#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
-#include "X86MCTargetExpr.h"
#include "X86TargetMachine.h"
#include "X86TargetObjectFile.h"
#include "llvm/CallingConv.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCContext.h"
+#include "llvm/MC/MCExpr.h"
#include "llvm/MC/MCSymbol.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/VectorExtras.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/Dwarf.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
+using namespace dwarf;
STATISTIC(NumTailCalls, "Number of tail calls");
case X86Subtarget::isDarwin:
if (TM.getSubtarget<X86Subtarget>().is64Bit())
return new X8664_MachoTargetObjectFile();
- return new X8632_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 (!VT.is128BitVector()) {
continue;
}
+
setOperationAction(ISD::AND, SVT, Promote);
AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
setOperationAction(ISD::OR, SVT, Promote);
// We have target-specific dag combine patterns for the following nodes:
setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
+ setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
setTargetDAGCombine(ISD::BUILD_VECTOR);
setTargetDAGCombine(ISD::SELECT);
setTargetDAGCombine(ISD::SHL);
// FIXME: These should be based on subtarget info. Plus, the values should
// be smaller when we are in optimizing for size mode.
maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
- maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
+ maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
setPrefLoopAlignment(16);
benefitFromCodePlacementOpt = true;
}
/// getOptimalMemOpType - Returns the target specific optimal type for load
-/// and store operations as a result of memset, memcpy, and memmove
-/// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
-/// determining it.
+/// and store operations as a result of memset, memcpy, and memmove lowering.
+/// If DstAlign is zero that means it's safe to destination alignment can
+/// satisfy any constraint. Similarly if SrcAlign is zero it means there
+/// isn't a need to check it against alignment requirement, probably because
+/// the source does not need to be loaded. If 'NonScalarIntSafe' is true, that
+/// means it's safe to return a non-scalar-integer type, e.g. constant string
+/// source or loaded from memory. It returns EVT::Other if SelectionDAG should
+/// be responsible for determining it.
EVT
-X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
- bool isSrcConst, bool isSrcStr,
+X86TargetLowering::getOptimalMemOpType(uint64_t Size,
+ unsigned DstAlign, unsigned SrcAlign,
+ bool NonScalarIntSafe,
SelectionDAG &DAG) const {
// FIXME: This turns off use of xmm stores for memset/memcpy on targets like
// linux. This is because the stack realignment code can't handle certain
// cases like PR2962. This should be removed when PR2962 is fixed.
const Function *F = DAG.getMachineFunction().getFunction();
- bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
- if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
- if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
- return MVT::v4i32;
- if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
- return MVT::v4f32;
+ if (NonScalarIntSafe &&
+ !F->hasFnAttr(Attribute::NoImplicitFloat)) {
+ if (Size >= 16 &&
+ (Subtarget->isUnalignedMemAccessFast() ||
+ ((DstAlign == 0 || DstAlign >= 16) &&
+ (SrcAlign == 0 || SrcAlign >= 16))) &&
+ Subtarget->getStackAlignment() >= 16) {
+ if (Subtarget->hasSSE2())
+ return MVT::v4i32;
+ if (Subtarget->hasSSE1())
+ return MVT::v4f32;
+ } else if (Size >= 8 &&
+ !Subtarget->is64Bit() &&
+ Subtarget->getStackAlignment() >= 8 &&
+ Subtarget->hasSSE2())
+ return MVT::f64;
}
if (Subtarget->is64Bit() && Size >= 8)
return MVT::i64;
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);
+ return MCSymbolRefExpr::Create(MBB->getSymbol(),
+ MCSymbolRefExpr::VK_GOTOFF, Ctx);
}
/// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
if (!Subtarget->is64Bit())
// This doesn't have DebugLoc associated with it, but is not really the
// same as a Register.
- return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
- getPointerTy());
+ return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
return Table;
}
return !Subtarget->is64Bit();
case CallingConv::Fast:
return GuaranteedTailCallOpt;
+ case CallingConv::GHC:
+ return GuaranteedTailCallOpt;
}
}
/// given CallingConvention value.
CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
if (Subtarget->is64Bit()) {
- if (Subtarget->isTargetWin64())
+ if (CC == CallingConv::GHC)
+ return CC_X86_64_GHC;
+ else if (Subtarget->isTargetWin64())
return CC_X86_Win64_C;
else
return CC_X86_64_C;
return CC_X86_32_FastCall;
else if (CC == CallingConv::Fast)
return CC_X86_32_FastCC;
+ else if (CC == CallingConv::GHC)
+ return CC_X86_32_GHC;
else
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
DebugLoc dl) {
SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
- /*AlwaysInline=*/true, NULL, 0, NULL, 0);
+ /*isVolatile*/false, /*AlwaysInline=*/true,
+ NULL, 0, NULL, 0);
+}
+
+/// IsTailCallConvention - Return true if the calling convention is one that
+/// supports tail call optimization.
+static bool IsTailCallConvention(CallingConv::ID CC) {
+ return (CC == CallingConv::Fast || CC == CallingConv::GHC);
}
/// 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;
+ return GuaranteedTailCallOpt && IsTailCallConvention(CC);
}
SDValue
VA.getLocMemOffset(), isImmutable, false);
SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
return DAG.getLoad(ValVT, dl, Chain, FIN,
- PseudoSourceValue::getFixedStack(FI), 0);
+ PseudoSourceValue::getFixedStack(FI), 0,
+ false, false, 0);
}
}
DebugLoc dl,
SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) {
-
MachineFunction &MF = DAG.getMachineFunction();
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
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();
- assert(!(isVarArg && CallConv == CallingConv::Fast) &&
- "Var args not supported with calling convention fastcc");
+ assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
+ "Var args not supported with calling convention fastcc or ghc");
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
// 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);
}
SDValue Store =
DAG.getStore(Val.getValue(1), dl, Val, FIN,
PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
- Offset);
+ Offset, false, false, 0);
MemOps.push_back(Store);
Offset += 8;
}
} 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))
+ if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
BytesToPopOnReturn = 4;
}
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
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, true,false);
+ 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;
}
if (isTailCall) {
// Check if it's really possible to do a tail call.
- isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
+ isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
+ isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
Outs, Ins, DAG);
// Sibcalls are automatically detected tailcalls which do not require
++NumTailCalls;
}
- assert(!(isVarArg && CallConv == CallingConv::Fast) &&
- "Var args not supported with calling convention fastcc");
+ assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
+ "Var args not supported with calling convention fastcc or ghc");
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
// This is a sibcall. The memory operands are available in caller's
// own caller's stack.
NumBytes = 0;
- else if (GuaranteedTailCallOpt && CallConv == CallingConv::Fast)
+ else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
int FPDiff = 0;
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 (!isTailCall) {
Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
DAG.getNode(X86ISD::GlobalBaseReg,
- DebugLoc::getUnknownLoc(),
- getPointerTy()),
+ DebugLoc(), getPointerTy()),
InFlag);
InFlag = Chain.getValue(1);
} else {
// Store relative to framepointer.
MemOpChains2.push_back(
DAG.getStore(ArgChain, dl, Arg, FIN,
- PseudoSourceValue::getFixedStack(FI), 0));
+ PseudoSourceValue::getFixedStack(FI), 0,
+ false, false, 0));
}
}
}
OpFlags);
}
- 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(TCReg, getPointerTy()),
- Callee,InFlag);
- 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 (RVLocs[i].isRegLoc())
MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
}
-
- assert(((Callee.getOpcode() == ISD::Register &&
- (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
- cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
- Callee.getOpcode() == ISD::TargetExternalSymbol ||
- Callee.getOpcode() == ISD::TargetGlobalAddress) &&
- "Expecting a global address, external symbol, or scratch register");
-
return DAG.getNode(X86ISD::TC_RETURN, dl,
NodeTys, &Ops[0], Ops.size());
}
unsigned NumBytesForCalleeToPush;
if (IsCalleePop(isVarArg, CallConv))
NumBytesForCalleeToPush = NumBytes; // Callee pops everything
- else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
+ else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
// 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.
bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
const X86InstrInfo *TII) {
- int FI;
+ unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
+ int FI = INT_MAX;
if (Arg.getOpcode() == ISD::CopyFromReg) {
unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
Def->getOperand(1).isFI()) {
FI = Def->getOperand(1).getIndex();
- if (MFI->getObjectSize(FI) != Flags.getByValSize())
- return false;
+ Bytes = Flags.getByValSize();
} else
return false;
}
- } else {
- LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
- if (!Ld)
+ } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
+ if (Flags.isByVal())
+ // ByVal argument is passed in as a pointer but it's now being
+ // dereferenced. e.g.
+ // define @foo(%struct.X* %A) {
+ // tail call @bar(%struct.X* byval %A)
+ // }
return false;
SDValue Ptr = Ld->getBasePtr();
FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
if (!FINode)
return false;
FI = FINode->getIndex();
- }
+ } else
+ return false;
+ assert(FI != INT_MAX);
if (!MFI->isFixedObjectIndex(FI))
return false;
- return Offset == MFI->getObjectOffset(FI);
+ return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
}
/// IsEligibleForTailCallOptimization - Check whether the call is eligible
X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
CallingConv::ID CalleeCC,
bool isVarArg,
+ bool isCalleeStructRet,
+ bool isCallerStructRet,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<ISD::InputArg> &Ins,
SelectionDAG& DAG) const {
- if (CalleeCC != CallingConv::Fast &&
+ if (!IsTailCallConvention(CalleeCC) &&
CalleeCC != CallingConv::C)
return false;
// If -tailcallopt is specified, make fastcc functions tail-callable.
+ const MachineFunction &MF = DAG.getMachineFunction();
const Function *CallerF = DAG.getMachineFunction().getFunction();
if (GuaranteedTailCallOpt) {
- if (CalleeCC == CallingConv::Fast &&
+ if (IsTailCallConvention(CalleeCC) &&
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)
+ // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
+ // emit a special epilogue.
+ if (RegInfo->needsStackRealignment(MF))
+ return false;
+
+ // Do not sibcall optimize vararg calls unless the call site is not passing any
+ // arguments.
+ if (isVarArg && !Outs.empty())
return false;
+ // Also avoid sibcall optimization if either caller or callee uses struct
+ // return semantics.
+ if (isCalleeStructRet || isCallerStructRet)
+ return false;
+
+ // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
+ // Therefore if it's not used by the call it is not safe to optimize this into
+ // a sibcall.
+ bool Unused = false;
+ for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
+ if (!Ins[i].Used) {
+ Unused = true;
+ break;
+ }
+ }
+ if (Unused) {
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState CCInfo(CalleeCC, false, getTargetMachine(),
+ RVLocs, *DAG.getContext());
+ CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
+ for (unsigned i = 0; i != RVLocs.size(); ++i) {
+ CCValAssign &VA = RVLocs[i];
+ if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
+ return false;
+ }
+ }
+
// If the callee takes no arguments then go on to check the results of the
// call.
if (!Outs.empty()) {
}
FastISel *
-X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
- DwarfWriter *dw,
+X86TargetLowering::createFastISel(MachineFunction &mf,
DenseMap<const Value *, unsigned> &vm,
DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
DenseMap<const AllocaInst *, int> &am
, SmallSet<Instruction*, 8> &cil
#endif
) {
- return X86::createFastISel(mf, mmo, dw, vm, bm, am
+ return X86::createFastISel(mf, vm, bm, am
#ifndef NDEBUG
, cil
#endif
// Set up a frame object for the return address.
uint64_t SlotSize = TD->getPointerSize();
ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
- true, false);
+ false, false);
FuncInfo->setRAIndex(ReturnAddrIndex);
}
bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
bool hasSymbolicDisplacement) {
// Offset should fit into 32 bit immediate field.
- if (!isInt32(Offset))
+ if (!isInt<32>(Offset))
return false;
// If we don't have a symbolic displacement - we don't have any extra
/// FIXME: split into pslldqi, psrldqi, palignr variants.
static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
- int NumElems = SVOp->getValueType(0).getVectorNumElements();
+ unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
isLeft = true;
unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
}
bool SeenV1 = false;
bool SeenV2 = false;
- for (int i = NumZeros; i < NumElems; ++i) {
- int Val = isLeft ? (i - NumZeros) : i;
- int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
- if (Idx < 0)
+ for (unsigned i = NumZeros; i < NumElems; ++i) {
+ unsigned Val = isLeft ? (i - NumZeros) : i;
+ int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
+ if (Idx_ < 0)
continue;
+ unsigned Idx = (unsigned) Idx_;
if (Idx < NumElems)
SeenV1 = true;
else {
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);
+ 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,
return SDValue();
}
+/// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
+/// vector of type 'VT', see if the elements can be replaced by a single large
+/// load which has the same value as a build_vector whose operands are 'elts'.
+///
+/// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
+///
+/// FIXME: we'd also like to handle the case where the last elements are zero
+/// rather than undef via VZEXT_LOAD, but we do not detect that case today.
+/// There's even a handy isZeroNode for that purpose.
+static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
+ DebugLoc &dl, SelectionDAG &DAG) {
+ EVT EltVT = VT.getVectorElementType();
+ unsigned NumElems = Elts.size();
+
+ LoadSDNode *LDBase = NULL;
+ unsigned LastLoadedElt = -1U;
+
+ // For each element in the initializer, see if we've found a load or an undef.
+ // If we don't find an initial load element, or later load elements are
+ // non-consecutive, bail out.
+ for (unsigned i = 0; i < NumElems; ++i) {
+ SDValue Elt = Elts[i];
+
+ if (!Elt.getNode() ||
+ (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
+ return SDValue();
+ if (!LDBase) {
+ if (Elt.getNode()->getOpcode() == ISD::UNDEF)
+ return SDValue();
+ LDBase = cast<LoadSDNode>(Elt.getNode());
+ LastLoadedElt = i;
+ continue;
+ }
+ if (Elt.getOpcode() == ISD::UNDEF)
+ continue;
+
+ LoadSDNode *LD = cast<LoadSDNode>(Elt);
+ if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
+ return SDValue();
+ LastLoadedElt = i;
+ }
+
+ // If we have found an entire vector of loads and undefs, then return a large
+ // load of the entire vector width starting at the base pointer. If we found
+ // consecutive loads for the low half, generate a vzext_load node.
+ if (LastLoadedElt == NumElems - 1) {
+ if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
+ return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
+ LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
+ LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
+ return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
+ LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
+ LDBase->isVolatile(), LDBase->isNonTemporal(),
+ LDBase->getAlignment());
+ } else if (NumElems == 4 && LastLoadedElt == 1) {
+ SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
+ SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
+ SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
+ }
+ return SDValue();
+}
+
SDValue
X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
DebugLoc dl = Op.getDebugLoc();
return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
}
- if (Values.size() > 2) {
- // If we have SSE 4.1, Expand into a number of inserts unless the number of
- // values to be inserted is equal to the number of elements, in which case
- // use the unpack code below in the hopes of matching the consecutive elts
- // load merge pattern for shuffles.
- // FIXME: We could probably just check that here directly.
- if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
- getSubtarget()->hasSSE41()) {
+ if (Values.size() > 1 && VT.getSizeInBits() == 128) {
+ // Check for a build vector of consecutive loads.
+ for (unsigned i = 0; i < NumElems; ++i)
+ V[i] = Op.getOperand(i);
+
+ // Check for elements which are consecutive loads.
+ SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
+ if (LD.getNode())
+ return LD;
+
+ // For SSE 4.1, use inserts into undef.
+ if (getSubtarget()->hasSSE41()) {
V[0] = DAG.getUNDEF(VT);
for (unsigned i = 0; i < NumElems; ++i)
if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
Op.getOperand(i), DAG.getIntPtrConstant(i));
return V[0];
}
- // Expand into a number of unpckl*.
+
+ // Otherwise, expand into a number of unpckl*
// e.g. for v4f32
// Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
// : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
}
return V[0];
}
-
return SDValue();
}
if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
isa<ConstantSDNode>(N2)) {
- unsigned Opc = (EltVT.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)
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();
}
if (OpFlag) {
Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
DAG.getNode(X86ISD::GlobalBaseReg,
- DebugLoc::getUnknownLoc(), getPointerTy()),
+ DebugLoc(), getPointerTy()),
Result);
}
if (OpFlag) {
Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
DAG.getNode(X86ISD::GlobalBaseReg,
- DebugLoc::getUnknownLoc(), getPointerTy()),
+ DebugLoc(), getPointerTy()),
Result);
}
!Subtarget->is64Bit()) {
Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
DAG.getNode(X86ISD::GlobalBaseReg,
- DebugLoc::getUnknownLoc(),
- getPointerTy()),
+ DebugLoc(), getPointerTy()),
Result);
}
// 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.
DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
DAG.getNode(X86ISD::GlobalBaseReg,
- DebugLoc::getUnknownLoc(),
- PtrVT), InFlag);
+ DebugLoc(), PtrVT), InFlag);
InFlag = Chain.getValue(1);
return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
DebugLoc dl = GA->getDebugLoc();
// Get the Thread Pointer
SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
- DebugLoc::getUnknownLoc(), PtrVT,
+ DebugLoc(), PtrVT,
DAG.getRegister(is64Bit? X86::FS : X86::GS,
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;
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);
}
};
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);
}
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())
// 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);
}
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.
/// 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,
+static SDValue LowerToBT(SDValue And, ISD::CondCode CC,
DebugLoc dl, SelectionDAG &DAG) {
+ SDValue Op0 = And.getOperand(0);
+ SDValue Op1 = And.getOperand(1);
+ if (Op0.getOpcode() == ISD::TRUNCATE)
+ Op0 = Op0.getOperand(0);
+ if (Op1.getOpcode() == ISD::TRUNCATE)
+ Op1 = Op1.getOperand(0);
+
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);
+ if (Op1.getOpcode() == ISD::SHL) {
+ if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
+ if (And10C->getZExtValue() == 1) {
+ LHS = Op0;
+ RHS = Op1.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.getOpcode() == ISD::SHL) {
+ if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
+ if (And00C->getZExtValue() == 1) {
+ LHS = Op1;
+ RHS = Op0.getOperand(1);
}
- } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
- ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
- SDValue AndLHS = Op0.getOperand(0);
+ } else if (Op1.getOpcode() == ISD::Constant) {
+ ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
+ SDValue AndLHS = Op0;
if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
LHS = AndLHS.getOperand(0);
RHS = AndLHS.getOperand(1);
return NewSetCC;
}
+ // Look for "(setcc) == / != 1" to avoid unncessary setcc.
+ if (Op0.getOpcode() == X86ISD::SETCC &&
+ Op1.getOpcode() == ISD::Constant &&
+ (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
+ cast<ConstantSDNode>(Op1)->isNullValue()) &&
+ (CC == ISD::SETEQ || CC == ISD::SETNE)) {
+ X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
+ bool Invert = (CC == ISD::SETNE) ^
+ cast<ConstantSDNode>(Op1)->isNullValue();
+ if (Invert)
+ CCode = X86::GetOppositeBranchCondition(CCode);
+ return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
+ }
+
bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
if (X86CC == X86::COND_INVALID)
N2C && N2C->isNullValue() &&
RHSC && RHSC->isNullValue()) {
SDValue CmpOp0 = Cmp.getOperand(0);
- Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
+ Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
EVT IntPtr = getPointerTy();
EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
- Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
-
Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
Flag = Chain.getValue(1);
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
- SDValue Ops[] = { Chain,
- DAG.getTargetExternalSymbol("_alloca", IntPtr),
- DAG.getRegister(X86::EAX, IntPtr),
- DAG.getRegister(X86StackPtr, SPTy),
- Flag };
- Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
- Flag = Chain.getValue(1);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
- Chain = DAG.getCALLSEQ_END(Chain,
- DAG.getIntPtrConstant(0, true),
- DAG.getIntPtrConstant(0, true),
- Flag);
+ Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
+ Flag = Chain.getValue(1);
Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
SDValue Chain,
SDValue Dst, SDValue Src,
SDValue Size, unsigned Align,
+ bool isVolatile,
const Value *DstSV,
uint64_t DstSVOff) {
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
false, false, false, false,
0, CallingConv::C, false, /*isReturnValueUsed=*/false,
- DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
- DAG.GetOrdering(Chain.getNode()));
+ DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
return CallResult.second;
}
DAG.getConstant(Offset, AddrVT)),
Src,
DAG.getConstant(BytesLeft, SizeVT),
- Align, DstSV, DstSVOff + Offset);
+ Align, isVolatile, DstSV, DstSVOff + Offset);
}
// TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
SDValue Chain, SDValue Dst, SDValue Src,
SDValue Size, unsigned Align,
- bool AlwaysInline,
+ bool isVolatile, bool AlwaysInline,
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff) {
// This requires the copy size to be a constant, preferrably
DAG.getNode(ISD::ADD, dl, SrcVT, Src,
DAG.getConstant(Offset, SrcVT)),
DAG.getConstant(BytesLeft, SizeVT),
- Align, AlwaysInline,
+ Align, isVolatile, AlwaysInline,
DstSV, DstSVOff + Offset,
SrcSV, SrcSVOff + Offset));
}
// 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());
DebugLoc dl = Op.getDebugLoc();
return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
- DAG.getIntPtrConstant(24), 8, false,
- DstSV, 0, SrcSV, 0);
+ DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
+ false, DstSV, 0, SrcSV, 0);
}
SDValue
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);
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) };
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 = 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) };
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 =
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::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::MUL_IMM: return "X86ISD::MUL_IMM";
case X86ISD::PTEST: return "X86ISD::PTEST";
case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
+ case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
}
}
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();
if (NumBits1 <= NumBits2)
return false;
- return Subtarget->is64Bit() || NumBits1 < 64;
+ return true;
}
bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
unsigned NumBits2 = VT2.getSizeInBits();
if (NumBits1 <= NumBits2)
return false;
- return Subtarget->is64Bit() || NumBits1 < 64;
+ return true;
}
bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
// x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
- return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
+ return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
}
bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
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;
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.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;
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);
}
return BB;
}
+MachineBasicBlock *
+X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
+ MachineBasicBlock *BB,
+ DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ DebugLoc DL = MI->getDebugLoc();
+ MachineFunction *F = BB->getParent();
+
+ // The lowering is pretty easy: we're just emitting the call to _alloca. The
+ // non-trivial part is impdef of ESP.
+ // FIXME: The code should be tweaked as soon as we'll try to do codegen for
+ // mingw-w64.
+
+ BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
+ .addExternalSymbol("_alloca")
+ .addReg(X86::EAX, RegState::Implicit)
+ .addReg(X86::ESP, RegState::Implicit)
+ .addReg(X86::EAX, RegState::Define | RegState::Implicit)
+ .addReg(X86::ESP, RegState::Define | RegState::Implicit);
+
+ F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
+ return BB;
+}
MachineBasicBlock *
X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
switch (MI->getOpcode()) {
default: assert(false && "Unexpected instr type to insert");
+ case X86::MINGW_ALLOCA:
+ return EmitLoweredMingwAlloca(MI, BB, EM);
case X86::CMOV_GR8:
case X86::CMOV_V1I64:
case X86::CMOV_FR32:
case X86::CMOV_V4F32:
case X86::CMOV_V2F64:
case X86::CMOV_V2I64:
+ case X86::CMOV_GR16:
+ case X86::CMOV_GR32:
+ case X86::CMOV_RFP32:
+ case X86::CMOV_RFP64:
+ case X86::CMOV_RFP80:
return EmitLoweredSelect(MI, BB, EM);
case X86::FP32_TO_INT16_IN_MEM:
F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
return BB;
}
+ // DBG_VALUE. Only the frame index case is done here.
+ case X86::DBG_VALUE: {
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ DebugLoc DL = MI->getDebugLoc();
+ X86AddressMode AM;
+ MachineFunction *F = BB->getParent();
+ AM.BaseType = X86AddressMode::FrameIndexBase;
+ AM.Base.FrameIndex = MI->getOperand(0).getImm();
+ addFullAddress(BuildMI(BB, DL, TII->get(X86::DBG_VALUE)), AM).
+ addImm(MI->getOperand(1).getImm()).
+ addMetadata(MI->getOperand(2).getMetadata());
+ F->DeleteMachineInstr(MI); // Remove pseudo.
+ return BB;
+ }
+
// String/text processing lowering.
case X86::PCMPISTRM128REG:
return EmitPCMP(MI, BB, 3, false /* in-mem */);
return TargetLowering::isGAPlusOffset(N, GA, Offset);
}
-static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
- EVT EltVT, LoadSDNode *&LDBase,
- unsigned &LastLoadedElt,
- SelectionDAG &DAG, MachineFrameInfo *MFI,
- const TargetLowering &TLI) {
- LDBase = NULL;
- LastLoadedElt = -1U;
- for (unsigned i = 0; i < NumElems; ++i) {
- if (N->getMaskElt(i) < 0) {
- if (!LDBase)
- return false;
- continue;
- }
-
- SDValue Elt = DAG.getShuffleScalarElt(N, i);
- if (!Elt.getNode() ||
- (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
- return false;
- if (!LDBase) {
- if (Elt.getNode()->getOpcode() == ISD::UNDEF)
- return false;
- LDBase = cast<LoadSDNode>(Elt.getNode());
- LastLoadedElt = i;
- continue;
- }
- if (Elt.getOpcode() == ISD::UNDEF)
- continue;
-
- LoadSDNode *LD = cast<LoadSDNode>(Elt);
- if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
- return false;
- LastLoadedElt = i;
- }
- return true;
-}
-
/// PerformShuffleCombine - Combine a vector_shuffle that is equal to
/// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
/// if the load addresses are consecutive, non-overlapping, and in the right
-/// order. In the case of v2i64, it will see if it can rewrite the
-/// shuffle to be an appropriate build vector so it can take advantage of
-// performBuildVectorCombine.
+/// order.
static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
const TargetLowering &TLI) {
DebugLoc dl = N->getDebugLoc();
EVT VT = N->getValueType(0);
- EVT EltVT = VT.getVectorElementType();
ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
- unsigned NumElems = VT.getVectorNumElements();
if (VT.getSizeInBits() != 128)
return SDValue();
- // Try to combine a vector_shuffle into a 128-bit load.
- MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
- LoadSDNode *LD = NULL;
- unsigned LastLoadedElt;
- if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
- MFI, TLI))
+ SmallVector<SDValue, 16> Elts;
+ for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
+ Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
+
+ return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
+}
+
+/// PerformShuffleCombine - Detect vector gather/scatter index generation
+/// and convert it from being a bunch of shuffles and extracts to a simple
+/// store and scalar loads to extract the elements.
+static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
+ const TargetLowering &TLI) {
+ SDValue InputVector = N->getOperand(0);
+
+ // Only operate on vectors of 4 elements, where the alternative shuffling
+ // gets to be more expensive.
+ if (InputVector.getValueType() != MVT::v4i32)
return SDValue();
- if (LastLoadedElt == NumElems - 1) {
- if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
- return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
- LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile());
- return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
- LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile(), LD->getAlignment());
- } else if (NumElems == 4 && LastLoadedElt == 1) {
- SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
- SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
- SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
+ // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
+ // single use which is a sign-extend or zero-extend, and all elements are
+ // used.
+ SmallVector<SDNode *, 4> Uses;
+ unsigned ExtractedElements = 0;
+ for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
+ UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
+ if (UI.getUse().getResNo() != InputVector.getResNo())
+ return SDValue();
+
+ SDNode *Extract = *UI;
+ if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
+ return SDValue();
+
+ if (Extract->getValueType(0) != MVT::i32)
+ return SDValue();
+ if (!Extract->hasOneUse())
+ return SDValue();
+ if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
+ Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
+ return SDValue();
+ if (!isa<ConstantSDNode>(Extract->getOperand(1)))
+ return SDValue();
+
+ // Record which element was extracted.
+ ExtractedElements |=
+ 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
+
+ Uses.push_back(Extract);
+ }
+
+ // If not all the elements were used, this may not be worthwhile.
+ if (ExtractedElements != 15)
+ return SDValue();
+
+ // Ok, we've now decided to do the transformation.
+ DebugLoc dl = InputVector.getDebugLoc();
+
+ // Store the value to a temporary stack slot.
+ SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
+ SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
+ false, false, 0);
+
+ // Replace each use (extract) with a load of the appropriate element.
+ for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
+ UE = Uses.end(); UI != UE; ++UI) {
+ SDNode *Extract = *UI;
+
+ // Compute the element's address.
+ SDValue Idx = Extract->getOperand(1);
+ unsigned EltSize =
+ InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
+ uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
+ SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
+
+ SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
+
+ // Load the scalar.
+ SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
+ NULL, 0, false, false, 0);
+
+ // Replace the exact with the load.
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
}
+
+ // The replacement was made in place; don't return anything.
return SDValue();
}
SDValue RHS = N->getOperand(2);
// If we have SSE[12] support, try to form min/max nodes. SSE min/max
- // instructions have the peculiarity that if either operand is a NaN,
- // they chose what we call the RHS operand (and as such are not symmetric).
- // It happens that this matches the semantics of the common C idiom
- // x<y?x:y and related forms, so we can recognize these cases.
+ // 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) {
unsigned Opcode = 0;
// Check for x CC y ? x : y.
- if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
+ if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
+ DAG.isEqualTo(RHS, Cond.getOperand(1))) {
switch (CC) {
default: break;
case ISD::SETULT:
- // This can be a min if we can prove that at least one of the operands
- // is not a nan.
- if (!FiniteOnlyFPMath()) {
- if (DAG.isKnownNeverNaN(RHS)) {
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- } else if (!DAG.isKnownNeverNaN(LHS))
+ // Converting this to a min would handle NaNs incorrectly, and swapping
+ // the operands would cause it to handle comparisons between positive
+ // and negative zero incorrectly.
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
+ if (!UnsafeFPMath &&
+ !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
break;
+ std::swap(LHS, RHS);
}
Opcode = X86ISD::FMIN;
break;
case ISD::SETOLE:
- // This can be a min if we can prove that at least one of the operands
- // is not a nan.
- if (!FiniteOnlyFPMath()) {
- if (DAG.isKnownNeverNaN(LHS)) {
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- } else if (!DAG.isKnownNeverNaN(RHS))
- break;
- }
+ // Converting this to a min would handle comparisons between positive
+ // and negative zero incorrectly.
+ if (!UnsafeFPMath &&
+ !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
+ break;
Opcode = X86ISD::FMIN;
break;
case ISD::SETULE:
- // This can be a min, but if either operand is a NaN we need it to
- // preserve the original LHS.
+ // Converting this to a min would handle both negative zeros and NaNs
+ // incorrectly, but we can swap the operands to fix both.
std::swap(LHS, RHS);
case ISD::SETOLT:
case ISD::SETLT:
break;
case ISD::SETOGE:
- // This can be a max if we can prove that at least one of the operands
- // is not a nan.
- if (!FiniteOnlyFPMath()) {
- if (DAG.isKnownNeverNaN(LHS)) {
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- } else if (!DAG.isKnownNeverNaN(RHS))
- break;
- }
+ // Converting this to a max would handle comparisons between positive
+ // and negative zero incorrectly.
+ if (!UnsafeFPMath &&
+ !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
+ break;
Opcode = X86ISD::FMAX;
break;
case ISD::SETUGT:
- // This can be a max if we can prove that at least one of the operands
- // is not a nan.
- if (!FiniteOnlyFPMath()) {
- if (DAG.isKnownNeverNaN(RHS)) {
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- } else if (!DAG.isKnownNeverNaN(LHS))
+ // Converting this to a max would handle NaNs incorrectly, and swapping
+ // the operands would cause it to handle comparisons between positive
+ // and negative zero incorrectly.
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
+ if (!UnsafeFPMath &&
+ !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
break;
+ std::swap(LHS, RHS);
}
Opcode = X86ISD::FMAX;
break;
case ISD::SETUGE:
- // This can be a max, but if either operand is a NaN we need it to
- // preserve the original LHS.
+ // Converting this to a max would handle both negative zeros and NaNs
+ // incorrectly, but we can swap the operands to fix both.
std::swap(LHS, RHS);
case ISD::SETOGT:
case ISD::SETGT:
break;
}
// Check for x CC y ? y : x -- a min/max with reversed arms.
- } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
+ } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
+ DAG.isEqualTo(RHS, Cond.getOperand(0))) {
switch (CC) {
default: break;
case ISD::SETOGE:
- // This can be a min if we can prove that at least one of the operands
- // is not a nan.
- if (!FiniteOnlyFPMath()) {
- if (DAG.isKnownNeverNaN(RHS)) {
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- } else if (!DAG.isKnownNeverNaN(LHS))
+ // Converting this to a min would handle comparisons between positive
+ // and negative zero incorrectly, and swapping the operands would
+ // cause it to handle NaNs incorrectly.
+ if (!UnsafeFPMath &&
+ !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
break;
+ std::swap(LHS, RHS);
}
Opcode = X86ISD::FMIN;
break;
case ISD::SETUGT:
- // This can be a min if we can prove that at least one of the operands
- // is not a nan.
- if (!FiniteOnlyFPMath()) {
- if (DAG.isKnownNeverNaN(LHS)) {
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- } else if (!DAG.isKnownNeverNaN(RHS))
- break;
- }
+ // Converting this to a min would handle NaNs incorrectly.
+ if (!UnsafeFPMath &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
+ break;
Opcode = X86ISD::FMIN;
break;
case ISD::SETUGE:
- // This can be a min, but if either operand is a NaN we need it to
- // preserve the original LHS.
+ // Converting this to a min would handle both negative zeros and NaNs
+ // incorrectly, but we can swap the operands to fix both.
std::swap(LHS, RHS);
case ISD::SETOGT:
case ISD::SETGT:
break;
case ISD::SETULT:
- // This can be a max if we can prove that at least one of the operands
- // is not a nan.
- if (!FiniteOnlyFPMath()) {
- if (DAG.isKnownNeverNaN(LHS)) {
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- } else if (!DAG.isKnownNeverNaN(RHS))
- break;
- }
+ // Converting this to a max would handle NaNs incorrectly.
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
+ break;
Opcode = X86ISD::FMAX;
break;
case ISD::SETOLE:
- // This can be a max if we can prove that at least one of the operands
- // is not a nan.
- if (!FiniteOnlyFPMath()) {
- if (DAG.isKnownNeverNaN(RHS)) {
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
- } else if (!DAG.isKnownNeverNaN(LHS))
+ // Converting this to a max would handle comparisons between positive
+ // and negative zero incorrectly, and swapping the operands would
+ // cause it to handle NaNs incorrectly.
+ if (!UnsafeFPMath &&
+ !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
break;
+ std::swap(LHS, RHS);
}
Opcode = X86ISD::FMAX;
break;
case ISD::SETULE:
- // This can be a max, but if either operand is a NaN we need it to
- // preserve the original LHS.
+ // Converting this to a max would handle both negative zeros and NaNs
+ // incorrectly, but we can swap the operands to fix both.
std::swap(LHS, RHS);
case ISD::SETOLT:
case ISD::SETLT:
/// LEA + SHL, LEA + LEA.
static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI) {
- if (DAG.getMachineFunction().
- getFunction()->hasFnAttr(Attribute::OptimizeForSize))
- return SDValue();
-
if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
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);
}
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);
}
switch (N->getOpcode()) {
default: break;
case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
+ case ISD::EXTRACT_VECTOR_ELT:
+ return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
// 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());
return LowerToBSwap(CI);
}
// rorw $$8, ${0:w} --> llvm.bswap.i16
- if (CI->getType()->isInteger(16) &&
+ if (CI->getType()->isIntegerTy(16) &&
AsmPieces.size() == 3 &&
- AsmPieces[0] == "rorw" &&
+ (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
AsmPieces[1] == "$$8," &&
AsmPieces[2] == "${0:w}" &&
- IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
- return LowerToBSwap(CI);
+ IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
+ AsmPieces.clear();
+ const std::string &Constraints = IA->getConstraintString();
+ SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
+ std::sort(AsmPieces.begin(), AsmPieces.end());
+ if (AsmPieces.size() == 4 &&
+ AsmPieces[0] == "~{cc}" &&
+ AsmPieces[1] == "~{dirflag}" &&
+ AsmPieces[2] == "~{flags}" &&
+ AsmPieces[3] == "~{fpsr}") {
+ return LowerToBSwap(CI);
+ }
}
break;
case 3:
- if (CI->getType()->isInteger(64) &&
+ 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") {
return Res;
}
-
-//===----------------------------------------------------------------------===//
-// X86 Widen vector type
-//===----------------------------------------------------------------------===//
-
-/// getWidenVectorType: given a vector type, returns the type to widen
-/// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
-/// If there is no vector type that we want to widen to, returns MVT::Other
-/// When and where to widen is target dependent based on the cost of
-/// scalarizing vs using the wider vector type.
-
-EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
- assert(VT.isVector());
- if (isTypeLegal(VT))
- return VT;
-
- // TODO: In computeRegisterProperty, we can compute the list of legal vector
- // type based on element type. This would speed up our search (though
- // it may not be worth it since the size of the list is relatively
- // small).
- EVT EltVT = VT.getVectorElementType();
- unsigned NElts = VT.getVectorNumElements();
-
- // On X86, it make sense to widen any vector wider than 1
- if (NElts <= 1)
- return MVT::Other;
-
- for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
- nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
- EVT SVT = (MVT::SimpleValueType)nVT;
-
- if (isTypeLegal(SVT) &&
- SVT.getVectorElementType() == EltVT &&
- SVT.getVectorNumElements() > NElts)
- return SVT;
- }
- return MVT::Other;
-}
projects / oota-llvm.git / blobdiff