return new X86LinuxTargetObjectFile();
if (Subtarget->isTargetELF())
return new TargetLoweringObjectFileELF();
+ if (Subtarget->isTargetWindows())
+ return new X86WindowsTargetObjectFile();
if (Subtarget->isTargetCOFF())
return new TargetLoweringObjectFileCOFF();
llvm_unreachable("unknown subtarget type");
}
bool
-X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
+X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
+ unsigned,
+ bool *Fast) const {
if (Fast)
*Fast = Subtarget->isUnalignedMemAccessFast();
return true;
X86Info->setTCReturnAddrDelta(FPDiff);
}
+ unsigned NumBytesToPush = NumBytes;
+ unsigned NumBytesToPop = NumBytes;
+
+ // If we have an inalloca argument, all stack space has already been allocated
+ // for us and be right at the top of the stack. We don't support multiple
+ // arguments passed in memory when using inalloca.
+ if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
+ NumBytesToPush = 0;
+ assert(ArgLocs.back().getLocMemOffset() == 0 &&
+ "an inalloca argument must be the only memory argument");
+ }
+
if (!IsSibcall)
- Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
- dl);
+ Chain = DAG.getCALLSEQ_START(
+ Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
SDValue RetAddrFrIdx;
// Load return address for tail calls.
const X86RegisterInfo *RegInfo =
static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
+ // Skip inalloca arguments, they have already been written.
+ ISD::ArgFlagsTy Flags = Outs[i].Flags;
+ if (Flags.isInAlloca())
+ continue;
+
CCValAssign &VA = ArgLocs[i];
EVT RegVT = VA.getLocVT();
SDValue Arg = OutVals[i];
- ISD::ArgFlagsTy Flags = Outs[i].Flags;
bool isByVal = Flags.isByVal();
// Promote the value if needed.
SmallVector<SDValue, 8> Ops;
if (!IsSibcall && isTailCall) {
- Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
- DAG.getIntPtrConstant(0, true), InFlag, dl);
+ Chain = DAG.getCALLSEQ_END(Chain,
+ DAG.getIntPtrConstant(NumBytesToPop, true),
+ DAG.getIntPtrConstant(0, true), InFlag, dl);
InFlag = Chain.getValue(1);
}
InFlag = Chain.getValue(1);
// Create the CALLSEQ_END node.
- unsigned NumBytesForCalleeToPush;
+ unsigned NumBytesForCalleeToPop;
if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
getTargetMachine().Options.GuaranteedTailCallOpt))
- NumBytesForCalleeToPush = NumBytes; // Callee pops everything
+ NumBytesForCalleeToPop = NumBytes; // Callee pops everything
else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
!Subtarget->getTargetTriple().isOSMSVCRT() &&
SR == StackStructReturn)
// pops the hidden struct pointer, so we have to push it back.
// This is common for Darwin/X86, Linux & Mingw32 targets.
// For MSVC Win32 targets, the caller pops the hidden struct pointer.
- NumBytesForCalleeToPush = 4;
+ NumBytesForCalleeToPop = 4;
else
- NumBytesForCalleeToPush = 0; // Callee pops nothing.
+ NumBytesForCalleeToPop = 0; // Callee pops nothing.
// Returns a flag for retval copy to use.
if (!IsSibcall) {
Chain = DAG.getCALLSEQ_END(Chain,
- DAG.getIntPtrConstant(NumBytes, true),
- DAG.getIntPtrConstant(NumBytesForCalleeToPush,
+ DAG.getIntPtrConstant(NumBytesToPop, true),
+ DAG.getIntPtrConstant(NumBytesForCalleeToPop,
true),
InFlag, dl);
InFlag = Chain.getValue(1);
return false;
}
+/// \brief Returns true if it is beneficial to convert a load of a constant
+/// to just the constant itself.
+bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
+ Type *Ty) const {
+ assert(Ty->isIntegerTy());
+
+ unsigned BitSize = Ty->getPrimitiveSizeInBits();
+ if (BitSize == 0 || BitSize > 64)
+ return false;
+ return true;
+}
+
/// isUndefOrInRange - Return true if Val is undef or if its value falls within
/// the specified range (L, H].
static bool isUndefOrInRange(int Val, int Low, int Hi) {
SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops, 16);
+ } else if (VT.getScalarType() == MVT::i1) {
+ assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
+ SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
+ SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
+ Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
+ Ops, VT.getVectorNumElements());
} else
llvm_unreachable("Unexpected vector type");
// For AVX-length vectors, build the individual 128-bit pieces and use
// shuffles to put them in place.
- if (VT.is256BitVector()) {
- SmallVector<SDValue, 32> V;
+ if (VT.is256BitVector() || VT.is512BitVector()) {
+ SmallVector<SDValue, 64> V;
for (unsigned i = 0; i != NumElems; ++i)
V.push_back(Op.getOperand(i));
NumElems/2);
// Recreate the wider vector with the lower and upper part.
- return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
+ if (VT.is256BitVector())
+ return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
+ return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
}
// Let legalizer expand 2-wide build_vectors.
// Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
return DAG.getNode(X86ISD::VPERMV, dl, VT,
DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
- return DAG.getNode(X86ISD::VPERMV3, dl, VT,
- DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1, V2);
+ return DAG.getNode(X86ISD::VPERMV3, dl, VT, V1,
+ DAG.getNode(ISD::BITCAST, dl, VT, Mask), V2);
}
//===--------------------------------------------------------------------===//
}
unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
- if (IdxVal) {
- unsigned MaxSift = VecVT.getSizeInBits() - 1;
- Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
- DAG.getConstant(MaxSift - IdxVal, MVT::i8));
- Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
- DAG.getConstant(MaxSift, MVT::i8));
- }
- return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i1, Vec,
+ unsigned MaxSift = VecVT.getSizeInBits() - 1;
+ Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
+ DAG.getConstant(MaxSift - IdxVal, MVT::i8));
+ Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
+ DAG.getConstant(MaxSift, MVT::i8));
+ return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
DAG.getIntPtrConstant(0));
}
return SDValue();
if (Subtarget->hasInt256())
- return DAG.getNode(X86ISD::VZEXT_MOVL, dl, VT, In);
+ return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
SDValue Undef = DAG.getUNDEF(InVT);
In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
InVT = ExtVT;
}
+
SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
SDValue CP = DAG.getConstantPool(C, getPointerTy());
SelectionDAG &DAG) const {
SDLoc dl(Op);
+ if (Op.getValueType() == MVT::i1)
+ // KORTEST instruction should be selected
+ return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
+ DAG.getConstant(0, Op.getValueType()));
+
// CF and OF aren't always set the way we want. Determine which
// of these we need.
bool NeedCF = false;
NeedOF = true;
break;
}
-
// See if we can use the EFLAGS value from the operand instead of
// doing a separate TEST. TEST always sets OF and CF to 0, so unless
// we prove that the arithmetic won't overflow, we can't use OF or CF.
if (Op.getResNo() != 0 || NeedOF || NeedCF) {
// Emit a CMP with 0, which is the TEST pattern.
- if (Op.getValueType() == MVT::i1)
- return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
- DAG.getConstant(0, MVT::i1));
+ //if (Op.getValueType() == MVT::i1)
+ // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
+ // DAG.getConstant(0, MVT::i1));
return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
DAG.getConstant(0, Op.getValueType()));
}
return EmitTest(Op0, X86CC, DAG);
if (Op0.getValueType() == MVT::i1) {
+ // invert the value
Op0 = DAG.getNode(ISD::XOR, dl, MVT::i1, Op0,
DAG.getConstant(-1, MVT::i1));
- return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op0,
- DAG.getConstant(0, MVT::i1));
+ return EmitTest(Op0, X86CC, DAG);
}
}
DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
}
-static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
+static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
SDValue CC = Op.getOperand(2);
MVT VT = Op.getSimpleValueType();
+ SDLoc dl(Op);
assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 32 &&
Op.getValueType().getScalarType() == MVT::i1 &&
"Cannot set masked compare for this operation");
ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
- SDLoc dl(Op);
-
+ unsigned Opc = 0;
bool Unsigned = false;
+ bool Swap = false;
unsigned SSECC;
switch (SetCCOpcode) {
default: llvm_unreachable("Unexpected SETCC condition");
case ISD::SETNE: SSECC = 4; break;
- case ISD::SETEQ: SSECC = 0; break;
- case ISD::SETUGT: Unsigned = true;
- case ISD::SETGT: SSECC = 6; break; // NLE
- case ISD::SETULT: Unsigned = true;
- case ISD::SETLT: SSECC = 1; break;
- case ISD::SETUGE: Unsigned = true;
- case ISD::SETGE: SSECC = 5; break; // NLT
- case ISD::SETULE: Unsigned = true;
+ case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
+ case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
+ case ISD::SETLT: Swap = true; //fall-through
+ case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
+ case ISD::SETULT: SSECC = 1; Unsigned = true; break;
+ case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
+ case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
+ case ISD::SETULE: Unsigned = true; //fall-through
case ISD::SETLE: SSECC = 2; break;
}
- unsigned Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
+
+ if (Swap)
+ std::swap(Op0, Op1);
+ if (Opc)
+ return DAG.getNode(Opc, dl, VT, Op0, Op1);
+ Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
return DAG.getNode(Opc, dl, VT, Op0, Op1,
DAG.getConstant(SSECC, MVT::i8));
-
}
static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
if (Subtarget->hasAVX512()) {
if (Op1.getValueType().is512BitVector() ||
(MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
- return LowerIntVSETCC_AVX512(Op, DAG);
+ return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
// In AVX-512 architecture setcc returns mask with i1 elements,
// But there is no compare instruction for i8 and i16 elements.
switch (SetCCOpcode) {
default: llvm_unreachable("Unexpected SETCC condition");
case ISD::SETNE: Invert = true;
- case ISD::SETEQ: Opc = MaskResult? X86ISD::PCMPEQM: X86ISD::PCMPEQ; break;
+ case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
case ISD::SETLT: Swap = true;
- case ISD::SETGT: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT; break;
+ case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
case ISD::SETGE: Swap = true;
- case ISD::SETLE: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT;
+ case ISD::SETLE: Opc = X86ISD::PCMPGT;
Invert = true; break;
case ISD::SETULT: Swap = true;
- case ISD::SETUGT: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT;
+ case ISD::SETUGT: Opc = X86ISD::PCMPGT;
FlipSigns = true; break;
case ISD::SETUGE: Swap = true;
- case ISD::SETULE: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT;
+ case ISD::SETULE: Opc = X86ISD::PCMPGT;
FlipSigns = true; Invert = true; break;
}
X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
bool Invert = (CC == ISD::SETNE) ^
cast<ConstantSDNode>(Op1)->isNullValue();
- if (!Invert) return Op0;
+ if (!Invert)
+ return Op0;
CCode = X86::GetOppositeBranchCondition(CCode);
SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
+ DAG.getConstant(CCode, MVT::i8),
+ Op0.getOperand(1));
if (VT == MVT::i1)
return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
return SetCC;
SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), EFLAGS);
+ DAG.getConstant(X86CC, MVT::i8), EFLAGS);
if (VT == MVT::i1)
return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
return SetCC;
return SDValue();
if (Subtarget->hasInt256())
- return DAG.getNode(X86ISD::VSEXT_MOVL, dl, VT, In);
+ return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
// Optimize vectors in AVX mode
// Sign extend v8i16 to v8i32 and
MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
VT.getVectorNumElements()/2);
- OpLo = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpLo);
- OpHi = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpHi);
+ OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
+ OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
}
unsigned X86Opcode;
unsigned X86Cond;
SDVTList VTs;
+ // Keep this in sync with LowerXALUO, otherwise we might create redundant
+ // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
+ // X86ISD::INC).
switch (CondOpcode) {
case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
- case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
+ case ISD::SADDO:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
+ if (C->isOne()) {
+ X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
+ break;
+ }
+ X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
- case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
+ case ISD::SSUBO:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
+ if (C->isOne()) {
+ X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
+ break;
+ }
+ X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
default: llvm_unreachable("unexpected overflowing operator");
case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
- case X86ISD::VSEXT_MOVL: return "X86ISD::VSEXT_MOVL";
case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
case X86ISD::VZEXT: return "X86ISD::VZEXT";
case X86ISD::VSEXT: return "X86ISD::VSEXT";
case X86ISD::OR: return "X86ISD::OR";
case X86ISD::XOR: return "X86ISD::XOR";
case X86ISD::AND: return "X86ISD::AND";
- case X86ISD::BLSI: return "X86ISD::BLSI";
- case X86ISD::BLSMSK: return "X86ISD::BLSMSK";
- case X86ISD::BLSR: return "X86ISD::BLSR";
case X86ISD::BZHI: return "X86ISD::BZHI";
case X86ISD::BEXTR: return "X86ISD::BEXTR";
case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
case X86ISD::PTEST: return "X86ISD::PTEST";
case X86ISD::TESTP: return "X86ISD::TESTP";
case X86ISD::TESTM: return "X86ISD::TESTM";
+ case X86ISD::TESTNM: return "X86ISD::TESTNM";
case X86ISD::KORTEST: return "X86ISD::KORTEST";
case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
case X86ISD::VPERMV: return "X86ISD::VPERMV";
case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
+ case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
case X86ISD::VPERMI: return "X86ISD::VPERMI";
case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
return MBB;
}
+// Replace 213-type (isel default) FMA3 instructions with 231-type for
+// accumulator loops. Writing back to the accumulator allows the coalescer
+// to remove extra copies in the loop.
+MachineBasicBlock *
+X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
+ MachineBasicBlock *MBB) const {
+ MachineOperand &AddendOp = MI->getOperand(3);
+
+ // Bail out early if the addend isn't a register - we can't switch these.
+ if (!AddendOp.isReg())
+ return MBB;
+
+ MachineFunction &MF = *MBB->getParent();
+ MachineRegisterInfo &MRI = MF.getRegInfo();
+
+ // Check whether the addend is defined by a PHI:
+ assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
+ MachineInstr &AddendDef = *MRI.def_begin(AddendOp.getReg());
+ if (!AddendDef.isPHI())
+ return MBB;
+
+ // Look for the following pattern:
+ // loop:
+ // %addend = phi [%entry, 0], [%loop, %result]
+ // ...
+ // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
+
+ // Replace with:
+ // loop:
+ // %addend = phi [%entry, 0], [%loop, %result]
+ // ...
+ // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
+
+ for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
+ assert(AddendDef.getOperand(i).isReg());
+ MachineOperand PHISrcOp = AddendDef.getOperand(i);
+ MachineInstr &PHISrcInst = *MRI.def_begin(PHISrcOp.getReg());
+ if (&PHISrcInst == MI) {
+ // Found a matching instruction.
+ unsigned NewFMAOpc = 0;
+ switch (MI->getOpcode()) {
+ case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
+ case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
+ case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
+ case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
+ case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
+ case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
+ case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
+ case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
+ case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
+ case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
+ case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
+ case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
+ case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
+ case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
+ case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
+ case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
+ case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
+ case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
+ case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
+ case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
+ case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
+ case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
+ case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
+ case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
+ default: llvm_unreachable("Unrecognized FMA variant.");
+ }
+
+ const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
+ MachineInstrBuilder MIB =
+ BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
+ .addOperand(MI->getOperand(0))
+ .addOperand(MI->getOperand(3))
+ .addOperand(MI->getOperand(2))
+ .addOperand(MI->getOperand(1));
+ MBB->insert(MachineBasicBlock::iterator(MI), MIB);
+ MI->eraseFromParent();
+ }
+ }
+
+ return MBB;
+}
+
MachineBasicBlock *
X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
MachineBasicBlock *BB) const {
case TargetOpcode::STACKMAP:
case TargetOpcode::PATCHPOINT:
return emitPatchPoint(MI, BB);
+
+ case X86::VFMADDPDr213r:
+ case X86::VFMADDPSr213r:
+ case X86::VFMADDSDr213r:
+ case X86::VFMADDSSr213r:
+ case X86::VFMSUBPDr213r:
+ case X86::VFMSUBPSr213r:
+ case X86::VFMSUBSDr213r:
+ case X86::VFMSUBSSr213r:
+ case X86::VFNMADDPDr213r:
+ case X86::VFNMADDPSr213r:
+ case X86::VFNMADDSDr213r:
+ case X86::VFNMADDSSr213r:
+ case X86::VFNMSUBPDr213r:
+ case X86::VFNMSUBPSr213r:
+ case X86::VFNMSUBSDr213r:
+ case X86::VFNMSUBSSr213r:
+ case X86::VFMADDPDr213rY:
+ case X86::VFMADDPSr213rY:
+ case X86::VFMSUBPDr213rY:
+ case X86::VFMSUBPSr213rY:
+ case X86::VFNMADDPDr213rY:
+ case X86::VFNMADDPSr213rY:
+ case X86::VFNMSUBPDr213rY:
+ case X86::VFNMSUBPSr213rY:
+ return emitFMA3Instr(MI, BB);
}
}
}
}
+ // Try to fold this VSELECT into a MOVSS/MOVSD
+ if (N->getOpcode() == ISD::VSELECT &&
+ Cond.getOpcode() == ISD::BUILD_VECTOR && !DCI.isBeforeLegalize()) {
+ if (VT == MVT::v4i32 || VT == MVT::v4f32 ||
+ (Subtarget->hasSSE2() && (VT == MVT::v2i64 || VT == MVT::v2f64))) {
+ bool CanFold = false;
+ unsigned NumElems = Cond.getNumOperands();
+ SDValue A = LHS;
+ SDValue B = RHS;
+
+ if (isZero(Cond.getOperand(0))) {
+ CanFold = true;
+
+ // fold (vselect <0,-1,-1,-1>, A, B) -> (movss A, B)
+ // fold (vselect <0,-1> -> (movsd A, B)
+ for (unsigned i = 1, e = NumElems; i != e && CanFold; ++i)
+ CanFold = isAllOnes(Cond.getOperand(i));
+ } else if (isAllOnes(Cond.getOperand(0))) {
+ CanFold = true;
+ std::swap(A, B);
+
+ // fold (vselect <-1,0,0,0>, A, B) -> (movss B, A)
+ // fold (vselect <-1,0> -> (movsd B, A)
+ for (unsigned i = 1, e = NumElems; i != e && CanFold; ++i)
+ CanFold = isZero(Cond.getOperand(i));
+ }
+
+ if (CanFold) {
+ if (VT == MVT::v4i32 || VT == MVT::v4f32)
+ return getTargetShuffleNode(X86ISD::MOVSS, DL, VT, A, B, DAG);
+ return getTargetShuffleNode(X86ISD::MOVSD, DL, VT, A, B, DAG);
+ }
+
+ if (Subtarget->hasSSE2() && (VT == MVT::v4i32 || VT == MVT::v4f32)) {
+ // fold (v4i32: vselect <0,0,-1,-1>, A, B) ->
+ // (v4i32 (bitcast (movsd (v2i64 (bitcast A)),
+ // (v2i64 (bitcast B)))))
+ //
+ // fold (v4f32: vselect <0,0,-1,-1>, A, B) ->
+ // (v4f32 (bitcast (movsd (v2f64 (bitcast A)),
+ // (v2f64 (bitcast B)))))
+ //
+ // fold (v4i32: vselect <-1,-1,0,0>, A, B) ->
+ // (v4i32 (bitcast (movsd (v2i64 (bitcast B)),
+ // (v2i64 (bitcast A)))))
+ //
+ // fold (v4f32: vselect <-1,-1,0,0>, A, B) ->
+ // (v4f32 (bitcast (movsd (v2f64 (bitcast B)),
+ // (v2f64 (bitcast A)))))
+
+ CanFold = (isZero(Cond.getOperand(0)) &&
+ isZero(Cond.getOperand(1)) &&
+ isAllOnes(Cond.getOperand(2)) &&
+ isAllOnes(Cond.getOperand(3)));
+
+ if (!CanFold && isAllOnes(Cond.getOperand(0)) &&
+ isAllOnes(Cond.getOperand(1)) &&
+ isZero(Cond.getOperand(2)) &&
+ isZero(Cond.getOperand(3))) {
+ CanFold = true;
+ std::swap(LHS, RHS);
+ }
+
+ if (CanFold) {
+ EVT NVT = (VT == MVT::v4i32) ? MVT::v2i64 : MVT::v2f64;
+ SDValue NewA = DAG.getNode(ISD::BITCAST, DL, NVT, LHS);
+ SDValue NewB = DAG.getNode(ISD::BITCAST, DL, NVT, RHS);
+ SDValue Select = getTargetShuffleNode(X86ISD::MOVSD, DL, NVT, NewA,
+ NewB, DAG);
+ return DAG.getNode(ISD::BITCAST, DL, VT, Select);
+ }
+ }
+ }
+ }
+
// If we know that this node is legal then we know that it is going to be
// matched by one of the SSE/AVX BLEND instructions. These instructions only
// depend on the highest bit in each word. Try to use SimplifyDemandedBits
// See X86ATTInstPrinter.cpp:printSSECC().
unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
if (Subtarget->hasAVX512()) {
- SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00, CMP01,
- DAG.getConstant(x86cc, MVT::i8));
+ SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
+ CMP01, DAG.getConstant(x86cc, MVT::i8));
if (N->getValueType(0) != MVT::i1)
- return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0), FSetCC);
+ return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
+ FSetCC);
return FSetCC;
}
- SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL, CMP00.getValueType(), CMP00, CMP01,
+ SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
+ CMP00.getValueType(), CMP00, CMP01,
DAG.getConstant(x86cc, MVT::i8));
MVT IntVT = (is64BitFP ? MVT::i64 : MVT::i32);
SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT,
if (R.getNode())
return R;
- // Create BLSI, BLSR, and BZHI instructions
- // BLSI is X & (-X)
- // BLSR is X & (X-1)
+ // Create BEXTR and BZHI instructions
// BZHI is X & ((1 << Y) - 1)
// BEXTR is ((X >> imm) & (2**size-1))
if (VT == MVT::i32 || VT == MVT::i64) {
SDValue N1 = N->getOperand(1);
SDLoc DL(N);
- if (Subtarget->hasBMI()) {
- // Check LHS for neg
- if (N0.getOpcode() == ISD::SUB && N0.getOperand(1) == N1 &&
- isZero(N0.getOperand(0)))
- return DAG.getNode(X86ISD::BLSI, DL, VT, N1);
-
- // Check RHS for neg
- if (N1.getOpcode() == ISD::SUB && N1.getOperand(1) == N0 &&
- isZero(N1.getOperand(0)))
- return DAG.getNode(X86ISD::BLSI, DL, VT, N0);
-
- // Check LHS for X-1
- if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
- isAllOnes(N0.getOperand(1)))
- return DAG.getNode(X86ISD::BLSR, DL, VT, N1);
-
- // Check RHS for X-1
- if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
- isAllOnes(N1.getOperand(1)))
- return DAG.getNode(X86ISD::BLSR, DL, VT, N0);
- }
-
if (Subtarget->hasBMI2()) {
// Check for (and (add (shl 1, Y), -1), X)
if (N0.getOpcode() == ISD::ADD && isAllOnes(N0.getOperand(1))) {
static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
- EVT VT = N->getValueType(0);
if (DCI.isBeforeLegalizeOps())
return SDValue();
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
+ EVT VT = N->getValueType(0);
// look for psign/blend
if (VT == MVT::v2i64 || VT == MVT::v4i64) {
static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
- EVT VT = N->getValueType(0);
if (DCI.isBeforeLegalizeOps())
return SDValue();
return RV;
}
- // Try forming BMI if it is available.
- if (!Subtarget->hasBMI())
- return SDValue();
-
- if (VT != MVT::i32 && VT != MVT::i64)
- return SDValue();
-
- assert(Subtarget->hasBMI() && "Creating BLSMSK requires BMI instructions");
-
- // Create BLSMSK instructions by finding X ^ (X-1)
- SDValue N0 = N->getOperand(0);
- SDValue N1 = N->getOperand(1);
- SDLoc DL(N);
-
- if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
- isAllOnes(N0.getOperand(1)))
- return DAG.getNode(X86ISD::BLSMSK, DL, VT, N1);
-
- if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
- isAllOnes(N1.getOperand(1)))
- return DAG.getNode(X86ISD::BLSMSK, DL, VT, N0);
-
return SDValue();
}
// Optimize x == -y --> x+y == 0
// x != -y --> x+y != 0
-static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG) {
+static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget* Subtarget) {
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
+ EVT VT = N->getValueType(0);
+ SDLoc DL(N);
if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
return DAG.getSetCC(SDLoc(N), N->getValueType(0),
addV, DAG.getConstant(0, addV.getValueType()), CC);
}
+
+ if (VT.getScalarType() == MVT::i1) {
+ bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
+ (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
+ bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
+ if (!IsSEXT0 && !IsVZero0)
+ return SDValue();
+ bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
+ (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
+ bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
+
+ if (!IsSEXT1 && !IsVZero1)
+ return SDValue();
+
+ if (IsSEXT0 && IsVZero1) {
+ assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
+ if (CC == ISD::SETEQ)
+ return DAG.getNOT(DL, LHS.getOperand(0), VT);
+ return LHS.getOperand(0);
+ }
+ if (IsSEXT1 && IsVZero0) {
+ assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
+ if (CC == ISD::SETEQ)
+ return DAG.getNOT(DL, RHS.getOperand(0), VT);
+ return RHS.getOperand(0);
+ }
+ }
+
return SDValue();
}
case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
case ISD::SIGN_EXTEND_INREG: return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
- case ISD::SETCC: return PerformISDSETCCCombine(N, DAG);
+ case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);