diff options
Diffstat (limited to 'lib/Target/ARM/ARMISelLowering.cpp')
| -rw-r--r-- | lib/Target/ARM/ARMISelLowering.cpp | 958 |
1 files changed, 786 insertions, 172 deletions
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp index ab9f9e1..0a31b87 100644 --- a/lib/Target/ARM/ARMISelLowering.cpp +++ b/lib/Target/ARM/ARMISelLowering.cpp @@ -72,6 +72,11 @@ ARMInterworking("arm-interworking", cl::Hidden, cl::desc("Enable / disable ARM interworking (for debugging only)"), cl::init(true)); +// The APCS parameter registers. +static const unsigned GPRArgRegs[] = { + ARM::R0, ARM::R1, ARM::R2, ARM::R3 +}; + void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT, EVT PromotedBitwiseVT) { if (VT != PromotedLdStVT) { @@ -393,6 +398,12 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS); } + // Use divmod iOS compiler-rt calls. + if (Subtarget->getTargetTriple().getOS() == Triple::IOS) { + setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4"); + setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4"); + } + if (Subtarget->isThumb1Only()) addRegisterClass(MVT::i32, ARM::tGPRRegisterClass); else @@ -461,6 +472,10 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setOperationAction(ISD::UDIV, MVT::v8i8, Custom); setOperationAction(ISD::VSETCC, MVT::v1i64, Expand); setOperationAction(ISD::VSETCC, 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. + setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom); + setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); setTargetDAGCombine(ISD::INTRINSIC_VOID); setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); @@ -502,18 +517,15 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) } // i64 operation support. + setOperationAction(ISD::MUL, MVT::i64, Expand); + setOperationAction(ISD::MULHU, MVT::i32, Expand); if (Subtarget->isThumb1Only()) { - setOperationAction(ISD::MUL, MVT::i64, Expand); - setOperationAction(ISD::MULHU, MVT::i32, Expand); - setOperationAction(ISD::MULHS, MVT::i32, Expand); setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); - } else { - setOperationAction(ISD::MUL, MVT::i64, Expand); - setOperationAction(ISD::MULHU, MVT::i32, Expand); - if (!Subtarget->hasV6Ops()) - setOperationAction(ISD::MULHS, MVT::i32, Expand); } + if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()) + setOperationAction(ISD::MULHS, MVT::i32, Expand); + setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); @@ -597,6 +609,18 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand); setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand); setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand); + setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i8, Expand); + setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i16, Expand); + setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand); + setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i8, Expand); + setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i16, Expand); + setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand); + setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i8, Expand); + setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i16, Expand); + setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand); + setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i8, Expand); + setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i16, Expand); + setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand); // Since the libcalls include locking, fold in the fences setShouldFoldAtomicFences(true); } @@ -716,7 +740,7 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) // pressure of the register class's representative and all of it's super // classes' representatives transitively. We have not implemented this because // of the difficulty prior to coalescing of modeling operand register classes -// due to the common occurence of cross class copies and subregister insertions +// due to the common occurrence of cross class copies and subregister insertions // and extractions. std::pair<const TargetRegisterClass*, uint8_t> ARMTargetLowering::findRepresentativeClass(EVT VT) const{ @@ -778,7 +802,6 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { case ARMISD::BCC_i64: return "ARMISD::BCC_i64"; case ARMISD::FMSTAT: return "ARMISD::FMSTAT"; case ARMISD::CMOV: return "ARMISD::CMOV"; - case ARMISD::CNEG: return "ARMISD::CNEG"; case ARMISD::RBIT: return "ARMISD::RBIT"; @@ -853,6 +876,8 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { case ARMISD::VZIP: return "ARMISD::VZIP"; case ARMISD::VUZP: return "ARMISD::VUZP"; case ARMISD::VTRN: return "ARMISD::VTRN"; + case ARMISD::VTBL1: return "ARMISD::VTBL1"; + case ARMISD::VTBL2: return "ARMISD::VTBL2"; case ARMISD::VMULLs: return "ARMISD::VMULLs"; case ARMISD::VMULLu: return "ARMISD::VMULLu"; case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR"; @@ -861,6 +886,7 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { case ARMISD::BFI: return "ARMISD::BFI"; case ARMISD::VORRIMM: return "ARMISD::VORRIMM"; case ARMISD::VBICIMM: return "ARMISD::VBICIMM"; + case ARMISD::VBSL: return "ARMISD::VBSL"; case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP"; case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP"; case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP"; @@ -946,27 +972,6 @@ Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const { return Sched::RegPressure; } -// FIXME: Move to RegInfo -unsigned -ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC, - MachineFunction &MF) const { - const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering(); - - switch (RC->getID()) { - default: - return 0; - case ARM::tGPRRegClassID: - return TFI->hasFP(MF) ? 4 : 5; - case ARM::GPRRegClassID: { - unsigned FP = TFI->hasFP(MF) ? 1 : 0; - return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0); - } - case ARM::SPRRegClassID: // Currently not used as 'rep' register class. - case ARM::DPRRegClassID: - return 32 - 10; - } -} - //===----------------------------------------------------------------------===// // Lowering Code //===----------------------------------------------------------------------===// @@ -1130,22 +1135,6 @@ ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, return Chain; } -/// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified -/// by "Src" to address "Dst" of size "Size". Alignment information is -/// specified by the specific parameter attribute. The copy will be passed as -/// a byval function parameter. -/// Sometimes what we are copying is the end of a larger object, the part that -/// does not fit in registers. -static SDValue -CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain, - ISD::ArgFlagsTy Flags, SelectionDAG &DAG, - DebugLoc dl) { - SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32); - return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(), - /*isVolatile=*/false, /*AlwaysInline=*/false, - MachinePointerInfo(0), MachinePointerInfo(0)); -} - /// LowerMemOpCallTo - Store the argument to the stack. SDValue ARMTargetLowering::LowerMemOpCallTo(SDValue Chain, @@ -1156,9 +1145,6 @@ ARMTargetLowering::LowerMemOpCallTo(SDValue Chain, unsigned LocMemOffset = VA.getLocMemOffset(); SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset); PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff); - if (Flags.isByVal()) - return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl); - return DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo::getStack(LocMemOffset), false, false, 0); @@ -1224,6 +1210,7 @@ ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, SmallVector<CCValAssign, 16> ArgLocs; CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext()); + CCInfo.setCallOrPrologue(Call); CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv, /* Return*/ false, isVarArg)); @@ -1253,6 +1240,7 @@ ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, CCValAssign &VA = ArgLocs[i]; SDValue Arg = OutVals[realArgIdx]; ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; + bool isByVal = Flags.isByVal(); // Promote the value if needed. switch (VA.getLocInfo()) { @@ -1299,6 +1287,43 @@ ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, } } else if (VA.isRegLoc()) { RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); + } else if (isByVal) { + assert(VA.isMemLoc()); + unsigned offset = 0; + + // True if this byval aggregate will be split between registers + // and memory. + if (CCInfo.isFirstByValRegValid()) { + EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); + unsigned int i, j; + for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) { + SDValue Const = DAG.getConstant(4*i, MVT::i32); + SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const); + SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg, + MachinePointerInfo(), + false, false, 0); + MemOpChains.push_back(Load.getValue(1)); + RegsToPass.push_back(std::make_pair(j, Load)); + } + offset = ARM::R4 - CCInfo.getFirstByValReg(); + CCInfo.clearFirstByValReg(); + } + + unsigned LocMemOffset = VA.getLocMemOffset(); + SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset); + SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, + StkPtrOff); + SDValue SrcOffset = DAG.getIntPtrConstant(4*offset); + SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset); + SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, + MVT::i32); + MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, + Flags.getByValAlign(), + /*isVolatile=*/false, + /*AlwaysInline=*/false, + MachinePointerInfo(0), + MachinePointerInfo(0))); + } else if (!IsSibCall) { assert(VA.isMemLoc()); @@ -1332,7 +1357,7 @@ ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, // than necessary, because it means that each store effectively depends // on every argument instead of just those arguments it would clobber. - // Do not flag preceeding copytoreg stuff together with the following stuff. + // Do not flag preceding copytoreg stuff together with the following stuff. InFlag = SDValue(); for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, @@ -1492,6 +1517,35 @@ ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, dl, DAG, InVals); } +/// HandleByVal - Every parameter *after* a byval parameter is passed +/// on the stack. Remember the next parameter register to allocate, +/// and then confiscate the rest of the parameter registers to insure +/// this. +void +llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const { + unsigned reg = State->AllocateReg(GPRArgRegs, 4); + assert((State->getCallOrPrologue() == Prologue || + State->getCallOrPrologue() == Call) && + "unhandled ParmContext"); + if ((!State->isFirstByValRegValid()) && + (ARM::R0 <= reg) && (reg <= ARM::R3)) { + State->setFirstByValReg(reg); + // At a call site, a byval parameter that is split between + // registers and memory needs its size truncated here. In a + // function prologue, such byval parameters are reassembled in + // memory, and are not truncated. + if (State->getCallOrPrologue() == Call) { + unsigned excess = 4 * (ARM::R4 - reg); + assert(size >= excess && "expected larger existing stack allocation"); + size -= excess; + } + } + // Confiscate any remaining parameter registers to preclude their + // assignment to subsequent parameters. + while (State->AllocateReg(GPRArgRegs, 4)) + ; +} + /// MatchingStackOffset - Return true if the given stack call argument is /// already available in the same position (relatively) of the caller's /// incoming argument stack. @@ -1813,6 +1867,16 @@ bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const { return HasRet; } +bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const { + if (!EnableARMTailCalls) + return false; + + if (!CI->isTailCall()) + return false; + + return !Subtarget->isThumb1Only(); +} + // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is // one of the above mentioned nodes. It has to be wrapped because otherwise @@ -2096,7 +2160,7 @@ ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG) const { DebugLoc dl = Op.getDebugLoc(); return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other, - Op.getOperand(0), Op.getOperand(1)); + Op.getOperand(0)); } SDValue @@ -2151,6 +2215,13 @@ ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG, } return Result; } + case Intrinsic::arm_neon_vmulls: + case Intrinsic::arm_neon_vmullu: { + unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls) + ? ARMISD::VMULLs : ARMISD::VMULLu; + return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(), + Op.getOperand(1), Op.getOperand(2)); + } } } @@ -2257,6 +2328,88 @@ ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2); } +void +ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF, + unsigned &VARegSize, unsigned &VARegSaveSize) + const { + unsigned NumGPRs; + if (CCInfo.isFirstByValRegValid()) + NumGPRs = ARM::R4 - CCInfo.getFirstByValReg(); + else { + unsigned int firstUnalloced; + firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs, + sizeof(GPRArgRegs) / + sizeof(GPRArgRegs[0])); + NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0; + } + + unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment(); + VARegSize = NumGPRs * 4; + VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1); +} + +// The remaining GPRs hold either the beginning of variable-argument +// data, or the beginning of an aggregate passed by value (usuall +// byval). Either way, we allocate stack slots adjacent to the data +// provided by our caller, and store the unallocated registers there. +// If this is a variadic function, the va_list pointer will begin with +// these values; otherwise, this reassembles a (byval) structure that +// was split between registers and memory. +void +ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, + DebugLoc dl, SDValue &Chain, + unsigned ArgOffset) const { + MachineFunction &MF = DAG.getMachineFunction(); + MachineFrameInfo *MFI = MF.getFrameInfo(); + ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); + unsigned firstRegToSaveIndex; + if (CCInfo.isFirstByValRegValid()) + firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0; + else { + firstRegToSaveIndex = CCInfo.getFirstUnallocated + (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0])); + } + + unsigned VARegSize, VARegSaveSize; + computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize); + if (VARegSaveSize) { + // If this function is vararg, store any remaining integer argument regs + // to their spots on the stack so that they may be loaded by deferencing + // the result of va_next. + AFI->setVarArgsRegSaveSize(VARegSaveSize); + AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize, + ArgOffset + VARegSaveSize + - VARegSize, + false)); + SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(), + getPointerTy()); + + SmallVector<SDValue, 4> MemOps; + for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) { + TargetRegisterClass *RC; + if (AFI->isThumb1OnlyFunction()) + RC = ARM::tGPRRegisterClass; + else + RC = ARM::GPRRegisterClass; + + unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC); + SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); + SDValue Store = + DAG.getStore(Val.getValue(1), dl, Val, FIN, + MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()), + false, false, 0); + MemOps.push_back(Store); + FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN, + DAG.getConstant(4, getPointerTy())); + } + if (!MemOps.empty()) + Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, + &MemOps[0], MemOps.size()); + } else + // This will point to the next argument passed via stack. + AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true)); +} + SDValue ARMTargetLowering::LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, @@ -2265,7 +2418,6 @@ ARMTargetLowering::LowerFormalArguments(SDValue Chain, DebugLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { - MachineFunction &MF = DAG.getMachineFunction(); MachineFrameInfo *MFI = MF.getFrameInfo(); @@ -2275,12 +2427,15 @@ ARMTargetLowering::LowerFormalArguments(SDValue Chain, SmallVector<CCValAssign, 16> ArgLocs; CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext()); + CCInfo.setCallOrPrologue(Prologue); CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv, /* Return*/ false, isVarArg)); SmallVector<SDValue, 16> ArgValues; + int lastInsIndex = -1; + SDValue ArgValue; for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { CCValAssign &VA = ArgLocs[i]; @@ -2288,7 +2443,6 @@ ARMTargetLowering::LowerFormalArguments(SDValue Chain, if (VA.isRegLoc()) { EVT RegVT = VA.getLocVT(); - SDValue ArgValue; if (VA.needsCustom()) { // f64 and vector types are split up into multiple registers or // combinations of registers and stack slots. @@ -2364,67 +2518,45 @@ ARMTargetLowering::LowerFormalArguments(SDValue Chain, assert(VA.isMemLoc()); assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered"); - unsigned ArgSize = VA.getLocVT().getSizeInBits()/8; - int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true); + int index = ArgLocs[i].getValNo(); + + // Some Ins[] entries become multiple ArgLoc[] entries. + // Process them only once. + if (index != lastInsIndex) + { + ISD::ArgFlagsTy Flags = Ins[index].Flags; + // FIXME: For now, all byval parameter objects are marked mutable. + // This can be changed with more analysis. + // In case of tail call optimization mark all arguments mutable. + // Since they could be overwritten by lowering of arguments in case of + // a tail call. + if (Flags.isByVal()) { + unsigned VARegSize, VARegSaveSize; + computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize); + VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0); + unsigned Bytes = Flags.getByValSize() - VARegSize; + if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects. + int FI = MFI->CreateFixedObject(Bytes, + VA.getLocMemOffset(), false); + InVals.push_back(DAG.getFrameIndex(FI, getPointerTy())); + } else { + int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8, + VA.getLocMemOffset(), true); - // Create load nodes to retrieve arguments from the stack. - SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); - InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, - MachinePointerInfo::getFixedStack(FI), - false, false, 0)); + // Create load nodes to retrieve arguments from the stack. + SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); + InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, + MachinePointerInfo::getFixedStack(FI), + false, false, 0)); + } + lastInsIndex = index; + } } } // varargs - if (isVarArg) { - static const unsigned GPRArgRegs[] = { - ARM::R0, ARM::R1, ARM::R2, ARM::R3 - }; - - unsigned NumGPRs = CCInfo.getFirstUnallocated - (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0])); - - unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment(); - unsigned VARegSize = (4 - NumGPRs) * 4; - unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1); - unsigned ArgOffset = CCInfo.getNextStackOffset(); - if (VARegSaveSize) { - // If this function is vararg, store any remaining integer argument regs - // to their spots on the stack so that they may be loaded by deferencing - // the result of va_next. - AFI->setVarArgsRegSaveSize(VARegSaveSize); - AFI->setVarArgsFrameIndex( - MFI->CreateFixedObject(VARegSaveSize, - ArgOffset + VARegSaveSize - VARegSize, - false)); - SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(), - getPointerTy()); - - SmallVector<SDValue, 4> MemOps; - for (; NumGPRs < 4; ++NumGPRs) { - TargetRegisterClass *RC; - if (AFI->isThumb1OnlyFunction()) - RC = ARM::tGPRRegisterClass; - else - RC = ARM::GPRRegisterClass; - - unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC); - SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); - SDValue Store = - DAG.getStore(Val.getValue(1), dl, Val, FIN, - MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()), - false, false, 0); - MemOps.push_back(Store); - FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN, - DAG.getConstant(4, getPointerTy())); - } - if (!MemOps.empty()) - Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, - &MemOps[0], MemOps.size()); - } else - // This will point to the next argument passed via stack. - AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true)); - } + if (isVarArg) + VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset()); return Chain; } @@ -2517,6 +2649,27 @@ ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG, return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp); } +/// duplicateCmp - Glue values can have only one use, so this function +/// duplicates a comparison node. +SDValue +ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const { + unsigned Opc = Cmp.getOpcode(); + DebugLoc DL = Cmp.getDebugLoc(); + if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ) + return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); + + assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation"); + Cmp = Cmp.getOperand(0); + Opc = Cmp.getOpcode(); + if (Opc == ARMISD::CMPFP) + Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); + else { + assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT"); + Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0)); + } + return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp); +} + SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { SDValue Cond = Op.getOperand(0); SDValue SelectTrue = Op.getOperand(1); @@ -2552,7 +2705,7 @@ SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { EVT VT = Cond.getValueType(); SDValue ARMcc = Cond.getOperand(2); SDValue CCR = Cond.getOperand(3); - SDValue Cmp = Cond.getOperand(4); + SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG); return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp); } } @@ -2681,8 +2834,8 @@ ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const { // 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)))) { - // If unsafe fp math optimization is enabled and there are no othter uses of - // the CMP operands, and the condition code is EQ oe NE, we can optimize it + // 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. if (CC == ISD::SETOEQ) CC = ISD::SETEQ; @@ -2811,8 +2964,39 @@ static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) { return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); } +static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) { + EVT VT = Op.getValueType(); + DebugLoc dl = Op.getDebugLoc(); + + EVT OperandVT = Op.getOperand(0).getValueType(); + assert(OperandVT == MVT::v4i16 && "Invalid type for custom lowering!"); + if (VT != MVT::v4f32) + return DAG.UnrollVectorOp(Op.getNode()); + + unsigned CastOpc; + unsigned Opc; + switch (Op.getOpcode()) { + default: + assert(0 && "Invalid opcode!"); + case ISD::SINT_TO_FP: + CastOpc = ISD::SIGN_EXTEND; + Opc = ISD::SINT_TO_FP; + break; + case ISD::UINT_TO_FP: + CastOpc = ISD::ZERO_EXTEND; + Opc = ISD::UINT_TO_FP; + break; + } + + Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0)); + return DAG.getNode(Opc, dl, VT, Op); +} + static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) { EVT VT = Op.getValueType(); + if (VT.isVector()) + return LowerVectorINT_TO_FP(Op, DAG); + DebugLoc dl = Op.getDebugLoc(); unsigned Opc; @@ -2860,7 +3044,10 @@ SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT, DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1), DAG.getConstant(32, MVT::i32)); - } + } else if (VT == MVT::f32) + Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64, + DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1), + DAG.getConstant(32, MVT::i32)); Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0); Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1); @@ -2869,11 +3056,11 @@ SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes); SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask, DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes)); - + SDValue Res = DAG.getNode(ISD::OR, dl, OpVT, DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask), DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot)); - if (SrcVT == MVT::f32) { + if (VT == MVT::f32) { Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res); Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res, DAG.getConstant(0, MVT::i32)); @@ -3508,6 +3695,13 @@ static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT, return true; } +static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) { + // We can handle <8 x i8> vector shuffles. If the index in the mask is out of + // range, then 0 is placed into the resulting vector. So pretty much any mask + // of 8 elements can work here. + return VT == MVT::v8i8 && M.size() == 8; +} + static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT, unsigned &WhichResult) { unsigned EltSz = VT.getVectorElementType().getSizeInBits(); @@ -3947,6 +4141,7 @@ ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M, isVREVMask(M, VT, 32) || isVREVMask(M, VT, 16) || isVEXTMask(M, VT, ReverseVEXT, Imm) || + isVTBLMask(M, VT) || isVTRNMask(M, VT, WhichResult) || isVUZPMask(M, VT, WhichResult) || isVZIPMask(M, VT, WhichResult) || @@ -4024,6 +4219,29 @@ static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS, } } +static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op, + SmallVectorImpl<int> &ShuffleMask, + SelectionDAG &DAG) { + // Check to see if we can use the VTBL instruction. + SDValue V1 = Op.getOperand(0); + SDValue V2 = Op.getOperand(1); + DebugLoc DL = Op.getDebugLoc(); + + SmallVector<SDValue, 8> VTBLMask; + for (SmallVectorImpl<int>::iterator + I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I) + VTBLMask.push_back(DAG.getConstant(*I, MVT::i32)); + + if (V2.getNode()->getOpcode() == ISD::UNDEF) + return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1, + DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, + &VTBLMask[0], 8)); + + return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2, + DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, + &VTBLMask[0], 8)); +} + static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) { SDValue V1 = Op.getOperand(0); SDValue V2 = Op.getOperand(1); @@ -4141,6 +4359,12 @@ static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) { return DAG.getNode(ISD::BITCAST, dl, VT, Val); } + if (VT == MVT::v8i8) { + SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG); + if (NewOp.getNode()) + return NewOp; + } + return SDValue(); } @@ -4290,6 +4514,28 @@ static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) { MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts); } +static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) { + unsigned Opcode = N->getOpcode(); + if (Opcode == ISD::ADD || Opcode == ISD::SUB) { + SDNode *N0 = N->getOperand(0).getNode(); + SDNode *N1 = N->getOperand(1).getNode(); + return N0->hasOneUse() && N1->hasOneUse() && + isSignExtended(N0, DAG) && isSignExtended(N1, DAG); + } + return false; +} + +static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) { + unsigned Opcode = N->getOpcode(); + if (Opcode == ISD::ADD || Opcode == ISD::SUB) { + SDNode *N0 = N->getOperand(0).getNode(); + SDNode *N1 = N->getOperand(1).getNode(); + return N0->hasOneUse() && N1->hasOneUse() && + isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG); + } + return false; +} + static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) { // Multiplications are only custom-lowered for 128-bit vectors so that // VMULL can be detected. Otherwise v2i64 multiplications are not legal. @@ -4298,29 +4544,73 @@ static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) { SDNode *N0 = Op.getOperand(0).getNode(); SDNode *N1 = Op.getOperand(1).getNode(); unsigned NewOpc = 0; - if (isSignExtended(N0, DAG) && isSignExtended(N1, DAG)) + bool isMLA = false; + bool isN0SExt = isSignExtended(N0, DAG); + bool isN1SExt = isSignExtended(N1, DAG); + if (isN0SExt && isN1SExt) NewOpc = ARMISD::VMULLs; - else if (isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG)) - NewOpc = ARMISD::VMULLu; - else if (VT == MVT::v2i64) - // Fall through to expand this. It is not legal. - return SDValue(); - else - // Other vector multiplications are legal. - return Op; + else { + bool isN0ZExt = isZeroExtended(N0, DAG); + bool isN1ZExt = isZeroExtended(N1, DAG); + if (isN0ZExt && isN1ZExt) + NewOpc = ARMISD::VMULLu; + else if (isN1SExt || isN1ZExt) { + // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these + // into (s/zext A * s/zext C) + (s/zext B * s/zext C) + if (isN1SExt && isAddSubSExt(N0, DAG)) { + NewOpc = ARMISD::VMULLs; + isMLA = true; + } else if (isN1ZExt && isAddSubZExt(N0, DAG)) { + NewOpc = ARMISD::VMULLu; + isMLA = true; + } else if (isN0ZExt && isAddSubZExt(N1, DAG)) { + std::swap(N0, N1); + NewOpc = ARMISD::VMULLu; + isMLA = true; + } + } + + if (!NewOpc) { + if (VT == MVT::v2i64) + // Fall through to expand this. It is not legal. + return SDValue(); + else + // Other vector multiplications are legal. + return Op; + } + } // Legalize to a VMULL instruction. DebugLoc DL = Op.getDebugLoc(); - SDValue Op0 = SkipExtension(N0, DAG); + SDValue Op0; SDValue Op1 = SkipExtension(N1, DAG); - - assert(Op0.getValueType().is64BitVector() && - Op1.getValueType().is64BitVector() && - "unexpected types for extended operands to VMULL"); - return DAG.getNode(NewOpc, DL, VT, Op0, Op1); + if (!isMLA) { + Op0 = SkipExtension(N0, DAG); + assert(Op0.getValueType().is64BitVector() && + Op1.getValueType().is64BitVector() && + "unexpected types for extended operands to VMULL"); + return DAG.getNode(NewOpc, DL, VT, Op0, Op1); + } + + // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during + // isel lowering to take advantage of no-stall back to back vmul + vmla. + // vmull q0, d4, d6 + // vmlal q0, d5, d6 + // is faster than + // vaddl q0, d4, d5 + // vmovl q1, d6 + // vmul q0, q0, q1 + SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG); + SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG); + EVT Op1VT = Op1.getValueType(); + return DAG.getNode(N0->getOpcode(), DL, VT, + DAG.getNode(NewOpc, DL, VT, + DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1), + DAG.getNode(NewOpc, DL, VT, + DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1)); } -static SDValue +static SDValue LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) { // Convert to float // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo)); @@ -4331,7 +4621,7 @@ LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) { Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y); // Get reciprocal estimate. // float4 recip = vrecpeq_f32(yf); - Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, + Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y); // Because char has a smaller range than uchar, we can actually get away // without any newton steps. This requires that we use a weird bias @@ -4349,7 +4639,7 @@ LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) { return X; } -static SDValue +static SDValue LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) { SDValue N2; // Convert to float. @@ -4359,13 +4649,13 @@ LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) { N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1); N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0); N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1); - + // Use reciprocal estimate and one refinement step. // float4 recip = vrecpeq_f32(yf); // recip *= vrecpsq_f32(yf, recip); - N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, + N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1); - N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, + N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), N1, N2); N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); @@ -4395,15 +4685,15 @@ static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) { SDValue N0 = Op.getOperand(0); SDValue N1 = Op.getOperand(1); SDValue N2, N3; - + if (VT == MVT::v8i8) { N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0); N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1); - + N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, DAG.getIntPtrConstant(4)); N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, - DAG.getIntPtrConstant(4)); + DAG.getIntPtrConstant(4)); N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, DAG.getIntPtrConstant(0)); N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, @@ -4414,7 +4704,7 @@ static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) { N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); N0 = LowerCONCAT_VECTORS(N0, DAG); - + N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0); return N0; } @@ -4430,32 +4720,32 @@ static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) { SDValue N0 = Op.getOperand(0); SDValue N1 = Op.getOperand(1); SDValue N2, N3; - + if (VT == MVT::v8i8) { N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0); N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1); - + N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, DAG.getIntPtrConstant(4)); N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, - DAG.getIntPtrConstant(4)); + DAG.getIntPtrConstant(4)); N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, DAG.getIntPtrConstant(0)); N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, DAG.getIntPtrConstant(0)); - + N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16 - + N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); N0 = LowerCONCAT_VECTORS(N0, DAG); - - N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8, + + N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8, DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32), N0); return N0; } - + // v4i16 sdiv ... Convert to float. // float4 yf = vcvt_f32_s32(vmovl_u16(y)); // float4 xf = vcvt_f32_s32(vmovl_u16(x)); @@ -4468,13 +4758,13 @@ static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) { // float4 recip = vrecpeq_f32(yf); // recip *= vrecpsq_f32(yf, recip); // recip *= vrecpsq_f32(yf, recip); - N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, + N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1); - N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, + N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), N1, N2); N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); - N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, + N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), N1, N2); N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); @@ -4503,7 +4793,7 @@ SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { case ISD::GlobalAddress: return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) : LowerGlobalAddressELF(Op, DAG); - case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); + case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); case ISD::SELECT: return LowerSELECT(Op, DAG); case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); case ISD::BR_CC: return LowerBR_CC(Op, DAG); @@ -4524,7 +4814,7 @@ SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG); case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG, Subtarget); - case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG); + case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG); case ISD::SHL: case ISD::SRL: case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget); @@ -4754,6 +5044,109 @@ ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, return BB; } +MachineBasicBlock * +ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI, + MachineBasicBlock *BB, + unsigned Size, + bool signExtend, + ARMCC::CondCodes Cond) const { + const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); + + const BasicBlock *LLVM_BB = BB->getBasicBlock(); + MachineFunction *MF = BB->getParent(); + MachineFunction::iterator It = BB; + ++It; + + unsigned dest = MI->getOperand(0).getReg(); + unsigned ptr = MI->getOperand(1).getReg(); + unsigned incr = MI->getOperand(2).getReg(); + unsigned oldval = dest; + DebugLoc dl = MI->getDebugLoc(); + + bool isThumb2 = Subtarget->isThumb2(); + unsigned ldrOpc, strOpc, extendOpc; + switch (Size) { + default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); + case 1: + ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; + strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; + extendOpc = isThumb2 ? ARM::t2SXTBr : ARM::SXTBr; + break; + case 2: + ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; + strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; + extendOpc = isThumb2 ? ARM::t2SXTHr : ARM::SXTHr; + break; + case 4: + ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; + strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; + extendOpc = 0; + break; + } + + MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); + MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); + MF->insert(It, loopMBB); + MF->insert(It, exitMBB); + + // Transfer the remainder of BB and its successor edges to exitMBB. + exitMBB->splice(exitMBB->begin(), BB, + llvm::next(MachineBasicBlock::iterator(MI)), + BB->end()); + exitMBB->transferSuccessorsAndUpdatePHIs(BB); + + MachineRegisterInfo &RegInfo = MF->getRegInfo(); + unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass); + unsigned scratch2 = RegInfo.createVirtualRegister(ARM::GPRRegisterClass); + + // thisMBB: + // ... + // fallthrough --> loopMBB + BB->addSuccessor(loopMBB); + + // loopMBB: + // ldrex dest, ptr + // (sign extend dest, if required) + // cmp dest, incr + // cmov.cond scratch2, dest, incr + // strex scratch, scratch2, ptr + // cmp scratch, #0 + // bne- loopMBB + // fallthrough --> exitMBB + BB = loopMBB; + AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr)); + + // Sign extend the value, if necessary. + if (signExtend && extendOpc) { + oldval = RegInfo.createVirtualRegister(ARM::GPRRegisterClass); + AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval).addReg(dest)); + } + + // Build compare and cmov instructions. + AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) + .addReg(oldval).addReg(incr)); + BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2) + .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR); + + AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2) + .addReg(ptr)); + AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) + .addReg(scratch).addImm(0)); + BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) + .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); + + BB->addSuccessor(loopMBB); + BB->addSuccessor(exitMBB); + + // exitMBB: + // ... + BB = exitMBB; + + MI->eraseFromParent(); // The instruction is gone now. + + return BB; +} + static MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) { for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(), @@ -4763,6 +5156,72 @@ MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) { llvm_unreachable("Expecting a BB with two successors!"); } +// FIXME: This opcode table should obviously be expressed in the target +// description. We probably just need a "machine opcode" value in the pseudo +// instruction. But the ideal solution maybe to simply remove the "S" version +// of the opcode altogether. +struct AddSubFlagsOpcodePair { + unsigned PseudoOpc; + unsigned MachineOpc; +}; + +static AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = { + {ARM::ADCSri, ARM::ADCri}, + {ARM::ADCSrr, ARM::ADCrr}, + {ARM::ADCSrs, ARM::ADCrs}, + {ARM::SBCSri, ARM::SBCri}, + {ARM::SBCSrr, ARM::SBCrr}, + {ARM::SBCSrs, ARM::SBCrs}, + {ARM::RSBSri, ARM::RSBri}, + {ARM::RSBSrr, ARM::RSBrr}, + {ARM::RSBSrs, ARM::RSBrs}, + {ARM::RSCSri, ARM::RSCri}, + {ARM::RSCSrs, ARM::RSCrs}, + {ARM::t2ADCSri, ARM::t2ADCri}, + {ARM::t2ADCSrr, ARM::t2ADCrr}, + {ARM::t2ADCSrs, ARM::t2ADCrs}, + {ARM::t2SBCSri, ARM::t2SBCri}, + {ARM::t2SBCSrr, ARM::t2SBCrr}, + {ARM::t2SBCSrs, ARM::t2SBCrs}, + {ARM::t2RSBSri, ARM::t2RSBri}, + {ARM::t2RSBSrs, ARM::t2RSBrs}, +}; + +// Convert and Add or Subtract with Carry and Flags to a generic opcode with +// CPSR<def> operand. e.g. ADCS (...) -> ADC (... CPSR<def>). +// +// FIXME: Somewhere we should assert that CPSR<def> is in the correct +// position to be recognized by the target descrition as the 'S' bit. +bool ARMTargetLowering::RemapAddSubWithFlags(MachineInstr *MI, + MachineBasicBlock *BB) const { + unsigned OldOpc = MI->getOpcode(); + unsigned NewOpc = 0; + + // This is only called for instructions that need remapping, so iterating over + // the tiny opcode table is not costly. + static const int NPairs = + sizeof(AddSubFlagsOpcodeMap) / sizeof(AddSubFlagsOpcodePair); + for (AddSubFlagsOpcodePair *Pair = &AddSubFlagsOpcodeMap[0], + *End = &AddSubFlagsOpcodeMap[NPairs]; Pair != End; ++Pair) { + if (OldOpc == Pair->PseudoOpc) { + NewOpc = Pair->MachineOpc; + break; + } + } + if (!NewOpc) + return false; + + const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); + DebugLoc dl = MI->getDebugLoc(); + MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc)); + for (unsigned i = 0; i < MI->getNumOperands(); ++i) + MIB.addOperand(MI->getOperand(i)); + AddDefaultPred(MIB); + MIB.addReg(ARM::CPSR, RegState::Define); // S bit + MI->eraseFromParent(); + return true; +} + MachineBasicBlock * ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *BB) const { @@ -4770,10 +5229,13 @@ ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, DebugLoc dl = MI->getDebugLoc(); bool isThumb2 = Subtarget->isThumb2(); switch (MI->getOpcode()) { - default: + default: { + if (RemapAddSubWithFlags(MI, BB)) + return BB; + MI->dump(); llvm_unreachable("Unexpected instr type to insert"); - + } case ARM::ATOMIC_LOAD_ADD_I8: return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); case ARM::ATOMIC_LOAD_ADD_I16: @@ -4816,6 +5278,34 @@ ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, case ARM::ATOMIC_LOAD_SUB_I32: return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); + case ARM::ATOMIC_LOAD_MIN_I8: + return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT); + case ARM::ATOMIC_LOAD_MIN_I16: + return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT); + case ARM::ATOMIC_LOAD_MIN_I32: + return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT); + + case ARM::ATOMIC_LOAD_MAX_I8: + return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT); + case ARM::ATOMIC_LOAD_MAX_I16: + return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT); + case ARM::ATOMIC_LOAD_MAX_I32: + return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT); + + case ARM::ATOMIC_LOAD_UMIN_I8: + return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO); + case ARM::ATOMIC_LOAD_UMIN_I16: + return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO); + case ARM::ATOMIC_LOAD_UMIN_I32: + return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO); + + case ARM::ATOMIC_LOAD_UMAX_I8: + return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI); + case ARM::ATOMIC_LOAD_UMAX_I16: + return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI); + case ARM::ATOMIC_LOAD_UMAX_I32: + return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI); + case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0); case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0); case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0); @@ -5034,6 +5524,42 @@ static SDValue PerformSUBCombine(SDNode *N, return SDValue(); } +/// PerformVMULCombine +/// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the +/// special multiplier accumulator forwarding. +/// vmul d3, d0, d2 +/// vmla d3, d1, d2 +/// is faster than +/// vadd d3, d0, d1 +/// vmul d3, d3, d2 +static SDValue PerformVMULCombine(SDNode *N, + TargetLowering::DAGCombinerInfo &DCI, + const ARMSubtarget *Subtarget) { + if (!Subtarget->hasVMLxForwarding()) + return SDValue(); + + SelectionDAG &DAG = DCI.DAG; + SDValue N0 = N->getOperand(0); + SDValue N1 = N->getOperand(1); + unsigned Opcode = N0.getOpcode(); + if (Opcode != ISD::ADD && Opcode != ISD::SUB && + Opcode != ISD::FADD && Opcode != ISD::FSUB) { + Opcode = N0.getOpcode(); + if (Opcode != ISD::ADD && Opcode != ISD::SUB && + Opcode != ISD::FADD && Opcode != ISD::FSUB) + return SDValue(); + std::swap(N0, N1); + } + + EVT VT = N->getValueType(0); + DebugLoc DL = N->getDebugLoc(); + SDValue N00 = N0->getOperand(0); + SDValue N01 = N0->getOperand(1); + return DAG.getNode(Opcode, DL, VT, + DAG.getNode(ISD::MUL, DL, VT, N00, N1), + DAG.getNode(ISD::MUL, DL, VT, N01, N1)); +} + static SDValue PerformMULCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI, const ARMSubtarget *Subtarget) { @@ -5046,6 +5572,8 @@ static SDValue PerformMULCombine(SDNode *N, return SDValue(); EVT VT = N->getValueType(0); + if (VT.is64BitVector() || VT.is128BitVector()) + return PerformVMULCombine(N, DCI, Subtarget); if (VT != MVT::i32) return SDValue(); @@ -5088,12 +5616,16 @@ static SDValue PerformMULCombine(SDNode *N, static SDValue PerformANDCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { + // Attempt to use immediate-form VBIC BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1)); DebugLoc dl = N->getDebugLoc(); EVT VT = N->getValueType(0); SelectionDAG &DAG = DCI.DAG; + if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) + return SDValue(); + APInt SplatBits, SplatUndef; unsigned SplatBitSize; bool HasAnyUndefs; @@ -5127,6 +5659,9 @@ static SDValue PerformORCombine(SDNode *N, EVT VT = N->getValueType(0); SelectionDAG &DAG = DCI.DAG; + if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) + return SDValue(); + APInt SplatBits, SplatUndef; unsigned SplatBitSize; bool HasAnyUndefs; @@ -5147,6 +5682,37 @@ static SDValue PerformORCombine(SDNode *N, } } + SDValue N0 = N->getOperand(0); + if (N0.getOpcode() != ISD::AND) + return SDValue(); + SDValue N1 = N->getOperand(1); + + // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant. + if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() && + DAG.getTargetLoweringInfo().isTypeLegal(VT)) { + APInt SplatUndef; + unsigned SplatBitSize; + bool HasAnyUndefs; + + BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1)); + APInt SplatBits0; + if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize, + HasAnyUndefs) && !HasAnyUndefs) { + BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1)); + APInt SplatBits1; + if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize, + HasAnyUndefs) && !HasAnyUndefs && + SplatBits0 == ~SplatBits1) { + // Canonicalize the vector type to make instruction selection simpler. + EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32; + SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT, + N0->getOperand(1), N0->getOperand(0), + N1->getOperand(0)); + return DAG.getNode(ISD::BITCAST, dl, VT, Result); + } + } + } + // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when // reasonable. @@ -5154,19 +5720,16 @@ static SDValue PerformORCombine(SDNode *N, if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops()) return SDValue(); - SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); DebugLoc DL = N->getDebugLoc(); // 1) or (and A, mask), val => ARMbfi A, val, mask // iff (val & mask) == val // // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2) - // && CountPopulation_32(mask) == CountPopulation_32(~mask2) + // && mask == ~mask2 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2) - // && CountPopulation_32(mask) == CountPopulation_32(~mask2) + // && ~mask == mask2 // (i.e., copy a bitfield value into another bitfield of the same width) - if (N0.getOpcode() != ISD::AND) - return SDValue(); if (VT != MVT::i32) return SDValue(); @@ -5209,26 +5772,26 @@ static SDValue PerformORCombine(SDNode *N, return SDValue(); unsigned Mask2 = N11C->getZExtValue(); + // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern + // as is to match. if (ARM::isBitFieldInvertedMask(Mask) && - ARM::isBitFieldInvertedMask(~Mask2) && - (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) { + (Mask == ~Mask2)) { // The pack halfword instruction works better for masks that fit it, // so use that when it's available. if (Subtarget->hasT2ExtractPack() && (Mask == 0xffff || Mask == 0xffff0000)) return SDValue(); // 2a - unsigned lsb = CountTrailingZeros_32(Mask2); + unsigned amt = CountTrailingZeros_32(Mask2); Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0), - DAG.getConstant(lsb, MVT::i32)); + DAG.getConstant(amt, MVT::i32)); Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res, DAG.getConstant(Mask, MVT::i32)); // Do not add new nodes to DAG combiner worklist. DCI.CombineTo(N, Res, false); return SDValue(); } else if (ARM::isBitFieldInvertedMask(~Mask) && - ARM::isBitFieldInvertedMask(Mask2) && - (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) { + (~Mask == Mask2)) { // The pack halfword instruction works better for masks that fit it, // so use that when it's available. if (Subtarget->hasT2ExtractPack() && @@ -5239,7 +5802,7 @@ static SDValue PerformORCombine(SDNode *N, Res = DAG.getNode(ISD::SRL, DL, VT, N00, DAG.getConstant(lsb, MVT::i32)); Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res, - DAG.getConstant(Mask2, MVT::i32)); + DAG.getConstant(Mask2, MVT::i32)); // Do not add new nodes to DAG combiner worklist. DCI.CombineTo(N, Res, false); return SDValue(); @@ -5294,6 +5857,37 @@ static SDValue PerformVMOVRRDCombine(SDNode *N, SDValue InDouble = N->getOperand(0); if (InDouble.getOpcode() == ARMISD::VMOVDRR) return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1)); + + // vmovrrd(load f64) -> (load i32), (load i32) + SDNode *InNode = InDouble.getNode(); + if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() && + InNode->getValueType(0) == MVT::f64 && + InNode->getOperand(1).getOpcode() == ISD::FrameIndex && + !cast<LoadSDNode>(InNode)->isVolatile()) { + // TODO: Should this be done for non-FrameIndex operands? + LoadSDNode *LD = cast<LoadSDNode>(InNode); + + SelectionDAG &DAG = DCI.DAG; + DebugLoc DL = LD->getDebugLoc(); + SDValue BasePtr = LD->getBasePtr(); + SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr, + LD->getPointerInfo(), LD->isVolatile(), + LD->isNonTemporal(), LD->getAlignment()); + + SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, + DAG.getConstant(4, MVT::i32)); + SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr, + LD->getPointerInfo(), LD->isVolatile(), + LD->isNonTemporal(), + std::min(4U, LD->getAlignment() / 2)); + + DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1)); + SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2); + DCI.RemoveFromWorklist(LD); + DAG.DeleteNode(LD); + return Result; + } + return SDValue(); } @@ -5323,8 +5917,28 @@ static SDValue PerformSTORECombine(SDNode *N, // Otherwise, the i64 value will be legalized to a pair of i32 values. StoreSDNode *St = cast<StoreSDNode>(N); SDValue StVal = St->getValue(); - if (!ISD::isNormalStore(St) || St->isVolatile() || - StVal.getValueType() != MVT::i64 || + if (!ISD::isNormalStore(St) || St->isVolatile()) + return SDValue(); + + if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR && + StVal.getNode()->hasOneUse() && !St->isVolatile()) { + SelectionDAG &DAG = DCI.DAG; + DebugLoc DL = St->getDebugLoc(); + SDValue BasePtr = St->getBasePtr(); + SDValue NewST1 = DAG.getStore(St->getChain(), DL, + StVal.getNode()->getOperand(0), BasePtr, + St->getPointerInfo(), St->isVolatile(), + St->isNonTemporal(), St->getAlignment()); + + SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, + DAG.getConstant(4, MVT::i32)); + return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1), + OffsetPtr, St->getPointerInfo(), St->isVolatile(), + St->isNonTemporal(), + std::min(4U, St->getAlignment() / 2)); + } + + if (StVal.getValueType() != MVT::i64 || StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT) return SDValue(); @@ -5553,7 +6167,7 @@ static SDValue CombineBaseUpdate(SDNode *N, EVT VecTy; if (isLoad) VecTy = N->getValueType(0); - else + else VecTy = N->getOperand(AddrOpIdx+1).getValueType(); unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8; if (isLaneOp) @@ -5603,7 +6217,7 @@ static SDValue CombineBaseUpdate(SDNode *N, DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs)); break; - } + } return SDValue(); } |
