//===----------------------------------------------------------------------===//
#include "InstCombine.h"
-#include "llvm/Support/PatternMatch.h"
+#include "llvm/IR/PatternMatch.h"
using namespace llvm;
using namespace PatternMatch;
+#define DEBUG_TYPE "instcombine"
+
/// CheapToScalarize - Return true if the value is cheaper to scalarize than it
/// is to leave as a vector operation. isConstant indicates whether we're
/// extracting one known element. If false we're extracting a variable index.
if (isConstant) return true;
// If all elts are the same, we can extract it and use any of the values.
- Constant *Op0 = C->getAggregateElement(0U);
- for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e; ++i)
- if (C->getAggregateElement(i) != Op0)
- return false;
- return true;
+ if (Constant *Op0 = C->getAggregateElement(0U)) {
+ for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e;
+ ++i)
+ if (C->getAggregateElement(i) != Op0)
+ return false;
+ return true;
+ }
}
Instruction *I = dyn_cast<Instruction>(V);
if (!I) return false;
if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
// If this is an insert to a variable element, we don't know what it is.
if (!isa<ConstantInt>(III->getOperand(2)))
- return 0;
+ return nullptr;
unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
// If this is an insert to the element we are looking for, return the
}
// Extract a value from a vector add operation with a constant zero.
- Value *Val = 0; Constant *Con = 0;
+ Value *Val = nullptr; Constant *Con = nullptr;
if (match(V, m_Add(m_Value(Val), m_Constant(Con)))) {
if (Con->getAggregateElement(EltNo)->isNullValue())
return FindScalarElement(Val, EltNo);
}
// Otherwise, we don't know.
- return 0;
+ return nullptr;
}
// 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
+// itself and be an operand of extractelement 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;
+ return nullptr;
// 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();
+ auto iu = PN->user_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) ||
+ if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) ||
!(isa<BinaryOperator>(PHIUser)) || !CheapToScalarize(PHIUser, true))
- return NULL;
+ return nullptr;
// Create a scalar PHI node that will replace the vector PHI node
// just before the current PHI node.
// 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
+ // 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;
Worklist.AddValue(EE);
return CastInst::Create(CI->getOpcode(), EE, EI.getType());
}
+ } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
+ if (SI->hasOneUse()) {
+ // TODO: For a select on vectors, it might be useful to do this if it
+ // has multiple extractelement uses. For vector select, that seems to
+ // 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.
+ Value *TrueVal = SI->getTrueValue();
+ Value *FalseVal = SI->getFalseValue();
+
+ Value *Cond = SI->getCondition();
+ if (Cond->getType()->isVectorTy()) {
+ Cond = Builder->CreateExtractElement(Cond,
+ EI.getIndexOperand(),
+ Cond->getName() + ".elt");
+ }
+
+ Value *V1Elem
+ = Builder->CreateExtractElement(TrueVal,
+ EI.getIndexOperand(),
+ TrueVal->getName() + ".elt");
+
+ Value *V2Elem
+ = Builder->CreateExtractElement(FalseVal,
+ EI.getIndexOperand(),
+ FalseVal->getName() + ".elt");
+ return SelectInst::Create(Cond,
+ V1Elem,
+ V2Elem,
+ SI->getName() + ".elt");
+ }
}
}
- return 0;
+ return nullptr;
}
/// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
/// Otherwise, return false.
static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
SmallVectorImpl<Constant*> &Mask) {
- assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
+ assert(LHS->getType() == RHS->getType() &&
"Invalid CollectSingleShuffleElements");
- unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
+ unsigned NumElts = V->getType()->getVectorNumElements();
if (isa<UndefValue>(V)) {
Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
- // Okay, we can handle this if the vector we are insertinting into is
+ // We can handle this if the vector we are inserting into is
// transitively ok.
if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
// If so, update the mask to reflect the inserted undef.
return true;
}
} else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
- if (isa<ConstantInt>(EI->getOperand(1)) &&
- EI->getOperand(0)->getType() == V->getType()) {
+ if (isa<ConstantInt>(EI->getOperand(1))) {
unsigned ExtractedIdx =
cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
+ unsigned NumLHSElts = LHS->getType()->getVectorNumElements();
// This must be extracting from either LHS or RHS.
if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
- // Okay, we can handle this if the vector we are insertinting into is
+ // We can handle this if the vector we are inserting into is
// transitively ok.
if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
// If so, update the mask to reflect the inserted value.
assert(EI->getOperand(0) == RHS);
Mask[InsertedIdx % NumElts] =
ConstantInt::get(Type::getInt32Ty(V->getContext()),
- ExtractedIdx+NumElts);
+ ExtractedIdx + NumLHSElts);
}
return true;
}
}
}
}
- // TODO: Handle shufflevector here!
return false;
}
-/// CollectShuffleElements - We are building a shuffle of V, using RHS as the
-/// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
-/// that computes V and the LHS value of the shuffle.
-static Value *CollectShuffleElements(Value *V, SmallVectorImpl<Constant*> &Mask,
- Value *&RHS) {
- assert(V->getType()->isVectorTy() &&
- (RHS == 0 || V->getType() == RHS->getType()) &&
- "Invalid shuffle!");
+
+/// 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
+/// not rely on the second vector source. Return a std::pair containing the
+/// left and right vectors of the proposed shuffle (or 0), and set the Mask
+/// parameter as required.
+///
+/// Note: we intentionally don't try to fold earlier shuffles since they have
+/// often been chosen carefully to be efficiently implementable on the target.
+typedef std::pair<Value *, Value *> ShuffleOps;
+
+static ShuffleOps CollectShuffleElements(Value *V,
+ SmallVectorImpl<Constant *> &Mask,
+ Value *PermittedRHS) {
+ assert(V->getType()->isVectorTy() && "Invalid shuffle!");
unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
if (isa<UndefValue>(V)) {
Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
- return V;
+ return std::make_pair(
+ PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
}
if (isa<ConstantAggregateZero>(V)) {
Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
- return V;
+ return std::make_pair(V, nullptr);
}
if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
Value *IdxOp = IEI->getOperand(2);
if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
- if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
- EI->getOperand(0)->getType() == V->getType()) {
+ if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
unsigned ExtractedIdx =
cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
// Either the extracted from or inserted into vector must be RHSVec,
// otherwise we'd end up with a shuffle of three inputs.
- if (EI->getOperand(0) == RHS || RHS == 0) {
- RHS = EI->getOperand(0);
- Value *V = CollectShuffleElements(VecOp, Mask, RHS);
+ if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
+ Value *RHS = EI->getOperand(0);
+ ShuffleOps LR = CollectShuffleElements(VecOp, Mask, RHS);
+ assert(LR.second == nullptr || LR.second == RHS);
+
+ if (LR.first->getType() != RHS->getType()) {
+ // 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)
+ Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
+ return std::make_pair(V, nullptr);
+ }
+
+ unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
Mask[InsertedIdx % NumElts] =
ConstantInt::get(Type::getInt32Ty(V->getContext()),
- NumElts+ExtractedIdx);
- return V;
+ NumLHSElts+ExtractedIdx);
+ return std::make_pair(LR.first, RHS);
}
- 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)
- Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()),
- NumElts+i);
- }
- return V;
+ if (VecOp == PermittedRHS) {
+ // We've gone as far as we can: anything on the other side of the
+ // extractelement will already have been converted into a shuffle.
+ unsigned NumLHSElts =
+ EI->getOperand(0)->getType()->getVectorNumElements();
+ for (unsigned i = 0; i != NumElts; ++i)
+ Mask.push_back(ConstantInt::get(
+ Type::getInt32Ty(V->getContext()),
+ i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
+ return std::make_pair(EI->getOperand(0), PermittedRHS);
}
// If this insertelement is a chain that comes from exactly these two
// vectors, return the vector and the effective shuffle.
- if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
- return EI->getOperand(0);
+ if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
+ CollectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
+ Mask))
+ return std::make_pair(EI->getOperand(0), PermittedRHS);
}
}
}
- // TODO: Handle shufflevector here!
// Otherwise, 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 V;
+ return std::make_pair(V, nullptr);
+}
+
+/// Try to find redundant insertvalue instructions, like the following ones:
+/// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
+/// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
+/// Here the second instruction inserts values at the same indices, as the
+/// first one, making the first one redundant.
+/// It should be transformed to:
+/// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
+Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
+ bool IsRedundant = false;
+ ArrayRef<unsigned int> FirstIndices = I.getIndices();
+
+ // If there is a chain of insertvalue instructions (each of them except the
+ // last one has only one use and it's another insertvalue insn from this
+ // chain), check if any of the 'children' uses the same indices as the first
+ // instruction. In this case, the first one is redundant.
+ Value *V = &I;
+ unsigned Depth = 0;
+ while (V->hasOneUse() && Depth < 10) {
+ User *U = V->user_back();
+ auto UserInsInst = dyn_cast<InsertValueInst>(U);
+ if (!UserInsInst || U->getOperand(0) != V)
+ break;
+ if (UserInsInst->getIndices() == FirstIndices) {
+ IsRedundant = true;
+ break;
+ }
+ V = UserInsInst;
+ Depth++;
+ }
+
+ if (IsRedundant)
+ return ReplaceInstUsesWith(I, I.getOperand(0));
+ return nullptr;
}
Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
// If the inserted element was extracted from some other vector, and if the
// indexes are constant, try to turn this into a shufflevector operation.
if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
- if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
- EI->getOperand(0)->getType() == IE.getType()) {
- unsigned NumVectorElts = IE.getType()->getNumElements();
+ if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
+ unsigned NumInsertVectorElts = IE.getType()->getNumElements();
+ unsigned NumExtractVectorElts =
+ EI->getOperand(0)->getType()->getVectorNumElements();
unsigned ExtractedIdx =
cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
- if (ExtractedIdx >= NumVectorElts) // Out of range extract.
+ if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
return ReplaceInstUsesWith(IE, VecOp);
- if (InsertedIdx >= NumVectorElts) // Out of range insert.
+ if (InsertedIdx >= NumInsertVectorElts) // Out of range insert.
return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
// If we are extracting a value from a vector, then inserting it right
// If this insertelement isn't used by some other insertelement, turn it
// (and any insertelements it points to), into one big shuffle.
- if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
+ if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
SmallVector<Constant*, 16> Mask;
- Value *RHS = 0;
- Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
- if (RHS == 0) RHS = UndefValue::get(LHS->getType());
- // We now have a shuffle of LHS, RHS, Mask.
- return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
+ ShuffleOps LR = CollectShuffleElements(&IE, Mask, nullptr);
+
+ // The proposed shuffle may be trivial, in which case we shouldn't
+ // perform the combine.
+ if (LR.first != &IE && LR.second != &IE) {
+ // We now have a shuffle of LHS, RHS, Mask.
+ if (LR.second == nullptr)
+ LR.second = UndefValue::get(LR.first->getType());
+ return new ShuffleVectorInst(LR.first, LR.second,
+ ConstantVector::get(Mask));
+ }
}
}
}
return &IE;
}
- return 0;
+ return nullptr;
}
/// 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 = 100) {
+ unsigned Depth = 5) {
// We can always reorder the elements of a constant.
if (isa<Constant>(V))
return true;
if (isa<PossiblyExactOperator>(BO)) {
New->setIsExact(BO->isExact());
}
+ if (isa<FPMathOperator>(BO))
+ New->copyFastMathFlags(I);
return New;
}
case Instruction::ICmp:
}
case Instruction::InsertElement: {
int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
- if (Element < 0 || Element >= (int)Mask.size()) {
- // Such instructions are valid and exhibit undefined behaviour.
- return UndefValue::get(I->getType());
- }
// 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)
+ for (int e = Mask.size(); Index != e; ++Index) {
+ if (Mask[Index] == Element) {
+ Found = true;
break;
+ }
+ }
+
+ // If element is not in Mask, no need to handle the operand 1 (element to
+ // be inserted). Just evaluate values in operand 0 according to Mask.
+ if (!Found)
+ return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
return InsertElementInst::Create(V, I->getOperand(1),
llvm_unreachable("failed to reorder elements of vector instruction!");
}
+static void RecognizeIdentityMask(const SmallVectorImpl<int> &Mask,
+ bool &isLHSID, bool &isRHSID) {
+ isLHSID = isRHSID = true;
+
+ for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
+ if (Mask[i] < 0) continue; // Ignore undef values.
+ // Is this an identity shuffle of the LHS value?
+ isLHSID &= (Mask[i] == (int)i);
+
+ // Is this an identity shuffle of the RHS value?
+ isRHSID &= (Mask[i]-e == i);
+ }
+}
+
Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
Value *LHS = SVI.getOperand(0);
Value *RHS = SVI.getOperand(1);
if (VWidth == LHSWidth) {
// Analyze the shuffle, are the LHS or RHS and identity shuffles?
- bool isLHSID = true, isRHSID = true;
-
- for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
- if (Mask[i] < 0) continue; // Ignore undef values.
- // Is this an identity shuffle of the LHS value?
- isLHSID &= (Mask[i] == (int)i);
-
- // Is this an identity shuffle of the RHS value?
- isRHSID &= (Mask[i]-e == i);
- }
+ bool isLHSID, isRHSID;
+ RecognizeIdentityMask(Mask, isLHSID, isRHSID);
// Eliminate identity shuffles.
if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
}
- if (isa<UndefValue>(RHS) &&
- // This isn't necessary for correctness, but the comment block below
- // claims that there are cases where folding two shuffles into one would
- // cause worse codegen on some targets.
- !isa<ShuffleVectorInst>(LHS) &&
- CanEvaluateShuffled(LHS, Mask)) {
+ if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
return ReplaceInstUsesWith(SVI, V);
}
ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
if (LHSShuffle)
if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
- LHSShuffle = NULL;
+ LHSShuffle = nullptr;
if (RHSShuffle)
if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
- RHSShuffle = NULL;
+ RHSShuffle = nullptr;
if (!LHSShuffle && !RHSShuffle)
- return MadeChange ? &SVI : 0;
+ return MadeChange ? &SVI : nullptr;
- Value* LHSOp0 = NULL;
- Value* LHSOp1 = NULL;
- Value* RHSOp0 = NULL;
+ Value* LHSOp0 = nullptr;
+ Value* LHSOp1 = nullptr;
+ Value* RHSOp0 = nullptr;
unsigned LHSOp0Width = 0;
unsigned RHSOp0Width = 0;
if (LHSShuffle) {
// case 4
if (LHSOp0 == RHSOp0) {
newLHS = LHSOp0;
- newRHS = NULL;
+ newRHS = nullptr;
}
if (newLHS == LHS && newRHS == RHS)
- return MadeChange ? &SVI : 0;
+ return MadeChange ? &SVI : nullptr;
SmallVector<int, 16> LHSMask;
SmallVector<int, 16> RHSMask;
// references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
// If newRHS == newLHS, we want to remap any references from newRHS to
// newLHS so that we can properly identify splats that may occur due to
- // obfuscation accross the two vectors.
- if (eltMask >= 0 && newRHS != NULL && newLHS != newRHS)
+ // obfuscation across the two vectors.
+ if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
eltMask += newLHSWidth;
}
Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
}
}
- if (newRHS == NULL)
+ if (!newRHS)
newRHS = UndefValue::get(newLHS->getType());
return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
}
- return MadeChange ? &SVI : 0;
+ // If the result mask is an identity, replace uses of this instruction with
+ // corresponding argument.
+ bool isLHSID, isRHSID;
+ RecognizeIdentityMask(newMask, isLHSID, isRHSID);
+ if (isLHSID && VWidth == LHSOp0Width) return ReplaceInstUsesWith(SVI, newLHS);
+ if (isRHSID && VWidth == RHSOp0Width) return ReplaceInstUsesWith(SVI, newRHS);
+
+ return MadeChange ? &SVI : nullptr;
}
projects / oota-llvm.git / blobdiff