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/DenseMap.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/VectorUtils.h"
19 #include "llvm/IR/PatternMatch.h"
21 using namespace PatternMatch;
23 #define DEBUG_TYPE "instcombine"
25 /// Return true if the value is cheaper to scalarize than it is to leave as a
26 /// vector operation. isConstant indicates whether we're extracting one known
27 /// element. If false we're extracting a variable index.
28 static bool cheapToScalarize(Value *V, bool isConstant) {
29 if (Constant *C = dyn_cast<Constant>(V)) {
30 if (isConstant) return true;
32 // If all elts are the same, we can extract it and use any of the values.
33 if (Constant *Op0 = C->getAggregateElement(0U)) {
34 for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e;
36 if (C->getAggregateElement(i) != Op0)
41 Instruction *I = dyn_cast<Instruction>(V);
44 // Insert element gets simplified to the inserted element or is deleted if
45 // this is constant idx extract element and its a constant idx insertelt.
46 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
47 isa<ConstantInt>(I->getOperand(2)))
49 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
51 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
52 if (BO->hasOneUse() &&
53 (cheapToScalarize(BO->getOperand(0), isConstant) ||
54 cheapToScalarize(BO->getOperand(1), isConstant)))
56 if (CmpInst *CI = dyn_cast<CmpInst>(I))
57 if (CI->hasOneUse() &&
58 (cheapToScalarize(CI->getOperand(0), isConstant) ||
59 cheapToScalarize(CI->getOperand(1), isConstant)))
65 // If we have a PHI node with a vector type that is only used to feed
66 // itself and be an operand of extractelement at a constant location,
67 // try to replace the PHI of the vector type with a PHI of a scalar type.
68 Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) {
69 SmallVector<Instruction *, 2> Extracts;
70 // The users we want the PHI to have are:
71 // 1) The EI ExtractElement (we already know this)
72 // 2) Possibly more ExtractElements with the same index.
73 // 3) Another operand, which will feed back into the PHI.
74 Instruction *PHIUser = nullptr;
75 for (auto U : PN->users()) {
76 if (ExtractElementInst *EU = dyn_cast<ExtractElementInst>(U)) {
77 if (EI.getIndexOperand() == EU->getIndexOperand())
78 Extracts.push_back(EU);
81 } else if (!PHIUser) {
82 PHIUser = cast<Instruction>(U);
91 // Verify that this PHI user has one use, which is the PHI itself,
92 // and that it is a binary operation which is cheap to scalarize.
93 // otherwise return NULL.
94 if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) ||
95 !(isa<BinaryOperator>(PHIUser)) || !cheapToScalarize(PHIUser, true))
98 // Create a scalar PHI node that will replace the vector PHI node
99 // just before the current PHI node.
100 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
101 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN));
102 // Scalarize each PHI operand.
103 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
104 Value *PHIInVal = PN->getIncomingValue(i);
105 BasicBlock *inBB = PN->getIncomingBlock(i);
106 Value *Elt = EI.getIndexOperand();
107 // If the operand is the PHI induction variable:
108 if (PHIInVal == PHIUser) {
109 // Scalarize the binary operation. Its first operand is the
110 // scalar PHI, and the second operand is extracted from the other
112 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
113 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
114 Value *Op = InsertNewInstWith(
115 ExtractElementInst::Create(B0->getOperand(opId), Elt,
116 B0->getOperand(opId)->getName() + ".Elt"),
118 Value *newPHIUser = InsertNewInstWith(
119 BinaryOperator::CreateWithCopiedFlags(B0->getOpcode(),
120 scalarPHI, Op, B0), *B0);
121 scalarPHI->addIncoming(newPHIUser, inBB);
123 // Scalarize PHI input:
124 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
125 // Insert the new instruction into the predecessor basic block.
126 Instruction *pos = dyn_cast<Instruction>(PHIInVal);
127 BasicBlock::iterator InsertPos;
128 if (pos && !isa<PHINode>(pos)) {
129 InsertPos = ++pos->getIterator();
131 InsertPos = inBB->getFirstInsertionPt();
134 InsertNewInstWith(newEI, *InsertPos);
136 scalarPHI->addIncoming(newEI, inBB);
140 for (auto E : Extracts)
141 replaceInstUsesWith(*E, scalarPHI);
146 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
147 if (Value *V = SimplifyExtractElementInst(EI.getVectorOperand(),
148 EI.getIndexOperand(),
149 SQ.getWithInstruction(&EI)))
150 return replaceInstUsesWith(EI, V);
152 // If vector val is constant with all elements the same, replace EI with
153 // that element. We handle a known element # below.
154 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0)))
155 if (cheapToScalarize(C, false))
156 return replaceInstUsesWith(EI, C->getAggregateElement(0U));
158 // If extracting a specified index from the vector, see if we can recursively
159 // find a previously computed scalar that was inserted into the vector.
160 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
161 unsigned IndexVal = IdxC->getZExtValue();
162 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
164 // InstSimplify handles cases where the index is invalid.
165 assert(IndexVal < VectorWidth);
167 // This instruction only demands the single element from the input vector.
168 // If the input vector has a single use, simplify it based on this use
170 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
171 APInt UndefElts(VectorWidth, 0);
172 APInt DemandedMask(VectorWidth, 0);
173 DemandedMask.setBit(IndexVal);
174 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), DemandedMask,
181 // If this extractelement is directly using a bitcast from a vector of
182 // the same number of elements, see if we can find the source element from
183 // it. In this case, we will end up needing to bitcast the scalars.
184 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
185 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
186 if (VT->getNumElements() == VectorWidth)
187 if (Value *Elt = findScalarElement(BCI->getOperand(0), IndexVal))
188 return new BitCastInst(Elt, EI.getType());
191 // If there's a vector PHI feeding a scalar use through this extractelement
192 // instruction, try to scalarize the PHI.
193 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
194 Instruction *scalarPHI = scalarizePHI(EI, PN);
200 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
201 // Push extractelement into predecessor operation if legal and
202 // profitable to do so.
203 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
204 if (I->hasOneUse() &&
205 cheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
207 Builder.CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
208 EI.getName()+".lhs");
210 Builder.CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
211 EI.getName()+".rhs");
212 return BinaryOperator::CreateWithCopiedFlags(BO->getOpcode(),
215 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
216 // Extracting the inserted element?
217 if (IE->getOperand(2) == EI.getOperand(1))
218 return replaceInstUsesWith(EI, IE->getOperand(1));
219 // If the inserted and extracted elements are constants, they must not
220 // be the same value, extract from the pre-inserted value instead.
221 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
222 Worklist.AddValue(EI.getOperand(0));
223 EI.setOperand(0, IE->getOperand(0));
226 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
227 // If this is extracting an element from a shufflevector, figure out where
228 // it came from and extract from the appropriate input element instead.
229 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
230 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
233 SVI->getOperand(0)->getType()->getVectorNumElements();
236 return replaceInstUsesWith(EI, UndefValue::get(EI.getType()));
237 if (SrcIdx < (int)LHSWidth)
238 Src = SVI->getOperand(0);
241 Src = SVI->getOperand(1);
243 Type *Int32Ty = Type::getInt32Ty(EI.getContext());
244 return ExtractElementInst::Create(Src,
245 ConstantInt::get(Int32Ty,
248 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
249 // Canonicalize extractelement(cast) -> cast(extractelement).
250 // Bitcasts can change the number of vector elements, and they cost
252 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
253 Value *EE = Builder.CreateExtractElement(CI->getOperand(0),
254 EI.getIndexOperand());
255 Worklist.AddValue(EE);
256 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
258 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
259 if (SI->hasOneUse()) {
260 // TODO: For a select on vectors, it might be useful to do this if it
261 // has multiple extractelement uses. For vector select, that seems to
262 // fight the vectorizer.
264 // If we are extracting an element from a vector select or a select on
265 // vectors, create a select on the scalars extracted from the vector
267 Value *TrueVal = SI->getTrueValue();
268 Value *FalseVal = SI->getFalseValue();
270 Value *Cond = SI->getCondition();
271 if (Cond->getType()->isVectorTy()) {
272 Cond = Builder.CreateExtractElement(Cond,
273 EI.getIndexOperand(),
274 Cond->getName() + ".elt");
278 = Builder.CreateExtractElement(TrueVal,
279 EI.getIndexOperand(),
280 TrueVal->getName() + ".elt");
283 = Builder.CreateExtractElement(FalseVal,
284 EI.getIndexOperand(),
285 FalseVal->getName() + ".elt");
286 return SelectInst::Create(Cond,
289 SI->getName() + ".elt");
296 /// If V is a shuffle of values that ONLY returns elements from either LHS or
297 /// RHS, return the shuffle mask and true. Otherwise, return false.
298 static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
299 SmallVectorImpl<Constant*> &Mask) {
300 assert(LHS->getType() == RHS->getType() &&
301 "Invalid CollectSingleShuffleElements");
302 unsigned NumElts = V->getType()->getVectorNumElements();
304 if (isa<UndefValue>(V)) {
305 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
310 for (unsigned i = 0; i != NumElts; ++i)
311 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
316 for (unsigned i = 0; i != NumElts; ++i)
317 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
322 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
323 // If this is an insert of an extract from some other vector, include it.
324 Value *VecOp = IEI->getOperand(0);
325 Value *ScalarOp = IEI->getOperand(1);
326 Value *IdxOp = IEI->getOperand(2);
328 if (!isa<ConstantInt>(IdxOp))
330 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
332 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
333 // We can handle this if the vector we are inserting into is
335 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
336 // If so, update the mask to reflect the inserted undef.
337 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
340 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
341 if (isa<ConstantInt>(EI->getOperand(1))) {
342 unsigned ExtractedIdx =
343 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
344 unsigned NumLHSElts = LHS->getType()->getVectorNumElements();
346 // This must be extracting from either LHS or RHS.
347 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
348 // We can handle this if the vector we are inserting into is
350 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
351 // If so, update the mask to reflect the inserted value.
352 if (EI->getOperand(0) == LHS) {
353 Mask[InsertedIdx % NumElts] =
354 ConstantInt::get(Type::getInt32Ty(V->getContext()),
357 assert(EI->getOperand(0) == RHS);
358 Mask[InsertedIdx % NumElts] =
359 ConstantInt::get(Type::getInt32Ty(V->getContext()),
360 ExtractedIdx + NumLHSElts);
372 /// If we have insertion into a vector that is wider than the vector that we
373 /// are extracting from, try to widen the source vector to allow a single
374 /// shufflevector to replace one or more insert/extract pairs.
375 static void replaceExtractElements(InsertElementInst *InsElt,
376 ExtractElementInst *ExtElt,
378 VectorType *InsVecType = InsElt->getType();
379 VectorType *ExtVecType = ExtElt->getVectorOperandType();
380 unsigned NumInsElts = InsVecType->getVectorNumElements();
381 unsigned NumExtElts = ExtVecType->getVectorNumElements();
383 // The inserted-to vector must be wider than the extracted-from vector.
384 if (InsVecType->getElementType() != ExtVecType->getElementType() ||
385 NumExtElts >= NumInsElts)
388 // Create a shuffle mask to widen the extended-from vector using undefined
389 // values. The mask selects all of the values of the original vector followed
390 // by as many undefined values as needed to create a vector of the same length
391 // as the inserted-to vector.
392 SmallVector<Constant *, 16> ExtendMask;
393 IntegerType *IntType = Type::getInt32Ty(InsElt->getContext());
394 for (unsigned i = 0; i < NumExtElts; ++i)
395 ExtendMask.push_back(ConstantInt::get(IntType, i));
396 for (unsigned i = NumExtElts; i < NumInsElts; ++i)
397 ExtendMask.push_back(UndefValue::get(IntType));
399 Value *ExtVecOp = ExtElt->getVectorOperand();
400 auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
401 BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
402 ? ExtVecOpInst->getParent()
403 : ExtElt->getParent();
405 // TODO: This restriction matches the basic block check below when creating
406 // new extractelement instructions. If that limitation is removed, this one
407 // could also be removed. But for now, we just bail out to ensure that we
408 // will replace the extractelement instruction that is feeding our
409 // insertelement instruction. This allows the insertelement to then be
410 // replaced by a shufflevector. If the insertelement is not replaced, we can
411 // induce infinite looping because there's an optimization for extractelement
412 // that will delete our widening shuffle. This would trigger another attempt
413 // here to create that shuffle, and we spin forever.
414 if (InsertionBlock != InsElt->getParent())
417 // TODO: This restriction matches the check in visitInsertElementInst() and
418 // prevents an infinite loop caused by not turning the extract/insert pair
419 // into a shuffle. We really should not need either check, but we're lacking
420 // folds for shufflevectors because we're afraid to generate shuffle masks
421 // that the backend can't handle.
422 if (InsElt->hasOneUse() && isa<InsertElementInst>(InsElt->user_back()))
425 auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
426 ConstantVector::get(ExtendMask));
428 // Insert the new shuffle after the vector operand of the extract is defined
429 // (as long as it's not a PHI) or at the start of the basic block of the
430 // extract, so any subsequent extracts in the same basic block can use it.
431 // TODO: Insert before the earliest ExtractElementInst that is replaced.
432 if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
433 WideVec->insertAfter(ExtVecOpInst);
435 IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
437 // Replace extracts from the original narrow vector with extracts from the new
439 for (User *U : ExtVecOp->users()) {
440 ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
441 if (!OldExt || OldExt->getParent() != WideVec->getParent())
443 auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
444 NewExt->insertAfter(OldExt);
445 IC.replaceInstUsesWith(*OldExt, NewExt);
449 /// We are building a shuffle to create V, which is a sequence of insertelement,
450 /// extractelement pairs. If PermittedRHS is set, then we must either use it or
451 /// not rely on the second vector source. Return a std::pair containing the
452 /// left and right vectors of the proposed shuffle (or 0), and set the Mask
453 /// parameter as required.
455 /// Note: we intentionally don't try to fold earlier shuffles since they have
456 /// often been chosen carefully to be efficiently implementable on the target.
457 typedef std::pair<Value *, Value *> ShuffleOps;
459 static ShuffleOps collectShuffleElements(Value *V,
460 SmallVectorImpl<Constant *> &Mask,
463 assert(V->getType()->isVectorTy() && "Invalid shuffle!");
464 unsigned NumElts = V->getType()->getVectorNumElements();
466 if (isa<UndefValue>(V)) {
467 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
468 return std::make_pair(
469 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
472 if (isa<ConstantAggregateZero>(V)) {
473 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
474 return std::make_pair(V, nullptr);
477 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
478 // If this is an insert of an extract from some other vector, include it.
479 Value *VecOp = IEI->getOperand(0);
480 Value *ScalarOp = IEI->getOperand(1);
481 Value *IdxOp = IEI->getOperand(2);
483 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
484 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
485 unsigned ExtractedIdx =
486 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
487 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
489 // Either the extracted from or inserted into vector must be RHSVec,
490 // otherwise we'd end up with a shuffle of three inputs.
491 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
492 Value *RHS = EI->getOperand(0);
493 ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC);
494 assert(LR.second == nullptr || LR.second == RHS);
496 if (LR.first->getType() != RHS->getType()) {
497 // Although we are giving up for now, see if we can create extracts
498 // that match the inserts for another round of combining.
499 replaceExtractElements(IEI, EI, IC);
501 // We tried our best, but we can't find anything compatible with RHS
502 // further up the chain. Return a trivial shuffle.
503 for (unsigned i = 0; i < NumElts; ++i)
504 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
505 return std::make_pair(V, nullptr);
508 unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
509 Mask[InsertedIdx % NumElts] =
510 ConstantInt::get(Type::getInt32Ty(V->getContext()),
511 NumLHSElts+ExtractedIdx);
512 return std::make_pair(LR.first, RHS);
515 if (VecOp == PermittedRHS) {
516 // We've gone as far as we can: anything on the other side of the
517 // extractelement will already have been converted into a shuffle.
518 unsigned NumLHSElts =
519 EI->getOperand(0)->getType()->getVectorNumElements();
520 for (unsigned i = 0; i != NumElts; ++i)
521 Mask.push_back(ConstantInt::get(
522 Type::getInt32Ty(V->getContext()),
523 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
524 return std::make_pair(EI->getOperand(0), PermittedRHS);
527 // If this insertelement is a chain that comes from exactly these two
528 // vectors, return the vector and the effective shuffle.
529 if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
530 collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
532 return std::make_pair(EI->getOperand(0), PermittedRHS);
537 // Otherwise, we can't do anything fancy. Return an identity vector.
538 for (unsigned i = 0; i != NumElts; ++i)
539 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
540 return std::make_pair(V, nullptr);
543 /// Try to find redundant insertvalue instructions, like the following ones:
544 /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
545 /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
546 /// Here the second instruction inserts values at the same indices, as the
547 /// first one, making the first one redundant.
548 /// It should be transformed to:
549 /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
550 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
551 bool IsRedundant = false;
552 ArrayRef<unsigned int> FirstIndices = I.getIndices();
554 // If there is a chain of insertvalue instructions (each of them except the
555 // last one has only one use and it's another insertvalue insn from this
556 // chain), check if any of the 'children' uses the same indices as the first
557 // instruction. In this case, the first one is redundant.
560 while (V->hasOneUse() && Depth < 10) {
561 User *U = V->user_back();
562 auto UserInsInst = dyn_cast<InsertValueInst>(U);
563 if (!UserInsInst || U->getOperand(0) != V)
565 if (UserInsInst->getIndices() == FirstIndices) {
574 return replaceInstUsesWith(I, I.getOperand(0));
578 static bool isShuffleEquivalentToSelect(ShuffleVectorInst &Shuf) {
579 int MaskSize = Shuf.getMask()->getType()->getVectorNumElements();
580 int VecSize = Shuf.getOperand(0)->getType()->getVectorNumElements();
582 // A vector select does not change the size of the operands.
583 if (MaskSize != VecSize)
586 // Each mask element must be undefined or choose a vector element from one of
587 // the source operands without crossing vector lanes.
588 for (int i = 0; i != MaskSize; ++i) {
589 int Elt = Shuf.getMaskValue(i);
590 if (Elt != -1 && Elt != i && Elt != i + VecSize)
597 // Turn a chain of inserts that splats a value into a canonical insert + shuffle
599 // insertelt(insertelt(insertelt(insertelt X, %k, 0), %k, 1), %k, 2) ... ->
600 // shufflevector(insertelt(X, %k, 0), undef, zero)
601 static Instruction *foldInsSequenceIntoBroadcast(InsertElementInst &InsElt) {
602 // We are interested in the last insert in a chain. So, if this insert
603 // has a single user, and that user is an insert, bail.
604 if (InsElt.hasOneUse() && isa<InsertElementInst>(InsElt.user_back()))
607 VectorType *VT = cast<VectorType>(InsElt.getType());
608 int NumElements = VT->getNumElements();
610 // Do not try to do this for a one-element vector, since that's a nop,
611 // and will cause an inf-loop.
612 if (NumElements == 1)
615 Value *SplatVal = InsElt.getOperand(1);
616 InsertElementInst *CurrIE = &InsElt;
617 SmallVector<bool, 16> ElementPresent(NumElements, false);
619 // Walk the chain backwards, keeping track of which indices we inserted into,
620 // until we hit something that isn't an insert of the splatted value.
622 ConstantInt *Idx = dyn_cast<ConstantInt>(CurrIE->getOperand(2));
623 if (!Idx || CurrIE->getOperand(1) != SplatVal)
626 // Check none of the intermediate steps have any additional uses.
627 if ((CurrIE != &InsElt) && !CurrIE->hasOneUse())
630 ElementPresent[Idx->getZExtValue()] = true;
631 CurrIE = dyn_cast<InsertElementInst>(CurrIE->getOperand(0));
634 // Make sure we've seen an insert into every element.
635 if (llvm::any_of(ElementPresent, [](bool Present) { return !Present; }))
638 // All right, create the insert + shuffle.
639 Instruction *InsertFirst = InsertElementInst::Create(
640 UndefValue::get(VT), SplatVal,
641 ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), 0), "", &InsElt);
643 Constant *ZeroMask = ConstantAggregateZero::get(
644 VectorType::get(Type::getInt32Ty(InsElt.getContext()), NumElements));
646 return new ShuffleVectorInst(InsertFirst, UndefValue::get(VT), ZeroMask);
649 /// If we have an insertelement instruction feeding into another insertelement
650 /// and the 2nd is inserting a constant into the vector, canonicalize that
651 /// constant insertion before the insertion of a variable:
653 /// insertelement (insertelement X, Y, IdxC1), ScalarC, IdxC2 -->
654 /// insertelement (insertelement X, ScalarC, IdxC2), Y, IdxC1
656 /// This has the potential of eliminating the 2nd insertelement instruction
657 /// via constant folding of the scalar constant into a vector constant.
658 static Instruction *hoistInsEltConst(InsertElementInst &InsElt2,
659 InstCombiner::BuilderTy &Builder) {
660 auto *InsElt1 = dyn_cast<InsertElementInst>(InsElt2.getOperand(0));
661 if (!InsElt1 || !InsElt1->hasOneUse())
666 ConstantInt *IdxC1, *IdxC2;
667 if (match(InsElt1->getOperand(0), m_Value(X)) &&
668 match(InsElt1->getOperand(1), m_Value(Y)) && !isa<Constant>(Y) &&
669 match(InsElt1->getOperand(2), m_ConstantInt(IdxC1)) &&
670 match(InsElt2.getOperand(1), m_Constant(ScalarC)) &&
671 match(InsElt2.getOperand(2), m_ConstantInt(IdxC2)) && IdxC1 != IdxC2) {
672 Value *NewInsElt1 = Builder.CreateInsertElement(X, ScalarC, IdxC2);
673 return InsertElementInst::Create(NewInsElt1, Y, IdxC1);
679 /// insertelt (shufflevector X, CVec, Mask|insertelt X, C1, CIndex1), C, CIndex
680 /// --> shufflevector X, CVec', Mask'
681 static Instruction *foldConstantInsEltIntoShuffle(InsertElementInst &InsElt) {
682 auto *Inst = dyn_cast<Instruction>(InsElt.getOperand(0));
683 // Bail out if the parent has more than one use. In that case, we'd be
684 // replacing the insertelt with a shuffle, and that's not a clear win.
685 if (!Inst || !Inst->hasOneUse())
687 if (auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0))) {
688 // The shuffle must have a constant vector operand. The insertelt must have
689 // a constant scalar being inserted at a constant position in the vector.
690 Constant *ShufConstVec, *InsEltScalar;
691 uint64_t InsEltIndex;
692 if (!match(Shuf->getOperand(1), m_Constant(ShufConstVec)) ||
693 !match(InsElt.getOperand(1), m_Constant(InsEltScalar)) ||
694 !match(InsElt.getOperand(2), m_ConstantInt(InsEltIndex)))
697 // Adding an element to an arbitrary shuffle could be expensive, but a
698 // shuffle that selects elements from vectors without crossing lanes is
700 // If we're just adding a constant into that shuffle, it will still be
702 if (!isShuffleEquivalentToSelect(*Shuf))
705 // From the above 'select' check, we know that the mask has the same number
706 // of elements as the vector input operands. We also know that each constant
707 // input element is used in its lane and can not be used more than once by
708 // the shuffle. Therefore, replace the constant in the shuffle's constant
709 // vector with the insertelt constant. Replace the constant in the shuffle's
710 // mask vector with the insertelt index plus the length of the vector
711 // (because the constant vector operand of a shuffle is always the 2nd
713 Constant *Mask = Shuf->getMask();
714 unsigned NumElts = Mask->getType()->getVectorNumElements();
715 SmallVector<Constant *, 16> NewShufElts(NumElts);
716 SmallVector<Constant *, 16> NewMaskElts(NumElts);
717 for (unsigned I = 0; I != NumElts; ++I) {
718 if (I == InsEltIndex) {
719 NewShufElts[I] = InsEltScalar;
720 Type *Int32Ty = Type::getInt32Ty(Shuf->getContext());
721 NewMaskElts[I] = ConstantInt::get(Int32Ty, InsEltIndex + NumElts);
723 // Copy over the existing values.
724 NewShufElts[I] = ShufConstVec->getAggregateElement(I);
725 NewMaskElts[I] = Mask->getAggregateElement(I);
729 // Create new operands for a shuffle that includes the constant of the
730 // original insertelt. The old shuffle will be dead now.
731 return new ShuffleVectorInst(Shuf->getOperand(0),
732 ConstantVector::get(NewShufElts),
733 ConstantVector::get(NewMaskElts));
734 } else if (auto *IEI = dyn_cast<InsertElementInst>(Inst)) {
735 // Transform sequences of insertelements ops with constant data/indexes into
736 // a single shuffle op.
737 unsigned NumElts = InsElt.getType()->getNumElements();
739 uint64_t InsertIdx[2];
741 if (!match(InsElt.getOperand(2), m_ConstantInt(InsertIdx[0])) ||
742 !match(InsElt.getOperand(1), m_Constant(Val[0])) ||
743 !match(IEI->getOperand(2), m_ConstantInt(InsertIdx[1])) ||
744 !match(IEI->getOperand(1), m_Constant(Val[1])))
746 SmallVector<Constant *, 16> Values(NumElts);
747 SmallVector<Constant *, 16> Mask(NumElts);
748 auto ValI = std::begin(Val);
749 // Generate new constant vector and mask.
750 // We have 2 values/masks from the insertelements instructions. Insert them
751 // into new value/mask vectors.
752 for (uint64_t I : InsertIdx) {
756 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()),
761 // Remaining values are filled with 'undef' values.
762 for (unsigned I = 0; I < NumElts; ++I) {
765 Values[I] = UndefValue::get(InsElt.getType()->getElementType());
766 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), I);
769 // Create new operands for a shuffle that includes the constant of the
770 // original insertelt.
771 return new ShuffleVectorInst(IEI->getOperand(0),
772 ConstantVector::get(Values),
773 ConstantVector::get(Mask));
778 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
779 Value *VecOp = IE.getOperand(0);
780 Value *ScalarOp = IE.getOperand(1);
781 Value *IdxOp = IE.getOperand(2);
783 // Inserting an undef or into an undefined place, remove this.
784 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
785 replaceInstUsesWith(IE, VecOp);
787 // If the inserted element was extracted from some other vector, and if the
788 // indexes are constant, try to turn this into a shufflevector operation.
789 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
790 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
791 unsigned NumInsertVectorElts = IE.getType()->getNumElements();
792 unsigned NumExtractVectorElts =
793 EI->getOperand(0)->getType()->getVectorNumElements();
794 unsigned ExtractedIdx =
795 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
796 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
798 if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
799 return replaceInstUsesWith(IE, VecOp);
801 if (InsertedIdx >= NumInsertVectorElts) // Out of range insert.
802 return replaceInstUsesWith(IE, UndefValue::get(IE.getType()));
804 // If we are extracting a value from a vector, then inserting it right
805 // back into the same place, just use the input vector.
806 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
807 return replaceInstUsesWith(IE, VecOp);
809 // If this insertelement isn't used by some other insertelement, turn it
810 // (and any insertelements it points to), into one big shuffle.
811 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
812 SmallVector<Constant*, 16> Mask;
813 ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this);
815 // The proposed shuffle may be trivial, in which case we shouldn't
816 // perform the combine.
817 if (LR.first != &IE && LR.second != &IE) {
818 // We now have a shuffle of LHS, RHS, Mask.
819 if (LR.second == nullptr)
820 LR.second = UndefValue::get(LR.first->getType());
821 return new ShuffleVectorInst(LR.first, LR.second,
822 ConstantVector::get(Mask));
828 unsigned VWidth = VecOp->getType()->getVectorNumElements();
829 APInt UndefElts(VWidth, 0);
830 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
831 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
833 return replaceInstUsesWith(IE, V);
837 if (Instruction *Shuf = foldConstantInsEltIntoShuffle(IE))
840 if (Instruction *NewInsElt = hoistInsEltConst(IE, Builder))
843 // Turn a sequence of inserts that broadcasts a scalar into a single
844 // insert + shufflevector.
845 if (Instruction *Broadcast = foldInsSequenceIntoBroadcast(IE))
851 /// Return true if we can evaluate the specified expression tree if the vector
852 /// elements were shuffled in a different order.
853 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
854 unsigned Depth = 5) {
855 // We can always reorder the elements of a constant.
856 if (isa<Constant>(V))
859 // We won't reorder vector arguments. No IPO here.
860 Instruction *I = dyn_cast<Instruction>(V);
861 if (!I) return false;
863 // Two users may expect different orders of the elements. Don't try it.
867 if (Depth == 0) return false;
869 switch (I->getOpcode()) {
870 case Instruction::Add:
871 case Instruction::FAdd:
872 case Instruction::Sub:
873 case Instruction::FSub:
874 case Instruction::Mul:
875 case Instruction::FMul:
876 case Instruction::UDiv:
877 case Instruction::SDiv:
878 case Instruction::FDiv:
879 case Instruction::URem:
880 case Instruction::SRem:
881 case Instruction::FRem:
882 case Instruction::Shl:
883 case Instruction::LShr:
884 case Instruction::AShr:
885 case Instruction::And:
886 case Instruction::Or:
887 case Instruction::Xor:
888 case Instruction::ICmp:
889 case Instruction::FCmp:
890 case Instruction::Trunc:
891 case Instruction::ZExt:
892 case Instruction::SExt:
893 case Instruction::FPToUI:
894 case Instruction::FPToSI:
895 case Instruction::UIToFP:
896 case Instruction::SIToFP:
897 case Instruction::FPTrunc:
898 case Instruction::FPExt:
899 case Instruction::GetElementPtr: {
900 for (Value *Operand : I->operands()) {
901 if (!CanEvaluateShuffled(Operand, Mask, Depth-1))
906 case Instruction::InsertElement: {
907 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
908 if (!CI) return false;
909 int ElementNumber = CI->getLimitedValue();
911 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
912 // can't put an element into multiple indices.
913 bool SeenOnce = false;
914 for (int i = 0, e = Mask.size(); i != e; ++i) {
915 if (Mask[i] == ElementNumber) {
921 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
927 /// Rebuild a new instruction just like 'I' but with the new operands given.
928 /// In the event of type mismatch, the type of the operands is correct.
929 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) {
930 // We don't want to use the IRBuilder here because we want the replacement
931 // instructions to appear next to 'I', not the builder's insertion point.
932 switch (I->getOpcode()) {
933 case Instruction::Add:
934 case Instruction::FAdd:
935 case Instruction::Sub:
936 case Instruction::FSub:
937 case Instruction::Mul:
938 case Instruction::FMul:
939 case Instruction::UDiv:
940 case Instruction::SDiv:
941 case Instruction::FDiv:
942 case Instruction::URem:
943 case Instruction::SRem:
944 case Instruction::FRem:
945 case Instruction::Shl:
946 case Instruction::LShr:
947 case Instruction::AShr:
948 case Instruction::And:
949 case Instruction::Or:
950 case Instruction::Xor: {
951 BinaryOperator *BO = cast<BinaryOperator>(I);
952 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
953 BinaryOperator *New =
954 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
955 NewOps[0], NewOps[1], "", BO);
956 if (isa<OverflowingBinaryOperator>(BO)) {
957 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
958 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
960 if (isa<PossiblyExactOperator>(BO)) {
961 New->setIsExact(BO->isExact());
963 if (isa<FPMathOperator>(BO))
964 New->copyFastMathFlags(I);
967 case Instruction::ICmp:
968 assert(NewOps.size() == 2 && "icmp with #ops != 2");
969 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
970 NewOps[0], NewOps[1]);
971 case Instruction::FCmp:
972 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
973 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
974 NewOps[0], NewOps[1]);
975 case Instruction::Trunc:
976 case Instruction::ZExt:
977 case Instruction::SExt:
978 case Instruction::FPToUI:
979 case Instruction::FPToSI:
980 case Instruction::UIToFP:
981 case Instruction::SIToFP:
982 case Instruction::FPTrunc:
983 case Instruction::FPExt: {
984 // It's possible that the mask has a different number of elements from
985 // the original cast. We recompute the destination type to match the mask.
987 VectorType::get(I->getType()->getScalarType(),
988 NewOps[0]->getType()->getVectorNumElements());
989 assert(NewOps.size() == 1 && "cast with #ops != 1");
990 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
993 case Instruction::GetElementPtr: {
994 Value *Ptr = NewOps[0];
995 ArrayRef<Value*> Idx = NewOps.slice(1);
996 GetElementPtrInst *GEP = GetElementPtrInst::Create(
997 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I);
998 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
1002 llvm_unreachable("failed to rebuild vector instructions");
1006 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
1007 // Mask.size() does not need to be equal to the number of vector elements.
1009 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
1010 if (isa<UndefValue>(V)) {
1011 return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
1014 if (isa<ConstantAggregateZero>(V)) {
1015 return ConstantAggregateZero::get(
1016 VectorType::get(V->getType()->getScalarType(),
1019 if (Constant *C = dyn_cast<Constant>(V)) {
1020 SmallVector<Constant *, 16> MaskValues;
1021 for (int i = 0, e = Mask.size(); i != e; ++i) {
1023 MaskValues.push_back(UndefValue::get(Builder.getInt32Ty()));
1025 MaskValues.push_back(Builder.getInt32(Mask[i]));
1027 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
1028 ConstantVector::get(MaskValues));
1031 Instruction *I = cast<Instruction>(V);
1032 switch (I->getOpcode()) {
1033 case Instruction::Add:
1034 case Instruction::FAdd:
1035 case Instruction::Sub:
1036 case Instruction::FSub:
1037 case Instruction::Mul:
1038 case Instruction::FMul:
1039 case Instruction::UDiv:
1040 case Instruction::SDiv:
1041 case Instruction::FDiv:
1042 case Instruction::URem:
1043 case Instruction::SRem:
1044 case Instruction::FRem:
1045 case Instruction::Shl:
1046 case Instruction::LShr:
1047 case Instruction::AShr:
1048 case Instruction::And:
1049 case Instruction::Or:
1050 case Instruction::Xor:
1051 case Instruction::ICmp:
1052 case Instruction::FCmp:
1053 case Instruction::Trunc:
1054 case Instruction::ZExt:
1055 case Instruction::SExt:
1056 case Instruction::FPToUI:
1057 case Instruction::FPToSI:
1058 case Instruction::UIToFP:
1059 case Instruction::SIToFP:
1060 case Instruction::FPTrunc:
1061 case Instruction::FPExt:
1062 case Instruction::Select:
1063 case Instruction::GetElementPtr: {
1064 SmallVector<Value*, 8> NewOps;
1065 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
1066 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
1067 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
1068 NewOps.push_back(V);
1069 NeedsRebuild |= (V != I->getOperand(i));
1072 return buildNew(I, NewOps);
1076 case Instruction::InsertElement: {
1077 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
1079 // The insertelement was inserting at Element. Figure out which element
1080 // that becomes after shuffling. The answer is guaranteed to be unique
1081 // by CanEvaluateShuffled.
1084 for (int e = Mask.size(); Index != e; ++Index) {
1085 if (Mask[Index] == Element) {
1091 // If element is not in Mask, no need to handle the operand 1 (element to
1092 // be inserted). Just evaluate values in operand 0 according to Mask.
1094 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
1096 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
1097 return InsertElementInst::Create(V, I->getOperand(1),
1098 Builder.getInt32(Index), "", I);
1101 llvm_unreachable("failed to reorder elements of vector instruction!");
1104 static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask,
1105 bool &isLHSID, bool &isRHSID) {
1106 isLHSID = isRHSID = true;
1108 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
1109 if (Mask[i] < 0) continue; // Ignore undef values.
1110 // Is this an identity shuffle of the LHS value?
1111 isLHSID &= (Mask[i] == (int)i);
1113 // Is this an identity shuffle of the RHS value?
1114 isRHSID &= (Mask[i]-e == i);
1118 // Returns true if the shuffle is extracting a contiguous range of values from
1119 // LHS, for example:
1120 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1121 // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
1122 // Shuffles to: |EE|FF|GG|HH|
1124 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
1125 SmallVector<int, 16> &Mask) {
1126 unsigned LHSElems = SVI.getOperand(0)->getType()->getVectorNumElements();
1127 unsigned MaskElems = Mask.size();
1128 unsigned BegIdx = Mask.front();
1129 unsigned EndIdx = Mask.back();
1130 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
1132 for (unsigned I = 0; I != MaskElems; ++I)
1133 if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
1138 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
1139 Value *LHS = SVI.getOperand(0);
1140 Value *RHS = SVI.getOperand(1);
1141 SmallVector<int, 16> Mask = SVI.getShuffleMask();
1142 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
1144 if (auto *V = SimplifyShuffleVectorInst(
1145 LHS, RHS, SVI.getMask(), SVI.getType(), SQ.getWithInstruction(&SVI)))
1146 return replaceInstUsesWith(SVI, V);
1148 bool MadeChange = false;
1149 unsigned VWidth = SVI.getType()->getVectorNumElements();
1151 APInt UndefElts(VWidth, 0);
1152 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1153 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
1155 return replaceInstUsesWith(SVI, V);
1156 LHS = SVI.getOperand(0);
1157 RHS = SVI.getOperand(1);
1161 unsigned LHSWidth = LHS->getType()->getVectorNumElements();
1163 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
1164 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
1165 if (LHS == RHS || isa<UndefValue>(LHS)) {
1166 if (isa<UndefValue>(LHS) && LHS == RHS) {
1167 // shuffle(undef,undef,mask) -> undef.
1168 Value *Result = (VWidth == LHSWidth)
1169 ? LHS : UndefValue::get(SVI.getType());
1170 return replaceInstUsesWith(SVI, Result);
1173 // Remap any references to RHS to use LHS.
1174 SmallVector<Constant*, 16> Elts;
1175 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
1177 Elts.push_back(UndefValue::get(Int32Ty));
1181 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
1182 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
1183 Mask[i] = -1; // Turn into undef.
1184 Elts.push_back(UndefValue::get(Int32Ty));
1186 Mask[i] = Mask[i] % e; // Force to LHS.
1187 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i]));
1190 SVI.setOperand(0, SVI.getOperand(1));
1191 SVI.setOperand(1, UndefValue::get(RHS->getType()));
1192 SVI.setOperand(2, ConstantVector::get(Elts));
1193 LHS = SVI.getOperand(0);
1194 RHS = SVI.getOperand(1);
1198 if (VWidth == LHSWidth) {
1199 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
1200 bool isLHSID, isRHSID;
1201 recognizeIdentityMask(Mask, isLHSID, isRHSID);
1203 // Eliminate identity shuffles.
1204 if (isLHSID) return replaceInstUsesWith(SVI, LHS);
1205 if (isRHSID) return replaceInstUsesWith(SVI, RHS);
1208 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
1209 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
1210 return replaceInstUsesWith(SVI, V);
1213 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
1214 // a non-vector type. We can instead bitcast the original vector followed by
1215 // an extract of the desired element:
1217 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
1218 // <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1219 // %1 = bitcast <4 x i8> %sroa to i32
1221 // %bc = bitcast <16 x i8> %in to <4 x i32>
1222 // %ext = extractelement <4 x i32> %bc, i32 0
1224 // If the shuffle is extracting a contiguous range of values from the input
1225 // vector then each use which is a bitcast of the extracted size can be
1226 // replaced. This will work if the vector types are compatible, and the begin
1227 // index is aligned to a value in the casted vector type. If the begin index
1228 // isn't aligned then we can shuffle the original vector (keeping the same
1229 // vector type) before extracting.
1231 // This code will bail out if the target type is fundamentally incompatible
1232 // with vectors of the source type.
1234 // Example of <16 x i8>, target type i32:
1235 // Index range [4,8): v-----------v Will work.
1236 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1237 // <16 x i8>: | | | | | | | | | | | | | | | | |
1238 // <4 x i32>: | | | | |
1239 // +-----------+-----------+-----------+-----------+
1240 // Index range [6,10): ^-----------^ Needs an extra shuffle.
1241 // Target type i40: ^--------------^ Won't work, bail.
1242 if (isShuffleExtractingFromLHS(SVI, Mask)) {
1244 unsigned MaskElems = Mask.size();
1245 VectorType *SrcTy = cast<VectorType>(V->getType());
1246 unsigned VecBitWidth = SrcTy->getBitWidth();
1247 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
1248 assert(SrcElemBitWidth && "vector elements must have a bitwidth");
1249 unsigned SrcNumElems = SrcTy->getNumElements();
1250 SmallVector<BitCastInst *, 8> BCs;
1251 DenseMap<Type *, Value *> NewBCs;
1252 for (User *U : SVI.users())
1253 if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
1254 if (!BC->use_empty())
1255 // Only visit bitcasts that weren't previously handled.
1257 for (BitCastInst *BC : BCs) {
1258 unsigned BegIdx = Mask.front();
1259 Type *TgtTy = BC->getDestTy();
1260 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
1261 if (!TgtElemBitWidth)
1263 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
1264 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
1265 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
1266 if (!VecBitWidthsEqual)
1268 if (!VectorType::isValidElementType(TgtTy))
1270 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems);
1271 if (!BegIsAligned) {
1272 // Shuffle the input so [0,NumElements) contains the output, and
1273 // [NumElems,SrcNumElems) is undef.
1274 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems,
1275 UndefValue::get(Int32Ty));
1276 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
1277 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx);
1278 V = Builder.CreateShuffleVector(V, UndefValue::get(V->getType()),
1279 ConstantVector::get(ShuffleMask),
1280 SVI.getName() + ".extract");
1283 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
1284 assert(SrcElemsPerTgtElem);
1285 BegIdx /= SrcElemsPerTgtElem;
1286 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end();
1290 : Builder.CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
1291 if (!BCAlreadyExists)
1292 NewBCs[CastSrcTy] = NewBC;
1293 auto *Ext = Builder.CreateExtractElement(
1294 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract");
1295 // The shufflevector isn't being replaced: the bitcast that used it
1296 // is. InstCombine will visit the newly-created instructions.
1297 replaceInstUsesWith(*BC, Ext);
1302 // If the LHS is a shufflevector itself, see if we can combine it with this
1303 // one without producing an unusual shuffle.
1304 // Cases that might be simplified:
1306 // x1=shuffle(v1,v2,mask1)
1307 // x=shuffle(x1,undef,mask)
1309 // x=shuffle(v1,undef,newMask)
1310 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
1312 // x1=shuffle(v1,undef,mask1)
1313 // x=shuffle(x1,x2,mask)
1314 // where v1.size() == mask1.size()
1316 // x=shuffle(v1,x2,newMask)
1317 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
1319 // x2=shuffle(v2,undef,mask2)
1320 // x=shuffle(x1,x2,mask)
1321 // where v2.size() == mask2.size()
1323 // x=shuffle(x1,v2,newMask)
1324 // newMask[i] = (mask[i] < x1.size())
1325 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
1327 // x1=shuffle(v1,undef,mask1)
1328 // x2=shuffle(v2,undef,mask2)
1329 // x=shuffle(x1,x2,mask)
1330 // where v1.size() == v2.size()
1332 // x=shuffle(v1,v2,newMask)
1333 // newMask[i] = (mask[i] < x1.size())
1334 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
1336 // Here we are really conservative:
1337 // we are absolutely afraid of producing a shuffle mask not in the input
1338 // program, because the code gen may not be smart enough to turn a merged
1339 // shuffle into two specific shuffles: it may produce worse code. As such,
1340 // we only merge two shuffles if the result is either a splat or one of the
1341 // input shuffle masks. In this case, merging the shuffles just removes
1342 // one instruction, which we know is safe. This is good for things like
1343 // turning: (splat(splat)) -> splat, or
1344 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
1345 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
1346 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
1348 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
1349 LHSShuffle = nullptr;
1351 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
1352 RHSShuffle = nullptr;
1353 if (!LHSShuffle && !RHSShuffle)
1354 return MadeChange ? &SVI : nullptr;
1356 Value* LHSOp0 = nullptr;
1357 Value* LHSOp1 = nullptr;
1358 Value* RHSOp0 = nullptr;
1359 unsigned LHSOp0Width = 0;
1360 unsigned RHSOp0Width = 0;
1362 LHSOp0 = LHSShuffle->getOperand(0);
1363 LHSOp1 = LHSShuffle->getOperand(1);
1364 LHSOp0Width = LHSOp0->getType()->getVectorNumElements();
1367 RHSOp0 = RHSShuffle->getOperand(0);
1368 RHSOp0Width = RHSOp0->getType()->getVectorNumElements();
1370 Value* newLHS = LHS;
1371 Value* newRHS = RHS;
1374 if (isa<UndefValue>(RHS)) {
1379 else if (LHSOp0Width == LHSWidth) {
1384 if (RHSShuffle && RHSOp0Width == LHSWidth) {
1388 if (LHSOp0 == RHSOp0) {
1393 if (newLHS == LHS && newRHS == RHS)
1394 return MadeChange ? &SVI : nullptr;
1396 SmallVector<int, 16> LHSMask;
1397 SmallVector<int, 16> RHSMask;
1399 LHSMask = LHSShuffle->getShuffleMask();
1400 if (RHSShuffle && newRHS != RHS)
1401 RHSMask = RHSShuffle->getShuffleMask();
1403 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
1404 SmallVector<int, 16> newMask;
1405 bool isSplat = true;
1407 // Create a new mask for the new ShuffleVectorInst so that the new
1408 // ShuffleVectorInst is equivalent to the original one.
1409 for (unsigned i = 0; i < VWidth; ++i) {
1412 // This element is an undef value.
1414 } else if (Mask[i] < (int)LHSWidth) {
1415 // This element is from left hand side vector operand.
1417 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
1418 // new mask value for the element.
1419 if (newLHS != LHS) {
1420 eltMask = LHSMask[Mask[i]];
1421 // If the value selected is an undef value, explicitly specify it
1422 // with a -1 mask value.
1423 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
1428 // This element is from right hand side vector operand
1430 // If the value selected is an undef value, explicitly specify it
1431 // with a -1 mask value. (case 1)
1432 if (isa<UndefValue>(RHS))
1434 // If RHS is going to be replaced (case 3 or 4), calculate the
1435 // new mask value for the element.
1436 else if (newRHS != RHS) {
1437 eltMask = RHSMask[Mask[i]-LHSWidth];
1438 // If the value selected is an undef value, explicitly specify it
1439 // with a -1 mask value.
1440 if (eltMask >= (int)RHSOp0Width) {
1441 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1442 && "should have been check above");
1446 eltMask = Mask[i]-LHSWidth;
1448 // If LHS's width is changed, shift the mask value accordingly.
1449 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1450 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1451 // If newRHS == newLHS, we want to remap any references from newRHS to
1452 // newLHS so that we can properly identify splats that may occur due to
1453 // obfuscation across the two vectors.
1454 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
1455 eltMask += newLHSWidth;
1458 // Check if this could still be a splat.
1460 if (SplatElt >= 0 && SplatElt != eltMask)
1465 newMask.push_back(eltMask);
1468 // If the result mask is equal to one of the original shuffle masks,
1469 // or is a splat, do the replacement.
1470 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1471 SmallVector<Constant*, 16> Elts;
1472 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1473 if (newMask[i] < 0) {
1474 Elts.push_back(UndefValue::get(Int32Ty));
1476 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1480 newRHS = UndefValue::get(newLHS->getType());
1481 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1484 // If the result mask is an identity, replace uses of this instruction with
1485 // corresponding argument.
1486 bool isLHSID, isRHSID;
1487 recognizeIdentityMask(newMask, isLHSID, isRHSID);
1488 if (isLHSID && VWidth == LHSOp0Width) return replaceInstUsesWith(SVI, newLHS);
1489 if (isRHSID && VWidth == RHSOp0Width) return replaceInstUsesWith(SVI, newRHS);
1491 return MadeChange ? &SVI : nullptr;