}
}
- // If the this extractelement is directly using a bitcast from a vector of
+ // If this extractelement is directly using a bitcast from a vector of
// the same number of elements, see if we can find the source element from
// it. In this case, we will end up needing to bitcast the scalars.
if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
// Push extractelement into predecessor operation if legal and
- // profitable to do so
+ // profitable to do so.
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
if (I->hasOneUse() &&
cheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
SrcIdx, false));
}
} else if (CastInst *CI = dyn_cast<CastInst>(I)) {
- // Canonicalize extractelement(cast) -> cast(extractelement)
- // bitcasts can change the number of vector elements and they cost nothing
+ // Canonicalize extractelement(cast) -> cast(extractelement).
+ // Bitcasts can change the number of vector elements, and they cost
+ // nothing.
if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
Value *EE = Builder->CreateExtractElement(CI->getOperand(0),
EI.getIndexOperand());
// fight the vectorizer.
// If we are extracting an element from a vector select or a select on
- // vectors, a select on the scalars extracted from the vector arguments.
+ // vectors, create a select on the scalars extracted from the vector
+ // arguments.
Value *TrueVal = SI->getTrueValue();
Value *FalseVal = SI->getFalseValue();
return false;
}
+/// If we have insertion into a vector that is wider than the vector that we
+/// are extracting from, try to widen the source vector to allow a single
+/// shufflevector to replace one or more insert/extract pairs.
+static void replaceExtractElements(InsertElementInst *InsElt,
+ ExtractElementInst *ExtElt,
+ InstCombiner &IC) {
+ VectorType *InsVecType = InsElt->getType();
+ VectorType *ExtVecType = ExtElt->getVectorOperandType();
+ unsigned NumInsElts = InsVecType->getVectorNumElements();
+ unsigned NumExtElts = ExtVecType->getVectorNumElements();
+
+ // The inserted-to vector must be wider than the extracted-from vector.
+ if (InsVecType->getElementType() != ExtVecType->getElementType() ||
+ NumExtElts >= NumInsElts)
+ return;
+
+ // Create a shuffle mask to widen the extended-from vector using undefined
+ // values. The mask selects all of the values of the original vector followed
+ // by as many undefined values as needed to create a vector of the same length
+ // as the inserted-to vector.
+ SmallVector<Constant *, 16> ExtendMask;
+ IntegerType *IntType = Type::getInt32Ty(InsElt->getContext());
+ for (unsigned i = 0; i < NumExtElts; ++i)
+ ExtendMask.push_back(ConstantInt::get(IntType, i));
+ for (unsigned i = NumExtElts; i < NumInsElts; ++i)
+ ExtendMask.push_back(UndefValue::get(IntType));
+
+ Value *ExtVecOp = ExtElt->getVectorOperand();
+ auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
+ ConstantVector::get(ExtendMask));
+
+ // Insert the new shuffle after the vector operand of the extract is defined
+ // (as long as it's not a PHI) or at the start of the basic block of the
+ // extract, so any subsequent extracts in the same basic block can use it.
+ // TODO: Insert before the earliest ExtractElementInst that is replaced.
+ auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
+ if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
+ WideVec->insertAfter(ExtVecOpInst);
+ else
+ IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
+
+ // Replace extracts from the original narrow vector with extracts from the new
+ // wide vector.
+ for (User *U : ExtVecOp->users()) {
+ ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
+ if (!OldExt || OldExt->getParent() != WideVec->getParent())
+ continue;
+ auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
+ NewExt->insertAfter(WideVec);
+ IC.ReplaceInstUsesWith(*OldExt, NewExt);
+ }
+}
/// We are building a shuffle to create V, which is a sequence of insertelement,
/// extractelement pairs. If PermittedRHS is set, then we must either use it or
static ShuffleOps collectShuffleElements(Value *V,
SmallVectorImpl<Constant *> &Mask,
- Value *PermittedRHS) {
+ Value *PermittedRHS,
+ InstCombiner &IC) {
assert(V->getType()->isVectorTy() && "Invalid shuffle!");
unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
// otherwise we'd end up with a shuffle of three inputs.
if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
Value *RHS = EI->getOperand(0);
- ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS);
+ ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC);
assert(LR.second == nullptr || LR.second == RHS);
if (LR.first->getType() != RHS->getType()) {
+ // Although we are giving up for now, see if we can create extracts
+ // that match the inserts for another round of combining.
+ replaceExtractElements(IEI, EI, IC);
+
// We tried our best, but we can't find anything compatible with RHS
// further up the chain. Return a trivial shuffle.
for (unsigned i = 0; i < NumElts; ++i)
}
}
- // Otherwise, can't do anything fancy. Return an identity vector.
+ // Otherwise, we can't do anything fancy. Return an identity vector.
for (unsigned i = 0; i != NumElts; ++i)
Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
return std::make_pair(V, nullptr);
// (and any insertelements it points to), into one big shuffle.
if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
SmallVector<Constant*, 16> Mask;
- ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr);
+ ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this);
// The proposed shuffle may be trivial, in which case we shouldn't
// perform the combine.