static SDValue Insert128BitVector(SDValue Result, SDValue Vec,
unsigned IdxVal, SelectionDAG &DAG,
DebugLoc dl) {
+ // Inserting UNDEF is Result
+ if (Vec.getOpcode() == ISD::UNDEF)
+ return Result;
+
EVT VT = Vec.getValueType();
assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
* ElemsPerChunk);
SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
- Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
- VecIdx);
- return Result;
+ return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
+ VecIdx);
}
/// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
if (Subtarget->isTargetEnvMacho()) {
if (is64Bit)
- return new X8664_MachoTargetObjectFile();
+ return new X86_64MachoTargetObjectFile();
return new TargetLoweringObjectFileMachO();
}
+ if (Subtarget->isTargetLinux())
+ return new X86LinuxTargetObjectFile();
if (Subtarget->isTargetELF())
return new TargetLoweringObjectFileELF();
if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
// We want to custom lower some of our intrinsics.
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
+ setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
// Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
setTargetDAGCombine(ISD::FP_TO_SINT);
if (Subtarget->is64Bit())
setTargetDAGCombine(ISD::MUL);
- if (Subtarget->hasBMI())
- setTargetDAGCombine(ISD::XOR);
+ setTargetDAGCombine(ISD::XOR);
computeRegisterProperties();
SDValue ValToCopy = OutVals[i];
EVT ValVT = ValToCopy.getValueType();
+ // Promote values to the appropriate types
+ if (VA.getLocInfo() == CCValAssign::SExt)
+ ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
+ else if (VA.getLocInfo() == CCValAssign::ZExt)
+ ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
+ else if (VA.getLocInfo() == CCValAssign::AExt)
+ ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
+ else if (VA.getLocInfo() == CCValAssign::BCvt)
+ ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
+
// If this is x86-64, and we disabled SSE, we can't return FP values,
// or SSE or MMX vectors.
if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
}
SDValue
-X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
- CallingConv::ID CallConv, bool isVarArg,
- bool doesNotRet, bool &isTailCall,
- const SmallVectorImpl<ISD::OutputArg> &Outs,
- const SmallVectorImpl<SDValue> &OutVals,
- const SmallVectorImpl<ISD::InputArg> &Ins,
- DebugLoc dl, SelectionDAG &DAG,
+X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
SmallVectorImpl<SDValue> &InVals) const {
+ SelectionDAG &DAG = CLI.DAG;
+ DebugLoc &dl = CLI.DL;
+ SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
+ SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
+ SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
+ SDValue Chain = CLI.Chain;
+ SDValue Callee = CLI.Callee;
+ CallingConv::ID CallConv = CLI.CallConv;
+ bool &isTailCall = CLI.IsTailCall;
+ bool isVarArg = CLI.IsVarArg;
+
MachineFunction &MF = DAG.getMachineFunction();
bool Is64Bit = Subtarget->is64Bit();
bool IsWin64 = Subtarget->isTargetWin64();
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&MemOpChains[0], MemOpChains.size());
- // Build a sequence of copy-to-reg nodes chained together with token chain
- // and flag operands which copy the outgoing args into registers.
- SDValue InFlag;
- // Tail call byval lowering might overwrite argument registers so in case of
- // tail call optimization the copies to registers are lowered later.
- if (!isTailCall)
- for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
- Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
- RegsToPass[i].second, InFlag);
- InFlag = Chain.getValue(1);
- }
-
if (Subtarget->isPICStyleGOT()) {
// ELF / PIC requires GOT in the EBX register before function calls via PLT
// GOT pointer.
if (!isTailCall) {
- Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
- DAG.getNode(X86ISD::GlobalBaseReg,
- DebugLoc(), getPointerTy()),
- InFlag);
- InFlag = Chain.getValue(1);
+ RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
+ DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy())));
} else {
// If we are tail calling and generating PIC/GOT style code load the
// address of the callee into ECX. The value in ecx is used as target of
assert((Subtarget->hasSSE1() || !NumXMMRegs)
&& "SSE registers cannot be used when SSE is disabled");
- Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
- DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
- InFlag = Chain.getValue(1);
+ RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
+ DAG.getConstant(NumXMMRegs, MVT::i8)));
}
-
// For tail calls lower the arguments to the 'real' stack slot.
if (isTailCall) {
// Force all the incoming stack arguments to be loaded from the stack
SmallVector<SDValue, 8> MemOpChains2;
SDValue FIN;
int FI = 0;
- // Do not flag preceding copytoreg stuff together with the following stuff.
- InFlag = SDValue();
if (getTargetMachine().Options.GuaranteedTailCallOpt) {
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&MemOpChains2[0], MemOpChains2.size());
- // Copy arguments to their registers.
- for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
- Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
- RegsToPass[i].second, InFlag);
- InFlag = Chain.getValue(1);
- }
- InFlag =SDValue();
-
// Store the return address to the appropriate stack slot.
Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
FPDiff, dl);
}
+ // Build a sequence of copy-to-reg nodes chained together with token chain
+ // and flag operands which copy the outgoing args into registers.
+ SDValue InFlag;
+ for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
+ Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
+ RegsToPass[i].second, InFlag);
+ InFlag = Chain.getValue(1);
+ }
+
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
Ops.push_back(DAG.getRegister(RegsToPass[i].first,
RegsToPass[i].second.getValueType()));
- // Add an implicit use GOT pointer in EBX.
- if (!isTailCall && Subtarget->isPICStyleGOT())
- Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
-
- // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
- if (Is64Bit && isVarArg && !IsWin64)
- Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
-
// Add a register mask operand representing the call-preserved registers.
const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
return false;
}
-/// isSequentialOrUndefInRange - Return true if every element in Mask, begining
+/// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
/// from position Pos and ending in Pos+Size, falls within the specified
/// sequential range (L, L+Pos]. or is undef.
static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
return true;
}
+//
+// Some special combinations that can be optimized.
+//
+static
+SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
+ SelectionDAG &DAG) {
+ EVT VT = SVOp->getValueType(0);
+ DebugLoc dl = SVOp->getDebugLoc();
+
+ if (VT != MVT::v8i32 && VT != MVT::v8f32)
+ return SDValue();
+
+ ArrayRef<int> Mask = SVOp->getMask();
+
+ // These are the special masks that may be optimized.
+ static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
+ static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
+ bool MatchEvenMask = true;
+ bool MatchOddMask = true;
+ for (int i=0; i<8; ++i) {
+ if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
+ MatchEvenMask = false;
+ if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
+ MatchOddMask = false;
+ }
+ static const int CompactionMaskEven[] = {0, 2, -1, -1, 4, 6, -1, -1};
+ static const int CompactionMaskOdd [] = {1, 3, -1, -1, 5, 7, -1, -1};
+
+ const int *CompactionMask;
+ if (MatchEvenMask)
+ CompactionMask = CompactionMaskEven;
+ else if (MatchOddMask)
+ CompactionMask = CompactionMaskOdd;
+ else
+ return SDValue();
+
+ SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
+
+ SDValue Op0 = DAG.getVectorShuffle(VT, dl, SVOp->getOperand(0),
+ UndefNode, CompactionMask);
+ SDValue Op1 = DAG.getVectorShuffle(VT, dl, SVOp->getOperand(1),
+ UndefNode, CompactionMask);
+ static const int UnpackMask[] = {0, 8, 1, 9, 4, 12, 5, 13};
+ return DAG.getVectorShuffle(VT, dl, Op0, Op1, UnpackMask);
+}
+
/// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to UNPCKL.
static bool isUNPCKLMask(ArrayRef<int> Mask, EVT VT,
SmallVector<int, 8> MaskVec;
for (unsigned i = 0; i != NumElems; ++i) {
- int idx = SVOp->getMaskElt(i);
- if (idx < 0)
- MaskVec.push_back(idx);
- else if (idx < (int)NumElems)
- MaskVec.push_back(idx + NumElems);
- else
- MaskVec.push_back(idx - NumElems);
+ int Idx = SVOp->getMaskElt(i);
+ if (Idx >= 0) {
+ if (Idx < (int)NumElems)
+ Idx += NumElems;
+ else
+ Idx -= NumElems;
+ }
+ MaskVec.push_back(Idx);
}
return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
SVOp->getOperand(0), &MaskVec[0]);
SDValue Sc = Op.getOperand(0);
if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
- Sc.getOpcode() != ISD::BUILD_VECTOR)
- return SDValue();
+ Sc.getOpcode() != ISD::BUILD_VECTOR) {
+
+ if (!Subtarget->hasAVX2())
+ return SDValue();
+
+ // Use the register form of the broadcast instruction available on AVX2.
+ if (VT.is256BitVector())
+ Sc = Extract128BitVector(Sc, 0, DAG, dl);
+ return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
+ }
Ld = Sc.getOperand(0);
ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
}
}
+ bool IsLoad = ISD::isNormalLoad(Ld.getNode());
+ unsigned ScalarSize = Ld.getValueType().getSizeInBits();
+
+ // Handle AVX2 in-register broadcasts.
+ if (!IsLoad && Subtarget->hasAVX2() &&
+ (ScalarSize == 32 || (Is256 && ScalarSize == 64)))
+ return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
+
// The scalar source must be a normal load.
- if (!ISD::isNormalLoad(Ld.getNode()))
+ if (!IsLoad)
return SDValue();
- unsigned ScalarSize = Ld.getValueType().getSizeInBits();
-
if (ScalarSize == 32 || (Is256 && ScalarSize == 64))
return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
bool TwoInputs = V1Used && V2Used;
for (unsigned i = 0; i != 8; ++i) {
int EltIdx = MaskVals[i] * 2;
- if (TwoInputs && (EltIdx >= 16)) {
- pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
- continue;
- }
- pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
+ int Idx0 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx;
+ int Idx1 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx+1;
+ pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
}
V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
pshufbMask.clear();
for (unsigned i = 0; i != 8; ++i) {
int EltIdx = MaskVals[i] * 2;
- if (EltIdx < 16) {
- pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
- continue;
- }
- pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
+ int Idx0 = (EltIdx < 16) ? 0x80 : EltIdx - 16;
+ int Idx1 = (EltIdx < 16) ? 0x80 : EltIdx - 15;
+ pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
}
V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
for (unsigned i = 0; i != 16; ++i) {
int EltIdx = MaskVals[i];
EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
- pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
}
V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
DAG.getNode(ISD::BUILD_VECTOR, dl,
/// which could not be matched by any known target speficic shuffle
static SDValue
LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
+
+ SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
+ if (NewOp.getNode())
+ return NewOp;
+
EVT VT = SVOp->getValueType(0);
unsigned NumElems = VT.getVectorNumElements();
DebugLoc dl = SVOp->getDebugLoc();
MVT EltVT = VT.getVectorElementType().getSimpleVT();
EVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
- SDValue Shufs[2];
+ SDValue Output[2];
SmallVector<int, 16> Mask;
for (unsigned l = 0; l < 2; ++l) {
// Build a shuffle mask for the output, discovering on the fly which
// input vectors to use as shuffle operands (recorded in InputUsed).
// If building a suitable shuffle vector proves too hard, then bail
- // out with useBuildVector set.
+ // out with UseBuildVector set.
+ bool UseBuildVector = false;
int InputUsed[2] = { -1, -1 }; // Not yet discovered.
unsigned LaneStart = l * NumLaneElems;
for (unsigned i = 0; i != NumLaneElems; ++i) {
}
if (OpNo >= array_lengthof(InputUsed)) {
- // More than two input vectors used! Give up.
- return SDValue();
+ // More than two input vectors used! Give up on trying to create a
+ // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
+ UseBuildVector = true;
+ break;
}
// Add the mask index for the new shuffle vector.
Mask.push_back(Idx + OpNo * NumLaneElems);
}
- if (InputUsed[0] < 0) {
+ if (UseBuildVector) {
+ SmallVector<SDValue, 16> SVOps;
+ for (unsigned i = 0; i != NumLaneElems; ++i) {
+ // The mask element. This indexes into the input.
+ int Idx = SVOp->getMaskElt(i+LaneStart);
+ if (Idx < 0) {
+ SVOps.push_back(DAG.getUNDEF(EltVT));
+ continue;
+ }
+
+ // The input vector this mask element indexes into.
+ int Input = Idx / NumElems;
+
+ // Turn the index into an offset from the start of the input vector.
+ Idx -= Input * NumElems;
+
+ // Extract the vector element by hand.
+ SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
+ SVOp->getOperand(Input),
+ DAG.getIntPtrConstant(Idx)));
+ }
+
+ // Construct the output using a BUILD_VECTOR.
+ Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, &SVOps[0],
+ SVOps.size());
+ } else if (InputUsed[0] < 0) {
// No input vectors were used! The result is undefined.
- Shufs[l] = DAG.getUNDEF(NVT);
+ Output[l] = DAG.getUNDEF(NVT);
} else {
SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
(InputUsed[0] % 2) * NumLaneElems,
Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
(InputUsed[1] % 2) * NumLaneElems, DAG, dl);
// At least one input vector was used. Create a new shuffle vector.
- Shufs[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
+ Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
}
Mask.clear();
}
// Concatenate the result back
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Shufs[0], Shufs[1]);
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
}
/// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
if (NumElems == 4)
- // If we don't care about the second element, procede to use movss.
+ // If we don't care about the second element, proceed to use movss.
if (SVOp->getMaskElt(1) != -1)
return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
}
static SDValue
GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
- unsigned char OperandFlags) {
+ unsigned char OperandFlags, bool LocalDynamic = false) {
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
DebugLoc dl = GA->getDebugLoc();
GA->getValueType(0),
GA->getOffset(),
OperandFlags);
+
+ X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
+ : X86ISD::TLSADDR;
+
if (InFlag) {
SDValue Ops[] = { Chain, TGA, *InFlag };
- Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
+ Chain = DAG.getNode(CallType, dl, NodeTys, Ops, 3);
} else {
SDValue Ops[] = { Chain, TGA };
- Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
+ Chain = DAG.getNode(CallType, dl, NodeTys, Ops, 2);
}
// TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
X86::RAX, X86II::MO_TLSGD);
}
+static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
+ SelectionDAG &DAG,
+ const EVT PtrVT,
+ bool is64Bit) {
+ DebugLoc dl = GA->getDebugLoc();
+
+ // Get the start address of the TLS block for this module.
+ X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
+ .getInfo<X86MachineFunctionInfo>();
+ MFI->incNumLocalDynamicTLSAccesses();
+
+ SDValue Base;
+ if (is64Bit) {
+ Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT, X86::RAX,
+ X86II::MO_TLSLD, /*LocalDynamic=*/true);
+ } else {
+ SDValue InFlag;
+ SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
+ DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), PtrVT), InFlag);
+ InFlag = Chain.getValue(1);
+ Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
+ X86II::MO_TLSLDM, /*LocalDynamic=*/true);
+ }
+
+ // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
+ // of Base.
+
+ // Build x@dtpoff.
+ unsigned char OperandFlags = X86II::MO_DTPOFF;
+ unsigned WrapperKind = X86ISD::Wrapper;
+ SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
+ GA->getValueType(0),
+ GA->getOffset(), OperandFlags);
+ SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
+
+ // Add x@dtpoff with the base.
+ return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
+}
+
// Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
const EVT PtrVT, TLSModel::Model model,
Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), PtrVT),
Offset);
- } else {
- Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
- MachinePointerInfo::getGOT(), false, false, false,
- 0);
}
+
+ Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
+ MachinePointerInfo::getGOT(), false, false, false,
+ 0);
}
// The address of the thread local variable is the add of the thread
const GlobalValue *GV = GA->getGlobal();
if (Subtarget->isTargetELF()) {
- // TODO: implement the "local dynamic" model
-
- // If GV is an alias then use the aliasee for determining
- // thread-localness.
- if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
- GV = GA->resolveAliasedGlobal(false);
-
TLSModel::Model model = getTargetMachine().getTLSModel(GV);
switch (model) {
case TLSModel::GeneralDynamic:
- case TLSModel::LocalDynamic: // not implemented
if (Subtarget->is64Bit())
return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
-
+ case TLSModel::LocalDynamic:
+ return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
+ Subtarget->is64Bit());
case TLSModel::InitialExec:
case TLSModel::LocalExec:
return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
// Otherwise use a regular EFLAGS-setting instruction.
switch (Op.getNode()->getOpcode()) {
default: llvm_unreachable("unexpected operator!");
- case ISD::SUB: Opcode = X86ISD::SUB; break;
+ case ISD::SUB:
+ // If the only use of SUB is EFLAGS, use CMP instead.
+ if (Op.hasOneUse())
+ Opcode = X86ISD::CMP;
+ else
+ 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;
return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
DAG.getConstant(0, Op.getValueType()));
+ if (Opcode == X86ISD::CMP) {
+ SDValue New = DAG.getNode(Opcode, dl, MVT::i32, Op.getOperand(0),
+ Op.getOperand(1));
+ // We can't replace usage of SUB with CMP.
+ // The SUB node will be removed later because there is no use of it.
+ return SDValue(New.getNode(), 0);
+ }
+
SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
SmallVector<SDValue, 4> Ops;
for (unsigned i = 0; i != NumOperands; ++i)
DAG.getConstant(X86CC, MVT::i8), Cond);
return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
}
- // XOP comparison intrinsics
- case Intrinsic::x86_xop_vpcomltb:
- case Intrinsic::x86_xop_vpcomltw:
- case Intrinsic::x86_xop_vpcomltd:
- case Intrinsic::x86_xop_vpcomltq:
- case Intrinsic::x86_xop_vpcomltub:
- case Intrinsic::x86_xop_vpcomltuw:
- case Intrinsic::x86_xop_vpcomltud:
- case Intrinsic::x86_xop_vpcomltuq:
- case Intrinsic::x86_xop_vpcomleb:
- case Intrinsic::x86_xop_vpcomlew:
- case Intrinsic::x86_xop_vpcomled:
- case Intrinsic::x86_xop_vpcomleq:
- case Intrinsic::x86_xop_vpcomleub:
- case Intrinsic::x86_xop_vpcomleuw:
- case Intrinsic::x86_xop_vpcomleud:
- case Intrinsic::x86_xop_vpcomleuq:
- case Intrinsic::x86_xop_vpcomgtb:
- case Intrinsic::x86_xop_vpcomgtw:
- case Intrinsic::x86_xop_vpcomgtd:
- case Intrinsic::x86_xop_vpcomgtq:
- case Intrinsic::x86_xop_vpcomgtub:
- case Intrinsic::x86_xop_vpcomgtuw:
- case Intrinsic::x86_xop_vpcomgtud:
- case Intrinsic::x86_xop_vpcomgtuq:
- case Intrinsic::x86_xop_vpcomgeb:
- case Intrinsic::x86_xop_vpcomgew:
- case Intrinsic::x86_xop_vpcomged:
- case Intrinsic::x86_xop_vpcomgeq:
- case Intrinsic::x86_xop_vpcomgeub:
- case Intrinsic::x86_xop_vpcomgeuw:
- case Intrinsic::x86_xop_vpcomgeud:
- case Intrinsic::x86_xop_vpcomgeuq:
- case Intrinsic::x86_xop_vpcomeqb:
- case Intrinsic::x86_xop_vpcomeqw:
- case Intrinsic::x86_xop_vpcomeqd:
- case Intrinsic::x86_xop_vpcomeqq:
- case Intrinsic::x86_xop_vpcomequb:
- case Intrinsic::x86_xop_vpcomequw:
- case Intrinsic::x86_xop_vpcomequd:
- case Intrinsic::x86_xop_vpcomequq:
- case Intrinsic::x86_xop_vpcomneb:
- case Intrinsic::x86_xop_vpcomnew:
- case Intrinsic::x86_xop_vpcomned:
- case Intrinsic::x86_xop_vpcomneq:
- case Intrinsic::x86_xop_vpcomneub:
- case Intrinsic::x86_xop_vpcomneuw:
- case Intrinsic::x86_xop_vpcomneud:
- case Intrinsic::x86_xop_vpcomneuq:
- case Intrinsic::x86_xop_vpcomfalseb:
- case Intrinsic::x86_xop_vpcomfalsew:
- case Intrinsic::x86_xop_vpcomfalsed:
- case Intrinsic::x86_xop_vpcomfalseq:
- case Intrinsic::x86_xop_vpcomfalseub:
- case Intrinsic::x86_xop_vpcomfalseuw:
- case Intrinsic::x86_xop_vpcomfalseud:
- case Intrinsic::x86_xop_vpcomfalseuq:
- case Intrinsic::x86_xop_vpcomtrueb:
- case Intrinsic::x86_xop_vpcomtruew:
- case Intrinsic::x86_xop_vpcomtrued:
- case Intrinsic::x86_xop_vpcomtrueq:
- case Intrinsic::x86_xop_vpcomtrueub:
- case Intrinsic::x86_xop_vpcomtrueuw:
- case Intrinsic::x86_xop_vpcomtrueud:
- case Intrinsic::x86_xop_vpcomtrueuq: {
- unsigned CC = 0;
- unsigned Opc = 0;
-
- switch (IntNo) {
- default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
- case Intrinsic::x86_xop_vpcomltb:
- case Intrinsic::x86_xop_vpcomltw:
- case Intrinsic::x86_xop_vpcomltd:
- case Intrinsic::x86_xop_vpcomltq:
- CC = 0;
- Opc = X86ISD::VPCOM;
- break;
- case Intrinsic::x86_xop_vpcomltub:
- case Intrinsic::x86_xop_vpcomltuw:
- case Intrinsic::x86_xop_vpcomltud:
- case Intrinsic::x86_xop_vpcomltuq:
- CC = 0;
- Opc = X86ISD::VPCOMU;
- break;
- case Intrinsic::x86_xop_vpcomleb:
- case Intrinsic::x86_xop_vpcomlew:
- case Intrinsic::x86_xop_vpcomled:
- case Intrinsic::x86_xop_vpcomleq:
- CC = 1;
- Opc = X86ISD::VPCOM;
- break;
- case Intrinsic::x86_xop_vpcomleub:
- case Intrinsic::x86_xop_vpcomleuw:
- case Intrinsic::x86_xop_vpcomleud:
- case Intrinsic::x86_xop_vpcomleuq:
- CC = 1;
- Opc = X86ISD::VPCOMU;
- break;
- case Intrinsic::x86_xop_vpcomgtb:
- case Intrinsic::x86_xop_vpcomgtw:
- case Intrinsic::x86_xop_vpcomgtd:
- case Intrinsic::x86_xop_vpcomgtq:
- CC = 2;
- Opc = X86ISD::VPCOM;
- break;
- case Intrinsic::x86_xop_vpcomgtub:
- case Intrinsic::x86_xop_vpcomgtuw:
- case Intrinsic::x86_xop_vpcomgtud:
- case Intrinsic::x86_xop_vpcomgtuq:
- CC = 2;
- Opc = X86ISD::VPCOMU;
- break;
- case Intrinsic::x86_xop_vpcomgeb:
- case Intrinsic::x86_xop_vpcomgew:
- case Intrinsic::x86_xop_vpcomged:
- case Intrinsic::x86_xop_vpcomgeq:
- CC = 3;
- Opc = X86ISD::VPCOM;
- break;
- case Intrinsic::x86_xop_vpcomgeub:
- case Intrinsic::x86_xop_vpcomgeuw:
- case Intrinsic::x86_xop_vpcomgeud:
- case Intrinsic::x86_xop_vpcomgeuq:
- CC = 3;
- Opc = X86ISD::VPCOMU;
- break;
- case Intrinsic::x86_xop_vpcomeqb:
- case Intrinsic::x86_xop_vpcomeqw:
- case Intrinsic::x86_xop_vpcomeqd:
- case Intrinsic::x86_xop_vpcomeqq:
- CC = 4;
- Opc = X86ISD::VPCOM;
- break;
- case Intrinsic::x86_xop_vpcomequb:
- case Intrinsic::x86_xop_vpcomequw:
- case Intrinsic::x86_xop_vpcomequd:
- case Intrinsic::x86_xop_vpcomequq:
- CC = 4;
- Opc = X86ISD::VPCOMU;
- break;
- case Intrinsic::x86_xop_vpcomneb:
- case Intrinsic::x86_xop_vpcomnew:
- case Intrinsic::x86_xop_vpcomned:
- case Intrinsic::x86_xop_vpcomneq:
- CC = 5;
- Opc = X86ISD::VPCOM;
- break;
- case Intrinsic::x86_xop_vpcomneub:
- case Intrinsic::x86_xop_vpcomneuw:
- case Intrinsic::x86_xop_vpcomneud:
- case Intrinsic::x86_xop_vpcomneuq:
- CC = 5;
- Opc = X86ISD::VPCOMU;
- break;
- case Intrinsic::x86_xop_vpcomfalseb:
- case Intrinsic::x86_xop_vpcomfalsew:
- case Intrinsic::x86_xop_vpcomfalsed:
- case Intrinsic::x86_xop_vpcomfalseq:
- CC = 6;
- Opc = X86ISD::VPCOM;
- break;
- case Intrinsic::x86_xop_vpcomfalseub:
- case Intrinsic::x86_xop_vpcomfalseuw:
- case Intrinsic::x86_xop_vpcomfalseud:
- case Intrinsic::x86_xop_vpcomfalseuq:
- CC = 6;
- Opc = X86ISD::VPCOMU;
- break;
- case Intrinsic::x86_xop_vpcomtrueb:
- case Intrinsic::x86_xop_vpcomtruew:
- case Intrinsic::x86_xop_vpcomtrued:
- case Intrinsic::x86_xop_vpcomtrueq:
- CC = 7;
- Opc = X86ISD::VPCOM;
- break;
- case Intrinsic::x86_xop_vpcomtrueub:
- case Intrinsic::x86_xop_vpcomtrueuw:
- case Intrinsic::x86_xop_vpcomtrueud:
- case Intrinsic::x86_xop_vpcomtrueuq:
- CC = 7;
- Opc = X86ISD::VPCOMU;
- break;
- }
-
- SDValue LHS = Op.getOperand(1);
- SDValue RHS = Op.getOperand(2);
- return DAG.getNode(Opc, dl, Op.getValueType(), LHS, RHS,
- DAG.getConstant(CC, MVT::i8));
- }
-
// Arithmetic intrinsics.
case Intrinsic::x86_sse2_pmulu_dq:
case Intrinsic::x86_avx2_pmulu_dq:
}
}
+SDValue
+X86TargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const {
+ DebugLoc dl = Op.getDebugLoc();
+ unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
+ switch (IntNo) {
+ default: return SDValue(); // Don't custom lower most intrinsics.
+
+ // RDRAND intrinsics.
+ case Intrinsic::x86_rdrand_16:
+ case Intrinsic::x86_rdrand_32:
+ case Intrinsic::x86_rdrand_64: {
+ // Emit the node with the right value type.
+ SDValue Result = DAG.getNode(X86ISD::RDRAND, dl,
+ DAG.getVTList(Op->getValueType(0), MVT::Glue));
+
+ // If the value returned by RDRAND was valid (CF=1), return 1. Otherwise
+ // return the value from Rand, which is always 0, casted to i32.
+ SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
+ DAG.getConstant(1, Op->getValueType(1)),
+ DAG.getConstant(X86::COND_B, MVT::i32),
+ SDValue(Result.getNode(), 1) };
+ SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
+ DAG.getVTList(Op->getValueType(1), MVT::Glue),
+ Ops, 4);
+
+ // Return { result, isValid, chain }.
+ return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
+ Op.getOperand(0));
+ }
+ }
+}
+
SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
SelectionDAG &DAG) const {
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
}
SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
- MachineFunction &MF = DAG.getMachineFunction();
SDValue Chain = Op.getOperand(0);
SDValue Offset = Op.getOperand(1);
SDValue Handler = Op.getOperand(2);
Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
false, false, 0);
Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
- MF.getRegInfo().addLiveOut(StoreAddrReg);
return DAG.getNode(X86ISD::EH_RETURN, dl,
MVT::Other,
case ISD::VAARG: return LowerVAARG(Op, DAG);
case ISD::VACOPY: return LowerVACOPY(Op, DAG);
case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
+ case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
case ISD::FRAME_TO_ARGS_OFFSET:
case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
case X86ISD::FRCP: return "X86ISD::FRCP";
case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
+ case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
case X86ISD::SAHF: return "X86ISD::SAHF";
+ case X86ISD::RDRAND: return "X86ISD::RDRAND";
}
}
BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
.addReg(sizeVReg);
BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
- .addExternalSymbol("__morestack_allocate_stack_space").addReg(X86::RDI)
+ .addExternalSymbol("__morestack_allocate_stack_space")
.addRegMask(RegMask)
+ .addReg(X86::RDI, RegState::Implicit)
.addReg(X86::RAX, RegState::ImplicitDefine);
} else {
BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
-
-
DebugLoc DL = N->getDebugLoc();
SDValue Cond = N->getOperand(0);
// Get the LHS/RHS of the select.
// to simplify previous instructions.
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
- !DCI.isBeforeLegalize() &&
- TLI.isOperationLegal(ISD::VSELECT, VT)) {
+ !DCI.isBeforeLegalize() && TLI.isOperationLegal(ISD::VSELECT, VT)) {
unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
+
+ // Don't optimize vector selects that map to mask-registers.
+ if (BitWidth == 1)
+ return SDValue();
+
assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
return SDValue();
}
+// Generate NEG and CMOV for integer abs.
+static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
+ EVT VT = N->getValueType(0);
+
+ // Since X86 does not have CMOV for 8-bit integer, we don't convert
+ // 8-bit integer abs to NEG and CMOV.
+ if (VT.isInteger() && VT.getSizeInBits() == 8)
+ return SDValue();
+
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ DebugLoc DL = N->getDebugLoc();
+
+ // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
+ // and change it to SUB and CMOV.
+ if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
+ N0.getOpcode() == ISD::ADD &&
+ N0.getOperand(1) == N1 &&
+ N1.getOpcode() == ISD::SRA &&
+ N1.getOperand(0) == N0.getOperand(0))
+ if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
+ if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
+ // Generate SUB & CMOV.
+ SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
+ DAG.getConstant(0, VT), N0.getOperand(0));
+
+ SDValue Ops[] = { N0.getOperand(0), Neg,
+ DAG.getConstant(X86::COND_GE, MVT::i8),
+ SDValue(Neg.getNode(), 1) };
+ return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue),
+ Ops, array_lengthof(Ops));
+ }
+ return SDValue();
+}
+
// PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
if (DCI.isBeforeLegalizeOps())
return SDValue();
+ if (Subtarget->hasCMov()) {
+ SDValue RV = performIntegerAbsCombine(N, DAG);
+ if (RV.getNode())
+ return RV;
+ }
+
+ // Try forming BMI if it is available.
+ if (!Subtarget->hasBMI())
+ return SDValue();
+
EVT VT = N->getValueType(0);
if (VT != MVT::i32 && VT != MVT::i64)
/// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
- const X86Subtarget *Subtarget) {
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
LoadSDNode *Ld = cast<LoadSDNode>(N);
EVT RegVT = Ld->getValueType(0);
EVT MemVT = Ld->getMemoryVT();
unsigned RegSz = RegVT.getSizeInBits();
unsigned MemSz = MemVT.getSizeInBits();
assert(RegSz > MemSz && "Register size must be greater than the mem size");
- // All sizes must be a power of two
- if (!isPowerOf2_32(RegSz * MemSz * NumElems)) return SDValue();
- // Attempt to load the original value using a single load op.
- // Find a scalar type which is equal to the loaded word size.
+ // All sizes must be a power of two.
+ if (!isPowerOf2_32(RegSz * MemSz * NumElems))
+ return SDValue();
+
+ // Attempt to load the original value using scalar loads.
+ // Find the largest scalar type that divides the total loaded size.
MVT SclrLoadTy = MVT::i8;
for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
MVT Tp = (MVT::SimpleValueType)tp;
- if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() == MemSz) {
+ if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
SclrLoadTy = Tp;
- break;
}
}
- // Proceed if a load word is found.
- if (SclrLoadTy.getSizeInBits() != MemSz) return SDValue();
+ // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
+ if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
+ (64 <= MemSz))
+ SclrLoadTy = MVT::f64;
+
+ // Calculate the number of scalar loads that we need to perform
+ // in order to load our vector from memory.
+ unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
+ // Represent our vector as a sequence of elements which are the
+ // largest scalar that we can load.
EVT LoadUnitVecVT = EVT::getVectorVT(*DAG.getContext(), SclrLoadTy,
RegSz/SclrLoadTy.getSizeInBits());
+ // Represent the data using the same element type that is stored in
+ // memory. In practice, we ''widen'' MemVT.
EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
RegSz/MemVT.getScalarType().getSizeInBits());
- // Can't shuffle using an illegal type.
- if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
- // Perform a single load.
- SDValue ScalarLoad = DAG.getLoad(SclrLoadTy, dl, Ld->getChain(),
- Ld->getBasePtr(),
- Ld->getPointerInfo(), Ld->isVolatile(),
- Ld->isNonTemporal(), Ld->isInvariant(),
- Ld->getAlignment());
+ assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
+ "Invalid vector type");
- // Insert the word loaded into a vector.
- SDValue ScalarInVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
- LoadUnitVecVT, ScalarLoad);
+ // We can't shuffle using an illegal type.
+ if (!TLI.isTypeLegal(WideVecVT))
+ return SDValue();
+
+ SmallVector<SDValue, 8> Chains;
+ SDValue Ptr = Ld->getBasePtr();
+ SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits()/8,
+ TLI.getPointerTy());
+ SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
+
+ for (unsigned i = 0; i < NumLoads; ++i) {
+ // Perform a single load.
+ SDValue ScalarLoad = DAG.getLoad(SclrLoadTy, dl, Ld->getChain(),
+ Ptr, Ld->getPointerInfo(),
+ Ld->isVolatile(), Ld->isNonTemporal(),
+ Ld->isInvariant(), Ld->getAlignment());
+ Chains.push_back(ScalarLoad.getValue(1));
+ // Create the first element type using SCALAR_TO_VECTOR in order to avoid
+ // another round of DAGCombining.
+ if (i == 0)
+ Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
+ else
+ Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
+ ScalarLoad, DAG.getIntPtrConstant(i));
+
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
+ }
+
+ SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
+ Chains.size());
// Bitcast the loaded value to a vector of the original element type, in
// the size of the target vector type.
- SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT,
- ScalarInVector);
+ SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
unsigned SizeRatio = RegSz/MemSz;
// Redistribute the loaded elements into the different locations.
Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
// Replace the original load with the new sequence
// and return the new chain.
- DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Shuff);
- return SDValue(ScalarLoad.getNode(), 1);
+ return DCI.CombineTo(N, Shuff, TF, true);
}
return SDValue();
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
// If we are saving a concatenation of two XMM registers, perform two stores.
- // This is better in Sandy Bridge cause one 256-bit mem op is done via two
- // 128-bit ones. If in the future the cost becomes only one memory access the
- // first version would be better.
- if (VT.getSizeInBits() == 256 &&
+ // On Sandy Bridge, 256-bit memory operations are executed by two
+ // 128-bit ports. However, on Haswell it is better to issue a single 256-bit
+ // memory operation.
+ if (VT.getSizeInBits() == 256 && !Subtarget->hasAVX2() &&
StoredVal.getNode()->getOpcode() == ISD::CONCAT_VECTORS &&
StoredVal.getNumOperands() == 2) {
-
SDValue Value0 = StoredVal.getOperand(0);
SDValue Value1 = StoredVal.getOperand(1);
for (unsigned i = 0; i != NumElems; ++i)
ShuffleVec[i] = i * SizeRatio;
- // Can't shuffle using an illegal type
- if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
+ // Can't shuffle using an illegal type.
+ if (!TLI.isTypeLegal(WideVecVT))
+ return SDValue();
SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
DAG.getUNDEF(WideVecVT),
for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
MVT Tp = (MVT::SimpleValueType)tp;
- if (TLI.isTypeLegal(Tp) && StoreType.getSizeInBits() < NumElems * ToSz)
+ if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
StoreType = Tp;
}
+ // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
+ if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
+ (64 <= NumElems * ToSz))
+ StoreType = MVT::f64;
+
// Bitcast the original vector into a vector of store-size units
EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
- StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
+ StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
SmallVector<SDValue, 8> Chains;
case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
- case ISD::LOAD: return PerformLOADCombine(N, DAG, Subtarget);
+ case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG);
case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
// wrong class. This can happen with constraints like {xmm0} where the
// target independent register mapper will just pick the first match it can
// find, ignoring the required type.
- if (VT == MVT::f32)
+
+ if (VT == MVT::f32 || VT == MVT::i32)
Res.second = &X86::FR32RegClass;
- else if (VT == MVT::f64)
+ else if (VT == MVT::f64 || VT == MVT::i64)
Res.second = &X86::FR64RegClass;
else if (X86::VR128RegClass.hasType(VT))
Res.second = &X86::VR128RegClass;
+ else if (X86::VR256RegClass.hasType(VT))
+ Res.second = &X86::VR256RegClass;
}
return Res;