//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "x86-isel"
#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
+#include "X86MCTargetExpr.h"
#include "X86TargetMachine.h"
+#include "X86TargetObjectFile.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Instructions.h"
#include "llvm/Intrinsics.h"
#include "llvm/LLVMContext.h"
-#include "llvm/ADT/BitVector.h"
-#include "llvm/ADT/VectorExtras.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
-#include "llvm/Support/MathExtras.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/Support/ErrorHandling.h"
-#include "llvm/Target/TargetLoweringObjectFile.h"
-#include "llvm/Target/TargetOptions.h"
+#include "llvm/MC/MCAsmInfo.h"
+#include "llvm/MC/MCContext.h"
+#include "llvm/MC/MCSymbol.h"
+#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringExtras.h"
+#include "llvm/ADT/VectorExtras.h"
#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
+STATISTIC(NumTailCalls, "Number of tail calls");
+
static cl::opt<bool>
DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
switch (TM.getSubtarget<X86Subtarget>().TargetType) {
default: llvm_unreachable("unknown subtarget type");
case X86Subtarget::isDarwin:
+ if (TM.getSubtarget<X86Subtarget>().is64Bit())
+ return new X8664_MachoTargetObjectFile();
return new TargetLoweringObjectFileMachO();
case X86Subtarget::isELF:
- return new TargetLoweringObjectFileELF();
+ if (TM.getSubtarget<X86Subtarget>().is64Bit())
+ return new X8664_ELFTargetObjectFile(TM);
+ return new X8632_ELFTargetObjectFile(TM);
case X86Subtarget::isMingw:
case X86Subtarget::isCygwin:
case X86Subtarget::isWindows:
return new TargetLoweringObjectFileCOFF();
}
-
}
X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
if (Subtarget->is64Bit())
setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
+ setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
if (Subtarget->is64Bit()) {
setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
+ setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
}
// 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
}
- // Use the default ISD::DBG_STOPPOINT.
- setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
// FIXME - use subtarget debug flags
if (!Subtarget->isTargetDarwin() &&
!Subtarget->isTargetELF() &&
!Subtarget->isTargetCygMing()) {
- setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
}
setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
+ setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
+ for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
+ InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
+ setTruncStoreAction((MVT::SimpleValueType)VT,
+ (MVT::SimpleValueType)InnerVT, Expand);
+ setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
+ setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
+ setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
}
// FIXME: In order to prevent SSE instructions being expanded to MMX ones
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
- setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
- setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
+
// Custom lower build_vector, vector_shuffle, and extract_vector_elt.
for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
EVT VT = (MVT::SimpleValueType)i;
setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
setTargetDAGCombine(ISD::BUILD_VECTOR);
setTargetDAGCombine(ISD::SELECT);
+ setTargetDAGCombine(ISD::AND);
setTargetDAGCombine(ISD::SHL);
setTargetDAGCombine(ISD::SRA);
setTargetDAGCombine(ISD::SRL);
+ setTargetDAGCombine(ISD::OR);
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::MEMBARRIER);
+ setTargetDAGCombine(ISD::ZERO_EXTEND);
if (Subtarget->is64Bit())
setTargetDAGCombine(ISD::MUL);
return MVT::i32;
}
+/// getJumpTableEncoding - Return the entry encoding for a jump table in the
+/// current function. The returned value is a member of the
+/// MachineJumpTableInfo::JTEntryKind enum.
+unsigned X86TargetLowering::getJumpTableEncoding() const {
+ // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
+ // symbol.
+ if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
+ Subtarget->isPICStyleGOT())
+ return MachineJumpTableInfo::EK_Custom32;
+
+ // Otherwise, use the normal jump table encoding heuristics.
+ return TargetLowering::getJumpTableEncoding();
+}
+
+/// getPICBaseSymbol - Return the X86-32 PIC base.
+MCSymbol *
+X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
+ MCContext &Ctx) const {
+ const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
+ return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
+ Twine(MF->getFunctionNumber())+"$pb");
+}
+
+
+const MCExpr *
+X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
+ const MachineBasicBlock *MBB,
+ unsigned uid,MCContext &Ctx) const{
+ assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
+ Subtarget->isPICStyleGOT());
+ // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
+ // entries.
+ return X86MCTargetExpr::Create(MBB->getSymbol(Ctx),
+ X86MCTargetExpr::GOTOFF, Ctx);
+}
+
/// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
/// jumptable.
SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
- SelectionDAG &DAG) const {
- if (usesGlobalOffsetTable())
- return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
+ SelectionDAG &DAG) const {
if (!Subtarget->is64Bit())
// This doesn't have DebugLoc associated with it, but is not really the
// same as a Register.
return Table;
}
+/// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
+/// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
+/// MCExpr.
+const MCExpr *X86TargetLowering::
+getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
+ MCContext &Ctx) const {
+ // X86-64 uses RIP relative addressing based on the jump table label.
+ if (Subtarget->isPICStyleRIPRel())
+ return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
+
+ // Otherwise, the reference is relative to the PIC base.
+ return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
+}
+
/// getFunctionAlignment - Return the Log2 alignment of this function.
unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
#include "X86GenCallingConv.inc"
+bool
+X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
+ const SmallVectorImpl<EVT> &OutTys,
+ const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
+ SelectionDAG &DAG) {
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
+ RVLocs, *DAG.getContext());
+ return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
+}
+
SDValue
X86TargetLowering::LowerReturn(SDValue Chain,
CallingConv::ID CallConv, bool isVarArg,
RVLocs, *DAG.getContext());
CCInfo.AnalyzeReturn(Outs, RetCC_X86);
- // If this is the first return lowered for this function, add the regs to the
- // liveout set for the function.
- if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
- for (unsigned i = 0; i != RVLocs.size(); ++i)
- if (RVLocs[i].isRegLoc())
- DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
- }
+ // Add the regs to the liveout set for the function.
+ MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
+ for (unsigned i = 0; i != RVLocs.size(); ++i)
+ if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
+ MRI.addLiveOut(RVLocs[i].getLocReg());
SDValue Flag;
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
unsigned Reg = FuncInfo->getSRetReturnReg();
if (!Reg) {
- Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
+ Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
FuncInfo->setSRetReturnReg(Reg);
}
SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
Flag = Chain.getValue(1);
+
+ // RAX now acts like a return value.
+ MRI.addLiveOut(X86::RAX);
}
RetOps[0] = Chain; // Update chain.
case CallingConv::X86_FastCall:
return !Subtarget->is64Bit();
case CallingConv::Fast:
- return PerformTailCallOpt;
+ return GuaranteedTailCallOpt;
}
}
return CC_X86_32_C;
}
-/// NameDecorationForCallConv - Selects the appropriate decoration to
-/// apply to a MachineFunction containing a given calling convention.
-NameDecorationStyle
-X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
- if (CallConv == CallingConv::X86_FastCall)
- return FastCall;
- else if (CallConv == CallingConv::X86_StdCall)
- return StdCall;
- return None;
-}
-
-
/// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
/// by "Src" to address "Dst" with size and alignment information specified by
/// the specific parameter attribute. The copy will be passed as a byval
/*AlwaysInline=*/true, NULL, 0, NULL, 0);
}
+/// FuncIsMadeTailCallSafe - Return true if the function is being made into
+/// a tailcall target by changing its ABI.
+static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
+ return GuaranteedTailCallOpt && CC == CallingConv::Fast;
+}
+
SDValue
X86TargetLowering::LowerMemArgument(SDValue Chain,
CallingConv::ID CallConv,
const CCValAssign &VA,
MachineFrameInfo *MFI,
unsigned i) {
-
// Create the nodes corresponding to a load from this parameter slot.
ISD::ArgFlagsTy Flags = Ins[i].Flags;
- bool AlwaysUseMutable = (CallConv==CallingConv::Fast) && PerformTailCallOpt;
+ bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
EVT ValVT;
// 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.
- int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
- VA.getLocMemOffset(), isImmutable);
- SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
- if (Flags.isByVal())
- return FIN;
- return DAG.getLoad(ValVT, dl, Chain, FIN,
- PseudoSourceValue::getFixedStack(FI), 0);
+ if (Flags.isByVal()) {
+ int FI = MFI->CreateFixedObject(Flags.getByValSize(),
+ VA.getLocMemOffset(), isImmutable, false);
+ return DAG.getFrameIndex(FI, getPointerTy());
+ } else {
+ int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
+ VA.getLocMemOffset(), isImmutable, false);
+ SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
+ return DAG.getLoad(ValVT, dl, Chain, FIN,
+ PseudoSourceValue::getFixedStack(FI), 0,
+ false, false, 0);
+ }
}
SDValue
Fn->getName() == "main")
FuncInfo->setForceFramePointer(true);
- // Decorate the function name.
- FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
-
MachineFrameInfo *MFI = MF.getFrameInfo();
bool Is64Bit = Subtarget->is64Bit();
bool IsWin64 = Subtarget->isTargetWin64();
// If value is passed via pointer - do a load.
if (VA.getLocInfo() == CCValAssign::Indirect)
- ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
+ ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
+ false, false, 0);
InVals.push_back(ArgValue);
}
}
unsigned StackSize = CCInfo.getNextStackOffset();
- // align stack specially for tail calls
- if (PerformTailCallOpt && CallConv == CallingConv::Fast)
+ // Align stack specially for tail calls.
+ if (FuncIsMadeTailCallSafe(CallConv))
StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
// If the function takes variable number of arguments, make a frame index for
// the start of the first vararg value... for expansion of llvm.va_start.
if (isVarArg) {
if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
- VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
+ VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
}
if (Is64Bit) {
unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
VarArgsGPOffset = NumIntRegs * 8;
VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
- TotalNumXMMRegs * 16, 16);
+ TotalNumXMMRegs * 16, 16,
+ false);
// Store the integer parameter registers.
SmallVector<SDValue, 8> MemOps;
SDValue Store =
DAG.getStore(Val.getValue(1), dl, Val, FIN,
PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
- Offset);
+ Offset, false, false, 0);
MemOps.push_back(Store);
Offset += 8;
}
// Some CCs need callee pop.
if (IsCalleePop(isVarArg, CallConv)) {
BytesToPopOnReturn = StackSize; // Callee pops everything.
- BytesCallerReserves = 0;
} else {
BytesToPopOnReturn = 0; // Callee pops nothing.
// If this is an sret function, the return should pop the hidden pointer.
if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
BytesToPopOnReturn = 4;
- BytesCallerReserves = StackSize;
}
if (!Is64Bit) {
return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
}
return DAG.getStore(Chain, dl, Arg, PtrOff,
- PseudoSourceValue::getStack(), LocMemOffset);
+ PseudoSourceValue::getStack(), LocMemOffset,
+ false, false, 0);
}
/// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
/// optimization is performed and it is required.
SDValue
X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
- SDValue &OutRetAddr,
- SDValue Chain,
- bool IsTailCall,
- bool Is64Bit,
- int FPDiff,
- DebugLoc dl) {
- if (!IsTailCall || FPDiff==0) return Chain;
-
+ SDValue &OutRetAddr, SDValue Chain,
+ bool IsTailCall, bool Is64Bit,
+ int FPDiff, DebugLoc dl) {
// Adjust the Return address stack slot.
EVT VT = getPointerTy();
OutRetAddr = getReturnAddressFrameIndex(DAG);
// Load the "old" Return address.
- OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
+ OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
return SDValue(OutRetAddr.getNode(), 1);
}
// Calculate the new stack slot for the return address.
int SlotSize = Is64Bit ? 8 : 4;
int NewReturnAddrFI =
- MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
+ MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
- PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
+ PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
+ false, false, 0);
return Chain;
}
SDValue
X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
CallingConv::ID CallConv, bool isVarArg,
- bool isTailCall,
+ bool &isTailCall,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<ISD::InputArg> &Ins,
DebugLoc dl, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) {
-
MachineFunction &MF = DAG.getMachineFunction();
bool Is64Bit = Subtarget->is64Bit();
bool IsStructRet = CallIsStructReturn(Outs);
+ bool IsSibcall = false;
+
+ if (isTailCall) {
+ // Check if it's really possible to do a tail call.
+ isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
+ Outs, Ins, DAG);
+
+ // Sibcalls are automatically detected tailcalls which do not require
+ // ABI changes.
+ if (!GuaranteedTailCallOpt && isTailCall)
+ IsSibcall = true;
+
+ if (isTailCall)
+ ++NumTailCalls;
+ }
- assert((!isTailCall ||
- (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
- "IsEligibleForTailCallOptimization missed a case!");
assert(!(isVarArg && CallConv == CallingConv::Fast) &&
"Var args not supported with calling convention fastcc");
// Get a count of how many bytes are to be pushed on the stack.
unsigned NumBytes = CCInfo.getNextStackOffset();
- if (PerformTailCallOpt && CallConv == CallingConv::Fast)
+ if (IsSibcall)
+ // This is a sibcall. The memory operands are available in caller's
+ // own caller's stack.
+ NumBytes = 0;
+ else if (GuaranteedTailCallOpt && CallConv == CallingConv::Fast)
NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
int FPDiff = 0;
- if (isTailCall) {
+ if (isTailCall && !IsSibcall) {
// Lower arguments at fp - stackoffset + fpdiff.
unsigned NumBytesCallerPushed =
MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
}
- Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
+ if (!IsSibcall)
+ Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
SDValue RetAddrFrIdx;
// Load return adress for tail calls.
- Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
- FPDiff, dl);
+ if (isTailCall && FPDiff)
+ Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
+ Is64Bit, FPDiff, dl);
SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
SmallVector<SDValue, 8> MemOpChains;
SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
- PseudoSourceValue::getFixedStack(FI), 0);
+ PseudoSourceValue::getFixedStack(FI), 0,
+ false, false, 0);
Arg = SpillSlot;
break;
}
if (VA.isRegLoc()) {
RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
- } else {
- if (!isTailCall || (isTailCall && isByVal)) {
- assert(VA.isMemLoc());
- if (StackPtr.getNode() == 0)
- StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
-
- MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
- dl, DAG, VA, Flags));
- }
+ } else if (!IsSibcall && (!isTailCall || isByVal)) {
+ assert(VA.isMemLoc());
+ if (StackPtr.getNode() == 0)
+ StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
+ MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
+ dl, DAG, VA, Flags));
}
}
InFlag = Chain.getValue(1);
}
-
if (Subtarget->isPICStyleGOT()) {
// ELF / PIC requires GOT in the EBX register before function calls via PLT
// GOT pointer.
int FI = 0;
// Do not flag preceeding copytoreg stuff together with the following stuff.
InFlag = SDValue();
- for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
- CCValAssign &VA = ArgLocs[i];
- if (!VA.isRegLoc()) {
+ if (GuaranteedTailCallOpt) {
+ for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
+ CCValAssign &VA = ArgLocs[i];
+ if (VA.isRegLoc())
+ continue;
assert(VA.isMemLoc());
SDValue Arg = Outs[i].Val;
ISD::ArgFlagsTy Flags = Outs[i].Flags;
// Create frame index.
int32_t Offset = VA.getLocMemOffset()+FPDiff;
uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
- FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
+ FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
FIN = DAG.getFrameIndex(FI, getPointerTy());
if (Flags.isByVal()) {
// Store relative to framepointer.
MemOpChains2.push_back(
DAG.getStore(ArgChain, dl, Arg, FIN,
- PseudoSourceValue::getFixedStack(FI), 0));
+ PseudoSourceValue::getFixedStack(FI), 0,
+ false, false, 0));
}
}
}
FPDiff, dl);
}
- // If the callee is a GlobalAddress node (quite common, every direct call is)
- // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
- if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
+ bool WasGlobalOrExternal = false;
+ if (getTargetMachine().getCodeModel() == CodeModel::Large) {
+ assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
+ // In the 64-bit large code model, we have to make all calls
+ // through a register, since the call instruction's 32-bit
+ // pc-relative offset may not be large enough to hold the whole
+ // address.
+ } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
+ WasGlobalOrExternal = true;
+ // If the callee is a GlobalAddress node (quite common, every direct call
+ // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
+ // it.
+
// We should use extra load for direct calls to dllimported functions in
// non-JIT mode.
GlobalValue *GV = G->getGlobal();
G->getOffset(), OpFlags);
}
} else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
+ WasGlobalOrExternal = true;
unsigned char OpFlags = 0;
// On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
OpFlags);
- } else if (isTailCall) {
- unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
+ }
+ if (isTailCall && !WasGlobalOrExternal) {
+ // Force the address into a (call preserved) caller-saved register since
+ // tailcall must happen after callee-saved registers are poped.
+ // FIXME: Give it a special register class that contains caller-saved
+ // register instead?
+ unsigned TCReg = Is64Bit ? X86::R11 : X86::EAX;
Chain = DAG.getCopyToReg(Chain, dl,
- DAG.getRegister(Opc, getPointerTy()),
+ DAG.getRegister(TCReg, getPointerTy()),
Callee,InFlag);
- Callee = DAG.getRegister(Opc, getPointerTy());
- // Add register as live out.
- MF.getRegInfo().addLiveOut(Opc);
+ Callee = DAG.getRegister(TCReg, getPointerTy());
}
// Returns a chain & a flag for retval copy to use.
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
SmallVector<SDValue, 8> Ops;
- if (isTailCall) {
+ if (!IsSibcall && isTailCall) {
Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
DAG.getIntPtrConstant(0, true), InFlag);
InFlag = Chain.getValue(1);
assert(((Callee.getOpcode() == ISD::Register &&
(cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
- cast<RegisterSDNode>(Callee)->getReg() == X86::R9)) ||
+ cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
Callee.getOpcode() == ISD::TargetExternalSymbol ||
Callee.getOpcode() == ISD::TargetGlobalAddress) &&
- "Expecting an global address, external symbol, or register");
+ "Expecting a global address, external symbol, or scratch register");
return DAG.getNode(X86ISD::TC_RETURN, dl,
NodeTys, &Ops[0], Ops.size());
if (IsCalleePop(isVarArg, CallConv))
NumBytesForCalleeToPush = NumBytes; // Callee pops everything
else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
- // If this is is a call to a struct-return function, the callee
+ // If this is a call to a struct-return function, the callee
// pops the hidden struct pointer, so we have to push it back.
// This is common for Darwin/X86, Linux & Mingw32 targets.
NumBytesForCalleeToPush = 4;
NumBytesForCalleeToPush = 0; // Callee pops nothing.
// Returns a flag for retval copy to use.
- Chain = DAG.getCALLSEQ_END(Chain,
- DAG.getIntPtrConstant(NumBytes, true),
- DAG.getIntPtrConstant(NumBytesForCalleeToPush,
- true),
- InFlag);
- InFlag = Chain.getValue(1);
+ if (!IsSibcall) {
+ Chain = DAG.getCALLSEQ_END(Chain,
+ DAG.getIntPtrConstant(NumBytes, true),
+ DAG.getIntPtrConstant(NumBytesForCalleeToPush,
+ true),
+ InFlag);
+ InFlag = Chain.getValue(1);
+ }
// Handle result values, copying them out of physregs into vregs that we
// return.
return Offset;
}
+/// MatchingStackOffset - Return true if the given stack call argument is
+/// already available in the same position (relatively) of the caller's
+/// incoming argument stack.
+static
+bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
+ MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
+ const X86InstrInfo *TII) {
+ int FI;
+ if (Arg.getOpcode() == ISD::CopyFromReg) {
+ unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
+ if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
+ return false;
+ MachineInstr *Def = MRI->getVRegDef(VR);
+ if (!Def)
+ return false;
+ if (!Flags.isByVal()) {
+ if (!TII->isLoadFromStackSlot(Def, FI))
+ return false;
+ } else {
+ unsigned Opcode = Def->getOpcode();
+ if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
+ Def->getOperand(1).isFI()) {
+ FI = Def->getOperand(1).getIndex();
+ if (MFI->getObjectSize(FI) != Flags.getByValSize())
+ return false;
+ } else
+ return false;
+ }
+ } else {
+ LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
+ if (!Ld)
+ return false;
+ SDValue Ptr = Ld->getBasePtr();
+ FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
+ if (!FINode)
+ return false;
+ FI = FINode->getIndex();
+ }
+
+ if (!MFI->isFixedObjectIndex(FI))
+ return false;
+ return Offset == MFI->getObjectOffset(FI);
+}
+
/// IsEligibleForTailCallOptimization - Check whether the call is eligible
/// for tail call optimization. Targets which want to do tail call
/// optimization should implement this function.
X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
CallingConv::ID CalleeCC,
bool isVarArg,
- const SmallVectorImpl<ISD::InputArg> &Ins,
+ const SmallVectorImpl<ISD::OutputArg> &Outs,
+ const SmallVectorImpl<ISD::InputArg> &Ins,
SelectionDAG& DAG) const {
- MachineFunction &MF = DAG.getMachineFunction();
- CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
- return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
+ if (CalleeCC != CallingConv::Fast &&
+ CalleeCC != CallingConv::C)
+ return false;
+
+ // If -tailcallopt is specified, make fastcc functions tail-callable.
+ const Function *CallerF = DAG.getMachineFunction().getFunction();
+ if (GuaranteedTailCallOpt) {
+ if (CalleeCC == CallingConv::Fast &&
+ CallerF->getCallingConv() == CalleeCC)
+ return true;
+ return false;
+ }
+
+ // Look for obvious safe cases to perform tail call optimization that does not
+ // requite ABI changes. This is what gcc calls sibcall.
+
+ // Do not tail call optimize vararg calls for now.
+ if (isVarArg)
+ return false;
+
+ // If the callee takes no arguments then go on to check the results of the
+ // call.
+ if (!Outs.empty()) {
+ // Check if stack adjustment is needed. For now, do not do this if any
+ // argument is passed on the stack.
+ SmallVector<CCValAssign, 16> ArgLocs;
+ CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
+ ArgLocs, *DAG.getContext());
+ CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
+ if (CCInfo.getNextStackOffset()) {
+ MachineFunction &MF = DAG.getMachineFunction();
+ if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
+ return false;
+ if (Subtarget->isTargetWin64())
+ // Win64 ABI has additional complications.
+ return false;
+
+ // Check if the arguments are already laid out in the right way as
+ // the caller's fixed stack objects.
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+ const MachineRegisterInfo *MRI = &MF.getRegInfo();
+ const X86InstrInfo *TII =
+ ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
+ for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
+ CCValAssign &VA = ArgLocs[i];
+ EVT RegVT = VA.getLocVT();
+ SDValue Arg = Outs[i].Val;
+ ISD::ArgFlagsTy Flags = Outs[i].Flags;
+ if (VA.getLocInfo() == CCValAssign::Indirect)
+ return false;
+ if (!VA.isRegLoc()) {
+ if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
+ MFI, MRI, TII))
+ return false;
+ }
+ }
+ }
+ }
+
+ return true;
}
FastISel *
-X86TargetLowering::createFastISel(MachineFunction &mf,
- MachineModuleInfo *mmo,
- DwarfWriter *dw,
- DenseMap<const Value *, unsigned> &vm,
- DenseMap<const BasicBlock *,
- MachineBasicBlock *> &bm,
- DenseMap<const AllocaInst *, int> &am
+X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
+ DwarfWriter *dw,
+ DenseMap<const Value *, unsigned> &vm,
+ DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
+ DenseMap<const AllocaInst *, int> &am
#ifndef NDEBUG
- , SmallSet<Instruction*, 8> &cil
+ , SmallSet<Instruction*, 8> &cil
#endif
) {
return X86::createFastISel(mf, mmo, dw, vm, bm, am
if (ReturnAddrIndex == 0) {
// Set up a frame object for the return address.
uint64_t SlotSize = TD->getPointerSize();
- ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
+ ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
+ false, false);
FuncInfo->setRAIndex(ReturnAddrIndex);
}
case ISD::SETNE: return X86::COND_NE;
case ISD::SETUO: return X86::COND_P;
case ISD::SETO: return X86::COND_NP;
+ case ISD::SETOEQ:
+ case ISD::SETUNE: return X86::COND_INVALID;
}
}
}
}
+/// isFPImmLegal - Returns true if the target can instruction select the
+/// specified FP immediate natively. If false, the legalizer will
+/// materialize the FP immediate as a load from a constant pool.
+bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
+ for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
+ if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
+ return true;
+ }
+ return false;
+}
+
/// isUndefOrInRange - Return true if Val is undef or if its value falls within
/// the specified range (L, H].
static bool isUndefOrInRange(int Val, int Low, int Hi) {
return ::isPSHUFLWMask(M, N->getValueType(0));
}
+/// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
+/// is suitable for input to PALIGNR.
+static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
+ bool hasSSSE3) {
+ int i, e = VT.getVectorNumElements();
+
+ // Do not handle v2i64 / v2f64 shuffles with palignr.
+ if (e < 4 || !hasSSSE3)
+ return false;
+
+ for (i = 0; i != e; ++i)
+ if (Mask[i] >= 0)
+ break;
+
+ // All undef, not a palignr.
+ if (i == e)
+ return false;
+
+ // Determine if it's ok to perform a palignr with only the LHS, since we
+ // don't have access to the actual shuffle elements to see if RHS is undef.
+ bool Unary = Mask[i] < (int)e;
+ bool NeedsUnary = false;
+
+ int s = Mask[i] - i;
+
+ // Check the rest of the elements to see if they are consecutive.
+ for (++i; i != e; ++i) {
+ int m = Mask[i];
+ if (m < 0)
+ continue;
+
+ Unary = Unary && (m < (int)e);
+ NeedsUnary = NeedsUnary || (m < s);
+
+ if (NeedsUnary && !Unary)
+ return false;
+ if (Unary && m != ((s+i) & (e-1)))
+ return false;
+ if (!Unary && m != (s+i))
+ return false;
+ }
+ return true;
+}
+
+bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isPALIGNRMask(M, N->getValueType(0), true);
+}
+
/// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to SHUFP*.
static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
isUndefOrEqual(N->getMaskElt(3), 3);
}
+/// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
+/// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
+/// <2, 3, 2, 3>
+bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
+ unsigned NumElems = N->getValueType(0).getVectorNumElements();
+
+ if (NumElems != 4)
+ return false;
+
+ return isUndefOrEqual(N->getMaskElt(0), 2) &&
+ isUndefOrEqual(N->getMaskElt(1), 3) &&
+ isUndefOrEqual(N->getMaskElt(2), 2) &&
+ isUndefOrEqual(N->getMaskElt(3), 3);
+}
+
/// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
return true;
}
-/// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
-/// and MOVLHPS.
-bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
+/// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
+/// specifies a shuffle of elements that is suitable for input to MOVLHPS.
+bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
unsigned NumElems = N->getValueType(0).getVectorNumElements();
if (NumElems != 2 && NumElems != 4)
return true;
}
-/// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
-/// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
-/// <2, 3, 2, 3>
-bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
- unsigned NumElems = N->getValueType(0).getVectorNumElements();
-
- if (NumElems != 4)
- return false;
-
- return isUndefOrEqual(N->getMaskElt(0), 2) &&
- isUndefOrEqual(N->getMaskElt(1), 3) &&
- isUndefOrEqual(N->getMaskElt(2), 2) &&
- isUndefOrEqual(N->getMaskElt(3), 3);
-}
-
/// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to UNPCKL.
static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
}
/// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
-/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
-/// instructions.
+/// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
int NumOperands = SVOp->getValueType(0).getVectorNumElements();
}
/// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
-/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
-/// instructions.
+/// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
unsigned Mask = 0;
}
/// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
-/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
-/// instructions.
+/// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
unsigned Mask = 0;
return Mask;
}
+/// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
+/// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
+unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
+ EVT VVT = N->getValueType(0);
+ unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
+ int Val = 0;
+
+ unsigned i, e;
+ for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
+ Val = SVOp->getMaskElt(i);
+ if (Val >= 0)
+ break;
+ }
+ return (Val - i) * EltSize;
+}
+
/// isZeroNode - Returns true if Elt is a constant zero or a floating point
/// constant +0.0.
bool X86::isZeroNode(SDValue Elt) {
DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
}
+SDValue
+X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
+ SelectionDAG &DAG) {
+
+ // Check if the scalar load can be widened into a vector load. And if
+ // the address is "base + cst" see if the cst can be "absorbed" into
+ // the shuffle mask.
+ if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
+ SDValue Ptr = LD->getBasePtr();
+ if (!ISD::isNormalLoad(LD) || LD->isVolatile())
+ return SDValue();
+ EVT PVT = LD->getValueType(0);
+ if (PVT != MVT::i32 && PVT != MVT::f32)
+ return SDValue();
+
+ int FI = -1;
+ int64_t Offset = 0;
+ if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
+ FI = FINode->getIndex();
+ Offset = 0;
+ } else if (Ptr.getOpcode() == ISD::ADD &&
+ isa<ConstantSDNode>(Ptr.getOperand(1)) &&
+ isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
+ FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
+ Offset = Ptr.getConstantOperandVal(1);
+ Ptr = Ptr.getOperand(0);
+ } else {
+ return SDValue();
+ }
+
+ SDValue Chain = LD->getChain();
+ // Make sure the stack object alignment is at least 16.
+ MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
+ if (DAG.InferPtrAlignment(Ptr) < 16) {
+ if (MFI->isFixedObjectIndex(FI)) {
+ // Can't change the alignment. FIXME: It's possible to compute
+ // the exact stack offset and reference FI + adjust offset instead.
+ // If someone *really* cares about this. That's the way to implement it.
+ return SDValue();
+ } else {
+ MFI->setObjectAlignment(FI, 16);
+ }
+ }
+
+ // (Offset % 16) must be multiple of 4. Then address is then
+ // Ptr + (Offset & ~15).
+ if (Offset < 0)
+ return SDValue();
+ if ((Offset % 16) & 3)
+ return SDValue();
+ int64_t StartOffset = Offset & ~15;
+ if (StartOffset)
+ Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
+ Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
+
+ int EltNo = (Offset - StartOffset) >> 2;
+ int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
+ EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
+ SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
+ false, false, 0);
+ // Canonicalize it to a v4i32 shuffle.
+ V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
+ DAG.getVectorShuffle(MVT::v4i32, dl, V1,
+ DAG.getUNDEF(MVT::v4i32), &Mask[0]));
+ }
+
+ return SDValue();
+}
+
SDValue
X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
DebugLoc dl = Op.getDebugLoc();
}
// Splat is obviously ok. Let legalizer expand it to a shuffle.
- if (Values.size() == 1)
+ if (Values.size() == 1) {
+ if (EVTBits == 32) {
+ // Instead of a shuffle like this:
+ // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
+ // Check if it's possible to issue this instead.
+ // shuffle (vload ptr)), undef, <1, 1, 1, 1>
+ unsigned Idx = CountTrailingZeros_32(NonZeros);
+ SDValue Item = Op.getOperand(Idx);
+ if (Op.getNode()->isOnlyUserOf(Item.getNode()))
+ return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
+ }
return SDValue();
+ }
// A vector full of immediates; various special cases are already
// handled, so this is best done with a single constant-pool load.
return SDValue();
}
+SDValue
+X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
+ // We support concatenate two MMX registers and place them in a MMX
+ // register. This is better than doing a stack convert.
+ DebugLoc dl = Op.getDebugLoc();
+ EVT ResVT = Op.getValueType();
+ assert(Op.getNumOperands() == 2);
+ assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
+ ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
+ int Mask[2];
+ SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
+ SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
+ InVec = Op.getOperand(1);
+ if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
+ unsigned NumElts = ResVT.getVectorNumElements();
+ VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
+ VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
+ InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
+ } else {
+ InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
+ SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
+ Mask[0] = 0; Mask[1] = 2;
+ VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
+ }
+ return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
+}
+
// v8i16 shuffles - Prefer shuffles in the following order:
// 1. [all] pshuflw, pshufhw, optional move
// 2. [ssse3] 1 x pshufb
unsigned ShAmt = 0;
SDValue ShVal;
bool isShift = getSubtarget()->hasSSE2() &&
- isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
+ isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
if (isShift && ShVal.hasOneUse()) {
// If the shifted value has multiple uses, it may be cheaper to use
// v_set0 + movlhps or movhlps, etc.
- EVT EVT = VT.getVectorElementType();
- ShAmt *= EVT.getSizeInBits();
+ EVT EltVT = VT.getVectorElementType();
+ ShAmt *= EltVT.getSizeInBits();
return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
}
if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
X86::isMOVSLDUPMask(SVOp) ||
X86::isMOVHLPSMask(SVOp) ||
- X86::isMOVHPMask(SVOp) ||
+ X86::isMOVLHPSMask(SVOp) ||
X86::isMOVLPMask(SVOp)))
return Op;
if (isShift) {
// No better options. Use a vshl / vsrl.
- EVT EVT = VT.getVectorElementType();
- ShAmt *= EVT.getSizeInBits();
+ EVT EltVT = VT.getVectorElementType();
+ ShAmt *= EltVT.getSizeInBits();
return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
}
MVT::v4i32, Vec),
Op.getOperand(1)));
// Transform it so it match pextrw which produces a 32-bit result.
- EVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy+1);
- SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
+ EVT EltVT = MVT::i32;
+ SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
Op.getOperand(0), Op.getOperand(1));
- SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
+ SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
DAG.getValueType(VT));
return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
} else if (VT.getSizeInBits() == 32) {
SDValue
X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
EVT VT = Op.getValueType();
- EVT EVT = VT.getVectorElementType();
+ EVT EltVT = VT.getVectorElementType();
DebugLoc dl = Op.getDebugLoc();
SDValue N0 = Op.getOperand(0);
SDValue N1 = Op.getOperand(1);
SDValue N2 = Op.getOperand(2);
- if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
+ if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
isa<ConstantSDNode>(N2)) {
- unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
- : X86ISD::PINSRW;
+ unsigned Opc;
+ if (VT == MVT::v8i16)
+ Opc = X86ISD::PINSRW;
+ else if (VT == MVT::v4i16)
+ Opc = X86ISD::MMX_PINSRW;
+ else if (VT == MVT::v16i8)
+ Opc = X86ISD::PINSRB;
+ else
+ Opc = X86ISD::PINSRB;
+
// Transform it so it match pinsr{b,w} which expects a GR32 as its second
// argument.
if (N1.getValueType() != MVT::i32)
if (N2.getValueType() != MVT::i32)
N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
return DAG.getNode(Opc, dl, VT, N0, N1, N2);
- } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
+ } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
// Bits [7:6] of the constant are the source select. This will always be
// zero here. The DAG Combiner may combine an extract_elt index into these
// bits. For example (insert (extract, 3), 2) could be matched by putting
// Create this as a scalar to vector..
N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
- } else if (EVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
+ } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
// PINSR* works with constant index.
return Op;
}
SDValue
X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
EVT VT = Op.getValueType();
- EVT EVT = VT.getVectorElementType();
+ EVT EltVT = VT.getVectorElementType();
if (Subtarget->hasSSE41())
return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
- if (EVT == MVT::i8)
+ if (EltVT == MVT::i8)
return SDValue();
DebugLoc dl = Op.getDebugLoc();
SDValue N1 = Op.getOperand(1);
SDValue N2 = Op.getOperand(2);
- if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
+ if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
// Transform it so it match pinsrw which expects a 16-bit value in a GR32
// as its second argument.
if (N1.getValueType() != MVT::i32)
N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
if (N2.getValueType() != MVT::i32)
N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
- return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
+ return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
+ dl, VT, N0, N1, N2);
}
return SDValue();
}
return Result;
}
+SDValue
+X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
+ // Create the TargetBlockAddressAddress node.
+ unsigned char OpFlags =
+ Subtarget->ClassifyBlockAddressReference();
+ CodeModel::Model M = getTargetMachine().getCodeModel();
+ BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
+ /*isTarget=*/true, OpFlags);
+
+ if (Subtarget->isPICStyleRIPRel() &&
+ (M == CodeModel::Small || M == CodeModel::Kernel))
+ Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
+ else
+ Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
+
+ // With PIC, the address is actually $g + Offset.
+ if (isGlobalRelativeToPICBase(OpFlags)) {
+ Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
+ DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
+ Result);
+ }
+
+ return Result;
+}
+
SDValue
X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
int64_t Offset,
// load.
if (isGlobalStubReference(OpFlags))
Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
- PseudoSourceValue::getGOT(), 0);
+ PseudoSourceValue::getGOT(), 0, false, false, 0);
// If there was a non-zero offset that we didn't fold, create an explicit
// addition for it.
GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
unsigned char OperandFlags) {
+ MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
DebugLoc dl = GA->getDebugLoc();
SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
SDValue Ops[] = { Chain, TGA };
Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
}
+
+ // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
+ MFI->setHasCalls(true);
+
SDValue Flag = Chain.getValue(1);
return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
}
MVT::i32));
SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
- NULL, 0);
+ NULL, 0, false, false, 0);
unsigned char OperandFlags = 0;
// Most TLS accesses are not RIP relative, even on x86-64. One exception is
if (model == TLSModel::InitialExec)
Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
- PseudoSourceValue::getGOT(), 0);
+ PseudoSourceValue::getGOT(), 0, false, false, 0);
// The address of the thread local variable is the add of the thread
// pointer with the offset of the variable.
SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
DAG.getConstant(VTBits, MVT::i8));
- SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
+ SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
AndNode, DAG.getConstant(0, MVT::i8));
SDValue Hi, Lo;
DebugLoc dl = Op.getDebugLoc();
unsigned Size = SrcVT.getSizeInBits()/8;
MachineFunction &MF = DAG.getMachineFunction();
- int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
+ int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
StackSlot,
- PseudoSourceValue::getFixedStack(SSFI), 0);
+ PseudoSourceValue::getFixedStack(SSFI), 0,
+ false, false, 0);
return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
}
Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
else
Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
- SmallVector<SDValue, 8> Ops;
- Ops.push_back(Chain);
- Ops.push_back(StackSlot);
- Ops.push_back(DAG.getValueType(SrcVT));
+ SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
- Tys, &Ops[0], Ops.size());
+ Tys, Ops, array_lengthof(Ops));
if (useSSE) {
Chain = Result.getValue(1);
// shouldn't be necessary except that RFP cannot be live across
// multiple blocks. When stackifier is fixed, they can be uncoupled.
MachineFunction &MF = DAG.getMachineFunction();
- int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
+ int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
Tys = DAG.getVTList(MVT::Other);
- SmallVector<SDValue, 8> Ops;
- Ops.push_back(Chain);
- Ops.push_back(Result);
- Ops.push_back(StackSlot);
- Ops.push_back(DAG.getValueType(Op.getValueType()));
- Ops.push_back(InFlag);
- Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
+ SDValue Ops[] = {
+ Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
+ };
+ Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
- PseudoSourceValue::getFixedStack(SSFI), 0);
+ PseudoSourceValue::getFixedStack(SSFI), 0,
+ false, false, 0);
}
return Result;
SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
PseudoSourceValue::getConstantPool(), 0,
- false, 16);
+ false, false, 16);
SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
PseudoSourceValue::getConstantPool(), 0,
- false, 16);
+ false, false, 16);
SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
// Add the halves; easiest way is to swap them into another reg first.
SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
getPointerTy(), StackSlot, WordOff);
SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
- StackSlot, NULL, 0);
+ StackSlot, NULL, 0, false, false, 0);
SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
- OffsetSlot, NULL, 0);
+ OffsetSlot, NULL, 0, false, false, 0);
return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
}
// stack slot.
MachineFunction &MF = DAG.getMachineFunction();
unsigned MemSize = DstTy.getSizeInBits()/8;
- int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
+ int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
unsigned Opc;
if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
Chain = DAG.getStore(Chain, dl, Value, StackSlot,
- PseudoSourceValue::getFixedStack(SSFI), 0);
+ PseudoSourceValue::getFixedStack(SSFI), 0,
+ false, false, 0);
SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
SDValue Ops[] = {
Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
};
Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
Chain = Value.getValue(1);
- SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
+ SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
}
// Load the result.
return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
- FIST, StackSlot, NULL, 0);
+ FIST, StackSlot, NULL, 0, false, false, 0);
}
SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
// Load the result.
return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
- FIST, StackSlot, NULL, 0);
+ FIST, StackSlot, NULL, 0, false, false, 0);
}
SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
Constant *C = ConstantVector::get(CV);
SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
- PseudoSourceValue::getConstantPool(), 0,
- false, 16);
+ PseudoSourceValue::getConstantPool(), 0,
+ false, false, 16);
return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
}
DebugLoc dl = Op.getDebugLoc();
EVT VT = Op.getValueType();
EVT EltVT = VT;
- unsigned EltNum = 1;
- if (VT.isVector()) {
+ if (VT.isVector())
EltVT = VT.getVectorElementType();
- EltNum = VT.getVectorNumElements();
- }
std::vector<Constant*> CV;
if (EltVT == MVT::f64) {
Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
Constant *C = ConstantVector::get(CV);
SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
- PseudoSourceValue::getConstantPool(), 0,
- false, 16);
+ PseudoSourceValue::getConstantPool(), 0,
+ false, false, 16);
if (VT.isVector()) {
return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
DAG.getNode(ISD::XOR, dl, MVT::v2i64,
Constant *C = ConstantVector::get(CV);
SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
- PseudoSourceValue::getConstantPool(), 0,
- false, 16);
+ PseudoSourceValue::getConstantPool(), 0,
+ false, false, 16);
SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
// Shift sign bit right or left if the two operands have different types.
C = ConstantVector::get(CV);
CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
- PseudoSourceValue::getConstantPool(), 0,
- false, 16);
+ PseudoSourceValue::getConstantPool(), 0,
+ false, false, 16);
SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
// Or the value with the sign bit.
Opcode = X86ISD::ADD;
NumOperands = 2;
break;
+ case ISD::AND: {
+ // If the primary and result isn't used, don't bother using X86ISD::AND,
+ // because a TEST instruction will be better.
+ bool NonFlagUse = false;
+ for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
+ UE = Op.getNode()->use_end(); UI != UE; ++UI) {
+ SDNode *User = *UI;
+ unsigned UOpNo = UI.getOperandNo();
+ if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
+ // Look pass truncate.
+ UOpNo = User->use_begin().getOperandNo();
+ User = *User->use_begin();
+ }
+ if (User->getOpcode() != ISD::BRCOND &&
+ User->getOpcode() != ISD::SETCC &&
+ (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
+ NonFlagUse = true;
+ break;
+ }
+ }
+ if (!NonFlagUse)
+ break;
+ }
+ // FALL THROUGH
case ISD::SUB:
- // Due to the ISEL shortcoming noted above, be conservative if this sub is
+ case ISD::OR:
+ case ISD::XOR:
+ // Due to the ISEL shortcoming noted above, be conservative if this op is
// likely to be selected as part of a load-modify-store instruction.
for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
UE = Op.getNode()->use_end(); UI != UE; ++UI)
if (UI->getOpcode() == ISD::STORE)
goto default_case;
- // Otherwise use a regular EFLAGS-setting sub.
- Opcode = X86ISD::SUB;
+ // Otherwise use a regular EFLAGS-setting instruction.
+ switch (Op.getNode()->getOpcode()) {
+ case ISD::SUB: Opcode = X86ISD::SUB; break;
+ case ISD::OR: Opcode = X86ISD::OR; break;
+ case ISD::XOR: Opcode = X86ISD::XOR; break;
+ case ISD::AND: Opcode = X86ISD::AND; break;
+ default: llvm_unreachable("unexpected operator!");
+ }
NumOperands = 2;
break;
case X86ISD::ADD:
case X86ISD::SUB:
case X86ISD::INC:
case X86ISD::DEC:
+ case X86ISD::OR:
+ case X86ISD::XOR:
+ case X86ISD::AND:
return SDValue(Op.getNode(), 1);
default:
default_case:
return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
}
+/// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
+/// if it's possible.
+static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
+ DebugLoc dl, SelectionDAG &DAG) {
+ SDValue LHS, RHS;
+ if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
+ if (ConstantSDNode *Op010C =
+ dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
+ if (Op010C->getZExtValue() == 1) {
+ LHS = Op0.getOperand(0);
+ RHS = Op0.getOperand(1).getOperand(1);
+ }
+ } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
+ if (ConstantSDNode *Op000C =
+ dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
+ if (Op000C->getZExtValue() == 1) {
+ LHS = Op0.getOperand(1);
+ RHS = Op0.getOperand(0).getOperand(1);
+ }
+ } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
+ ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
+ SDValue AndLHS = Op0.getOperand(0);
+ if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
+ LHS = AndLHS.getOperand(0);
+ RHS = AndLHS.getOperand(1);
+ }
+ }
+
+ if (LHS.getNode()) {
+ // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
+ // instruction. Since the shift amount is in-range-or-undefined, we know
+ // that doing a bittest on the i16 value is ok. We extend to i32 because
+ // the encoding for the i16 version is larger than the i32 version.
+ if (LHS.getValueType() == MVT::i8)
+ LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
+
+ // If the operand types disagree, extend the shift amount to match. Since
+ // BT ignores high bits (like shifts) we can use anyextend.
+ if (LHS.getValueType() != RHS.getValueType())
+ RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
+
+ SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
+ unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
+ return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(Cond, MVT::i8), BT);
+ }
+
+ return SDValue();
+}
+
SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
SDValue Op0 = Op.getOperand(0);
DebugLoc dl = Op.getDebugLoc();
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
+ // Optimize to BT if possible.
// Lower (X & (1 << N)) == 0 to BT(X, N).
// Lower ((X >>u N) & 1) != 0 to BT(X, N).
// Lower ((X >>s N) & 1) != 0 to BT(X, N).
Op1.getOpcode() == ISD::Constant &&
cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
- SDValue LHS, RHS;
- if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
- if (ConstantSDNode *Op010C =
- dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
- if (Op010C->getZExtValue() == 1) {
- LHS = Op0.getOperand(0);
- RHS = Op0.getOperand(1).getOperand(1);
- }
- } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
- if (ConstantSDNode *Op000C =
- dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
- if (Op000C->getZExtValue() == 1) {
- LHS = Op0.getOperand(1);
- RHS = Op0.getOperand(0).getOperand(1);
- }
- } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
- ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
- SDValue AndLHS = Op0.getOperand(0);
- if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
- LHS = AndLHS.getOperand(0);
- RHS = AndLHS.getOperand(1);
- }
- }
-
- if (LHS.getNode()) {
- // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
- // instruction. Since the shift amount is in-range-or-undefined, we know
- // that doing a bittest on the i16 value is ok. We extend to i32 because
- // the encoding for the i16 version is larger than the i32 version.
- if (LHS.getValueType() == MVT::i8)
- LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
-
- // If the operand types disagree, extend the shift amount to match. Since
- // BT ignores high bits (like shifts) we can use anyextend.
- if (LHS.getValueType() != RHS.getValueType())
- RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
-
- SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
- unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
- return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(Cond, MVT::i8), BT);
- }
+ SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
+ if (NewSetCC.getNode())
+ return NewSetCC;
}
bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
+ if (X86CC == X86::COND_INVALID)
+ return SDValue();
SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
+
+ // Use sbb x, x to materialize carry bit into a GPR.
+ if (X86CC == X86::COND_B)
+ return DAG.getNode(ISD::AND, dl, MVT::i8,
+ DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), Cond),
+ DAG.getConstant(1, MVT::i8));
+
return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
DAG.getConstant(X86CC, MVT::i8), Cond);
}
Opc == X86ISD::SMUL ||
Opc == X86ISD::UMUL ||
Opc == X86ISD::INC ||
- Opc == X86ISD::DEC))
+ Opc == X86ISD::DEC ||
+ Opc == X86ISD::OR ||
+ Opc == X86ISD::XOR ||
+ Opc == X86ISD::AND))
return true;
return false;
DebugLoc dl = Op.getDebugLoc();
SDValue CC;
- if (Cond.getOpcode() == ISD::SETCC)
- Cond = LowerSETCC(Cond, DAG);
+ if (Cond.getOpcode() == ISD::SETCC) {
+ SDValue NewCond = LowerSETCC(Cond, DAG);
+ if (NewCond.getNode())
+ Cond = NewCond;
+ }
+
+ // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
+ SDValue Op1 = Op.getOperand(1);
+ SDValue Op2 = Op.getOperand(2);
+ if (Cond.getOpcode() == X86ISD::SETCC &&
+ cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
+ SDValue Cmp = Cond.getOperand(1);
+ if (Cmp.getOpcode() == X86ISD::CMP) {
+ ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
+ ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
+ ConstantSDNode *RHSC =
+ dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
+ if (N1C && N1C->isAllOnesValue() &&
+ N2C && N2C->isNullValue() &&
+ RHSC && RHSC->isNullValue()) {
+ SDValue CmpOp0 = Cmp.getOperand(0);
+ Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
+ CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
+ return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
+ DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
+ }
+ }
+ }
+
+ // Look pass (and (setcc_carry (cmp ...)), 1).
+ if (Cond.getOpcode() == ISD::AND &&
+ Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
+ if (C && C->getAPIntValue() == 1)
+ Cond = Cond.getOperand(0);
+ }
// If condition flag is set by a X86ISD::CMP, then use it as the condition
// setting operand in place of the X86ISD::SETCC.
- if (Cond.getOpcode() == X86ISD::SETCC) {
+ if (Cond.getOpcode() == X86ISD::SETCC ||
+ Cond.getOpcode() == X86ISD::SETCC_CARRY) {
CC = Cond.getOperand(0);
SDValue Cmp = Cond.getOperand(1);
}
}
+ if (addTest) {
+ // Look pass the truncate.
+ if (Cond.getOpcode() == ISD::TRUNCATE)
+ Cond = Cond.getOperand(0);
+
+ // We know the result of AND is compared against zero. Try to match
+ // it to BT.
+ if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
+ SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
+ if (NewSetCC.getNode()) {
+ CC = NewSetCC.getOperand(0);
+ Cond = NewSetCC.getOperand(1);
+ addTest = false;
+ }
+ }
+ }
+
if (addTest) {
CC = DAG.getConstant(X86::COND_NE, MVT::i8);
Cond = EmitTest(Cond, X86::COND_NE, DAG);
}
- SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
- SmallVector<SDValue, 4> Ops;
// X86ISD::CMOV means set the result (which is operand 1) to the RHS if
// condition is true.
- Ops.push_back(Op.getOperand(2));
- Ops.push_back(Op.getOperand(1));
- Ops.push_back(CC);
- Ops.push_back(Cond);
- return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
+ SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
+ SDValue Ops[] = { Op2, Op1, CC, Cond };
+ return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
}
// isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
DebugLoc dl = Op.getDebugLoc();
SDValue CC;
- if (Cond.getOpcode() == ISD::SETCC)
- Cond = LowerSETCC(Cond, DAG);
+ if (Cond.getOpcode() == ISD::SETCC) {
+ SDValue NewCond = LowerSETCC(Cond, DAG);
+ if (NewCond.getNode())
+ Cond = NewCond;
+ }
#if 0
// FIXME: LowerXALUO doesn't handle these!!
else if (Cond.getOpcode() == X86ISD::ADD ||
Cond = LowerXALUO(Cond, DAG);
#endif
+ // Look pass (and (setcc_carry (cmp ...)), 1).
+ if (Cond.getOpcode() == ISD::AND &&
+ Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
+ if (C && C->getAPIntValue() == 1)
+ Cond = Cond.getOperand(0);
+ }
+
// If condition flag is set by a X86ISD::CMP, then use it as the condition
// setting operand in place of the X86ISD::SETCC.
- if (Cond.getOpcode() == X86ISD::SETCC) {
+ if (Cond.getOpcode() == X86ISD::SETCC ||
+ Cond.getOpcode() == X86ISD::SETCC_CARRY) {
CC = Cond.getOperand(0);
SDValue Cmp = Cond.getOperand(1);
}
}
+ if (addTest) {
+ // Look pass the truncate.
+ if (Cond.getOpcode() == ISD::TRUNCATE)
+ Cond = Cond.getOperand(0);
+
+ // We know the result of AND is compared against zero. Try to match
+ // it to BT.
+ if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
+ SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
+ if (NewSetCC.getNode()) {
+ CC = NewSetCC.getOperand(0);
+ Cond = NewSetCC.getOperand(1);
+ addTest = false;
+ }
+ }
+ }
+
if (addTest) {
CC = DAG.getConstant(X86::COND_NE, MVT::i8);
Cond = EmitTest(Cond, X86::COND_NE, DAG);
LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
false, false, false, false,
0, CallingConv::C, false, /*isReturnValueUsed=*/false,
- DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
+ DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
+ DAG.GetOrdering(Chain.getNode()));
return CallResult.second;
}
InFlag = Chain.getValue(1);
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
- SmallVector<SDValue, 8> Ops;
- Ops.push_back(Chain);
- Ops.push_back(DAG.getValueType(AVT));
- Ops.push_back(InFlag);
- Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
+ SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
+ Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
if (TwoRepStos) {
InFlag = Chain.getValue(1);
Left, InFlag);
InFlag = Chain.getValue(1);
Tys = DAG.getVTList(MVT::Other, MVT::Flag);
- Ops.clear();
- Ops.push_back(Chain);
- Ops.push_back(DAG.getValueType(MVT::i8));
- Ops.push_back(InFlag);
- Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
+ SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
+ Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
} else if (BytesLeft) {
// Handle the last 1 - 7 bytes.
unsigned Offset = SizeVal - BytesLeft;
InFlag = Chain.getValue(1);
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
- SmallVector<SDValue, 8> Ops;
- Ops.push_back(Chain);
- Ops.push_back(DAG.getValueType(AVT));
- Ops.push_back(InFlag);
- SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
+ SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
+ SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
+ array_lengthof(Ops));
SmallVector<SDValue, 4> Results;
Results.push_back(RepMovs);
// vastart just stores the address of the VarArgsFrameIndex slot into the
// memory location argument.
SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
- return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
+ return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
+ false, false, 0);
}
// __va_list_tag:
SDValue FIN = Op.getOperand(1);
// Store gp_offset
SDValue Store = DAG.getStore(Op.getOperand(0), dl,
- DAG.getConstant(VarArgsGPOffset, MVT::i32),
- FIN, SV, 0);
+ DAG.getConstant(VarArgsGPOffset, MVT::i32),
+ FIN, SV, 0, false, false, 0);
MemOps.push_back(Store);
// Store fp_offset
FIN, DAG.getIntPtrConstant(4));
Store = DAG.getStore(Op.getOperand(0), dl,
DAG.getConstant(VarArgsFPOffset, MVT::i32),
- FIN, SV, 0);
+ FIN, SV, 0, false, false, 0);
MemOps.push_back(Store);
// Store ptr to overflow_arg_area
FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
FIN, DAG.getIntPtrConstant(4));
SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
- Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
+ Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
+ false, false, 0);
MemOps.push_back(Store);
// Store ptr to reg_save_area.
FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
FIN, DAG.getIntPtrConstant(8));
SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
- Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
+ Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
+ false, false, 0);
MemOps.push_back(Store);
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&MemOps[0], MemOps.size());
SDValue LHS = Op.getOperand(1);
SDValue RHS = Op.getOperand(2);
unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
+ assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
DAG.getConstant(X86CC, MVT::i8), Cond);
return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
DAG.getNode(ISD::ADD, dl, getPointerTy(),
FrameAddr, Offset),
- NULL, 0);
+ NULL, 0, false, false, 0);
}
// Just load the return address.
SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
- RetAddrFI, NULL, 0);
+ RetAddrFI, NULL, 0, false, false, 0);
}
SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
while (Depth--)
- FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
+ FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
+ false, false, 0);
return FrameAddr;
}
SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
DAG.getIntPtrConstant(-TD->getPointerSize()));
StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
- Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
+ Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
MF.getRegInfo().addLiveOut(StoreAddrReg);
const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
- const X86InstrInfo *TII =
- ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
-
if (Subtarget->is64Bit()) {
SDValue OutChains[6];
// Large code-model.
-
- const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
- const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
+ const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
+ const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
SDValue Addr = Trmp;
OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
- Addr, TrmpAddr, 0);
+ Addr, TrmpAddr, 0, false, false, 0);
Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(2, MVT::i64));
- OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
+ OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
+ false, false, 2);
// Load the 'nest' parameter value into R10.
// R10 is specified in X86CallingConv.td
Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(10, MVT::i64));
OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
- Addr, TrmpAddr, 10);
+ Addr, TrmpAddr, 10, false, false, 0);
Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(12, MVT::i64));
- OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
+ OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
+ false, false, 2);
// Jump to the nested function.
OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(20, MVT::i64));
OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
- Addr, TrmpAddr, 20);
+ Addr, TrmpAddr, 20, false, false, 0);
unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(22, MVT::i64));
OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
- TrmpAddr, 22);
+ TrmpAddr, 22, false, false, 0);
SDValue Ops[] =
{ Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
DAG.getConstant(10, MVT::i32));
Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
- const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
+ // This is storing the opcode for MOV32ri.
+ const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
OutChains[0] = DAG.getStore(Root, dl,
DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
- Trmp, TrmpAddr, 0);
+ Trmp, TrmpAddr, 0, false, false, 0);
Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
DAG.getConstant(1, MVT::i32));
- OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
+ OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
+ false, false, 1);
- const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
+ const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
DAG.getConstant(5, MVT::i32));
OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
- TrmpAddr, 5, false, 1);
+ TrmpAddr, 5, false, false, 1);
Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
DAG.getConstant(6, MVT::i32));
- OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
+ OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
+ false, false, 1);
SDValue Ops[] =
{ Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
DebugLoc dl = Op.getDebugLoc();
// Save FP Control Word to stack slot
- int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
+ int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
DAG.getEntryNode(), StackSlot);
// Load FP Control Word from stack slot
- SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
+ SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
+ false, false, 0);
// Transform as necessary
SDValue CWD1 =
Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
// If src is zero (i.e. bsr sets ZF), returns NumBits.
- SmallVector<SDValue, 4> Ops;
- Ops.push_back(Op);
- Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
- Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
- Ops.push_back(Op.getValue(1));
- Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
+ SDValue Ops[] = {
+ Op,
+ DAG.getConstant(NumBits+NumBits-1, OpVT),
+ DAG.getConstant(X86::COND_E, MVT::i8),
+ Op.getValue(1)
+ };
+ Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
// Finally xor with NumBits-1.
Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
// If src is zero (i.e. bsf sets ZF), returns NumBits.
- SmallVector<SDValue, 4> Ops;
- Ops.push_back(Op);
- Ops.push_back(DAG.getConstant(NumBits, OpVT));
- Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
- Ops.push_back(Op.getValue(1));
- Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
+ SDValue Ops[] = {
+ Op,
+ DAG.getConstant(NumBits, OpVT),
+ DAG.getConstant(X86::COND_E, MVT::i8),
+ Op.getValue(1)
+ };
+ Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
if (VT == MVT::i8)
Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
+ case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
+ case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
case ISD::SHL_PARTS:
case ISD::SRA_PARTS:
case ISD::SRL_PARTS: return LowerShift(Op, DAG);
Node->getOperand(2), DAG.getIntPtrConstant(0));
SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
Node->getOperand(2), DAG.getIntPtrConstant(1));
- // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
- // have a MemOperand. Pass the info through as a normal operand.
- SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
- SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
+ SDValue Ops[] = { Chain, In1, In2L, In2H };
SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
- SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
+ SDValue Result =
+ DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
+ cast<MemSDNode>(Node)->getMemOperand());
SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
Results.push_back(Result.getValue(2));
if (FIST.getNode() != 0) {
EVT VT = N->getValueType(0);
// Return a load from the stack slot.
- Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
+ Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
+ false, false, 0));
}
return;
}
case X86ISD::COMI: return "X86ISD::COMI";
case X86ISD::UCOMI: return "X86ISD::UCOMI";
case X86ISD::SETCC: return "X86ISD::SETCC";
+ case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
case X86ISD::CMOV: return "X86ISD::CMOV";
case X86ISD::BRCOND: return "X86ISD::BRCOND";
case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
case X86ISD::PINSRB: return "X86ISD::PINSRB";
case X86ISD::PINSRW: return "X86ISD::PINSRW";
+ case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
case X86ISD::FMAX: return "X86ISD::FMAX";
case X86ISD::FMIN: return "X86ISD::FMIN";
case X86ISD::UMUL: return "X86ISD::UMUL";
case X86ISD::INC: return "X86ISD::INC";
case X86ISD::DEC: return "X86ISD::DEC";
+ case X86ISD::OR: return "X86ISD::OR";
+ case X86ISD::XOR: return "X86ISD::XOR";
+ case X86ISD::AND: return "X86ISD::AND";
case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
case X86ISD::PTEST: return "X86ISD::PTEST";
case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
- if (!Ty1->isInteger() || !Ty2->isInteger())
+ if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
return false;
unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
// x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
- return Ty1 == Type::getInt32Ty(Ty1->getContext()) &&
- Ty2 == Type::getInt64Ty(Ty1->getContext()) && Subtarget->is64Bit();
+ return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
}
bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
if (VT.getSizeInBits() == 64)
return false;
- // FIXME: pshufb, blends, palignr, shifts.
+ // FIXME: pshufb, blends, shifts.
return (VT.getVectorNumElements() == 2 ||
ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
isMOVLMask(M, VT) ||
isPSHUFDMask(M, VT) ||
isPSHUFHWMask(M, VT) ||
isPSHUFLWMask(M, VT) ||
+ isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
isUNPCKLMask(M, VT) ||
isUNPCKHMask(M, VT) ||
isUNPCKL_v_undef_Mask(M, VT) ||
(*MIB).addOperand(*argOpers[i]);
MIB.addReg(t2);
assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
- (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
+ (*MIB).setMemRefs(bInstr->memoperands_begin(),
+ bInstr->memoperands_end());
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
MIB.addReg(EAXreg);
// insert branch
- BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
+ BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
return nextMBB;
for (int i=0; i < 2 + X86AddrNumOperands; ++i)
argOpers[i] = &bInstr->getOperand(i+2);
- // x86 address has 4 operands: base, index, scale, and displacement
+ // x86 address has 5 operands: base, index, scale, displacement, and segment.
int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
.addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
- unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
- unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
+ // The subsequent operations should be using the destination registers of
+ //the PHI instructions.
if (invSrc) {
- MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
- MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
+ t1 = F->getRegInfo().createVirtualRegister(RC);
+ t2 = F->getRegInfo().createVirtualRegister(RC);
+ MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
+ MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
} else {
- tt1 = t1;
- tt2 = t2;
+ t1 = dest1Oper.getReg();
+ t2 = dest2Oper.getReg();
}
int valArgIndx = lastAddrIndx + 1;
else
MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
if (regOpcL != X86::MOV32rr)
- MIB.addReg(tt1);
+ MIB.addReg(t1);
(*MIB).addOperand(*argOpers[valArgIndx]);
assert(argOpers[valArgIndx + 1]->isReg() ==
argOpers[valArgIndx]->isReg());
else
MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
if (regOpcH != X86::MOV32rr)
- MIB.addReg(tt2);
+ MIB.addReg(t2);
(*MIB).addOperand(*argOpers[valArgIndx + 1]);
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
(*MIB).addOperand(*argOpers[i]);
assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
- (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
+ (*MIB).setMemRefs(bInstr->memoperands_begin(),
+ bInstr->memoperands_end());
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
MIB.addReg(X86::EAX);
MIB.addReg(X86::EDX);
// insert branch
- BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
+ BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
return nextMBB;
(*MIB).addOperand(*argOpers[i]);
MIB.addReg(t3);
assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
- (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
+ (*MIB).setMemRefs(mInstr->memoperands_begin(),
+ mInstr->memoperands_end());
MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
MIB.addReg(X86::EAX);
// insert branch
- BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
+ BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
return nextMBB;
// all of this code can be replaced with that in the .td file.
MachineBasicBlock *
X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
- unsigned numArgs, bool memArg) const {
+ unsigned numArgs, bool memArg) const {
MachineFunction *F = BB->getParent();
DebugLoc dl = MI->getDebugLoc();
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
unsigned Opc;
-
- if (memArg) {
- Opc = numArgs == 3 ?
- X86::PCMPISTRM128rm :
- X86::PCMPESTRM128rm;
- } else {
- Opc = numArgs == 3 ?
- X86::PCMPISTRM128rr :
- X86::PCMPESTRM128rr;
- }
+ if (memArg)
+ Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
+ else
+ Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
if (!Subtarget->isTargetWin64()) {
// If %al is 0, branch around the XMM save block.
BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
- BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
+ BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
MBB->addSuccessor(EndMBB);
}
// In the XMM save block, save all the XMM argument registers.
for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
+ MachineMemOperand *MMO =
+ F->getMachineMemOperand(
+ PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
+ MachineMemOperand::MOStore, Offset,
+ /*Size=*/16, /*Align=*/16);
BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
.addFrameIndex(RegSaveFrameIndex)
.addImm(/*Scale=*/1)
.addImm(/*Disp=*/Offset)
.addReg(/*Segment=*/0)
.addReg(MI->getOperand(i).getReg())
- .addMemOperand(MachineMemOperand(
- PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
- MachineMemOperand::MOStore, Offset,
- /*Size=*/16, /*Align=*/16));
+ .addMemOperand(MMO);
}
F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
MachineBasicBlock *
X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
- MachineBasicBlock *BB) const {
+ MachineBasicBlock *BB,
+ DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
-
+
// To "insert" a SELECT_CC instruction, we actually have to insert the
// diamond control-flow pattern. The incoming instruction knows the
// destination vreg to set, the condition code register to branch on, the
const BasicBlock *LLVM_BB = BB->getBasicBlock();
MachineFunction::iterator It = BB;
++It;
-
+
// thisMBB:
// ...
// TrueVal = ...
BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
F->insert(It, copy0MBB);
F->insert(It, sinkMBB);
- // Update machine-CFG edges by transferring all successors of the current
+ // Update machine-CFG edges by first adding all successors of the current
// block to the new block which will contain the Phi node for the select.
- sinkMBB->transferSuccessors(BB);
-
+ // Also inform sdisel of the edge changes.
+ for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
+ E = BB->succ_end(); I != E; ++I) {
+ EM->insert(std::make_pair(*I, sinkMBB));
+ sinkMBB->addSuccessor(*I);
+ }
+ // Next, remove all successors of the current block, and add the true
+ // and fallthrough blocks as its successors.
+ while (!BB->succ_empty())
+ BB->removeSuccessor(BB->succ_begin());
// Add the true and fallthrough blocks as its successors.
BB->addSuccessor(copy0MBB);
BB->addSuccessor(sinkMBB);
-
+
// copy0MBB:
// %FalseValue = ...
// # fallthrough to sinkMBB
BB = copy0MBB;
-
+
// Update machine-CFG edges
BB->addSuccessor(sinkMBB);
-
+
// sinkMBB:
// %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
// ...
MachineBasicBlock *
X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
- MachineBasicBlock *BB) const {
+ MachineBasicBlock *BB,
+ DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
switch (MI->getOpcode()) {
default: assert(false && "Unexpected instr type to insert");
case X86::CMOV_GR8:
case X86::CMOV_V4F32:
case X86::CMOV_V2F64:
case X86::CMOV_V2I64:
- return EmitLoweredSelect(MI, BB);
+ return EmitLoweredSelect(MI, BB, EM);
case X86::FP32_TO_INT16_IN_MEM:
case X86::FP32_TO_INT32_IN_MEM:
// Change the floating point control register to use "round towards zero"
// mode when truncating to an integer value.
MachineFunction *F = BB->getParent();
- int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
+ int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
// Load the old value of the high byte of the control word...
case X86ISD::UMUL:
case X86ISD::INC:
case X86ISD::DEC:
+ case X86ISD::OR:
+ case X86ISD::XOR:
+ case X86ISD::AND:
// These nodes' second result is a boolean.
if (Op.getResNo() == 0)
break;
return TargetLowering::isGAPlusOffset(N, GA, Offset);
}
-static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
- const TargetLowering &TLI) {
- GlobalValue *GV;
- int64_t Offset = 0;
- if (TLI.isGAPlusOffset(Base, GV, Offset))
- return (GV->getAlignment() >= N && (Offset % N) == 0);
- // DAG combine handles the stack object case.
- return false;
-}
-
static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
- EVT EVT, LoadSDNode *&LDBase,
+ EVT EltVT, LoadSDNode *&LDBase,
unsigned &LastLoadedElt,
SelectionDAG &DAG, MachineFrameInfo *MFI,
const TargetLowering &TLI) {
continue;
LoadSDNode *LD = cast<LoadSDNode>(Elt);
- if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
+ if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
return false;
LastLoadedElt = i;
}
const TargetLowering &TLI) {
DebugLoc dl = N->getDebugLoc();
EVT VT = N->getValueType(0);
- EVT EVT = VT.getVectorElementType();
+ EVT EltVT = VT.getVectorElementType();
ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
unsigned NumElems = VT.getVectorNumElements();
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
LoadSDNode *LD = NULL;
unsigned LastLoadedElt;
- if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
+ if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
MFI, TLI))
return SDValue();
if (LastLoadedElt == NumElems - 1) {
- if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
+ if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile());
+ LD->isVolatile(), LD->isNonTemporal(), 0);
return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile(), LD->getAlignment());
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment());
} else if (NumElems == 4 && LastLoadedElt == 1) {
SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
SDValue LHS = N->getOperand(1);
SDValue RHS = N->getOperand(2);
- // If we have SSE[12] support, try to form min/max nodes.
+ // If we have SSE[12] support, try to form min/max nodes. SSE min/max
+ // instructions match the semantics of the common C idiom x<y?x:y but not
+ // x<=y?x:y, because of how they handle negative zero (which can be
+ // ignored in unsafe-math mode).
if (Subtarget->hasSSE2() &&
(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
Cond.getOpcode() == ISD::SETCC) {
ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
unsigned Opcode = 0;
+ // Check for x CC y ? x : y.
if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
switch (CC) {
default: break;
- case ISD::SETOLE: // (X <= Y) ? X : Y -> min
- case ISD::SETULE:
- case ISD::SETLE:
+ case ISD::SETULT:
+ if (!UnsafeFPMath) break;
+ Opcode = X86ISD::FMIN;
+ break;
+ case ISD::SETOLE:
if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
+ Opcode = X86ISD::FMIN;
+ break;
+ case ISD::SETULE:
+ case ISD::SETOLT:
case ISD::SETLT:
+ case ISD::SETLE:
Opcode = X86ISD::FMIN;
break;
- case ISD::SETOGT: // (X > Y) ? X : Y -> max
+ case ISD::SETOGE:
+ if (!UnsafeFPMath) break;
+ Opcode = X86ISD::FMAX;
+ break;
case ISD::SETUGT:
- case ISD::SETGT:
if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
+ Opcode = X86ISD::FMAX;
+ break;
+ case ISD::SETUGE:
+ case ISD::SETOGT:
+ case ISD::SETGT:
case ISD::SETGE:
Opcode = X86ISD::FMAX;
break;
}
+ // Check for x CC y ? y : x -- a min/max with reversed arms.
} else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
switch (CC) {
default: break;
- case ISD::SETOGT:
- // This can use a min only if the LHS isn't NaN.
- if (DAG.isKnownNeverNaN(LHS))
- Opcode = X86ISD::FMIN;
- else if (DAG.isKnownNeverNaN(RHS)) {
- Opcode = X86ISD::FMIN;
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- }
+ case ISD::SETOGE:
+ if (!UnsafeFPMath) break;
+ Opcode = X86ISD::FMIN;
break;
-
- case ISD::SETUGT: // (X > Y) ? Y : X -> min
- case ISD::SETGT:
+ case ISD::SETUGT:
if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
+ Opcode = X86ISD::FMIN;
+ break;
+ case ISD::SETUGE:
+ case ISD::SETOGT:
+ case ISD::SETGT:
case ISD::SETGE:
Opcode = X86ISD::FMIN;
break;
- case ISD::SETULE:
- // This can use a max only if the LHS isn't NaN.
- if (DAG.isKnownNeverNaN(LHS))
- Opcode = X86ISD::FMAX;
- else if (DAG.isKnownNeverNaN(RHS)) {
- Opcode = X86ISD::FMAX;
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- }
+ case ISD::SETULT:
+ if (!UnsafeFPMath) break;
+ Opcode = X86ISD::FMAX;
break;
-
- case ISD::SETOLE: // (X <= Y) ? Y : X -> max
- case ISD::SETLE:
+ case ISD::SETOLE:
if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
+ Opcode = X86ISD::FMAX;
+ break;
+ case ISD::SETULE:
+ case ISD::SETOLT:
case ISD::SETLT:
+ case ISD::SETLE:
Opcode = X86ISD::FMAX;
break;
}
return SDValue();
}
+/// PerformANDCombine - Look for SSE and instructions of this form:
+/// (and x, (build_vector c1,c2,c3,c4)). If there exists a use of a build_vector
+/// that's the bitwise complement of the mask, then transform the node to
+/// (and (xor x, (build_vector -1,-1,-1,-1)), (build_vector ~c1,~c2,~c3,~c4)).
+static SDValue PerformANDCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI) {
+ EVT VT = N->getValueType(0);
+ if (!VT.isVector() || !VT.isInteger())
+ return SDValue();
+
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ if (N0.getOpcode() == ISD::XOR || !N1.hasOneUse())
+ return SDValue();
+
+ if (N1.getOpcode() == ISD::BUILD_VECTOR) {
+ unsigned NumElts = VT.getVectorNumElements();
+ EVT EltVT = VT.getVectorElementType();
+ SmallVector<SDValue, 8> Mask;
+ Mask.reserve(NumElts);
+ for (unsigned i = 0; i != NumElts; ++i) {
+ SDValue Arg = N1.getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) {
+ Mask.push_back(Arg);
+ continue;
+ }
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(Arg);
+ if (!C) return SDValue();
+ Mask.push_back(DAG.getConstant(~C->getAPIntValue(), EltVT));
+ }
+ N1 = DAG.getNode(ISD::BUILD_VECTOR, N1.getDebugLoc(), VT,
+ &Mask[0], NumElts);
+ if (!N1.use_empty()) {
+ unsigned Bits = EltVT.getSizeInBits();
+ Mask.clear();
+ for (unsigned i = 0; i != NumElts; ++i)
+ Mask.push_back(DAG.getConstant(APInt::getAllOnesValue(Bits), EltVT));
+ SDValue NewMask = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
+ VT, &Mask[0], NumElts);
+ return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
+ DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
+ N0, NewMask), N1);
+ }
+ }
+
+ return SDValue();
+}
/// PerformMulCombine - Optimize a single multiply with constant into two
/// in order to implement it with two cheaper instructions, e.g.
return SDValue();
}
+static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
+ EVT VT = N0.getValueType();
+
+ // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
+ // since the result of setcc_c is all zero's or all ones.
+ if (N1C && N0.getOpcode() == ISD::AND &&
+ N0.getOperand(1).getOpcode() == ISD::Constant) {
+ SDValue N00 = N0.getOperand(0);
+ if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
+ ((N00.getOpcode() == ISD::ANY_EXTEND ||
+ N00.getOpcode() == ISD::ZERO_EXTEND) &&
+ N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
+ APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
+ APInt ShAmt = N1C->getAPIntValue();
+ Mask = Mask.shl(ShAmt);
+ if (Mask != 0)
+ return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
+ N00, DAG.getConstant(Mask, VT));
+ }
+ }
+
+ return SDValue();
+}
/// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
/// when possible.
static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
const X86Subtarget *Subtarget) {
+ EVT VT = N->getValueType(0);
+ if (!VT.isVector() && VT.isInteger() &&
+ N->getOpcode() == ISD::SHL)
+ return PerformSHLCombine(N, DAG);
+
// On X86 with SSE2 support, we can transform this to a vector shift if
// all elements are shifted by the same amount. We can't do this in legalize
// because the a constant vector is typically transformed to a constant pool
if (!Subtarget->hasSSE2())
return SDValue();
- EVT VT = N->getValueType(0);
if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
return SDValue();
}
} else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
- unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
+ unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
if (C->getZExtValue() == SplatIdx)
BaseShAmt = InVec.getOperand(1);
}
return SDValue();
}
+static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ EVT VT = N->getValueType(0);
+ if (VT != MVT::i64 || !Subtarget->is64Bit())
+ return SDValue();
+
+ // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
+ std::swap(N0, N1);
+ if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
+ return SDValue();
+
+ SDValue ShAmt0 = N0.getOperand(1);
+ if (ShAmt0.getValueType() != MVT::i8)
+ return SDValue();
+ SDValue ShAmt1 = N1.getOperand(1);
+ if (ShAmt1.getValueType() != MVT::i8)
+ return SDValue();
+ if (ShAmt0.getOpcode() == ISD::TRUNCATE)
+ ShAmt0 = ShAmt0.getOperand(0);
+ if (ShAmt1.getOpcode() == ISD::TRUNCATE)
+ ShAmt1 = ShAmt1.getOperand(0);
+
+ DebugLoc DL = N->getDebugLoc();
+ unsigned Opc = X86ISD::SHLD;
+ SDValue Op0 = N0.getOperand(0);
+ SDValue Op1 = N1.getOperand(0);
+ if (ShAmt0.getOpcode() == ISD::SUB) {
+ Opc = X86ISD::SHRD;
+ std::swap(Op0, Op1);
+ std::swap(ShAmt0, ShAmt1);
+ }
+
+ if (ShAmt1.getOpcode() == ISD::SUB) {
+ SDValue Sum = ShAmt1.getOperand(0);
+ if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
+ if (SumC->getSExtValue() == 64 &&
+ ShAmt1.getOperand(1) == ShAmt0)
+ return DAG.getNode(Opc, DL, VT,
+ Op0, Op1,
+ DAG.getNode(ISD::TRUNCATE, DL,
+ MVT::i8, ShAmt0));
+ }
+ } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
+ ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
+ if (ShAmt0C &&
+ ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
+ return DAG.getNode(Opc, DL, VT,
+ N0.getOperand(0), N1.getOperand(0),
+ DAG.getNode(ISD::TRUNCATE, DL,
+ MVT::i8, ShAmt0));
+ }
+
+ return SDValue();
+}
+
/// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
const X86Subtarget *Subtarget) {
SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
Ld->getBasePtr(), Ld->getSrcValue(),
Ld->getSrcValueOffset(), Ld->isVolatile(),
- Ld->getAlignment());
+ Ld->isNonTemporal(), Ld->getAlignment());
SDValue NewChain = NewLd.getValue(1);
if (TokenFactorIndex != -1) {
Ops.push_back(NewChain);
}
return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
St->getSrcValue(), St->getSrcValueOffset(),
- St->isVolatile(), St->getAlignment());
+ St->isVolatile(), St->isNonTemporal(),
+ St->getAlignment());
}
// Otherwise, lower to two pairs of 32-bit loads / stores.
SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
Ld->getSrcValue(), Ld->getSrcValueOffset(),
- Ld->isVolatile(), Ld->getAlignment());
+ Ld->isVolatile(), Ld->isNonTemporal(),
+ Ld->getAlignment());
SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
- Ld->isVolatile(),
+ Ld->isVolatile(), Ld->isNonTemporal(),
MinAlign(Ld->getAlignment(), 4));
SDValue NewChain = LoLd.getValue(1);
SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
St->getSrcValue(), St->getSrcValueOffset(),
- St->isVolatile(), St->getAlignment());
+ St->isVolatile(), St->isNonTemporal(),
+ St->getAlignment());
SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
St->getSrcValue(),
St->getSrcValueOffset() + 4,
St->isVolatile(),
+ St->isNonTemporal(),
MinAlign(St->getAlignment(), 4));
return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
}
}
}
+static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
+ // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
+ // (and (i32 x86isd::setcc_carry), 1)
+ // This eliminates the zext. This transformation is necessary because
+ // ISD::SETCC is always legalized to i8.
+ DebugLoc dl = N->getDebugLoc();
+ SDValue N0 = N->getOperand(0);
+ EVT VT = N->getValueType(0);
+ if (N0.getOpcode() == ISD::AND &&
+ N0.hasOneUse() &&
+ N0.getOperand(0).hasOneUse()) {
+ SDValue N00 = N0.getOperand(0);
+ if (N00.getOpcode() != X86ISD::SETCC_CARRY)
+ return SDValue();
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+ if (!C || C->getZExtValue() != 1)
+ return SDValue();
+ return DAG.getNode(ISD::AND, dl, VT,
+ DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
+ N00.getOperand(0), N00.getOperand(1)),
+ DAG.getConstant(1, VT));
+ }
+
+ return SDValue();
+}
+
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
+ case ISD::AND: return PerformANDCombine(N, DAG, DCI);
case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
case ISD::SHL:
case ISD::SRA:
case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
+ case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
case X86ISD::FXOR:
case X86ISD::FOR: return PerformFORCombine(N, DAG);
case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
+ case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
}
return SDValue();
// Verify this is a simple bswap.
if (CI->getNumOperands() != 2 ||
CI->getType() != CI->getOperand(1)->getType() ||
- !CI->getType()->isInteger())
+ !CI->getType()->isIntegerTy())
return false;
const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
std::string AsmStr = IA->getAsmString();
// TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
- std::vector<std::string> AsmPieces;
+ SmallVector<StringRef, 4> AsmPieces;
SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
switch (AsmPieces.size()) {
return LowerToBSwap(CI);
}
// rorw $$8, ${0:w} --> llvm.bswap.i16
- if (CI->getType() == Type::getInt16Ty(CI->getContext()) &&
+ if (CI->getType()->isIntegerTy(16) &&
AsmPieces.size() == 3 &&
AsmPieces[0] == "rorw" &&
AsmPieces[1] == "$$8," &&
}
break;
case 3:
- if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
+ if (CI->getType()->isIntegerTy(64) &&
Constraints.size() >= 2 &&
Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
// bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
- std::vector<std::string> Words;
+ SmallVector<StringRef, 4> Words;
SplitString(AsmPieces[0], Words, " \t");
if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
Words.clear();
switch (Constraint[0]) {
default: break;
case 'r': // GENERAL_REGS
- case 'R': // LEGACY_REGS
case 'l': // INDEX_REGS
if (VT == MVT::i8)
return std::make_pair(0U, X86::GR8RegisterClass);
if (VT == MVT::i32 || !Subtarget->is64Bit())
return std::make_pair(0U, X86::GR32RegisterClass);
return std::make_pair(0U, X86::GR64RegisterClass);
+ case 'R': // LEGACY_REGS
+ if (VT == MVT::i8)
+ return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
+ if (VT == MVT::i16)
+ return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
+ if (VT == MVT::i32 || !Subtarget->is64Bit())
+ return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
+ return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
case 'f': // FP Stack registers.
// If SSE is enabled for this VT, use f80 to ensure the isel moves the
// value to the correct fpstack register class.
// Not found as a standard register?
if (Res.second == 0) {
- // GCC calls "st(0)" just plain "st".
- if (StringsEqualNoCase("{st}", Constraint)) {
+ // Map st(0) -> st(7) -> ST0
+ if (Constraint.size() == 7 && Constraint[0] == '{' &&
+ tolower(Constraint[1]) == 's' &&
+ tolower(Constraint[2]) == 't' &&
+ Constraint[3] == '(' &&
+ (Constraint[4] >= '0' && Constraint[4] <= '7') &&
+ Constraint[5] == ')' &&
+ Constraint[6] == '}') {
+
+ Res.first = X86::ST0+Constraint[4]-'0';
+ Res.second = X86::RFP80RegisterClass;
+ return Res;
+ }
+
+ // GCC allows "st(0)" to be called just plain "st".
+ if (StringRef("{st}").equals_lower(Constraint)) {
Res.first = X86::ST0;
Res.second = X86::RFP80RegisterClass;
+ return Res;
+ }
+
+ // flags -> EFLAGS
+ if (StringRef("{flags}").equals_lower(Constraint)) {
+ Res.first = X86::EFLAGS;
+ Res.second = X86::CCRRegisterClass;
+ return Res;
}
+
// 'A' means EAX + EDX.
if (Constraint == "A") {
Res.first = X86::EAX;
Res.second = X86::GR32_ADRegisterClass;
+ return Res;
}
return Res;
}