setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
- if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
+ if (Subtarget->hasAnyFMA()) {
setOperationAction(ISD::FMA, MVT::v8f32, Legal);
setOperationAction(ISD::FMA, MVT::v4f64, Legal);
setOperationAction(ISD::FMA, MVT::v4f32, Legal);
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);
return getInsertVINSERTImmediate(N, 256);
}
-/// Returns true if V is a constant integer zero.
-static bool isZero(SDValue V) {
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
- return C && C->isNullValue();
-}
-
/// Returns true if Elt is a constant zero or a floating point constant +0.0.
bool X86::isZeroNode(SDValue Elt) {
- if (isZero(Elt))
- return true;
- if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
- return CFP->getValueAPF().isPosZero();
- return false;
+ return isNullConstant(Elt) || isNullFPConstant(Elt);
}
// Build a vector of constants
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
SmallVector<SDValue, 8> V(NumElems);
if (NumElems == 4 && NumZero > 0) {
for (unsigned i = 0; i < 4; ++i) {
- bool isZero = !(NonZeros & (1 << i));
+ bool isZero = !(NonZeros & (1ULL << i));
if (isZero)
V[i] = getZeroVector(VT, Subtarget, DAG, dl);
else
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))) {
if (Subtarget->hasAVX2())
return DAG.getNode(
X86ISD::VPERMV, DL, MVT::v8f32,
- DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
- MVT::v8i32, VPermMask)),
- V1);
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
// Otherwise, fall back.
return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
int Lane1Cond = -1, Lane2Cond = -1;
if (isa<ConstantSDNode>(EltCond))
- Lane1Cond = !isZero(EltCond);
+ Lane1Cond = !isNullConstant(EltCond);
if (isa<ConstantSDNode>(SndLaneEltCond))
- Lane2Cond = !isZero(SndLaneEltCond);
+ Lane2Cond = !isNullConstant(SndLaneEltCond);
unsigned LaneMask = 0;
if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
SDValue CondElt = CondBV->getOperand(i);
Mask.push_back(
- isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
+ isa<ConstantSDNode>(CondElt) ? i + (isNullConstant(CondElt) ? Size : 0)
+ : -1);
}
return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
}
}
if (VT.getSizeInBits() == 16) {
- unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
// If Idx is 0, it's cheaper to do a move instead of a pextrw.
- if (Idx == 0)
+ if (isNullConstant(Op.getOperand(1)))
return DAG.getNode(
ISD::TRUNCATE, dl, MVT::i16,
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
return SDValue();
SDNode *User = *Op.getNode()->use_begin();
if ((User->getOpcode() != ISD::STORE ||
- (isa<ConstantSDNode>(Op.getOperand(1)) &&
- cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
+ isNullConstant(Op.getOperand(1))) &&
(User->getOpcode() != ISD::BITCAST ||
User->getValueType(0) != MVT::i32))
return SDValue();
// TODO: handle v16i8.
if (VT.getSizeInBits() == 16) {
SDValue Vec = Op.getOperand(0);
- unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
- if (Idx == 0)
+ if (isNullConstant(Op.getOperand(1)))
return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
DAG.getBitcast(MVT::v4i32, Vec),
// FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
// FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
// to match extract_elt for f64.
- unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
- if (Idx == 0)
+ if (isNullConstant(Op.getOperand(1)))
return Op;
// UNPCKHPD the element to the lowest double word, then movsd.
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();
}
if (ConstantSDNode *C =
dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
// An add of one will be selected as an INC.
- if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
+ if (C->isOne() && !Subtarget->slowIncDec()) {
Opcode = X86ISD::INC;
NumOperands = 1;
break;
}
// An add of negative one (subtract of one) will be selected as a DEC.
- if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
+ if (C->isAllOnesValue() && !Subtarget->slowIncDec()) {
Opcode = X86ISD::DEC;
NumOperands = 1;
break;
/// equivalent.
SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
SDLoc dl, SelectionDAG &DAG) const {
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
- if (C->getAPIntValue() == 0)
- return EmitTest(Op0, X86CC, dl, DAG);
+ if (isNullConstant(Op1))
+ return EmitTest(Op0, X86CC, dl, DAG);
- assert(Op0.getValueType() != MVT::i1 &&
- "Unexpected comparison operation for MVT::i1 operands");
- }
+ assert(!(isa<ConstantSDNode>(Op1) && Op0.getValueType() == MVT::i1) &&
+ "Unexpected comparison operation for MVT::i1 operands");
if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
return 2;
}
-static bool isAllOnes(SDValue V) {
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
- return C && C->isAllOnesValue();
-}
-
/// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
/// if it's possible.
SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
if (Op1.getOpcode() == ISD::SHL)
std::swap(Op0, Op1);
if (Op0.getOpcode() == ISD::SHL) {
- if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
- if (And00C->getZExtValue() == 1) {
+ if (isOneConstant(Op0.getOperand(0))) {
// If we looked past a truncate, check that it's only truncating away
// known zeros.
unsigned BitWidth = Op0.getValueSizeInBits();
// Lower ((X >>u N) & 1) != 0 to BT(X, N).
// Lower ((X >>s N) & 1) != 0 to BT(X, N).
if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
- Op1.getOpcode() == ISD::Constant &&
- cast<ConstantSDNode>(Op1)->isNullValue() &&
+ isNullConstant(Op1) &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) {
if (VT == MVT::i1)
// Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
// these.
- if (Op1.getOpcode() == ISD::Constant &&
- (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
- cast<ConstantSDNode>(Op1)->isNullValue()) &&
+ if ((isOneConstant(Op1) || isNullConstant(Op1)) &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
// If the input is a setcc, then reuse the input setcc or use a new one with
// the inverted condition.
if (Op0.getOpcode() == X86ISD::SETCC) {
X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
- bool Invert = (CC == ISD::SETNE) ^
- cast<ConstantSDNode>(Op1)->isNullValue();
+ bool Invert = (CC == ISD::SETNE) ^ isNullConstant(Op1);
if (!Invert)
return Op0;
return SetCC;
}
}
- if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
- (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
+ if ((Op0.getValueType() == MVT::i1) && isOneConstant(Op1) &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
// (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
if (Cond.getOpcode() == X86ISD::SETCC &&
Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
- isZero(Cond.getOperand(1).getOperand(1))) {
+ isNullConstant(Cond.getOperand(1).getOperand(1))) {
SDValue Cmp = Cond.getOperand(1);
unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
- if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
+ if ((isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
(CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
- SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
+ SDValue Y = isAllOnesConstant(Op2) ? Op1 : Op2;
SDValue CmpOp0 = Cmp.getOperand(0);
// Apply further optimizations for special cases
// (select (x != 0), -1, 0) -> neg & sbb
// (select (x == 0), 0, -1) -> neg & sbb
- if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
- if (YC->isNullValue() &&
- (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
+ if (isNullConstant(Y) &&
+ (isAllOnesConstant(Op1) == (CondCode == X86::COND_NE))) {
SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
DAG.getConstant(0, DL,
DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
- if (isAllOnes(Op1) != (CondCode == X86::COND_E))
+ if (isAllOnesConstant(Op1) != (CondCode == X86::COND_E))
Res = DAG.getNOT(DL, Res, Res.getValueType());
- ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
- if (!N2C || !N2C->isNullValue())
+ if (!isNullConstant(Op2))
Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
return Res;
}
// Look past (and (setcc_carry (cmp ...)), 1).
if (Cond.getOpcode() == ISD::AND &&
- Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
- if (C && C->getAPIntValue() == 1)
- Cond = Cond.getOperand(0);
- }
+ Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
+ isOneConstant(Cond.getOperand(1)))
+ Cond = Cond.getOperand(0);
// If condition flag is set by a X86ISD::CMP, then use it as the condition
// setting operand in place of the X86ISD::SETCC.
unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
- (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
+ (isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
+ (isNullConstant(Op1) || isNullConstant(Op2))) {
SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
DAG.getConstant(X86::COND_B, DL, MVT::i8),
Cond);
- if (isAllOnes(Op1) != (CondCode == X86::COND_B))
+ if (isAllOnesConstant(Op1) != (CondCode == X86::COND_B))
return DAG.getNOT(DL, Res, Res.getValueType());
return Res;
}
static bool isXor1OfSetCC(SDValue Op) {
if (Op.getOpcode() != ISD::XOR)
return false;
- ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
- if (N1C && N1C->getAPIntValue() == 1) {
+ if (isOneConstant(Op.getOperand(1)))
return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
- Op.getOperand(0).hasOneUse();
- }
+ Op.getOperand(0).hasOneUse();
return false;
}
if (Cond.getOpcode() == ISD::SETCC) {
// Check for setcc([su]{add,sub,mul}o == 0).
if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
- isa<ConstantSDNode>(Cond.getOperand(1)) &&
- cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
+ isNullConstant(Cond.getOperand(1)) &&
Cond.getOperand(0).getResNo() == 1 &&
(Cond.getOperand(0).getOpcode() == ISD::SADDO ||
Cond.getOperand(0).getOpcode() == ISD::UADDO ||
// Look pass (and (setcc_carry (cmp ...)), 1).
if (Cond.getOpcode() == ISD::AND &&
- Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
- if (C && C->getAPIntValue() == 1)
- Cond = Cond.getOperand(0);
- }
+ Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
+ isOneConstant(Cond.getOperand(1)))
+ Cond = Cond.getOperand(0);
// If condition flag is set by a X86ISD::CMP, then use it as the condition
// setting operand in place of the X86ISD::SETCC.
switch (CondOpcode) {
case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
case ISD::SADDO:
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
- if (C->isOne()) {
+ if (isOneConstant(RHS)) {
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:
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
- if (C->isOne()) {
+ if (isOneConstant(RHS)) {
X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
break;
}
unsigned OpcodeSelect = ISD::VSELECT;
SDLoc dl(Op);
- if (isAllOnes(Mask))
+ if (isAllOnesConstant(Mask))
return Op;
if (MaskVT.bitsGT(Mask.getSimpleValueType())) {
case X86ISD::CMPMU:
return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
case X86ISD::VFPCLASS:
+ case X86ISD::VFPCLASSS:
return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
case X86ISD::VTRUNC:
case X86ISD::VTRUNCS:
SDValue PreservedSrc,
const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
- if (isAllOnes(Mask))
+ if (isAllOnesConstant(Mask))
return Op;
MVT VT = Op.getSimpleValueType();
if (Op.getOpcode() == X86ISD::FSETCC)
return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
- if (Op.getOpcode() == X86ISD::VFPCLASS)
+ if (Op.getOpcode() == X86ISD::VFPCLASS ||
+ Op.getOpcode() == X86ISD::VFPCLASSS)
return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
if (PreservedSrc.getOpcode() == ISD::UNDEF)
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 = DAG.getBitcast(VT, 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;
SDValue Mask = Op.getOperand(3);
SDValue DataToCompress = Op.getOperand(1);
SDValue PassThru = Op.getOperand(2);
- if (isAllOnes(Mask)) // return data as is
+ if (isAllOnesConstant(Mask)) // return data as is
return Op.getOperand(1);
return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
MVT VT = DataToTruncate.getSimpleValueType();
MVT SVT = MVT::getVectorVT(ElementType, VT.getVectorNumElements());
- if (isAllOnes(Mask)) // return just a truncate store
+ if (isAllOnesConstant(Mask)) // return just a truncate store
return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
MachinePointerInfo(), SVT, false, false,
SVT.getScalarSizeInBits()/8);
SDValue Chain = Op.getOperand(0);
MVT VT = DataToCompress.getSimpleValueType();
- if (isAllOnes(Mask)) // return just a store
+ if (isAllOnesConstant(Mask)) // return just a store
return DAG.getStore(Chain, dl, DataToCompress, Addr,
MachinePointerInfo(), false, false,
VT.getScalarSizeInBits()/8);
SDValue Chain = Op.getOperand(0);
MVT VT = Op.getSimpleValueType();
- if (isAllOnes(Mask)) // return just a load
+ if (isAllOnesConstant(Mask)) // return just a load
return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
false, VT.getScalarSizeInBits()/8);
case ISD::SADDO:
// A subtract of one will be selected as a INC. Note that INC doesn't
// set CF, so we can't do this for UADDO.
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
- if (C->isOne()) {
+ if (isOneConstant(RHS)) {
BaseOp = X86ISD::INC;
Cond = X86::COND_O;
break;
case ISD::SSUBO:
// A subtract of one will be selected as a DEC. Note that DEC doesn't
// set CF, so we can't do this for USUBO.
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
- if (C->isOne()) {
+ if (isOneConstant(RHS)) {
BaseOp = X86ISD::DEC;
Cond = X86::COND_O;
break;
// 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.
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::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
+ case X86ISD::VFPCLASSS: return "X86ISD::VFPCLASSS";
}
return nullptr;
}
bool
X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
- if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
+ if (!Subtarget->hasAnyFMA())
return false;
VT = VT.getScalarType();
return SDValue();
auto *SVN = cast<ShuffleVectorSDNode>(N);
- ArrayRef<int> Mask = SVN->getMask();
+ SmallVector<int, 8> Mask;
+ for (int M : SVN->getMask())
+ Mask.push_back(M);
+
SDValue V1 = N->getOperand(0);
SDValue V2 = N->getOperand(1);
- // We require the first shuffle operand to be the SUB node, and the second to
- // be the ADD node.
- // FIXME: We should support the commuted patterns.
- if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
+ // We require the first shuffle operand to be the FSUB node, and the second to
+ // be the FADD node.
+ if (V1.getOpcode() == ISD::FADD && V2.getOpcode() == ISD::FSUB) {
+ ShuffleVectorSDNode::commuteMask(Mask);
+ std::swap(V1, V2);
+ } else if (V1.getOpcode() != ISD::FSUB || V2.getOpcode() != ISD::FADD)
return SDValue();
// If there are other uses of these operations we can't fold them.
// store-load conversions.
if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
N0.getValueType() == MVT::v2i32 &&
- isa<ConstantSDNode>(N0.getOperand(1))) {
+ isNullConstant(N0.getOperand(1))) {
SDValue N00 = N0->getOperand(0);
- if (N0.getConstantOperandVal(1) == 0 && N00.getValueType() == MVT::i32)
+ if (N00.getValueType() == MVT::i32)
return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
}
SetCC.getOpcode() == ISD::AND) {
if (SetCC.getOpcode() == ISD::AND) {
int OpIdx = -1;
- ConstantSDNode *CS;
- if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
- CS->getZExtValue() == 1)
+ if (isOneConstant(SetCC.getOperand(0)))
OpIdx = 1;
- if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
- CS->getZExtValue() == 1)
+ if (isOneConstant(SetCC.getOperand(1)))
OpIdx = 0;
if (OpIdx == -1)
break;
X86::CondCode &CC1, SDValue &Flags,
bool &isAnd) {
if (Cond->getOpcode() == X86ISD::CMP) {
- ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
- if (!CondOp1C || !CondOp1C->isNullValue())
+ if (!isNullConstant(Cond->getOperand(1)))
return false;
Cond = Cond->getOperand(0);
return SDValue();
// Make sure we are performing an xor against one.
- if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
+ if (!isOneConstant(N1))
return SDValue();
// SetCC on x86 zero extends so only act on this if it's a logical shift.
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,
SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
for (unsigned i = 0; i != NumElems; ++i)
ShuffleVec[i] = i * SizeRatio;
- for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
- ShuffleVec[i] = NumElems*SizeRatio;
+ for (unsigned i = NumElems; i != NumElems * SizeRatio; ++i)
+ ShuffleVec[i] = NumElems * SizeRatio;
NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
DAG.getConstant(0, dl, WideVecVT),
&ShuffleVec[0]);
"WideVecVT should be legal");
SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
- DAG.getUNDEF(WideVecVT),
- &ShuffleVec[0]);
+ DAG.getUNDEF(WideVecVT),
+ &ShuffleVec[0]);
SDValue NewMask;
SDValue Mask = Mst->getMask();
NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
}
- return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
- NewMask, StVT, Mst->getMemOperand(), false);
+ return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal,
+ Mst->getBasePtr(), NewMask, StVT,
+ Mst->getMemOperand(), false);
}
/// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
// 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) {
return SDValue();
EVT ScalarVT = VT.getScalarType();
- if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
- (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
- !Subtarget->hasAVX512()))
+ if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) || !Subtarget->hasAnyFMA())
return SDValue();
SDValue A = N->getOperand(0);
N0.getOperand(0).hasOneUse()) {
SDValue N00 = N0.getOperand(0);
if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
- if (!C || C->getZExtValue() != 1)
+ if (!isOneConstant(N0.getOperand(1)))
return SDValue();
return DAG.getNode(ISD::AND, dl, VT,
DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
SDLoc DL(N);
if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
- if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
+ if (isNullConstant(LHS.getOperand(0)) && LHS.hasOneUse()) {
SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
LHS.getOperand(1));
return DAG.getSetCC(DL, N->getValueType(0), addV,
DAG.getConstant(0, DL, addV.getValueType()), CC);
}
if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
- if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
+ if (isNullConstant(RHS.getOperand(0)) && RHS.hasOneUse()) {
SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
RHS.getOperand(1));
return DAG.getSetCC(DL, N->getValueType(0), addV,
V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
SDValue ExtractedV = V.getOperand(0);
SDValue OrigV = ExtractedV.getOperand(0);
- if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
- if (ExtractIdx->getZExtValue() == 0) {
+ if (isNullConstant(ExtractedV.getOperand(1))) {
MVT OrigVT = OrigV.getSimpleValueType();
// Extract a subvector if necessary...
if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
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: