setOperationAction(ISD::SETCC , MVT::f32 , Custom);
setOperationAction(ISD::SETCC , MVT::f64 , Custom);
setOperationAction(ISD::SETCC , MVT::f80 , Custom);
+ setOperationAction(ISD::SETCCE , MVT::i8 , Custom);
+ setOperationAction(ISD::SETCCE , MVT::i16 , Custom);
+ setOperationAction(ISD::SETCCE , MVT::i32 , Custom);
if (Subtarget->is64Bit()) {
setOperationAction(ISD::SELECT , MVT::i64 , Custom);
setOperationAction(ISD::SETCC , MVT::i64 , Custom);
+ setOperationAction(ISD::SETCCE , MVT::i64 , Custom);
}
setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
// NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
- setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Custom);
setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
setOperationAction(ISD::MUL, MVT::v32i16, Legal);
setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
- setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Legal);
- setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Legal);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
setTargetDAGCombine(ISD::ADD);
setTargetDAGCombine(ISD::FADD);
setTargetDAGCombine(ISD::FSUB);
+ setTargetDAGCombine(ISD::FNEG);
setTargetDAGCombine(ISD::FMA);
setTargetDAGCombine(ISD::SUB);
setTargetDAGCombine(ISD::LOAD);
setTargetDAGCombine(ISD::MLOAD);
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::MSTORE);
+ setTargetDAGCombine(ISD::TRUNCATE);
setTargetDAGCombine(ISD::ZERO_EXTEND);
setTargetDAGCombine(ISD::ANY_EXTEND);
setTargetDAGCombine(ISD::SIGN_EXTEND);
MFI->CreateFixedObject(1, StackSize, true));
}
- MachineModuleInfo &MMI = MF.getMMI();
-
// Figure out if XMM registers are in use.
assert(!(Subtarget->useSoftFloat() &&
Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
FuncInfo->setArgumentStackSize(StackSize);
- if (MMI.hasWinEHFuncInfo(Fn)) {
+ if (WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo()) {
EHPersonality Personality = classifyEHPersonality(Fn->getPersonalityFn());
if (Personality == EHPersonality::CoreCLR) {
assert(Is64Bit);
// how far this slot is from the bottom (since they allocate just enough
// space to accomodate holding this slot at the correct offset).
int PSPSymFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
- MMI.getWinEHFuncInfo(MF.getFunction()).PSPSymFrameIdx = PSPSymFI;
+ EHInfo->PSPSymFrameIdx = PSPSymFI;
}
}
}
}
+static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) {
+ switch (SetCCOpcode) {
+ default: llvm_unreachable("Invalid integer condition!");
+ case ISD::SETEQ: return X86::COND_E;
+ case ISD::SETGT: return X86::COND_G;
+ case ISD::SETGE: return X86::COND_GE;
+ case ISD::SETLT: return X86::COND_L;
+ case ISD::SETLE: return X86::COND_LE;
+ case ISD::SETNE: return X86::COND_NE;
+ case ISD::SETULT: return X86::COND_B;
+ case ISD::SETUGT: return X86::COND_A;
+ case ISD::SETULE: return X86::COND_BE;
+ case ISD::SETUGE: return X86::COND_AE;
+ }
+}
+
/// Do a one-to-one translation of a ISD::CondCode to the X86-specific
/// condition code, returning the condition code and the LHS/RHS of the
/// comparison to make.
}
}
- switch (SetCCOpcode) {
- default: llvm_unreachable("Invalid integer condition!");
- case ISD::SETEQ: return X86::COND_E;
- case ISD::SETGT: return X86::COND_G;
- case ISD::SETGE: return X86::COND_GE;
- case ISD::SETLT: return X86::COND_L;
- case ISD::SETLE: return X86::COND_LE;
- case ISD::SETNE: return X86::COND_NE;
- case ISD::SETULT: return X86::COND_B;
- case ISD::SETUGT: return X86::COND_A;
- case ISD::SETULE: return X86::COND_BE;
- case ISD::SETUGE: return X86::COND_AE;
- }
+ return TranslateIntegerX86CC(SetCCOpcode);
}
// First determine if it is required or is profitable to flip the operands.
return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
}
+/// Insert i1-subvector to i1-vector.
+static SDValue Insert1BitVector(SDValue Op, SelectionDAG &DAG) {
+
+ SDLoc dl(Op);
+ SDValue Vec = Op.getOperand(0);
+ SDValue SubVec = Op.getOperand(1);
+ SDValue Idx = Op.getOperand(2);
+
+ if (!isa<ConstantSDNode>(Idx))
+ return SDValue();
+
+ unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+ if (IdxVal == 0 && Vec.isUndef()) // the operation is legal
+ return Op;
+
+ MVT OpVT = Op.getSimpleValueType();
+ MVT SubVecVT = SubVec.getSimpleValueType();
+ unsigned NumElems = OpVT.getVectorNumElements();
+ unsigned SubVecNumElems = SubVecVT.getVectorNumElements();
+
+ assert(IdxVal + SubVecNumElems <= NumElems &&
+ IdxVal % SubVecVT.getSizeInBits() == 0 &&
+ "Unexpected index value in INSERT_SUBVECTOR");
+
+ // There are 3 possible cases:
+ // 1. Subvector should be inserted in the lower part (IdxVal == 0)
+ // 2. Subvector should be inserted in the upper part
+ // (IdxVal + SubVecNumElems == NumElems)
+ // 3. Subvector should be inserted in the middle (for example v2i1
+ // to v16i1, index 2)
+
+ SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
+ SDValue Undef = DAG.getUNDEF(OpVT);
+ SDValue WideSubVec =
+ DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef, SubVec, ZeroIdx);
+ if (Vec.isUndef())
+ return DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
+ DAG.getConstant(IdxVal, dl, MVT::i8));
+
+ if (ISD::isBuildVectorAllZeros(Vec.getNode())) {
+ unsigned ShiftLeft = NumElems - SubVecNumElems;
+ unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal;
+ WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
+ DAG.getConstant(ShiftLeft, dl, MVT::i8));
+ return ShiftRight ? DAG.getNode(X86ISD::VSRLI, dl, OpVT, WideSubVec,
+ DAG.getConstant(ShiftRight, dl, MVT::i8)) : WideSubVec;
+ }
+
+ if (IdxVal == 0) {
+ // Zero lower bits of the Vec
+ SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
+ Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
+ Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
+ // Merge them together
+ return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
+ }
+
+ // Simple case when we put subvector in the upper part
+ if (IdxVal + SubVecNumElems == NumElems) {
+ // Zero upper bits of the Vec
+ WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec,
+ DAG.getConstant(IdxVal, dl, MVT::i8));
+ SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
+ Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
+ Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
+ return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
+ }
+ // Subvector should be inserted in the middle - use shuffle
+ SmallVector<int, 64> Mask;
+ for (unsigned i = 0; i < NumElems; ++i)
+ Mask.push_back(i >= IdxVal && i < IdxVal + SubVecNumElems ?
+ i : i + NumElems);
+ return DAG.getVectorShuffle(OpVT, dl, WideSubVec, Vec, Mask);
+}
+
/// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
/// instructions. This is used because creating CONCAT_VECTOR nodes of
/// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
unsigned NumZero = 0;
unsigned NumNonZero = 0;
- unsigned NonZeros = 0;
+ uint64_t NonZeros = 0;
bool IsAllConstants = true;
SmallSet<SDValue, 8> Values;
for (unsigned i = 0; i < NumElems; ++i) {
if (X86::isZeroNode(Elt))
NumZero++;
else {
- NonZeros |= (1 << i);
+ assert(i < sizeof(NonZeros) * 8); // Make sure the shift is within range.
+ NonZeros |= ((uint64_t)1 << i);
NumNonZero++;
}
}
// One half is zero or undef.
unsigned Idx = countTrailingZeros(NonZeros);
SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
- Op.getOperand(Idx));
+ Op.getOperand(Idx));
return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
}
return SDValue();
// If element VT is < 32 bits, convert it to inserts into a zero vector.
if (EVTBits == 8 && NumElems == 16)
- if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
- Subtarget, *this))
+ if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros, NumNonZero, NumZero,
+ DAG, Subtarget, *this))
return V;
if (EVTBits == 16 && NumElems == 8)
- if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
- Subtarget, *this))
+ if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros, NumNonZero, NumZero,
+ DAG, Subtarget, *this))
return V;
// If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
assert(isPowerOf2_32(NumOfOperands) &&
"Unexpected number of operands in CONCAT_VECTORS");
+ SDValue Undef = DAG.getUNDEF(ResVT);
if (NumOfOperands > 2) {
+ // Specialize the cases when all, or all but one, of the operands are undef.
+ unsigned NumOfDefinedOps = 0;
+ unsigned OpIdx = 0;
+ for (unsigned i = 0; i < NumOfOperands; i++)
+ if (!Op.getOperand(i).isUndef()) {
+ NumOfDefinedOps++;
+ OpIdx = i;
+ }
+ if (NumOfDefinedOps == 0)
+ return Undef;
+ if (NumOfDefinedOps == 1) {
+ unsigned SubVecNumElts =
+ Op.getOperand(OpIdx).getValueType().getVectorNumElements();
+ SDValue IdxVal = DAG.getIntPtrConstant(SubVecNumElts * OpIdx, dl);
+ return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef,
+ Op.getOperand(OpIdx), IdxVal);
+ }
+
MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
ResVT.getVectorNumElements()/2);
SmallVector<SDValue, 2> Ops;
return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
}
+ // 2 operands
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
+ unsigned NumElems = ResVT.getVectorNumElements();
+ assert(V1.getValueType() == V2.getValueType() &&
+ V1.getValueType().getVectorNumElements() == NumElems/2 &&
+ "Unexpected operands in CONCAT_VECTORS");
+
+ if (ResVT.getSizeInBits() >= 16)
+ return Op; // The operation is legal with KUNPCK
+
bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
-
+ SDValue ZeroVec = getZeroVector(ResVT, Subtarget, DAG, dl);
if (IsZeroV1 && IsZeroV2)
- return getZeroVector(ResVT, Subtarget, DAG, dl);
+ return ZeroVec;
SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
- SDValue Undef = DAG.getUNDEF(ResVT);
- unsigned NumElems = ResVT.getVectorNumElements();
- SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
+ if (V2.isUndef())
+ return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
+ if (IsZeroV2)
+ return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V1, ZeroIdx);
+
+ SDValue IdxVal = DAG.getIntPtrConstant(NumElems/2, dl);
+ if (V1.isUndef())
+ V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, IdxVal);
- V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
- V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
if (IsZeroV1)
- return V2;
+ return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V2, IdxVal);
V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
- // Zero the upper bits of V1
- V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
- V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
- if (IsZeroV2)
- return V1;
- return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
+ return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, V1, V2, IdxVal);
}
static SDValue LowerCONCAT_VECTORS(SDValue Op,
SDValue &V = (M < Size ? V1 : V2);
M = M % Size;
- // All mask elements must be in the lower half.
- if (M >= HalfSize)
+ // The extracted elements must start at a valid index and all mask
+ // elements must be in the lower half.
+ if (i > M || M >= HalfSize)
return SDValue();
if (Idx < 0 || (Src == V && Idx == (M - i))) {
unsigned &MaskValue) {
MaskValue = 0;
unsigned NumElems = BuildVector->getNumOperands();
-
+
// There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
// We don't handle the >2 lanes case right now.
unsigned NumLanes = (NumElems - 1) / 8 + 1;
if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
- if (OpVT.getVectorElementType() == MVT::i1) {
- if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
- return Op;
- SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
- SDValue Undef = DAG.getUNDEF(OpVT);
- unsigned NumElems = OpVT.getVectorNumElements();
- SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
-
- if (IdxVal == OpVT.getVectorNumElements() / 2) {
- // Zero upper bits of the Vec
- Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
- Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
-
- SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
- SubVec, ZeroIdx);
- Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
- return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
- }
- if (IdxVal == 0) {
- SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
- SubVec, ZeroIdx);
- // Zero upper bits of the Vec2
- Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
- Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
- // Zero lower bits of the Vec
- Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
- Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
- // Merge them together
- return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
- }
- }
+ if (OpVT.getVectorElementType() == MVT::i1)
+ return Insert1BitVector(Op, DAG);
+
return SDValue();
}
Op1.getOpcode() == ISD::Constant &&
cast<ConstantSDNode>(Op1)->isNullValue() &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
- SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
- if (NewSetCC.getNode()) {
+ if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) {
if (VT == MVT::i1)
return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
return NewSetCC;
return SetCC;
}
+SDValue X86TargetLowering::LowerSETCCE(SDValue Op, SelectionDAG &DAG) const {
+ SDValue LHS = Op.getOperand(0);
+ SDValue RHS = Op.getOperand(1);
+ SDValue Carry = Op.getOperand(2);
+ SDValue Cond = Op.getOperand(3);
+ SDLoc DL(Op);
+
+ assert(LHS.getSimpleValueType().isInteger() && "SETCCE is integer only.");
+ X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get());
+
+ assert(Carry.getOpcode() != ISD::CARRY_FALSE);
+ SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
+ SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry);
+ return DAG.getNode(X86ISD::SETCC, DL, Op.getValueType(),
+ DAG.getConstant(CC, DL, MVT::i8), Cmp.getValue(1));
+}
+
// isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
static bool isX86LogicalCmp(SDValue Op) {
unsigned Opc = Op.getNode()->getOpcode();
// We know the result of AND is compared against zero. Try to match
// it to BT.
if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
- SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
- if (NewSetCC.getNode()) {
+ if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG)) {
CC = NewSetCC.getOperand(0);
Cond = NewSetCC.getOperand(1);
addTest = false;
// We know the result of AND is compared against zero. Try to match
// it to BT.
if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
- SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
- if (NewSetCC.getNode()) {
+ if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG)) {
CC = NewSetCC.getOperand(0);
Cond = NewSetCC.getOperand(1);
addTest = false;
Mask, PassThru, Subtarget, DAG);
}
case VPERM_3OP_MASKZ:
- case VPERM_3OP_MASK:
+ case VPERM_3OP_MASK:{
+ // Src2 is the PassThru
+ SDValue Src1 = Op.getOperand(1);
+ SDValue Src2 = Op.getOperand(2);
+ SDValue Src3 = Op.getOperand(3);
+ SDValue Mask = Op.getOperand(4);
+ MVT VT = Op.getSimpleValueType();
+ SDValue PassThru = SDValue();
+
+ // set PassThru element
+ if (IntrData->Type == VPERM_3OP_MASKZ)
+ PassThru = getZeroVector(VT, Subtarget, DAG, dl);
+ else
+ PassThru = Src2;
+
+ // Swap Src1 and Src2 in the node creation
+ return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
+ dl, Op.getValueType(),
+ Src2, Src1, Src3),
+ Mask, PassThru, Subtarget, DAG);
+ }
case FMA_OP_MASK3:
case FMA_OP_MASKZ:
case FMA_OP_MASK: {
SDValue PassThru = SDValue();
// set PassThru element
- if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
+ if (IntrData->Type == FMA_OP_MASKZ)
PassThru = getZeroVector(VT, Subtarget, DAG, dl);
else if (IntrData->Type == FMA_OP_MASK3)
PassThru = Src3;
DataToCompress),
Mask, PassThru, Subtarget, DAG);
}
+ case BROADCASTM: {
+ SDValue Mask = Op.getOperand(1);
+ MVT MaskVT = MVT::getVectorVT(MVT::i1, Mask.getSimpleValueType().getSizeInBits());
+ Mask = DAG.getBitcast(MaskVT, Mask);
+ return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Mask);
+ }
case BLEND: {
SDValue Mask = Op.getOperand(3);
MVT VT = Op.getSimpleValueType();
return Chain;
}
+static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) {
+ MachineFunction &MF = DAG.getMachineFunction();
+ SDValue Chain = Op.getOperand(0);
+ SDValue RegNode = Op.getOperand(2);
+ WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo();
+ if (!EHInfo)
+ report_fatal_error("EH registrations only live in functions using WinEH");
+
+ // Cast the operand to an alloca, and remember the frame index.
+ auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode);
+ if (!FINode)
+ report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca");
+ EHInfo->EHRegNodeFrameIndex = FINode->getIndex();
+
+ // Return the chain operand without making any DAG nodes.
+ return Chain;
+}
+
/// \brief Lower intrinsics for TRUNCATE_TO_MEM case
/// return truncate Store/MaskedStore Node
static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
if (!IntrData) {
if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
+ else if (IntNo == llvm::Intrinsic::x86_seh_ehregnode)
+ return MarkEHRegistrationNode(Op, DAG);
return SDValue();
}
// chunks, thus directly computes the pop count for v2i64 and v4i64.
if (EltVT == MVT::i64) {
SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
- V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
+ MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
+ V = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, V, Zeros);
return DAG.getBitcast(VT, V);
}
// Do the horizontal sums into two v2i64s.
Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
- Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
+ MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
+ Low = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
DAG.getBitcast(ByteVecVT, Low), Zeros);
- High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
+ High = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
DAG.getBitcast(ByteVecVT, High), Zeros);
// Merge them together.
case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
case ISD::SETCC: return LowerSETCC(Op, DAG);
+ case ISD::SETCCE: return LowerSETCCE(Op, DAG);
case ISD::SELECT: return LowerSELECT(Op, DAG);
case ISD::BRCOND: return LowerBRCOND(Op, DAG);
case ISD::JumpTable: return LowerJumpTable(Op, DAG);
switch (N->getOpcode()) {
default:
llvm_unreachable("Do not know how to custom type legalize this operation!");
+ case X86ISD::AVG: {
+ // Legalize types for X86ISD::AVG by expanding vectors.
+ assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
+
+ auto InVT = N->getValueType(0);
+ auto InVTSize = InVT.getSizeInBits();
+ const unsigned RegSize =
+ (InVTSize > 128) ? ((InVTSize > 256) ? 512 : 256) : 128;
+ assert((!Subtarget->hasAVX512() || RegSize < 512) &&
+ "512-bit vector requires AVX512");
+ assert((!Subtarget->hasAVX2() || RegSize < 256) &&
+ "256-bit vector requires AVX2");
+
+ auto ElemVT = InVT.getVectorElementType();
+ auto RegVT = EVT::getVectorVT(*DAG.getContext(), ElemVT,
+ RegSize / ElemVT.getSizeInBits());
+ assert(RegSize % InVT.getSizeInBits() == 0);
+ unsigned NumConcat = RegSize / InVT.getSizeInBits();
+
+ SmallVector<SDValue, 16> Ops(NumConcat, DAG.getUNDEF(InVT));
+ Ops[0] = N->getOperand(0);
+ SDValue InVec0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
+ Ops[0] = N->getOperand(1);
+ SDValue InVec1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
+
+ SDValue Res = DAG.getNode(X86ISD::AVG, dl, RegVT, InVec0, InVec1);
+ Results.push_back(DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InVT, Res,
+ DAG.getIntPtrConstant(0, dl)));
+ return;
+ }
// We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
case X86ISD::FMINC:
case X86ISD::FMIN:
case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
+ case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
return SDValue();
}
+/// This function detects the AVG pattern between vectors of unsigned i8/i16,
+/// which is c = (a + b + 1) / 2, and replace this operation with the efficient
+/// X86ISD::AVG instruction.
+static SDValue detectAVGPattern(SDValue In, EVT VT, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget, SDLoc DL) {
+ if (!VT.isVector() || !VT.isSimple())
+ return SDValue();
+ EVT InVT = In.getValueType();
+ unsigned NumElems = VT.getVectorNumElements();
+
+ EVT ScalarVT = VT.getVectorElementType();
+ if (!((ScalarVT == MVT::i8 || ScalarVT == MVT::i16) &&
+ isPowerOf2_32(NumElems)))
+ return SDValue();
+
+ // InScalarVT is the intermediate type in AVG pattern and it should be greater
+ // than the original input type (i8/i16).
+ EVT InScalarVT = InVT.getVectorElementType();
+ if (InScalarVT.getSizeInBits() <= ScalarVT.getSizeInBits())
+ return SDValue();
+
+ if (Subtarget->hasAVX512()) {
+ if (VT.getSizeInBits() > 512)
+ return SDValue();
+ } else if (Subtarget->hasAVX2()) {
+ if (VT.getSizeInBits() > 256)
+ return SDValue();
+ } else {
+ if (VT.getSizeInBits() > 128)
+ return SDValue();
+ }
+
+ // Detect the following pattern:
+ //
+ // %1 = zext <N x i8> %a to <N x i32>
+ // %2 = zext <N x i8> %b to <N x i32>
+ // %3 = add nuw nsw <N x i32> %1, <i32 1 x N>
+ // %4 = add nuw nsw <N x i32> %3, %2
+ // %5 = lshr <N x i32> %N, <i32 1 x N>
+ // %6 = trunc <N x i32> %5 to <N x i8>
+ //
+ // In AVX512, the last instruction can also be a trunc store.
+
+ if (In.getOpcode() != ISD::SRL)
+ return SDValue();
+
+ // A lambda checking the given SDValue is a constant vector and each element
+ // is in the range [Min, Max].
+ auto IsConstVectorInRange = [](SDValue V, unsigned Min, unsigned Max) {
+ BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(V);
+ if (!BV || !BV->isConstant())
+ return false;
+ for (unsigned i = 0, e = V.getNumOperands(); i < e; i++) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(i));
+ if (!C)
+ return false;
+ uint64_t Val = C->getZExtValue();
+ if (Val < Min || Val > Max)
+ return false;
+ }
+ return true;
+ };
+
+ // Check if each element of the vector is left-shifted by one.
+ auto LHS = In.getOperand(0);
+ auto RHS = In.getOperand(1);
+ if (!IsConstVectorInRange(RHS, 1, 1))
+ return SDValue();
+ if (LHS.getOpcode() != ISD::ADD)
+ return SDValue();
+
+ // Detect a pattern of a + b + 1 where the order doesn't matter.
+ SDValue Operands[3];
+ Operands[0] = LHS.getOperand(0);
+ Operands[1] = LHS.getOperand(1);
+
+ // Take care of the case when one of the operands is a constant vector whose
+ // element is in the range [1, 256].
+ if (IsConstVectorInRange(Operands[1], 1, ScalarVT == MVT::i8 ? 256 : 65536) &&
+ Operands[0].getOpcode() == ISD::ZERO_EXTEND &&
+ Operands[0].getOperand(0).getValueType() == VT) {
+ // The pattern is detected. Subtract one from the constant vector, then
+ // demote it and emit X86ISD::AVG instruction.
+ SDValue One = DAG.getConstant(1, DL, InScalarVT);
+ SDValue Ones = DAG.getNode(ISD::BUILD_VECTOR, DL, InVT,
+ SmallVector<SDValue, 8>(NumElems, One));
+ Operands[1] = DAG.getNode(ISD::SUB, DL, InVT, Operands[1], Ones);
+ Operands[1] = DAG.getNode(ISD::TRUNCATE, DL, VT, Operands[1]);
+ return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
+ Operands[1]);
+ }
+
+ if (Operands[0].getOpcode() == ISD::ADD)
+ std::swap(Operands[0], Operands[1]);
+ else if (Operands[1].getOpcode() != ISD::ADD)
+ return SDValue();
+ Operands[2] = Operands[1].getOperand(0);
+ Operands[1] = Operands[1].getOperand(1);
+
+ // Now we have three operands of two additions. Check that one of them is a
+ // constant vector with ones, and the other two are promoted from i8/i16.
+ for (int i = 0; i < 3; ++i) {
+ if (!IsConstVectorInRange(Operands[i], 1, 1))
+ continue;
+ std::swap(Operands[i], Operands[2]);
+
+ // Check if Operands[0] and Operands[1] are results of type promotion.
+ for (int j = 0; j < 2; ++j)
+ if (Operands[j].getOpcode() != ISD::ZERO_EXTEND ||
+ Operands[j].getOperand(0).getValueType() != VT)
+ return SDValue();
+
+ // The pattern is detected, emit X86ISD::AVG instruction.
+ return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
+ Operands[1].getOperand(0));
+ }
+
+ return SDValue();
+}
+
+static SDValue PerformTRUNCATECombine(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ return detectAVGPattern(N->getOperand(0), N->getValueType(0), DAG, Subtarget,
+ SDLoc(N));
+}
+
/// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
// First, pack all of the elements in one place. Next, store to memory
// in fewer chunks.
if (St->isTruncatingStore() && VT.isVector()) {
+ // Check if we can detect an AVG pattern from the truncation. If yes,
+ // replace the trunc store by a normal store with the result of X86ISD::AVG
+ // instruction.
+ SDValue Avg =
+ detectAVGPattern(St->getValue(), St->getMemoryVT(), DAG, Subtarget, dl);
+ if (Avg.getNode())
+ return DAG.getStore(St->getChain(), dl, Avg, St->getBasePtr(),
+ St->getPointerInfo(), St->isVolatile(),
+ St->isNonTemporal(), St->getAlignment());
+
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
unsigned NumElems = VT.getVectorNumElements();
assert(StVT != VT && "Cannot truncate to the same type");
return SDValue();
}
+/// Do target-specific dag combines on floating point negations.
+static SDValue PerformFNEGCombine(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ EVT VT = N->getValueType(0);
+ SDValue Arg = N->getOperand(0);
+
+ // If we're negating a FMA node, then we can adjust the
+ // instruction to include the extra negation.
+ if (Arg.hasOneUse()) {
+ switch (Arg.getOpcode()) {
+ case X86ISD::FMADD:
+ return DAG.getNode(X86ISD::FNMSUB, SDLoc(N), VT, Arg.getOperand(0),
+ Arg.getOperand(1), Arg.getOperand(2));
+ case X86ISD::FMSUB:
+ return DAG.getNode(X86ISD::FNMADD, SDLoc(N), VT, Arg.getOperand(0),
+ Arg.getOperand(1), Arg.getOperand(2));
+ case X86ISD::FNMADD:
+ return DAG.getNode(X86ISD::FMSUB, SDLoc(N), VT, Arg.getOperand(0),
+ Arg.getOperand(1), Arg.getOperand(2));
+ case X86ISD::FNMSUB:
+ return DAG.getNode(X86ISD::FMADD, SDLoc(N), VT, Arg.getOperand(0),
+ Arg.getOperand(1), Arg.getOperand(2));
+ }
+ }
+ return SDValue();
+}
+
/// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
const X86Subtarget *Subtarget) {
}
// Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
- // a 32-bit target where SSE doesn't support i64->FP operations.
+ // a 32-bit target where SSE doesn't support i64->FP operations.
if (!Subtarget->useSoftFloat() && Op0.getOpcode() == ISD::LOAD) {
LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
EVT LdVT = Ld->getValueType(0);
case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
+ case ISD::FNEG: return PerformFNEGCombine(N, DAG, Subtarget);
+ case ISD::TRUNCATE: return PerformTRUNCATECombine(N, DAG, Subtarget);
case X86ISD::FXOR:
case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
case X86ISD::FMIN: