[X86] Use APInt instead of SmallBitVector tracking undef elements from getTargetConstantBitsFromNode and getConstVector.

[craig.topper]

Summary

SmallBitVector uses a malloc for more than 58 bits on a 64-bit target and more than 27 bits on a 32-bit target. Some of the vector types we deal with here use more than those number of elements and therefore cause a malloc.

APInt on the other hand supports up to 64 bits without a malloc. That's the maximum number of bits we need here so we can avoid a malloc for all cases by using APInt.

[RKSimon]

LGTM

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

if (Split)
                     DAG.getConstant(0, dl, EltVT));
   }
   SDValue ConstsNode = DAG.getBuildVector(ConstVecVT, dl, Ops);
   if (Split)
     ConstsNode = DAG.getBitcast(VT, ConstsNode);
   return ConstsNode;
 }
 
-static SDValue getConstVector(ArrayRef<APInt> Bits, SmallBitVector &Undefs,
+static SDValue getConstVector(ArrayRef<APInt> Bits, APInt &Undefs,
                               MVT VT, SelectionDAG &DAG, const SDLoc &dl) {
-  assert(Bits.size() == Undefs.size() && "Unequal constant and undef arrays");
+  assert(Bits.size() == Undefs.getBitWidth() &&
+         "Unequal constant and undef arrays");
   SmallVector<SDValue, 32> Ops;
   bool Split = false;
 
   MVT ConstVecVT = VT;
   unsigned NumElts = VT.getVectorNumElements();
   bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
   if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
     ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
static const Constant *getTargetConstantFromNode(SDValue Op) {
   if (!CNode || CNode->isMachineConstantPoolEntry())
     return nullptr;
 
   return dyn_cast<Constant>(CNode->getConstVal());
 }
 
 // Extract raw constant bits from constant pools.
 static bool getTargetConstantBitsFromNode(SDValue Op, unsigned EltSizeInBits,
-                                          SmallBitVector &UndefElts,
+                                          APInt &UndefElts,
                                           SmallVectorImpl<APInt> &EltBits,
                                           bool AllowWholeUndefs = true,
                                           bool AllowPartialUndefs = true) {
-  assert(UndefElts.empty() && "Expected an empty UndefElts vector");
   assert(EltBits.empty() && "Expected an empty EltBits vector");
 
   Op = peekThroughBitcasts(Op);
 
   EVT VT = Op.getValueType();
   unsigned SizeInBits = VT.getSizeInBits();
   assert((SizeInBits % EltSizeInBits) == 0 && "Can't split constant!");
   unsigned NumElts = SizeInBits / EltSizeInBits;
 
   unsigned SrcEltSizeInBits = VT.getScalarSizeInBits();
   unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits;
 
   // Extract all the undef/constant element data and pack into single bitsets.
   APInt UndefBits(SizeInBits, 0);
   APInt MaskBits(SizeInBits, 0);
 
   // Split the undef/constant single bitset data into the target elements.
   auto SplitBitData = [&]() {
     // Don't split if we don't allow undef bits.
     bool AllowUndefs = AllowWholeUndefs || AllowPartialUndefs;
     if (UndefBits.getBoolValue() && !AllowUndefs)
       return false;
 
-    UndefElts = SmallBitVector(NumElts, false);
+    UndefElts = APInt(NumElts, 0);
     EltBits.resize(NumElts, APInt(EltSizeInBits, 0));
 
     for (unsigned i = 0; i != NumElts; ++i) {
       unsigned BitOffset = i * EltSizeInBits;
       APInt UndefEltBits = UndefBits.extractBits(EltSizeInBits, BitOffset);
 
       // Only treat an element as UNDEF if all bits are UNDEF.
       if (UndefEltBits.isAllOnesValue()) {
         if (!AllowWholeUndefs)
           return false;
-        UndefElts[i] = true;
+        UndefElts.setBit(i);
         continue;
       }
 
       // If only some bits are UNDEF then treat them as zero (or bail if not
       // supported).
       if (UndefEltBits.getBoolValue() && !AllowPartialUndefs)
         return false;
 
static bool getTargetConstantBitsFromNode(SDValue Op, unsigned EltSizeInBits,
   }
 
   return false;
 }
 
 static bool getTargetShuffleMaskIndices(SDValue MaskNode,
                                         unsigned MaskEltSizeInBits,
                                         SmallVectorImpl<uint64_t> &RawMask) {
-  SmallBitVector UndefElts;
+  APInt UndefElts;
   SmallVector<APInt, 64> EltBits;
 
   // Extract the raw target constant bits.
   // FIXME: We currently don't support UNDEF bits or mask entries.
   if (!getTargetConstantBitsFromNode(MaskNode, MaskEltSizeInBits, UndefElts,
                                      EltBits, /* AllowWholeUndefs */ false,
                                      /* AllowPartialUndefs */ false))
     return false;
static bool setTargetShuffleZeroElements(SDValue N,
   V1 = peekThroughBitcasts(V1);
   V2 = peekThroughBitcasts(V2);
 
   assert((VT.getSizeInBits() % Mask.size()) == 0 &&
          "Illegal split of shuffle value type");
   unsigned EltSizeInBits = VT.getSizeInBits() / Mask.size();
 
   // Extract known constant input data.
-  SmallBitVector UndefSrcElts[2];
+  APInt UndefSrcElts[2];
   SmallVector<APInt, 32> SrcEltBits[2];
   bool IsSrcConstant[2] = {
       getTargetConstantBitsFromNode(V1, EltSizeInBits, UndefSrcElts[0],
                                     SrcEltBits[0], true, false),
       getTargetConstantBitsFromNode(V2, EltSizeInBits, UndefSrcElts[1],
                                     SrcEltBits[1], true, false)};
 
   for (int i = 0, Size = Mask.size(); i < Size; ++i) {
static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask,
   assert((NumBitsPerElt % 8) == 0 && (NumSizeInBits % 8) == 0 &&
          "Expected byte aligned value types");
 
   unsigned Opcode = N.getOpcode();
   switch (Opcode) {
   case ISD::AND:
   case X86ISD::ANDNP: {
     // Attempt to decode as a per-byte mask.
-    SmallBitVector UndefElts;
+    APInt UndefElts;
     SmallVector<APInt, 32> EltBits;
     SDValue N0 = N.getOperand(0);
     SDValue N1 = N.getOperand(1);
     bool IsAndN = (X86ISD::ANDNP == Opcode);
     uint64_t ZeroMask = IsAndN ? 255 : 0;
     if (!getTargetConstantBitsFromNode(IsAndN ? N0 : N1, 8, UndefElts, EltBits))
       return false;
     for (int i = 0, e = (int)EltBits.size(); i != e; ++i) {
static bool combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root,
 
   // See if we can combine a single input shuffle with zeros to a bit-mask,
   // which is much simpler than any shuffle.
   if (UnaryShuffle && MaskContainsZeros && (Depth >= 3 || HasVariableMask) &&
       isSequentialOrUndefOrZeroInRange(Mask, 0, NumMaskElts, 0) &&
       DAG.getTargetLoweringInfo().isTypeLegal(MaskVT)) {
     APInt Zero = APInt::getNullValue(MaskEltSizeInBits);
     APInt AllOnes = APInt::getAllOnesValue(MaskEltSizeInBits);
-    SmallBitVector UndefElts(NumMaskElts, false);
+    APInt UndefElts(NumMaskElts, 0);
     SmallVector<APInt, 64> EltBits(NumMaskElts, Zero);
     for (unsigned i = 0; i != NumMaskElts; ++i) {
       int M = Mask[i];
       if (M == SM_SentinelUndef) {
-        UndefElts[i] = true;
+        UndefElts.setBit(i);
         continue;
       }
       if (M == SM_SentinelZero)
         continue;
       EltBits[i] = AllOnes;
     }
     SDValue BitMask = getConstVector(EltBits, UndefElts, MaskVT, DAG, DL);
     DCI.AddToWorklist(BitMask.getNode());
static bool combineX86ShufflesConstants(const SmallVectorImpl<SDValue> &Ops,
 
   unsigned SizeInBits = VT.getSizeInBits();
   unsigned NumMaskElts = Mask.size();
   unsigned MaskSizeInBits = SizeInBits / NumMaskElts;
   unsigned NumOps = Ops.size();
 
   // Extract constant bits from each source op.
   bool OneUseConstantOp = false;
-  SmallVector<SmallBitVector, 16> UndefEltsOps(NumOps);
+  SmallVector<APInt, 16> UndefEltsOps(NumOps);
   SmallVector<SmallVector<APInt, 16>, 16> RawBitsOps(NumOps);
   for (unsigned i = 0; i != NumOps; ++i) {
     SDValue SrcOp = Ops[i];
     OneUseConstantOp |= SrcOp.hasOneUse();
     if (!getTargetConstantBitsFromNode(SrcOp, MaskSizeInBits, UndefEltsOps[i],
                                        RawBitsOps[i]))
       return false;
   }
 
   // Only fold if at least one of the constants is only used once or
   // the combined shuffle has included a variable mask shuffle, this
   // is to avoid constant pool bloat.
   if (!OneUseConstantOp && !HasVariableMask)
     return false;
 
   // Shuffle the constant bits according to the mask.
-  SmallBitVector UndefElts(NumMaskElts, false);
-  SmallBitVector ZeroElts(NumMaskElts, false);
-  SmallBitVector ConstantElts(NumMaskElts, false);
+  APInt UndefElts(NumMaskElts, 0);
+  APInt ZeroElts(NumMaskElts, 0);
+  APInt ConstantElts(NumMaskElts, 0);
   SmallVector<APInt, 8> ConstantBitData(NumMaskElts,
                                         APInt::getNullValue(MaskSizeInBits));
   for (unsigned i = 0; i != NumMaskElts; ++i) {
     int M = Mask[i];
     if (M == SM_SentinelUndef) {
-      UndefElts[i] = true;
+      UndefElts.setBit(i);
       continue;
     } else if (M == SM_SentinelZero) {
-      ZeroElts[i] = true;
+      ZeroElts.setBit(i);
       continue;
     }
     assert(0 <= M && M < (int)(NumMaskElts * NumOps));
 
     unsigned SrcOpIdx = (unsigned)M / NumMaskElts;
     unsigned SrcMaskIdx = (unsigned)M % NumMaskElts;
 
     auto &SrcUndefElts = UndefEltsOps[SrcOpIdx];
     if (SrcUndefElts[SrcMaskIdx]) {
-      UndefElts[i] = true;
+      UndefElts.setBit(i);
       continue;
     }
 
     auto &SrcEltBits = RawBitsOps[SrcOpIdx];
     APInt &Bits = SrcEltBits[SrcMaskIdx];
     if (!Bits) {
-      ZeroElts[i] = true;
+      ZeroElts.setBit(i);
       continue;
     }
 
-    ConstantElts[i] = true;
+    ConstantElts.setBit(i);
     ConstantBitData[i] = Bits;
   }
-  assert((UndefElts | ZeroElts | ConstantElts).count() == NumMaskElts);
+  assert((UndefElts | ZeroElts | ConstantElts).isAllOnesValue());
 
   // Create the constant data.
   MVT MaskSVT;
   if (VT.isFloatingPoint() && (MaskSizeInBits == 32 || MaskSizeInBits == 64))
     MaskSVT = MVT::getFloatingPointVT(MaskSizeInBits);
   else
     MaskSVT = MVT::getIntegerVT(MaskSizeInBits);
 
if (LogicalShift && (ShiftVal.getZExtValue() % 8) == 0) {
     NonceMask.push_back(0);
     if (combineX86ShufflesRecursively({Op}, 0, Op, NonceMask, {},
                                       /*Depth*/ 1, /*HasVarMask*/ false, DAG,
                                       DCI, Subtarget))
       return SDValue(); // This routine will use CombineTo to replace N.
   }
 
   // Constant Folding.
-  SmallBitVector UndefElts;
+  APInt UndefElts;
   SmallVector<APInt, 32> EltBits;
   if (N->isOnlyUserOf(N0.getNode()) &&
       getTargetConstantBitsFromNode(N0, NumBitsPerElt, UndefElts, EltBits)) {
     assert(EltBits.size() == VT.getVectorNumElements() &&
            "Unexpected shift value type");
     unsigned ShiftImm = ShiftVal.getZExtValue();
     for (APInt &Elt : EltBits) {
       if (X86ISD::VSHLI == Opcode)
static SDValue combineVSZext(SDNode *N, SelectionDAG &DAG,
   SDValue Op = N->getOperand(0);
   MVT OpVT = Op.getSimpleValueType();
   MVT OpEltVT = OpVT.getVectorElementType();
   unsigned OpEltSizeInBits = OpEltVT.getSizeInBits();
   unsigned InputBits = OpEltSizeInBits * NumElts;
 
   // Perform any constant folding.
   // FIXME: Reduce constant pool usage and don't fold when OptSize is enabled.
-  SmallBitVector UndefElts;
+  APInt UndefElts;
   SmallVector<APInt, 64> EltBits;
   if (getTargetConstantBitsFromNode(Op, OpEltSizeInBits, UndefElts, EltBits)) {
-    SmallBitVector Undefs(NumElts, false);
+    APInt Undefs(NumElts, 0);
     SmallVector<APInt, 4> Vals(NumElts, APInt(EltSizeInBits, 0));
     bool IsZEXT =
         (Opcode == X86ISD::VZEXT) || (Opcode == ISD::ZERO_EXTEND_VECTOR_INREG);
     for (unsigned i = 0; i != NumElts; ++i) {
       if (UndefElts[i]) {
-        Undefs[i] = true;
+        Undefs.setBit(i);
         continue;
       }
       Vals[i] = IsZEXT ? EltBits[i].zextOrTrunc(EltSizeInBits)
                        : EltBits[i].sextOrTrunc(EltSizeInBits);
     }
     return getConstVector(Vals, Undefs, VT, DAG, DL);
   }