1 //===- InstCombineVectorOps.cpp -------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements instcombine for ExtractElement, InsertElement and
13 //===----------------------------------------------------------------------===//
15 #include "InstCombineInternal.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/Analysis/InstructionSimplify.h"
22 #include "llvm/Analysis/VectorUtils.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/Constant.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/InstrTypes.h"
28 #include "llvm/IR/Instruction.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/Operator.h"
31 #include "llvm/IR/PatternMatch.h"
32 #include "llvm/IR/Type.h"
33 #include "llvm/IR/User.h"
34 #include "llvm/IR/Value.h"
35 #include "llvm/Support/Casting.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h"
44 using namespace PatternMatch;
46 #define DEBUG_TYPE "instcombine"
48 /// Return true if the value is cheaper to scalarize than it is to leave as a
49 /// vector operation. isConstant indicates whether we're extracting one known
50 /// element. If false we're extracting a variable index.
51 static bool cheapToScalarize(Value *V, bool isConstant) {
52 if (Constant *C = dyn_cast<Constant>(V)) {
53 if (isConstant) return true;
55 // If all elts are the same, we can extract it and use any of the values.
56 if (Constant *Op0 = C->getAggregateElement(0U)) {
57 for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e;
59 if (C->getAggregateElement(i) != Op0)
64 Instruction *I = dyn_cast<Instruction>(V);
67 // Insert element gets simplified to the inserted element or is deleted if
68 // this is constant idx extract element and its a constant idx insertelt.
69 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
70 isa<ConstantInt>(I->getOperand(2)))
72 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
74 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
75 if (BO->hasOneUse() &&
76 (cheapToScalarize(BO->getOperand(0), isConstant) ||
77 cheapToScalarize(BO->getOperand(1), isConstant)))
79 if (CmpInst *CI = dyn_cast<CmpInst>(I))
80 if (CI->hasOneUse() &&
81 (cheapToScalarize(CI->getOperand(0), isConstant) ||
82 cheapToScalarize(CI->getOperand(1), isConstant)))
88 // If we have a PHI node with a vector type that is only used to feed
89 // itself and be an operand of extractelement at a constant location,
90 // try to replace the PHI of the vector type with a PHI of a scalar type.
91 Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) {
92 SmallVector<Instruction *, 2> Extracts;
93 // The users we want the PHI to have are:
94 // 1) The EI ExtractElement (we already know this)
95 // 2) Possibly more ExtractElements with the same index.
96 // 3) Another operand, which will feed back into the PHI.
97 Instruction *PHIUser = nullptr;
98 for (auto U : PN->users()) {
99 if (ExtractElementInst *EU = dyn_cast<ExtractElementInst>(U)) {
100 if (EI.getIndexOperand() == EU->getIndexOperand())
101 Extracts.push_back(EU);
104 } else if (!PHIUser) {
105 PHIUser = cast<Instruction>(U);
114 // Verify that this PHI user has one use, which is the PHI itself,
115 // and that it is a binary operation which is cheap to scalarize.
116 // otherwise return nullptr.
117 if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) ||
118 !(isa<BinaryOperator>(PHIUser)) || !cheapToScalarize(PHIUser, true))
121 // Create a scalar PHI node that will replace the vector PHI node
122 // just before the current PHI node.
123 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
124 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN));
125 // Scalarize each PHI operand.
126 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
127 Value *PHIInVal = PN->getIncomingValue(i);
128 BasicBlock *inBB = PN->getIncomingBlock(i);
129 Value *Elt = EI.getIndexOperand();
130 // If the operand is the PHI induction variable:
131 if (PHIInVal == PHIUser) {
132 // Scalarize the binary operation. Its first operand is the
133 // scalar PHI, and the second operand is extracted from the other
135 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
136 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
137 Value *Op = InsertNewInstWith(
138 ExtractElementInst::Create(B0->getOperand(opId), Elt,
139 B0->getOperand(opId)->getName() + ".Elt"),
141 Value *newPHIUser = InsertNewInstWith(
142 BinaryOperator::CreateWithCopiedFlags(B0->getOpcode(),
143 scalarPHI, Op, B0), *B0);
144 scalarPHI->addIncoming(newPHIUser, inBB);
146 // Scalarize PHI input:
147 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
148 // Insert the new instruction into the predecessor basic block.
149 Instruction *pos = dyn_cast<Instruction>(PHIInVal);
150 BasicBlock::iterator InsertPos;
151 if (pos && !isa<PHINode>(pos)) {
152 InsertPos = ++pos->getIterator();
154 InsertPos = inBB->getFirstInsertionPt();
157 InsertNewInstWith(newEI, *InsertPos);
159 scalarPHI->addIncoming(newEI, inBB);
163 for (auto E : Extracts)
164 replaceInstUsesWith(*E, scalarPHI);
169 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
170 if (Value *V = SimplifyExtractElementInst(EI.getVectorOperand(),
171 EI.getIndexOperand(),
172 SQ.getWithInstruction(&EI)))
173 return replaceInstUsesWith(EI, V);
175 // If vector val is constant with all elements the same, replace EI with
176 // that element. We handle a known element # below.
177 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0)))
178 if (cheapToScalarize(C, false))
179 return replaceInstUsesWith(EI, C->getAggregateElement(0U));
181 // If extracting a specified index from the vector, see if we can recursively
182 // find a previously computed scalar that was inserted into the vector.
183 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
184 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
186 // InstSimplify should handle cases where the index is invalid.
187 if (!IdxC->getValue().ule(VectorWidth))
190 unsigned IndexVal = IdxC->getZExtValue();
192 // This instruction only demands the single element from the input vector.
193 // If the input vector has a single use, simplify it based on this use
195 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
196 APInt UndefElts(VectorWidth, 0);
197 APInt DemandedMask(VectorWidth, 0);
198 DemandedMask.setBit(IndexVal);
199 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), DemandedMask,
206 // If this extractelement is directly using a bitcast from a vector of
207 // the same number of elements, see if we can find the source element from
208 // it. In this case, we will end up needing to bitcast the scalars.
209 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
210 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
211 if (VT->getNumElements() == VectorWidth)
212 if (Value *Elt = findScalarElement(BCI->getOperand(0), IndexVal))
213 return new BitCastInst(Elt, EI.getType());
216 // If there's a vector PHI feeding a scalar use through this extractelement
217 // instruction, try to scalarize the PHI.
218 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
219 Instruction *scalarPHI = scalarizePHI(EI, PN);
225 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
226 // Push extractelement into predecessor operation if legal and
227 // profitable to do so.
228 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
229 if (I->hasOneUse() &&
230 cheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
232 Builder.CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
233 EI.getName()+".lhs");
235 Builder.CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
236 EI.getName()+".rhs");
237 return BinaryOperator::CreateWithCopiedFlags(BO->getOpcode(),
240 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
241 // Extracting the inserted element?
242 if (IE->getOperand(2) == EI.getOperand(1))
243 return replaceInstUsesWith(EI, IE->getOperand(1));
244 // If the inserted and extracted elements are constants, they must not
245 // be the same value, extract from the pre-inserted value instead.
246 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
247 Worklist.AddValue(EI.getOperand(0));
248 EI.setOperand(0, IE->getOperand(0));
251 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
252 // If this is extracting an element from a shufflevector, figure out where
253 // it came from and extract from the appropriate input element instead.
254 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
255 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
258 SVI->getOperand(0)->getType()->getVectorNumElements();
261 return replaceInstUsesWith(EI, UndefValue::get(EI.getType()));
262 if (SrcIdx < (int)LHSWidth)
263 Src = SVI->getOperand(0);
266 Src = SVI->getOperand(1);
268 Type *Int32Ty = Type::getInt32Ty(EI.getContext());
269 return ExtractElementInst::Create(Src,
270 ConstantInt::get(Int32Ty,
273 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
274 // Canonicalize extractelement(cast) -> cast(extractelement).
275 // Bitcasts can change the number of vector elements, and they cost
277 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
278 Value *EE = Builder.CreateExtractElement(CI->getOperand(0),
279 EI.getIndexOperand());
280 Worklist.AddValue(EE);
281 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
288 /// If V is a shuffle of values that ONLY returns elements from either LHS or
289 /// RHS, return the shuffle mask and true. Otherwise, return false.
290 static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
291 SmallVectorImpl<Constant*> &Mask) {
292 assert(LHS->getType() == RHS->getType() &&
293 "Invalid CollectSingleShuffleElements");
294 unsigned NumElts = V->getType()->getVectorNumElements();
296 if (isa<UndefValue>(V)) {
297 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
302 for (unsigned i = 0; i != NumElts; ++i)
303 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
308 for (unsigned i = 0; i != NumElts; ++i)
309 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
314 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
315 // If this is an insert of an extract from some other vector, include it.
316 Value *VecOp = IEI->getOperand(0);
317 Value *ScalarOp = IEI->getOperand(1);
318 Value *IdxOp = IEI->getOperand(2);
320 if (!isa<ConstantInt>(IdxOp))
322 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
324 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
325 // We can handle this if the vector we are inserting into is
327 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
328 // If so, update the mask to reflect the inserted undef.
329 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
332 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
333 if (isa<ConstantInt>(EI->getOperand(1))) {
334 unsigned ExtractedIdx =
335 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
336 unsigned NumLHSElts = LHS->getType()->getVectorNumElements();
338 // This must be extracting from either LHS or RHS.
339 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
340 // We can handle this if the vector we are inserting into is
342 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
343 // If so, update the mask to reflect the inserted value.
344 if (EI->getOperand(0) == LHS) {
345 Mask[InsertedIdx % NumElts] =
346 ConstantInt::get(Type::getInt32Ty(V->getContext()),
349 assert(EI->getOperand(0) == RHS);
350 Mask[InsertedIdx % NumElts] =
351 ConstantInt::get(Type::getInt32Ty(V->getContext()),
352 ExtractedIdx + NumLHSElts);
364 /// If we have insertion into a vector that is wider than the vector that we
365 /// are extracting from, try to widen the source vector to allow a single
366 /// shufflevector to replace one or more insert/extract pairs.
367 static void replaceExtractElements(InsertElementInst *InsElt,
368 ExtractElementInst *ExtElt,
370 VectorType *InsVecType = InsElt->getType();
371 VectorType *ExtVecType = ExtElt->getVectorOperandType();
372 unsigned NumInsElts = InsVecType->getVectorNumElements();
373 unsigned NumExtElts = ExtVecType->getVectorNumElements();
375 // The inserted-to vector must be wider than the extracted-from vector.
376 if (InsVecType->getElementType() != ExtVecType->getElementType() ||
377 NumExtElts >= NumInsElts)
380 // Create a shuffle mask to widen the extended-from vector using undefined
381 // values. The mask selects all of the values of the original vector followed
382 // by as many undefined values as needed to create a vector of the same length
383 // as the inserted-to vector.
384 SmallVector<Constant *, 16> ExtendMask;
385 IntegerType *IntType = Type::getInt32Ty(InsElt->getContext());
386 for (unsigned i = 0; i < NumExtElts; ++i)
387 ExtendMask.push_back(ConstantInt::get(IntType, i));
388 for (unsigned i = NumExtElts; i < NumInsElts; ++i)
389 ExtendMask.push_back(UndefValue::get(IntType));
391 Value *ExtVecOp = ExtElt->getVectorOperand();
392 auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
393 BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
394 ? ExtVecOpInst->getParent()
395 : ExtElt->getParent();
397 // TODO: This restriction matches the basic block check below when creating
398 // new extractelement instructions. If that limitation is removed, this one
399 // could also be removed. But for now, we just bail out to ensure that we
400 // will replace the extractelement instruction that is feeding our
401 // insertelement instruction. This allows the insertelement to then be
402 // replaced by a shufflevector. If the insertelement is not replaced, we can
403 // induce infinite looping because there's an optimization for extractelement
404 // that will delete our widening shuffle. This would trigger another attempt
405 // here to create that shuffle, and we spin forever.
406 if (InsertionBlock != InsElt->getParent())
409 // TODO: This restriction matches the check in visitInsertElementInst() and
410 // prevents an infinite loop caused by not turning the extract/insert pair
411 // into a shuffle. We really should not need either check, but we're lacking
412 // folds for shufflevectors because we're afraid to generate shuffle masks
413 // that the backend can't handle.
414 if (InsElt->hasOneUse() && isa<InsertElementInst>(InsElt->user_back()))
417 auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
418 ConstantVector::get(ExtendMask));
420 // Insert the new shuffle after the vector operand of the extract is defined
421 // (as long as it's not a PHI) or at the start of the basic block of the
422 // extract, so any subsequent extracts in the same basic block can use it.
423 // TODO: Insert before the earliest ExtractElementInst that is replaced.
424 if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
425 WideVec->insertAfter(ExtVecOpInst);
427 IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
429 // Replace extracts from the original narrow vector with extracts from the new
431 for (User *U : ExtVecOp->users()) {
432 ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
433 if (!OldExt || OldExt->getParent() != WideVec->getParent())
435 auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
436 NewExt->insertAfter(OldExt);
437 IC.replaceInstUsesWith(*OldExt, NewExt);
441 /// We are building a shuffle to create V, which is a sequence of insertelement,
442 /// extractelement pairs. If PermittedRHS is set, then we must either use it or
443 /// not rely on the second vector source. Return a std::pair containing the
444 /// left and right vectors of the proposed shuffle (or 0), and set the Mask
445 /// parameter as required.
447 /// Note: we intentionally don't try to fold earlier shuffles since they have
448 /// often been chosen carefully to be efficiently implementable on the target.
449 using ShuffleOps = std::pair<Value *, Value *>;
451 static ShuffleOps collectShuffleElements(Value *V,
452 SmallVectorImpl<Constant *> &Mask,
455 assert(V->getType()->isVectorTy() && "Invalid shuffle!");
456 unsigned NumElts = V->getType()->getVectorNumElements();
458 if (isa<UndefValue>(V)) {
459 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
460 return std::make_pair(
461 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
464 if (isa<ConstantAggregateZero>(V)) {
465 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
466 return std::make_pair(V, nullptr);
469 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
470 // If this is an insert of an extract from some other vector, include it.
471 Value *VecOp = IEI->getOperand(0);
472 Value *ScalarOp = IEI->getOperand(1);
473 Value *IdxOp = IEI->getOperand(2);
475 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
476 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
477 unsigned ExtractedIdx =
478 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
479 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
481 // Either the extracted from or inserted into vector must be RHSVec,
482 // otherwise we'd end up with a shuffle of three inputs.
483 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
484 Value *RHS = EI->getOperand(0);
485 ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC);
486 assert(LR.second == nullptr || LR.second == RHS);
488 if (LR.first->getType() != RHS->getType()) {
489 // Although we are giving up for now, see if we can create extracts
490 // that match the inserts for another round of combining.
491 replaceExtractElements(IEI, EI, IC);
493 // We tried our best, but we can't find anything compatible with RHS
494 // further up the chain. Return a trivial shuffle.
495 for (unsigned i = 0; i < NumElts; ++i)
496 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
497 return std::make_pair(V, nullptr);
500 unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
501 Mask[InsertedIdx % NumElts] =
502 ConstantInt::get(Type::getInt32Ty(V->getContext()),
503 NumLHSElts+ExtractedIdx);
504 return std::make_pair(LR.first, RHS);
507 if (VecOp == PermittedRHS) {
508 // We've gone as far as we can: anything on the other side of the
509 // extractelement will already have been converted into a shuffle.
510 unsigned NumLHSElts =
511 EI->getOperand(0)->getType()->getVectorNumElements();
512 for (unsigned i = 0; i != NumElts; ++i)
513 Mask.push_back(ConstantInt::get(
514 Type::getInt32Ty(V->getContext()),
515 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
516 return std::make_pair(EI->getOperand(0), PermittedRHS);
519 // If this insertelement is a chain that comes from exactly these two
520 // vectors, return the vector and the effective shuffle.
521 if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
522 collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
524 return std::make_pair(EI->getOperand(0), PermittedRHS);
529 // Otherwise, we can't do anything fancy. Return an identity vector.
530 for (unsigned i = 0; i != NumElts; ++i)
531 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
532 return std::make_pair(V, nullptr);
535 /// Try to find redundant insertvalue instructions, like the following ones:
536 /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
537 /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
538 /// Here the second instruction inserts values at the same indices, as the
539 /// first one, making the first one redundant.
540 /// It should be transformed to:
541 /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
542 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
543 bool IsRedundant = false;
544 ArrayRef<unsigned int> FirstIndices = I.getIndices();
546 // If there is a chain of insertvalue instructions (each of them except the
547 // last one has only one use and it's another insertvalue insn from this
548 // chain), check if any of the 'children' uses the same indices as the first
549 // instruction. In this case, the first one is redundant.
552 while (V->hasOneUse() && Depth < 10) {
553 User *U = V->user_back();
554 auto UserInsInst = dyn_cast<InsertValueInst>(U);
555 if (!UserInsInst || U->getOperand(0) != V)
557 if (UserInsInst->getIndices() == FirstIndices) {
566 return replaceInstUsesWith(I, I.getOperand(0));
570 static bool isShuffleEquivalentToSelect(ShuffleVectorInst &Shuf) {
571 int MaskSize = Shuf.getMask()->getType()->getVectorNumElements();
572 int VecSize = Shuf.getOperand(0)->getType()->getVectorNumElements();
574 // A vector select does not change the size of the operands.
575 if (MaskSize != VecSize)
578 // Each mask element must be undefined or choose a vector element from one of
579 // the source operands without crossing vector lanes.
580 for (int i = 0; i != MaskSize; ++i) {
581 int Elt = Shuf.getMaskValue(i);
582 if (Elt != -1 && Elt != i && Elt != i + VecSize)
589 // Turn a chain of inserts that splats a value into a canonical insert + shuffle
591 // insertelt(insertelt(insertelt(insertelt X, %k, 0), %k, 1), %k, 2) ... ->
592 // shufflevector(insertelt(X, %k, 0), undef, zero)
593 static Instruction *foldInsSequenceIntoBroadcast(InsertElementInst &InsElt) {
594 // We are interested in the last insert in a chain. So, if this insert
595 // has a single user, and that user is an insert, bail.
596 if (InsElt.hasOneUse() && isa<InsertElementInst>(InsElt.user_back()))
599 VectorType *VT = cast<VectorType>(InsElt.getType());
600 int NumElements = VT->getNumElements();
602 // Do not try to do this for a one-element vector, since that's a nop,
603 // and will cause an inf-loop.
604 if (NumElements == 1)
607 Value *SplatVal = InsElt.getOperand(1);
608 InsertElementInst *CurrIE = &InsElt;
609 SmallVector<bool, 16> ElementPresent(NumElements, false);
610 InsertElementInst *FirstIE = nullptr;
612 // Walk the chain backwards, keeping track of which indices we inserted into,
613 // until we hit something that isn't an insert of the splatted value.
615 auto *Idx = dyn_cast<ConstantInt>(CurrIE->getOperand(2));
616 if (!Idx || CurrIE->getOperand(1) != SplatVal)
619 auto *NextIE = dyn_cast<InsertElementInst>(CurrIE->getOperand(0));
620 // Check none of the intermediate steps have any additional uses, except
621 // for the root insertelement instruction, which can be re-used, if it
622 // inserts at position 0.
623 if (CurrIE != &InsElt &&
624 (!CurrIE->hasOneUse() && (NextIE != nullptr || !Idx->isZero())))
627 ElementPresent[Idx->getZExtValue()] = true;
632 // Make sure we've seen an insert into every element.
633 if (llvm::any_of(ElementPresent, [](bool Present) { return !Present; }))
636 // All right, create the insert + shuffle.
637 Instruction *InsertFirst;
638 if (cast<ConstantInt>(FirstIE->getOperand(2))->isZero())
639 InsertFirst = FirstIE;
641 InsertFirst = InsertElementInst::Create(
642 UndefValue::get(VT), SplatVal,
643 ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), 0),
646 Constant *ZeroMask = ConstantAggregateZero::get(
647 VectorType::get(Type::getInt32Ty(InsElt.getContext()), NumElements));
649 return new ShuffleVectorInst(InsertFirst, UndefValue::get(VT), ZeroMask);
652 /// If we have an insertelement instruction feeding into another insertelement
653 /// and the 2nd is inserting a constant into the vector, canonicalize that
654 /// constant insertion before the insertion of a variable:
656 /// insertelement (insertelement X, Y, IdxC1), ScalarC, IdxC2 -->
657 /// insertelement (insertelement X, ScalarC, IdxC2), Y, IdxC1
659 /// This has the potential of eliminating the 2nd insertelement instruction
660 /// via constant folding of the scalar constant into a vector constant.
661 static Instruction *hoistInsEltConst(InsertElementInst &InsElt2,
662 InstCombiner::BuilderTy &Builder) {
663 auto *InsElt1 = dyn_cast<InsertElementInst>(InsElt2.getOperand(0));
664 if (!InsElt1 || !InsElt1->hasOneUse())
669 ConstantInt *IdxC1, *IdxC2;
670 if (match(InsElt1->getOperand(0), m_Value(X)) &&
671 match(InsElt1->getOperand(1), m_Value(Y)) && !isa<Constant>(Y) &&
672 match(InsElt1->getOperand(2), m_ConstantInt(IdxC1)) &&
673 match(InsElt2.getOperand(1), m_Constant(ScalarC)) &&
674 match(InsElt2.getOperand(2), m_ConstantInt(IdxC2)) && IdxC1 != IdxC2) {
675 Value *NewInsElt1 = Builder.CreateInsertElement(X, ScalarC, IdxC2);
676 return InsertElementInst::Create(NewInsElt1, Y, IdxC1);
682 /// insertelt (shufflevector X, CVec, Mask|insertelt X, C1, CIndex1), C, CIndex
683 /// --> shufflevector X, CVec', Mask'
684 static Instruction *foldConstantInsEltIntoShuffle(InsertElementInst &InsElt) {
685 auto *Inst = dyn_cast<Instruction>(InsElt.getOperand(0));
686 // Bail out if the parent has more than one use. In that case, we'd be
687 // replacing the insertelt with a shuffle, and that's not a clear win.
688 if (!Inst || !Inst->hasOneUse())
690 if (auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0))) {
691 // The shuffle must have a constant vector operand. The insertelt must have
692 // a constant scalar being inserted at a constant position in the vector.
693 Constant *ShufConstVec, *InsEltScalar;
694 uint64_t InsEltIndex;
695 if (!match(Shuf->getOperand(1), m_Constant(ShufConstVec)) ||
696 !match(InsElt.getOperand(1), m_Constant(InsEltScalar)) ||
697 !match(InsElt.getOperand(2), m_ConstantInt(InsEltIndex)))
700 // Adding an element to an arbitrary shuffle could be expensive, but a
701 // shuffle that selects elements from vectors without crossing lanes is
703 // If we're just adding a constant into that shuffle, it will still be
705 if (!isShuffleEquivalentToSelect(*Shuf))
708 // From the above 'select' check, we know that the mask has the same number
709 // of elements as the vector input operands. We also know that each constant
710 // input element is used in its lane and can not be used more than once by
711 // the shuffle. Therefore, replace the constant in the shuffle's constant
712 // vector with the insertelt constant. Replace the constant in the shuffle's
713 // mask vector with the insertelt index plus the length of the vector
714 // (because the constant vector operand of a shuffle is always the 2nd
716 Constant *Mask = Shuf->getMask();
717 unsigned NumElts = Mask->getType()->getVectorNumElements();
718 SmallVector<Constant *, 16> NewShufElts(NumElts);
719 SmallVector<Constant *, 16> NewMaskElts(NumElts);
720 for (unsigned I = 0; I != NumElts; ++I) {
721 if (I == InsEltIndex) {
722 NewShufElts[I] = InsEltScalar;
723 Type *Int32Ty = Type::getInt32Ty(Shuf->getContext());
724 NewMaskElts[I] = ConstantInt::get(Int32Ty, InsEltIndex + NumElts);
726 // Copy over the existing values.
727 NewShufElts[I] = ShufConstVec->getAggregateElement(I);
728 NewMaskElts[I] = Mask->getAggregateElement(I);
732 // Create new operands for a shuffle that includes the constant of the
733 // original insertelt. The old shuffle will be dead now.
734 return new ShuffleVectorInst(Shuf->getOperand(0),
735 ConstantVector::get(NewShufElts),
736 ConstantVector::get(NewMaskElts));
737 } else if (auto *IEI = dyn_cast<InsertElementInst>(Inst)) {
738 // Transform sequences of insertelements ops with constant data/indexes into
739 // a single shuffle op.
740 unsigned NumElts = InsElt.getType()->getNumElements();
742 uint64_t InsertIdx[2];
744 if (!match(InsElt.getOperand(2), m_ConstantInt(InsertIdx[0])) ||
745 !match(InsElt.getOperand(1), m_Constant(Val[0])) ||
746 !match(IEI->getOperand(2), m_ConstantInt(InsertIdx[1])) ||
747 !match(IEI->getOperand(1), m_Constant(Val[1])))
749 SmallVector<Constant *, 16> Values(NumElts);
750 SmallVector<Constant *, 16> Mask(NumElts);
751 auto ValI = std::begin(Val);
752 // Generate new constant vector and mask.
753 // We have 2 values/masks from the insertelements instructions. Insert them
754 // into new value/mask vectors.
755 for (uint64_t I : InsertIdx) {
759 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()),
764 // Remaining values are filled with 'undef' values.
765 for (unsigned I = 0; I < NumElts; ++I) {
768 Values[I] = UndefValue::get(InsElt.getType()->getElementType());
769 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), I);
772 // Create new operands for a shuffle that includes the constant of the
773 // original insertelt.
774 return new ShuffleVectorInst(IEI->getOperand(0),
775 ConstantVector::get(Values),
776 ConstantVector::get(Mask));
781 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
782 Value *VecOp = IE.getOperand(0);
783 Value *ScalarOp = IE.getOperand(1);
784 Value *IdxOp = IE.getOperand(2);
786 if (auto *V = SimplifyInsertElementInst(
787 VecOp, ScalarOp, IdxOp, SQ.getWithInstruction(&IE)))
788 return replaceInstUsesWith(IE, V);
790 // Inserting an undef or into an undefined place, remove this.
791 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
792 replaceInstUsesWith(IE, VecOp);
794 // If the inserted element was extracted from some other vector, and if the
795 // indexes are constant, try to turn this into a shufflevector operation.
796 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
797 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
798 unsigned NumInsertVectorElts = IE.getType()->getNumElements();
799 unsigned NumExtractVectorElts =
800 EI->getOperand(0)->getType()->getVectorNumElements();
801 unsigned ExtractedIdx =
802 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
803 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
805 if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
806 return replaceInstUsesWith(IE, VecOp);
808 if (InsertedIdx >= NumInsertVectorElts) // Out of range insert.
809 return replaceInstUsesWith(IE, UndefValue::get(IE.getType()));
811 // If we are extracting a value from a vector, then inserting it right
812 // back into the same place, just use the input vector.
813 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
814 return replaceInstUsesWith(IE, VecOp);
816 // If this insertelement isn't used by some other insertelement, turn it
817 // (and any insertelements it points to), into one big shuffle.
818 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
819 SmallVector<Constant*, 16> Mask;
820 ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this);
822 // The proposed shuffle may be trivial, in which case we shouldn't
823 // perform the combine.
824 if (LR.first != &IE && LR.second != &IE) {
825 // We now have a shuffle of LHS, RHS, Mask.
826 if (LR.second == nullptr)
827 LR.second = UndefValue::get(LR.first->getType());
828 return new ShuffleVectorInst(LR.first, LR.second,
829 ConstantVector::get(Mask));
835 unsigned VWidth = VecOp->getType()->getVectorNumElements();
836 APInt UndefElts(VWidth, 0);
837 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
838 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
840 return replaceInstUsesWith(IE, V);
844 if (Instruction *Shuf = foldConstantInsEltIntoShuffle(IE))
847 if (Instruction *NewInsElt = hoistInsEltConst(IE, Builder))
850 // Turn a sequence of inserts that broadcasts a scalar into a single
851 // insert + shufflevector.
852 if (Instruction *Broadcast = foldInsSequenceIntoBroadcast(IE))
858 /// Return true if we can evaluate the specified expression tree if the vector
859 /// elements were shuffled in a different order.
860 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
861 unsigned Depth = 5) {
862 // We can always reorder the elements of a constant.
863 if (isa<Constant>(V))
866 // We won't reorder vector arguments. No IPO here.
867 Instruction *I = dyn_cast<Instruction>(V);
868 if (!I) return false;
870 // Two users may expect different orders of the elements. Don't try it.
874 if (Depth == 0) return false;
876 switch (I->getOpcode()) {
877 case Instruction::Add:
878 case Instruction::FAdd:
879 case Instruction::Sub:
880 case Instruction::FSub:
881 case Instruction::Mul:
882 case Instruction::FMul:
883 case Instruction::UDiv:
884 case Instruction::SDiv:
885 case Instruction::FDiv:
886 case Instruction::URem:
887 case Instruction::SRem:
888 case Instruction::FRem:
889 case Instruction::Shl:
890 case Instruction::LShr:
891 case Instruction::AShr:
892 case Instruction::And:
893 case Instruction::Or:
894 case Instruction::Xor:
895 case Instruction::ICmp:
896 case Instruction::FCmp:
897 case Instruction::Trunc:
898 case Instruction::ZExt:
899 case Instruction::SExt:
900 case Instruction::FPToUI:
901 case Instruction::FPToSI:
902 case Instruction::UIToFP:
903 case Instruction::SIToFP:
904 case Instruction::FPTrunc:
905 case Instruction::FPExt:
906 case Instruction::GetElementPtr: {
907 for (Value *Operand : I->operands()) {
908 if (!CanEvaluateShuffled(Operand, Mask, Depth-1))
913 case Instruction::InsertElement: {
914 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
915 if (!CI) return false;
916 int ElementNumber = CI->getLimitedValue();
918 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
919 // can't put an element into multiple indices.
920 bool SeenOnce = false;
921 for (int i = 0, e = Mask.size(); i != e; ++i) {
922 if (Mask[i] == ElementNumber) {
928 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
934 /// Rebuild a new instruction just like 'I' but with the new operands given.
935 /// In the event of type mismatch, the type of the operands is correct.
936 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) {
937 // We don't want to use the IRBuilder here because we want the replacement
938 // instructions to appear next to 'I', not the builder's insertion point.
939 switch (I->getOpcode()) {
940 case Instruction::Add:
941 case Instruction::FAdd:
942 case Instruction::Sub:
943 case Instruction::FSub:
944 case Instruction::Mul:
945 case Instruction::FMul:
946 case Instruction::UDiv:
947 case Instruction::SDiv:
948 case Instruction::FDiv:
949 case Instruction::URem:
950 case Instruction::SRem:
951 case Instruction::FRem:
952 case Instruction::Shl:
953 case Instruction::LShr:
954 case Instruction::AShr:
955 case Instruction::And:
956 case Instruction::Or:
957 case Instruction::Xor: {
958 BinaryOperator *BO = cast<BinaryOperator>(I);
959 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
960 BinaryOperator *New =
961 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
962 NewOps[0], NewOps[1], "", BO);
963 if (isa<OverflowingBinaryOperator>(BO)) {
964 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
965 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
967 if (isa<PossiblyExactOperator>(BO)) {
968 New->setIsExact(BO->isExact());
970 if (isa<FPMathOperator>(BO))
971 New->copyFastMathFlags(I);
974 case Instruction::ICmp:
975 assert(NewOps.size() == 2 && "icmp with #ops != 2");
976 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
977 NewOps[0], NewOps[1]);
978 case Instruction::FCmp:
979 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
980 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
981 NewOps[0], NewOps[1]);
982 case Instruction::Trunc:
983 case Instruction::ZExt:
984 case Instruction::SExt:
985 case Instruction::FPToUI:
986 case Instruction::FPToSI:
987 case Instruction::UIToFP:
988 case Instruction::SIToFP:
989 case Instruction::FPTrunc:
990 case Instruction::FPExt: {
991 // It's possible that the mask has a different number of elements from
992 // the original cast. We recompute the destination type to match the mask.
994 VectorType::get(I->getType()->getScalarType(),
995 NewOps[0]->getType()->getVectorNumElements());
996 assert(NewOps.size() == 1 && "cast with #ops != 1");
997 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
1000 case Instruction::GetElementPtr: {
1001 Value *Ptr = NewOps[0];
1002 ArrayRef<Value*> Idx = NewOps.slice(1);
1003 GetElementPtrInst *GEP = GetElementPtrInst::Create(
1004 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I);
1005 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
1009 llvm_unreachable("failed to rebuild vector instructions");
1013 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
1014 // Mask.size() does not need to be equal to the number of vector elements.
1016 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
1017 Type *EltTy = V->getType()->getScalarType();
1018 if (isa<UndefValue>(V))
1019 return UndefValue::get(VectorType::get(EltTy, Mask.size()));
1021 if (isa<ConstantAggregateZero>(V))
1022 return ConstantAggregateZero::get(VectorType::get(EltTy, Mask.size()));
1024 if (Constant *C = dyn_cast<Constant>(V)) {
1025 SmallVector<Constant *, 16> MaskValues;
1026 for (int i = 0, e = Mask.size(); i != e; ++i) {
1028 MaskValues.push_back(UndefValue::get(Builder.getInt32Ty()));
1030 MaskValues.push_back(Builder.getInt32(Mask[i]));
1032 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
1033 ConstantVector::get(MaskValues));
1036 Instruction *I = cast<Instruction>(V);
1037 switch (I->getOpcode()) {
1038 case Instruction::Add:
1039 case Instruction::FAdd:
1040 case Instruction::Sub:
1041 case Instruction::FSub:
1042 case Instruction::Mul:
1043 case Instruction::FMul:
1044 case Instruction::UDiv:
1045 case Instruction::SDiv:
1046 case Instruction::FDiv:
1047 case Instruction::URem:
1048 case Instruction::SRem:
1049 case Instruction::FRem:
1050 case Instruction::Shl:
1051 case Instruction::LShr:
1052 case Instruction::AShr:
1053 case Instruction::And:
1054 case Instruction::Or:
1055 case Instruction::Xor:
1056 case Instruction::ICmp:
1057 case Instruction::FCmp:
1058 case Instruction::Trunc:
1059 case Instruction::ZExt:
1060 case Instruction::SExt:
1061 case Instruction::FPToUI:
1062 case Instruction::FPToSI:
1063 case Instruction::UIToFP:
1064 case Instruction::SIToFP:
1065 case Instruction::FPTrunc:
1066 case Instruction::FPExt:
1067 case Instruction::Select:
1068 case Instruction::GetElementPtr: {
1069 SmallVector<Value*, 8> NewOps;
1070 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
1071 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
1072 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
1073 NewOps.push_back(V);
1074 NeedsRebuild |= (V != I->getOperand(i));
1077 return buildNew(I, NewOps);
1081 case Instruction::InsertElement: {
1082 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
1084 // The insertelement was inserting at Element. Figure out which element
1085 // that becomes after shuffling. The answer is guaranteed to be unique
1086 // by CanEvaluateShuffled.
1089 for (int e = Mask.size(); Index != e; ++Index) {
1090 if (Mask[Index] == Element) {
1096 // If element is not in Mask, no need to handle the operand 1 (element to
1097 // be inserted). Just evaluate values in operand 0 according to Mask.
1099 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
1101 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
1102 return InsertElementInst::Create(V, I->getOperand(1),
1103 Builder.getInt32(Index), "", I);
1106 llvm_unreachable("failed to reorder elements of vector instruction!");
1109 static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask,
1110 bool &isLHSID, bool &isRHSID) {
1111 isLHSID = isRHSID = true;
1113 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
1114 if (Mask[i] < 0) continue; // Ignore undef values.
1115 // Is this an identity shuffle of the LHS value?
1116 isLHSID &= (Mask[i] == (int)i);
1118 // Is this an identity shuffle of the RHS value?
1119 isRHSID &= (Mask[i]-e == i);
1123 // Returns true if the shuffle is extracting a contiguous range of values from
1124 // LHS, for example:
1125 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1126 // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
1127 // Shuffles to: |EE|FF|GG|HH|
1129 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
1130 SmallVector<int, 16> &Mask) {
1131 unsigned LHSElems = SVI.getOperand(0)->getType()->getVectorNumElements();
1132 unsigned MaskElems = Mask.size();
1133 unsigned BegIdx = Mask.front();
1134 unsigned EndIdx = Mask.back();
1135 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
1137 for (unsigned I = 0; I != MaskElems; ++I)
1138 if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
1143 /// These are the ingredients in an alternate form binary operator as described
1146 BinaryOperator::BinaryOps Opcode;
1149 BinopElts(BinaryOperator::BinaryOps Opc = (BinaryOperator::BinaryOps)0,
1150 Value *V0 = nullptr, Value *V1 = nullptr) :
1151 Opcode(Opc), Op0(V0), Op1(V1) {}
1152 operator bool() const { return Opcode != 0; }
1155 /// Binops may be transformed into binops with different opcodes and operands.
1156 /// Reverse the usual canonicalization to enable folds with the non-canonical
1157 /// form of the binop. If a transform is possible, return the elements of the
1158 /// new binop. If not, return invalid elements.
1159 static BinopElts getAlternateBinop(BinaryOperator *BO, const DataLayout &DL) {
1160 Value *BO0 = BO->getOperand(0), *BO1 = BO->getOperand(1);
1161 Type *Ty = BO->getType();
1162 switch (BO->getOpcode()) {
1163 case Instruction::Shl: {
1164 // shl X, C --> mul X, (1 << C)
1166 if (match(BO1, m_Constant(C))) {
1167 Constant *ShlOne = ConstantExpr::getShl(ConstantInt::get(Ty, 1), C);
1168 return { Instruction::Mul, BO0, ShlOne };
1172 case Instruction::Or: {
1173 // or X, C --> add X, C (when X and C have no common bits set)
1175 if (match(BO1, m_APInt(C)) && MaskedValueIsZero(BO0, *C, DL))
1176 return { Instruction::Add, BO0, BO1 };
1185 static Instruction *foldSelectShuffleWith1Binop(ShuffleVectorInst &Shuf) {
1186 assert(Shuf.isSelect() && "Must have select-equivalent shuffle");
1188 // Are we shuffling together some value and that same value after it has been
1189 // modified by a binop with a constant?
1190 Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1);
1193 if (match(Op0, m_BinOp(m_Specific(Op1), m_Constant(C))))
1195 else if (match(Op1, m_BinOp(m_Specific(Op0), m_Constant(C))))
1200 // The identity constant for a binop leaves a variable operand unchanged. For
1201 // a vector, this is a splat of something like 0, -1, or 1.
1202 // If there's no identity constant for this binop, we're done.
1203 auto *BO = cast<BinaryOperator>(Op0IsBinop ? Op0 : Op1);
1204 BinaryOperator::BinaryOps BOpcode = BO->getOpcode();
1205 Constant *IdC = ConstantExpr::getBinOpIdentity(BOpcode, Shuf.getType(), true);
1209 // Shuffle identity constants into the lanes that return the original value.
1210 // Example: shuf (mul X, {-1,-2,-3,-4}), X, {0,5,6,3} --> mul X, {-1,1,1,-4}
1211 // Example: shuf X, (add X, {-1,-2,-3,-4}), {0,1,6,7} --> add X, {0,0,-3,-4}
1212 // The existing binop constant vector remains in the same operand position.
1213 Constant *Mask = Shuf.getMask();
1214 Constant *NewC = Op0IsBinop ? ConstantExpr::getShuffleVector(C, IdC, Mask) :
1215 ConstantExpr::getShuffleVector(IdC, C, Mask);
1217 bool MightCreatePoisonOrUB =
1218 Mask->containsUndefElement() &&
1219 (Instruction::isIntDivRem(BOpcode) || Instruction::isShift(BOpcode));
1220 if (MightCreatePoisonOrUB)
1221 NewC = getSafeVectorConstantForBinop(BOpcode, NewC, true);
1223 // shuf (bop X, C), X, M --> bop X, C'
1224 // shuf X, (bop X, C), M --> bop X, C'
1225 Value *X = Op0IsBinop ? Op1 : Op0;
1226 Instruction *NewBO = BinaryOperator::Create(BOpcode, X, NewC);
1227 NewBO->copyIRFlags(BO);
1229 // An undef shuffle mask element may propagate as an undef constant element in
1230 // the new binop. That would produce poison where the original code might not.
1231 // If we already made a safe constant, then there's no danger.
1232 if (Mask->containsUndefElement() && !MightCreatePoisonOrUB)
1233 NewBO->dropPoisonGeneratingFlags();
1237 /// Try to fold shuffles that are the equivalent of a vector select.
1238 static Instruction *foldSelectShuffle(ShuffleVectorInst &Shuf,
1239 InstCombiner::BuilderTy &Builder,
1240 const DataLayout &DL) {
1241 if (!Shuf.isSelect())
1244 if (Instruction *I = foldSelectShuffleWith1Binop(Shuf))
1247 BinaryOperator *B0, *B1;
1248 if (!match(Shuf.getOperand(0), m_BinOp(B0)) ||
1249 !match(Shuf.getOperand(1), m_BinOp(B1)))
1254 bool ConstantsAreOp1;
1255 if (match(B0, m_BinOp(m_Value(X), m_Constant(C0))) &&
1256 match(B1, m_BinOp(m_Value(Y), m_Constant(C1))))
1257 ConstantsAreOp1 = true;
1258 else if (match(B0, m_BinOp(m_Constant(C0), m_Value(X))) &&
1259 match(B1, m_BinOp(m_Constant(C1), m_Value(Y))))
1260 ConstantsAreOp1 = false;
1264 // We need matching binops to fold the lanes together.
1265 BinaryOperator::BinaryOps Opc0 = B0->getOpcode();
1266 BinaryOperator::BinaryOps Opc1 = B1->getOpcode();
1267 bool DropNSW = false;
1268 if (ConstantsAreOp1 && Opc0 != Opc1) {
1269 // TODO: We drop "nsw" if shift is converted into multiply because it may
1270 // not be correct when the shift amount is BitWidth - 1. We could examine
1271 // each vector element to determine if it is safe to keep that flag.
1272 if (Opc0 == Instruction::Shl || Opc1 == Instruction::Shl)
1274 if (BinopElts AltB0 = getAlternateBinop(B0, DL)) {
1275 assert(isa<Constant>(AltB0.Op1) && "Expecting constant with alt binop");
1276 Opc0 = AltB0.Opcode;
1277 C0 = cast<Constant>(AltB0.Op1);
1278 } else if (BinopElts AltB1 = getAlternateBinop(B1, DL)) {
1279 assert(isa<Constant>(AltB1.Op1) && "Expecting constant with alt binop");
1280 Opc1 = AltB1.Opcode;
1281 C1 = cast<Constant>(AltB1.Op1);
1288 // The opcodes must be the same. Use a new name to make that clear.
1289 BinaryOperator::BinaryOps BOpc = Opc0;
1291 // Select the constant elements needed for the single binop.
1292 Constant *Mask = Shuf.getMask();
1293 Constant *NewC = ConstantExpr::getShuffleVector(C0, C1, Mask);
1295 // We are moving a binop after a shuffle. When a shuffle has an undefined
1296 // mask element, the result is undefined, but it is not poison or undefined
1297 // behavior. That is not necessarily true for div/rem/shift.
1298 bool MightCreatePoisonOrUB =
1299 Mask->containsUndefElement() &&
1300 (Instruction::isIntDivRem(BOpc) || Instruction::isShift(BOpc));
1301 if (MightCreatePoisonOrUB)
1302 NewC = getSafeVectorConstantForBinop(BOpc, NewC, ConstantsAreOp1);
1306 // Remove a binop and the shuffle by rearranging the constant:
1307 // shuffle (op V, C0), (op V, C1), M --> op V, C'
1308 // shuffle (op C0, V), (op C1, V), M --> op C', V
1311 // If there are 2 different variable operands, we must create a new shuffle
1312 // (select) first, so check uses to ensure that we don't end up with more
1313 // instructions than we started with.
1314 if (!B0->hasOneUse() && !B1->hasOneUse())
1317 // If we use the original shuffle mask and op1 is *variable*, we would be
1318 // putting an undef into operand 1 of div/rem/shift. This is either UB or
1319 // poison. We do not have to guard against UB when *constants* are op1
1320 // because safe constants guarantee that we do not overflow sdiv/srem (and
1321 // there's no danger for other opcodes).
1322 // TODO: To allow this case, create a new shuffle mask with no undefs.
1323 if (MightCreatePoisonOrUB && !ConstantsAreOp1)
1326 // Note: In general, we do not create new shuffles in InstCombine because we
1327 // do not know if a target can lower an arbitrary shuffle optimally. In this
1328 // case, the shuffle uses the existing mask, so there is no additional risk.
1330 // Select the variable vectors first, then perform the binop:
1331 // shuffle (op X, C0), (op Y, C1), M --> op (shuffle X, Y, M), C'
1332 // shuffle (op C0, X), (op C1, Y), M --> op C', (shuffle X, Y, M)
1333 V = Builder.CreateShuffleVector(X, Y, Mask);
1336 Instruction *NewBO = ConstantsAreOp1 ? BinaryOperator::Create(BOpc, V, NewC) :
1337 BinaryOperator::Create(BOpc, NewC, V);
1339 // Flags are intersected from the 2 source binops. But there are 2 exceptions:
1340 // 1. If we changed an opcode, poison conditions might have changed.
1341 // 2. If the shuffle had undef mask elements, the new binop might have undefs
1342 // where the original code did not. But if we already made a safe constant,
1343 // then there's no danger.
1344 NewBO->copyIRFlags(B0);
1345 NewBO->andIRFlags(B1);
1347 NewBO->setHasNoSignedWrap(false);
1348 if (Mask->containsUndefElement() && !MightCreatePoisonOrUB)
1349 NewBO->dropPoisonGeneratingFlags();
1353 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
1354 Value *LHS = SVI.getOperand(0);
1355 Value *RHS = SVI.getOperand(1);
1356 SmallVector<int, 16> Mask = SVI.getShuffleMask();
1357 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
1359 if (auto *V = SimplifyShuffleVectorInst(
1360 LHS, RHS, SVI.getMask(), SVI.getType(), SQ.getWithInstruction(&SVI)))
1361 return replaceInstUsesWith(SVI, V);
1363 if (Instruction *I = foldSelectShuffle(SVI, Builder, DL))
1366 bool MadeChange = false;
1367 unsigned VWidth = SVI.getType()->getVectorNumElements();
1369 APInt UndefElts(VWidth, 0);
1370 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1371 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
1373 return replaceInstUsesWith(SVI, V);
1377 unsigned LHSWidth = LHS->getType()->getVectorNumElements();
1379 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
1380 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
1381 if (LHS == RHS || isa<UndefValue>(LHS)) {
1382 if (isa<UndefValue>(LHS) && LHS == RHS) {
1383 // shuffle(undef,undef,mask) -> undef.
1384 Value *Result = (VWidth == LHSWidth)
1385 ? LHS : UndefValue::get(SVI.getType());
1386 return replaceInstUsesWith(SVI, Result);
1389 // Remap any references to RHS to use LHS.
1390 SmallVector<Constant*, 16> Elts;
1391 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
1393 Elts.push_back(UndefValue::get(Int32Ty));
1397 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
1398 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
1399 Mask[i] = -1; // Turn into undef.
1400 Elts.push_back(UndefValue::get(Int32Ty));
1402 Mask[i] = Mask[i] % e; // Force to LHS.
1403 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i]));
1406 SVI.setOperand(0, SVI.getOperand(1));
1407 SVI.setOperand(1, UndefValue::get(RHS->getType()));
1408 SVI.setOperand(2, ConstantVector::get(Elts));
1409 LHS = SVI.getOperand(0);
1410 RHS = SVI.getOperand(1);
1414 if (VWidth == LHSWidth) {
1415 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
1416 bool isLHSID, isRHSID;
1417 recognizeIdentityMask(Mask, isLHSID, isRHSID);
1419 // Eliminate identity shuffles.
1420 if (isLHSID) return replaceInstUsesWith(SVI, LHS);
1421 if (isRHSID) return replaceInstUsesWith(SVI, RHS);
1424 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
1425 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
1426 return replaceInstUsesWith(SVI, V);
1429 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
1430 // a non-vector type. We can instead bitcast the original vector followed by
1431 // an extract of the desired element:
1433 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
1434 // <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1435 // %1 = bitcast <4 x i8> %sroa to i32
1437 // %bc = bitcast <16 x i8> %in to <4 x i32>
1438 // %ext = extractelement <4 x i32> %bc, i32 0
1440 // If the shuffle is extracting a contiguous range of values from the input
1441 // vector then each use which is a bitcast of the extracted size can be
1442 // replaced. This will work if the vector types are compatible, and the begin
1443 // index is aligned to a value in the casted vector type. If the begin index
1444 // isn't aligned then we can shuffle the original vector (keeping the same
1445 // vector type) before extracting.
1447 // This code will bail out if the target type is fundamentally incompatible
1448 // with vectors of the source type.
1450 // Example of <16 x i8>, target type i32:
1451 // Index range [4,8): v-----------v Will work.
1452 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1453 // <16 x i8>: | | | | | | | | | | | | | | | | |
1454 // <4 x i32>: | | | | |
1455 // +-----------+-----------+-----------+-----------+
1456 // Index range [6,10): ^-----------^ Needs an extra shuffle.
1457 // Target type i40: ^--------------^ Won't work, bail.
1458 if (isShuffleExtractingFromLHS(SVI, Mask)) {
1460 unsigned MaskElems = Mask.size();
1461 VectorType *SrcTy = cast<VectorType>(V->getType());
1462 unsigned VecBitWidth = SrcTy->getBitWidth();
1463 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
1464 assert(SrcElemBitWidth && "vector elements must have a bitwidth");
1465 unsigned SrcNumElems = SrcTy->getNumElements();
1466 SmallVector<BitCastInst *, 8> BCs;
1467 DenseMap<Type *, Value *> NewBCs;
1468 for (User *U : SVI.users())
1469 if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
1470 if (!BC->use_empty())
1471 // Only visit bitcasts that weren't previously handled.
1473 for (BitCastInst *BC : BCs) {
1474 unsigned BegIdx = Mask.front();
1475 Type *TgtTy = BC->getDestTy();
1476 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
1477 if (!TgtElemBitWidth)
1479 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
1480 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
1481 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
1482 if (!VecBitWidthsEqual)
1484 if (!VectorType::isValidElementType(TgtTy))
1486 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems);
1487 if (!BegIsAligned) {
1488 // Shuffle the input so [0,NumElements) contains the output, and
1489 // [NumElems,SrcNumElems) is undef.
1490 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems,
1491 UndefValue::get(Int32Ty));
1492 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
1493 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx);
1494 V = Builder.CreateShuffleVector(V, UndefValue::get(V->getType()),
1495 ConstantVector::get(ShuffleMask),
1496 SVI.getName() + ".extract");
1499 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
1500 assert(SrcElemsPerTgtElem);
1501 BegIdx /= SrcElemsPerTgtElem;
1502 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end();
1506 : Builder.CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
1507 if (!BCAlreadyExists)
1508 NewBCs[CastSrcTy] = NewBC;
1509 auto *Ext = Builder.CreateExtractElement(
1510 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract");
1511 // The shufflevector isn't being replaced: the bitcast that used it
1512 // is. InstCombine will visit the newly-created instructions.
1513 replaceInstUsesWith(*BC, Ext);
1518 // If the LHS is a shufflevector itself, see if we can combine it with this
1519 // one without producing an unusual shuffle.
1520 // Cases that might be simplified:
1522 // x1=shuffle(v1,v2,mask1)
1523 // x=shuffle(x1,undef,mask)
1525 // x=shuffle(v1,undef,newMask)
1526 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
1528 // x1=shuffle(v1,undef,mask1)
1529 // x=shuffle(x1,x2,mask)
1530 // where v1.size() == mask1.size()
1532 // x=shuffle(v1,x2,newMask)
1533 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
1535 // x2=shuffle(v2,undef,mask2)
1536 // x=shuffle(x1,x2,mask)
1537 // where v2.size() == mask2.size()
1539 // x=shuffle(x1,v2,newMask)
1540 // newMask[i] = (mask[i] < x1.size())
1541 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
1543 // x1=shuffle(v1,undef,mask1)
1544 // x2=shuffle(v2,undef,mask2)
1545 // x=shuffle(x1,x2,mask)
1546 // where v1.size() == v2.size()
1548 // x=shuffle(v1,v2,newMask)
1549 // newMask[i] = (mask[i] < x1.size())
1550 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
1552 // Here we are really conservative:
1553 // we are absolutely afraid of producing a shuffle mask not in the input
1554 // program, because the code gen may not be smart enough to turn a merged
1555 // shuffle into two specific shuffles: it may produce worse code. As such,
1556 // we only merge two shuffles if the result is either a splat or one of the
1557 // input shuffle masks. In this case, merging the shuffles just removes
1558 // one instruction, which we know is safe. This is good for things like
1559 // turning: (splat(splat)) -> splat, or
1560 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
1561 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
1562 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
1564 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
1565 LHSShuffle = nullptr;
1567 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
1568 RHSShuffle = nullptr;
1569 if (!LHSShuffle && !RHSShuffle)
1570 return MadeChange ? &SVI : nullptr;
1572 Value* LHSOp0 = nullptr;
1573 Value* LHSOp1 = nullptr;
1574 Value* RHSOp0 = nullptr;
1575 unsigned LHSOp0Width = 0;
1576 unsigned RHSOp0Width = 0;
1578 LHSOp0 = LHSShuffle->getOperand(0);
1579 LHSOp1 = LHSShuffle->getOperand(1);
1580 LHSOp0Width = LHSOp0->getType()->getVectorNumElements();
1583 RHSOp0 = RHSShuffle->getOperand(0);
1584 RHSOp0Width = RHSOp0->getType()->getVectorNumElements();
1586 Value* newLHS = LHS;
1587 Value* newRHS = RHS;
1590 if (isa<UndefValue>(RHS)) {
1595 else if (LHSOp0Width == LHSWidth) {
1600 if (RHSShuffle && RHSOp0Width == LHSWidth) {
1604 if (LHSOp0 == RHSOp0) {
1609 if (newLHS == LHS && newRHS == RHS)
1610 return MadeChange ? &SVI : nullptr;
1612 SmallVector<int, 16> LHSMask;
1613 SmallVector<int, 16> RHSMask;
1615 LHSMask = LHSShuffle->getShuffleMask();
1616 if (RHSShuffle && newRHS != RHS)
1617 RHSMask = RHSShuffle->getShuffleMask();
1619 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
1620 SmallVector<int, 16> newMask;
1621 bool isSplat = true;
1623 // Create a new mask for the new ShuffleVectorInst so that the new
1624 // ShuffleVectorInst is equivalent to the original one.
1625 for (unsigned i = 0; i < VWidth; ++i) {
1628 // This element is an undef value.
1630 } else if (Mask[i] < (int)LHSWidth) {
1631 // This element is from left hand side vector operand.
1633 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
1634 // new mask value for the element.
1635 if (newLHS != LHS) {
1636 eltMask = LHSMask[Mask[i]];
1637 // If the value selected is an undef value, explicitly specify it
1638 // with a -1 mask value.
1639 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
1644 // This element is from right hand side vector operand
1646 // If the value selected is an undef value, explicitly specify it
1647 // with a -1 mask value. (case 1)
1648 if (isa<UndefValue>(RHS))
1650 // If RHS is going to be replaced (case 3 or 4), calculate the
1651 // new mask value for the element.
1652 else if (newRHS != RHS) {
1653 eltMask = RHSMask[Mask[i]-LHSWidth];
1654 // If the value selected is an undef value, explicitly specify it
1655 // with a -1 mask value.
1656 if (eltMask >= (int)RHSOp0Width) {
1657 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1658 && "should have been check above");
1662 eltMask = Mask[i]-LHSWidth;
1664 // If LHS's width is changed, shift the mask value accordingly.
1665 // If newRHS == nullptr, i.e. LHSOp0 == RHSOp0, we want to remap any
1666 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1667 // If newRHS == newLHS, we want to remap any references from newRHS to
1668 // newLHS so that we can properly identify splats that may occur due to
1669 // obfuscation across the two vectors.
1670 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
1671 eltMask += newLHSWidth;
1674 // Check if this could still be a splat.
1676 if (SplatElt >= 0 && SplatElt != eltMask)
1681 newMask.push_back(eltMask);
1684 // If the result mask is equal to one of the original shuffle masks,
1685 // or is a splat, do the replacement.
1686 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1687 SmallVector<Constant*, 16> Elts;
1688 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1689 if (newMask[i] < 0) {
1690 Elts.push_back(UndefValue::get(Int32Ty));
1692 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1696 newRHS = UndefValue::get(newLHS->getType());
1697 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1700 // If the result mask is an identity, replace uses of this instruction with
1701 // corresponding argument.
1702 bool isLHSID, isRHSID;
1703 recognizeIdentityMask(newMask, isLHSID, isRHSID);
1704 if (isLHSID && VWidth == LHSOp0Width) return replaceInstUsesWith(SVI, newLHS);
1705 if (isRHSID && VWidth == RHSOp0Width) return replaceInstUsesWith(SVI, newRHS);
1707 return MadeChange ? &SVI : nullptr;