X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FARM%2FARMISelLowering.cpp;h=e193672ef7a97c8a7066926feac84bfe7fd0c456;hb=420761a0f193e87d08ee1c51b26bba23ab4bac7f;hp=f23dd8903857232d99954d8e4659ab06fa7ff779;hpb=6edd5884c91aea72661ed899b0c91dfb4f0ea80f;p=oota-llvm.git diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp index f23dd890385..e193672ef7a 100644 --- a/lib/Target/ARM/ARMISelLowering.cpp +++ b/lib/Target/ARM/ARMISelLowering.cpp @@ -13,13 +13,12 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "arm-isel" +#include "ARMISelLowering.h" #include "ARM.h" #include "ARMCallingConv.h" #include "ARMConstantPoolValue.h" -#include "ARMISelLowering.h" #include "ARMMachineFunctionInfo.h" #include "ARMPerfectShuffle.h" -#include "ARMRegisterInfo.h" #include "ARMSubtarget.h" #include "ARMTargetMachine.h" #include "ARMTargetObjectFile.h" @@ -49,7 +48,6 @@ #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" -#include using namespace llvm; STATISTIC(NumTailCalls, "Number of tail calls"); @@ -87,7 +85,7 @@ namespace { } // The APCS parameter registers. -static const unsigned GPRArgRegs[] = { +static const uint16_t GPRArgRegs[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 }; @@ -155,12 +153,12 @@ void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT, } void ARMTargetLowering::addDRTypeForNEON(EVT VT) { - addRegisterClass(VT, ARM::DPRRegisterClass); + addRegisterClass(VT, &ARM::DPRRegClass); addTypeForNEON(VT, MVT::f64, MVT::v2i32); } void ARMTargetLowering::addQRTypeForNEON(EVT VT) { - addRegisterClass(VT, ARM::QPRRegisterClass); + addRegisterClass(VT, &ARM::QPRRegClass); addTypeForNEON(VT, MVT::v2f64, MVT::v4i32); } @@ -261,7 +259,7 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setLibcallName(RTLIB::SRL_I128, 0); setLibcallName(RTLIB::SRA_I128, 0); - if (Subtarget->isAAPCS_ABI()) { + if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) { // Double-precision floating-point arithmetic helper functions // RTABI chapter 4.1.2, Table 2 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd"); @@ -433,14 +431,14 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) } if (Subtarget->isThumb1Only()) - addRegisterClass(MVT::i32, ARM::tGPRRegisterClass); + addRegisterClass(MVT::i32, &ARM::tGPRRegClass); else - addRegisterClass(MVT::i32, ARM::GPRRegisterClass); + addRegisterClass(MVT::i32, &ARM::GPRRegClass); if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) { - addRegisterClass(MVT::f32, ARM::SPRRegisterClass); + addRegisterClass(MVT::f32, &ARM::SPRRegClass); if (!Subtarget->isFPOnlySP()) - addRegisterClass(MVT::f64, ARM::DPRRegisterClass); + addRegisterClass(MVT::f64, &ARM::DPRRegClass); setTruncStoreAction(MVT::f64, MVT::f32, Expand); } @@ -456,6 +454,8 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand); } + setOperationAction(ISD::ConstantFP, MVT::f32, Custom); + if (Subtarget->hasNEON()) { addDRTypeForNEON(MVT::v2f32); addDRTypeForNEON(MVT::v8i8); @@ -507,7 +507,7 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setOperationAction(ISD::FRINT, MVT::v2f64, Expand); setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand); setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand); - + setOperationAction(ISD::FSQRT, MVT::v4f32, Expand); setOperationAction(ISD::FSIN, MVT::v4f32, Expand); setOperationAction(ISD::FCOS, MVT::v4f32, Expand); @@ -533,9 +533,13 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setOperationAction(ISD::SETCC, MVT::v1i64, Expand); setOperationAction(ISD::SETCC, MVT::v2i64, Expand); // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with - // a destination type that is wider than the source. + // a destination type that is wider than the source, and nor does + // it have a FP_TO_[SU]INT instruction with a narrower destination than + // source. setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom); setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); + setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom); + setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom); setTargetDAGCombine(ISD::INTRINSIC_VOID); setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); @@ -555,7 +559,15 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setTargetDAGCombine(ISD::FP_TO_UINT); setTargetDAGCombine(ISD::FDIV); - setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand); + // It is legal to extload from v4i8 to v4i16 or v4i32. + MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8, + MVT::v4i16, MVT::v2i16, + MVT::v2i32}; + for (unsigned i = 0; i < 6; ++i) { + setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal); + setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal); + setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal); + } } computeRegisterProperties(); @@ -647,10 +659,15 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setOperationAction(ISD::VAEND, MVT::Other, Expand); setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); - setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); - setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand); - setExceptionPointerRegister(ARM::R0); - setExceptionSelectorRegister(ARM::R1); + + if (!Subtarget->isTargetDarwin()) { + // Non-Darwin platforms may return values in these registers via the + // personality function. + setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); + setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand); + setExceptionPointerRegister(ARM::R0); + setExceptionSelectorRegister(ARM::R1); + } setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use @@ -752,8 +769,10 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setOperationAction(ISD::FPOW, MVT::f64, Expand); setOperationAction(ISD::FPOW, MVT::f32, Expand); - setOperationAction(ISD::FMA, MVT::f64, Expand); - setOperationAction(ISD::FMA, MVT::f32, Expand); + if (!Subtarget->hasVFP4()) { + setOperationAction(ISD::FMA, MVT::f64, Expand); + setOperationAction(ISD::FMA, MVT::f32, Expand); + } // Various VFP goodness if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) { @@ -777,10 +796,14 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setTargetDAGCombine(ISD::SUB); setTargetDAGCombine(ISD::MUL); - if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) - setTargetDAGCombine(ISD::OR); - if (Subtarget->hasNEON()) + if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) { setTargetDAGCombine(ISD::AND); + setTargetDAGCombine(ISD::OR); + setTargetDAGCombine(ISD::XOR); + } + + if (Subtarget->hasV6Ops()) + setTargetDAGCombine(ISD::SRL); setStackPointerRegisterToSaveRestore(ARM::SP); @@ -826,7 +849,7 @@ ARMTargetLowering::findRepresentativeClass(EVT VT) const{ // the cost is 1 for both f32 and f64. case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16: case MVT::v2i32: case MVT::v1i64: case MVT::v2f32: - RRC = ARM::DPRRegisterClass; + RRC = &ARM::DPRRegClass; // When NEON is used for SP, only half of the register file is available // because operations that define both SP and DP results will be constrained // to the VFP2 class (D0-D15). We currently model this constraint prior to @@ -836,15 +859,15 @@ ARMTargetLowering::findRepresentativeClass(EVT VT) const{ break; case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64: case MVT::v4f32: case MVT::v2f64: - RRC = ARM::DPRRegisterClass; + RRC = &ARM::DPRRegClass; Cost = 2; break; case MVT::v4i64: - RRC = ARM::DPRRegisterClass; + RRC = &ARM::DPRRegClass; Cost = 4; break; case MVT::v8i64: - RRC = ARM::DPRRegisterClass; + RRC = &ARM::DPRRegClass; Cost = 8; break; } @@ -873,7 +896,11 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0"; case ARMISD::BCC_i64: return "ARMISD::BCC_i64"; case ARMISD::FMSTAT: return "ARMISD::FMSTAT"; + case ARMISD::CMOV: return "ARMISD::CMOV"; + case ARMISD::CAND: return "ARMISD::CAND"; + case ARMISD::COR: return "ARMISD::COR"; + case ARMISD::CXOR: return "ARMISD::CXOR"; case ARMISD::RBIT: return "ARMISD::RBIT"; @@ -994,15 +1021,15 @@ EVT ARMTargetLowering::getSetCCResultType(EVT VT) const { /// getRegClassFor - Return the register class that should be used for the /// specified value type. -TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const { +const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const { // Map v4i64 to QQ registers but do not make the type legal. Similarly map // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to // load / store 4 to 8 consecutive D registers. if (Subtarget->hasNEON()) { if (VT == MVT::v4i64) - return ARM::QQPRRegisterClass; - else if (VT == MVT::v8i64) - return ARM::QQQQPRRegisterClass; + return &ARM::QQPRRegClass; + if (VT == MVT::v8i64) + return &ARM::QQQQPRRegClass; } return TargetLowering::getRegClassFor(VT); } @@ -1261,7 +1288,7 @@ void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG, SDValue ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg, - bool &isTailCall, + bool doesNotRet, bool &isTailCall, const SmallVectorImpl &Outs, const SmallVectorImpl &OutVals, const SmallVectorImpl &Ins, @@ -1557,12 +1584,20 @@ ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, if (Subtarget->isThumb()) { if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps()) CallOpc = ARMISD::CALL_NOLINK; + else if (doesNotRet && isDirect && !isARMFunc && + Subtarget->hasRAS() && !Subtarget->isThumb1Only()) + // "mov lr, pc; b _foo" to avoid confusing the RSP + CallOpc = ARMISD::CALL_NOLINK; else CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL; } else { - CallOpc = (isDirect || Subtarget->hasV5TOps()) - ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL) - : ARMISD::CALL_NOLINK; + if (!isDirect && !Subtarget->hasV5TOps()) { + CallOpc = ARMISD::CALL_NOLINK; + } else if (doesNotRet && isDirect && Subtarget->hasRAS()) + // "mov lr, pc; b _foo" to avoid confusing the RSP + CallOpc = ARMISD::CALL_NOLINK; + else + CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL; } std::vector Ops; @@ -1575,6 +1610,12 @@ ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, Ops.push_back(DAG.getRegister(RegsToPass[i].first, RegsToPass[i].second.getValueType())); + // Add a register mask operand representing the call-preserved registers. + const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); + const uint32_t *Mask = TRI->getCallPreservedMask(CallConv); + assert(Mask && "Missing call preserved mask for calling convention"); + Ops.push_back(DAG.getRegisterMask(Mask)); + if (InFlag.getNode()) Ops.push_back(InFlag); @@ -1602,7 +1643,7 @@ ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, /// and then confiscate the rest of the parameter registers to insure /// this. void -llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const { +ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const { unsigned reg = State->AllocateReg(GPRArgRegs, 4); assert((State->getCallOrPrologue() == Prologue || State->getCallOrPrologue() == Call) && @@ -1632,7 +1673,7 @@ llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const { static bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags, MachineFrameInfo *MFI, const MachineRegisterInfo *MRI, - const ARMInstrInfo *TII) { + const TargetInstrInfo *TII) { unsigned Bytes = Arg.getValueType().getSizeInBits() / 8; int FI = INT_MAX; if (Arg.getOpcode() == ISD::CopyFromReg) { @@ -1767,8 +1808,7 @@ ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee, // the caller's fixed stack objects. MachineFrameInfo *MFI = MF.getFrameInfo(); const MachineRegisterInfo *MRI = &MF.getRegInfo(); - const ARMInstrInfo *TII = - ((ARMTargetMachine&)getTargetMachine()).getInstrInfo(); + const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e; ++i, ++realArgIdx) { @@ -1896,63 +1936,72 @@ ARMTargetLowering::LowerReturn(SDValue Chain, return result; } -bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const { +bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const { if (N->getNumValues() != 1) return false; if (!N->hasNUsesOfValue(1, 0)) return false; - unsigned NumCopies = 0; - SDNode* Copies[2]; - SDNode *Use = *N->use_begin(); - if (Use->getOpcode() == ISD::CopyToReg) { - Copies[NumCopies++] = Use; - } else if (Use->getOpcode() == ARMISD::VMOVRRD) { + SDValue TCChain = Chain; + SDNode *Copy = *N->use_begin(); + if (Copy->getOpcode() == ISD::CopyToReg) { + // If the copy has a glue operand, we conservatively assume it isn't safe to + // perform a tail call. + if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue) + return false; + TCChain = Copy->getOperand(0); + } else if (Copy->getOpcode() == ARMISD::VMOVRRD) { + SDNode *VMov = Copy; // f64 returned in a pair of GPRs. - for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end(); + SmallPtrSet Copies; + for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); UI != UE; ++UI) { if (UI->getOpcode() != ISD::CopyToReg) return false; - Copies[UI.getUse().getResNo()] = *UI; - ++NumCopies; + Copies.insert(*UI); } - } else if (Use->getOpcode() == ISD::BITCAST) { + if (Copies.size() > 2) + return false; + + for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); + UI != UE; ++UI) { + SDValue UseChain = UI->getOperand(0); + if (Copies.count(UseChain.getNode())) + // Second CopyToReg + Copy = *UI; + else + // First CopyToReg + TCChain = UseChain; + } + } else if (Copy->getOpcode() == ISD::BITCAST) { // f32 returned in a single GPR. - if (!Use->hasNUsesOfValue(1, 0)) + if (!Copy->hasOneUse()) return false; - Use = *Use->use_begin(); - if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0)) + Copy = *Copy->use_begin(); + if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0)) return false; - Copies[NumCopies++] = Use; + Chain = Copy->getOperand(0); } else { return false; } - if (NumCopies != 1 && NumCopies != 2) - return false; - bool HasRet = false; - for (unsigned i = 0; i < NumCopies; ++i) { - SDNode *Copy = Copies[i]; - for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end(); - UI != UE; ++UI) { - if (UI->getOpcode() == ISD::CopyToReg) { - SDNode *Use = *UI; - if (Use == Copies[0] || Use == Copies[1]) - continue; - return false; - } - if (UI->getOpcode() != ARMISD::RET_FLAG) - return false; - HasRet = true; - } + for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end(); + UI != UE; ++UI) { + if (UI->getOpcode() != ARMISD::RET_FLAG) + return false; + HasRet = true; } - return HasRet; + if (!HasRet) + return false; + + Chain = TCChain; + return true; } bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const { - if (!EnableARMTailCalls) + if (!EnableARMTailCalls && !Subtarget->supportsTailCall()) return false; if (!CI->isTailCall()) @@ -2049,7 +2098,8 @@ ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, std::pair CallResult = LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()), false, false, false, false, - 0, CallingConv::C, false, /*isReturnValueUsed=*/true, + 0, CallingConv::C, /*isTailCall=*/false, + /*doesNotRet=*/false, /*isReturnValueUsed=*/true, DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl); return CallResult.first; } @@ -2405,11 +2455,11 @@ ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, MachineFunction &MF = DAG.getMachineFunction(); ARMFunctionInfo *AFI = MF.getInfo(); - TargetRegisterClass *RC; + const TargetRegisterClass *RC; if (AFI->isThumb1OnlyFunction()) - RC = ARM::tGPRRegisterClass; + RC = &ARM::tGPRRegClass; else - RC = ARM::GPRRegisterClass; + RC = &ARM::GPRRegClass; // Transform the arguments stored in physical registers into virtual ones. unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); @@ -2491,11 +2541,11 @@ ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, SmallVector MemOps; for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) { - TargetRegisterClass *RC; + const TargetRegisterClass *RC; if (AFI->isThumb1OnlyFunction()) - RC = ARM::tGPRRegisterClass; + RC = &ARM::tGPRRegClass; else - RC = ARM::GPRRegisterClass; + RC = &ARM::GPRRegClass; unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC); SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); @@ -2574,17 +2624,18 @@ ARMTargetLowering::LowerFormalArguments(SDValue Chain, ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl); } else { - TargetRegisterClass *RC; + const TargetRegisterClass *RC; if (RegVT == MVT::f32) - RC = ARM::SPRRegisterClass; + RC = &ARM::SPRRegClass; else if (RegVT == MVT::f64) - RC = ARM::DPRRegisterClass; + RC = &ARM::DPRRegClass; else if (RegVT == MVT::v2f64) - RC = ARM::QPRRegisterClass; + RC = &ARM::QPRRegClass; else if (RegVT == MVT::i32) - RC = (AFI->isThumb1OnlyFunction() ? - ARM::tGPRRegisterClass : ARM::GPRRegisterClass); + RC = AFI->isThumb1OnlyFunction() ? + (const TargetRegisterClass*)&ARM::tGPRRegClass : + (const TargetRegisterClass*)&ARM::GPRRegClass; else llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering"); @@ -2816,6 +2867,11 @@ SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { } } + // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the + // undefined bits before doing a full-word comparison with zero. + Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond, + DAG.getConstant(1, Cond.getValueType())); + return DAG.getSelectCC(dl, Cond, DAG.getConstant(0, Cond.getValueType()), SelectTrue, SelectFalse, ISD::SETNE); @@ -2933,12 +2989,11 @@ ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const { SDValue Dest = Op.getOperand(4); DebugLoc dl = Op.getDebugLoc(); - bool SeenZero = false; - if (canChangeToInt(LHS, SeenZero, Subtarget) && - canChangeToInt(RHS, SeenZero, Subtarget) && - // If one of the operand is zero, it's safe to ignore the NaN case since - // we only care about equality comparisons. - (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) { + bool LHSSeenZero = false; + bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget); + bool RHSSeenZero = false; + bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget); + if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) { // If unsafe fp math optimization is enabled and there are no other uses of // the CMP operands, and the condition code is EQ or NE, we can optimize it // to an integer comparison. @@ -2947,10 +3002,13 @@ ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const { else if (CC == ISD::SETUNE) CC = ISD::SETNE; + SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32); SDValue ARMcc; if (LHS.getValueType() == MVT::f32) { - LHS = bitcastf32Toi32(LHS, DAG); - RHS = bitcastf32Toi32(RHS, DAG); + LHS = DAG.getNode(ISD::AND, dl, MVT::i32, + bitcastf32Toi32(LHS, DAG), Mask); + RHS = DAG.getNode(ISD::AND, dl, MVT::i32, + bitcastf32Toi32(RHS, DAG), Mask); SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, @@ -2961,6 +3019,8 @@ ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const { SDValue RHS1, RHS2; expandf64Toi32(LHS, DAG, LHS1, LHS2); expandf64Toi32(RHS, DAG, RHS1, RHS2); + LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask); + RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask); ARMCC::CondCodes CondCode = IntCCToARMCC(CC); ARMcc = DAG.getConstant(CondCode, MVT::i32); SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue); @@ -3053,12 +3113,22 @@ SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const { } static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) { - assert(Op.getValueType().getVectorElementType() == MVT::i32 - && "Unexpected custom lowering"); + EVT VT = Op.getValueType(); + DebugLoc dl = Op.getDebugLoc(); - if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32) - return Op; - return DAG.UnrollVectorOp(Op.getNode()); + if (Op.getValueType().getVectorElementType() == MVT::i32) { + if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32) + return Op; + return DAG.UnrollVectorOp(Op.getNode()); + } + + assert(Op.getOperand(0).getValueType() == MVT::v4f32 && + "Invalid type for custom lowering!"); + if (VT != MVT::v4i16) + return DAG.UnrollVectorOp(Op.getNode()); + + Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0)); + return DAG.getNode(ISD::TRUNCATE, dl, VT, Op); } static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) { @@ -3070,8 +3140,7 @@ static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) { unsigned Opc; switch (Op.getOpcode()) { - default: - assert(0 && "Invalid opcode!"); + default: llvm_unreachable("Invalid opcode!"); case ISD::FP_TO_SINT: Opc = ARMISD::FTOSI; break; @@ -3101,8 +3170,7 @@ static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) { unsigned CastOpc; unsigned Opc; switch (Op.getOpcode()) { - default: - assert(0 && "Invalid opcode!"); + default: llvm_unreachable("Invalid opcode!"); case ISD::SINT_TO_FP: CastOpc = ISD::SIGN_EXTEND; Opc = ISD::SINT_TO_FP; @@ -3126,8 +3194,7 @@ static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) { unsigned Opc; switch (Op.getOpcode()) { - default: - assert(0 && "Invalid opcode!"); + default: llvm_unreachable("Invalid opcode!"); case ISD::SINT_TO_FP: Opc = ARMISD::SITOF; break; @@ -3753,6 +3820,58 @@ static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef, return DAG.getTargetConstant(EncodedVal, MVT::i32); } +SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG, + const ARMSubtarget *ST) const { + if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16()) + return SDValue(); + + ConstantFPSDNode *CFP = cast(Op); + assert(Op.getValueType() == MVT::f32 && + "ConstantFP custom lowering should only occur for f32."); + + // Try splatting with a VMOV.f32... + APFloat FPVal = CFP->getValueAPF(); + int ImmVal = ARM_AM::getFP32Imm(FPVal); + if (ImmVal != -1) { + DebugLoc DL = Op.getDebugLoc(); + SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32); + SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32, + NewVal); + return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant, + DAG.getConstant(0, MVT::i32)); + } + + // If that fails, try a VMOV.i32 + EVT VMovVT; + unsigned iVal = FPVal.bitcastToAPInt().getZExtValue(); + SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false, + VMOVModImm); + if (NewVal != SDValue()) { + DebugLoc DL = Op.getDebugLoc(); + SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT, + NewVal); + SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, + VecConstant); + return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, + DAG.getConstant(0, MVT::i32)); + } + + // Finally, try a VMVN.i32 + NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false, + VMVNModImm); + if (NewVal != SDValue()) { + DebugLoc DL = Op.getDebugLoc(); + SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal); + SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, + VecConstant); + return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, + DAG.getConstant(0, MVT::i32)); + } + + return SDValue(); +} + + static bool isVEXTMask(ArrayRef M, EVT VT, bool &ReverseVEXT, unsigned &Imm) { unsigned NumElts = VT.getVectorNumElements(); @@ -4957,7 +5076,7 @@ static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) { unsigned Opc; bool ExtraOp = false; switch (Op.getOpcode()) { - default: assert(0 && "Invalid code"); + default: llvm_unreachable("Invalid code"); case ISD::ADDC: Opc = ARMISD::ADDC; break; case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break; case ISD::SUBC: Opc = ARMISD::SUBC; break; @@ -5053,6 +5172,7 @@ SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG); case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget); case ISD::SETCC: return LowerVSETCC(Op, DAG); + case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget); case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget); case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); @@ -5133,14 +5253,14 @@ ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI, bool isThumb2 = Subtarget->isThumb2(); MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); - unsigned scratch = - MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass - : ARM::GPRRegisterClass); + unsigned scratch = MRI.createVirtualRegister(isThumb2 ? + (const TargetRegisterClass*)&ARM::rGPRRegClass : + (const TargetRegisterClass*)&ARM::GPRRegClass); if (isThumb2) { - MRI.constrainRegClass(dest, ARM::rGPRRegisterClass); - MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass); - MRI.constrainRegClass(newval, ARM::rGPRRegisterClass); + MRI.constrainRegClass(dest, &ARM::rGPRRegClass); + MRI.constrainRegClass(oldval, &ARM::rGPRRegClass); + MRI.constrainRegClass(newval, &ARM::rGPRRegClass); } unsigned ldrOpc, strOpc; @@ -5243,8 +5363,8 @@ ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); if (isThumb2) { - MRI.constrainRegClass(dest, ARM::rGPRRegisterClass); - MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass); + MRI.constrainRegClass(dest, &ARM::rGPRRegClass); + MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); } unsigned ldrOpc, strOpc; @@ -5275,8 +5395,9 @@ ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, BB->end()); exitMBB->transferSuccessorsAndUpdatePHIs(BB); - TargetRegisterClass *TRC = - isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; + const TargetRegisterClass *TRC = isThumb2 ? + (const TargetRegisterClass*)&ARM::tGPRRegClass : + (const TargetRegisterClass*)&ARM::GPRRegClass; unsigned scratch = MRI.createVirtualRegister(TRC); unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC); @@ -5350,8 +5471,8 @@ ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI, MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); if (isThumb2) { - MRI.constrainRegClass(dest, ARM::rGPRRegisterClass); - MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass); + MRI.constrainRegClass(dest, &ARM::rGPRRegClass); + MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); } unsigned ldrOpc, strOpc, extendOpc; @@ -5385,8 +5506,9 @@ ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI, BB->end()); exitMBB->transferSuccessorsAndUpdatePHIs(BB); - TargetRegisterClass *TRC = - isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; + const TargetRegisterClass *TRC = isThumb2 ? + (const TargetRegisterClass*)&ARM::tGPRRegClass : + (const TargetRegisterClass*)&ARM::GPRRegClass; unsigned scratch = MRI.createVirtualRegister(TRC); unsigned scratch2 = MRI.createVirtualRegister(TRC); @@ -5412,7 +5534,7 @@ ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI, // Sign extend the value, if necessary. if (signExtend && extendOpc) { - oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass); + oldval = MRI.createVirtualRegister(&ARM::GPRRegClass); AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval) .addReg(dest) .addImm(0)); @@ -5467,9 +5589,9 @@ ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB, MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); if (isThumb2) { - MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass); - MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass); - MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass); + MRI.constrainRegClass(destlo, &ARM::rGPRRegClass); + MRI.constrainRegClass(desthi, &ARM::rGPRRegClass); + MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); } unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD; @@ -5495,8 +5617,9 @@ ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB, BB->end()); exitMBB->transferSuccessorsAndUpdatePHIs(BB); - TargetRegisterClass *TRC = - isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; + const TargetRegisterClass *TRC = isThumb2 ? + (const TargetRegisterClass*)&ARM::tGPRRegClass : + (const TargetRegisterClass*)&ARM::GPRRegClass; unsigned storesuccess = MRI.createVirtualRegister(TRC); // thisMBB: @@ -5603,8 +5726,9 @@ SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB, ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj); unsigned CPI = MCP->getConstantPoolIndex(CPV, 4); - const TargetRegisterClass *TRC = - isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; + const TargetRegisterClass *TRC = isThumb ? + (const TargetRegisterClass*)&ARM::tGPRRegClass : + (const TargetRegisterClass*)&ARM::GPRRegClass; // Grab constant pool and fixed stack memory operands. MachineMemOperand *CPMMO = @@ -5708,15 +5832,17 @@ EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const { MachineFrameInfo *MFI = MF->getFrameInfo(); int FI = MFI->getFunctionContextIndex(); - const TargetRegisterClass *TRC = - Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; + const TargetRegisterClass *TRC = Subtarget->isThumb() ? + (const TargetRegisterClass*)&ARM::tGPRRegClass : + (const TargetRegisterClass*)&ARM::GPRRegClass; // Get a mapping of the call site numbers to all of the landing pads they're // associated with. DenseMap > CallSiteNumToLPad; unsigned MaxCSNum = 0; MachineModuleInfo &MMI = MF->getMMI(); - for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) { + for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; + ++BB) { if (!BB->isLandingPad()) continue; // FIXME: We should assert that the EH_LABEL is the first MI in the landing @@ -5792,7 +5918,7 @@ EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const { BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup)); else if (!Subtarget->hasVFP2()) BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp)); - else + else BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup)); unsigned NumLPads = LPadList.size(); @@ -6015,7 +6141,7 @@ EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const { // N.B. the order the invoke BBs are processed in doesn't matter here. const ARMBaseInstrInfo *AII = static_cast(TII); const ARMBaseRegisterInfo &RI = AII->getRegisterInfo(); - const unsigned *SavedRegs = RI.getCalleeSavedRegs(MF); + const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF); SmallVector MBBLPads; for (SmallPtrSet::iterator I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) { @@ -6056,14 +6182,12 @@ EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const { for (unsigned i = 0; SavedRegs[i] != 0; ++i) { unsigned Reg = SavedRegs[i]; if (Subtarget->isThumb2() && - !ARM::tGPRRegisterClass->contains(Reg) && - !ARM::hGPRRegisterClass->contains(Reg)) + !ARM::tGPRRegClass.contains(Reg) && + !ARM::hGPRRegClass.contains(Reg)) continue; - else if (Subtarget->isThumb1Only() && - !ARM::tGPRRegisterClass->contains(Reg)) + if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg)) continue; - else if (!Subtarget->isThumb() && - !ARM::GPRRegisterClass->contains(Reg)) + if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg)) continue; if (!DefRegs[Reg]) MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead); @@ -6397,10 +6521,12 @@ ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, MachineRegisterInfo &MRI = Fn->getRegInfo(); // In Thumb mode S must not be specified if source register is the SP or // PC and if destination register is the SP, so restrict register class - unsigned NewMovDstReg = MRI.createVirtualRegister( - isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass); - unsigned NewRsbDstReg = MRI.createVirtualRegister( - isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass); + unsigned NewMovDstReg = MRI.createVirtualRegister(isThumb2 ? + (const TargetRegisterClass*)&ARM::rGPRRegClass : + (const TargetRegisterClass*)&ARM::GPRRegClass); + unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ? + (const TargetRegisterClass*)&ARM::rGPRRegClass : + (const TargetRegisterClass*)&ARM::GPRRegClass); // Transfer the remainder of BB and its successor edges to sinkMBB. SinkBB->splice(SinkBB->begin(), BB, @@ -6667,7 +6793,7 @@ static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1, case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break; case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break; default: - assert(0 && "Invalid vector element type for padd optimization."); + llvm_unreachable("Invalid vector element type for padd optimization."); } SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), @@ -6786,41 +6912,115 @@ static SDValue PerformMULCombine(SDNode *N, if (!C) return SDValue(); - uint64_t MulAmt = C->getZExtValue(); + int64_t MulAmt = C->getSExtValue(); unsigned ShiftAmt = CountTrailingZeros_64(MulAmt); + ShiftAmt = ShiftAmt & (32 - 1); SDValue V = N->getOperand(0); DebugLoc DL = N->getDebugLoc(); SDValue Res; MulAmt >>= ShiftAmt; - if (isPowerOf2_32(MulAmt - 1)) { - // (mul x, 2^N + 1) => (add (shl x, N), x) - Res = DAG.getNode(ISD::ADD, DL, VT, - V, DAG.getNode(ISD::SHL, DL, VT, - V, DAG.getConstant(Log2_32(MulAmt-1), - MVT::i32))); - } else if (isPowerOf2_32(MulAmt + 1)) { - // (mul x, 2^N - 1) => (sub (shl x, N), x) - Res = DAG.getNode(ISD::SUB, DL, VT, - DAG.getNode(ISD::SHL, DL, VT, - V, DAG.getConstant(Log2_32(MulAmt+1), - MVT::i32)), - V); - } else - return SDValue(); + + if (MulAmt >= 0) { + if (isPowerOf2_32(MulAmt - 1)) { + // (mul x, 2^N + 1) => (add (shl x, N), x) + Res = DAG.getNode(ISD::ADD, DL, VT, + V, + DAG.getNode(ISD::SHL, DL, VT, + V, + DAG.getConstant(Log2_32(MulAmt - 1), + MVT::i32))); + } else if (isPowerOf2_32(MulAmt + 1)) { + // (mul x, 2^N - 1) => (sub (shl x, N), x) + Res = DAG.getNode(ISD::SUB, DL, VT, + DAG.getNode(ISD::SHL, DL, VT, + V, + DAG.getConstant(Log2_32(MulAmt + 1), + MVT::i32)), + V); + } else + return SDValue(); + } else { + uint64_t MulAmtAbs = -MulAmt; + if (isPowerOf2_32(MulAmtAbs + 1)) { + // (mul x, -(2^N - 1)) => (sub x, (shl x, N)) + Res = DAG.getNode(ISD::SUB, DL, VT, + V, + DAG.getNode(ISD::SHL, DL, VT, + V, + DAG.getConstant(Log2_32(MulAmtAbs + 1), + MVT::i32))); + } else if (isPowerOf2_32(MulAmtAbs - 1)) { + // (mul x, -(2^N + 1)) => - (add (shl x, N), x) + Res = DAG.getNode(ISD::ADD, DL, VT, + V, + DAG.getNode(ISD::SHL, DL, VT, + V, + DAG.getConstant(Log2_32(MulAmtAbs-1), + MVT::i32))); + Res = DAG.getNode(ISD::SUB, DL, VT, + DAG.getConstant(0, MVT::i32),Res); + + } else + return SDValue(); + } if (ShiftAmt != 0) - Res = DAG.getNode(ISD::SHL, DL, VT, Res, - DAG.getConstant(ShiftAmt, MVT::i32)); + Res = DAG.getNode(ISD::SHL, DL, VT, + Res, DAG.getConstant(ShiftAmt, MVT::i32)); // Do not add new nodes to DAG combiner worklist. DCI.CombineTo(N, Res, false); return SDValue(); } +static bool isCMOVWithZeroOrAllOnesLHS(SDValue N, bool AllOnes) { + if (N.getOpcode() != ARMISD::CMOV || !N.getNode()->hasOneUse()) + return false; + + SDValue FalseVal = N.getOperand(0); + ConstantSDNode *C = dyn_cast(FalseVal); + if (!C) + return false; + if (AllOnes) + return C->isAllOnesValue(); + return C->isNullValue(); +} + +/// formConditionalOp - Combine an operation with a conditional move operand +/// to form a conditional op. e.g. (or x, (cmov 0, y, cond)) => (or.cond x, y) +/// (and x, (cmov -1, y, cond)) => (and.cond, x, y) +static SDValue formConditionalOp(SDNode *N, SelectionDAG &DAG, + bool Commutable) { + SDValue N0 = N->getOperand(0); + SDValue N1 = N->getOperand(1); + + bool isAND = N->getOpcode() == ISD::AND; + bool isCand = isCMOVWithZeroOrAllOnesLHS(N1, isAND); + if (!isCand && Commutable) { + isCand = isCMOVWithZeroOrAllOnesLHS(N0, isAND); + if (isCand) + std::swap(N0, N1); + } + if (!isCand) + return SDValue(); + + unsigned Opc = 0; + switch (N->getOpcode()) { + default: llvm_unreachable("Unexpected node"); + case ISD::AND: Opc = ARMISD::CAND; break; + case ISD::OR: Opc = ARMISD::COR; break; + case ISD::XOR: Opc = ARMISD::CXOR; break; + } + return DAG.getNode(Opc, N->getDebugLoc(), N->getValueType(0), N0, + N1.getOperand(1), N1.getOperand(2), N1.getOperand(3), + N1.getOperand(4)); +} + static SDValue PerformANDCombine(SDNode *N, - TargetLowering::DAGCombinerInfo &DCI) { + TargetLowering::DAGCombinerInfo &DCI, + const ARMSubtarget *Subtarget) { // Attempt to use immediate-form VBIC BuildVectorSDNode *BVN = dyn_cast(N->getOperand(1)); @@ -6851,6 +7051,13 @@ static SDValue PerformANDCombine(SDNode *N, } } + if (!Subtarget->isThumb1Only()) { + // (and x, (cmov -1, y, cond)) => (and.cond x, y) + SDValue CAND = formConditionalOp(N, DAG, true); + if (CAND.getNode()) + return CAND; + } + return SDValue(); } @@ -6887,6 +7094,13 @@ static SDValue PerformORCombine(SDNode *N, } } + if (!Subtarget->isThumb1Only()) { + // (or x, (cmov 0, y, cond)) => (or.cond x, y) + SDValue COR = formConditionalOp(N, DAG, true); + if (COR.getNode()) + return COR; + } + SDValue N0 = N->getOperand(0); if (N0.getOpcode() != ISD::AND) return SDValue(); @@ -7035,6 +7249,25 @@ static SDValue PerformORCombine(SDNode *N, return SDValue(); } +static SDValue PerformXORCombine(SDNode *N, + TargetLowering::DAGCombinerInfo &DCI, + const ARMSubtarget *Subtarget) { + EVT VT = N->getValueType(0); + SelectionDAG &DAG = DCI.DAG; + + if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) + return SDValue(); + + if (!Subtarget->isThumb1Only()) { + // (xor x, (cmov 0, y, cond)) => (xor.cond x, y) + SDValue CXOR = formConditionalOp(N, DAG, true); + if (CXOR.getNode()) + return CXOR; + } + + return SDValue(); +} + /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff /// the bits being cleared by the AND are not demanded by the BFI. static SDValue PerformBFICombine(SDNode *N, @@ -7122,15 +7355,99 @@ static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) { /// ISD::STORE. static SDValue PerformSTORECombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { - // Bitcast an i64 store extracted from a vector to f64. - // Otherwise, the i64 value will be legalized to a pair of i32 values. StoreSDNode *St = cast(N); + if (St->isVolatile()) + return SDValue(); + + // Optimize trunc store (of multiple scalars) to shuffle and store. First, + // pack all of the elements in one place. Next, store to memory in fewer + // chunks. SDValue StVal = St->getValue(); - if (!ISD::isNormalStore(St) || St->isVolatile()) + EVT VT = StVal.getValueType(); + if (St->isTruncatingStore() && VT.isVector()) { + SelectionDAG &DAG = DCI.DAG; + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); + EVT StVT = St->getMemoryVT(); + unsigned NumElems = VT.getVectorNumElements(); + assert(StVT != VT && "Cannot truncate to the same type"); + unsigned FromEltSz = VT.getVectorElementType().getSizeInBits(); + unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits(); + + // From, To sizes and ElemCount must be pow of two + if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue(); + + // We are going to use the original vector elt for storing. + // Accumulated smaller vector elements must be a multiple of the store size. + if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue(); + + unsigned SizeRatio = FromEltSz / ToEltSz; + assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits()); + + // Create a type on which we perform the shuffle. + EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(), + NumElems*SizeRatio); + assert(WideVecVT.getSizeInBits() == VT.getSizeInBits()); + + DebugLoc DL = St->getDebugLoc(); + SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal); + SmallVector ShuffleVec(NumElems * SizeRatio, -1); + for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio; + + // Can't shuffle using an illegal type. + if (!TLI.isTypeLegal(WideVecVT)) return SDValue(); + + SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec, + DAG.getUNDEF(WideVec.getValueType()), + ShuffleVec.data()); + // At this point all of the data is stored at the bottom of the + // register. We now need to save it to mem. + + // Find the largest store unit + MVT StoreType = MVT::i8; + for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE; + tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) { + MVT Tp = (MVT::SimpleValueType)tp; + if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz) + StoreType = Tp; + } + // Didn't find a legal store type. + if (!TLI.isTypeLegal(StoreType)) + return SDValue(); + + // Bitcast the original vector into a vector of store-size units + EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(), + StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits()); + assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits()); + SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff); + SmallVector Chains; + SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, + TLI.getPointerTy()); + SDValue BasePtr = St->getBasePtr(); + + // Perform one or more big stores into memory. + unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits(); + for (unsigned I = 0; I < E; I++) { + SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, + StoreType, ShuffWide, + DAG.getIntPtrConstant(I)); + SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr, + St->getPointerInfo(), St->isVolatile(), + St->isNonTemporal(), St->getAlignment()); + BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr, + Increment); + Chains.push_back(Ch); + } + return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0], + Chains.size()); + } + + if (!ISD::isNormalStore(St)) return SDValue(); + // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and + // ARM stores of arguments in the same cache line. if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR && - StVal.getNode()->hasOneUse() && !St->isVolatile()) { + StVal.getNode()->hasOneUse()) { SelectionDAG &DAG = DCI.DAG; DebugLoc DL = St->getDebugLoc(); SDValue BasePtr = St->getBasePtr(); @@ -7151,6 +7468,8 @@ static SDValue PerformSTORECombine(SDNode *N, StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT) return SDValue(); + // Bitcast an i64 store extracted from a vector to f64. + // Otherwise, the i64 value will be legalized to a pair of i32 values. SelectionDAG &DAG = DCI.DAG; DebugLoc dl = StVal.getDebugLoc(); SDValue IntVec = StVal.getOperand(0); @@ -7332,7 +7651,7 @@ static SDValue CombineBaseUpdate(SDNode *N, if (isIntrinsic) { unsigned IntNo = cast(N->getOperand(1))->getZExtValue(); switch (IntNo) { - default: assert(0 && "unexpected intrinsic for Neon base update"); + default: llvm_unreachable("unexpected intrinsic for Neon base update"); case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD; NumVecs = 1; break; case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD; @@ -7365,7 +7684,7 @@ static SDValue CombineBaseUpdate(SDNode *N, } else { isLaneOp = true; switch (N->getOpcode()) { - default: assert(0 && "unexpected opcode for Neon base update"); + default: llvm_unreachable("unexpected opcode for Neon base update"); case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break; case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break; case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break; @@ -7858,6 +8177,18 @@ static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) { static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG, const ARMSubtarget *ST) { EVT VT = N->getValueType(0); + if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) { + // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high + // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16. + SDValue N1 = N->getOperand(1); + if (ConstantSDNode *C = dyn_cast(N1)) { + SDValue N0 = N->getOperand(0); + if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP && + DAG.MaskedValueIsZero(N0.getOperand(0), + APInt::getHighBitsSet(32, 16))) + return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1); + } + } // Nothing to be done for scalar shifts. const TargetLowering &TLI = DAG.getTargetLoweringInfo(); @@ -8061,8 +8392,7 @@ ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const { if (Res.getNode()) { APInt KnownZero, KnownOne; - APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()); - DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne); + DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne); // Capture demanded bits information that would be otherwise lost. if (KnownZero == 0xfffffffe) Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, @@ -8086,7 +8416,8 @@ SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N, case ISD::SUB: return PerformSUBCombine(N, DCI); case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget); case ISD::OR: return PerformORCombine(N, DCI, Subtarget); - case ISD::AND: return PerformANDCombine(N, DCI); + case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget); + case ISD::AND: return PerformANDCombine(N, DCI, Subtarget); case ARMISD::BFI: return PerformBFICombine(N, DCI); case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI); case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG); @@ -8387,10 +8718,12 @@ bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM, /// a register against the immediate without having to materialize the /// immediate into a register. bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const { + // Thumb2 and ARM modes can use cmn for negative immediates. if (!Subtarget->isThumb()) - return ARM_AM::getSOImmVal(Imm) != -1; + return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1; if (Subtarget->isThumb2()) - return ARM_AM::getT2SOImmVal(Imm) != -1; + return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1; + // Thumb1 doesn't have cmn, and only 8-bit immediates. return Imm >= 0 && Imm <= 255; } @@ -8577,22 +8910,20 @@ bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op, } void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, - const APInt &Mask, APInt &KnownZero, APInt &KnownOne, const SelectionDAG &DAG, unsigned Depth) const { - KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); + KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); switch (Op.getOpcode()) { default: break; case ARMISD::CMOV: { // Bits are known zero/one if known on the LHS and RHS. - DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1); + DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1); if (KnownZero == 0 && KnownOne == 0) return; APInt KnownZeroRHS, KnownOneRHS; - DAG.ComputeMaskedBits(Op.getOperand(1), Mask, - KnownZeroRHS, KnownOneRHS, Depth+1); + DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1); KnownZero &= KnownZeroRHS; KnownOne &= KnownOneRHS; return; @@ -8705,39 +9036,38 @@ ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, switch (Constraint[0]) { case 'l': // Low regs or general regs. if (Subtarget->isThumb()) - return RCPair(0U, ARM::tGPRRegisterClass); - else - return RCPair(0U, ARM::GPRRegisterClass); + return RCPair(0U, &ARM::tGPRRegClass); + return RCPair(0U, &ARM::GPRRegClass); case 'h': // High regs or no regs. if (Subtarget->isThumb()) - return RCPair(0U, ARM::hGPRRegisterClass); + return RCPair(0U, &ARM::hGPRRegClass); break; case 'r': - return RCPair(0U, ARM::GPRRegisterClass); + return RCPair(0U, &ARM::GPRRegClass); case 'w': if (VT == MVT::f32) - return RCPair(0U, ARM::SPRRegisterClass); + return RCPair(0U, &ARM::SPRRegClass); if (VT.getSizeInBits() == 64) - return RCPair(0U, ARM::DPRRegisterClass); + return RCPair(0U, &ARM::DPRRegClass); if (VT.getSizeInBits() == 128) - return RCPair(0U, ARM::QPRRegisterClass); + return RCPair(0U, &ARM::QPRRegClass); break; case 'x': if (VT == MVT::f32) - return RCPair(0U, ARM::SPR_8RegisterClass); + return RCPair(0U, &ARM::SPR_8RegClass); if (VT.getSizeInBits() == 64) - return RCPair(0U, ARM::DPR_8RegisterClass); + return RCPair(0U, &ARM::DPR_8RegClass); if (VT.getSizeInBits() == 128) - return RCPair(0U, ARM::QPR_8RegisterClass); + return RCPair(0U, &ARM::QPR_8RegClass); break; case 't': if (VT == MVT::f32) - return RCPair(0U, ARM::SPRRegisterClass); + return RCPair(0U, &ARM::SPRRegClass); break; } } if (StringRef("{cc}").equals_lower(Constraint)) - return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass); + return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass); return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); }