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 IndexVal = IdxC->getZExtValue();
185 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
187 // InstSimplify handles cases where the index is invalid.
188 assert(IndexVal < VectorWidth);
190 // This instruction only demands the single element from the input vector.
191 // If the input vector has a single use, simplify it based on this use
193 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
194 APInt UndefElts(VectorWidth, 0);
195 APInt DemandedMask(VectorWidth, 0);
196 DemandedMask.setBit(IndexVal);
197 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), DemandedMask,
204 // If this extractelement is directly using a bitcast from a vector of
205 // the same number of elements, see if we can find the source element from
206 // it. In this case, we will end up needing to bitcast the scalars.
207 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
208 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
209 if (VT->getNumElements() == VectorWidth)
210 if (Value *Elt = findScalarElement(BCI->getOperand(0), IndexVal))
211 return new BitCastInst(Elt, EI.getType());
214 // If there's a vector PHI feeding a scalar use through this extractelement
215 // instruction, try to scalarize the PHI.
216 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
217 Instruction *scalarPHI = scalarizePHI(EI, PN);
223 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
224 // Push extractelement into predecessor operation if legal and
225 // profitable to do so.
226 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
227 if (I->hasOneUse() &&
228 cheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
230 Builder.CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
231 EI.getName()+".lhs");
233 Builder.CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
234 EI.getName()+".rhs");
235 return BinaryOperator::CreateWithCopiedFlags(BO->getOpcode(),
238 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
239 // Extracting the inserted element?
240 if (IE->getOperand(2) == EI.getOperand(1))
241 return replaceInstUsesWith(EI, IE->getOperand(1));
242 // If the inserted and extracted elements are constants, they must not
243 // be the same value, extract from the pre-inserted value instead.
244 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
245 Worklist.AddValue(EI.getOperand(0));
246 EI.setOperand(0, IE->getOperand(0));
249 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
250 // If this is extracting an element from a shufflevector, figure out where
251 // it came from and extract from the appropriate input element instead.
252 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
253 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
256 SVI->getOperand(0)->getType()->getVectorNumElements();
259 return replaceInstUsesWith(EI, UndefValue::get(EI.getType()));
260 if (SrcIdx < (int)LHSWidth)
261 Src = SVI->getOperand(0);
264 Src = SVI->getOperand(1);
266 Type *Int32Ty = Type::getInt32Ty(EI.getContext());
267 return ExtractElementInst::Create(Src,
268 ConstantInt::get(Int32Ty,
271 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
272 // Canonicalize extractelement(cast) -> cast(extractelement).
273 // Bitcasts can change the number of vector elements, and they cost
275 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
276 Value *EE = Builder.CreateExtractElement(CI->getOperand(0),
277 EI.getIndexOperand());
278 Worklist.AddValue(EE);
279 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
286 /// If V is a shuffle of values that ONLY returns elements from either LHS or
287 /// RHS, return the shuffle mask and true. Otherwise, return false.
288 static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
289 SmallVectorImpl<Constant*> &Mask) {
290 assert(LHS->getType() == RHS->getType() &&
291 "Invalid CollectSingleShuffleElements");
292 unsigned NumElts = V->getType()->getVectorNumElements();
294 if (isa<UndefValue>(V)) {
295 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
300 for (unsigned i = 0; i != NumElts; ++i)
301 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
306 for (unsigned i = 0; i != NumElts; ++i)
307 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
312 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
313 // If this is an insert of an extract from some other vector, include it.
314 Value *VecOp = IEI->getOperand(0);
315 Value *ScalarOp = IEI->getOperand(1);
316 Value *IdxOp = IEI->getOperand(2);
318 if (!isa<ConstantInt>(IdxOp))
320 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
322 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
323 // We can handle this if the vector we are inserting into is
325 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
326 // If so, update the mask to reflect the inserted undef.
327 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
330 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
331 if (isa<ConstantInt>(EI->getOperand(1))) {
332 unsigned ExtractedIdx =
333 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
334 unsigned NumLHSElts = LHS->getType()->getVectorNumElements();
336 // This must be extracting from either LHS or RHS.
337 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
338 // We can handle this if the vector we are inserting into is
340 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
341 // If so, update the mask to reflect the inserted value.
342 if (EI->getOperand(0) == LHS) {
343 Mask[InsertedIdx % NumElts] =
344 ConstantInt::get(Type::getInt32Ty(V->getContext()),
347 assert(EI->getOperand(0) == RHS);
348 Mask[InsertedIdx % NumElts] =
349 ConstantInt::get(Type::getInt32Ty(V->getContext()),
350 ExtractedIdx + NumLHSElts);
362 /// If we have insertion into a vector that is wider than the vector that we
363 /// are extracting from, try to widen the source vector to allow a single
364 /// shufflevector to replace one or more insert/extract pairs.
365 static void replaceExtractElements(InsertElementInst *InsElt,
366 ExtractElementInst *ExtElt,
368 VectorType *InsVecType = InsElt->getType();
369 VectorType *ExtVecType = ExtElt->getVectorOperandType();
370 unsigned NumInsElts = InsVecType->getVectorNumElements();
371 unsigned NumExtElts = ExtVecType->getVectorNumElements();
373 // The inserted-to vector must be wider than the extracted-from vector.
374 if (InsVecType->getElementType() != ExtVecType->getElementType() ||
375 NumExtElts >= NumInsElts)
378 // Create a shuffle mask to widen the extended-from vector using undefined
379 // values. The mask selects all of the values of the original vector followed
380 // by as many undefined values as needed to create a vector of the same length
381 // as the inserted-to vector.
382 SmallVector<Constant *, 16> ExtendMask;
383 IntegerType *IntType = Type::getInt32Ty(InsElt->getContext());
384 for (unsigned i = 0; i < NumExtElts; ++i)
385 ExtendMask.push_back(ConstantInt::get(IntType, i));
386 for (unsigned i = NumExtElts; i < NumInsElts; ++i)
387 ExtendMask.push_back(UndefValue::get(IntType));
389 Value *ExtVecOp = ExtElt->getVectorOperand();
390 auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
391 BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
392 ? ExtVecOpInst->getParent()
393 : ExtElt->getParent();
395 // TODO: This restriction matches the basic block check below when creating
396 // new extractelement instructions. If that limitation is removed, this one
397 // could also be removed. But for now, we just bail out to ensure that we
398 // will replace the extractelement instruction that is feeding our
399 // insertelement instruction. This allows the insertelement to then be
400 // replaced by a shufflevector. If the insertelement is not replaced, we can
401 // induce infinite looping because there's an optimization for extractelement
402 // that will delete our widening shuffle. This would trigger another attempt
403 // here to create that shuffle, and we spin forever.
404 if (InsertionBlock != InsElt->getParent())
407 // TODO: This restriction matches the check in visitInsertElementInst() and
408 // prevents an infinite loop caused by not turning the extract/insert pair
409 // into a shuffle. We really should not need either check, but we're lacking
410 // folds for shufflevectors because we're afraid to generate shuffle masks
411 // that the backend can't handle.
412 if (InsElt->hasOneUse() && isa<InsertElementInst>(InsElt->user_back()))
415 auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
416 ConstantVector::get(ExtendMask));
418 // Insert the new shuffle after the vector operand of the extract is defined
419 // (as long as it's not a PHI) or at the start of the basic block of the
420 // extract, so any subsequent extracts in the same basic block can use it.
421 // TODO: Insert before the earliest ExtractElementInst that is replaced.
422 if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
423 WideVec->insertAfter(ExtVecOpInst);
425 IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
427 // Replace extracts from the original narrow vector with extracts from the new
429 for (User *U : ExtVecOp->users()) {
430 ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
431 if (!OldExt || OldExt->getParent() != WideVec->getParent())
433 auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
434 NewExt->insertAfter(OldExt);
435 IC.replaceInstUsesWith(*OldExt, NewExt);
439 /// We are building a shuffle to create V, which is a sequence of insertelement,
440 /// extractelement pairs. If PermittedRHS is set, then we must either use it or
441 /// not rely on the second vector source. Return a std::pair containing the
442 /// left and right vectors of the proposed shuffle (or 0), and set the Mask
443 /// parameter as required.
445 /// Note: we intentionally don't try to fold earlier shuffles since they have
446 /// often been chosen carefully to be efficiently implementable on the target.
447 using ShuffleOps = std::pair<Value *, Value *>;
449 static ShuffleOps collectShuffleElements(Value *V,
450 SmallVectorImpl<Constant *> &Mask,
453 assert(V->getType()->isVectorTy() && "Invalid shuffle!");
454 unsigned NumElts = V->getType()->getVectorNumElements();
456 if (isa<UndefValue>(V)) {
457 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
458 return std::make_pair(
459 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
462 if (isa<ConstantAggregateZero>(V)) {
463 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
464 return std::make_pair(V, nullptr);
467 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
468 // If this is an insert of an extract from some other vector, include it.
469 Value *VecOp = IEI->getOperand(0);
470 Value *ScalarOp = IEI->getOperand(1);
471 Value *IdxOp = IEI->getOperand(2);
473 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
474 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
475 unsigned ExtractedIdx =
476 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
477 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
479 // Either the extracted from or inserted into vector must be RHSVec,
480 // otherwise we'd end up with a shuffle of three inputs.
481 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
482 Value *RHS = EI->getOperand(0);
483 ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC);
484 assert(LR.second == nullptr || LR.second == RHS);
486 if (LR.first->getType() != RHS->getType()) {
487 // Although we are giving up for now, see if we can create extracts
488 // that match the inserts for another round of combining.
489 replaceExtractElements(IEI, EI, IC);
491 // We tried our best, but we can't find anything compatible with RHS
492 // further up the chain. Return a trivial shuffle.
493 for (unsigned i = 0; i < NumElts; ++i)
494 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
495 return std::make_pair(V, nullptr);
498 unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
499 Mask[InsertedIdx % NumElts] =
500 ConstantInt::get(Type::getInt32Ty(V->getContext()),
501 NumLHSElts+ExtractedIdx);
502 return std::make_pair(LR.first, RHS);
505 if (VecOp == PermittedRHS) {
506 // We've gone as far as we can: anything on the other side of the
507 // extractelement will already have been converted into a shuffle.
508 unsigned NumLHSElts =
509 EI->getOperand(0)->getType()->getVectorNumElements();
510 for (unsigned i = 0; i != NumElts; ++i)
511 Mask.push_back(ConstantInt::get(
512 Type::getInt32Ty(V->getContext()),
513 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
514 return std::make_pair(EI->getOperand(0), PermittedRHS);
517 // If this insertelement is a chain that comes from exactly these two
518 // vectors, return the vector and the effective shuffle.
519 if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
520 collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
522 return std::make_pair(EI->getOperand(0), PermittedRHS);
527 // Otherwise, we can't do anything fancy. Return an identity vector.
528 for (unsigned i = 0; i != NumElts; ++i)
529 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
530 return std::make_pair(V, nullptr);
533 /// Try to find redundant insertvalue instructions, like the following ones:
534 /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
535 /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
536 /// Here the second instruction inserts values at the same indices, as the
537 /// first one, making the first one redundant.
538 /// It should be transformed to:
539 /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
540 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
541 bool IsRedundant = false;
542 ArrayRef<unsigned int> FirstIndices = I.getIndices();
544 // If there is a chain of insertvalue instructions (each of them except the
545 // last one has only one use and it's another insertvalue insn from this
546 // chain), check if any of the 'children' uses the same indices as the first
547 // instruction. In this case, the first one is redundant.
550 while (V->hasOneUse() && Depth < 10) {
551 User *U = V->user_back();
552 auto UserInsInst = dyn_cast<InsertValueInst>(U);
553 if (!UserInsInst || U->getOperand(0) != V)
555 if (UserInsInst->getIndices() == FirstIndices) {
564 return replaceInstUsesWith(I, I.getOperand(0));
568 static bool isShuffleEquivalentToSelect(ShuffleVectorInst &Shuf) {
569 int MaskSize = Shuf.getMask()->getType()->getVectorNumElements();
570 int VecSize = Shuf.getOperand(0)->getType()->getVectorNumElements();
572 // A vector select does not change the size of the operands.
573 if (MaskSize != VecSize)
576 // Each mask element must be undefined or choose a vector element from one of
577 // the source operands without crossing vector lanes.
578 for (int i = 0; i != MaskSize; ++i) {
579 int Elt = Shuf.getMaskValue(i);
580 if (Elt != -1 && Elt != i && Elt != i + VecSize)
587 // Turn a chain of inserts that splats a value into a canonical insert + shuffle
589 // insertelt(insertelt(insertelt(insertelt X, %k, 0), %k, 1), %k, 2) ... ->
590 // shufflevector(insertelt(X, %k, 0), undef, zero)
591 static Instruction *foldInsSequenceIntoBroadcast(InsertElementInst &InsElt) {
592 // We are interested in the last insert in a chain. So, if this insert
593 // has a single user, and that user is an insert, bail.
594 if (InsElt.hasOneUse() && isa<InsertElementInst>(InsElt.user_back()))
597 VectorType *VT = cast<VectorType>(InsElt.getType());
598 int NumElements = VT->getNumElements();
600 // Do not try to do this for a one-element vector, since that's a nop,
601 // and will cause an inf-loop.
602 if (NumElements == 1)
605 Value *SplatVal = InsElt.getOperand(1);
606 InsertElementInst *CurrIE = &InsElt;
607 SmallVector<bool, 16> ElementPresent(NumElements, false);
608 InsertElementInst *FirstIE = nullptr;
610 // Walk the chain backwards, keeping track of which indices we inserted into,
611 // until we hit something that isn't an insert of the splatted value.
613 auto *Idx = dyn_cast<ConstantInt>(CurrIE->getOperand(2));
614 if (!Idx || CurrIE->getOperand(1) != SplatVal)
617 auto *NextIE = dyn_cast<InsertElementInst>(CurrIE->getOperand(0));
618 // Check none of the intermediate steps have any additional uses, except
619 // for the root insertelement instruction, which can be re-used, if it
620 // inserts at position 0.
621 if (CurrIE != &InsElt &&
622 (!CurrIE->hasOneUse() && (NextIE != nullptr || !Idx->isZero())))
625 ElementPresent[Idx->getZExtValue()] = true;
630 // Make sure we've seen an insert into every element.
631 if (llvm::any_of(ElementPresent, [](bool Present) { return !Present; }))
634 // All right, create the insert + shuffle.
635 Instruction *InsertFirst;
636 if (cast<ConstantInt>(FirstIE->getOperand(2))->isZero())
637 InsertFirst = FirstIE;
639 InsertFirst = InsertElementInst::Create(
640 UndefValue::get(VT), SplatVal,
641 ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), 0),
644 Constant *ZeroMask = ConstantAggregateZero::get(
645 VectorType::get(Type::getInt32Ty(InsElt.getContext()), NumElements));
647 return new ShuffleVectorInst(InsertFirst, UndefValue::get(VT), ZeroMask);
650 /// If we have an insertelement instruction feeding into another insertelement
651 /// and the 2nd is inserting a constant into the vector, canonicalize that
652 /// constant insertion before the insertion of a variable:
654 /// insertelement (insertelement X, Y, IdxC1), ScalarC, IdxC2 -->
655 /// insertelement (insertelement X, ScalarC, IdxC2), Y, IdxC1
657 /// This has the potential of eliminating the 2nd insertelement instruction
658 /// via constant folding of the scalar constant into a vector constant.
659 static Instruction *hoistInsEltConst(InsertElementInst &InsElt2,
660 InstCombiner::BuilderTy &Builder) {
661 auto *InsElt1 = dyn_cast<InsertElementInst>(InsElt2.getOperand(0));
662 if (!InsElt1 || !InsElt1->hasOneUse())
667 ConstantInt *IdxC1, *IdxC2;
668 if (match(InsElt1->getOperand(0), m_Value(X)) &&
669 match(InsElt1->getOperand(1), m_Value(Y)) && !isa<Constant>(Y) &&
670 match(InsElt1->getOperand(2), m_ConstantInt(IdxC1)) &&
671 match(InsElt2.getOperand(1), m_Constant(ScalarC)) &&
672 match(InsElt2.getOperand(2), m_ConstantInt(IdxC2)) && IdxC1 != IdxC2) {
673 Value *NewInsElt1 = Builder.CreateInsertElement(X, ScalarC, IdxC2);
674 return InsertElementInst::Create(NewInsElt1, Y, IdxC1);
680 /// insertelt (shufflevector X, CVec, Mask|insertelt X, C1, CIndex1), C, CIndex
681 /// --> shufflevector X, CVec', Mask'
682 static Instruction *foldConstantInsEltIntoShuffle(InsertElementInst &InsElt) {
683 auto *Inst = dyn_cast<Instruction>(InsElt.getOperand(0));
684 // Bail out if the parent has more than one use. In that case, we'd be
685 // replacing the insertelt with a shuffle, and that's not a clear win.
686 if (!Inst || !Inst->hasOneUse())
688 if (auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0))) {
689 // The shuffle must have a constant vector operand. The insertelt must have
690 // a constant scalar being inserted at a constant position in the vector.
691 Constant *ShufConstVec, *InsEltScalar;
692 uint64_t InsEltIndex;
693 if (!match(Shuf->getOperand(1), m_Constant(ShufConstVec)) ||
694 !match(InsElt.getOperand(1), m_Constant(InsEltScalar)) ||
695 !match(InsElt.getOperand(2), m_ConstantInt(InsEltIndex)))
698 // Adding an element to an arbitrary shuffle could be expensive, but a
699 // shuffle that selects elements from vectors without crossing lanes is
701 // If we're just adding a constant into that shuffle, it will still be
703 if (!isShuffleEquivalentToSelect(*Shuf))
706 // From the above 'select' check, we know that the mask has the same number
707 // of elements as the vector input operands. We also know that each constant
708 // input element is used in its lane and can not be used more than once by
709 // the shuffle. Therefore, replace the constant in the shuffle's constant
710 // vector with the insertelt constant. Replace the constant in the shuffle's
711 // mask vector with the insertelt index plus the length of the vector
712 // (because the constant vector operand of a shuffle is always the 2nd
714 Constant *Mask = Shuf->getMask();
715 unsigned NumElts = Mask->getType()->getVectorNumElements();
716 SmallVector<Constant *, 16> NewShufElts(NumElts);
717 SmallVector<Constant *, 16> NewMaskElts(NumElts);
718 for (unsigned I = 0; I != NumElts; ++I) {
719 if (I == InsEltIndex) {
720 NewShufElts[I] = InsEltScalar;
721 Type *Int32Ty = Type::getInt32Ty(Shuf->getContext());
722 NewMaskElts[I] = ConstantInt::get(Int32Ty, InsEltIndex + NumElts);
724 // Copy over the existing values.
725 NewShufElts[I] = ShufConstVec->getAggregateElement(I);
726 NewMaskElts[I] = Mask->getAggregateElement(I);
730 // Create new operands for a shuffle that includes the constant of the
731 // original insertelt. The old shuffle will be dead now.
732 return new ShuffleVectorInst(Shuf->getOperand(0),
733 ConstantVector::get(NewShufElts),
734 ConstantVector::get(NewMaskElts));
735 } else if (auto *IEI = dyn_cast<InsertElementInst>(Inst)) {
736 // Transform sequences of insertelements ops with constant data/indexes into
737 // a single shuffle op.
738 unsigned NumElts = InsElt.getType()->getNumElements();
740 uint64_t InsertIdx[2];
742 if (!match(InsElt.getOperand(2), m_ConstantInt(InsertIdx[0])) ||
743 !match(InsElt.getOperand(1), m_Constant(Val[0])) ||
744 !match(IEI->getOperand(2), m_ConstantInt(InsertIdx[1])) ||
745 !match(IEI->getOperand(1), m_Constant(Val[1])))
747 SmallVector<Constant *, 16> Values(NumElts);
748 SmallVector<Constant *, 16> Mask(NumElts);
749 auto ValI = std::begin(Val);
750 // Generate new constant vector and mask.
751 // We have 2 values/masks from the insertelements instructions. Insert them
752 // into new value/mask vectors.
753 for (uint64_t I : InsertIdx) {
757 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()),
762 // Remaining values are filled with 'undef' values.
763 for (unsigned I = 0; I < NumElts; ++I) {
766 Values[I] = UndefValue::get(InsElt.getType()->getElementType());
767 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), I);
770 // Create new operands for a shuffle that includes the constant of the
771 // original insertelt.
772 return new ShuffleVectorInst(IEI->getOperand(0),
773 ConstantVector::get(Values),
774 ConstantVector::get(Mask));
779 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
780 Value *VecOp = IE.getOperand(0);
781 Value *ScalarOp = IE.getOperand(1);
782 Value *IdxOp = IE.getOperand(2);
784 if (auto *V = SimplifyInsertElementInst(
785 VecOp, ScalarOp, IdxOp, SQ.getWithInstruction(&IE)))
786 return replaceInstUsesWith(IE, V);
788 // Inserting an undef or into an undefined place, remove this.
789 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
790 replaceInstUsesWith(IE, VecOp);
792 // If the inserted element was extracted from some other vector, and if the
793 // indexes are constant, try to turn this into a shufflevector operation.
794 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
795 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
796 unsigned NumInsertVectorElts = IE.getType()->getNumElements();
797 unsigned NumExtractVectorElts =
798 EI->getOperand(0)->getType()->getVectorNumElements();
799 unsigned ExtractedIdx =
800 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
801 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
803 if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
804 return replaceInstUsesWith(IE, VecOp);
806 if (InsertedIdx >= NumInsertVectorElts) // Out of range insert.
807 return replaceInstUsesWith(IE, UndefValue::get(IE.getType()));
809 // If we are extracting a value from a vector, then inserting it right
810 // back into the same place, just use the input vector.
811 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
812 return replaceInstUsesWith(IE, VecOp);
814 // If this insertelement isn't used by some other insertelement, turn it
815 // (and any insertelements it points to), into one big shuffle.
816 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
817 SmallVector<Constant*, 16> Mask;
818 ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this);
820 // The proposed shuffle may be trivial, in which case we shouldn't
821 // perform the combine.
822 if (LR.first != &IE && LR.second != &IE) {
823 // We now have a shuffle of LHS, RHS, Mask.
824 if (LR.second == nullptr)
825 LR.second = UndefValue::get(LR.first->getType());
826 return new ShuffleVectorInst(LR.first, LR.second,
827 ConstantVector::get(Mask));
833 unsigned VWidth = VecOp->getType()->getVectorNumElements();
834 APInt UndefElts(VWidth, 0);
835 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
836 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
838 return replaceInstUsesWith(IE, V);
842 if (Instruction *Shuf = foldConstantInsEltIntoShuffle(IE))
845 if (Instruction *NewInsElt = hoistInsEltConst(IE, Builder))
848 // Turn a sequence of inserts that broadcasts a scalar into a single
849 // insert + shufflevector.
850 if (Instruction *Broadcast = foldInsSequenceIntoBroadcast(IE))
856 /// Return true if we can evaluate the specified expression tree if the vector
857 /// elements were shuffled in a different order.
858 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
859 unsigned Depth = 5) {
860 // We can always reorder the elements of a constant.
861 if (isa<Constant>(V))
864 // We won't reorder vector arguments. No IPO here.
865 Instruction *I = dyn_cast<Instruction>(V);
866 if (!I) return false;
868 // Two users may expect different orders of the elements. Don't try it.
872 if (Depth == 0) return false;
874 switch (I->getOpcode()) {
875 case Instruction::Add:
876 case Instruction::FAdd:
877 case Instruction::Sub:
878 case Instruction::FSub:
879 case Instruction::Mul:
880 case Instruction::FMul:
881 case Instruction::UDiv:
882 case Instruction::SDiv:
883 case Instruction::FDiv:
884 case Instruction::URem:
885 case Instruction::SRem:
886 case Instruction::FRem:
887 case Instruction::Shl:
888 case Instruction::LShr:
889 case Instruction::AShr:
890 case Instruction::And:
891 case Instruction::Or:
892 case Instruction::Xor:
893 case Instruction::ICmp:
894 case Instruction::FCmp:
895 case Instruction::Trunc:
896 case Instruction::ZExt:
897 case Instruction::SExt:
898 case Instruction::FPToUI:
899 case Instruction::FPToSI:
900 case Instruction::UIToFP:
901 case Instruction::SIToFP:
902 case Instruction::FPTrunc:
903 case Instruction::FPExt:
904 case Instruction::GetElementPtr: {
905 for (Value *Operand : I->operands()) {
906 if (!CanEvaluateShuffled(Operand, Mask, Depth-1))
911 case Instruction::InsertElement: {
912 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
913 if (!CI) return false;
914 int ElementNumber = CI->getLimitedValue();
916 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
917 // can't put an element into multiple indices.
918 bool SeenOnce = false;
919 for (int i = 0, e = Mask.size(); i != e; ++i) {
920 if (Mask[i] == ElementNumber) {
926 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
932 /// Rebuild a new instruction just like 'I' but with the new operands given.
933 /// In the event of type mismatch, the type of the operands is correct.
934 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) {
935 // We don't want to use the IRBuilder here because we want the replacement
936 // instructions to appear next to 'I', not the builder's insertion point.
937 switch (I->getOpcode()) {
938 case Instruction::Add:
939 case Instruction::FAdd:
940 case Instruction::Sub:
941 case Instruction::FSub:
942 case Instruction::Mul:
943 case Instruction::FMul:
944 case Instruction::UDiv:
945 case Instruction::SDiv:
946 case Instruction::FDiv:
947 case Instruction::URem:
948 case Instruction::SRem:
949 case Instruction::FRem:
950 case Instruction::Shl:
951 case Instruction::LShr:
952 case Instruction::AShr:
953 case Instruction::And:
954 case Instruction::Or:
955 case Instruction::Xor: {
956 BinaryOperator *BO = cast<BinaryOperator>(I);
957 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
958 BinaryOperator *New =
959 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
960 NewOps[0], NewOps[1], "", BO);
961 if (isa<OverflowingBinaryOperator>(BO)) {
962 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
963 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
965 if (isa<PossiblyExactOperator>(BO)) {
966 New->setIsExact(BO->isExact());
968 if (isa<FPMathOperator>(BO))
969 New->copyFastMathFlags(I);
972 case Instruction::ICmp:
973 assert(NewOps.size() == 2 && "icmp with #ops != 2");
974 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
975 NewOps[0], NewOps[1]);
976 case Instruction::FCmp:
977 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
978 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
979 NewOps[0], NewOps[1]);
980 case Instruction::Trunc:
981 case Instruction::ZExt:
982 case Instruction::SExt:
983 case Instruction::FPToUI:
984 case Instruction::FPToSI:
985 case Instruction::UIToFP:
986 case Instruction::SIToFP:
987 case Instruction::FPTrunc:
988 case Instruction::FPExt: {
989 // It's possible that the mask has a different number of elements from
990 // the original cast. We recompute the destination type to match the mask.
992 VectorType::get(I->getType()->getScalarType(),
993 NewOps[0]->getType()->getVectorNumElements());
994 assert(NewOps.size() == 1 && "cast with #ops != 1");
995 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
998 case Instruction::GetElementPtr: {
999 Value *Ptr = NewOps[0];
1000 ArrayRef<Value*> Idx = NewOps.slice(1);
1001 GetElementPtrInst *GEP = GetElementPtrInst::Create(
1002 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I);
1003 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
1007 llvm_unreachable("failed to rebuild vector instructions");
1011 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
1012 // Mask.size() does not need to be equal to the number of vector elements.
1014 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
1015 Type *EltTy = V->getType()->getScalarType();
1016 if (isa<UndefValue>(V))
1017 return UndefValue::get(VectorType::get(EltTy, Mask.size()));
1019 if (isa<ConstantAggregateZero>(V))
1020 return ConstantAggregateZero::get(VectorType::get(EltTy, Mask.size()));
1022 if (Constant *C = dyn_cast<Constant>(V)) {
1023 SmallVector<Constant *, 16> MaskValues;
1024 for (int i = 0, e = Mask.size(); i != e; ++i) {
1026 MaskValues.push_back(UndefValue::get(Builder.getInt32Ty()));
1028 MaskValues.push_back(Builder.getInt32(Mask[i]));
1030 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
1031 ConstantVector::get(MaskValues));
1034 Instruction *I = cast<Instruction>(V);
1035 switch (I->getOpcode()) {
1036 case Instruction::Add:
1037 case Instruction::FAdd:
1038 case Instruction::Sub:
1039 case Instruction::FSub:
1040 case Instruction::Mul:
1041 case Instruction::FMul:
1042 case Instruction::UDiv:
1043 case Instruction::SDiv:
1044 case Instruction::FDiv:
1045 case Instruction::URem:
1046 case Instruction::SRem:
1047 case Instruction::FRem:
1048 case Instruction::Shl:
1049 case Instruction::LShr:
1050 case Instruction::AShr:
1051 case Instruction::And:
1052 case Instruction::Or:
1053 case Instruction::Xor:
1054 case Instruction::ICmp:
1055 case Instruction::FCmp:
1056 case Instruction::Trunc:
1057 case Instruction::ZExt:
1058 case Instruction::SExt:
1059 case Instruction::FPToUI:
1060 case Instruction::FPToSI:
1061 case Instruction::UIToFP:
1062 case Instruction::SIToFP:
1063 case Instruction::FPTrunc:
1064 case Instruction::FPExt:
1065 case Instruction::Select:
1066 case Instruction::GetElementPtr: {
1067 SmallVector<Value*, 8> NewOps;
1068 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
1069 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
1070 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
1071 NewOps.push_back(V);
1072 NeedsRebuild |= (V != I->getOperand(i));
1075 return buildNew(I, NewOps);
1079 case Instruction::InsertElement: {
1080 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
1082 // The insertelement was inserting at Element. Figure out which element
1083 // that becomes after shuffling. The answer is guaranteed to be unique
1084 // by CanEvaluateShuffled.
1087 for (int e = Mask.size(); Index != e; ++Index) {
1088 if (Mask[Index] == Element) {
1094 // If element is not in Mask, no need to handle the operand 1 (element to
1095 // be inserted). Just evaluate values in operand 0 according to Mask.
1097 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
1099 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
1100 return InsertElementInst::Create(V, I->getOperand(1),
1101 Builder.getInt32(Index), "", I);
1104 llvm_unreachable("failed to reorder elements of vector instruction!");
1107 static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask,
1108 bool &isLHSID, bool &isRHSID) {
1109 isLHSID = isRHSID = true;
1111 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
1112 if (Mask[i] < 0) continue; // Ignore undef values.
1113 // Is this an identity shuffle of the LHS value?
1114 isLHSID &= (Mask[i] == (int)i);
1116 // Is this an identity shuffle of the RHS value?
1117 isRHSID &= (Mask[i]-e == i);
1121 // Returns true if the shuffle is extracting a contiguous range of values from
1122 // LHS, for example:
1123 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1124 // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
1125 // Shuffles to: |EE|FF|GG|HH|
1127 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
1128 SmallVector<int, 16> &Mask) {
1129 unsigned LHSElems = SVI.getOperand(0)->getType()->getVectorNumElements();
1130 unsigned MaskElems = Mask.size();
1131 unsigned BegIdx = Mask.front();
1132 unsigned EndIdx = Mask.back();
1133 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
1135 for (unsigned I = 0; I != MaskElems; ++I)
1136 if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
1141 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
1142 Value *LHS = SVI.getOperand(0);
1143 Value *RHS = SVI.getOperand(1);
1144 SmallVector<int, 16> Mask = SVI.getShuffleMask();
1145 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
1147 if (auto *V = SimplifyShuffleVectorInst(
1148 LHS, RHS, SVI.getMask(), SVI.getType(), SQ.getWithInstruction(&SVI)))
1149 return replaceInstUsesWith(SVI, V);
1151 bool MadeChange = false;
1152 unsigned VWidth = SVI.getType()->getVectorNumElements();
1154 APInt UndefElts(VWidth, 0);
1155 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1156 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
1158 return replaceInstUsesWith(SVI, V);
1162 unsigned LHSWidth = LHS->getType()->getVectorNumElements();
1164 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
1165 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
1166 if (LHS == RHS || isa<UndefValue>(LHS)) {
1167 if (isa<UndefValue>(LHS) && LHS == RHS) {
1168 // shuffle(undef,undef,mask) -> undef.
1169 Value *Result = (VWidth == LHSWidth)
1170 ? LHS : UndefValue::get(SVI.getType());
1171 return replaceInstUsesWith(SVI, Result);
1174 // Remap any references to RHS to use LHS.
1175 SmallVector<Constant*, 16> Elts;
1176 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
1178 Elts.push_back(UndefValue::get(Int32Ty));
1182 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
1183 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
1184 Mask[i] = -1; // Turn into undef.
1185 Elts.push_back(UndefValue::get(Int32Ty));
1187 Mask[i] = Mask[i] % e; // Force to LHS.
1188 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i]));
1191 SVI.setOperand(0, SVI.getOperand(1));
1192 SVI.setOperand(1, UndefValue::get(RHS->getType()));
1193 SVI.setOperand(2, ConstantVector::get(Elts));
1194 LHS = SVI.getOperand(0);
1195 RHS = SVI.getOperand(1);
1199 if (VWidth == LHSWidth) {
1200 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
1201 bool isLHSID, isRHSID;
1202 recognizeIdentityMask(Mask, isLHSID, isRHSID);
1204 // Eliminate identity shuffles.
1205 if (isLHSID) return replaceInstUsesWith(SVI, LHS);
1206 if (isRHSID) return replaceInstUsesWith(SVI, RHS);
1209 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
1210 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
1211 return replaceInstUsesWith(SVI, V);
1214 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
1215 // a non-vector type. We can instead bitcast the original vector followed by
1216 // an extract of the desired element:
1218 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
1219 // <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1220 // %1 = bitcast <4 x i8> %sroa to i32
1222 // %bc = bitcast <16 x i8> %in to <4 x i32>
1223 // %ext = extractelement <4 x i32> %bc, i32 0
1225 // If the shuffle is extracting a contiguous range of values from the input
1226 // vector then each use which is a bitcast of the extracted size can be
1227 // replaced. This will work if the vector types are compatible, and the begin
1228 // index is aligned to a value in the casted vector type. If the begin index
1229 // isn't aligned then we can shuffle the original vector (keeping the same
1230 // vector type) before extracting.
1232 // This code will bail out if the target type is fundamentally incompatible
1233 // with vectors of the source type.
1235 // Example of <16 x i8>, target type i32:
1236 // Index range [4,8): v-----------v Will work.
1237 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1238 // <16 x i8>: | | | | | | | | | | | | | | | | |
1239 // <4 x i32>: | | | | |
1240 // +-----------+-----------+-----------+-----------+
1241 // Index range [6,10): ^-----------^ Needs an extra shuffle.
1242 // Target type i40: ^--------------^ Won't work, bail.
1243 if (isShuffleExtractingFromLHS(SVI, Mask)) {
1245 unsigned MaskElems = Mask.size();
1246 VectorType *SrcTy = cast<VectorType>(V->getType());
1247 unsigned VecBitWidth = SrcTy->getBitWidth();
1248 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
1249 assert(SrcElemBitWidth && "vector elements must have a bitwidth");
1250 unsigned SrcNumElems = SrcTy->getNumElements();
1251 SmallVector<BitCastInst *, 8> BCs;
1252 DenseMap<Type *, Value *> NewBCs;
1253 for (User *U : SVI.users())
1254 if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
1255 if (!BC->use_empty())
1256 // Only visit bitcasts that weren't previously handled.
1258 for (BitCastInst *BC : BCs) {
1259 unsigned BegIdx = Mask.front();
1260 Type *TgtTy = BC->getDestTy();
1261 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
1262 if (!TgtElemBitWidth)
1264 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
1265 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
1266 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
1267 if (!VecBitWidthsEqual)
1269 if (!VectorType::isValidElementType(TgtTy))
1271 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems);
1272 if (!BegIsAligned) {
1273 // Shuffle the input so [0,NumElements) contains the output, and
1274 // [NumElems,SrcNumElems) is undef.
1275 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems,
1276 UndefValue::get(Int32Ty));
1277 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
1278 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx);
1279 V = Builder.CreateShuffleVector(V, UndefValue::get(V->getType()),
1280 ConstantVector::get(ShuffleMask),
1281 SVI.getName() + ".extract");
1284 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
1285 assert(SrcElemsPerTgtElem);
1286 BegIdx /= SrcElemsPerTgtElem;
1287 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end();
1291 : Builder.CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
1292 if (!BCAlreadyExists)
1293 NewBCs[CastSrcTy] = NewBC;
1294 auto *Ext = Builder.CreateExtractElement(
1295 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract");
1296 // The shufflevector isn't being replaced: the bitcast that used it
1297 // is. InstCombine will visit the newly-created instructions.
1298 replaceInstUsesWith(*BC, Ext);
1303 // If the LHS is a shufflevector itself, see if we can combine it with this
1304 // one without producing an unusual shuffle.
1305 // Cases that might be simplified:
1307 // x1=shuffle(v1,v2,mask1)
1308 // x=shuffle(x1,undef,mask)
1310 // x=shuffle(v1,undef,newMask)
1311 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
1313 // x1=shuffle(v1,undef,mask1)
1314 // x=shuffle(x1,x2,mask)
1315 // where v1.size() == mask1.size()
1317 // x=shuffle(v1,x2,newMask)
1318 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
1320 // x2=shuffle(v2,undef,mask2)
1321 // x=shuffle(x1,x2,mask)
1322 // where v2.size() == mask2.size()
1324 // x=shuffle(x1,v2,newMask)
1325 // newMask[i] = (mask[i] < x1.size())
1326 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
1328 // x1=shuffle(v1,undef,mask1)
1329 // x2=shuffle(v2,undef,mask2)
1330 // x=shuffle(x1,x2,mask)
1331 // where v1.size() == v2.size()
1333 // x=shuffle(v1,v2,newMask)
1334 // newMask[i] = (mask[i] < x1.size())
1335 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
1337 // Here we are really conservative:
1338 // we are absolutely afraid of producing a shuffle mask not in the input
1339 // program, because the code gen may not be smart enough to turn a merged
1340 // shuffle into two specific shuffles: it may produce worse code. As such,
1341 // we only merge two shuffles if the result is either a splat or one of the
1342 // input shuffle masks. In this case, merging the shuffles just removes
1343 // one instruction, which we know is safe. This is good for things like
1344 // turning: (splat(splat)) -> splat, or
1345 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
1346 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
1347 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
1349 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
1350 LHSShuffle = nullptr;
1352 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
1353 RHSShuffle = nullptr;
1354 if (!LHSShuffle && !RHSShuffle)
1355 return MadeChange ? &SVI : nullptr;
1357 Value* LHSOp0 = nullptr;
1358 Value* LHSOp1 = nullptr;
1359 Value* RHSOp0 = nullptr;
1360 unsigned LHSOp0Width = 0;
1361 unsigned RHSOp0Width = 0;
1363 LHSOp0 = LHSShuffle->getOperand(0);
1364 LHSOp1 = LHSShuffle->getOperand(1);
1365 LHSOp0Width = LHSOp0->getType()->getVectorNumElements();
1368 RHSOp0 = RHSShuffle->getOperand(0);
1369 RHSOp0Width = RHSOp0->getType()->getVectorNumElements();
1371 Value* newLHS = LHS;
1372 Value* newRHS = RHS;
1375 if (isa<UndefValue>(RHS)) {
1380 else if (LHSOp0Width == LHSWidth) {
1385 if (RHSShuffle && RHSOp0Width == LHSWidth) {
1389 if (LHSOp0 == RHSOp0) {
1394 if (newLHS == LHS && newRHS == RHS)
1395 return MadeChange ? &SVI : nullptr;
1397 SmallVector<int, 16> LHSMask;
1398 SmallVector<int, 16> RHSMask;
1400 LHSMask = LHSShuffle->getShuffleMask();
1401 if (RHSShuffle && newRHS != RHS)
1402 RHSMask = RHSShuffle->getShuffleMask();
1404 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
1405 SmallVector<int, 16> newMask;
1406 bool isSplat = true;
1408 // Create a new mask for the new ShuffleVectorInst so that the new
1409 // ShuffleVectorInst is equivalent to the original one.
1410 for (unsigned i = 0; i < VWidth; ++i) {
1413 // This element is an undef value.
1415 } else if (Mask[i] < (int)LHSWidth) {
1416 // This element is from left hand side vector operand.
1418 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
1419 // new mask value for the element.
1420 if (newLHS != LHS) {
1421 eltMask = LHSMask[Mask[i]];
1422 // If the value selected is an undef value, explicitly specify it
1423 // with a -1 mask value.
1424 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
1429 // This element is from right hand side vector operand
1431 // If the value selected is an undef value, explicitly specify it
1432 // with a -1 mask value. (case 1)
1433 if (isa<UndefValue>(RHS))
1435 // If RHS is going to be replaced (case 3 or 4), calculate the
1436 // new mask value for the element.
1437 else if (newRHS != RHS) {
1438 eltMask = RHSMask[Mask[i]-LHSWidth];
1439 // If the value selected is an undef value, explicitly specify it
1440 // with a -1 mask value.
1441 if (eltMask >= (int)RHSOp0Width) {
1442 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1443 && "should have been check above");
1447 eltMask = Mask[i]-LHSWidth;
1449 // If LHS's width is changed, shift the mask value accordingly.
1450 // If newRHS == nullptr, i.e. LHSOp0 == RHSOp0, we want to remap any
1451 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1452 // If newRHS == newLHS, we want to remap any references from newRHS to
1453 // newLHS so that we can properly identify splats that may occur due to
1454 // obfuscation across the two vectors.
1455 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
1456 eltMask += newLHSWidth;
1459 // Check if this could still be a splat.
1461 if (SplatElt >= 0 && SplatElt != eltMask)
1466 newMask.push_back(eltMask);
1469 // If the result mask is equal to one of the original shuffle masks,
1470 // or is a splat, do the replacement.
1471 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1472 SmallVector<Constant*, 16> Elts;
1473 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1474 if (newMask[i] < 0) {
1475 Elts.push_back(UndefValue::get(Int32Ty));
1477 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1481 newRHS = UndefValue::get(newLHS->getType());
1482 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1485 // If the result mask is an identity, replace uses of this instruction with
1486 // corresponding argument.
1487 bool isLHSID, isRHSID;
1488 recognizeIdentityMask(newMask, isLHSID, isRHSID);
1489 if (isLHSID && VWidth == LHSOp0Width) return replaceInstUsesWith(SVI, newLHS);
1490 if (isRHSID && VWidth == RHSOp0Width) return replaceInstUsesWith(SVI, newRHS);
1492 return MadeChange ? &SVI : nullptr;