"rather than promotion."),
cl::Hidden);
-static cl::opt<int> ReciprocalEstimateRefinementSteps(
- "x86-recip-refinement-steps", cl::init(1),
- cl::desc("Specify the number of Newton-Raphson iterations applied to the "
- "result of the hardware reciprocal estimate instruction."),
- cl::NotHidden);
-
// Forward declarations.
static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
SDValue V2);
}
}
+static SDValue lowerVectorShuffleWithVALIGN(SDLoc DL, MVT VT,
+ ArrayRef<int> Mask, SDValue V1,
+ SDValue V2, SelectionDAG &DAG) {
+
+ assert(VT.getScalarSizeInBits() >= 32 && "Unexpected data type for VALIGN");
+ // VALIGN pattern 2, 3, 4, 5, .. (sequential, shifted right)
+ int AlignVal = -1;
+ for (int i = 0; i < (signed)VT.getVectorNumElements(); ++i) {
+ if (Mask[i] < 0)
+ continue;
+ if (Mask[i] < i)
+ return SDValue();
+ if (AlignVal == -1)
+ AlignVal = Mask[i] - i;
+ else if (Mask[i] - i != AlignVal)
+ return SDValue();
+ }
+ return DAG.getNode(X86ISD::VALIGN, DL, VT, V1, V2,
+ DAG.getConstant(AlignVal, DL, MVT::i8));
+}
+
+static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
+ ArrayRef<int> Mask, SDValue V1,
+ SDValue V2, SelectionDAG &DAG) {
+
+ assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
+
+ MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
+ MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
+
+ SmallVector<SDValue, 32> VPermMask;
+ for (unsigned i = 0; i < VT.getVectorNumElements(); ++i)
+ VPermMask.push_back(Mask[i] < 0 ? DAG.getUNDEF(MaskEltVT) :
+ DAG.getConstant(Mask[i], DL,MaskEltVT));
+ SDValue MaskNode = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecVT,
+ VPermMask);
+ if (isSingleInputShuffleMask(Mask))
+ return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
+
+ return DAG.getNode(X86ISD::VPERMV3, DL, VT, MaskNode, V1, V2);
+}
+
+
/// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
static SDValue lowerV8X64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
const X86Subtarget *Subtarget,
if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
+ if (SDValue Op = lowerVectorShuffleWithVALIGN(DL, VT, Mask, V1, V2, DAG))
+ return Op;
+
// VSHUFPD instruction - mask 0/1, 8/9, 2/3, 10/11, 4/5, 12/13, 6/7, 14/15
bool ShufpdMask = true;
unsigned Immediate = 0;
Immediate |= (Mask[i]%2) << i;
}
if (ShufpdMask)
- return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
- DAG.getConstant(Immediate, DL, MVT::i8));
+ return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
+ DAG.getConstant(Immediate, DL, MVT::i8));
// PERMILPD instruction - mask 0/1, 0/1, 2/3, 2/3, 4/5, 4/5, 6/7, 6/7
if (isSingleInputShuffleMask(Mask)) {
DAG.getConstant(Immediate, DL, MVT::i8));
}
}
- SDValue VPermMask[8];
- for (int i = 0; i < 8; ++i)
- VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i64)
- : DAG.getConstant(Mask[i], DL, MVT::i64);
- SDValue MaskNode = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i64,
- VPermMask);
- if (isSingleInputShuffleMask(Mask))
- return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
-
- return DAG.getNode(X86ISD::VPERMV3, DL, VT, MaskNode, V1, V2);
+ return lowerVectorShuffleWithPERMV(DL, VT, Mask, V1, V2, DAG);
}
-/// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
-static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+/// \brief Handle lowering of 16-lane 32-bit integer shuffles.
+static SDValue lowerV16X32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
+ MVT VT = Op.getSimpleValueType();
SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
+ assert((V1.getSimpleValueType() == MVT::v16i32 ||
+ V1.getSimpleValueType() == MVT::v16f32) && "Bad operand type!");
+ assert((V2.getSimpleValueType() == MVT::v16i32 ||
+ V2.getSimpleValueType() == MVT::v16f32) && "Bad operand type!");
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
0, 16, 1, 17, 4, 20, 5, 21,
// Second 128-bit lane.
8, 24, 9, 25, 12, 28, 13, 29}))
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
+ return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
if (isShuffleEquivalent(V1, V2, Mask,
{// First 128-bit lane.
2, 18, 3, 19, 6, 22, 7, 23,
// Second 128-bit lane.
10, 26, 11, 27, 14, 30, 15, 31}))
- return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
+ return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
- // FIXME: Implement direct support for this type!
- return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
-}
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6, 8, 8, 10, 10,
+ 12, 12, 14, 14}))
+ return DAG.getNode(X86ISD::MOVSLDUP, DL, VT, V1);
+ if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11,
+ 13, 13, 15, 15}))
+ return DAG.getNode(X86ISD::MOVSHDUP, DL, VT, V1);
-/// \brief Handle lowering of 16-lane 32-bit integer shuffles.
-static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
+ SmallVector<int, 4> RepeatedMask;
+ if (is128BitLaneRepeatedShuffleMask(VT, Mask, RepeatedMask)) {
+ if (isSingleInputShuffleMask(Mask)) {
+ unsigned Opc = VT.isInteger() ? X86ISD::PSHUFD : X86ISD::VPERMILPI;
+ return DAG.getNode(Opc, DL, VT, V1,
+ getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
+ }
- // Use dedicated unpack instructions for masks that match their pattern.
- if (isShuffleEquivalent(V1, V2, Mask,
- {// First 128-bit lane.
- 0, 16, 1, 17, 4, 20, 5, 21,
- // Second 128-bit lane.
- 8, 24, 9, 25, 12, 28, 13, 29}))
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
- if (isShuffleEquivalent(V1, V2, Mask,
- {// First 128-bit lane.
- 2, 18, 3, 19, 6, 22, 7, 23,
- // Second 128-bit lane.
- 10, 26, 11, 27, 14, 30, 15, 31}))
- return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
+ for (int i = 0; i < 4; ++i) {
+ if (RepeatedMask[i] >= 16)
+ RepeatedMask[i] -= 12;
+ }
+ return lowerVectorShuffleWithSHUFPS(DL, VT, RepeatedMask, V1, V2, DAG);
+ }
- // FIXME: Implement direct support for this type!
- return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
+ if (SDValue Op = lowerVectorShuffleWithVALIGN(DL, VT, Mask, V1, V2, DAG))
+ return Op;
+
+ return lowerVectorShuffleWithPERMV(DL, VT, Mask, V1, V2, DAG);
}
/// \brief Handle lowering of 32-lane 16-bit integer shuffles.
case MVT::v8i64:
return lowerV8X64VectorShuffle(Op, V1, V2, Subtarget, DAG);
case MVT::v16f32:
- return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
case MVT::v16i32:
- return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ return lowerV16X32VectorShuffle(Op, V1, V2, Subtarget, DAG);
case MVT::v32i16:
if (Subtarget->hasBWI())
return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
DAGCombinerInfo &DCI,
unsigned &RefinementSteps,
bool &UseOneConstNR) const {
- // FIXME: We should use instruction latency models to calculate the cost of
- // each potential sequence, but this is very hard to do reliably because
- // at least Intel's Core* chips have variable timing based on the number of
- // significant digits in the divisor and/or sqrt operand.
- if (!Subtarget->useSqrtEst())
- return SDValue();
-
EVT VT = Op.getValueType();
+ const char *RecipOp;
- // SSE1 has rsqrtss and rsqrtps.
+ // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
// TODO: Add support for AVX512 (v16f32).
// It is likely not profitable to do this for f64 because a double-precision
// rsqrt estimate with refinement on x86 prior to FMA requires at least 16
// instructions: convert to single, rsqrtss, convert back to double, refine
// (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
// along with FMA, this could be a throughput win.
- if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
- (Subtarget->hasAVX() && VT == MVT::v8f32)) {
- RefinementSteps = 1;
- UseOneConstNR = false;
- return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
- }
- return SDValue();
+ if (VT == MVT::f32 && Subtarget->hasSSE1())
+ RecipOp = "sqrtf";
+ else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
+ (VT == MVT::v8f32 && Subtarget->hasAVX()))
+ RecipOp = "vec-sqrtf";
+ else
+ return SDValue();
+
+ TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
+ if (!Recips.isEnabled(RecipOp))
+ return SDValue();
+
+ RefinementSteps = Recips.getRefinementSteps(RecipOp);
+ UseOneConstNR = false;
+ return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
}
/// The minimum architected relative accuracy is 2^-12. We need one
SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
DAGCombinerInfo &DCI,
unsigned &RefinementSteps) const {
- // FIXME: We should use instruction latency models to calculate the cost of
- // each potential sequence, but this is very hard to do reliably because
- // at least Intel's Core* chips have variable timing based on the number of
- // significant digits in the divisor.
- if (!Subtarget->useReciprocalEst())
- return SDValue();
-
EVT VT = Op.getValueType();
-
+ const char *RecipOp;
+
// SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
// TODO: Add support for AVX512 (v16f32).
// It is likely not profitable to do this for f64 because a double-precision
// 15 instructions: convert to single, rcpss, convert back to double, refine
// (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
// along with FMA, this could be a throughput win.
- if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
- (Subtarget->hasAVX() && VT == MVT::v8f32)) {
- RefinementSteps = ReciprocalEstimateRefinementSteps;
- return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
- }
- return SDValue();
+ if (VT == MVT::f32 && Subtarget->hasSSE1())
+ RecipOp = "divf";
+ else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
+ (VT == MVT::v8f32 && Subtarget->hasAVX()))
+ RecipOp = "vec-divf";
+ else
+ return SDValue();
+
+ TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
+ if (!Recips.isEnabled(RecipOp))
+ return SDValue();
+
+ RefinementSteps = Recips.getRefinementSteps(RecipOp);
+ return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
}
/// If we have at least two divisions that use the same divisor, convert to
case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
case X86ISD::VPERMI: return "X86ISD::VPERMI";
case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
+ case X86ISD::VRANGE: return "X86ISD::VRANGE";
case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
case X86ISD::PSADBW: return "X86ISD::PSADBW";