N->getOperand(0); EVT SHVT = getShiftAmountTy(VT, DAG.getDataLayout()); unsigned Sz = VT.getScalarSizeInBits(); SDValue Tmp, Tmp2, Tmp3; // If we can, perform BSWAP first and then the mask+swap the i4, then i2 // and finally the i1 pairs. // TODO: We can easily support i4/i2 legal types if any target ever does. if (Sz >= 8 && isPowerOf2_32(Sz)) { // Create the masks - repeating the pattern every byte. APInt Mask4 = APInt::getSplat(Sz, APInt(8, 0x0F)); APInt Mask2 = APInt::getSplat(Sz, APInt(8, 0x33)); APInt Mask1 = APInt::getSplat(Sz, APInt(8, 0x55)); // BSWAP if the type is wider than a single byte. Tmp = (Sz > 8 ? DAG.getNode(ISD::BSWAP, dl, VT, Op) : Op); // swap i4: ((V >> 4) & 0x0F) | ((V & 0x0F) << 4) Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp, DAG.getConstant(4, dl, SHVT)); Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Mask4, dl, VT)); Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(Mask4, dl, VT)); Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(4, dl, SHVT)); Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3); // swap i2: ((V >> 2) & 0x33) | ((V & 0x33) << 2) Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp, DAG.getConstant(2, dl, SHVT)); Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Mask2, dl, VT)); Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(Mask2, dl, VT)); Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(2, dl, SHVT)); Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3); // swap i1: ((V >> 1) & 0x55) | ((V & 0x55) << 1) Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp, DAG.getConstant(1, dl, SHVT)); Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Mask1, dl, VT)); Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(Mask1, dl, VT)); Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(1, dl, SHVT)); Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3); return Tmp; } Tmp = DAG.getConstant(0, dl, VT); for (unsigned I = 0, J = Sz-1; I < Sz; ++I, --J) { if (I < J) Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(J - I, dl, SHVT)); else Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(I - J, dl, SHVT)); APInt Shift = APInt::getOneBitSet(Sz, J); Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Shift, dl, VT)); Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp, Tmp2); } return Tmp; } SDValue TargetLowering::expandVPBITREVERSE(SDNode *N, SelectionDAG &DAG) const { assert(N->getOpcode() == ISD::VP_BITREVERSE); SDLoc dl(N); EVT VT = N->getValueType(0); SDValue Op = N->getOperand(0); SDValue Mask = N->getOperand(1); SDValue EVL = N->getOperand(2); EVT SHVT = getShiftAmountTy(VT, DAG.getDataLayout()); unsigned Sz = VT.getScalarSizeInBits(); SDValue Tmp, Tmp2, Tmp3; // If we can, perform BSWAP first and then the mask+swap the i4, then i2 // and finally the i1 pairs. // TODO: We can easily support i4/i2 legal types if any target ever does. if (Sz >= 8 && isPowerOf2_32(Sz)) { // Create the masks - repeating the pattern every byte. APInt Mask4 = APInt::getSplat(Sz, APInt(8, 0x0F)); APInt Mask2 = APInt::getSplat(Sz, APInt(8, 0x33)); APInt Mask1 = APInt::getSplat(Sz, APInt(8, 0x55)); // BSWAP if the type is wider than a single byte. Tmp = (Sz > 8 ? DAG.getNode(ISD::VP_BSWAP, dl, VT, Op, Mask, EVL) : Op); // swap i4: ((V >> 4) & 0x0F) | ((V & 0x0F) << 4) Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Tmp, DAG.getConstant(4, dl, SHVT), Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2, DAG.getConstant(Mask4, dl, VT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp, DAG.getConstant(Mask4, dl, VT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp3, DAG.getConstant(4, dl, SHVT), Mask, EVL); Tmp = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp3, Mask, EVL); // swap i2: ((V >> 2) & 0x33) | ((V & 0x33) << 2) Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Tmp, DAG.getConstant(2, dl, SHVT), Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2, DAG.getConstant(Mask2, dl, VT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp, DAG.getConstant(Mask2, dl, VT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp3, DAG.getConstant(2, dl, SHVT), Mask, EVL); Tmp = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp3, Mask, EVL); // swap i1: ((V >> 1) & 0x55) | ((V & 0x55) << 1) Tmp2 = DAG.getNode(ISD::VP_SRL, dl, VT, Tmp, DAG.getConstant(1, dl, SHVT), Mask, EVL); Tmp2 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp2, DAG.getConstant(Mask1, dl, VT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_AND, dl, VT, Tmp, DAG.getConstant(Mask1, dl, VT), Mask, EVL); Tmp3 = DAG.getNode(ISD::VP_SHL, dl, VT, Tmp3, DAG.getConstant(1, dl, SHVT), Mask, EVL); Tmp = DAG.getNode(ISD::VP_OR, dl, VT, Tmp2, Tmp3, Mask, EVL); return Tmp; } return SDValue(); } std::pair TargetLowering::scalarizeVectorLoad(LoadSDNode *LD, SelectionDAG &DAG) const { SDLoc SL(LD); SDValue Chain = LD->getChain(); SDValue BasePTR = LD->getBasePtr(); EVT SrcVT = LD->getMemoryVT(); EVT DstVT = LD->getValueType(0); ISD::LoadExtType ExtType = LD->getExtensionType(); if (SrcVT.isScalableVector()) report_fatal_error("Cannot scalarize scalable vector loads"); unsigned NumElem = SrcVT.getVectorNumElements(); EVT SrcEltVT = SrcVT.getScalarType(); EVT DstEltVT = DstVT.getScalarType(); // A vector must always be stored in memory as-is, i.e. without any padding // between the elements, since various code depend on it, e.g. in the // handling of a bitcast of a vector type to int, which may be done with a // vector store followed by an integer load. A vector that does not have // elements that are byte-sized must therefore be stored as an integer // built out of the extracted vector elements. if (!SrcEltVT.isByteSized()) { unsigned NumLoadBits = SrcVT.getStoreSizeInBits(); EVT LoadVT = EVT::getIntegerVT(*DAG.getContext(), NumLoadBits); unsigned NumSrcBits = SrcVT.getSizeInBits(); EVT SrcIntVT = EVT::getIntegerVT(*DAG.getContext(), NumSrcBits); unsigned SrcEltBits = SrcEltVT.getSizeInBits(); SDValue SrcEltBitMask = DAG.getConstant( APInt::getLowBitsSet(NumLoadBits, SrcEltBits), SL, LoadVT); // Load the whole vector and avoid masking off the top bits as it makes // the codegen worse. SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, SL, LoadVT, Chain, BasePTR, LD->getPointerInfo(), SrcIntVT, LD->getOriginalAlign(), LD->getMemOperand()->getFlags(), LD->getAAInfo()); SmallVector Vals; for (unsigned Idx = 0; Idx < NumElem; ++Idx) { unsigned ShiftIntoIdx = (DAG.getDataLayout().isBigEndian() ? (NumElem - 1) - Idx : Idx); SDValue ShiftAmount = DAG.getShiftAmountConstant( ShiftIntoIdx * SrcEltVT.getSizeInBits(), LoadVT, SL); SDValue ShiftedElt = DAG.getNode(ISD::SRL, SL, LoadVT, Load, ShiftAmount); SDValue Elt = DAG.getNode(ISD::AND, SL, LoadVT, ShiftedElt, SrcEltBitMask); SDValue Scalar = DAG.getNode(ISD::TRUNCATE, SL, SrcEltVT, Elt); if (ExtType != ISD::NON_EXTLOAD) { unsigned ExtendOp = ISD::getExtForLoadExtType(false, ExtType); Scalar = DAG.getNode(ExtendOp, SL, DstEltVT, Scalar); } Vals.push_back(Scalar); } SDValue Value = DAG.getBuildVector(DstVT, SL, Vals); return std::make_pair(Value, Load.getValue(1)); } unsigned Stride = SrcEltVT.getSizeInBits() / 8; assert(SrcEltVT.isByteSized()); SmallVector Vals; SmallVector LoadChains; for (unsigned Idx = 0; Idx < NumElem; ++Idx) { SDValue ScalarLoad = DAG.getExtLoad(ExtType, SL, DstEltVT, Chain, BasePTR, LD->getPointerInfo().getWithOffset(Idx * Stride), SrcEltVT, LD->getOriginalAlign(), LD->getMemOperand()->getFlags(), LD->getAAInfo()); BasePTR = DAG.getObjectPtrOffset(SL, BasePTR, TypeSize::getFixed(Stride)); Vals.push_back(ScalarLoad.getValue(0)); LoadChains.push_back(ScalarLoad.getValue(1)); } SDValue NewChain = DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoadChains); SDValue Value = DAG.getBuildVector(DstVT, SL, Vals); return std::make_pair(Value, NewChain); } SDValue TargetLowering::scalarizeVectorStore(StoreSDNode *ST, SelectionDAG &DAG) const { SDLoc SL(ST); SDValue Chain = ST->getChain(); SDValue BasePtr = ST->getBasePtr(); SDValue Value = ST->getValue(); EVT StVT = ST->getMemoryVT(); if (StVT.isScalableVector()) report_fatal_error("Cannot scalarize scalable vector stores"); // The type of the data we want to save EVT RegVT = Value.getValueType(); EVT RegSclVT = RegVT.getScalarType(); // The type of data as saved in memory. EVT MemSclVT = StVT.getScalarType(); unsigned NumElem = StVT.getVectorNumElements(); // A vector must always be stored in memory as-is, i.e. without any padding // between the elements, since various code depend on it, e.g. in the // handling of a bitcast of a vector type to int, which may be done with a // vector store followed by an integer load. A vector that does not have // elements that are byte-sized must therefore be stored as an integer // built out of the extracted vector elements. if (!MemSclVT.isByteSized()) { unsigned NumBits = StVT.getSizeInBits(); EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), NumBits); SDValue CurrVal = DAG.getConstant(0, SL, IntVT); for (unsigned Idx = 0; Idx < NumElem; ++Idx) { SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, RegSclVT, Value, DAG.getVectorIdxConstant(Idx, SL)); SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, MemSclVT, Elt); SDValue ExtElt = DAG.getNode(ISD::ZERO_EXTEND, SL, IntVT, Trunc); unsigned ShiftIntoIdx = (DAG.getDataLayout().isBigEndian() ? (NumElem - 1) - Idx : Idx); SDValue ShiftAmount = DAG.getConstant(ShiftIntoIdx * MemSclVT.getSizeInBits(), SL, IntVT); SDValue ShiftedElt = DAG.getNode(ISD::SHL, SL, IntVT, ExtElt, ShiftAmount); CurrVal = DAG.getNode(ISD::OR, SL, IntVT, CurrVal, ShiftedElt); } return DAG.getStore(Chain, SL, CurrVal, BasePtr, ST->getPointerInfo(), ST->getOriginalAlign(), ST->getMemOperand()->getFlags(), ST->getAAInfo()); } // Store Stride in bytes unsigned Stride = MemSclVT.getSizeInBits() / 8; assert(Stride && "Zero stride!"); // Extract each of the elements from the original vector and save them into // memory individually. SmallVector Stores; for (unsigned Idx = 0; Idx < NumElem; ++Idx) { SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, RegSclVT, Value, DAG.getVectorIdxConstant(Idx, SL)); SDValue Ptr = DAG.getObjectPtrOffset(SL, BasePtr, TypeSize::getFixed(Idx * Stride)); // This scalar TruncStore may be illegal, but we legalize it later. SDValue Store = DAG.getTruncStore( Chain, SL, Elt, Ptr, ST->getPointerInfo().getWithOffset(Idx * Stride), MemSclVT, ST->getOriginalAlign(), ST->getMemOperand()->getFlags(), ST->getAAInfo()); Stores.push_back(Store); } return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Stores); } std::pair TargetLowering::expandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG) const { assert(LD->getAddressingMode() == ISD::UNINDEXED && "unaligned indexed loads not implemented!"); SDValue Chain = LD->getChain(); SDValue Ptr = LD->getBasePtr(); EVT VT = LD->getValueType(0); EVT LoadedVT = LD->getMemoryVT(); SDLoc dl(LD); auto &MF = DAG.getMachineFunction(); if (VT.isFloatingPoint() || VT.isVector()) { EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits()); if (isTypeLegal(intVT) && isTypeLegal(LoadedVT)) { if (!isOperationLegalOrCustom(ISD::LOAD, intVT) && LoadedVT.isVector()) { // Scalarize the load and let the individual components be handled. return scalarizeVectorLoad(LD, DAG); } // Expand to a (misaligned) integer load of the same size, // then bitconvert to floating point or vector. SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr, LD->getMemOperand()); SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad); if (LoadedVT != VT) Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND : ISD::ANY_EXTEND, dl, VT, Result); return std::make_pair(Result, newLoad.getValue(1)); } // Copy the value to a (aligned) stack slot using (unaligned) integer // loads and stores, then do a (aligned) load from the stack slot. MVT RegVT = getRegisterType(*DAG.getContext(), intVT); unsigned LoadedBytes = LoadedVT.getStoreSize(); unsigned RegBytes = RegVT.getSizeInBits() / 8; unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes; // Make sure the stack slot is also aligned for the register type. SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT); auto FrameIndex = cast(StackBase.getNode())->getIndex(); SmallVector Stores; SDValue StackPtr = StackBase; unsigned Offset = 0; EVT PtrVT = Ptr.getValueType(); EVT StackPtrVT = StackPtr.getValueType(); SDValue PtrIncrement = DAG.getConstant(RegBytes, dl, PtrVT); SDValue StackPtrIncrement = DAG.getConstant(RegBytes, dl, StackPtrVT); // Do all but one copies using the full register width. for (unsigned i = 1; i < NumRegs; i++) { // Load one integer register's worth from the original location. SDValue Load = DAG.getLoad( RegVT, dl, Chain, Ptr, LD->getPointerInfo().getWithOffset(Offset), LD->getOriginalAlign(), LD->getMemOperand()->getFlags(), LD->getAAInfo()); // Follow the load with a store to the stack slot. Remember the store. Stores.push_back(DAG.getStore( Load.getValue(1), dl, Load, StackPtr, MachinePointerInfo::getFixedStack(MF, FrameIndex, Offset))); // Increment the pointers. Offset += RegBytes; Ptr = DAG.getObjectPtrOffset(dl, Ptr, PtrIncrement); StackPtr = DAG.getObjectPtrOffset(dl, StackPtr, StackPtrIncrement); } // The last copy may be partial. Do an extending load. EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), 8 * (LoadedBytes - Offset)); SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr, LD->getPointerInfo().getWithOffset(Offset), MemVT, LD->getOriginalAlign(), LD->getMemOperand()->getFlags(), LD->getAAInfo()); // Follow the load with a store to the stack slot. Remember the store. // On big-endian machines this requires a truncating store to ensure // that the bits end up in the right place. Stores.push_back(DAG.getTruncStore( Load.getValue(1), dl, Load, StackPtr, MachinePointerInfo::getFixedStack(MF, FrameIndex, Offset), MemVT)); // The order of the stores doesn't matter - say it with a TokenFactor. SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores); // Finally, perform the original load only redirected to the stack slot. Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase, MachinePointerInfo::getFixedStack(MF, FrameIndex, 0), LoadedVT); // Callers expect a MERGE_VALUES node. return std::make_pair(Load, TF); } assert(LoadedVT.isInteger() && !LoadedVT.isVector() && "Unaligned load of unsupported type."); // Compute the new VT that is half the size of the old one. This is an // integer MVT. unsigned NumBits = LoadedVT.getSizeInBits(); EVT NewLoadedVT; NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2); NumBits >>= 1; Align Alignment = LD->getOriginalAlign(); unsigned IncrementSize = NumBits / 8; ISD::LoadExtType HiExtType = LD->getExtensionType(); // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD. if (HiExtType == ISD::NON_EXTLOAD) HiExtType = ISD::ZEXTLOAD; // Load the value in two parts SDValue Lo, Hi; if (DAG.getDataLayout().isLittleEndian()) { Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(), NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(), LD->getAAInfo()); Ptr = DAG.getObjectPtrOffset(dl, Ptr, TypeSize::getFixed(IncrementSize)); Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo().getWithOffset(IncrementSize), NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(), LD->getAAInfo()); } else { Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(), NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(), LD->getAAInfo()); Ptr = DAG.getObjectPtrOffset(dl, Ptr, TypeSize::getFixed(IncrementSize)); Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo().getWithOffset(IncrementSize), NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(), LD->getAAInfo()); } // aggregate the two parts SDValue ShiftAmount = DAG.getShiftAmountConstant(NumBits, VT, dl); SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount); Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo); SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1), Hi.getValue(1)); return std::make_pair(Result, TF); } SDValue TargetLowering::expandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG) const { assert(ST->getAddressingMode() == ISD::UNINDEXED && "unaligned indexed stores not implemented!"); SDValue Chain = ST->getChain(); SDValue Ptr = ST->getBasePtr(); SDValue Val = ST->getValue(); EVT VT = Val.getValueType(); Align Alignment = ST->getOriginalAlign(); auto &MF = DAG.getMachineFunction(); EVT StoreMemVT = ST->getMemoryVT(); SDLoc dl(ST); if (StoreMemVT.isFloatingPoint() || StoreMemVT.isVector()) { EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits()); if (isTypeLegal(intVT)) { if (!isOperationLegalOrCustom(ISD::STORE, intVT) && StoreMemVT.isVector()) { // Scalarize the store and let the individual components be handled. SDValue Result = scalarizeVectorStore(ST, DAG); return Result; } // Expand to a bitconvert of the value to the integer type of the // same size, then a (misaligned) int store. // FIXME: Does not handle truncating floating point stores! SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val); Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(), Alignment, ST->getMemOperand()->getFlags()); return Result; } // Do a (aligned) store to a stack slot, then copy from the stack slot // to the final destination using (unaligned) integer loads and stores. MVT RegVT = getRegisterType( *DAG.getContext(), EVT::getIntegerVT(*DAG.getContext(), StoreMemVT.getSizeInBits())); EVT PtrVT = Ptr.getValueType(); unsigned StoredBytes = StoreMemVT.getStoreSize(); unsigned RegBytes = RegVT.getSizeInBits() / 8; unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes; // Make sure the stack slot is also aligned for the register type. SDValue StackPtr = DAG.CreateStackTemporary(StoreMemVT, RegVT); auto FrameIndex = cast(StackPtr.getNode())->getIndex(); // Perform the original store, only redirected to the stack slot. SDValue Store = DAG.getTruncStore( Chain, dl, Val, StackPtr, MachinePointerInfo::getFixedStack(MF, FrameIndex, 0), StoreMemVT); EVT StackPtrVT = StackPtr.getValueType(); SDValue PtrIncrement = DAG.getConstant(RegBytes, dl, PtrVT); SDValue StackPtrIncrement = DAG.getConstant(RegBytes, dl, StackPtrVT); SmallVector Stores; unsigned Offset = 0; // Do all but one copies using the full register width. for (unsigned i = 1; i < NumRegs; i++) { // Load one integer register's worth from the stack slot. SDValue Load = DAG.getLoad( RegVT, dl, Store, StackPtr, MachinePointerInfo::getFixedStack(MF, FrameIndex, Offset)); // Store it to the final location. Remember the store. Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr, ST->getPointerInfo().getWithOffset(Offset), ST->getOriginalAlign(), ST->getMemOperand()->getFlags())); // Increment the pointers. Offset += RegBytes; StackPtr = DAG.getObjectPtrOffset(dl, StackPtr, StackPtrIncrement); Ptr = DAG.getObjectPtrOffset(dl, Ptr, PtrIncrement); } // The last store may be partial. Do a truncating store. On big-endian // machines this requires an extending load from the stack slot to ensure // that the bits are in the right place. EVT LoadMemVT = EVT::getIntegerVT(*DAG.getContext(), 8 * (StoredBytes - Offset)); // Load from the stack slot. SDValue Load = DAG.getExtLoad( ISD::EXTLOAD, dl, RegVT, Store, StackPtr, MachinePointerInfo::getFixedStack(MF, FrameIndex, Offset), LoadMemVT); Stores.push_back( DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr, ST->getPointerInfo().getWithOffset(Offset), LoadMemVT, ST->getOriginalAlign(), ST->getMemOperand()->getFlags(), ST->getAAInfo())); // The order of the stores doesn't matter - say it with a TokenFactor. SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores); return Result; } assert(StoreMemVT.isInteger() && !StoreMemVT.isVector() && "Unaligned store of unknown type."); // Get the half-size VT EVT NewStoredVT = StoreMemVT.getHalfSizedIntegerVT(*DAG.getContext()); unsigned NumBits = NewStoredVT.getFixedSizeInBits(); unsigned IncrementSize = NumBits / 8; // Divide the stored value in two parts. SDValue ShiftAmount = DAG.getShiftAmountConstant(NumBits, Val.getValueType(), dl); SDValue Lo = Val; // If Val is a constant, replace the upper bits with 0. The SRL will constant // fold and not use the upper bits. A smaller constant may be easier to // materialize. if (auto *C = dyn_cast(Lo); C && !C->isOpaque()) Lo = DAG.getNode( ISD::AND, dl, VT, Lo, DAG.getConstant(APInt::getLowBitsSet(VT.getSizeInBits(), NumBits), dl, VT)); SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount); // Store the two parts SDValue Store1, Store2; Store1 = DAG.getTruncStore(Chain, dl, DAG.getDataLayout().isLittleEndian() ? Lo : Hi, Ptr, ST->getPointerInfo(), NewStoredVT, Alignment, ST->getMemOperand()->getFlags()); Ptr = DAG.getObjectPtrOffset(dl, Ptr, TypeSize::getFixed(IncrementSize)); Store2 = DAG.getTruncStore( Chain, dl, DAG.getDataLayout().isLittleEndian() ? Hi : Lo, Ptr, ST->getPointerInfo().getWithOffset(IncrementSize), NewStoredVT, Alignment, ST->getMemOperand()->getFlags(), ST->getAAInfo()); SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2); return Result; } SDValue TargetLowering::IncrementMemoryAddress(SDValue Addr, SDValue Mask, const SDLoc &DL, EVT DataVT, SelectionDAG &DAG, bool IsCompressedMemory) const { SDValue Increment; EVT AddrVT = Addr.getValueType(); EVT MaskVT = Mask.getValueType(); assert(DataVT.getVectorElementCount() == MaskVT.getVectorElementCount() && "Incompatible types of Data and Mask"); if (IsCompressedMemory) { if (DataVT.isScalableVector()) report_fatal_error( "Cannot currently handle compressed memory with scalable vectors"); // Incrementing the pointer according to number of '1's in the mask. EVT MaskIntVT = EVT::getIntegerVT(*DAG.getContext(), MaskVT.getSizeInBits()); SDValue MaskInIntReg = DAG.getBitcast(MaskIntVT, Mask); if (MaskIntVT.getSizeInBits() < 32) { MaskInIntReg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, MaskInIntReg); MaskIntVT = MVT::i32; } // Count '1's with POPCNT. Increment = DAG.getNode(ISD::CTPOP, DL, MaskIntVT, MaskInIntReg); Increment = DAG.getZExtOrTrunc(Increment, DL, AddrVT); // Scale is an element size in bytes. SDValue Scale = DAG.getConstant(DataVT.getScalarSizeInBits() / 8, DL, AddrVT); Increment = DAG.getNode(ISD::MUL, DL, AddrVT, Increment, Scale); } else if (DataVT.isScalableVector()) { Increment = DAG.getVScale(DL, AddrVT, APInt(AddrVT.getFixedSizeInBits(), DataVT.getStoreSize().getKnownMinValue())); } else Increment = DAG.getConstant(DataVT.getStoreSize(), DL, AddrVT); return DAG.getNode(ISD::ADD, DL, AddrVT, Addr, Increment); } static SDValue clampDynamicVectorIndex(SelectionDAG &DAG, SDValue Idx, EVT VecVT, const SDLoc &dl, ElementCount SubEC) { assert(!(SubEC.isScalable() && VecVT.isFixedLengthVector()) && "Cannot index a scalable vector within a fixed-width vector"); unsigned NElts = VecVT.getVectorMinNumElements(); unsigned NumSubElts = SubEC.getKnownMinValue(); EVT IdxVT = Idx.getValueType(); if (VecVT.isScalableVector() && !SubEC.isScalable()) { // If this is a constant index and we know the value plus the number of the // elements in the subvector minus one is less than the minimum number of // elements then it's safe to return Idx. if (auto *IdxCst = dyn_cast(Idx)) if (IdxCst->getZExtValue() + (NumSubElts - 1) < NElts) return Idx; SDValue VS = DAG.getVScale(dl, IdxVT, APInt(IdxVT.getFixedSizeInBits(), NElts)); unsigned SubOpcode = NumSubElts <= NElts ? ISD::SUB : ISD::USUBSAT; SDValue Sub = DAG.getNode(SubOpcode, dl, IdxVT, VS, DAG.getConstant(NumSubElts, dl, IdxVT)); return DAG.getNode(ISD::UMIN, dl, IdxVT, Idx, Sub); } if (isPowerOf2_32(NElts) && NumSubElts == 1) { APInt Imm = APInt::getLowBitsSet(IdxVT.getSizeInBits(), Log2_32(NElts)); return DAG.getNode(ISD::AND, dl, IdxVT, Idx, DAG.getConstant(Imm, dl, IdxVT)); } unsigned MaxIndex = NumSubElts < NElts ? NElts - NumSubElts : 0; return DAG.getNode(ISD::UMIN, dl, IdxVT, Idx, DAG.getConstant(MaxIndex, dl, IdxVT)); } SDValue TargetLowering::getVectorElementPointer(SelectionDAG &DAG, SDValue VecPtr, EVT VecVT, SDValue Index) const { return getVectorSubVecPointer( DAG, VecPtr, VecVT, EVT::getVectorVT(*DAG.getContext(), VecVT.getVectorElementType(), 1), Index); } SDValue TargetLowering::getVectorSubVecPointer(SelectionDAG &DAG, SDValue VecPtr, EVT VecVT, EVT SubVecVT, SDValue Index) const { SDLoc dl(Index); // Make sure the index type is big enough to compute in. Index = DAG.getZExtOrTrunc(Index, dl, VecPtr.getValueType()); EVT EltVT = VecVT.getVectorElementType(); // Calculate the element offset and add it to the pointer. unsigned EltSize = EltVT.getFixedSizeInBits() / 8; // FIXME: should be ABI size. assert(EltSize * 8 == EltVT.getFixedSizeInBits() && "Converting bits to bytes lost precision"); assert(SubVecVT.getVectorElementType() == EltVT && "Sub-vector must be a vector with matching element type"); Index = clampDynamicVectorIndex(DAG, Index, VecVT, dl, SubVecVT.getVectorElementCount()); EVT IdxVT = Index.getValueType(); if (SubVecVT.isScalableVector()) Index = DAG.getNode(ISD::MUL, dl, IdxVT, Index, DAG.getVScale(dl, IdxVT, APInt(IdxVT.getSizeInBits(), 1))); Index = DAG.getNode(ISD::MUL, dl, IdxVT, Index, DAG.getConstant(EltSize, dl, IdxVT)); return DAG.getMemBasePlusOffset(VecPtr, Index, dl); } //===----------------------------------------------------------------------===// // Implementation of Emulated TLS Model //===----------------------------------------------------------------------===// SDValue TargetLowering::LowerToTLSEmulatedModel(const GlobalAddressSDNode *GA, SelectionDAG &DAG) const { // Access to address of TLS varialbe xyz is lowered to a function call: // __emutls_get_address( address of global variable named "__emutls_v.xyz" ) EVT PtrVT = getPointerTy(DAG.getDataLayout()); PointerType *VoidPtrType = PointerType::get(*DAG.getContext(), 0); SDLoc dl(GA); ArgListTy Args; ArgListEntry Entry; std::string NameString = ("__emutls_v." + GA->getGlobal()->getName()).str(); Module *VariableModule = const_cast(GA->getGlobal()->getParent()); StringRef EmuTlsVarName(NameString); GlobalVariable *EmuTlsVar = VariableModule->getNamedGlobal(EmuTlsVarName); assert(EmuTlsVar && "Cannot find EmuTlsVar "); Entry.Node = DAG.getGlobalAddress(EmuTlsVar, dl, PtrVT); Entry.Ty = VoidPtrType; Args.push_back(Entry); SDValue EmuTlsGetAddr = DAG.getExternalSymbol("__emutls_get_address", PtrVT); TargetLowering::CallLoweringInfo CLI(DAG); CLI.setDebugLoc(dl).setChain(DAG.getEntryNode()); CLI.setLibCallee(CallingConv::C, VoidPtrType, EmuTlsGetAddr, std::move(Args)); std::pair CallResult = LowerCallTo(CLI); // TLSADDR will be codegen'ed as call. Inform MFI that function has calls. // At last for X86 targets, maybe good for other targets too? MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); MFI.setAdjustsStack(true); // Is this only for X86 target? MFI.setHasCalls(true); assert((GA->getOffset() == 0) && "Emulated TLS must have zero offset in GlobalAddressSDNode"); return CallResult.first; } SDValue TargetLowering::lowerCmpEqZeroToCtlzSrl(SDValue Op, SelectionDAG &DAG) const { assert((Op->getOpcode() == ISD::SETCC) && "Input has to be a SETCC node."); if (!isCtlzFast()) return SDValue(); ISD::CondCode CC = cast(Op.getOperand(2))->get(); SDLoc dl(Op); if (isNullConstant(Op.getOperand(1)) && CC == ISD::SETEQ) { EVT VT = Op.getOperand(0).getValueType(); SDValue Zext = Op.getOperand(0); if (VT.bitsLT(MVT::i32)) { VT = MVT::i32; Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0)); } unsigned Log2b = Log2_32(VT.getSizeInBits()); SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext); SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz, DAG.getConstant(Log2b, dl, MVT::i32)); return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc); } return SDValue(); } SDValue TargetLowering::expandIntMINMAX(SDNode *Node, SelectionDAG &DAG) const { SDValue Op0 = Node->getOperand(0); SDValue Op1 = Node->getOperand(1); EVT VT = Op0.getValueType(); EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); unsigned Opcode = Node->getOpcode(); SDLoc DL(Node); // umax(x,1) --> sub(x,cmpeq(x,0)) iff cmp result is allbits if (Opcode == ISD::UMAX && llvm::isOneOrOneSplat(Op1, true) && BoolVT == VT && getBooleanContents(VT) == ZeroOrNegativeOneBooleanContent) { Op0 = DAG.getFreeze(Op0); SDValue Zero = DAG.getConstant(0, DL, VT); return DAG.getNode(ISD::SUB, DL, VT, Op0, DAG.getSetCC(DL, VT, Op0, Zero, ISD::SETEQ)); } // umin(x,y) -> sub(x,usubsat(x,y)) // TODO: Missing freeze(Op0)? if (Opcode == ISD::UMIN && isOperationLegal(ISD::SUB, VT) && isOperationLegal(ISD::USUBSAT, VT)) { return DAG.getNode(ISD::SUB, DL, VT, Op0, DAG.getNode(ISD::USUBSAT, DL, VT, Op0, Op1)); } // umax(x,y) -> add(x,usubsat(y,x)) // TODO: Missing freeze(Op0)? if (Opcode == ISD::UMAX && isOperationLegal(ISD::ADD, VT) && isOperationLegal(ISD::USUBSAT, VT)) { return DAG.getNode(ISD::ADD, DL, VT, Op0, DAG.getNode(ISD::USUBSAT, DL, VT, Op1, Op0)); } // FIXME: Should really try to split the vector in case it's legal on a // subvector. if (VT.isVector() && !isOperationLegalOrCustom(ISD::VSELECT, VT)) return DAG.UnrollVectorOp(Node); // Attempt to find an existing SETCC node that we can reuse. // TODO: Do we need a generic doesSETCCNodeExist? // TODO: Missing freeze(Op0)/freeze(Op1)? auto buildMinMax = [&](ISD::CondCode PrefCC, ISD::CondCode AltCC, ISD::CondCode PrefCommuteCC, ISD::CondCode AltCommuteCC) { SDVTList BoolVTList = DAG.getVTList(BoolVT); for (ISD::CondCode CC : {PrefCC, AltCC}) { if (DAG.doesNodeExist(ISD::SETCC, BoolVTList, {Op0, Op1, DAG.getCondCode(CC)})) { SDValue Cond = DAG.getSetCC(DL, BoolVT, Op0, Op1, CC); return DAG.getSelect(DL, VT, Cond, Op0, Op1); } } for (ISD::CondCode CC : {PrefCommuteCC, AltCommuteCC}) { if (DAG.doesNodeExist(ISD::SETCC, BoolVTList, {Op0, Op1, DAG.getCondCode(CC)})) { SDValue Cond = DAG.getSetCC(DL, BoolVT, Op0, Op1, CC); return DAG.getSelect(DL, VT, Cond, Op1, Op0); } } SDValue Cond = DAG.getSetCC(DL, BoolVT, Op0, Op1, PrefCC); return DAG.getSelect(DL, VT, Cond, Op0, Op1); }; // Expand Y = MAX(A, B) -> Y = (A > B) ? A : B // -> Y = (A < B) ? B : A // -> Y = (A >= B) ? A : B // -> Y = (A <= B) ? B : A switch (Opcode) { case ISD::SMAX: return buildMinMax(ISD::SETGT, ISD::SETGE, ISD::SETLT, ISD::SETLE); case ISD::SMIN: return buildMinMax(ISD::SETLT, ISD::SETLE, ISD::SETGT, ISD::SETGE); case ISD::UMAX: return buildMinMax(ISD::SETUGT, ISD::SETUGE, ISD::SETULT, ISD::SETULE); case ISD::UMIN: return buildMinMax(ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE); } llvm_unreachable("How did we get here?"); } SDValue TargetLowering::expandAddSubSat(SDNode *Node, SelectionDAG &DAG) const { unsigned Opcode = Node->getOpcode(); SDValue LHS = Node->getOperand(0); SDValue RHS = Node->getOperand(1); EVT VT = LHS.getValueType(); SDLoc dl(Node); assert(VT == RHS.getValueType() && "Expected operands to be the same type"); assert(VT.isInteger() && "Expected operands to be integers"); // usub.sat(a, b) -> umax(a, b) - b if (Opcode == ISD::USUBSAT && isOperationLegal(ISD::UMAX, VT)) { SDValue Max = DAG.getNode(ISD::UMAX, dl, VT, LHS, RHS); return DAG.getNode(ISD::SUB, dl, VT, Max, RHS); } // uadd.sat(a, b) -> umin(a, ~b) + b if (Opcode == ISD::UADDSAT && isOperationLegal(ISD::UMIN, VT)) { SDValue InvRHS = DAG.getNOT(dl, RHS, VT); SDValue Min = DAG.getNode(ISD::UMIN, dl, VT, LHS, InvRHS); return DAG.getNode(ISD::ADD, dl, VT, Min, RHS); } unsigned OverflowOp; switch (Opcode) { case ISD::SADDSAT: OverflowOp = ISD::SADDO; break; case ISD::UADDSAT: OverflowOp = ISD::UADDO; break; case ISD::SSUBSAT: OverflowOp = ISD::SSUBO; break; case ISD::USUBSAT: OverflowOp = ISD::USUBO; break; default: llvm_unreachable("Expected method to receive signed or unsigned saturation " "addition or subtraction node."); } // FIXME: Should really try to split the vector in case it's legal on a // subvector. if (VT.isVector() && !isOperationLegalOrCustom(ISD::VSELECT, VT)) return DAG.UnrollVectorOp(Node); unsigned BitWidth = LHS.getScalarValueSizeInBits(); EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDValue Result = DAG.getNode(OverflowOp, dl, DAG.getVTList(VT, BoolVT), LHS, RHS); SDValue SumDiff = Result.getValue(0); SDValue Overflow = Result.getValue(1); SDValue Zero = DAG.getConstant(0, dl, VT); SDValue AllOnes = DAG.getAllOnesConstant(dl, VT); if (Opcode == ISD::UADDSAT) { if (getBooleanContents(VT) == ZeroOrNegativeOneBooleanContent) { // (LHS + RHS) | OverflowMask SDValue OverflowMask = DAG.getSExtOrTrunc(Overflow, dl, VT); return DAG.getNode(ISD::OR, dl, VT, SumDiff, OverflowMask); } // Overflow ? 0xffff.... : (LHS + RHS) return DAG.getSelect(dl, VT, Overflow, AllOnes, SumDiff); } if (Opcode == ISD::USUBSAT) { if (getBooleanContents(VT) == ZeroOrNegativeOneBooleanContent) { // (LHS - RHS) & ~OverflowMask SDValue OverflowMask = DAG.getSExtOrTrunc(Overflow, dl, VT); SDValue Not = DAG.getNOT(dl, OverflowMask, VT); return DAG.getNode(ISD::AND, dl, VT, SumDiff, Not); } // Overflow ? 0 : (LHS - RHS) return DAG.getSelect(dl, VT, Overflow, Zero, SumDiff); } if (Opcode == ISD::SADDSAT || Opcode == ISD::SSUBSAT) { APInt MinVal = APInt::getSignedMinValue(BitWidth); APInt MaxVal = APInt::getSignedMaxValue(BitWidth); KnownBits KnownLHS = DAG.computeKnownBits(LHS); KnownBits KnownRHS = DAG.computeKnownBits(RHS); // If either of the operand signs are known, then they are guaranteed to // only saturate in one direction. If non-negative they will saturate // towards SIGNED_MAX, if negative they will saturate towards SIGNED_MIN. // // In the case of ISD::SSUBSAT, 'x - y' is equivalent to 'x + (-y)', so the // sign of 'y' has to be flipped. bool LHSIsNonNegative = KnownLHS.isNonNegative(); bool RHSIsNonNegative = Opcode == ISD::SADDSAT ? KnownRHS.isNonNegative() : KnownRHS.isNegative(); if (LHSIsNonNegative || RHSIsNonNegative) { SDValue SatMax = DAG.getConstant(MaxVal, dl, VT); return DAG.getSelect(dl, VT, Overflow, SatMax, SumDiff); } bool LHSIsNegative = KnownLHS.isNegative(); bool RHSIsNegative = Opcode == ISD::SADDSAT ? KnownRHS.isNegative() : KnownRHS.isNonNegative(); if (LHSIsNegative || RHSIsNegative) { SDValue SatMin = DAG.getConstant(MinVal, dl, VT); return DAG.getSelect(dl, VT, Overflow, SatMin, SumDiff); } } // Overflow ? (SumDiff >> BW) ^ MinVal : SumDiff APInt MinVal = APInt::getSignedMinValue(BitWidth); SDValue SatMin = DAG.getConstant(MinVal, dl, VT); SDValue Shift = DAG.getNode(ISD::SRA, dl, VT, SumDiff, DAG.getConstant(BitWidth - 1, dl, VT)); Result = DAG.getNode(ISD::XOR, dl, VT, Shift, SatMin); return DAG.getSelect(dl, VT, Overflow, Result, SumDiff); } SDValue TargetLowering::expandCMP(SDNode *Node, SelectionDAG &DAG) const { unsigned Opcode = Node->getOpcode(); SDValue LHS = Node->getOperand(0); SDValue RHS = Node->getOperand(1); EVT VT = LHS.getValueType(); EVT ResVT = Node->getValueType(0); EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDLoc dl(Node); auto LTPredicate = (Opcode == ISD::UCMP ? ISD::SETULT : ISD::SETLT); auto GTPredicate = (Opcode == ISD::UCMP ? ISD::SETUGT : ISD::SETGT); SDValue IsLT = DAG.getSetCC(dl, BoolVT, LHS, RHS, LTPredicate); SDValue IsGT = DAG.getSetCC(dl, BoolVT, LHS, RHS, GTPredicate); // We can't perform arithmetic on i1 values. Extending them would // probably result in worse codegen, so let's just use two selects instead. // Some targets are also just better off using selects rather than subtraction // because one of the conditions can be merged with one of the selects. // And finally, if we don't know the contents of high bits of a boolean value // we can't perform any arithmetic either. if (shouldExpandCmpUsingSelects() || BoolVT.getScalarSizeInBits() == 1 || getBooleanContents(BoolVT) == UndefinedBooleanContent) { SDValue SelectZeroOrOne = DAG.getSelect(dl, ResVT, IsGT, DAG.getConstant(1, dl, ResVT), DAG.getConstant(0, dl, ResVT)); return DAG.getSelect(dl, ResVT, IsLT, DAG.getConstant(-1, dl, ResVT), SelectZeroOrOne); } if (getBooleanContents(BoolVT) == ZeroOrNegativeOneBooleanContent) std::swap(IsGT, IsLT); return DAG.getSExtOrTrunc(DAG.getNode(ISD::SUB, dl, BoolVT, IsGT, IsLT), dl, ResVT); } SDValue TargetLowering::expandShlSat(SDNode *Node, SelectionDAG &DAG) const { unsigned Opcode = Node->getOpcode(); bool IsSigned = Opcode == ISD::SSHLSAT; SDValue LHS = Node->getOperand(0); SDValue RHS = Node->getOperand(1); EVT VT = LHS.getValueType(); SDLoc dl(Node); assert((Node->getOpcode() == ISD::SSHLSAT || Node->getOpcode() == ISD::USHLSAT) && "Expected a SHLSAT opcode"); assert(VT == RHS.getValueType() && "Expected operands to be the same type"); assert(VT.isInteger() && "Expected operands to be integers"); if (VT.isVector() && !isOperationLegalOrCustom(ISD::VSELECT, VT)) return DAG.UnrollVectorOp(Node); // If LHS != (LHS << RHS) >> RHS, we have overflow and must saturate. unsigned BW = VT.getScalarSizeInBits(); EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDValue Result = DAG.getNode(ISD::SHL, dl, VT, LHS, RHS); SDValue Orig = DAG.getNode(IsSigned ? ISD::SRA : ISD::SRL, dl, VT, Result, RHS); SDValue SatVal; if (IsSigned) { SDValue SatMin = DAG.getConstant(APInt::getSignedMinValue(BW), dl, VT); SDValue SatMax = DAG.getConstant(APInt::getSignedMaxValue(BW), dl, VT); SDValue Cond = DAG.getSetCC(dl, BoolVT, LHS, DAG.getConstant(0, dl, VT), ISD::SETLT); SatVal = DAG.getSelect(dl, VT, Cond, SatMin, SatMax); } else { SatVal = DAG.getConstant(APInt::getMaxValue(BW), dl, VT); } SDValue Cond = DAG.getSetCC(dl, BoolVT, LHS, Orig, ISD::SETNE); return DAG.getSelect(dl, VT, Cond, SatVal, Result); } void TargetLowering::forceExpandWideMUL(SelectionDAG &DAG, const SDLoc &dl, bool Signed, EVT WideVT, const SDValue LL, const SDValue LH, const SDValue RL, const SDValue RH, SDValue &Lo, SDValue &Hi) const { // We can fall back to a libcall with an illegal type for the MUL if we // have a libcall big enough. // Also, we can fall back to a division in some cases, but that's a big // performance hit in the general case. RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL; if (WideVT == MVT::i16) LC = RTLIB::MUL_I16; else if (WideVT == MVT::i32) LC = RTLIB::MUL_I32; else if (WideVT == MVT::i64) LC = RTLIB::MUL_I64; else if (WideVT == MVT::i128) LC = RTLIB::MUL_I128; if (LC == RTLIB::UNKNOWN_LIBCALL || !getLibcallName(LC)) { // We'll expand the multiplication by brute force because we have no other // options. This is a trivially-generalized version of the code from // Hacker's Delight (itself derived from Knuth's Algorithm M from section // 4.3.1). EVT VT = LL.getValueType(); unsigned Bits = VT.getSizeInBits(); unsigned HalfBits = Bits >> 1; SDValue Mask = DAG.getConstant(APInt::getLowBitsSet(Bits, HalfBits), dl, VT); SDValue LLL = DAG.getNode(ISD::AND, dl, VT, LL, Mask); SDValue RLL = DAG.getNode(ISD::AND, dl, VT, RL, Mask); SDValue T = DAG.getNode(ISD::MUL, dl, VT, LLL, RLL); SDValue TL = DAG.getNode(ISD::AND, dl, VT, T, Mask); SDValue Shift = DAG.getShiftAmountConstant(HalfBits, VT, dl); SDValue TH = DAG.getNode(ISD::SRL, dl, VT, T, Shift); SDValue LLH = DAG.getNode(ISD::SRL, dl, VT, LL, Shift); SDValue RLH = DAG.getNode(ISD::SRL, dl, VT, RL, Shift); SDValue U = DAG.getNode(ISD::ADD, dl, VT, DAG.getNode(ISD::MUL, dl, VT, LLH, RLL), TH); SDValue UL = DAG.getNode(ISD::AND, dl, VT, U, Mask); SDValue UH = DAG.getNode(ISD::SRL, dl, VT, U, Shift); SDValue V = DAG.getNode(ISD::ADD, dl, VT, DAG.getNode(ISD::MUL, dl, VT, LLL, RLH), UL); SDValue VH = DAG.getNode(ISD::SRL, dl, VT, V, Shift); SDValue W = DAG.getNode(ISD::ADD, dl, VT, DAG.getNode(ISD::MUL, dl, VT, LLH, RLH), DAG.getNode(ISD::ADD, dl, VT, UH, VH)); Lo = DAG.getNode(ISD::ADD, dl, VT, TL, DAG.getNode(ISD::SHL, dl, VT, V, Shift)); Hi = DAG.getNode(ISD::ADD, dl, VT, W, DAG.getNode(ISD::ADD, dl, VT, DAG.getNode(ISD::MUL, dl, VT, RH, LL), DAG.getNode(ISD::MUL, dl, VT, RL, LH))); } else { // Attempt a libcall. SDValue Ret; TargetLowering::MakeLibCallOptions CallOptions; CallOptions.setSExt(Signed); CallOptions.setIsPostTypeLegalization(true); if (shouldSplitFunctionArgumentsAsLittleEndian(DAG.getDataLayout())) { // Halves of WideVT are packed into registers in different order // depending on platform endianness. This is usually handled by // the C calling convention, but we can't defer to it in // the legalizer. SDValue Args[] = {LL, LH, RL, RH}; Ret = makeLibCall(DAG, LC, WideVT, Args, CallOptions, dl).first; } else { SDValue Args[] = {LH, LL, RH, RL}; Ret = makeLibCall(DAG, LC, WideVT, Args, CallOptions, dl).first; } assert(Ret.getOpcode() == ISD::MERGE_VALUES && "Ret value is a collection of constituent nodes holding result."); if (DAG.getDataLayout().isLittleEndian()) { // Same as above. Lo = Ret.getOperand(0); Hi = Ret.getOperand(1); } else { Lo = Ret.getOperand(1); Hi = Ret.getOperand(0); } } } void TargetLowering::forceExpandWideMUL(SelectionDAG &DAG, const SDLoc &dl, bool Signed, const SDValue LHS, const SDValue RHS, SDValue &Lo, SDValue &Hi) const { EVT VT = LHS.getValueType(); assert(RHS.getValueType() == VT && "Mismatching operand types"); SDValue HiLHS; SDValue HiRHS; if (Signed) { // The high part is obtained by SRA'ing all but one of the bits of low // part. unsigned LoSize = VT.getFixedSizeInBits(); HiLHS = DAG.getNode( ISD::SRA, dl, VT, LHS, DAG.getConstant(LoSize - 1, dl, getPointerTy(DAG.getDataLayout()))); HiRHS = DAG.getNode( ISD::SRA, dl, VT, RHS, DAG.getConstant(LoSize - 1, dl, getPointerTy(DAG.getDataLayout()))); } else { HiLHS = DAG.getConstant(0, dl, VT); HiRHS = DAG.getConstant(0, dl, VT); } EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2); forceExpandWideMUL(DAG, dl, Signed, WideVT, LHS, HiLHS, RHS, HiRHS, Lo, Hi); } SDValue TargetLowering::expandFixedPointMul(SDNode *Node, SelectionDAG &DAG) const { assert((Node->getOpcode() == ISD::SMULFIX || Node->getOpcode() == ISD::UMULFIX || Node->getOpcode() == ISD::SMULFIXSAT || Node->getOpcode() == ISD::UMULFIXSAT) && "Expected a fixed point multiplication opcode"); SDLoc dl(Node); SDValue LHS = Node->getOperand(0); SDValue RHS = Node->getOperand(1); EVT VT = LHS.getValueType(); unsigned Scale = Node->getConstantOperandVal(2); bool Saturating = (Node->getOpcode() == ISD::SMULFIXSAT || Node->getOpcode() == ISD::UMULFIXSAT); bool Signed = (Node->getOpcode() == ISD::SMULFIX || Node->getOpcode() == ISD::SMULFIXSAT); EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); unsigned VTSize = VT.getScalarSizeInBits(); if (!Scale) { // [us]mul.fix(a, b, 0) -> mul(a, b) if (!Saturating) { if (isOperationLegalOrCustom(ISD::MUL, VT)) return DAG.getNode(ISD::MUL, dl, VT, LHS, RHS); } else if (Signed && isOperationLegalOrCustom(ISD::SMULO, VT)) { SDValue Result = DAG.getNode(ISD::SMULO, dl, DAG.getVTList(VT, BoolVT), LHS, RHS); SDValue Product = Result.getValue(0); SDValue Overflow = Result.getValue(1); SDValue Zero = DAG.getConstant(0, dl, VT); APInt MinVal = APInt::getSignedMinValue(VTSize); APInt MaxVal = APInt::getSignedMaxValue(VTSize); SDValue SatMin = DAG.getConstant(MinVal, dl, VT); SDValue SatMax = DAG.getConstant(MaxVal, dl, VT); // Xor the inputs, if resulting sign bit is 0 the product will be // positive, else negative. SDValue Xor = DAG.getNode(ISD::XOR, dl, VT, LHS, RHS); SDValue ProdNeg = DAG.getSetCC(dl, BoolVT, Xor, Zero, ISD::SETLT); Result = DAG.getSelect(dl, VT, ProdNeg, SatMin, SatMax); return DAG.getSelect(dl, VT, Overflow, Result, Product); } else if (!Signed && isOperationLegalOrCustom(ISD::UMULO, VT)) { SDValue Result = DAG.getNode(ISD::UMULO, dl, DAG.getVTList(VT, BoolVT), LHS, RHS); SDValue Product = Result.getValue(0); SDValue Overflow = Result.getValue(1); APInt MaxVal = APInt::getMaxValue(VTSize); SDValue SatMax = DAG.getConstant(MaxVal, dl, VT); return DAG.getSelect(dl, VT, Overflow, SatMax, Product); } } assert(((Signed && Scale < VTSize) || (!Signed && Scale <= VTSize)) && "Expected scale to be less than the number of bits if signed or at " "most the number of bits if unsigned."); assert(LHS.getValueType() == RHS.getValueType() && "Expected both operands to be the same type"); // Get the upper and lower bits of the result. SDValue Lo, Hi; unsigned LoHiOp = Signed ? ISD::SMUL_LOHI : ISD::UMUL_LOHI; unsigned HiOp = Signed ? ISD::MULHS : ISD::MULHU; EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VTSize * 2); if (isOperationLegalOrCustom(LoHiOp, VT)) { SDValue Result = DAG.getNode(LoHiOp, dl, DAG.getVTList(VT, VT), LHS, RHS); Lo = Result.getValue(0); Hi = Result.getValue(1); } else if (isOperationLegalOrCustom(HiOp, VT)) { Lo = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS); Hi = DAG.getNode(HiOp, dl, VT, LHS, RHS); } else if (isOperationLegalOrCustom(ISD::MUL, WideVT)) { // Try for a multiplication using a wider type. unsigned Ext = Signed ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; SDValue LHSExt = DAG.getNode(Ext, dl, WideVT, LHS); SDValue RHSExt = DAG.getNode(Ext, dl, WideVT, RHS); SDValue Res = DAG.getNode(ISD::MUL, dl, WideVT, LHSExt, RHSExt); Lo = DAG.getNode(ISD::TRUNCATE, dl, VT, Res); SDValue Shifted = DAG.getNode(ISD::SRA, dl, WideVT, Res, DAG.getShiftAmountConstant(VTSize, WideVT, dl)); Hi = DAG.getNode(ISD::TRUNCATE, dl, VT, Shifted); } else if (VT.isVector()) { return SDValue(); } else { forceExpandWideMUL(DAG, dl, Signed, LHS, RHS, Lo, Hi); } if (Scale == VTSize) // Result is just the top half since we'd be shifting by the width of the // operand. Overflow impossible so this works for both UMULFIX and // UMULFIXSAT. return Hi; // The result will need to be shifted right by the scale since both operands // are scaled. The result is given to us in 2 halves, so we only want part of // both in the result. SDValue Result = DAG.getNode(ISD::FSHR, dl, VT, Hi, Lo, DAG.getShiftAmountConstant(Scale, VT, dl)); if (!Saturating) return Result; if (!Signed) { // Unsigned overflow happened if the upper (VTSize - Scale) bits (of the // widened multiplication) aren't all zeroes. // Saturate to max if ((Hi >> Scale) != 0), // which is the same as if (Hi > ((1 << Scale) - 1)) APInt MaxVal = APInt::getMaxValue(VTSize); SDValue LowMask = DAG.getConstant(APInt::getLowBitsSet(VTSize, Scale), dl, VT); Result = DAG.getSelectCC(dl, Hi, LowMask, DAG.getConstant(MaxVal, dl, VT), Result, ISD::SETUGT); return Result; } // Signed overflow happened if the upper (VTSize - Scale + 1) bits (of the // widened multiplication) aren't all ones or all zeroes. SDValue SatMin = DAG.getConstant(APInt::getSignedMinValue(VTSize), dl, VT); SDValue SatMax = DAG.getConstant(APInt::getSignedMaxValue(VTSize), dl, VT); if (Scale == 0) { SDValue Sign = DAG.getNode(ISD::SRA, dl, VT, Lo, DAG.getShiftAmountConstant(VTSize - 1, VT, dl)); SDValue Overflow = DAG.getSetCC(dl, BoolVT, Hi, Sign, ISD::SETNE); // Saturated to SatMin if wide product is negative, and SatMax if wide // product is positive ... SDValue Zero = DAG.getConstant(0, dl, VT); SDValue ResultIfOverflow = DAG.getSelectCC(dl, Hi, Zero, SatMin, SatMax, ISD::SETLT); // ... but only if we overflowed. return DAG.getSelect(dl, VT, Overflow, ResultIfOverflow, Result); } // We handled Scale==0 above so all the bits to examine is in Hi. // Saturate to max if ((Hi >> (Scale - 1)) > 0), // which is the same as if (Hi > (1 << (Scale - 1)) - 1) SDValue LowMask = DAG.getConstant(APInt::getLowBitsSet(VTSize, Scale - 1), dl, VT); Result = DAG.getSelectCC(dl, Hi, LowMask, SatMax, Result, ISD::SETGT); // Saturate to min if (Hi >> (Scale - 1)) < -1), // which is the same as if (HI < (-1 << (Scale - 1)) SDValue HighMask = DAG.getConstant(APInt::getHighBitsSet(VTSize, VTSize - Scale + 1), dl, VT); Result = DAG.getSelectCC(dl, Hi, HighMask, SatMin, Result, ISD::SETLT); return Result; } SDValue TargetLowering::expandFixedPointDiv(unsigned Opcode, const SDLoc &dl, SDValue LHS, SDValue RHS, unsigned Scale, SelectionDAG &DAG) const { assert((Opcode == ISD::SDIVFIX || Opcode == ISD::SDIVFIXSAT || Opcode == ISD::UDIVFIX || Opcode == ISD::UDIVFIXSAT) && "Expected a fixed point division opcode"); EVT VT = LHS.getValueType(); bool Signed = Opcode == ISD::SDIVFIX || Opcode == ISD::SDIVFIXSAT; bool Saturating = Opcode == ISD::SDIVFIXSAT || Opcode == ISD::UDIVFIXSAT; EVT BoolVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); // If there is enough room in the type to upscale the LHS or downscale the // RHS before the division, we can perform it in this type without having to // resize. For signed operations, the LHS headroom is the number of // redundant sign bits, and for unsigned ones it is the number of zeroes. // The headroom for the RHS is the number of trailing zeroes. unsigned LHSLead = Signed ? DAG.ComputeNumSignBits(LHS) - 1 : DAG.computeKnownBits(LHS).countMinLeadingZeros(); unsigned RHSTrail = DAG.computeKnownBits(RHS).countMinTrailingZeros(); // For signed saturating operations, we need to be able to detect true integer // division overflow; that is, when you have MIN / -EPS. However, this // is undefined behavior and if we emit divisions that could take such // values it may cause undesired behavior (arithmetic exceptions on x86, for // example). // Avoid this by requiring an extra bit so that we never get this case. // FIXME: This is a bit unfortunate as it means that for an 8-bit 7-scale // signed saturating division, we need to emit a whopping 32-bit division. if (LHSLead + RHSTrail < Scale + (unsigned)(Saturating && Signed)) return SDValue(); unsigned LHSShift = std::min(LHSLead, Scale); unsigned RHSShift = Scale - LHSShift; // At this point, we know that if we shift the LHS up by LHSShift and the // RHS down by RHSShift, we can emit a regular division with a final scaling // factor of Scale. if (LHSShift) LHS = DAG.getNode(ISD::SHL, dl, VT, LHS, DAG.getShiftAmountConstant(LHSShift, VT, dl)); if (RHSShift) RHS = DAG.getNode(Signed ? ISD::SRA : ISD::SRL, dl, VT, RHS, DAG.getShiftAmountConstant(RHSShift, VT, dl)); SDValue Quot; if (Signed) { // For signed operations, if the resulting quotient is negative and the // remainder is nonzero, subtract 1 from the quotient to round towards // negative infinity. SDValue Rem; // FIXME: Ideally we would always produce an SDIVREM here, but if the // type isn't legal, SDIVREM cannot be expanded. There is no reason why // we couldn't just form a libcall, but the type legalizer doesn't do it. if (isTypeLegal(VT) && isOperationLegalOrCustom(ISD::SDIVREM, VT)) { Quot = DAG.getNode(ISD::SDIVREM, dl, DAG.getVTList(VT, VT), LHS, RHS); Rem = Quot.getValue(1); Quot = Quot.getValue(0); } else { Quot = DAG.getNode(ISD::SDIV, dl, VT, LHS, RHS); Rem = DAG.getNode(ISD::SREM, dl, VT, LHS, RHS); } SDValue Zero = DAG.getConstant(0, dl, VT); SDValue RemNonZero = DAG.getSetCC(dl, BoolVT, Rem, Zero, ISD::SETNE); SDValue LHSNeg = DAG.getSetCC(dl, BoolVT, LHS, Zero, ISD::SETLT); SDValue RHSNeg = DAG.getSetCC(dl, BoolVT, RHS, Zero, ISD::SETLT); SDValue QuotNeg = DAG.getNode(ISD::XOR, dl, BoolVT, LHSNeg, RHSNeg); SDValue Sub1 = DAG.getNode(ISD::SUB, dl, VT, Quot, DAG.getConstant(1, dl, VT)); Quot = DAG.getSelect(dl, VT, DAG.getNode(ISD::AND, dl, BoolVT, RemNonZero, QuotNeg), Sub1, Quot); } else Quot = DAG.getNode(ISD::UDIV, dl, VT, LHS, RHS); return Quot; } void TargetLowering::expandUADDSUBO( SDNode *Node, SDValue &Result, SDValue &Overflow, SelectionDAG &DAG) const { SDLoc dl(Node); SDValue LHS = Node->getOperand(0); SDValue RHS = Node->getOperand(1); bool IsAdd = Node->getOpcode() == ISD::UADDO; // If UADDO_CARRY/SUBO_CARRY is legal, use that instead. unsigned OpcCarry = IsAdd ? ISD::UADDO_CARRY : ISD::USUBO_CARRY; if (isOperationLegalOrCustom(OpcCarry, Node->getValueType(0))) { SDValue CarryIn = DAG.getConstant(0, dl, Node->getValueType(1)); SDValue NodeCarry = DAG.getNode(OpcCarry, dl, Node->getVTList(), { LHS, RHS, CarryIn }); Result = SDValue(NodeCarry.getNode(), 0); Overflow = SDValue(NodeCarry.getNode(), 1); return; } Result = DAG.getNode(IsAdd ? ISD::ADD : ISD::SUB, dl, LHS.getValueType(), LHS, RHS); EVT ResultType = Node->getValueType(1); EVT SetCCType = getSetCCResultType( DAG.getDataLayout(), *DAG.getContext(), Node->getValueType(0)); SDValue SetCC; if (IsAdd && isOneConstant(RHS)) { // Special case: uaddo X, 1 overflowed if X+1 is 0. This potential reduces // the live range of X. We assume comparing with 0 is cheap. // The general case (X + C) < C is not necessarily beneficial. Although we // reduce the live range of X, we may introduce the materialization of // constant C. SetCC = DAG.getSetCC(dl, SetCCType, Result, DAG.getConstant(0, dl, Node->getValueType(0)), ISD::SETEQ); } else if (IsAdd && isAllOnesConstant(RHS)) { // Special case: uaddo X, -1 overflows if X != 0. SetCC = DAG.getSetCC(dl, SetCCType, LHS, DAG.getConstant(0, dl, Node->getValueType(0)), ISD::SETNE); } else { ISD::CondCode CC = IsAdd ? ISD::SETULT : ISD::SETUGT; SetCC = DAG.getSetCC(dl, SetCCType, Result, LHS, CC); } Overflow = DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType); } void TargetLowering::expandSADDSUBO( SDNode *Node, SDValue &Result, SDValue &Overflow, SelectionDAG &DAG) const { SDLoc dl(Node); SDValue LHS = Node->getOperand(0); SDValue RHS = Node->getOperand(1); bool IsAdd = Node->getOpcode() == ISD::SADDO; Result = DAG.getNode(IsAdd ? ISD::ADD : ISD::SUB, dl, LHS.getValueType(), LHS, RHS); EVT ResultType = Node->getValueType(1); EVT OType = getSetCCResultType( DAG.getDataLayout(), *DAG.getContext(), Node->getValueType(0)); // If SADDSAT/SSUBSAT is legal, compare results to detect overflow. unsigned OpcSat = IsAdd ? ISD::SADDSAT : ISD::SSUBSAT; if (isOperationLegal(OpcSat, LHS.getValueType())) { SDValue Sat = DAG.getNode(OpcSat, dl, LHS.getValueType(), LHS, RHS); SDValue SetCC = DAG.getSetCC(dl, OType, Result, Sat, ISD::SETNE); Overflow = DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType); return; } SDValue Zero = DAG.getConstant(0, dl, LHS.getValueType()); // For an addition, the result should be less than one of the operands (LHS) // if and only if the other operand (RHS) is negative, otherwise there will // be overflow. // For a subtraction, the result should be less than one of the operands // (LHS) if and only if the other operand (RHS) is (non-zero) positive, // otherwise there will be overflow. SDValue ResultLowerThanLHS = DAG.getSetCC(dl, OType, Result, LHS, ISD::SETLT); SDValue ConditionRHS = DAG.getSetCC(dl, OType, RHS, Zero, IsAdd ? ISD::SETLT : ISD::SETGT); Overflow = DAG.getBoolExtOrTrunc( DAG.getNode(ISD::XOR, dl, OType, ConditionRHS, ResultLowerThanLHS), dl, ResultType, ResultType); } bool TargetLowering::expandMULO(SDNode *Node, SDValue &Result, SDValue &Overflow, SelectionDAG &DAG) const { SDLoc dl(Node); EVT VT = Node->getValueType(0); EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); SDValue LHS = Node->getOperand(0); SDValue RHS = Node->getOperand(1); bool isSigned = Node->getOpcode() == ISD::SMULO; // For power-of-two multiplications we can use a simpler shift expansion. if (ConstantSDNode *RHSC = isConstOrConstSplat(RHS)) { const APInt &C = RHSC->getAPIntValue(); // mulo(X, 1 << S) -> { X << S, (X << S) >> S != X } if (C.isPowerOf2()) { // smulo(x, signed_min) is same as umulo(x, signed_min). bool UseArithShift = isSigned && !C.isMinSignedValue(); SDValue ShiftAmt = DAG.getShiftAmountConstant(C.logBase2(), VT, dl); Result = DAG.getNode(ISD::SHL, dl, VT, LHS, ShiftAmt); Overflow = DAG.getSetCC(dl, SetCCVT, DAG.getNode(UseArithShift ? ISD::SRA : ISD::SRL, dl, VT, Result, ShiftAmt), LHS, ISD::SETNE); return true; } } EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getScalarSizeInBits() * 2); if (VT.isVector()) WideVT = EVT::getVectorVT(*DAG.getContext(), WideVT, VT.getVectorElementCount()); SDValue BottomHalf; SDValue TopHalf; static const unsigned Ops[2][3] = { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND }, { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }}; if (isOperationLegalOrCustom(Ops[isSigned][0], VT)) { BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS); TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS); } else if (isOperationLegalOrCustom(Ops[isSigned][1], VT)) { BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS, RHS); TopHalf = BottomHalf.getValue(1); } else if (isTypeLegal(WideVT)) { LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS); RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS); SDValue Mul = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS); BottomHalf = DAG.getNode(ISD::TRUNCATE, dl, VT, Mul); SDValue ShiftAmt = DAG.getShiftAmountConstant(VT.getScalarSizeInBits(), WideVT, dl); TopHalf = DAG.getNode(ISD::TRUNCATE, dl, VT, DAG.getNode(ISD::SRL, dl, WideVT, Mul, ShiftAmt)); } else { if (VT.isVector()) return false; forceExpandWideMUL(DAG, dl, isSigned, LHS, RHS, BottomHalf, TopHalf); } Result = BottomHalf; if (isSigned) { SDValue ShiftAmt = DAG.getShiftAmountConstant( VT.getScalarSizeInBits() - 1, BottomHalf.getValueType(), dl); SDValue Sign = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, ShiftAmt); Overflow = DAG.getSetCC(dl, SetCCVT, TopHalf, Sign, ISD::SETNE); } else { Overflow = DAG.getSetCC(dl, SetCCVT, TopHalf, DAG.getConstant(0, dl, VT), ISD::SETNE); } // Truncate the result if SetCC returns a larger type than needed. EVT RType = Node->getValueType(1); if (RType.bitsLT(Overflow.getValueType())) Overflow = DAG.getNode(ISD::TRUNCATE, dl, RType, Overflow); assert(RType.getSizeInBits() == Overflow.getValueSizeInBits() && "Unexpected result type for S/UMULO legalization"); return true; } SDValue TargetLowering::expandVecReduce(SDNode *Node, SelectionDAG &DAG) const { SDLoc dl(Node); unsigned BaseOpcode = ISD::getVecReduceBaseOpcode(Node->getOpcode()); SDValue Op = Node->getOperand(0); EVT VT = Op.getValueType(); if (VT.isScalableVector()) report_fatal_error( "Expanding reductions for scalable vectors is undefined."); // Try to use a shuffle reduction for power of two vectors. if (VT.isPow2VectorType()) { while (VT.getVectorNumElements() > 1) { EVT HalfVT = VT.getHalfNumVectorElementsVT(*DAG.getContext()); if (!isOperationLegalOrCustom(BaseOpcode, HalfVT)) break; SDValue Lo, Hi; std::tie(Lo, Hi) = DAG.SplitVector(Op, dl); Op = DAG.getNode(BaseOpcode, dl, HalfVT, Lo, Hi, Node->getFlags()); VT = HalfVT; } } EVT EltVT = VT.getVectorElementType(); unsigned NumElts = VT.getVectorNumElements(); SmallVector Ops; DAG.ExtractVectorElements(Op, Ops, 0, NumElts); SDValue Res = Ops[0]; for (unsigned i = 1; i < NumElts; i++) Res = DAG.getNode(BaseOpcode, dl, EltVT, Res, Ops[i], Node->getFlags()); // Result type may be wider than element type. if (EltVT != Node->getValueType(0)) Res = DAG.getNode(ISD::ANY_EXTEND, dl, Node->getValueType(0), Res); return Res; } SDValue TargetLowering::expandVecReduceSeq(SDNode *Node, SelectionDAG &DAG) const { SDLoc dl(Node); SDValue AccOp = Node->getOperand(0); SDValue VecOp = Node->getOperand(1); SDNodeFlags Flags = Node->getFlags(); EVT VT = VecOp.getValueType(); EVT EltVT = VT.getVectorElementType(); if (VT.isScalableVector()) report_fatal_error( "Expanding reductions for scalable vectors is undefined."); unsigned NumElts = VT.getVectorNumElements(); SmallVector Ops; DAG.ExtractVectorElements(VecOp, Ops, 0, NumElts); unsigned BaseOpcode = ISD::getVecReduceBaseOpcode(Node->getOpcode()); SDValue Res = AccOp; for (unsigned i = 0; i < NumElts; i++) Res = DAG.getNode(BaseOpcode, dl, EltVT, Res, Ops[i], Flags); return Res; } bool TargetLowering::expandREM(SDNode *Node, SDValue &Result, SelectionDAG &DAG) const { EVT VT = Node->getValueType(0); SDLoc dl(Node); bool isSigned = Node->getOpcode() == ISD::SREM; unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV; unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM; SDValue Dividend = Node->getOperand(0); SDValue Divisor = Node->getOperand(1); if (isOperationLegalOrCustom(DivRemOpc, VT)) { SDVTList VTs = DAG.getVTList(VT, VT); Result = DAG.getNode(DivRemOpc, dl, VTs, Dividend, Divisor).getValue(1); return true; } if (isOperationLegalOrCustom(DivOpc, VT)) { // X % Y -> X-X/Y*Y SDValue Divide = DAG.getNode(DivOpc, dl, VT, Dividend, Divisor); SDValue Mul = DAG.getNode(ISD::MUL, dl, VT, Divide, Divisor); Result = DAG.getNode(ISD::SUB, dl, VT, Dividend, Mul); return true; } return false; } SDValue TargetLowering::expandFP_TO_INT_SAT(SDNode *Node, SelectionDAG &DAG) const { bool IsSigned = Node->getOpcode() == ISD::FP_TO_SINT_SAT; SDLoc dl(SDValue(Node, 0)); SDValue Src = Node->getOperand(0); // DstVT is the result type, while SatVT is the size to which we saturate EVT SrcVT = Src.getValueType(); EVT DstVT = Node->getValueType(0); EVT SatVT = cast(Node->getOperand(1))->getVT(); unsigned SatWidth = SatVT.getScalarSizeInBits(); unsigned DstWidth = DstVT.getScalarSizeInBits(); assert(SatWidth <= DstWidth && "Expected saturation width smaller than result width"); // Determine minimum and maximum integer values and their corresponding // floating-point values. APInt MinInt, MaxInt; if (IsSigned) { MinInt = APInt::getSignedMinValue(SatWidth).sext(DstWidth); MaxInt = APInt::getSignedMaxValue(SatWidth).sext(DstWidth); } else { MinInt = APInt::getMinValue(SatWidth).zext(DstWidth); MaxInt = APInt::getMaxValue(SatWidth).zext(DstWidth); } // We cannot risk emitting FP_TO_XINT nodes with a source VT of [b]f16, as // libcall emission cannot handle this. Large result types will fail. if (SrcVT == MVT::f16 || SrcVT == MVT::bf16) { Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Src); SrcVT = Src.getValueType(); } APFloat MinFloat(DAG.EVTToAPFloatSemantics(SrcVT)); APFloat MaxFloat(DAG.EVTToAPFloatSemantics(SrcVT)); APFloat::opStatus MinStatus = MinFloat.convertFromAPInt(MinInt, IsSigned, APFloat::rmTowardZero); APFloat::opStatus MaxStatus = MaxFloat.convertFromAPInt(MaxInt, IsSigned, APFloat::rmTowardZero); bool AreExactFloatBounds = !(MinStatus & APFloat::opStatus::opInexact) && !(MaxStatus & APFloat::opStatus::opInexact); SDValue MinFloatNode = DAG.getConstantFP(MinFloat, dl, SrcVT); SDValue MaxFloatNode = DAG.getConstantFP(MaxFloat, dl, SrcVT); // If the integer bounds are exactly representable as floats and min/max are // legal, emit a min+max+fptoi sequence. Otherwise we have to use a sequence // of comparisons and selects. bool MinMaxLegal = isOperationLegal(ISD::FMINNUM, SrcVT) && isOperationLegal(ISD::FMAXNUM, SrcVT); if (AreExactFloatBounds && MinMaxLegal) { SDValue Clamped = Src; // Clamp Src by MinFloat from below. If Src is NaN the result is MinFloat. Clamped = DAG.getNode(ISD::FMAXNUM, dl, SrcVT, Clamped, MinFloatNode); // Clamp by MaxFloat from above. NaN cannot occur. Clamped = DAG.getNode(ISD::FMINNUM, dl, SrcVT, Clamped, MaxFloatNode); // Convert clamped value to integer. SDValue FpToInt = DAG.getNode(IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, dl, DstVT, Clamped); // In the unsigned case we're done, because we mapped NaN to MinFloat, // which will cast to zero. if (!IsSigned) return FpToInt; // Otherwise, select 0 if Src is NaN. SDValue ZeroInt = DAG.getConstant(0, dl, DstVT); EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT); SDValue IsNan = DAG.getSetCC(dl, SetCCVT, Src, Src, ISD::CondCode::SETUO); return DAG.getSelect(dl, DstVT, IsNan, ZeroInt, FpToInt); } SDValue MinIntNode = DAG.getConstant(MinInt, dl, DstVT); SDValue MaxIntNode = DAG.getConstant(MaxInt, dl, DstVT); // Result of direct conversion. The assumption here is that the operation is // non-trapping and it's fine to apply it to an out-of-range value if we // select it away later. SDValue FpToInt = DAG.getNode(IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, dl, DstVT, Src); SDValue Select = FpToInt; EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT); // If Src ULT MinFloat, select MinInt. In particular, this also selects // MinInt if Src is NaN. SDValue ULT = DAG.getSetCC(dl, SetCCVT, Src, MinFloatNode, ISD::SETULT); Select = DAG.getSelect(dl, DstVT, ULT, MinIntNode, Select); // If Src OGT MaxFloat, select MaxInt. SDValue OGT = DAG.getSetCC(dl, SetCCVT, Src, MaxFloatNode, ISD::SETOGT); Select = DAG.getSelect(dl, DstVT, OGT, MaxIntNode, Select); // In the unsigned case we are done, because we mapped NaN to MinInt, which // is already zero. if (!IsSigned) return Select; // Otherwise, select 0 if Src is NaN. SDValue ZeroInt = DAG.getConstant(0, dl, DstVT); SDValue IsNan = DAG.getSetCC(dl, SetCCVT, Src, Src, ISD::CondCode::SETUO); return DAG.getSelect(dl, DstVT, IsNan, ZeroInt, Select); } SDValue TargetLowering::expandRoundInexactToOdd(EVT ResultVT, SDValue Op, const SDLoc &dl, SelectionDAG &DAG) const { EVT OperandVT = Op.getValueType(); if (OperandVT.getScalarType() == ResultVT.getScalarType()) return Op; EVT ResultIntVT = ResultVT.changeTypeToInteger(); // We are rounding binary64/binary128 -> binary32 -> bfloat16. This // can induce double-rounding which may alter the results. We can // correct for this using a trick explained in: Boldo, Sylvie, and // Guillaume Melquiond. "When double rounding is odd." 17th IMACS // World Congress. 2005. unsigned BitSize = OperandVT.getScalarSizeInBits(); EVT WideIntVT = OperandVT.changeTypeToInteger(); SDValue OpAsInt = DAG.getBitcast(WideIntVT, Op); SDValue SignBit = DAG.getNode(ISD::AND, dl, WideIntVT, OpAsInt, DAG.getConstant(APInt::getSignMask(BitSize), dl, WideIntVT)); SDValue AbsWide; if (isOperationLegalOrCustom(ISD::FABS, OperandVT)) { AbsWide = DAG.getNode(ISD::FABS, dl, OperandVT, Op); } else { SDValue ClearedSign = DAG.getNode( ISD::AND, dl, WideIntVT, OpAsInt, DAG.getConstant(APInt::getSignedMaxValue(BitSize), dl, WideIntVT)); AbsWide = DAG.getBitcast(OperandVT, ClearedSign); } SDValue AbsNarrow = DAG.getFPExtendOrRound(AbsWide, dl, ResultVT); SDValue AbsNarrowAsWide = DAG.getFPExtendOrRound(AbsNarrow, dl, OperandVT); // We can keep the narrow value as-is if narrowing was exact (no // rounding error), the wide value was NaN (the narrow value is also // NaN and should be preserved) or if we rounded to the odd value. SDValue NarrowBits = DAG.getNode(ISD::BITCAST, dl, ResultIntVT, AbsNarrow); SDValue One = DAG.getConstant(1, dl, ResultIntVT); SDValue NegativeOne = DAG.getAllOnesConstant(dl, ResultIntVT); SDValue And = DAG.getNode(ISD::AND, dl, ResultIntVT, NarrowBits, One); EVT ResultIntVTCCVT = getSetCCResultType( DAG.getDataLayout(), *DAG.getContext(), And.getValueType()); SDValue Zero = DAG.getConstant(0, dl, ResultIntVT); // The result is already odd so we don't need to do anything. SDValue AlreadyOdd = DAG.getSetCC(dl, ResultIntVTCCVT, And, Zero, ISD::SETNE); EVT WideSetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), AbsWide.getValueType()); // We keep results which are exact, odd or NaN. SDValue KeepNarrow = DAG.getSetCC(dl, WideSetCCVT, AbsWide, AbsNarrowAsWide, ISD::SETUEQ); KeepNarrow = DAG.getNode(ISD::OR, dl, WideSetCCVT, KeepNarrow, AlreadyOdd); // We morally performed a round-down if AbsNarrow is smaller than // AbsWide. SDValue NarrowIsRd = DAG.getSetCC(dl, WideSetCCVT, AbsWide, AbsNarrowAsWide, ISD::SETOGT); // If the narrow value is odd or exact, pick it. // Otherwise, narrow is even and corresponds to either the rounded-up // or rounded-down value. If narrow is the rounded-down value, we want // the rounded-up value as it will be odd. SDValue Adjust = DAG.getSelect(dl, ResultIntVT, NarrowIsRd, One, NegativeOne); SDValue Adjusted = DAG.getNode(ISD::ADD, dl, ResultIntVT, NarrowBits, Adjust); Op = DAG.getSelect(dl, ResultIntVT, KeepNarrow, NarrowBits, Adjusted); int ShiftAmount = BitSize - ResultVT.getScalarSizeInBits(); SDValue ShiftCnst = DAG.getShiftAmountConstant(ShiftAmount, WideIntVT, dl); SignBit = DAG.getNode(ISD::SRL, dl, WideIntVT, SignBit, ShiftCnst); SignBit = DAG.getNode(ISD::TRUNCATE, dl, ResultIntVT, SignBit); Op = DAG.getNode(ISD::OR, dl, ResultIntVT, Op, SignBit); return DAG.getNode(ISD::BITCAST, dl, ResultVT, Op); } SDValue TargetLowering::expandFP_ROUND(SDNode *Node, SelectionDAG &DAG) const { assert(Node->getOpcode() == ISD::FP_ROUND && "Unexpected opcode!"); SDValue Op = Node->getOperand(0); EVT VT = Node->getValueType(0); SDLoc dl(Node); if (VT.getScalarType() == MVT::bf16) { if (Node->getConstantOperandVal(1) == 1) { return DAG.getNode(ISD::FP_TO_BF16, dl, VT, Node->getOperand(0)); } EVT OperandVT = Op.getValueType(); SDValue IsNaN = DAG.getSetCC( dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), OperandVT), Op, Op, ISD::SETUO); // We are rounding binary64/binary128 -> binary32 -> bfloat16. This // can induce double-rounding which may alter the results. We can // correct for this using a trick explained in: Boldo, Sylvie, and // Guillaume Melquiond. "When double rounding is odd." 17th IMACS // World Congress. 2005. EVT F32 = VT.isVector() ? VT.changeVectorElementType(MVT::f32) : MVT::f32; EVT I32 = F32.changeTypeToInteger(); Op = expandRoundInexactToOdd(F32, Op, dl, DAG); Op = DAG.getNode(ISD::BITCAST, dl, I32, Op); // Conversions should set NaN's quiet bit. This also prevents NaNs from // turning into infinities. SDValue NaN = DAG.getNode(ISD::OR, dl, I32, Op, DAG.getConstant(0x400000, dl, I32)); // Factor in the contribution of the low 16 bits. SDValue One = DAG.getConstant(1, dl, I32); SDValue Lsb = DAG.getNode(ISD::SRL, dl, I32, Op, DAG.getShiftAmountConstant(16, I32, dl)); Lsb = DAG.getNode(ISD::AND, dl, I32, Lsb, One); SDValue RoundingBias = DAG.getNode(ISD::ADD, dl, I32, DAG.getConstant(0x7fff, dl, I32), Lsb); SDValue Add = DAG.getNode(ISD::ADD, dl, I32, Op, RoundingBias); // Don't round if we had a NaN, we don't want to turn 0x7fffffff into // 0x80000000. Op = DAG.getSelect(dl, I32, IsNaN, NaN, Add); // Now that we have rounded, shift the bits into position. Op = DAG.getNode(ISD::SRL, dl, I32, Op, DAG.getShiftAmountConstant(16, I32, dl)); Op = DAG.getNode(ISD::BITCAST, dl, I32, Op); EVT I16 = I32.isVector() ? I32.changeVectorElementType(MVT::i16) : MVT::i16; Op = DAG.getNode(ISD::TRUNCATE, dl, I16, Op); return DAG.getNode(ISD::BITCAST, dl, VT, Op); } return SDValue(); } SDValue TargetLowering::expandVectorSplice(SDNode *Node, SelectionDAG &DAG) const { assert(Node->getOpcode() == ISD::VECTOR_SPLICE && "Unexpected opcode!"); assert(Node->getValueType(0).isScalableVector() && "Fixed length vector types expected to use SHUFFLE_VECTOR!"); EVT VT = Node->getValueType(0); SDValue V1 = Node->getOperand(0); SDValue V2 = Node->getOperand(1); int64_t Imm = cast(Node->getOperand(2))->getSExtValue(); SDLoc DL(Node); // Expand through memory thusly: // Alloca CONCAT_VECTORS_TYPES(V1, V2) Ptr // Store V1, Ptr // Store V2, Ptr + sizeof(V1) // If (Imm < 0) // TrailingElts = -Imm // Ptr = Ptr + sizeof(V1) - (TrailingElts * sizeof(VT.Elt)) // else // Ptr = Ptr + (Imm * sizeof(VT.Elt)) // Res = Load Ptr Align Alignment = DAG.getReducedAlign(VT, /*UseABI=*/false); EVT MemVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), VT.getVectorElementCount() * 2); SDValue StackPtr = DAG.CreateStackTemporary(MemVT.getStoreSize(), Alignment); EVT PtrVT = StackPtr.getValueType(); auto &MF = DAG.getMachineFunction(); auto FrameIndex = cast(StackPtr.getNode())->getIndex(); auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FrameIndex); // Store the lo part of CONCAT_VECTORS(V1, V2) SDValue StoreV1 = DAG.getStore(DAG.getEntryNode(), DL, V1, StackPtr, PtrInfo); // Store the hi part of CONCAT_VECTORS(V1, V2) SDValue OffsetToV2 = DAG.getVScale( DL, PtrVT, APInt(PtrVT.getFixedSizeInBits(), VT.getStoreSize().getKnownMinValue())); SDValue StackPtr2 = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, OffsetToV2); SDValue StoreV2 = DAG.getStore(StoreV1, DL, V2, StackPtr2, PtrInfo); if (Imm >= 0) { // Load back the required element. getVectorElementPointer takes care of // clamping the index if it's out-of-bounds. StackPtr = getVectorElementPointer(DAG, StackPtr, VT, Node->getOperand(2)); // Load the spliced result return DAG.getLoad(VT, DL, StoreV2, StackPtr, MachinePointerInfo::getUnknownStack(MF)); } uint64_t TrailingElts = -Imm; // NOTE: TrailingElts must be clamped so as not to read outside of V1:V2. TypeSize EltByteSize = VT.getVectorElementType().getStoreSize(); SDValue TrailingBytes = DAG.getConstant(TrailingElts * EltByteSize, DL, PtrVT); if (TrailingElts > VT.getVectorMinNumElements()) { SDValue VLBytes = DAG.getVScale(DL, PtrVT, APInt(PtrVT.getFixedSizeInBits(), VT.getStoreSize().getKnownMinValue())); TrailingBytes = DAG.getNode(ISD::UMIN, DL, PtrVT, TrailingBytes, VLBytes); } // Calculate the start address of the spliced result. StackPtr2 = DAG.getNode(ISD::SUB, DL, PtrVT, StackPtr2, TrailingBytes); // Load the spliced result return DAG.getLoad(VT, DL, StoreV2, StackPtr2, MachinePointerInfo::getUnknownStack(MF)); } SDValue TargetLowering::expandVECTOR_COMPRESS(SDNode *Node, SelectionDAG &DAG) const { SDLoc DL(Node); SDValue Vec = Node->getOperand(0); SDValue Mask = Node->getOperand(1); SDValue Passthru = Node->getOperand(2); EVT VecVT = Vec.getValueType(); EVT ScalarVT = VecVT.getScalarType(); EVT MaskVT = Mask.getValueType(); EVT MaskScalarVT = MaskVT.getScalarType(); // Needs to be handled by targets that have scalable vector types. if (VecVT.isScalableVector()) report_fatal_error("Cannot expand masked_compress for scalable vectors."); SDValue StackPtr = DAG.CreateStackTemporary( VecVT.getStoreSize(), DAG.getReducedAlign(VecVT, /*UseABI=*/false)); int FI = cast(StackPtr.getNode())->getIndex(); MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI); MVT PositionVT = getVectorIdxTy(DAG.getDataLayout()); SDValue Chain = DAG.getEntryNode(); SDValue OutPos = DAG.getConstant(0, DL, PositionVT); bool HasPassthru = !Passthru.isUndef(); // If we have a passthru vector, store it on the stack, overwrite the matching // positions and then re-write the last element that was potentially // overwritten even though mask[i] = false. if (HasPassthru) Chain = DAG.getStore(Chain, DL, Passthru, StackPtr, PtrInfo); SDValue LastWriteVal; APInt PassthruSplatVal; bool IsSplatPassthru = ISD::isConstantSplatVector(Passthru.getNode(), PassthruSplatVal); if (IsSplatPassthru) { // As we do not know which position we wrote to last, we cannot simply // access that index from the passthru vector. So we first check if passthru // is a splat vector, to use any element ... LastWriteVal = DAG.getConstant(PassthruSplatVal, DL, ScalarVT); } else if (HasPassthru) { // ... if it is not a splat vector, we need to get the passthru value at // position = popcount(mask) and re-load it from the stack before it is // overwritten in the loop below. SDValue Popcount = DAG.getNode( ISD::TRUNCATE, DL, MaskVT.changeVectorElementType(MVT::i1), Mask); Popcount = DAG.getNode(ISD::ZERO_EXTEND, DL, MaskVT.changeVectorElementType(ScalarVT), Popcount); Popcount = DAG.getNode(ISD::VECREDUCE_ADD, DL, ScalarVT, Popcount); SDValue LastElmtPtr = getVectorElementPointer(DAG, StackPtr, VecVT, Popcount); LastWriteVal = DAG.getLoad( ScalarVT, DL, Chain, LastElmtPtr, MachinePointerInfo::getUnknownStack(DAG.getMachineFunction())); Chain = LastWriteVal.getValue(1); } unsigned NumElms = VecVT.getVectorNumElements(); for (unsigned I = 0; I < NumElms; I++) { SDValue Idx = DAG.getVectorIdxConstant(I, DL); SDValue ValI = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ScalarVT, Vec, Idx); SDValue OutPtr = getVectorElementPointer(DAG, StackPtr, VecVT, OutPos); Chain = DAG.getStore( Chain, DL, ValI, OutPtr, MachinePointerInfo::getUnknownStack(DAG.getMachineFunction())); // Get the mask value and add it to the current output position. This // either increments by 1 if MaskI is true or adds 0 otherwise. // Freeze in case we have poison/undef mask entries. SDValue MaskI = DAG.getFreeze( DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MaskScalarVT, Mask, Idx)); MaskI = DAG.getFreeze(MaskI); MaskI = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, MaskI); MaskI = DAG.getNode(ISD::ZERO_EXTEND, DL, PositionVT, MaskI); OutPos = DAG.getNode(ISD::ADD, DL, PositionVT, OutPos, MaskI); if (HasPassthru && I == NumElms - 1) { SDValue EndOfVector = DAG.getConstant(VecVT.getVectorNumElements() - 1, DL, PositionVT); SDValue AllLanesSelected = DAG.getSetCC(DL, MVT::i1, OutPos, EndOfVector, ISD::CondCode::SETUGT); OutPos = DAG.getNode(ISD::UMIN, DL, PositionVT, OutPos, EndOfVector); OutPtr = getVectorElementPointer(DAG, StackPtr, VecVT, OutPos); // Re-write the last ValI if all lanes were selected. Otherwise, // overwrite the last write it with the passthru value. LastWriteVal = DAG.getSelect(DL, ScalarVT, AllLanesSelected, ValI, LastWriteVal); Chain = DAG.getStore( Chain, DL, LastWriteVal, OutPtr, MachinePointerInfo::getUnknownStack(DAG.getMachineFunction())); } } return DAG.getLoad(VecVT, DL, Chain, StackPtr, PtrInfo); } bool TargetLowering::LegalizeSetCCCondCode(SelectionDAG &DAG, EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC, SDValue Mask, SDValue EVL, bool &NeedInvert, const SDLoc &dl, SDValue &Chain, bool IsSignaling) const { const TargetLowering &TLI = DAG.getTargetLoweringInfo(); MVT OpVT = LHS.getSimpleValueType(); ISD::CondCode CCCode = cast(CC)->get(); NeedInvert = false; assert(!EVL == !Mask && "VP Mask and EVL must either both be set or unset"); bool IsNonVP = !EVL; switch (TLI.getCondCodeAction(CCCode, OpVT)) { default: llvm_unreachable("Unknown condition code action!"); case TargetLowering::Legal: // Nothing to do. break; case TargetLowering::Expand: { ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode); if (TLI.isCondCodeLegalOrCustom(InvCC, OpVT)) { std::swap(LHS, RHS); CC = DAG.getCondCode(InvCC); return true; } // Swapping operands didn't work. Try inverting the condition. bool NeedSwap = false; InvCC = getSetCCInverse(CCCode, OpVT); if (!TLI.isCondCodeLegalOrCustom(InvCC, OpVT)) { // If inverting the condition is not enough, try swapping operands // on top of it. InvCC = ISD::getSetCCSwappedOperands(InvCC); NeedSwap = true; } if (TLI.isCondCodeLegalOrCustom(InvCC, OpVT)) { CC = DAG.getCondCode(InvCC); NeedInvert = true; if (NeedSwap) std::swap(LHS, RHS); return true; } ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID; unsigned Opc = 0; switch (CCCode) { default: llvm_unreachable("Don't know how to expand this condition!"); case ISD::SETUO: if (TLI.isCondCodeLegal(ISD::SETUNE, OpVT)) { CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break; } assert(TLI.isCondCodeLegal(ISD::SETOEQ, OpVT) && "If SETUE is expanded, SETOEQ or SETUNE must be legal!"); NeedInvert = true; [[fallthrough]]; case ISD::SETO: assert(TLI.isCondCodeLegal(ISD::SETOEQ, OpVT) && "If SETO is expanded, SETOEQ must be legal!"); CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break; case ISD::SETONE: case ISD::SETUEQ: // If the SETUO or SETO CC isn't legal, we might be able to use // SETOGT || SETOLT, inverting the result for SETUEQ. We only need one // of SETOGT/SETOLT to be legal, the other can be emulated by swapping // the operands. CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO; if (!TLI.isCondCodeLegal(CC2, OpVT) && (TLI.isCondCodeLegal(ISD::SETOGT, OpVT) || TLI.isCondCodeLegal(ISD::SETOLT, OpVT))) { CC1 = ISD::SETOGT; CC2 = ISD::SETOLT; Opc = ISD::OR; NeedInvert = ((unsigned)CCCode & 0x8U); break; } [[fallthrough]]; case ISD::SETOEQ: case ISD::SETOGT: case ISD::SETOGE: case ISD::SETOLT: case ISD::SETOLE: case ISD::SETUNE: case ISD::SETUGT: case ISD::SETUGE: case ISD::SETULT: case ISD::SETULE: // If we are floating point, assign and break, otherwise fall through. if (!OpVT.isInteger()) { // We can use the 4th bit to tell if we are the unordered // or ordered version of the opcode. CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO; Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND; CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10); break; } // Fallthrough if we are unsigned integer. [[fallthrough]]; case ISD::SETLE: case ISD::SETGT: case ISD::SETGE: case ISD::SETLT: case ISD::SETNE: case ISD::SETEQ: // If all combinations of inverting the condition and swapping operands // didn't work then we have no means to expand the condition. llvm_unreachable("Don't know how to expand this condition!"); } SDValue SetCC1, SetCC2; if (CCCode != ISD::SETO && CCCode != ISD::SETUO) { // If we aren't the ordered or unorder operation, // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS). if (IsNonVP) { SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1, Chain, IsSignaling); SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2, Chain, IsSignaling); } else { SetCC1 = DAG.getSetCCVP(dl, VT, LHS, RHS, CC1, Mask, EVL); SetCC2 = DAG.getSetCCVP(dl, VT, LHS, RHS, CC2, Mask, EVL); } } else { // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS) if (IsNonVP) { SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1, Chain, IsSignaling); SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2, Chain, IsSignaling); } else { SetCC1 = DAG.getSetCCVP(dl, VT, LHS, LHS, CC1, Mask, EVL); SetCC2 = DAG.getSetCCVP(dl, VT, RHS, RHS, CC2, Mask, EVL); } } if (Chain) Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, SetCC1.getValue(1), SetCC2.getValue(1)); if (IsNonVP) LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2); else { // Transform the binary opcode to the VP equivalent. assert((Opc == ISD::OR || Opc == ISD::AND) && "Unexpected opcode"); Opc = Opc == ISD::OR ? ISD::VP_OR : ISD::VP_AND; LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2, Mask, EVL); } RHS = SDValue(); CC = SDValue(); return true; } } return false; }