diff options
Diffstat (limited to 'lib/Transforms/InstCombine/InstCombineVectorOps.cpp')
| -rw-r--r-- | lib/Transforms/InstCombine/InstCombineVectorOps.cpp | 338 |
1 files changed, 334 insertions, 4 deletions
diff --git a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp index 4f71db1..d43093d 100644 --- a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp +++ b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp @@ -105,6 +105,73 @@ static Value *FindScalarElement(Value *V, unsigned EltNo) { return 0; } +// If we have a PHI node with a vector type that has only 2 uses: feed +// itself and be an operand of extractelemnt at a constant location, +// try to replace the PHI of the vector type with a PHI of a scalar type +Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) { + // Verify that the PHI node has exactly 2 uses. Otherwise return NULL. + if (!PN->hasNUses(2)) + return NULL; + + // If so, it's known at this point that one operand is PHI and the other is + // an extractelement node. Find the PHI user that is not the extractelement + // node. + Value::use_iterator iu = PN->use_begin(); + Instruction *PHIUser = dyn_cast<Instruction>(*iu); + if (PHIUser == cast<Instruction>(&EI)) + PHIUser = cast<Instruction>(*(++iu)); + + // Verify that this PHI user has one use, which is the PHI itself, + // and that it is a binary operation which is cheap to scalarize. + // otherwise return NULL. + if (!PHIUser->hasOneUse() || !(PHIUser->use_back() == PN) || + !(isa<BinaryOperator>(PHIUser)) || !CheapToScalarize(PHIUser, true)) + return NULL; + + // Create a scalar PHI node that will replace the vector PHI node + // just before the current PHI node. + PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith( + PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN)); + // Scalarize each PHI operand. + for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) { + Value *PHIInVal = PN->getIncomingValue(i); + BasicBlock *inBB = PN->getIncomingBlock(i); + Value *Elt = EI.getIndexOperand(); + // If the operand is the PHI induction variable: + if (PHIInVal == PHIUser) { + // Scalarize the binary operation. Its first operand is the + // scalar PHI and the second operand is extracted from the other + // vector operand. + BinaryOperator *B0 = cast<BinaryOperator>(PHIUser); + unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0; + Value *Op = InsertNewInstWith( + ExtractElementInst::Create(B0->getOperand(opId), Elt, + B0->getOperand(opId)->getName() + ".Elt"), + *B0); + Value *newPHIUser = InsertNewInstWith( + BinaryOperator::Create(B0->getOpcode(), scalarPHI, Op), *B0); + scalarPHI->addIncoming(newPHIUser, inBB); + } else { + // Scalarize PHI input: + Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, ""); + // Insert the new instruction into the predecessor basic block. + Instruction *pos = dyn_cast<Instruction>(PHIInVal); + BasicBlock::iterator InsertPos; + if (pos && !isa<PHINode>(pos)) { + InsertPos = pos; + ++InsertPos; + } else { + InsertPos = inBB->getFirstInsertionPt(); + } + + InsertNewInstWith(newEI, *InsertPos); + + scalarPHI->addIncoming(newEI, inBB); + } + } + return ReplaceInstUsesWith(EI, scalarPHI); +} + Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) { // If vector val is constant with all elements the same, replace EI with // that element. We handle a known element # below. @@ -149,6 +216,14 @@ Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) { if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal)) return new BitCastInst(Elt, EI.getType()); } + + // If there's a vector PHI feeding a scalar use through this extractelement + // instruction, try to scalarize the PHI. + if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) { + Instruction *scalarPHI = scalarizePHI(EI, PN); + if (scalarPHI) + return scalarPHI; + } } if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) { @@ -201,10 +276,10 @@ Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) { } 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 - if (CI->hasOneUse() && EI.hasOneUse() && - (CI->getOpcode() != Instruction::BitCast)) { + if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) { Value *EE = Builder->CreateExtractElement(CI->getOperand(0), EI.getIndexOperand()); + Worklist.AddValue(EE); return CastInst::Create(CI->getOpcode(), EE, EI.getType()); } } @@ -336,6 +411,10 @@ static Value *CollectShuffleElements(Value *V, SmallVectorImpl<Constant*> &Mask, if (VecOp == RHS) { Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS); + // Update Mask to reflect that `ScalarOp' has been inserted at + // position `InsertedIdx' within the vector returned by IEI. + Mask[InsertedIdx % NumElts] = Mask[ExtractedIdx]; + // Everything but the extracted element is replaced with the RHS. for (unsigned i = 0; i != NumElts; ++i) { if (i != InsertedIdx) @@ -415,6 +494,252 @@ Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) { return 0; } +/// Return true if we can evaluate the specified expression tree if the vector +/// elements were shuffled in a different order. +static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask, + unsigned Depth = 5) { + // We can always reorder the elements of a constant. + if (isa<Constant>(V)) + return true; + + // We won't reorder vector arguments. No IPO here. + Instruction *I = dyn_cast<Instruction>(V); + if (!I) return false; + + // Two users may expect different orders of the elements. Don't try it. + if (!I->hasOneUse()) + return false; + + if (Depth == 0) return false; + + switch (I->getOpcode()) { + case Instruction::Add: + case Instruction::FAdd: + case Instruction::Sub: + case Instruction::FSub: + case Instruction::Mul: + case Instruction::FMul: + case Instruction::UDiv: + case Instruction::SDiv: + case Instruction::FDiv: + case Instruction::URem: + case Instruction::SRem: + case Instruction::FRem: + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + case Instruction::ICmp: + case Instruction::FCmp: + case Instruction::Trunc: + case Instruction::ZExt: + case Instruction::SExt: + case Instruction::FPToUI: + case Instruction::FPToSI: + case Instruction::UIToFP: + case Instruction::SIToFP: + case Instruction::FPTrunc: + case Instruction::FPExt: + case Instruction::GetElementPtr: { + for (int i = 0, e = I->getNumOperands(); i != e; ++i) { + if (!CanEvaluateShuffled(I->getOperand(i), Mask, Depth-1)) + return false; + } + return true; + } + case Instruction::InsertElement: { + ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2)); + if (!CI) return false; + int ElementNumber = CI->getLimitedValue(); + + // Verify that 'CI' does not occur twice in Mask. A single 'insertelement' + // can't put an element into multiple indices. + bool SeenOnce = false; + for (int i = 0, e = Mask.size(); i != e; ++i) { + if (Mask[i] == ElementNumber) { + if (SeenOnce) + return false; + SeenOnce = true; + } + } + return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1); + } + } + return false; +} + +/// Rebuild a new instruction just like 'I' but with the new operands given. +/// In the event of type mismatch, the type of the operands is correct. +static Value *BuildNew(Instruction *I, ArrayRef<Value*> NewOps) { + // We don't want to use the IRBuilder here because we want the replacement + // instructions to appear next to 'I', not the builder's insertion point. + switch (I->getOpcode()) { + case Instruction::Add: + case Instruction::FAdd: + case Instruction::Sub: + case Instruction::FSub: + case Instruction::Mul: + case Instruction::FMul: + case Instruction::UDiv: + case Instruction::SDiv: + case Instruction::FDiv: + case Instruction::URem: + case Instruction::SRem: + case Instruction::FRem: + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: { + BinaryOperator *BO = cast<BinaryOperator>(I); + assert(NewOps.size() == 2 && "binary operator with #ops != 2"); + BinaryOperator *New = + BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(), + NewOps[0], NewOps[1], "", BO); + if (isa<OverflowingBinaryOperator>(BO)) { + New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap()); + New->setHasNoSignedWrap(BO->hasNoSignedWrap()); + } + if (isa<PossiblyExactOperator>(BO)) { + New->setIsExact(BO->isExact()); + } + return New; + } + case Instruction::ICmp: + assert(NewOps.size() == 2 && "icmp with #ops != 2"); + return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(), + NewOps[0], NewOps[1]); + case Instruction::FCmp: + assert(NewOps.size() == 2 && "fcmp with #ops != 2"); + return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(), + NewOps[0], NewOps[1]); + case Instruction::Trunc: + case Instruction::ZExt: + case Instruction::SExt: + case Instruction::FPToUI: + case Instruction::FPToSI: + case Instruction::UIToFP: + case Instruction::SIToFP: + case Instruction::FPTrunc: + case Instruction::FPExt: { + // It's possible that the mask has a different number of elements from + // the original cast. We recompute the destination type to match the mask. + Type *DestTy = + VectorType::get(I->getType()->getScalarType(), + NewOps[0]->getType()->getVectorNumElements()); + assert(NewOps.size() == 1 && "cast with #ops != 1"); + return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy, + "", I); + } + case Instruction::GetElementPtr: { + Value *Ptr = NewOps[0]; + ArrayRef<Value*> Idx = NewOps.slice(1); + GetElementPtrInst *GEP = GetElementPtrInst::Create(Ptr, Idx, "", I); + GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds()); + return GEP; + } + } + llvm_unreachable("failed to rebuild vector instructions"); +} + +Value * +InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) { + // Mask.size() does not need to be equal to the number of vector elements. + + assert(V->getType()->isVectorTy() && "can't reorder non-vector elements"); + if (isa<UndefValue>(V)) { + return UndefValue::get(VectorType::get(V->getType()->getScalarType(), + Mask.size())); + } + if (isa<ConstantAggregateZero>(V)) { + return ConstantAggregateZero::get( + VectorType::get(V->getType()->getScalarType(), + Mask.size())); + } + if (Constant *C = dyn_cast<Constant>(V)) { + SmallVector<Constant *, 16> MaskValues; + for (int i = 0, e = Mask.size(); i != e; ++i) { + if (Mask[i] == -1) + MaskValues.push_back(UndefValue::get(Builder->getInt32Ty())); + else + MaskValues.push_back(Builder->getInt32(Mask[i])); + } + return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()), + ConstantVector::get(MaskValues)); + } + + Instruction *I = cast<Instruction>(V); + switch (I->getOpcode()) { + case Instruction::Add: + case Instruction::FAdd: + case Instruction::Sub: + case Instruction::FSub: + case Instruction::Mul: + case Instruction::FMul: + case Instruction::UDiv: + case Instruction::SDiv: + case Instruction::FDiv: + case Instruction::URem: + case Instruction::SRem: + case Instruction::FRem: + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + case Instruction::ICmp: + case Instruction::FCmp: + case Instruction::Trunc: + case Instruction::ZExt: + case Instruction::SExt: + case Instruction::FPToUI: + case Instruction::FPToSI: + case Instruction::UIToFP: + case Instruction::SIToFP: + case Instruction::FPTrunc: + case Instruction::FPExt: + case Instruction::Select: + case Instruction::GetElementPtr: { + SmallVector<Value*, 8> NewOps; + bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements()); + for (int i = 0, e = I->getNumOperands(); i != e; ++i) { + Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask); + NewOps.push_back(V); + NeedsRebuild |= (V != I->getOperand(i)); + } + if (NeedsRebuild) { + return BuildNew(I, NewOps); + } + return I; + } + case Instruction::InsertElement: { + int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue(); + + // The insertelement was inserting at Element. Figure out which element + // that becomes after shuffling. The answer is guaranteed to be unique + // by CanEvaluateShuffled. + bool Found = false; + int Index = 0; + for (int e = Mask.size(); Index != e; ++Index) { + if (Mask[Index] == Element) { + Found = true; + break; + } + } + + if (!Found) + return UndefValue::get(I->getType()); + Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask); + return InsertElementInst::Create(V, I->getOperand(1), + Builder->getInt32(Index), "", I); + } + } + llvm_unreachable("failed to reorder elements of vector instruction!"); +} Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) { Value *LHS = SVI.getOperand(0); @@ -446,9 +771,9 @@ Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) { if (LHS == RHS || isa<UndefValue>(LHS)) { if (isa<UndefValue>(LHS) && LHS == RHS) { // shuffle(undef,undef,mask) -> undef. - Value* result = (VWidth == LHSWidth) + Value *Result = (VWidth == LHSWidth) ? LHS : UndefValue::get(SVI.getType()); - return ReplaceInstUsesWith(SVI, result); + return ReplaceInstUsesWith(SVI, Result); } // Remap any references to RHS to use LHS. @@ -495,6 +820,11 @@ Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) { if (isRHSID) return ReplaceInstUsesWith(SVI, RHS); } + if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) { + Value *V = EvaluateInDifferentElementOrder(LHS, Mask); + return ReplaceInstUsesWith(SVI, V); + } + // If the LHS is a shufflevector itself, see if we can combine it with this // one without producing an unusual shuffle. // Cases that might be simplified: |