xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1  //===- InstCombineVectorOps.cpp -------------------------------------------===//
2  //
3  // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4  // See https://llvm.org/LICENSE.txt for license information.
5  // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6  //
7  //===----------------------------------------------------------------------===//
8  //
9  // This file implements instcombine for ExtractElement, InsertElement and
10  // ShuffleVector.
11  //
12  //===----------------------------------------------------------------------===//
13  
14  #include "InstCombineInternal.h"
15  #include "llvm/ADT/APInt.h"
16  #include "llvm/ADT/ArrayRef.h"
17  #include "llvm/ADT/DenseMap.h"
18  #include "llvm/ADT/STLExtras.h"
19  #include "llvm/ADT/SmallBitVector.h"
20  #include "llvm/ADT/SmallVector.h"
21  #include "llvm/ADT/Statistic.h"
22  #include "llvm/Analysis/InstructionSimplify.h"
23  #include "llvm/Analysis/VectorUtils.h"
24  #include "llvm/IR/BasicBlock.h"
25  #include "llvm/IR/Constant.h"
26  #include "llvm/IR/Constants.h"
27  #include "llvm/IR/DerivedTypes.h"
28  #include "llvm/IR/InstrTypes.h"
29  #include "llvm/IR/Instruction.h"
30  #include "llvm/IR/Instructions.h"
31  #include "llvm/IR/Operator.h"
32  #include "llvm/IR/PatternMatch.h"
33  #include "llvm/IR/Type.h"
34  #include "llvm/IR/User.h"
35  #include "llvm/IR/Value.h"
36  #include "llvm/Support/Casting.h"
37  #include "llvm/Support/ErrorHandling.h"
38  #include "llvm/Transforms/InstCombine/InstCombiner.h"
39  #include <cassert>
40  #include <cstdint>
41  #include <iterator>
42  #include <utility>
43  
44  #define DEBUG_TYPE "instcombine"
45  
46  using namespace llvm;
47  using namespace PatternMatch;
48  
49  STATISTIC(NumAggregateReconstructionsSimplified,
50            "Number of aggregate reconstructions turned into reuse of the "
51            "original aggregate");
52  
53  /// Return true if the value is cheaper to scalarize than it is to leave as a
54  /// vector operation. If the extract index \p EI is a constant integer then
55  /// some operations may be cheap to scalarize.
56  ///
57  /// FIXME: It's possible to create more instructions than previously existed.
cheapToScalarize(Value * V,Value * EI)58  static bool cheapToScalarize(Value *V, Value *EI) {
59    ConstantInt *CEI = dyn_cast<ConstantInt>(EI);
60  
61    // If we can pick a scalar constant value out of a vector, that is free.
62    if (auto *C = dyn_cast<Constant>(V))
63      return CEI || C->getSplatValue();
64  
65    if (CEI && match(V, m_Intrinsic<Intrinsic::experimental_stepvector>())) {
66      ElementCount EC = cast<VectorType>(V->getType())->getElementCount();
67      // Index needs to be lower than the minimum size of the vector, because
68      // for scalable vector, the vector size is known at run time.
69      return CEI->getValue().ult(EC.getKnownMinValue());
70    }
71  
72    // An insertelement to the same constant index as our extract will simplify
73    // to the scalar inserted element. An insertelement to a different constant
74    // index is irrelevant to our extract.
75    if (match(V, m_InsertElt(m_Value(), m_Value(), m_ConstantInt())))
76      return CEI;
77  
78    if (match(V, m_OneUse(m_Load(m_Value()))))
79      return true;
80  
81    if (match(V, m_OneUse(m_UnOp())))
82      return true;
83  
84    Value *V0, *V1;
85    if (match(V, m_OneUse(m_BinOp(m_Value(V0), m_Value(V1)))))
86      if (cheapToScalarize(V0, EI) || cheapToScalarize(V1, EI))
87        return true;
88  
89    CmpInst::Predicate UnusedPred;
90    if (match(V, m_OneUse(m_Cmp(UnusedPred, m_Value(V0), m_Value(V1)))))
91      if (cheapToScalarize(V0, EI) || cheapToScalarize(V1, EI))
92        return true;
93  
94    return false;
95  }
96  
97  // If we have a PHI node with a vector type that is only used to feed
98  // itself and be an operand of extractelement at a constant location,
99  // try to replace the PHI of the vector type with a PHI of a scalar type.
scalarizePHI(ExtractElementInst & EI,PHINode * PN)100  Instruction *InstCombinerImpl::scalarizePHI(ExtractElementInst &EI,
101                                              PHINode *PN) {
102    SmallVector<Instruction *, 2> Extracts;
103    // The users we want the PHI to have are:
104    // 1) The EI ExtractElement (we already know this)
105    // 2) Possibly more ExtractElements with the same index.
106    // 3) Another operand, which will feed back into the PHI.
107    Instruction *PHIUser = nullptr;
108    for (auto *U : PN->users()) {
109      if (ExtractElementInst *EU = dyn_cast<ExtractElementInst>(U)) {
110        if (EI.getIndexOperand() == EU->getIndexOperand())
111          Extracts.push_back(EU);
112        else
113          return nullptr;
114      } else if (!PHIUser) {
115        PHIUser = cast<Instruction>(U);
116      } else {
117        return nullptr;
118      }
119    }
120  
121    if (!PHIUser)
122      return nullptr;
123  
124    // Verify that this PHI user has one use, which is the PHI itself,
125    // and that it is a binary operation which is cheap to scalarize.
126    // otherwise return nullptr.
127    if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) ||
128        !(isa<BinaryOperator>(PHIUser)) ||
129        !cheapToScalarize(PHIUser, EI.getIndexOperand()))
130      return nullptr;
131  
132    // Create a scalar PHI node that will replace the vector PHI node
133    // just before the current PHI node.
134    PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
135        PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), PN->getIterator()));
136    // Scalarize each PHI operand.
137    for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
138      Value *PHIInVal = PN->getIncomingValue(i);
139      BasicBlock *inBB = PN->getIncomingBlock(i);
140      Value *Elt = EI.getIndexOperand();
141      // If the operand is the PHI induction variable:
142      if (PHIInVal == PHIUser) {
143        // Scalarize the binary operation. Its first operand is the
144        // scalar PHI, and the second operand is extracted from the other
145        // vector operand.
146        BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
147        unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
148        Value *Op = InsertNewInstWith(
149            ExtractElementInst::Create(B0->getOperand(opId), Elt,
150                                       B0->getOperand(opId)->getName() + ".Elt"),
151            B0->getIterator());
152        Value *newPHIUser = InsertNewInstWith(
153            BinaryOperator::CreateWithCopiedFlags(B0->getOpcode(),
154                                                  scalarPHI, Op, B0), B0->getIterator());
155        scalarPHI->addIncoming(newPHIUser, inBB);
156      } else {
157        // Scalarize PHI input:
158        Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
159        // Insert the new instruction into the predecessor basic block.
160        Instruction *pos = dyn_cast<Instruction>(PHIInVal);
161        BasicBlock::iterator InsertPos;
162        if (pos && !isa<PHINode>(pos)) {
163          InsertPos = ++pos->getIterator();
164        } else {
165          InsertPos = inBB->getFirstInsertionPt();
166        }
167  
168        InsertNewInstWith(newEI, InsertPos);
169  
170        scalarPHI->addIncoming(newEI, inBB);
171      }
172    }
173  
174    for (auto *E : Extracts) {
175      replaceInstUsesWith(*E, scalarPHI);
176      // Add old extract to worklist for DCE.
177      addToWorklist(E);
178    }
179  
180    return &EI;
181  }
182  
foldBitcastExtElt(ExtractElementInst & Ext)183  Instruction *InstCombinerImpl::foldBitcastExtElt(ExtractElementInst &Ext) {
184    Value *X;
185    uint64_t ExtIndexC;
186    if (!match(Ext.getVectorOperand(), m_BitCast(m_Value(X))) ||
187        !match(Ext.getIndexOperand(), m_ConstantInt(ExtIndexC)))
188      return nullptr;
189  
190    ElementCount NumElts =
191        cast<VectorType>(Ext.getVectorOperandType())->getElementCount();
192    Type *DestTy = Ext.getType();
193    unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
194    bool IsBigEndian = DL.isBigEndian();
195  
196    // If we are casting an integer to vector and extracting a portion, that is
197    // a shift-right and truncate.
198    if (X->getType()->isIntegerTy()) {
199      assert(isa<FixedVectorType>(Ext.getVectorOperand()->getType()) &&
200             "Expected fixed vector type for bitcast from scalar integer");
201  
202      // Big endian requires adjusting the extract index since MSB is at index 0.
203      // LittleEndian: extelt (bitcast i32 X to v4i8), 0 -> trunc i32 X to i8
204      // BigEndian: extelt (bitcast i32 X to v4i8), 0 -> trunc i32 (X >> 24) to i8
205      if (IsBigEndian)
206        ExtIndexC = NumElts.getKnownMinValue() - 1 - ExtIndexC;
207      unsigned ShiftAmountC = ExtIndexC * DestWidth;
208      if (!ShiftAmountC ||
209          (isDesirableIntType(X->getType()->getPrimitiveSizeInBits()) &&
210          Ext.getVectorOperand()->hasOneUse())) {
211        if (ShiftAmountC)
212          X = Builder.CreateLShr(X, ShiftAmountC, "extelt.offset");
213        if (DestTy->isFloatingPointTy()) {
214          Type *DstIntTy = IntegerType::getIntNTy(X->getContext(), DestWidth);
215          Value *Trunc = Builder.CreateTrunc(X, DstIntTy);
216          return new BitCastInst(Trunc, DestTy);
217        }
218        return new TruncInst(X, DestTy);
219      }
220    }
221  
222    if (!X->getType()->isVectorTy())
223      return nullptr;
224  
225    // If this extractelement is using a bitcast from a vector of the same number
226    // of elements, see if we can find the source element from the source vector:
227    // extelt (bitcast VecX), IndexC --> bitcast X[IndexC]
228    auto *SrcTy = cast<VectorType>(X->getType());
229    ElementCount NumSrcElts = SrcTy->getElementCount();
230    if (NumSrcElts == NumElts)
231      if (Value *Elt = findScalarElement(X, ExtIndexC))
232        return new BitCastInst(Elt, DestTy);
233  
234    assert(NumSrcElts.isScalable() == NumElts.isScalable() &&
235           "Src and Dst must be the same sort of vector type");
236  
237    // If the source elements are wider than the destination, try to shift and
238    // truncate a subset of scalar bits of an insert op.
239    if (NumSrcElts.getKnownMinValue() < NumElts.getKnownMinValue()) {
240      Value *Scalar;
241      Value *Vec;
242      uint64_t InsIndexC;
243      if (!match(X, m_InsertElt(m_Value(Vec), m_Value(Scalar),
244                                m_ConstantInt(InsIndexC))))
245        return nullptr;
246  
247      // The extract must be from the subset of vector elements that we inserted
248      // into. Example: if we inserted element 1 of a <2 x i64> and we are
249      // extracting an i16 (narrowing ratio = 4), then this extract must be from 1
250      // of elements 4-7 of the bitcasted vector.
251      unsigned NarrowingRatio =
252          NumElts.getKnownMinValue() / NumSrcElts.getKnownMinValue();
253  
254      if (ExtIndexC / NarrowingRatio != InsIndexC) {
255        // Remove insertelement, if we don't use the inserted element.
256        // extractelement (bitcast (insertelement (Vec, b)), a) ->
257        // extractelement (bitcast (Vec), a)
258        // FIXME: this should be removed to SimplifyDemandedVectorElts,
259        // once scale vectors are supported.
260        if (X->hasOneUse() && Ext.getVectorOperand()->hasOneUse()) {
261          Value *NewBC = Builder.CreateBitCast(Vec, Ext.getVectorOperandType());
262          return ExtractElementInst::Create(NewBC, Ext.getIndexOperand());
263        }
264        return nullptr;
265      }
266  
267      // We are extracting part of the original scalar. How that scalar is
268      // inserted into the vector depends on the endian-ness. Example:
269      //              Vector Byte Elt Index:    0  1  2  3  4  5  6  7
270      //                                       +--+--+--+--+--+--+--+--+
271      // inselt <2 x i32> V, <i32> S, 1:       |V0|V1|V2|V3|S0|S1|S2|S3|
272      // extelt <4 x i16> V', 3:               |                 |S2|S3|
273      //                                       +--+--+--+--+--+--+--+--+
274      // If this is little-endian, S2|S3 are the MSB of the 32-bit 'S' value.
275      // If this is big-endian, S2|S3 are the LSB of the 32-bit 'S' value.
276      // In this example, we must right-shift little-endian. Big-endian is just a
277      // truncate.
278      unsigned Chunk = ExtIndexC % NarrowingRatio;
279      if (IsBigEndian)
280        Chunk = NarrowingRatio - 1 - Chunk;
281  
282      // Bail out if this is an FP vector to FP vector sequence. That would take
283      // more instructions than we started with unless there is no shift, and it
284      // may not be handled as well in the backend.
285      bool NeedSrcBitcast = SrcTy->getScalarType()->isFloatingPointTy();
286      bool NeedDestBitcast = DestTy->isFloatingPointTy();
287      if (NeedSrcBitcast && NeedDestBitcast)
288        return nullptr;
289  
290      unsigned SrcWidth = SrcTy->getScalarSizeInBits();
291      unsigned ShAmt = Chunk * DestWidth;
292  
293      // TODO: This limitation is more strict than necessary. We could sum the
294      // number of new instructions and subtract the number eliminated to know if
295      // we can proceed.
296      if (!X->hasOneUse() || !Ext.getVectorOperand()->hasOneUse())
297        if (NeedSrcBitcast || NeedDestBitcast)
298          return nullptr;
299  
300      if (NeedSrcBitcast) {
301        Type *SrcIntTy = IntegerType::getIntNTy(Scalar->getContext(), SrcWidth);
302        Scalar = Builder.CreateBitCast(Scalar, SrcIntTy);
303      }
304  
305      if (ShAmt) {
306        // Bail out if we could end with more instructions than we started with.
307        if (!Ext.getVectorOperand()->hasOneUse())
308          return nullptr;
309        Scalar = Builder.CreateLShr(Scalar, ShAmt);
310      }
311  
312      if (NeedDestBitcast) {
313        Type *DestIntTy = IntegerType::getIntNTy(Scalar->getContext(), DestWidth);
314        return new BitCastInst(Builder.CreateTrunc(Scalar, DestIntTy), DestTy);
315      }
316      return new TruncInst(Scalar, DestTy);
317    }
318  
319    return nullptr;
320  }
321  
322  /// Find elements of V demanded by UserInstr.
findDemandedEltsBySingleUser(Value * V,Instruction * UserInstr)323  static APInt findDemandedEltsBySingleUser(Value *V, Instruction *UserInstr) {
324    unsigned VWidth = cast<FixedVectorType>(V->getType())->getNumElements();
325  
326    // Conservatively assume that all elements are needed.
327    APInt UsedElts(APInt::getAllOnes(VWidth));
328  
329    switch (UserInstr->getOpcode()) {
330    case Instruction::ExtractElement: {
331      ExtractElementInst *EEI = cast<ExtractElementInst>(UserInstr);
332      assert(EEI->getVectorOperand() == V);
333      ConstantInt *EEIIndexC = dyn_cast<ConstantInt>(EEI->getIndexOperand());
334      if (EEIIndexC && EEIIndexC->getValue().ult(VWidth)) {
335        UsedElts = APInt::getOneBitSet(VWidth, EEIIndexC->getZExtValue());
336      }
337      break;
338    }
339    case Instruction::ShuffleVector: {
340      ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(UserInstr);
341      unsigned MaskNumElts =
342          cast<FixedVectorType>(UserInstr->getType())->getNumElements();
343  
344      UsedElts = APInt(VWidth, 0);
345      for (unsigned i = 0; i < MaskNumElts; i++) {
346        unsigned MaskVal = Shuffle->getMaskValue(i);
347        if (MaskVal == -1u || MaskVal >= 2 * VWidth)
348          continue;
349        if (Shuffle->getOperand(0) == V && (MaskVal < VWidth))
350          UsedElts.setBit(MaskVal);
351        if (Shuffle->getOperand(1) == V &&
352            ((MaskVal >= VWidth) && (MaskVal < 2 * VWidth)))
353          UsedElts.setBit(MaskVal - VWidth);
354      }
355      break;
356    }
357    default:
358      break;
359    }
360    return UsedElts;
361  }
362  
363  /// Find union of elements of V demanded by all its users.
364  /// If it is known by querying findDemandedEltsBySingleUser that
365  /// no user demands an element of V, then the corresponding bit
366  /// remains unset in the returned value.
findDemandedEltsByAllUsers(Value * V)367  static APInt findDemandedEltsByAllUsers(Value *V) {
368    unsigned VWidth = cast<FixedVectorType>(V->getType())->getNumElements();
369  
370    APInt UnionUsedElts(VWidth, 0);
371    for (const Use &U : V->uses()) {
372      if (Instruction *I = dyn_cast<Instruction>(U.getUser())) {
373        UnionUsedElts |= findDemandedEltsBySingleUser(V, I);
374      } else {
375        UnionUsedElts = APInt::getAllOnes(VWidth);
376        break;
377      }
378  
379      if (UnionUsedElts.isAllOnes())
380        break;
381    }
382  
383    return UnionUsedElts;
384  }
385  
386  /// Given a constant index for a extractelement or insertelement instruction,
387  /// return it with the canonical type if it isn't already canonical.  We
388  /// arbitrarily pick 64 bit as our canonical type.  The actual bitwidth doesn't
389  /// matter, we just want a consistent type to simplify CSE.
getPreferredVectorIndex(ConstantInt * IndexC)390  static ConstantInt *getPreferredVectorIndex(ConstantInt *IndexC) {
391    const unsigned IndexBW = IndexC->getBitWidth();
392    if (IndexBW == 64 || IndexC->getValue().getActiveBits() > 64)
393      return nullptr;
394    return ConstantInt::get(IndexC->getContext(),
395                            IndexC->getValue().zextOrTrunc(64));
396  }
397  
visitExtractElementInst(ExtractElementInst & EI)398  Instruction *InstCombinerImpl::visitExtractElementInst(ExtractElementInst &EI) {
399    Value *SrcVec = EI.getVectorOperand();
400    Value *Index = EI.getIndexOperand();
401    if (Value *V = simplifyExtractElementInst(SrcVec, Index,
402                                              SQ.getWithInstruction(&EI)))
403      return replaceInstUsesWith(EI, V);
404  
405    // extractelt (select %x, %vec1, %vec2), %const ->
406    // select %x, %vec1[%const], %vec2[%const]
407    // TODO: Support constant folding of multiple select operands:
408    // extractelt (select %x, %vec1, %vec2), (select %x, %c1, %c2)
409    // If the extractelement will for instance try to do out of bounds accesses
410    // because of the values of %c1 and/or %c2, the sequence could be optimized
411    // early. This is currently not possible because constant folding will reach
412    // an unreachable assertion if it doesn't find a constant operand.
413    if (SelectInst *SI = dyn_cast<SelectInst>(EI.getVectorOperand()))
414      if (SI->getCondition()->getType()->isIntegerTy() &&
415          isa<Constant>(EI.getIndexOperand()))
416        if (Instruction *R = FoldOpIntoSelect(EI, SI))
417          return R;
418  
419    // If extracting a specified index from the vector, see if we can recursively
420    // find a previously computed scalar that was inserted into the vector.
421    auto *IndexC = dyn_cast<ConstantInt>(Index);
422    bool HasKnownValidIndex = false;
423    if (IndexC) {
424      // Canonicalize type of constant indices to i64 to simplify CSE
425      if (auto *NewIdx = getPreferredVectorIndex(IndexC))
426        return replaceOperand(EI, 1, NewIdx);
427  
428      ElementCount EC = EI.getVectorOperandType()->getElementCount();
429      unsigned NumElts = EC.getKnownMinValue();
430      HasKnownValidIndex = IndexC->getValue().ult(NumElts);
431  
432      if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(SrcVec)) {
433        Intrinsic::ID IID = II->getIntrinsicID();
434        // Index needs to be lower than the minimum size of the vector, because
435        // for scalable vector, the vector size is known at run time.
436        if (IID == Intrinsic::experimental_stepvector &&
437            IndexC->getValue().ult(NumElts)) {
438          Type *Ty = EI.getType();
439          unsigned BitWidth = Ty->getIntegerBitWidth();
440          Value *Idx;
441          // Return index when its value does not exceed the allowed limit
442          // for the element type of the vector, otherwise return undefined.
443          if (IndexC->getValue().getActiveBits() <= BitWidth)
444            Idx = ConstantInt::get(Ty, IndexC->getValue().zextOrTrunc(BitWidth));
445          else
446            Idx = PoisonValue::get(Ty);
447          return replaceInstUsesWith(EI, Idx);
448        }
449      }
450  
451      // InstSimplify should handle cases where the index is invalid.
452      // For fixed-length vector, it's invalid to extract out-of-range element.
453      if (!EC.isScalable() && IndexC->getValue().uge(NumElts))
454        return nullptr;
455  
456      if (Instruction *I = foldBitcastExtElt(EI))
457        return I;
458  
459      // If there's a vector PHI feeding a scalar use through this extractelement
460      // instruction, try to scalarize the PHI.
461      if (auto *Phi = dyn_cast<PHINode>(SrcVec))
462        if (Instruction *ScalarPHI = scalarizePHI(EI, Phi))
463          return ScalarPHI;
464    }
465  
466    // TODO come up with a n-ary matcher that subsumes both unary and
467    // binary matchers.
468    UnaryOperator *UO;
469    if (match(SrcVec, m_UnOp(UO)) && cheapToScalarize(SrcVec, Index)) {
470      // extelt (unop X), Index --> unop (extelt X, Index)
471      Value *X = UO->getOperand(0);
472      Value *E = Builder.CreateExtractElement(X, Index);
473      return UnaryOperator::CreateWithCopiedFlags(UO->getOpcode(), E, UO);
474    }
475  
476    // If the binop is not speculatable, we cannot hoist the extractelement if
477    // it may make the operand poison.
478    BinaryOperator *BO;
479    if (match(SrcVec, m_BinOp(BO)) && cheapToScalarize(SrcVec, Index) &&
480        (HasKnownValidIndex || isSafeToSpeculativelyExecute(BO))) {
481      // extelt (binop X, Y), Index --> binop (extelt X, Index), (extelt Y, Index)
482      Value *X = BO->getOperand(0), *Y = BO->getOperand(1);
483      Value *E0 = Builder.CreateExtractElement(X, Index);
484      Value *E1 = Builder.CreateExtractElement(Y, Index);
485      return BinaryOperator::CreateWithCopiedFlags(BO->getOpcode(), E0, E1, BO);
486    }
487  
488    Value *X, *Y;
489    CmpInst::Predicate Pred;
490    if (match(SrcVec, m_Cmp(Pred, m_Value(X), m_Value(Y))) &&
491        cheapToScalarize(SrcVec, Index)) {
492      // extelt (cmp X, Y), Index --> cmp (extelt X, Index), (extelt Y, Index)
493      Value *E0 = Builder.CreateExtractElement(X, Index);
494      Value *E1 = Builder.CreateExtractElement(Y, Index);
495      CmpInst *SrcCmpInst = cast<CmpInst>(SrcVec);
496      return CmpInst::CreateWithCopiedFlags(SrcCmpInst->getOpcode(), Pred, E0, E1,
497                                            SrcCmpInst);
498    }
499  
500    if (auto *I = dyn_cast<Instruction>(SrcVec)) {
501      if (auto *IE = dyn_cast<InsertElementInst>(I)) {
502        // instsimplify already handled the case where the indices are constants
503        // and equal by value, if both are constants, they must not be the same
504        // value, extract from the pre-inserted value instead.
505        if (isa<Constant>(IE->getOperand(2)) && IndexC)
506          return replaceOperand(EI, 0, IE->getOperand(0));
507      } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
508        auto *VecType = cast<VectorType>(GEP->getType());
509        ElementCount EC = VecType->getElementCount();
510        uint64_t IdxVal = IndexC ? IndexC->getZExtValue() : 0;
511        if (IndexC && IdxVal < EC.getKnownMinValue() && GEP->hasOneUse()) {
512          // Find out why we have a vector result - these are a few examples:
513          //  1. We have a scalar pointer and a vector of indices, or
514          //  2. We have a vector of pointers and a scalar index, or
515          //  3. We have a vector of pointers and a vector of indices, etc.
516          // Here we only consider combining when there is exactly one vector
517          // operand, since the optimization is less obviously a win due to
518          // needing more than one extractelements.
519  
520          unsigned VectorOps =
521              llvm::count_if(GEP->operands(), [](const Value *V) {
522                return isa<VectorType>(V->getType());
523              });
524          if (VectorOps == 1) {
525            Value *NewPtr = GEP->getPointerOperand();
526            if (isa<VectorType>(NewPtr->getType()))
527              NewPtr = Builder.CreateExtractElement(NewPtr, IndexC);
528  
529            SmallVector<Value *> NewOps;
530            for (unsigned I = 1; I != GEP->getNumOperands(); ++I) {
531              Value *Op = GEP->getOperand(I);
532              if (isa<VectorType>(Op->getType()))
533                NewOps.push_back(Builder.CreateExtractElement(Op, IndexC));
534              else
535                NewOps.push_back(Op);
536            }
537  
538            GetElementPtrInst *NewGEP = GetElementPtrInst::Create(
539                GEP->getSourceElementType(), NewPtr, NewOps);
540            NewGEP->setIsInBounds(GEP->isInBounds());
541            return NewGEP;
542          }
543        }
544      } else if (auto *SVI = dyn_cast<ShuffleVectorInst>(I)) {
545        // If this is extracting an element from a shufflevector, figure out where
546        // it came from and extract from the appropriate input element instead.
547        // Restrict the following transformation to fixed-length vector.
548        if (isa<FixedVectorType>(SVI->getType()) && isa<ConstantInt>(Index)) {
549          int SrcIdx =
550              SVI->getMaskValue(cast<ConstantInt>(Index)->getZExtValue());
551          Value *Src;
552          unsigned LHSWidth = cast<FixedVectorType>(SVI->getOperand(0)->getType())
553                                  ->getNumElements();
554  
555          if (SrcIdx < 0)
556            return replaceInstUsesWith(EI, PoisonValue::get(EI.getType()));
557          if (SrcIdx < (int)LHSWidth)
558            Src = SVI->getOperand(0);
559          else {
560            SrcIdx -= LHSWidth;
561            Src = SVI->getOperand(1);
562          }
563          Type *Int64Ty = Type::getInt64Ty(EI.getContext());
564          return ExtractElementInst::Create(
565              Src, ConstantInt::get(Int64Ty, SrcIdx, false));
566        }
567      } else if (auto *CI = dyn_cast<CastInst>(I)) {
568        // Canonicalize extractelement(cast) -> cast(extractelement).
569        // Bitcasts can change the number of vector elements, and they cost
570        // nothing.
571        if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
572          Value *EE = Builder.CreateExtractElement(CI->getOperand(0), Index);
573          return CastInst::Create(CI->getOpcode(), EE, EI.getType());
574        }
575      }
576    }
577  
578    // Run demanded elements after other transforms as this can drop flags on
579    // binops.  If there's two paths to the same final result, we prefer the
580    // one which doesn't force us to drop flags.
581    if (IndexC) {
582      ElementCount EC = EI.getVectorOperandType()->getElementCount();
583      unsigned NumElts = EC.getKnownMinValue();
584      // This instruction only demands the single element from the input vector.
585      // Skip for scalable type, the number of elements is unknown at
586      // compile-time.
587      if (!EC.isScalable() && NumElts != 1) {
588        // If the input vector has a single use, simplify it based on this use
589        // property.
590        if (SrcVec->hasOneUse()) {
591          APInt PoisonElts(NumElts, 0);
592          APInt DemandedElts(NumElts, 0);
593          DemandedElts.setBit(IndexC->getZExtValue());
594          if (Value *V =
595                  SimplifyDemandedVectorElts(SrcVec, DemandedElts, PoisonElts))
596            return replaceOperand(EI, 0, V);
597        } else {
598          // If the input vector has multiple uses, simplify it based on a union
599          // of all elements used.
600          APInt DemandedElts = findDemandedEltsByAllUsers(SrcVec);
601          if (!DemandedElts.isAllOnes()) {
602            APInt PoisonElts(NumElts, 0);
603            if (Value *V = SimplifyDemandedVectorElts(
604                    SrcVec, DemandedElts, PoisonElts, 0 /* Depth */,
605                    true /* AllowMultipleUsers */)) {
606              if (V != SrcVec) {
607                Worklist.addValue(SrcVec);
608                SrcVec->replaceAllUsesWith(V);
609                return &EI;
610              }
611            }
612          }
613        }
614      }
615    }
616    return nullptr;
617  }
618  
619  /// If V is a shuffle of values that ONLY returns elements from either LHS or
620  /// RHS, return the shuffle mask and true. Otherwise, return false.
collectSingleShuffleElements(Value * V,Value * LHS,Value * RHS,SmallVectorImpl<int> & Mask)621  static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
622                                           SmallVectorImpl<int> &Mask) {
623    assert(LHS->getType() == RHS->getType() &&
624           "Invalid CollectSingleShuffleElements");
625    unsigned NumElts = cast<FixedVectorType>(V->getType())->getNumElements();
626  
627    if (match(V, m_Poison())) {
628      Mask.assign(NumElts, -1);
629      return true;
630    }
631  
632    if (V == LHS) {
633      for (unsigned i = 0; i != NumElts; ++i)
634        Mask.push_back(i);
635      return true;
636    }
637  
638    if (V == RHS) {
639      for (unsigned i = 0; i != NumElts; ++i)
640        Mask.push_back(i + NumElts);
641      return true;
642    }
643  
644    if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
645      // If this is an insert of an extract from some other vector, include it.
646      Value *VecOp    = IEI->getOperand(0);
647      Value *ScalarOp = IEI->getOperand(1);
648      Value *IdxOp    = IEI->getOperand(2);
649  
650      if (!isa<ConstantInt>(IdxOp))
651        return false;
652      unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
653  
654      if (isa<PoisonValue>(ScalarOp)) {  // inserting poison into vector.
655        // We can handle this if the vector we are inserting into is
656        // transitively ok.
657        if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
658          // If so, update the mask to reflect the inserted poison.
659          Mask[InsertedIdx] = -1;
660          return true;
661        }
662      } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
663        if (isa<ConstantInt>(EI->getOperand(1))) {
664          unsigned ExtractedIdx =
665          cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
666          unsigned NumLHSElts =
667              cast<FixedVectorType>(LHS->getType())->getNumElements();
668  
669          // This must be extracting from either LHS or RHS.
670          if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
671            // We can handle this if the vector we are inserting into is
672            // transitively ok.
673            if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
674              // If so, update the mask to reflect the inserted value.
675              if (EI->getOperand(0) == LHS) {
676                Mask[InsertedIdx % NumElts] = ExtractedIdx;
677              } else {
678                assert(EI->getOperand(0) == RHS);
679                Mask[InsertedIdx % NumElts] = ExtractedIdx + NumLHSElts;
680              }
681              return true;
682            }
683          }
684        }
685      }
686    }
687  
688    return false;
689  }
690  
691  /// If we have insertion into a vector that is wider than the vector that we
692  /// are extracting from, try to widen the source vector to allow a single
693  /// shufflevector to replace one or more insert/extract pairs.
replaceExtractElements(InsertElementInst * InsElt,ExtractElementInst * ExtElt,InstCombinerImpl & IC)694  static bool replaceExtractElements(InsertElementInst *InsElt,
695                                     ExtractElementInst *ExtElt,
696                                     InstCombinerImpl &IC) {
697    auto *InsVecType = cast<FixedVectorType>(InsElt->getType());
698    auto *ExtVecType = cast<FixedVectorType>(ExtElt->getVectorOperandType());
699    unsigned NumInsElts = InsVecType->getNumElements();
700    unsigned NumExtElts = ExtVecType->getNumElements();
701  
702    // The inserted-to vector must be wider than the extracted-from vector.
703    if (InsVecType->getElementType() != ExtVecType->getElementType() ||
704        NumExtElts >= NumInsElts)
705      return false;
706  
707    // Create a shuffle mask to widen the extended-from vector using poison
708    // values. The mask selects all of the values of the original vector followed
709    // by as many poison values as needed to create a vector of the same length
710    // as the inserted-to vector.
711    SmallVector<int, 16> ExtendMask;
712    for (unsigned i = 0; i < NumExtElts; ++i)
713      ExtendMask.push_back(i);
714    for (unsigned i = NumExtElts; i < NumInsElts; ++i)
715      ExtendMask.push_back(-1);
716  
717    Value *ExtVecOp = ExtElt->getVectorOperand();
718    auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
719    BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
720                                     ? ExtVecOpInst->getParent()
721                                     : ExtElt->getParent();
722  
723    // TODO: This restriction matches the basic block check below when creating
724    // new extractelement instructions. If that limitation is removed, this one
725    // could also be removed. But for now, we just bail out to ensure that we
726    // will replace the extractelement instruction that is feeding our
727    // insertelement instruction. This allows the insertelement to then be
728    // replaced by a shufflevector. If the insertelement is not replaced, we can
729    // induce infinite looping because there's an optimization for extractelement
730    // that will delete our widening shuffle. This would trigger another attempt
731    // here to create that shuffle, and we spin forever.
732    if (InsertionBlock != InsElt->getParent())
733      return false;
734  
735    // TODO: This restriction matches the check in visitInsertElementInst() and
736    // prevents an infinite loop caused by not turning the extract/insert pair
737    // into a shuffle. We really should not need either check, but we're lacking
738    // folds for shufflevectors because we're afraid to generate shuffle masks
739    // that the backend can't handle.
740    if (InsElt->hasOneUse() && isa<InsertElementInst>(InsElt->user_back()))
741      return false;
742  
743    auto *WideVec = new ShuffleVectorInst(ExtVecOp, ExtendMask);
744  
745    // Insert the new shuffle after the vector operand of the extract is defined
746    // (as long as it's not a PHI) or at the start of the basic block of the
747    // extract, so any subsequent extracts in the same basic block can use it.
748    // TODO: Insert before the earliest ExtractElementInst that is replaced.
749    if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
750      WideVec->insertAfter(ExtVecOpInst);
751    else
752      IC.InsertNewInstWith(WideVec, ExtElt->getParent()->getFirstInsertionPt());
753  
754    // Replace extracts from the original narrow vector with extracts from the new
755    // wide vector.
756    for (User *U : ExtVecOp->users()) {
757      ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
758      if (!OldExt || OldExt->getParent() != WideVec->getParent())
759        continue;
760      auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
761      IC.InsertNewInstWith(NewExt, OldExt->getIterator());
762      IC.replaceInstUsesWith(*OldExt, NewExt);
763      // Add the old extracts to the worklist for DCE. We can't remove the
764      // extracts directly, because they may still be used by the calling code.
765      IC.addToWorklist(OldExt);
766    }
767  
768    return true;
769  }
770  
771  /// We are building a shuffle to create V, which is a sequence of insertelement,
772  /// extractelement pairs. If PermittedRHS is set, then we must either use it or
773  /// not rely on the second vector source. Return a std::pair containing the
774  /// left and right vectors of the proposed shuffle (or 0), and set the Mask
775  /// parameter as required.
776  ///
777  /// Note: we intentionally don't try to fold earlier shuffles since they have
778  /// often been chosen carefully to be efficiently implementable on the target.
779  using ShuffleOps = std::pair<Value *, Value *>;
780  
collectShuffleElements(Value * V,SmallVectorImpl<int> & Mask,Value * PermittedRHS,InstCombinerImpl & IC,bool & Rerun)781  static ShuffleOps collectShuffleElements(Value *V, SmallVectorImpl<int> &Mask,
782                                           Value *PermittedRHS,
783                                           InstCombinerImpl &IC, bool &Rerun) {
784    assert(V->getType()->isVectorTy() && "Invalid shuffle!");
785    unsigned NumElts = cast<FixedVectorType>(V->getType())->getNumElements();
786  
787    if (match(V, m_Poison())) {
788      Mask.assign(NumElts, -1);
789      return std::make_pair(
790          PermittedRHS ? PoisonValue::get(PermittedRHS->getType()) : V, nullptr);
791    }
792  
793    if (isa<ConstantAggregateZero>(V)) {
794      Mask.assign(NumElts, 0);
795      return std::make_pair(V, nullptr);
796    }
797  
798    if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
799      // If this is an insert of an extract from some other vector, include it.
800      Value *VecOp    = IEI->getOperand(0);
801      Value *ScalarOp = IEI->getOperand(1);
802      Value *IdxOp    = IEI->getOperand(2);
803  
804      if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
805        if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
806          unsigned ExtractedIdx =
807            cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
808          unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
809  
810          // Either the extracted from or inserted into vector must be RHSVec,
811          // otherwise we'd end up with a shuffle of three inputs.
812          if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
813            Value *RHS = EI->getOperand(0);
814            ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC, Rerun);
815            assert(LR.second == nullptr || LR.second == RHS);
816  
817            if (LR.first->getType() != RHS->getType()) {
818              // Although we are giving up for now, see if we can create extracts
819              // that match the inserts for another round of combining.
820              if (replaceExtractElements(IEI, EI, IC))
821                Rerun = true;
822  
823              // We tried our best, but we can't find anything compatible with RHS
824              // further up the chain. Return a trivial shuffle.
825              for (unsigned i = 0; i < NumElts; ++i)
826                Mask[i] = i;
827              return std::make_pair(V, nullptr);
828            }
829  
830            unsigned NumLHSElts =
831                cast<FixedVectorType>(RHS->getType())->getNumElements();
832            Mask[InsertedIdx % NumElts] = NumLHSElts + ExtractedIdx;
833            return std::make_pair(LR.first, RHS);
834          }
835  
836          if (VecOp == PermittedRHS) {
837            // We've gone as far as we can: anything on the other side of the
838            // extractelement will already have been converted into a shuffle.
839            unsigned NumLHSElts =
840                cast<FixedVectorType>(EI->getOperand(0)->getType())
841                    ->getNumElements();
842            for (unsigned i = 0; i != NumElts; ++i)
843              Mask.push_back(i == InsertedIdx ? ExtractedIdx : NumLHSElts + i);
844            return std::make_pair(EI->getOperand(0), PermittedRHS);
845          }
846  
847          // If this insertelement is a chain that comes from exactly these two
848          // vectors, return the vector and the effective shuffle.
849          if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
850              collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
851                                           Mask))
852            return std::make_pair(EI->getOperand(0), PermittedRHS);
853        }
854      }
855    }
856  
857    // Otherwise, we can't do anything fancy. Return an identity vector.
858    for (unsigned i = 0; i != NumElts; ++i)
859      Mask.push_back(i);
860    return std::make_pair(V, nullptr);
861  }
862  
863  /// Look for chain of insertvalue's that fully define an aggregate, and trace
864  /// back the values inserted, see if they are all were extractvalue'd from
865  /// the same source aggregate from the exact same element indexes.
866  /// If they were, just reuse the source aggregate.
867  /// This potentially deals with PHI indirections.
foldAggregateConstructionIntoAggregateReuse(InsertValueInst & OrigIVI)868  Instruction *InstCombinerImpl::foldAggregateConstructionIntoAggregateReuse(
869      InsertValueInst &OrigIVI) {
870    Type *AggTy = OrigIVI.getType();
871    unsigned NumAggElts;
872    switch (AggTy->getTypeID()) {
873    case Type::StructTyID:
874      NumAggElts = AggTy->getStructNumElements();
875      break;
876    case Type::ArrayTyID:
877      NumAggElts = AggTy->getArrayNumElements();
878      break;
879    default:
880      llvm_unreachable("Unhandled aggregate type?");
881    }
882  
883    // Arbitrary aggregate size cut-off. Motivation for limit of 2 is to be able
884    // to handle clang C++ exception struct (which is hardcoded as {i8*, i32}),
885    // FIXME: any interesting patterns to be caught with larger limit?
886    assert(NumAggElts > 0 && "Aggregate should have elements.");
887    if (NumAggElts > 2)
888      return nullptr;
889  
890    static constexpr auto NotFound = std::nullopt;
891    static constexpr auto FoundMismatch = nullptr;
892  
893    // Try to find a value of each element of an aggregate.
894    // FIXME: deal with more complex, not one-dimensional, aggregate types
895    SmallVector<std::optional<Instruction *>, 2> AggElts(NumAggElts, NotFound);
896  
897    // Do we know values for each element of the aggregate?
898    auto KnowAllElts = [&AggElts]() {
899      return !llvm::is_contained(AggElts, NotFound);
900    };
901  
902    int Depth = 0;
903  
904    // Arbitrary `insertvalue` visitation depth limit. Let's be okay with
905    // every element being overwritten twice, which should never happen.
906    static const int DepthLimit = 2 * NumAggElts;
907  
908    // Recurse up the chain of `insertvalue` aggregate operands until either we've
909    // reconstructed full initializer or can't visit any more `insertvalue`'s.
910    for (InsertValueInst *CurrIVI = &OrigIVI;
911         Depth < DepthLimit && CurrIVI && !KnowAllElts();
912         CurrIVI = dyn_cast<InsertValueInst>(CurrIVI->getAggregateOperand()),
913                         ++Depth) {
914      auto *InsertedValue =
915          dyn_cast<Instruction>(CurrIVI->getInsertedValueOperand());
916      if (!InsertedValue)
917        return nullptr; // Inserted value must be produced by an instruction.
918  
919      ArrayRef<unsigned int> Indices = CurrIVI->getIndices();
920  
921      // Don't bother with more than single-level aggregates.
922      if (Indices.size() != 1)
923        return nullptr; // FIXME: deal with more complex aggregates?
924  
925      // Now, we may have already previously recorded the value for this element
926      // of an aggregate. If we did, that means the CurrIVI will later be
927      // overwritten with the already-recorded value. But if not, let's record it!
928      std::optional<Instruction *> &Elt = AggElts[Indices.front()];
929      Elt = Elt.value_or(InsertedValue);
930  
931      // FIXME: should we handle chain-terminating undef base operand?
932    }
933  
934    // Was that sufficient to deduce the full initializer for the aggregate?
935    if (!KnowAllElts())
936      return nullptr; // Give up then.
937  
938    // We now want to find the source[s] of the aggregate elements we've found.
939    // And with "source" we mean the original aggregate[s] from which
940    // the inserted elements were extracted. This may require PHI translation.
941  
942    enum class AggregateDescription {
943      /// When analyzing the value that was inserted into an aggregate, we did
944      /// not manage to find defining `extractvalue` instruction to analyze.
945      NotFound,
946      /// When analyzing the value that was inserted into an aggregate, we did
947      /// manage to find defining `extractvalue` instruction[s], and everything
948      /// matched perfectly - aggregate type, element insertion/extraction index.
949      Found,
950      /// When analyzing the value that was inserted into an aggregate, we did
951      /// manage to find defining `extractvalue` instruction, but there was
952      /// a mismatch: either the source type from which the extraction was didn't
953      /// match the aggregate type into which the insertion was,
954      /// or the extraction/insertion channels mismatched,
955      /// or different elements had different source aggregates.
956      FoundMismatch
957    };
958    auto Describe = [](std::optional<Value *> SourceAggregate) {
959      if (SourceAggregate == NotFound)
960        return AggregateDescription::NotFound;
961      if (*SourceAggregate == FoundMismatch)
962        return AggregateDescription::FoundMismatch;
963      return AggregateDescription::Found;
964    };
965  
966    // Given the value \p Elt that was being inserted into element \p EltIdx of an
967    // aggregate AggTy, see if \p Elt was originally defined by an
968    // appropriate extractvalue (same element index, same aggregate type).
969    // If found, return the source aggregate from which the extraction was.
970    // If \p PredBB is provided, does PHI translation of an \p Elt first.
971    auto FindSourceAggregate =
972        [&](Instruction *Elt, unsigned EltIdx, std::optional<BasicBlock *> UseBB,
973            std::optional<BasicBlock *> PredBB) -> std::optional<Value *> {
974      // For now(?), only deal with, at most, a single level of PHI indirection.
975      if (UseBB && PredBB)
976        Elt = dyn_cast<Instruction>(Elt->DoPHITranslation(*UseBB, *PredBB));
977      // FIXME: deal with multiple levels of PHI indirection?
978  
979      // Did we find an extraction?
980      auto *EVI = dyn_cast_or_null<ExtractValueInst>(Elt);
981      if (!EVI)
982        return NotFound;
983  
984      Value *SourceAggregate = EVI->getAggregateOperand();
985  
986      // Is the extraction from the same type into which the insertion was?
987      if (SourceAggregate->getType() != AggTy)
988        return FoundMismatch;
989      // And the element index doesn't change between extraction and insertion?
990      if (EVI->getNumIndices() != 1 || EltIdx != EVI->getIndices().front())
991        return FoundMismatch;
992  
993      return SourceAggregate; // AggregateDescription::Found
994    };
995  
996    // Given elements AggElts that were constructing an aggregate OrigIVI,
997    // see if we can find appropriate source aggregate for each of the elements,
998    // and see it's the same aggregate for each element. If so, return it.
999    auto FindCommonSourceAggregate =
1000        [&](std::optional<BasicBlock *> UseBB,
1001            std::optional<BasicBlock *> PredBB) -> std::optional<Value *> {
1002      std::optional<Value *> SourceAggregate;
1003  
1004      for (auto I : enumerate(AggElts)) {
1005        assert(Describe(SourceAggregate) != AggregateDescription::FoundMismatch &&
1006               "We don't store nullptr in SourceAggregate!");
1007        assert((Describe(SourceAggregate) == AggregateDescription::Found) ==
1008                   (I.index() != 0) &&
1009               "SourceAggregate should be valid after the first element,");
1010  
1011        // For this element, is there a plausible source aggregate?
1012        // FIXME: we could special-case undef element, IFF we know that in the
1013        //        source aggregate said element isn't poison.
1014        std::optional<Value *> SourceAggregateForElement =
1015            FindSourceAggregate(*I.value(), I.index(), UseBB, PredBB);
1016  
1017        // Okay, what have we found? Does that correlate with previous findings?
1018  
1019        // Regardless of whether or not we have previously found source
1020        // aggregate for previous elements (if any), if we didn't find one for
1021        // this element, passthrough whatever we have just found.
1022        if (Describe(SourceAggregateForElement) != AggregateDescription::Found)
1023          return SourceAggregateForElement;
1024  
1025        // Okay, we have found source aggregate for this element.
1026        // Let's see what we already know from previous elements, if any.
1027        switch (Describe(SourceAggregate)) {
1028        case AggregateDescription::NotFound:
1029          // This is apparently the first element that we have examined.
1030          SourceAggregate = SourceAggregateForElement; // Record the aggregate!
1031          continue; // Great, now look at next element.
1032        case AggregateDescription::Found:
1033          // We have previously already successfully examined other elements.
1034          // Is this the same source aggregate we've found for other elements?
1035          if (*SourceAggregateForElement != *SourceAggregate)
1036            return FoundMismatch;
1037          continue; // Still the same aggregate, look at next element.
1038        case AggregateDescription::FoundMismatch:
1039          llvm_unreachable("Can't happen. We would have early-exited then.");
1040        };
1041      }
1042  
1043      assert(Describe(SourceAggregate) == AggregateDescription::Found &&
1044             "Must be a valid Value");
1045      return *SourceAggregate;
1046    };
1047  
1048    std::optional<Value *> SourceAggregate;
1049  
1050    // Can we find the source aggregate without looking at predecessors?
1051    SourceAggregate = FindCommonSourceAggregate(/*UseBB=*/std::nullopt,
1052                                                /*PredBB=*/std::nullopt);
1053    if (Describe(SourceAggregate) != AggregateDescription::NotFound) {
1054      if (Describe(SourceAggregate) == AggregateDescription::FoundMismatch)
1055        return nullptr; // Conflicting source aggregates!
1056      ++NumAggregateReconstructionsSimplified;
1057      return replaceInstUsesWith(OrigIVI, *SourceAggregate);
1058    }
1059  
1060    // Okay, apparently we need to look at predecessors.
1061  
1062    // We should be smart about picking the "use" basic block, which will be the
1063    // merge point for aggregate, where we'll insert the final PHI that will be
1064    // used instead of OrigIVI. Basic block of OrigIVI is *not* the right choice.
1065    // We should look in which blocks each of the AggElts is being defined,
1066    // they all should be defined in the same basic block.
1067    BasicBlock *UseBB = nullptr;
1068  
1069    for (const std::optional<Instruction *> &I : AggElts) {
1070      BasicBlock *BB = (*I)->getParent();
1071      // If it's the first instruction we've encountered, record the basic block.
1072      if (!UseBB) {
1073        UseBB = BB;
1074        continue;
1075      }
1076      // Otherwise, this must be the same basic block we've seen previously.
1077      if (UseBB != BB)
1078        return nullptr;
1079    }
1080  
1081    // If *all* of the elements are basic-block-independent, meaning they are
1082    // either function arguments, or constant expressions, then if we didn't
1083    // handle them without predecessor-aware handling, we won't handle them now.
1084    if (!UseBB)
1085      return nullptr;
1086  
1087    // If we didn't manage to find source aggregate without looking at
1088    // predecessors, and there are no predecessors to look at, then we're done.
1089    if (pred_empty(UseBB))
1090      return nullptr;
1091  
1092    // Arbitrary predecessor count limit.
1093    static const int PredCountLimit = 64;
1094  
1095    // Cache the (non-uniqified!) list of predecessors in a vector,
1096    // checking the limit at the same time for efficiency.
1097    SmallVector<BasicBlock *, 4> Preds; // May have duplicates!
1098    for (BasicBlock *Pred : predecessors(UseBB)) {
1099      // Don't bother if there are too many predecessors.
1100      if (Preds.size() >= PredCountLimit) // FIXME: only count duplicates once?
1101        return nullptr;
1102      Preds.emplace_back(Pred);
1103    }
1104  
1105    // For each predecessor, what is the source aggregate,
1106    // from which all the elements were originally extracted from?
1107    // Note that we want for the map to have stable iteration order!
1108    SmallDenseMap<BasicBlock *, Value *, 4> SourceAggregates;
1109    for (BasicBlock *Pred : Preds) {
1110      std::pair<decltype(SourceAggregates)::iterator, bool> IV =
1111          SourceAggregates.insert({Pred, nullptr});
1112      // Did we already evaluate this predecessor?
1113      if (!IV.second)
1114        continue;
1115  
1116      // Let's hope that when coming from predecessor Pred, all elements of the
1117      // aggregate produced by OrigIVI must have been originally extracted from
1118      // the same aggregate. Is that so? Can we find said original aggregate?
1119      SourceAggregate = FindCommonSourceAggregate(UseBB, Pred);
1120      if (Describe(SourceAggregate) != AggregateDescription::Found)
1121        return nullptr; // Give up.
1122      IV.first->second = *SourceAggregate;
1123    }
1124  
1125    // All good! Now we just need to thread the source aggregates here.
1126    // Note that we have to insert the new PHI here, ourselves, because we can't
1127    // rely on InstCombinerImpl::run() inserting it into the right basic block.
1128    // Note that the same block can be a predecessor more than once,
1129    // and we need to preserve that invariant for the PHI node.
1130    BuilderTy::InsertPointGuard Guard(Builder);
1131    Builder.SetInsertPoint(UseBB, UseBB->getFirstNonPHIIt());
1132    auto *PHI =
1133        Builder.CreatePHI(AggTy, Preds.size(), OrigIVI.getName() + ".merged");
1134    for (BasicBlock *Pred : Preds)
1135      PHI->addIncoming(SourceAggregates[Pred], Pred);
1136  
1137    ++NumAggregateReconstructionsSimplified;
1138    return replaceInstUsesWith(OrigIVI, PHI);
1139  }
1140  
1141  /// Try to find redundant insertvalue instructions, like the following ones:
1142  ///  %0 = insertvalue { i8, i32 } undef, i8 %x, 0
1143  ///  %1 = insertvalue { i8, i32 } %0,    i8 %y, 0
1144  /// Here the second instruction inserts values at the same indices, as the
1145  /// first one, making the first one redundant.
1146  /// It should be transformed to:
1147  ///  %0 = insertvalue { i8, i32 } undef, i8 %y, 0
visitInsertValueInst(InsertValueInst & I)1148  Instruction *InstCombinerImpl::visitInsertValueInst(InsertValueInst &I) {
1149    if (Value *V = simplifyInsertValueInst(
1150            I.getAggregateOperand(), I.getInsertedValueOperand(), I.getIndices(),
1151            SQ.getWithInstruction(&I)))
1152      return replaceInstUsesWith(I, V);
1153  
1154    bool IsRedundant = false;
1155    ArrayRef<unsigned int> FirstIndices = I.getIndices();
1156  
1157    // If there is a chain of insertvalue instructions (each of them except the
1158    // last one has only one use and it's another insertvalue insn from this
1159    // chain), check if any of the 'children' uses the same indices as the first
1160    // instruction. In this case, the first one is redundant.
1161    Value *V = &I;
1162    unsigned Depth = 0;
1163    while (V->hasOneUse() && Depth < 10) {
1164      User *U = V->user_back();
1165      auto UserInsInst = dyn_cast<InsertValueInst>(U);
1166      if (!UserInsInst || U->getOperand(0) != V)
1167        break;
1168      if (UserInsInst->getIndices() == FirstIndices) {
1169        IsRedundant = true;
1170        break;
1171      }
1172      V = UserInsInst;
1173      Depth++;
1174    }
1175  
1176    if (IsRedundant)
1177      return replaceInstUsesWith(I, I.getOperand(0));
1178  
1179    if (Instruction *NewI = foldAggregateConstructionIntoAggregateReuse(I))
1180      return NewI;
1181  
1182    return nullptr;
1183  }
1184  
isShuffleEquivalentToSelect(ShuffleVectorInst & Shuf)1185  static bool isShuffleEquivalentToSelect(ShuffleVectorInst &Shuf) {
1186    // Can not analyze scalable type, the number of elements is not a compile-time
1187    // constant.
1188    if (isa<ScalableVectorType>(Shuf.getOperand(0)->getType()))
1189      return false;
1190  
1191    int MaskSize = Shuf.getShuffleMask().size();
1192    int VecSize =
1193        cast<FixedVectorType>(Shuf.getOperand(0)->getType())->getNumElements();
1194  
1195    // A vector select does not change the size of the operands.
1196    if (MaskSize != VecSize)
1197      return false;
1198  
1199    // Each mask element must be undefined or choose a vector element from one of
1200    // the source operands without crossing vector lanes.
1201    for (int i = 0; i != MaskSize; ++i) {
1202      int Elt = Shuf.getMaskValue(i);
1203      if (Elt != -1 && Elt != i && Elt != i + VecSize)
1204        return false;
1205    }
1206  
1207    return true;
1208  }
1209  
1210  /// Turn a chain of inserts that splats a value into an insert + shuffle:
1211  /// insertelt(insertelt(insertelt(insertelt X, %k, 0), %k, 1), %k, 2) ... ->
1212  /// shufflevector(insertelt(X, %k, 0), poison, zero)
foldInsSequenceIntoSplat(InsertElementInst & InsElt)1213  static Instruction *foldInsSequenceIntoSplat(InsertElementInst &InsElt) {
1214    // We are interested in the last insert in a chain. So if this insert has a
1215    // single user and that user is an insert, bail.
1216    if (InsElt.hasOneUse() && isa<InsertElementInst>(InsElt.user_back()))
1217      return nullptr;
1218  
1219    VectorType *VecTy = InsElt.getType();
1220    // Can not handle scalable type, the number of elements is not a compile-time
1221    // constant.
1222    if (isa<ScalableVectorType>(VecTy))
1223      return nullptr;
1224    unsigned NumElements = cast<FixedVectorType>(VecTy)->getNumElements();
1225  
1226    // Do not try to do this for a one-element vector, since that's a nop,
1227    // and will cause an inf-loop.
1228    if (NumElements == 1)
1229      return nullptr;
1230  
1231    Value *SplatVal = InsElt.getOperand(1);
1232    InsertElementInst *CurrIE = &InsElt;
1233    SmallBitVector ElementPresent(NumElements, false);
1234    InsertElementInst *FirstIE = nullptr;
1235  
1236    // Walk the chain backwards, keeping track of which indices we inserted into,
1237    // until we hit something that isn't an insert of the splatted value.
1238    while (CurrIE) {
1239      auto *Idx = dyn_cast<ConstantInt>(CurrIE->getOperand(2));
1240      if (!Idx || CurrIE->getOperand(1) != SplatVal)
1241        return nullptr;
1242  
1243      auto *NextIE = dyn_cast<InsertElementInst>(CurrIE->getOperand(0));
1244      // Check none of the intermediate steps have any additional uses, except
1245      // for the root insertelement instruction, which can be re-used, if it
1246      // inserts at position 0.
1247      if (CurrIE != &InsElt &&
1248          (!CurrIE->hasOneUse() && (NextIE != nullptr || !Idx->isZero())))
1249        return nullptr;
1250  
1251      ElementPresent[Idx->getZExtValue()] = true;
1252      FirstIE = CurrIE;
1253      CurrIE = NextIE;
1254    }
1255  
1256    // If this is just a single insertelement (not a sequence), we are done.
1257    if (FirstIE == &InsElt)
1258      return nullptr;
1259  
1260    // If we are not inserting into a poison vector, make sure we've seen an
1261    // insert into every element.
1262    // TODO: If the base vector is not undef, it might be better to create a splat
1263    //       and then a select-shuffle (blend) with the base vector.
1264    if (!match(FirstIE->getOperand(0), m_Poison()))
1265      if (!ElementPresent.all())
1266        return nullptr;
1267  
1268    // Create the insert + shuffle.
1269    Type *Int64Ty = Type::getInt64Ty(InsElt.getContext());
1270    PoisonValue *PoisonVec = PoisonValue::get(VecTy);
1271    Constant *Zero = ConstantInt::get(Int64Ty, 0);
1272    if (!cast<ConstantInt>(FirstIE->getOperand(2))->isZero())
1273      FirstIE = InsertElementInst::Create(PoisonVec, SplatVal, Zero, "",
1274                                          InsElt.getIterator());
1275  
1276    // Splat from element 0, but replace absent elements with poison in the mask.
1277    SmallVector<int, 16> Mask(NumElements, 0);
1278    for (unsigned i = 0; i != NumElements; ++i)
1279      if (!ElementPresent[i])
1280        Mask[i] = -1;
1281  
1282    return new ShuffleVectorInst(FirstIE, Mask);
1283  }
1284  
1285  /// Try to fold an insert element into an existing splat shuffle by changing
1286  /// the shuffle's mask to include the index of this insert element.
foldInsEltIntoSplat(InsertElementInst & InsElt)1287  static Instruction *foldInsEltIntoSplat(InsertElementInst &InsElt) {
1288    // Check if the vector operand of this insert is a canonical splat shuffle.
1289    auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0));
1290    if (!Shuf || !Shuf->isZeroEltSplat())
1291      return nullptr;
1292  
1293    // Bail out early if shuffle is scalable type. The number of elements in
1294    // shuffle mask is unknown at compile-time.
1295    if (isa<ScalableVectorType>(Shuf->getType()))
1296      return nullptr;
1297  
1298    // Check for a constant insertion index.
1299    uint64_t IdxC;
1300    if (!match(InsElt.getOperand(2), m_ConstantInt(IdxC)))
1301      return nullptr;
1302  
1303    // Check if the splat shuffle's input is the same as this insert's scalar op.
1304    Value *X = InsElt.getOperand(1);
1305    Value *Op0 = Shuf->getOperand(0);
1306    if (!match(Op0, m_InsertElt(m_Undef(), m_Specific(X), m_ZeroInt())))
1307      return nullptr;
1308  
1309    // Replace the shuffle mask element at the index of this insert with a zero.
1310    // For example:
1311    // inselt (shuf (inselt undef, X, 0), _, <0,undef,0,undef>), X, 1
1312    //   --> shuf (inselt undef, X, 0), poison, <0,0,0,undef>
1313    unsigned NumMaskElts =
1314        cast<FixedVectorType>(Shuf->getType())->getNumElements();
1315    SmallVector<int, 16> NewMask(NumMaskElts);
1316    for (unsigned i = 0; i != NumMaskElts; ++i)
1317      NewMask[i] = i == IdxC ? 0 : Shuf->getMaskValue(i);
1318  
1319    return new ShuffleVectorInst(Op0, NewMask);
1320  }
1321  
1322  /// Try to fold an extract+insert element into an existing identity shuffle by
1323  /// changing the shuffle's mask to include the index of this insert element.
foldInsEltIntoIdentityShuffle(InsertElementInst & InsElt)1324  static Instruction *foldInsEltIntoIdentityShuffle(InsertElementInst &InsElt) {
1325    // Check if the vector operand of this insert is an identity shuffle.
1326    auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0));
1327    if (!Shuf || !match(Shuf->getOperand(1), m_Poison()) ||
1328        !(Shuf->isIdentityWithExtract() || Shuf->isIdentityWithPadding()))
1329      return nullptr;
1330  
1331    // Bail out early if shuffle is scalable type. The number of elements in
1332    // shuffle mask is unknown at compile-time.
1333    if (isa<ScalableVectorType>(Shuf->getType()))
1334      return nullptr;
1335  
1336    // Check for a constant insertion index.
1337    uint64_t IdxC;
1338    if (!match(InsElt.getOperand(2), m_ConstantInt(IdxC)))
1339      return nullptr;
1340  
1341    // Check if this insert's scalar op is extracted from the identity shuffle's
1342    // input vector.
1343    Value *Scalar = InsElt.getOperand(1);
1344    Value *X = Shuf->getOperand(0);
1345    if (!match(Scalar, m_ExtractElt(m_Specific(X), m_SpecificInt(IdxC))))
1346      return nullptr;
1347  
1348    // Replace the shuffle mask element at the index of this extract+insert with
1349    // that same index value.
1350    // For example:
1351    // inselt (shuf X, IdMask), (extelt X, IdxC), IdxC --> shuf X, IdMask'
1352    unsigned NumMaskElts =
1353        cast<FixedVectorType>(Shuf->getType())->getNumElements();
1354    SmallVector<int, 16> NewMask(NumMaskElts);
1355    ArrayRef<int> OldMask = Shuf->getShuffleMask();
1356    for (unsigned i = 0; i != NumMaskElts; ++i) {
1357      if (i != IdxC) {
1358        // All mask elements besides the inserted element remain the same.
1359        NewMask[i] = OldMask[i];
1360      } else if (OldMask[i] == (int)IdxC) {
1361        // If the mask element was already set, there's nothing to do
1362        // (demanded elements analysis may unset it later).
1363        return nullptr;
1364      } else {
1365        assert(OldMask[i] == PoisonMaskElem &&
1366               "Unexpected shuffle mask element for identity shuffle");
1367        NewMask[i] = IdxC;
1368      }
1369    }
1370  
1371    return new ShuffleVectorInst(X, Shuf->getOperand(1), NewMask);
1372  }
1373  
1374  /// If we have an insertelement instruction feeding into another insertelement
1375  /// and the 2nd is inserting a constant into the vector, canonicalize that
1376  /// constant insertion before the insertion of a variable:
1377  ///
1378  /// insertelement (insertelement X, Y, IdxC1), ScalarC, IdxC2 -->
1379  /// insertelement (insertelement X, ScalarC, IdxC2), Y, IdxC1
1380  ///
1381  /// This has the potential of eliminating the 2nd insertelement instruction
1382  /// via constant folding of the scalar constant into a vector constant.
hoistInsEltConst(InsertElementInst & InsElt2,InstCombiner::BuilderTy & Builder)1383  static Instruction *hoistInsEltConst(InsertElementInst &InsElt2,
1384                                       InstCombiner::BuilderTy &Builder) {
1385    auto *InsElt1 = dyn_cast<InsertElementInst>(InsElt2.getOperand(0));
1386    if (!InsElt1 || !InsElt1->hasOneUse())
1387      return nullptr;
1388  
1389    Value *X, *Y;
1390    Constant *ScalarC;
1391    ConstantInt *IdxC1, *IdxC2;
1392    if (match(InsElt1->getOperand(0), m_Value(X)) &&
1393        match(InsElt1->getOperand(1), m_Value(Y)) && !isa<Constant>(Y) &&
1394        match(InsElt1->getOperand(2), m_ConstantInt(IdxC1)) &&
1395        match(InsElt2.getOperand(1), m_Constant(ScalarC)) &&
1396        match(InsElt2.getOperand(2), m_ConstantInt(IdxC2)) && IdxC1 != IdxC2) {
1397      Value *NewInsElt1 = Builder.CreateInsertElement(X, ScalarC, IdxC2);
1398      return InsertElementInst::Create(NewInsElt1, Y, IdxC1);
1399    }
1400  
1401    return nullptr;
1402  }
1403  
1404  /// insertelt (shufflevector X, CVec, Mask|insertelt X, C1, CIndex1), C, CIndex
1405  /// --> shufflevector X, CVec', Mask'
foldConstantInsEltIntoShuffle(InsertElementInst & InsElt)1406  static Instruction *foldConstantInsEltIntoShuffle(InsertElementInst &InsElt) {
1407    auto *Inst = dyn_cast<Instruction>(InsElt.getOperand(0));
1408    // Bail out if the parent has more than one use. In that case, we'd be
1409    // replacing the insertelt with a shuffle, and that's not a clear win.
1410    if (!Inst || !Inst->hasOneUse())
1411      return nullptr;
1412    if (auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0))) {
1413      // The shuffle must have a constant vector operand. The insertelt must have
1414      // a constant scalar being inserted at a constant position in the vector.
1415      Constant *ShufConstVec, *InsEltScalar;
1416      uint64_t InsEltIndex;
1417      if (!match(Shuf->getOperand(1), m_Constant(ShufConstVec)) ||
1418          !match(InsElt.getOperand(1), m_Constant(InsEltScalar)) ||
1419          !match(InsElt.getOperand(2), m_ConstantInt(InsEltIndex)))
1420        return nullptr;
1421  
1422      // Adding an element to an arbitrary shuffle could be expensive, but a
1423      // shuffle that selects elements from vectors without crossing lanes is
1424      // assumed cheap.
1425      // If we're just adding a constant into that shuffle, it will still be
1426      // cheap.
1427      if (!isShuffleEquivalentToSelect(*Shuf))
1428        return nullptr;
1429  
1430      // From the above 'select' check, we know that the mask has the same number
1431      // of elements as the vector input operands. We also know that each constant
1432      // input element is used in its lane and can not be used more than once by
1433      // the shuffle. Therefore, replace the constant in the shuffle's constant
1434      // vector with the insertelt constant. Replace the constant in the shuffle's
1435      // mask vector with the insertelt index plus the length of the vector
1436      // (because the constant vector operand of a shuffle is always the 2nd
1437      // operand).
1438      ArrayRef<int> Mask = Shuf->getShuffleMask();
1439      unsigned NumElts = Mask.size();
1440      SmallVector<Constant *, 16> NewShufElts(NumElts);
1441      SmallVector<int, 16> NewMaskElts(NumElts);
1442      for (unsigned I = 0; I != NumElts; ++I) {
1443        if (I == InsEltIndex) {
1444          NewShufElts[I] = InsEltScalar;
1445          NewMaskElts[I] = InsEltIndex + NumElts;
1446        } else {
1447          // Copy over the existing values.
1448          NewShufElts[I] = ShufConstVec->getAggregateElement(I);
1449          NewMaskElts[I] = Mask[I];
1450        }
1451  
1452        // Bail if we failed to find an element.
1453        if (!NewShufElts[I])
1454          return nullptr;
1455      }
1456  
1457      // Create new operands for a shuffle that includes the constant of the
1458      // original insertelt. The old shuffle will be dead now.
1459      return new ShuffleVectorInst(Shuf->getOperand(0),
1460                                   ConstantVector::get(NewShufElts), NewMaskElts);
1461    } else if (auto *IEI = dyn_cast<InsertElementInst>(Inst)) {
1462      // Transform sequences of insertelements ops with constant data/indexes into
1463      // a single shuffle op.
1464      // Can not handle scalable type, the number of elements needed to create
1465      // shuffle mask is not a compile-time constant.
1466      if (isa<ScalableVectorType>(InsElt.getType()))
1467        return nullptr;
1468      unsigned NumElts =
1469          cast<FixedVectorType>(InsElt.getType())->getNumElements();
1470  
1471      uint64_t InsertIdx[2];
1472      Constant *Val[2];
1473      if (!match(InsElt.getOperand(2), m_ConstantInt(InsertIdx[0])) ||
1474          !match(InsElt.getOperand(1), m_Constant(Val[0])) ||
1475          !match(IEI->getOperand(2), m_ConstantInt(InsertIdx[1])) ||
1476          !match(IEI->getOperand(1), m_Constant(Val[1])))
1477        return nullptr;
1478      SmallVector<Constant *, 16> Values(NumElts);
1479      SmallVector<int, 16> Mask(NumElts);
1480      auto ValI = std::begin(Val);
1481      // Generate new constant vector and mask.
1482      // We have 2 values/masks from the insertelements instructions. Insert them
1483      // into new value/mask vectors.
1484      for (uint64_t I : InsertIdx) {
1485        if (!Values[I]) {
1486          Values[I] = *ValI;
1487          Mask[I] = NumElts + I;
1488        }
1489        ++ValI;
1490      }
1491      // Remaining values are filled with 'poison' values.
1492      for (unsigned I = 0; I < NumElts; ++I) {
1493        if (!Values[I]) {
1494          Values[I] = PoisonValue::get(InsElt.getType()->getElementType());
1495          Mask[I] = I;
1496        }
1497      }
1498      // Create new operands for a shuffle that includes the constant of the
1499      // original insertelt.
1500      return new ShuffleVectorInst(IEI->getOperand(0),
1501                                   ConstantVector::get(Values), Mask);
1502    }
1503    return nullptr;
1504  }
1505  
1506  /// If both the base vector and the inserted element are extended from the same
1507  /// type, do the insert element in the narrow source type followed by extend.
1508  /// TODO: This can be extended to include other cast opcodes, but particularly
1509  ///       if we create a wider insertelement, make sure codegen is not harmed.
narrowInsElt(InsertElementInst & InsElt,InstCombiner::BuilderTy & Builder)1510  static Instruction *narrowInsElt(InsertElementInst &InsElt,
1511                                   InstCombiner::BuilderTy &Builder) {
1512    // We are creating a vector extend. If the original vector extend has another
1513    // use, that would mean we end up with 2 vector extends, so avoid that.
1514    // TODO: We could ease the use-clause to "if at least one op has one use"
1515    //       (assuming that the source types match - see next TODO comment).
1516    Value *Vec = InsElt.getOperand(0);
1517    if (!Vec->hasOneUse())
1518      return nullptr;
1519  
1520    Value *Scalar = InsElt.getOperand(1);
1521    Value *X, *Y;
1522    CastInst::CastOps CastOpcode;
1523    if (match(Vec, m_FPExt(m_Value(X))) && match(Scalar, m_FPExt(m_Value(Y))))
1524      CastOpcode = Instruction::FPExt;
1525    else if (match(Vec, m_SExt(m_Value(X))) && match(Scalar, m_SExt(m_Value(Y))))
1526      CastOpcode = Instruction::SExt;
1527    else if (match(Vec, m_ZExt(m_Value(X))) && match(Scalar, m_ZExt(m_Value(Y))))
1528      CastOpcode = Instruction::ZExt;
1529    else
1530      return nullptr;
1531  
1532    // TODO: We can allow mismatched types by creating an intermediate cast.
1533    if (X->getType()->getScalarType() != Y->getType())
1534      return nullptr;
1535  
1536    // inselt (ext X), (ext Y), Index --> ext (inselt X, Y, Index)
1537    Value *NewInsElt = Builder.CreateInsertElement(X, Y, InsElt.getOperand(2));
1538    return CastInst::Create(CastOpcode, NewInsElt, InsElt.getType());
1539  }
1540  
1541  /// If we are inserting 2 halves of a value into adjacent elements of a vector,
1542  /// try to convert to a single insert with appropriate bitcasts.
foldTruncInsEltPair(InsertElementInst & InsElt,bool IsBigEndian,InstCombiner::BuilderTy & Builder)1543  static Instruction *foldTruncInsEltPair(InsertElementInst &InsElt,
1544                                          bool IsBigEndian,
1545                                          InstCombiner::BuilderTy &Builder) {
1546    Value *VecOp    = InsElt.getOperand(0);
1547    Value *ScalarOp = InsElt.getOperand(1);
1548    Value *IndexOp  = InsElt.getOperand(2);
1549  
1550    // Pattern depends on endian because we expect lower index is inserted first.
1551    // Big endian:
1552    // inselt (inselt BaseVec, (trunc (lshr X, BW/2), Index0), (trunc X), Index1
1553    // Little endian:
1554    // inselt (inselt BaseVec, (trunc X), Index0), (trunc (lshr X, BW/2)), Index1
1555    // Note: It is not safe to do this transform with an arbitrary base vector
1556    //       because the bitcast of that vector to fewer/larger elements could
1557    //       allow poison to spill into an element that was not poison before.
1558    // TODO: Detect smaller fractions of the scalar.
1559    // TODO: One-use checks are conservative.
1560    auto *VTy = dyn_cast<FixedVectorType>(InsElt.getType());
1561    Value *Scalar0, *BaseVec;
1562    uint64_t Index0, Index1;
1563    if (!VTy || (VTy->getNumElements() & 1) ||
1564        !match(IndexOp, m_ConstantInt(Index1)) ||
1565        !match(VecOp, m_InsertElt(m_Value(BaseVec), m_Value(Scalar0),
1566                                  m_ConstantInt(Index0))) ||
1567        !match(BaseVec, m_Undef()))
1568      return nullptr;
1569  
1570    // The first insert must be to the index one less than this one, and
1571    // the first insert must be to an even index.
1572    if (Index0 + 1 != Index1 || Index0 & 1)
1573      return nullptr;
1574  
1575    // For big endian, the high half of the value should be inserted first.
1576    // For little endian, the low half of the value should be inserted first.
1577    Value *X;
1578    uint64_t ShAmt;
1579    if (IsBigEndian) {
1580      if (!match(ScalarOp, m_Trunc(m_Value(X))) ||
1581          !match(Scalar0, m_Trunc(m_LShr(m_Specific(X), m_ConstantInt(ShAmt)))))
1582        return nullptr;
1583    } else {
1584      if (!match(Scalar0, m_Trunc(m_Value(X))) ||
1585          !match(ScalarOp, m_Trunc(m_LShr(m_Specific(X), m_ConstantInt(ShAmt)))))
1586        return nullptr;
1587    }
1588  
1589    Type *SrcTy = X->getType();
1590    unsigned ScalarWidth = SrcTy->getScalarSizeInBits();
1591    unsigned VecEltWidth = VTy->getScalarSizeInBits();
1592    if (ScalarWidth != VecEltWidth * 2 || ShAmt != VecEltWidth)
1593      return nullptr;
1594  
1595    // Bitcast the base vector to a vector type with the source element type.
1596    Type *CastTy = FixedVectorType::get(SrcTy, VTy->getNumElements() / 2);
1597    Value *CastBaseVec = Builder.CreateBitCast(BaseVec, CastTy);
1598  
1599    // Scale the insert index for a vector with half as many elements.
1600    // bitcast (inselt (bitcast BaseVec), X, NewIndex)
1601    uint64_t NewIndex = IsBigEndian ? Index1 / 2 : Index0 / 2;
1602    Value *NewInsert = Builder.CreateInsertElement(CastBaseVec, X, NewIndex);
1603    return new BitCastInst(NewInsert, VTy);
1604  }
1605  
visitInsertElementInst(InsertElementInst & IE)1606  Instruction *InstCombinerImpl::visitInsertElementInst(InsertElementInst &IE) {
1607    Value *VecOp    = IE.getOperand(0);
1608    Value *ScalarOp = IE.getOperand(1);
1609    Value *IdxOp    = IE.getOperand(2);
1610  
1611    if (auto *V = simplifyInsertElementInst(
1612            VecOp, ScalarOp, IdxOp, SQ.getWithInstruction(&IE)))
1613      return replaceInstUsesWith(IE, V);
1614  
1615    // Canonicalize type of constant indices to i64 to simplify CSE
1616    if (auto *IndexC = dyn_cast<ConstantInt>(IdxOp)) {
1617      if (auto *NewIdx = getPreferredVectorIndex(IndexC))
1618        return replaceOperand(IE, 2, NewIdx);
1619  
1620      Value *BaseVec, *OtherScalar;
1621      uint64_t OtherIndexVal;
1622      if (match(VecOp, m_OneUse(m_InsertElt(m_Value(BaseVec),
1623                                            m_Value(OtherScalar),
1624                                            m_ConstantInt(OtherIndexVal)))) &&
1625          !isa<Constant>(OtherScalar) && OtherIndexVal > IndexC->getZExtValue()) {
1626        Value *NewIns = Builder.CreateInsertElement(BaseVec, ScalarOp, IdxOp);
1627        return InsertElementInst::Create(NewIns, OtherScalar,
1628                                         Builder.getInt64(OtherIndexVal));
1629      }
1630    }
1631  
1632    // If the scalar is bitcast and inserted into undef, do the insert in the
1633    // source type followed by bitcast.
1634    // TODO: Generalize for insert into any constant, not just undef?
1635    Value *ScalarSrc;
1636    if (match(VecOp, m_Undef()) &&
1637        match(ScalarOp, m_OneUse(m_BitCast(m_Value(ScalarSrc)))) &&
1638        (ScalarSrc->getType()->isIntegerTy() ||
1639         ScalarSrc->getType()->isFloatingPointTy())) {
1640      // inselt undef, (bitcast ScalarSrc), IdxOp -->
1641      //   bitcast (inselt undef, ScalarSrc, IdxOp)
1642      Type *ScalarTy = ScalarSrc->getType();
1643      Type *VecTy = VectorType::get(ScalarTy, IE.getType()->getElementCount());
1644      Constant *NewUndef = isa<PoisonValue>(VecOp) ? PoisonValue::get(VecTy)
1645                                                   : UndefValue::get(VecTy);
1646      Value *NewInsElt = Builder.CreateInsertElement(NewUndef, ScalarSrc, IdxOp);
1647      return new BitCastInst(NewInsElt, IE.getType());
1648    }
1649  
1650    // If the vector and scalar are both bitcast from the same element type, do
1651    // the insert in that source type followed by bitcast.
1652    Value *VecSrc;
1653    if (match(VecOp, m_BitCast(m_Value(VecSrc))) &&
1654        match(ScalarOp, m_BitCast(m_Value(ScalarSrc))) &&
1655        (VecOp->hasOneUse() || ScalarOp->hasOneUse()) &&
1656        VecSrc->getType()->isVectorTy() && !ScalarSrc->getType()->isVectorTy() &&
1657        cast<VectorType>(VecSrc->getType())->getElementType() ==
1658            ScalarSrc->getType()) {
1659      // inselt (bitcast VecSrc), (bitcast ScalarSrc), IdxOp -->
1660      //   bitcast (inselt VecSrc, ScalarSrc, IdxOp)
1661      Value *NewInsElt = Builder.CreateInsertElement(VecSrc, ScalarSrc, IdxOp);
1662      return new BitCastInst(NewInsElt, IE.getType());
1663    }
1664  
1665    // If the inserted element was extracted from some other fixed-length vector
1666    // and both indexes are valid constants, try to turn this into a shuffle.
1667    // Can not handle scalable vector type, the number of elements needed to
1668    // create shuffle mask is not a compile-time constant.
1669    uint64_t InsertedIdx, ExtractedIdx;
1670    Value *ExtVecOp;
1671    if (isa<FixedVectorType>(IE.getType()) &&
1672        match(IdxOp, m_ConstantInt(InsertedIdx)) &&
1673        match(ScalarOp,
1674              m_ExtractElt(m_Value(ExtVecOp), m_ConstantInt(ExtractedIdx))) &&
1675        isa<FixedVectorType>(ExtVecOp->getType()) &&
1676        ExtractedIdx <
1677            cast<FixedVectorType>(ExtVecOp->getType())->getNumElements()) {
1678      // TODO: Looking at the user(s) to determine if this insert is a
1679      // fold-to-shuffle opportunity does not match the usual instcombine
1680      // constraints. We should decide if the transform is worthy based only
1681      // on this instruction and its operands, but that may not work currently.
1682      //
1683      // Here, we are trying to avoid creating shuffles before reaching
1684      // the end of a chain of extract-insert pairs. This is complicated because
1685      // we do not generally form arbitrary shuffle masks in instcombine
1686      // (because those may codegen poorly), but collectShuffleElements() does
1687      // exactly that.
1688      //
1689      // The rules for determining what is an acceptable target-independent
1690      // shuffle mask are fuzzy because they evolve based on the backend's
1691      // capabilities and real-world impact.
1692      auto isShuffleRootCandidate = [](InsertElementInst &Insert) {
1693        if (!Insert.hasOneUse())
1694          return true;
1695        auto *InsertUser = dyn_cast<InsertElementInst>(Insert.user_back());
1696        if (!InsertUser)
1697          return true;
1698        return false;
1699      };
1700  
1701      // Try to form a shuffle from a chain of extract-insert ops.
1702      if (isShuffleRootCandidate(IE)) {
1703        bool Rerun = true;
1704        while (Rerun) {
1705          Rerun = false;
1706  
1707          SmallVector<int, 16> Mask;
1708          ShuffleOps LR =
1709              collectShuffleElements(&IE, Mask, nullptr, *this, Rerun);
1710  
1711          // The proposed shuffle may be trivial, in which case we shouldn't
1712          // perform the combine.
1713          if (LR.first != &IE && LR.second != &IE) {
1714            // We now have a shuffle of LHS, RHS, Mask.
1715            if (LR.second == nullptr)
1716              LR.second = PoisonValue::get(LR.first->getType());
1717            return new ShuffleVectorInst(LR.first, LR.second, Mask);
1718          }
1719        }
1720      }
1721    }
1722  
1723    if (auto VecTy = dyn_cast<FixedVectorType>(VecOp->getType())) {
1724      unsigned VWidth = VecTy->getNumElements();
1725      APInt PoisonElts(VWidth, 0);
1726      APInt AllOnesEltMask(APInt::getAllOnes(VWidth));
1727      if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask,
1728                                                PoisonElts)) {
1729        if (V != &IE)
1730          return replaceInstUsesWith(IE, V);
1731        return &IE;
1732      }
1733    }
1734  
1735    if (Instruction *Shuf = foldConstantInsEltIntoShuffle(IE))
1736      return Shuf;
1737  
1738    if (Instruction *NewInsElt = hoistInsEltConst(IE, Builder))
1739      return NewInsElt;
1740  
1741    if (Instruction *Broadcast = foldInsSequenceIntoSplat(IE))
1742      return Broadcast;
1743  
1744    if (Instruction *Splat = foldInsEltIntoSplat(IE))
1745      return Splat;
1746  
1747    if (Instruction *IdentityShuf = foldInsEltIntoIdentityShuffle(IE))
1748      return IdentityShuf;
1749  
1750    if (Instruction *Ext = narrowInsElt(IE, Builder))
1751      return Ext;
1752  
1753    if (Instruction *Ext = foldTruncInsEltPair(IE, DL.isBigEndian(), Builder))
1754      return Ext;
1755  
1756    return nullptr;
1757  }
1758  
1759  /// Return true if we can evaluate the specified expression tree if the vector
1760  /// elements were shuffled in a different order.
canEvaluateShuffled(Value * V,ArrayRef<int> Mask,unsigned Depth=5)1761  static bool canEvaluateShuffled(Value *V, ArrayRef<int> Mask,
1762                                  unsigned Depth = 5) {
1763    // We can always reorder the elements of a constant.
1764    if (isa<Constant>(V))
1765      return true;
1766  
1767    // We won't reorder vector arguments. No IPO here.
1768    Instruction *I = dyn_cast<Instruction>(V);
1769    if (!I) return false;
1770  
1771    // Two users may expect different orders of the elements. Don't try it.
1772    if (!I->hasOneUse())
1773      return false;
1774  
1775    if (Depth == 0) return false;
1776  
1777    switch (I->getOpcode()) {
1778      case Instruction::UDiv:
1779      case Instruction::SDiv:
1780      case Instruction::URem:
1781      case Instruction::SRem:
1782        // Propagating an undefined shuffle mask element to integer div/rem is not
1783        // allowed because those opcodes can create immediate undefined behavior
1784        // from an undefined element in an operand.
1785        if (llvm::is_contained(Mask, -1))
1786          return false;
1787        [[fallthrough]];
1788      case Instruction::Add:
1789      case Instruction::FAdd:
1790      case Instruction::Sub:
1791      case Instruction::FSub:
1792      case Instruction::Mul:
1793      case Instruction::FMul:
1794      case Instruction::FDiv:
1795      case Instruction::FRem:
1796      case Instruction::Shl:
1797      case Instruction::LShr:
1798      case Instruction::AShr:
1799      case Instruction::And:
1800      case Instruction::Or:
1801      case Instruction::Xor:
1802      case Instruction::ICmp:
1803      case Instruction::FCmp:
1804      case Instruction::Trunc:
1805      case Instruction::ZExt:
1806      case Instruction::SExt:
1807      case Instruction::FPToUI:
1808      case Instruction::FPToSI:
1809      case Instruction::UIToFP:
1810      case Instruction::SIToFP:
1811      case Instruction::FPTrunc:
1812      case Instruction::FPExt:
1813      case Instruction::GetElementPtr: {
1814        // Bail out if we would create longer vector ops. We could allow creating
1815        // longer vector ops, but that may result in more expensive codegen.
1816        Type *ITy = I->getType();
1817        if (ITy->isVectorTy() &&
1818            Mask.size() > cast<FixedVectorType>(ITy)->getNumElements())
1819          return false;
1820        for (Value *Operand : I->operands()) {
1821          if (!canEvaluateShuffled(Operand, Mask, Depth - 1))
1822            return false;
1823        }
1824        return true;
1825      }
1826      case Instruction::InsertElement: {
1827        ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
1828        if (!CI) return false;
1829        int ElementNumber = CI->getLimitedValue();
1830  
1831        // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
1832        // can't put an element into multiple indices.
1833        bool SeenOnce = false;
1834        for (int I : Mask) {
1835          if (I == ElementNumber) {
1836            if (SeenOnce)
1837              return false;
1838            SeenOnce = true;
1839          }
1840        }
1841        return canEvaluateShuffled(I->getOperand(0), Mask, Depth - 1);
1842      }
1843    }
1844    return false;
1845  }
1846  
1847  /// Rebuild a new instruction just like 'I' but with the new operands given.
1848  /// In the event of type mismatch, the type of the operands is correct.
buildNew(Instruction * I,ArrayRef<Value * > NewOps,IRBuilderBase & Builder)1849  static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps,
1850                         IRBuilderBase &Builder) {
1851    Builder.SetInsertPoint(I);
1852    switch (I->getOpcode()) {
1853      case Instruction::Add:
1854      case Instruction::FAdd:
1855      case Instruction::Sub:
1856      case Instruction::FSub:
1857      case Instruction::Mul:
1858      case Instruction::FMul:
1859      case Instruction::UDiv:
1860      case Instruction::SDiv:
1861      case Instruction::FDiv:
1862      case Instruction::URem:
1863      case Instruction::SRem:
1864      case Instruction::FRem:
1865      case Instruction::Shl:
1866      case Instruction::LShr:
1867      case Instruction::AShr:
1868      case Instruction::And:
1869      case Instruction::Or:
1870      case Instruction::Xor: {
1871        BinaryOperator *BO = cast<BinaryOperator>(I);
1872        assert(NewOps.size() == 2 && "binary operator with #ops != 2");
1873        Value *New = Builder.CreateBinOp(cast<BinaryOperator>(I)->getOpcode(),
1874                                         NewOps[0], NewOps[1]);
1875        if (auto *NewI = dyn_cast<Instruction>(New)) {
1876          if (isa<OverflowingBinaryOperator>(BO)) {
1877            NewI->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
1878            NewI->setHasNoSignedWrap(BO->hasNoSignedWrap());
1879          }
1880          if (isa<PossiblyExactOperator>(BO)) {
1881            NewI->setIsExact(BO->isExact());
1882          }
1883          if (isa<FPMathOperator>(BO))
1884            NewI->copyFastMathFlags(I);
1885        }
1886        return New;
1887      }
1888      case Instruction::ICmp:
1889        assert(NewOps.size() == 2 && "icmp with #ops != 2");
1890        return Builder.CreateICmp(cast<ICmpInst>(I)->getPredicate(), NewOps[0],
1891                                  NewOps[1]);
1892      case Instruction::FCmp:
1893        assert(NewOps.size() == 2 && "fcmp with #ops != 2");
1894        return Builder.CreateFCmp(cast<FCmpInst>(I)->getPredicate(), NewOps[0],
1895                                  NewOps[1]);
1896      case Instruction::Trunc:
1897      case Instruction::ZExt:
1898      case Instruction::SExt:
1899      case Instruction::FPToUI:
1900      case Instruction::FPToSI:
1901      case Instruction::UIToFP:
1902      case Instruction::SIToFP:
1903      case Instruction::FPTrunc:
1904      case Instruction::FPExt: {
1905        // It's possible that the mask has a different number of elements from
1906        // the original cast. We recompute the destination type to match the mask.
1907        Type *DestTy = VectorType::get(
1908            I->getType()->getScalarType(),
1909            cast<VectorType>(NewOps[0]->getType())->getElementCount());
1910        assert(NewOps.size() == 1 && "cast with #ops != 1");
1911        return Builder.CreateCast(cast<CastInst>(I)->getOpcode(), NewOps[0],
1912                                  DestTy);
1913      }
1914      case Instruction::GetElementPtr: {
1915        Value *Ptr = NewOps[0];
1916        ArrayRef<Value*> Idx = NewOps.slice(1);
1917        return Builder.CreateGEP(cast<GEPOperator>(I)->getSourceElementType(),
1918                                 Ptr, Idx, "",
1919                                 cast<GEPOperator>(I)->isInBounds());
1920      }
1921    }
1922    llvm_unreachable("failed to rebuild vector instructions");
1923  }
1924  
evaluateInDifferentElementOrder(Value * V,ArrayRef<int> Mask,IRBuilderBase & Builder)1925  static Value *evaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask,
1926                                                IRBuilderBase &Builder) {
1927    // Mask.size() does not need to be equal to the number of vector elements.
1928  
1929    assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
1930    Type *EltTy = V->getType()->getScalarType();
1931  
1932    if (isa<PoisonValue>(V))
1933      return PoisonValue::get(FixedVectorType::get(EltTy, Mask.size()));
1934  
1935    if (match(V, m_Undef()))
1936      return UndefValue::get(FixedVectorType::get(EltTy, Mask.size()));
1937  
1938    if (isa<ConstantAggregateZero>(V))
1939      return ConstantAggregateZero::get(FixedVectorType::get(EltTy, Mask.size()));
1940  
1941    if (Constant *C = dyn_cast<Constant>(V))
1942      return ConstantExpr::getShuffleVector(C, PoisonValue::get(C->getType()),
1943                                            Mask);
1944  
1945    Instruction *I = cast<Instruction>(V);
1946    switch (I->getOpcode()) {
1947      case Instruction::Add:
1948      case Instruction::FAdd:
1949      case Instruction::Sub:
1950      case Instruction::FSub:
1951      case Instruction::Mul:
1952      case Instruction::FMul:
1953      case Instruction::UDiv:
1954      case Instruction::SDiv:
1955      case Instruction::FDiv:
1956      case Instruction::URem:
1957      case Instruction::SRem:
1958      case Instruction::FRem:
1959      case Instruction::Shl:
1960      case Instruction::LShr:
1961      case Instruction::AShr:
1962      case Instruction::And:
1963      case Instruction::Or:
1964      case Instruction::Xor:
1965      case Instruction::ICmp:
1966      case Instruction::FCmp:
1967      case Instruction::Trunc:
1968      case Instruction::ZExt:
1969      case Instruction::SExt:
1970      case Instruction::FPToUI:
1971      case Instruction::FPToSI:
1972      case Instruction::UIToFP:
1973      case Instruction::SIToFP:
1974      case Instruction::FPTrunc:
1975      case Instruction::FPExt:
1976      case Instruction::Select:
1977      case Instruction::GetElementPtr: {
1978        SmallVector<Value*, 8> NewOps;
1979        bool NeedsRebuild =
1980            (Mask.size() !=
1981             cast<FixedVectorType>(I->getType())->getNumElements());
1982        for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
1983          Value *V;
1984          // Recursively call evaluateInDifferentElementOrder on vector arguments
1985          // as well. E.g. GetElementPtr may have scalar operands even if the
1986          // return value is a vector, so we need to examine the operand type.
1987          if (I->getOperand(i)->getType()->isVectorTy())
1988            V = evaluateInDifferentElementOrder(I->getOperand(i), Mask, Builder);
1989          else
1990            V = I->getOperand(i);
1991          NewOps.push_back(V);
1992          NeedsRebuild |= (V != I->getOperand(i));
1993        }
1994        if (NeedsRebuild)
1995          return buildNew(I, NewOps, Builder);
1996        return I;
1997      }
1998      case Instruction::InsertElement: {
1999        int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
2000  
2001        // The insertelement was inserting at Element. Figure out which element
2002        // that becomes after shuffling. The answer is guaranteed to be unique
2003        // by CanEvaluateShuffled.
2004        bool Found = false;
2005        int Index = 0;
2006        for (int e = Mask.size(); Index != e; ++Index) {
2007          if (Mask[Index] == Element) {
2008            Found = true;
2009            break;
2010          }
2011        }
2012  
2013        // If element is not in Mask, no need to handle the operand 1 (element to
2014        // be inserted). Just evaluate values in operand 0 according to Mask.
2015        if (!Found)
2016          return evaluateInDifferentElementOrder(I->getOperand(0), Mask, Builder);
2017  
2018        Value *V = evaluateInDifferentElementOrder(I->getOperand(0), Mask,
2019                                                   Builder);
2020        Builder.SetInsertPoint(I);
2021        return Builder.CreateInsertElement(V, I->getOperand(1), Index);
2022      }
2023    }
2024    llvm_unreachable("failed to reorder elements of vector instruction!");
2025  }
2026  
2027  // Returns true if the shuffle is extracting a contiguous range of values from
2028  // LHS, for example:
2029  //                 +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
2030  //   Input:        |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
2031  //   Shuffles to:  |EE|FF|GG|HH|
2032  //                 +--+--+--+--+
isShuffleExtractingFromLHS(ShuffleVectorInst & SVI,ArrayRef<int> Mask)2033  static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
2034                                         ArrayRef<int> Mask) {
2035    unsigned LHSElems =
2036        cast<FixedVectorType>(SVI.getOperand(0)->getType())->getNumElements();
2037    unsigned MaskElems = Mask.size();
2038    unsigned BegIdx = Mask.front();
2039    unsigned EndIdx = Mask.back();
2040    if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
2041      return false;
2042    for (unsigned I = 0; I != MaskElems; ++I)
2043      if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
2044        return false;
2045    return true;
2046  }
2047  
2048  /// These are the ingredients in an alternate form binary operator as described
2049  /// below.
2050  struct BinopElts {
2051    BinaryOperator::BinaryOps Opcode;
2052    Value *Op0;
2053    Value *Op1;
BinopEltsBinopElts2054    BinopElts(BinaryOperator::BinaryOps Opc = (BinaryOperator::BinaryOps)0,
2055              Value *V0 = nullptr, Value *V1 = nullptr) :
2056        Opcode(Opc), Op0(V0), Op1(V1) {}
operator boolBinopElts2057    operator bool() const { return Opcode != 0; }
2058  };
2059  
2060  /// Binops may be transformed into binops with different opcodes and operands.
2061  /// Reverse the usual canonicalization to enable folds with the non-canonical
2062  /// form of the binop. If a transform is possible, return the elements of the
2063  /// new binop. If not, return invalid elements.
getAlternateBinop(BinaryOperator * BO,const DataLayout & DL)2064  static BinopElts getAlternateBinop(BinaryOperator *BO, const DataLayout &DL) {
2065    Value *BO0 = BO->getOperand(0), *BO1 = BO->getOperand(1);
2066    Type *Ty = BO->getType();
2067    switch (BO->getOpcode()) {
2068    case Instruction::Shl: {
2069      // shl X, C --> mul X, (1 << C)
2070      Constant *C;
2071      if (match(BO1, m_ImmConstant(C))) {
2072        Constant *ShlOne = ConstantFoldBinaryOpOperands(
2073            Instruction::Shl, ConstantInt::get(Ty, 1), C, DL);
2074        assert(ShlOne && "Constant folding of immediate constants failed");
2075        return {Instruction::Mul, BO0, ShlOne};
2076      }
2077      break;
2078    }
2079    case Instruction::Or: {
2080      // or disjoin X, C --> add X, C
2081      if (cast<PossiblyDisjointInst>(BO)->isDisjoint())
2082        return {Instruction::Add, BO0, BO1};
2083      break;
2084    }
2085    case Instruction::Sub:
2086      // sub 0, X --> mul X, -1
2087      if (match(BO0, m_ZeroInt()))
2088        return {Instruction::Mul, BO1, ConstantInt::getAllOnesValue(Ty)};
2089      break;
2090    default:
2091      break;
2092    }
2093    return {};
2094  }
2095  
2096  /// A select shuffle of a select shuffle with a shared operand can be reduced
2097  /// to a single select shuffle. This is an obvious improvement in IR, and the
2098  /// backend is expected to lower select shuffles efficiently.
foldSelectShuffleOfSelectShuffle(ShuffleVectorInst & Shuf)2099  static Instruction *foldSelectShuffleOfSelectShuffle(ShuffleVectorInst &Shuf) {
2100    assert(Shuf.isSelect() && "Must have select-equivalent shuffle");
2101  
2102    Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1);
2103    SmallVector<int, 16> Mask;
2104    Shuf.getShuffleMask(Mask);
2105    unsigned NumElts = Mask.size();
2106  
2107    // Canonicalize a select shuffle with common operand as Op1.
2108    auto *ShufOp = dyn_cast<ShuffleVectorInst>(Op0);
2109    if (ShufOp && ShufOp->isSelect() &&
2110        (ShufOp->getOperand(0) == Op1 || ShufOp->getOperand(1) == Op1)) {
2111      std::swap(Op0, Op1);
2112      ShuffleVectorInst::commuteShuffleMask(Mask, NumElts);
2113    }
2114  
2115    ShufOp = dyn_cast<ShuffleVectorInst>(Op1);
2116    if (!ShufOp || !ShufOp->isSelect() ||
2117        (ShufOp->getOperand(0) != Op0 && ShufOp->getOperand(1) != Op0))
2118      return nullptr;
2119  
2120    Value *X = ShufOp->getOperand(0), *Y = ShufOp->getOperand(1);
2121    SmallVector<int, 16> Mask1;
2122    ShufOp->getShuffleMask(Mask1);
2123    assert(Mask1.size() == NumElts && "Vector size changed with select shuffle");
2124  
2125    // Canonicalize common operand (Op0) as X (first operand of first shuffle).
2126    if (Y == Op0) {
2127      std::swap(X, Y);
2128      ShuffleVectorInst::commuteShuffleMask(Mask1, NumElts);
2129    }
2130  
2131    // If the mask chooses from X (operand 0), it stays the same.
2132    // If the mask chooses from the earlier shuffle, the other mask value is
2133    // transferred to the combined select shuffle:
2134    // shuf X, (shuf X, Y, M1), M --> shuf X, Y, M'
2135    SmallVector<int, 16> NewMask(NumElts);
2136    for (unsigned i = 0; i != NumElts; ++i)
2137      NewMask[i] = Mask[i] < (signed)NumElts ? Mask[i] : Mask1[i];
2138  
2139    // A select mask with undef elements might look like an identity mask.
2140    assert((ShuffleVectorInst::isSelectMask(NewMask, NumElts) ||
2141            ShuffleVectorInst::isIdentityMask(NewMask, NumElts)) &&
2142           "Unexpected shuffle mask");
2143    return new ShuffleVectorInst(X, Y, NewMask);
2144  }
2145  
foldSelectShuffleWith1Binop(ShuffleVectorInst & Shuf,const SimplifyQuery & SQ)2146  static Instruction *foldSelectShuffleWith1Binop(ShuffleVectorInst &Shuf,
2147                                                  const SimplifyQuery &SQ) {
2148    assert(Shuf.isSelect() && "Must have select-equivalent shuffle");
2149  
2150    // Are we shuffling together some value and that same value after it has been
2151    // modified by a binop with a constant?
2152    Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1);
2153    Constant *C;
2154    bool Op0IsBinop;
2155    if (match(Op0, m_BinOp(m_Specific(Op1), m_Constant(C))))
2156      Op0IsBinop = true;
2157    else if (match(Op1, m_BinOp(m_Specific(Op0), m_Constant(C))))
2158      Op0IsBinop = false;
2159    else
2160      return nullptr;
2161  
2162    // The identity constant for a binop leaves a variable operand unchanged. For
2163    // a vector, this is a splat of something like 0, -1, or 1.
2164    // If there's no identity constant for this binop, we're done.
2165    auto *BO = cast<BinaryOperator>(Op0IsBinop ? Op0 : Op1);
2166    BinaryOperator::BinaryOps BOpcode = BO->getOpcode();
2167    Constant *IdC = ConstantExpr::getBinOpIdentity(BOpcode, Shuf.getType(), true);
2168    if (!IdC)
2169      return nullptr;
2170  
2171    Value *X = Op0IsBinop ? Op1 : Op0;
2172  
2173    // Prevent folding in the case the non-binop operand might have NaN values.
2174    // If X can have NaN elements then we have that the floating point math
2175    // operation in the transformed code may not preserve the exact NaN
2176    // bit-pattern -- e.g. `fadd sNaN, 0.0 -> qNaN`.
2177    // This makes the transformation incorrect since the original program would
2178    // have preserved the exact NaN bit-pattern.
2179    // Avoid the folding if X can have NaN elements.
2180    if (Shuf.getType()->getElementType()->isFloatingPointTy() &&
2181        !isKnownNeverNaN(X, 0, SQ))
2182      return nullptr;
2183  
2184    // Shuffle identity constants into the lanes that return the original value.
2185    // Example: shuf (mul X, {-1,-2,-3,-4}), X, {0,5,6,3} --> mul X, {-1,1,1,-4}
2186    // Example: shuf X, (add X, {-1,-2,-3,-4}), {0,1,6,7} --> add X, {0,0,-3,-4}
2187    // The existing binop constant vector remains in the same operand position.
2188    ArrayRef<int> Mask = Shuf.getShuffleMask();
2189    Constant *NewC = Op0IsBinop ? ConstantExpr::getShuffleVector(C, IdC, Mask) :
2190                                  ConstantExpr::getShuffleVector(IdC, C, Mask);
2191  
2192    bool MightCreatePoisonOrUB =
2193        is_contained(Mask, PoisonMaskElem) &&
2194        (Instruction::isIntDivRem(BOpcode) || Instruction::isShift(BOpcode));
2195    if (MightCreatePoisonOrUB)
2196      NewC = InstCombiner::getSafeVectorConstantForBinop(BOpcode, NewC, true);
2197  
2198    // shuf (bop X, C), X, M --> bop X, C'
2199    // shuf X, (bop X, C), M --> bop X, C'
2200    Instruction *NewBO = BinaryOperator::Create(BOpcode, X, NewC);
2201    NewBO->copyIRFlags(BO);
2202  
2203    // An undef shuffle mask element may propagate as an undef constant element in
2204    // the new binop. That would produce poison where the original code might not.
2205    // If we already made a safe constant, then there's no danger.
2206    if (is_contained(Mask, PoisonMaskElem) && !MightCreatePoisonOrUB)
2207      NewBO->dropPoisonGeneratingFlags();
2208    return NewBO;
2209  }
2210  
2211  /// If we have an insert of a scalar to a non-zero element of an undefined
2212  /// vector and then shuffle that value, that's the same as inserting to the zero
2213  /// element and shuffling. Splatting from the zero element is recognized as the
2214  /// canonical form of splat.
canonicalizeInsertSplat(ShuffleVectorInst & Shuf,InstCombiner::BuilderTy & Builder)2215  static Instruction *canonicalizeInsertSplat(ShuffleVectorInst &Shuf,
2216                                              InstCombiner::BuilderTy &Builder) {
2217    Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1);
2218    ArrayRef<int> Mask = Shuf.getShuffleMask();
2219    Value *X;
2220    uint64_t IndexC;
2221  
2222    // Match a shuffle that is a splat to a non-zero element.
2223    if (!match(Op0, m_OneUse(m_InsertElt(m_Poison(), m_Value(X),
2224                                         m_ConstantInt(IndexC)))) ||
2225        !match(Op1, m_Poison()) || match(Mask, m_ZeroMask()) || IndexC == 0)
2226      return nullptr;
2227  
2228    // Insert into element 0 of a poison vector.
2229    PoisonValue *PoisonVec = PoisonValue::get(Shuf.getType());
2230    Value *NewIns = Builder.CreateInsertElement(PoisonVec, X, (uint64_t)0);
2231  
2232    // Splat from element 0. Any mask element that is poison remains poison.
2233    // For example:
2234    // shuf (inselt poison, X, 2), _, <2,2,undef>
2235    //   --> shuf (inselt poison, X, 0), poison, <0,0,undef>
2236    unsigned NumMaskElts =
2237        cast<FixedVectorType>(Shuf.getType())->getNumElements();
2238    SmallVector<int, 16> NewMask(NumMaskElts, 0);
2239    for (unsigned i = 0; i != NumMaskElts; ++i)
2240      if (Mask[i] == PoisonMaskElem)
2241        NewMask[i] = Mask[i];
2242  
2243    return new ShuffleVectorInst(NewIns, NewMask);
2244  }
2245  
2246  /// Try to fold shuffles that are the equivalent of a vector select.
foldSelectShuffle(ShuffleVectorInst & Shuf)2247  Instruction *InstCombinerImpl::foldSelectShuffle(ShuffleVectorInst &Shuf) {
2248    if (!Shuf.isSelect())
2249      return nullptr;
2250  
2251    // Canonicalize to choose from operand 0 first unless operand 1 is undefined.
2252    // Commuting undef to operand 0 conflicts with another canonicalization.
2253    unsigned NumElts = cast<FixedVectorType>(Shuf.getType())->getNumElements();
2254    if (!match(Shuf.getOperand(1), m_Undef()) &&
2255        Shuf.getMaskValue(0) >= (int)NumElts) {
2256      // TODO: Can we assert that both operands of a shuffle-select are not undef
2257      // (otherwise, it would have been folded by instsimplify?
2258      Shuf.commute();
2259      return &Shuf;
2260    }
2261  
2262    if (Instruction *I = foldSelectShuffleOfSelectShuffle(Shuf))
2263      return I;
2264  
2265    if (Instruction *I = foldSelectShuffleWith1Binop(
2266            Shuf, getSimplifyQuery().getWithInstruction(&Shuf)))
2267      return I;
2268  
2269    BinaryOperator *B0, *B1;
2270    if (!match(Shuf.getOperand(0), m_BinOp(B0)) ||
2271        !match(Shuf.getOperand(1), m_BinOp(B1)))
2272      return nullptr;
2273  
2274    // If one operand is "0 - X", allow that to be viewed as "X * -1"
2275    // (ConstantsAreOp1) by getAlternateBinop below. If the neg is not paired
2276    // with a multiply, we will exit because C0/C1 will not be set.
2277    Value *X, *Y;
2278    Constant *C0 = nullptr, *C1 = nullptr;
2279    bool ConstantsAreOp1;
2280    if (match(B0, m_BinOp(m_Constant(C0), m_Value(X))) &&
2281        match(B1, m_BinOp(m_Constant(C1), m_Value(Y))))
2282      ConstantsAreOp1 = false;
2283    else if (match(B0, m_CombineOr(m_BinOp(m_Value(X), m_Constant(C0)),
2284                                   m_Neg(m_Value(X)))) &&
2285             match(B1, m_CombineOr(m_BinOp(m_Value(Y), m_Constant(C1)),
2286                                   m_Neg(m_Value(Y)))))
2287      ConstantsAreOp1 = true;
2288    else
2289      return nullptr;
2290  
2291    // We need matching binops to fold the lanes together.
2292    BinaryOperator::BinaryOps Opc0 = B0->getOpcode();
2293    BinaryOperator::BinaryOps Opc1 = B1->getOpcode();
2294    bool DropNSW = false;
2295    if (ConstantsAreOp1 && Opc0 != Opc1) {
2296      // TODO: We drop "nsw" if shift is converted into multiply because it may
2297      // not be correct when the shift amount is BitWidth - 1. We could examine
2298      // each vector element to determine if it is safe to keep that flag.
2299      if (Opc0 == Instruction::Shl || Opc1 == Instruction::Shl)
2300        DropNSW = true;
2301      if (BinopElts AltB0 = getAlternateBinop(B0, DL)) {
2302        assert(isa<Constant>(AltB0.Op1) && "Expecting constant with alt binop");
2303        Opc0 = AltB0.Opcode;
2304        C0 = cast<Constant>(AltB0.Op1);
2305      } else if (BinopElts AltB1 = getAlternateBinop(B1, DL)) {
2306        assert(isa<Constant>(AltB1.Op1) && "Expecting constant with alt binop");
2307        Opc1 = AltB1.Opcode;
2308        C1 = cast<Constant>(AltB1.Op1);
2309      }
2310    }
2311  
2312    if (Opc0 != Opc1 || !C0 || !C1)
2313      return nullptr;
2314  
2315    // The opcodes must be the same. Use a new name to make that clear.
2316    BinaryOperator::BinaryOps BOpc = Opc0;
2317  
2318    // Select the constant elements needed for the single binop.
2319    ArrayRef<int> Mask = Shuf.getShuffleMask();
2320    Constant *NewC = ConstantExpr::getShuffleVector(C0, C1, Mask);
2321  
2322    // We are moving a binop after a shuffle. When a shuffle has an undefined
2323    // mask element, the result is undefined, but it is not poison or undefined
2324    // behavior. That is not necessarily true for div/rem/shift.
2325    bool MightCreatePoisonOrUB =
2326        is_contained(Mask, PoisonMaskElem) &&
2327        (Instruction::isIntDivRem(BOpc) || Instruction::isShift(BOpc));
2328    if (MightCreatePoisonOrUB)
2329      NewC = InstCombiner::getSafeVectorConstantForBinop(BOpc, NewC,
2330                                                         ConstantsAreOp1);
2331  
2332    Value *V;
2333    if (X == Y) {
2334      // Remove a binop and the shuffle by rearranging the constant:
2335      // shuffle (op V, C0), (op V, C1), M --> op V, C'
2336      // shuffle (op C0, V), (op C1, V), M --> op C', V
2337      V = X;
2338    } else {
2339      // If there are 2 different variable operands, we must create a new shuffle
2340      // (select) first, so check uses to ensure that we don't end up with more
2341      // instructions than we started with.
2342      if (!B0->hasOneUse() && !B1->hasOneUse())
2343        return nullptr;
2344  
2345      // If we use the original shuffle mask and op1 is *variable*, we would be
2346      // putting an undef into operand 1 of div/rem/shift. This is either UB or
2347      // poison. We do not have to guard against UB when *constants* are op1
2348      // because safe constants guarantee that we do not overflow sdiv/srem (and
2349      // there's no danger for other opcodes).
2350      // TODO: To allow this case, create a new shuffle mask with no undefs.
2351      if (MightCreatePoisonOrUB && !ConstantsAreOp1)
2352        return nullptr;
2353  
2354      // Note: In general, we do not create new shuffles in InstCombine because we
2355      // do not know if a target can lower an arbitrary shuffle optimally. In this
2356      // case, the shuffle uses the existing mask, so there is no additional risk.
2357  
2358      // Select the variable vectors first, then perform the binop:
2359      // shuffle (op X, C0), (op Y, C1), M --> op (shuffle X, Y, M), C'
2360      // shuffle (op C0, X), (op C1, Y), M --> op C', (shuffle X, Y, M)
2361      V = Builder.CreateShuffleVector(X, Y, Mask);
2362    }
2363  
2364    Value *NewBO = ConstantsAreOp1 ? Builder.CreateBinOp(BOpc, V, NewC) :
2365                                     Builder.CreateBinOp(BOpc, NewC, V);
2366  
2367    // Flags are intersected from the 2 source binops. But there are 2 exceptions:
2368    // 1. If we changed an opcode, poison conditions might have changed.
2369    // 2. If the shuffle had undef mask elements, the new binop might have undefs
2370    //    where the original code did not. But if we already made a safe constant,
2371    //    then there's no danger.
2372    if (auto *NewI = dyn_cast<Instruction>(NewBO)) {
2373      NewI->copyIRFlags(B0);
2374      NewI->andIRFlags(B1);
2375      if (DropNSW)
2376        NewI->setHasNoSignedWrap(false);
2377      if (is_contained(Mask, PoisonMaskElem) && !MightCreatePoisonOrUB)
2378        NewI->dropPoisonGeneratingFlags();
2379    }
2380    return replaceInstUsesWith(Shuf, NewBO);
2381  }
2382  
2383  /// Convert a narrowing shuffle of a bitcasted vector into a vector truncate.
2384  /// Example (little endian):
2385  /// shuf (bitcast <4 x i16> X to <8 x i8>), <0, 2, 4, 6> --> trunc X to <4 x i8>
foldTruncShuffle(ShuffleVectorInst & Shuf,bool IsBigEndian)2386  static Instruction *foldTruncShuffle(ShuffleVectorInst &Shuf,
2387                                       bool IsBigEndian) {
2388    // This must be a bitcasted shuffle of 1 vector integer operand.
2389    Type *DestType = Shuf.getType();
2390    Value *X;
2391    if (!match(Shuf.getOperand(0), m_BitCast(m_Value(X))) ||
2392        !match(Shuf.getOperand(1), m_Poison()) || !DestType->isIntOrIntVectorTy())
2393      return nullptr;
2394  
2395    // The source type must have the same number of elements as the shuffle,
2396    // and the source element type must be larger than the shuffle element type.
2397    Type *SrcType = X->getType();
2398    if (!SrcType->isVectorTy() || !SrcType->isIntOrIntVectorTy() ||
2399        cast<FixedVectorType>(SrcType)->getNumElements() !=
2400            cast<FixedVectorType>(DestType)->getNumElements() ||
2401        SrcType->getScalarSizeInBits() % DestType->getScalarSizeInBits() != 0)
2402      return nullptr;
2403  
2404    assert(Shuf.changesLength() && !Shuf.increasesLength() &&
2405           "Expected a shuffle that decreases length");
2406  
2407    // Last, check that the mask chooses the correct low bits for each narrow
2408    // element in the result.
2409    uint64_t TruncRatio =
2410        SrcType->getScalarSizeInBits() / DestType->getScalarSizeInBits();
2411    ArrayRef<int> Mask = Shuf.getShuffleMask();
2412    for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
2413      if (Mask[i] == PoisonMaskElem)
2414        continue;
2415      uint64_t LSBIndex = IsBigEndian ? (i + 1) * TruncRatio - 1 : i * TruncRatio;
2416      assert(LSBIndex <= INT32_MAX && "Overflowed 32-bits");
2417      if (Mask[i] != (int)LSBIndex)
2418        return nullptr;
2419    }
2420  
2421    return new TruncInst(X, DestType);
2422  }
2423  
2424  /// Match a shuffle-select-shuffle pattern where the shuffles are widening and
2425  /// narrowing (concatenating with poison and extracting back to the original
2426  /// length). This allows replacing the wide select with a narrow select.
narrowVectorSelect(ShuffleVectorInst & Shuf,InstCombiner::BuilderTy & Builder)2427  static Instruction *narrowVectorSelect(ShuffleVectorInst &Shuf,
2428                                         InstCombiner::BuilderTy &Builder) {
2429    // This must be a narrowing identity shuffle. It extracts the 1st N elements
2430    // of the 1st vector operand of a shuffle.
2431    if (!match(Shuf.getOperand(1), m_Poison()) || !Shuf.isIdentityWithExtract())
2432      return nullptr;
2433  
2434    // The vector being shuffled must be a vector select that we can eliminate.
2435    // TODO: The one-use requirement could be eased if X and/or Y are constants.
2436    Value *Cond, *X, *Y;
2437    if (!match(Shuf.getOperand(0),
2438               m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y)))))
2439      return nullptr;
2440  
2441    // We need a narrow condition value. It must be extended with poison elements
2442    // and have the same number of elements as this shuffle.
2443    unsigned NarrowNumElts =
2444        cast<FixedVectorType>(Shuf.getType())->getNumElements();
2445    Value *NarrowCond;
2446    if (!match(Cond, m_OneUse(m_Shuffle(m_Value(NarrowCond), m_Poison()))) ||
2447        cast<FixedVectorType>(NarrowCond->getType())->getNumElements() !=
2448            NarrowNumElts ||
2449        !cast<ShuffleVectorInst>(Cond)->isIdentityWithPadding())
2450      return nullptr;
2451  
2452    // shuf (sel (shuf NarrowCond, poison, WideMask), X, Y), poison, NarrowMask)
2453    // -->
2454    // sel NarrowCond, (shuf X, poison, NarrowMask), (shuf Y, poison, NarrowMask)
2455    Value *NarrowX = Builder.CreateShuffleVector(X, Shuf.getShuffleMask());
2456    Value *NarrowY = Builder.CreateShuffleVector(Y, Shuf.getShuffleMask());
2457    return SelectInst::Create(NarrowCond, NarrowX, NarrowY);
2458  }
2459  
2460  /// Canonicalize FP negate/abs after shuffle.
foldShuffleOfUnaryOps(ShuffleVectorInst & Shuf,InstCombiner::BuilderTy & Builder)2461  static Instruction *foldShuffleOfUnaryOps(ShuffleVectorInst &Shuf,
2462                                            InstCombiner::BuilderTy &Builder) {
2463    auto *S0 = dyn_cast<Instruction>(Shuf.getOperand(0));
2464    Value *X;
2465    if (!S0 || !match(S0, m_CombineOr(m_FNeg(m_Value(X)), m_FAbs(m_Value(X)))))
2466      return nullptr;
2467  
2468    bool IsFNeg = S0->getOpcode() == Instruction::FNeg;
2469  
2470    // Match 1-input (unary) shuffle.
2471    // shuffle (fneg/fabs X), Mask --> fneg/fabs (shuffle X, Mask)
2472    if (S0->hasOneUse() && match(Shuf.getOperand(1), m_Poison())) {
2473      Value *NewShuf = Builder.CreateShuffleVector(X, Shuf.getShuffleMask());
2474      if (IsFNeg)
2475        return UnaryOperator::CreateFNegFMF(NewShuf, S0);
2476  
2477      Function *FAbs = Intrinsic::getDeclaration(Shuf.getModule(),
2478                                                 Intrinsic::fabs, Shuf.getType());
2479      CallInst *NewF = CallInst::Create(FAbs, {NewShuf});
2480      NewF->setFastMathFlags(S0->getFastMathFlags());
2481      return NewF;
2482    }
2483  
2484    // Match 2-input (binary) shuffle.
2485    auto *S1 = dyn_cast<Instruction>(Shuf.getOperand(1));
2486    Value *Y;
2487    if (!S1 || !match(S1, m_CombineOr(m_FNeg(m_Value(Y)), m_FAbs(m_Value(Y)))) ||
2488        S0->getOpcode() != S1->getOpcode() ||
2489        (!S0->hasOneUse() && !S1->hasOneUse()))
2490      return nullptr;
2491  
2492    // shuf (fneg/fabs X), (fneg/fabs Y), Mask --> fneg/fabs (shuf X, Y, Mask)
2493    Value *NewShuf = Builder.CreateShuffleVector(X, Y, Shuf.getShuffleMask());
2494    Instruction *NewF;
2495    if (IsFNeg) {
2496      NewF = UnaryOperator::CreateFNeg(NewShuf);
2497    } else {
2498      Function *FAbs = Intrinsic::getDeclaration(Shuf.getModule(),
2499                                                 Intrinsic::fabs, Shuf.getType());
2500      NewF = CallInst::Create(FAbs, {NewShuf});
2501    }
2502    NewF->copyIRFlags(S0);
2503    NewF->andIRFlags(S1);
2504    return NewF;
2505  }
2506  
2507  /// Canonicalize casts after shuffle.
foldCastShuffle(ShuffleVectorInst & Shuf,InstCombiner::BuilderTy & Builder)2508  static Instruction *foldCastShuffle(ShuffleVectorInst &Shuf,
2509                                      InstCombiner::BuilderTy &Builder) {
2510    // Do we have 2 matching cast operands?
2511    auto *Cast0 = dyn_cast<CastInst>(Shuf.getOperand(0));
2512    auto *Cast1 = dyn_cast<CastInst>(Shuf.getOperand(1));
2513    if (!Cast0 || !Cast1 || Cast0->getOpcode() != Cast1->getOpcode() ||
2514        Cast0->getSrcTy() != Cast1->getSrcTy())
2515      return nullptr;
2516  
2517    // TODO: Allow other opcodes? That would require easing the type restrictions
2518    //       below here.
2519    CastInst::CastOps CastOpcode = Cast0->getOpcode();
2520    switch (CastOpcode) {
2521    case Instruction::FPToSI:
2522    case Instruction::FPToUI:
2523    case Instruction::SIToFP:
2524    case Instruction::UIToFP:
2525      break;
2526    default:
2527      return nullptr;
2528    }
2529  
2530    VectorType *ShufTy = Shuf.getType();
2531    VectorType *ShufOpTy = cast<VectorType>(Shuf.getOperand(0)->getType());
2532    VectorType *CastSrcTy = cast<VectorType>(Cast0->getSrcTy());
2533  
2534    // TODO: Allow length-increasing shuffles?
2535    if (ShufTy->getElementCount().getKnownMinValue() >
2536        ShufOpTy->getElementCount().getKnownMinValue())
2537      return nullptr;
2538  
2539    // TODO: Allow element-size-decreasing casts (ex: fptosi float to i8)?
2540    assert(isa<FixedVectorType>(CastSrcTy) && isa<FixedVectorType>(ShufOpTy) &&
2541           "Expected fixed vector operands for casts and binary shuffle");
2542    if (CastSrcTy->getPrimitiveSizeInBits() > ShufOpTy->getPrimitiveSizeInBits())
2543      return nullptr;
2544  
2545    // At least one of the operands must have only one use (the shuffle).
2546    if (!Cast0->hasOneUse() && !Cast1->hasOneUse())
2547      return nullptr;
2548  
2549    // shuffle (cast X), (cast Y), Mask --> cast (shuffle X, Y, Mask)
2550    Value *X = Cast0->getOperand(0);
2551    Value *Y = Cast1->getOperand(0);
2552    Value *NewShuf = Builder.CreateShuffleVector(X, Y, Shuf.getShuffleMask());
2553    return CastInst::Create(CastOpcode, NewShuf, ShufTy);
2554  }
2555  
2556  /// Try to fold an extract subvector operation.
foldIdentityExtractShuffle(ShuffleVectorInst & Shuf)2557  static Instruction *foldIdentityExtractShuffle(ShuffleVectorInst &Shuf) {
2558    Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1);
2559    if (!Shuf.isIdentityWithExtract() || !match(Op1, m_Poison()))
2560      return nullptr;
2561  
2562    // Check if we are extracting all bits of an inserted scalar:
2563    // extract-subvec (bitcast (inselt ?, X, 0) --> bitcast X to subvec type
2564    Value *X;
2565    if (match(Op0, m_BitCast(m_InsertElt(m_Value(), m_Value(X), m_Zero()))) &&
2566        X->getType()->getPrimitiveSizeInBits() ==
2567            Shuf.getType()->getPrimitiveSizeInBits())
2568      return new BitCastInst(X, Shuf.getType());
2569  
2570    // Try to combine 2 shuffles into 1 shuffle by concatenating a shuffle mask.
2571    Value *Y;
2572    ArrayRef<int> Mask;
2573    if (!match(Op0, m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask))))
2574      return nullptr;
2575  
2576    // Be conservative with shuffle transforms. If we can't kill the 1st shuffle,
2577    // then combining may result in worse codegen.
2578    if (!Op0->hasOneUse())
2579      return nullptr;
2580  
2581    // We are extracting a subvector from a shuffle. Remove excess elements from
2582    // the 1st shuffle mask to eliminate the extract.
2583    //
2584    // This transform is conservatively limited to identity extracts because we do
2585    // not allow arbitrary shuffle mask creation as a target-independent transform
2586    // (because we can't guarantee that will lower efficiently).
2587    //
2588    // If the extracting shuffle has an poison mask element, it transfers to the
2589    // new shuffle mask. Otherwise, copy the original mask element. Example:
2590    //   shuf (shuf X, Y, <C0, C1, C2, poison, C4>), poison, <0, poison, 2, 3> -->
2591    //   shuf X, Y, <C0, poison, C2, poison>
2592    unsigned NumElts = cast<FixedVectorType>(Shuf.getType())->getNumElements();
2593    SmallVector<int, 16> NewMask(NumElts);
2594    assert(NumElts < Mask.size() &&
2595           "Identity with extract must have less elements than its inputs");
2596  
2597    for (unsigned i = 0; i != NumElts; ++i) {
2598      int ExtractMaskElt = Shuf.getMaskValue(i);
2599      int MaskElt = Mask[i];
2600      NewMask[i] = ExtractMaskElt == PoisonMaskElem ? ExtractMaskElt : MaskElt;
2601    }
2602    return new ShuffleVectorInst(X, Y, NewMask);
2603  }
2604  
2605  /// Try to replace a shuffle with an insertelement or try to replace a shuffle
2606  /// operand with the operand of an insertelement.
foldShuffleWithInsert(ShuffleVectorInst & Shuf,InstCombinerImpl & IC)2607  static Instruction *foldShuffleWithInsert(ShuffleVectorInst &Shuf,
2608                                            InstCombinerImpl &IC) {
2609    Value *V0 = Shuf.getOperand(0), *V1 = Shuf.getOperand(1);
2610    SmallVector<int, 16> Mask;
2611    Shuf.getShuffleMask(Mask);
2612  
2613    int NumElts = Mask.size();
2614    int InpNumElts = cast<FixedVectorType>(V0->getType())->getNumElements();
2615  
2616    // This is a specialization of a fold in SimplifyDemandedVectorElts. We may
2617    // not be able to handle it there if the insertelement has >1 use.
2618    // If the shuffle has an insertelement operand but does not choose the
2619    // inserted scalar element from that value, then we can replace that shuffle
2620    // operand with the source vector of the insertelement.
2621    Value *X;
2622    uint64_t IdxC;
2623    if (match(V0, m_InsertElt(m_Value(X), m_Value(), m_ConstantInt(IdxC)))) {
2624      // shuf (inselt X, ?, IdxC), ?, Mask --> shuf X, ?, Mask
2625      if (!is_contained(Mask, (int)IdxC))
2626        return IC.replaceOperand(Shuf, 0, X);
2627    }
2628    if (match(V1, m_InsertElt(m_Value(X), m_Value(), m_ConstantInt(IdxC)))) {
2629      // Offset the index constant by the vector width because we are checking for
2630      // accesses to the 2nd vector input of the shuffle.
2631      IdxC += InpNumElts;
2632      // shuf ?, (inselt X, ?, IdxC), Mask --> shuf ?, X, Mask
2633      if (!is_contained(Mask, (int)IdxC))
2634        return IC.replaceOperand(Shuf, 1, X);
2635    }
2636    // For the rest of the transform, the shuffle must not change vector sizes.
2637    // TODO: This restriction could be removed if the insert has only one use
2638    //       (because the transform would require a new length-changing shuffle).
2639    if (NumElts != InpNumElts)
2640      return nullptr;
2641  
2642    // shuffle (insert ?, Scalar, IndexC), V1, Mask --> insert V1, Scalar, IndexC'
2643    auto isShufflingScalarIntoOp1 = [&](Value *&Scalar, ConstantInt *&IndexC) {
2644      // We need an insertelement with a constant index.
2645      if (!match(V0, m_InsertElt(m_Value(), m_Value(Scalar),
2646                                 m_ConstantInt(IndexC))))
2647        return false;
2648  
2649      // Test the shuffle mask to see if it splices the inserted scalar into the
2650      // operand 1 vector of the shuffle.
2651      int NewInsIndex = -1;
2652      for (int i = 0; i != NumElts; ++i) {
2653        // Ignore undef mask elements.
2654        if (Mask[i] == -1)
2655          continue;
2656  
2657        // The shuffle takes elements of operand 1 without lane changes.
2658        if (Mask[i] == NumElts + i)
2659          continue;
2660  
2661        // The shuffle must choose the inserted scalar exactly once.
2662        if (NewInsIndex != -1 || Mask[i] != IndexC->getSExtValue())
2663          return false;
2664  
2665        // The shuffle is placing the inserted scalar into element i.
2666        NewInsIndex = i;
2667      }
2668  
2669      assert(NewInsIndex != -1 && "Did not fold shuffle with unused operand?");
2670  
2671      // Index is updated to the potentially translated insertion lane.
2672      IndexC = ConstantInt::get(IndexC->getIntegerType(), NewInsIndex);
2673      return true;
2674    };
2675  
2676    // If the shuffle is unnecessary, insert the scalar operand directly into
2677    // operand 1 of the shuffle. Example:
2678    // shuffle (insert ?, S, 1), V1, <1, 5, 6, 7> --> insert V1, S, 0
2679    Value *Scalar;
2680    ConstantInt *IndexC;
2681    if (isShufflingScalarIntoOp1(Scalar, IndexC))
2682      return InsertElementInst::Create(V1, Scalar, IndexC);
2683  
2684    // Try again after commuting shuffle. Example:
2685    // shuffle V0, (insert ?, S, 0), <0, 1, 2, 4> -->
2686    // shuffle (insert ?, S, 0), V0, <4, 5, 6, 0> --> insert V0, S, 3
2687    std::swap(V0, V1);
2688    ShuffleVectorInst::commuteShuffleMask(Mask, NumElts);
2689    if (isShufflingScalarIntoOp1(Scalar, IndexC))
2690      return InsertElementInst::Create(V1, Scalar, IndexC);
2691  
2692    return nullptr;
2693  }
2694  
foldIdentityPaddedShuffles(ShuffleVectorInst & Shuf)2695  static Instruction *foldIdentityPaddedShuffles(ShuffleVectorInst &Shuf) {
2696    // Match the operands as identity with padding (also known as concatenation
2697    // with undef) shuffles of the same source type. The backend is expected to
2698    // recreate these concatenations from a shuffle of narrow operands.
2699    auto *Shuffle0 = dyn_cast<ShuffleVectorInst>(Shuf.getOperand(0));
2700    auto *Shuffle1 = dyn_cast<ShuffleVectorInst>(Shuf.getOperand(1));
2701    if (!Shuffle0 || !Shuffle0->isIdentityWithPadding() ||
2702        !Shuffle1 || !Shuffle1->isIdentityWithPadding())
2703      return nullptr;
2704  
2705    // We limit this transform to power-of-2 types because we expect that the
2706    // backend can convert the simplified IR patterns to identical nodes as the
2707    // original IR.
2708    // TODO: If we can verify the same behavior for arbitrary types, the
2709    //       power-of-2 checks can be removed.
2710    Value *X = Shuffle0->getOperand(0);
2711    Value *Y = Shuffle1->getOperand(0);
2712    if (X->getType() != Y->getType() ||
2713        !isPowerOf2_32(cast<FixedVectorType>(Shuf.getType())->getNumElements()) ||
2714        !isPowerOf2_32(
2715            cast<FixedVectorType>(Shuffle0->getType())->getNumElements()) ||
2716        !isPowerOf2_32(cast<FixedVectorType>(X->getType())->getNumElements()) ||
2717        match(X, m_Undef()) || match(Y, m_Undef()))
2718      return nullptr;
2719    assert(match(Shuffle0->getOperand(1), m_Undef()) &&
2720           match(Shuffle1->getOperand(1), m_Undef()) &&
2721           "Unexpected operand for identity shuffle");
2722  
2723    // This is a shuffle of 2 widening shuffles. We can shuffle the narrow source
2724    // operands directly by adjusting the shuffle mask to account for the narrower
2725    // types:
2726    // shuf (widen X), (widen Y), Mask --> shuf X, Y, Mask'
2727    int NarrowElts = cast<FixedVectorType>(X->getType())->getNumElements();
2728    int WideElts = cast<FixedVectorType>(Shuffle0->getType())->getNumElements();
2729    assert(WideElts > NarrowElts && "Unexpected types for identity with padding");
2730  
2731    ArrayRef<int> Mask = Shuf.getShuffleMask();
2732    SmallVector<int, 16> NewMask(Mask.size(), -1);
2733    for (int i = 0, e = Mask.size(); i != e; ++i) {
2734      if (Mask[i] == -1)
2735        continue;
2736  
2737      // If this shuffle is choosing an undef element from 1 of the sources, that
2738      // element is undef.
2739      if (Mask[i] < WideElts) {
2740        if (Shuffle0->getMaskValue(Mask[i]) == -1)
2741          continue;
2742      } else {
2743        if (Shuffle1->getMaskValue(Mask[i] - WideElts) == -1)
2744          continue;
2745      }
2746  
2747      // If this shuffle is choosing from the 1st narrow op, the mask element is
2748      // the same. If this shuffle is choosing from the 2nd narrow op, the mask
2749      // element is offset down to adjust for the narrow vector widths.
2750      if (Mask[i] < WideElts) {
2751        assert(Mask[i] < NarrowElts && "Unexpected shuffle mask");
2752        NewMask[i] = Mask[i];
2753      } else {
2754        assert(Mask[i] < (WideElts + NarrowElts) && "Unexpected shuffle mask");
2755        NewMask[i] = Mask[i] - (WideElts - NarrowElts);
2756      }
2757    }
2758    return new ShuffleVectorInst(X, Y, NewMask);
2759  }
2760  
2761  // Splatting the first element of the result of a BinOp, where any of the
2762  // BinOp's operands are the result of a first element splat can be simplified to
2763  // splatting the first element of the result of the BinOp
simplifyBinOpSplats(ShuffleVectorInst & SVI)2764  Instruction *InstCombinerImpl::simplifyBinOpSplats(ShuffleVectorInst &SVI) {
2765    if (!match(SVI.getOperand(1), m_Poison()) ||
2766        !match(SVI.getShuffleMask(), m_ZeroMask()) ||
2767        !SVI.getOperand(0)->hasOneUse())
2768      return nullptr;
2769  
2770    Value *Op0 = SVI.getOperand(0);
2771    Value *X, *Y;
2772    if (!match(Op0, m_BinOp(m_Shuffle(m_Value(X), m_Poison(), m_ZeroMask()),
2773                            m_Value(Y))) &&
2774        !match(Op0, m_BinOp(m_Value(X),
2775                            m_Shuffle(m_Value(Y), m_Poison(), m_ZeroMask()))))
2776      return nullptr;
2777    if (X->getType() != Y->getType())
2778      return nullptr;
2779  
2780    auto *BinOp = cast<BinaryOperator>(Op0);
2781    if (!isSafeToSpeculativelyExecute(BinOp))
2782      return nullptr;
2783  
2784    Value *NewBO = Builder.CreateBinOp(BinOp->getOpcode(), X, Y);
2785    if (auto NewBOI = dyn_cast<Instruction>(NewBO))
2786      NewBOI->copyIRFlags(BinOp);
2787  
2788    return new ShuffleVectorInst(NewBO, SVI.getShuffleMask());
2789  }
2790  
visitShuffleVectorInst(ShuffleVectorInst & SVI)2791  Instruction *InstCombinerImpl::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
2792    Value *LHS = SVI.getOperand(0);
2793    Value *RHS = SVI.getOperand(1);
2794    SimplifyQuery ShufQuery = SQ.getWithInstruction(&SVI);
2795    if (auto *V = simplifyShuffleVectorInst(LHS, RHS, SVI.getShuffleMask(),
2796                                            SVI.getType(), ShufQuery))
2797      return replaceInstUsesWith(SVI, V);
2798  
2799    if (Instruction *I = simplifyBinOpSplats(SVI))
2800      return I;
2801  
2802    // Canonicalize splat shuffle to use poison RHS. Handle this explicitly in
2803    // order to support scalable vectors.
2804    if (match(SVI.getShuffleMask(), m_ZeroMask()) && !isa<PoisonValue>(RHS))
2805      return replaceOperand(SVI, 1, PoisonValue::get(RHS->getType()));
2806  
2807    if (isa<ScalableVectorType>(LHS->getType()))
2808      return nullptr;
2809  
2810    unsigned VWidth = cast<FixedVectorType>(SVI.getType())->getNumElements();
2811    unsigned LHSWidth = cast<FixedVectorType>(LHS->getType())->getNumElements();
2812  
2813    // shuffle (bitcast X), (bitcast Y), Mask --> bitcast (shuffle X, Y, Mask)
2814    //
2815    // if X and Y are of the same (vector) type, and the element size is not
2816    // changed by the bitcasts, we can distribute the bitcasts through the
2817    // shuffle, hopefully reducing the number of instructions. We make sure that
2818    // at least one bitcast only has one use, so we don't *increase* the number of
2819    // instructions here.
2820    Value *X, *Y;
2821    if (match(LHS, m_BitCast(m_Value(X))) && match(RHS, m_BitCast(m_Value(Y))) &&
2822        X->getType()->isVectorTy() && X->getType() == Y->getType() &&
2823        X->getType()->getScalarSizeInBits() ==
2824            SVI.getType()->getScalarSizeInBits() &&
2825        (LHS->hasOneUse() || RHS->hasOneUse())) {
2826      Value *V = Builder.CreateShuffleVector(X, Y, SVI.getShuffleMask(),
2827                                             SVI.getName() + ".uncasted");
2828      return new BitCastInst(V, SVI.getType());
2829    }
2830  
2831    ArrayRef<int> Mask = SVI.getShuffleMask();
2832  
2833    // Peek through a bitcasted shuffle operand by scaling the mask. If the
2834    // simulated shuffle can simplify, then this shuffle is unnecessary:
2835    // shuf (bitcast X), undef, Mask --> bitcast X'
2836    // TODO: This could be extended to allow length-changing shuffles.
2837    //       The transform might also be obsoleted if we allowed canonicalization
2838    //       of bitcasted shuffles.
2839    if (match(LHS, m_BitCast(m_Value(X))) && match(RHS, m_Undef()) &&
2840        X->getType()->isVectorTy() && VWidth == LHSWidth) {
2841      // Try to create a scaled mask constant.
2842      auto *XType = cast<FixedVectorType>(X->getType());
2843      unsigned XNumElts = XType->getNumElements();
2844      SmallVector<int, 16> ScaledMask;
2845      if (scaleShuffleMaskElts(XNumElts, Mask, ScaledMask)) {
2846        // If the shuffled source vector simplifies, cast that value to this
2847        // shuffle's type.
2848        if (auto *V = simplifyShuffleVectorInst(X, UndefValue::get(XType),
2849                                                ScaledMask, XType, ShufQuery))
2850          return BitCastInst::Create(Instruction::BitCast, V, SVI.getType());
2851      }
2852    }
2853  
2854    // shuffle x, x, mask --> shuffle x, undef, mask'
2855    if (LHS == RHS) {
2856      assert(!match(RHS, m_Undef()) &&
2857             "Shuffle with 2 undef ops not simplified?");
2858      return new ShuffleVectorInst(LHS, createUnaryMask(Mask, LHSWidth));
2859    }
2860  
2861    // shuffle undef, x, mask --> shuffle x, undef, mask'
2862    if (match(LHS, m_Undef())) {
2863      SVI.commute();
2864      return &SVI;
2865    }
2866  
2867    if (Instruction *I = canonicalizeInsertSplat(SVI, Builder))
2868      return I;
2869  
2870    if (Instruction *I = foldSelectShuffle(SVI))
2871      return I;
2872  
2873    if (Instruction *I = foldTruncShuffle(SVI, DL.isBigEndian()))
2874      return I;
2875  
2876    if (Instruction *I = narrowVectorSelect(SVI, Builder))
2877      return I;
2878  
2879    if (Instruction *I = foldShuffleOfUnaryOps(SVI, Builder))
2880      return I;
2881  
2882    if (Instruction *I = foldCastShuffle(SVI, Builder))
2883      return I;
2884  
2885    APInt PoisonElts(VWidth, 0);
2886    APInt AllOnesEltMask(APInt::getAllOnes(VWidth));
2887    if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, PoisonElts)) {
2888      if (V != &SVI)
2889        return replaceInstUsesWith(SVI, V);
2890      return &SVI;
2891    }
2892  
2893    if (Instruction *I = foldIdentityExtractShuffle(SVI))
2894      return I;
2895  
2896    // These transforms have the potential to lose undef knowledge, so they are
2897    // intentionally placed after SimplifyDemandedVectorElts().
2898    if (Instruction *I = foldShuffleWithInsert(SVI, *this))
2899      return I;
2900    if (Instruction *I = foldIdentityPaddedShuffles(SVI))
2901      return I;
2902  
2903    if (match(RHS, m_Poison()) && canEvaluateShuffled(LHS, Mask)) {
2904      Value *V = evaluateInDifferentElementOrder(LHS, Mask, Builder);
2905      return replaceInstUsesWith(SVI, V);
2906    }
2907  
2908    // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
2909    // a non-vector type. We can instead bitcast the original vector followed by
2910    // an extract of the desired element:
2911    //
2912    //   %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
2913    //                         <4 x i32> <i32 0, i32 1, i32 2, i32 3>
2914    //   %1 = bitcast <4 x i8> %sroa to i32
2915    // Becomes:
2916    //   %bc = bitcast <16 x i8> %in to <4 x i32>
2917    //   %ext = extractelement <4 x i32> %bc, i32 0
2918    //
2919    // If the shuffle is extracting a contiguous range of values from the input
2920    // vector then each use which is a bitcast of the extracted size can be
2921    // replaced. This will work if the vector types are compatible, and the begin
2922    // index is aligned to a value in the casted vector type. If the begin index
2923    // isn't aligned then we can shuffle the original vector (keeping the same
2924    // vector type) before extracting.
2925    //
2926    // This code will bail out if the target type is fundamentally incompatible
2927    // with vectors of the source type.
2928    //
2929    // Example of <16 x i8>, target type i32:
2930    // Index range [4,8):         v-----------v Will work.
2931    //                +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
2932    //     <16 x i8>: |  |  |  |  |  |  |  |  |  |  |  |  |  |  |  |  |
2933    //     <4 x i32>: |           |           |           |           |
2934    //                +-----------+-----------+-----------+-----------+
2935    // Index range [6,10):              ^-----------^ Needs an extra shuffle.
2936    // Target type i40:           ^--------------^ Won't work, bail.
2937    bool MadeChange = false;
2938    if (isShuffleExtractingFromLHS(SVI, Mask)) {
2939      Value *V = LHS;
2940      unsigned MaskElems = Mask.size();
2941      auto *SrcTy = cast<FixedVectorType>(V->getType());
2942      unsigned VecBitWidth = SrcTy->getPrimitiveSizeInBits().getFixedValue();
2943      unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
2944      assert(SrcElemBitWidth && "vector elements must have a bitwidth");
2945      unsigned SrcNumElems = SrcTy->getNumElements();
2946      SmallVector<BitCastInst *, 8> BCs;
2947      DenseMap<Type *, Value *> NewBCs;
2948      for (User *U : SVI.users())
2949        if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
2950          if (!BC->use_empty())
2951            // Only visit bitcasts that weren't previously handled.
2952            BCs.push_back(BC);
2953      for (BitCastInst *BC : BCs) {
2954        unsigned BegIdx = Mask.front();
2955        Type *TgtTy = BC->getDestTy();
2956        unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
2957        if (!TgtElemBitWidth)
2958          continue;
2959        unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
2960        bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
2961        bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
2962        if (!VecBitWidthsEqual)
2963          continue;
2964        if (!VectorType::isValidElementType(TgtTy))
2965          continue;
2966        auto *CastSrcTy = FixedVectorType::get(TgtTy, TgtNumElems);
2967        if (!BegIsAligned) {
2968          // Shuffle the input so [0,NumElements) contains the output, and
2969          // [NumElems,SrcNumElems) is undef.
2970          SmallVector<int, 16> ShuffleMask(SrcNumElems, -1);
2971          for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
2972            ShuffleMask[I] = Idx;
2973          V = Builder.CreateShuffleVector(V, ShuffleMask,
2974                                          SVI.getName() + ".extract");
2975          BegIdx = 0;
2976        }
2977        unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
2978        assert(SrcElemsPerTgtElem);
2979        BegIdx /= SrcElemsPerTgtElem;
2980        bool BCAlreadyExists = NewBCs.contains(CastSrcTy);
2981        auto *NewBC =
2982            BCAlreadyExists
2983                ? NewBCs[CastSrcTy]
2984                : Builder.CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
2985        if (!BCAlreadyExists)
2986          NewBCs[CastSrcTy] = NewBC;
2987        auto *Ext = Builder.CreateExtractElement(NewBC, BegIdx,
2988                                                 SVI.getName() + ".extract");
2989        // The shufflevector isn't being replaced: the bitcast that used it
2990        // is. InstCombine will visit the newly-created instructions.
2991        replaceInstUsesWith(*BC, Ext);
2992        MadeChange = true;
2993      }
2994    }
2995  
2996    // If the LHS is a shufflevector itself, see if we can combine it with this
2997    // one without producing an unusual shuffle.
2998    // Cases that might be simplified:
2999    // 1.
3000    // x1=shuffle(v1,v2,mask1)
3001    //  x=shuffle(x1,undef,mask)
3002    //        ==>
3003    //  x=shuffle(v1,undef,newMask)
3004    // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
3005    // 2.
3006    // x1=shuffle(v1,undef,mask1)
3007    //  x=shuffle(x1,x2,mask)
3008    // where v1.size() == mask1.size()
3009    //        ==>
3010    //  x=shuffle(v1,x2,newMask)
3011    // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
3012    // 3.
3013    // x2=shuffle(v2,undef,mask2)
3014    //  x=shuffle(x1,x2,mask)
3015    // where v2.size() == mask2.size()
3016    //        ==>
3017    //  x=shuffle(x1,v2,newMask)
3018    // newMask[i] = (mask[i] < x1.size())
3019    //              ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
3020    // 4.
3021    // x1=shuffle(v1,undef,mask1)
3022    // x2=shuffle(v2,undef,mask2)
3023    //  x=shuffle(x1,x2,mask)
3024    // where v1.size() == v2.size()
3025    //        ==>
3026    //  x=shuffle(v1,v2,newMask)
3027    // newMask[i] = (mask[i] < x1.size())
3028    //              ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
3029    //
3030    // Here we are really conservative:
3031    // we are absolutely afraid of producing a shuffle mask not in the input
3032    // program, because the code gen may not be smart enough to turn a merged
3033    // shuffle into two specific shuffles: it may produce worse code.  As such,
3034    // we only merge two shuffles if the result is either a splat or one of the
3035    // input shuffle masks.  In this case, merging the shuffles just removes
3036    // one instruction, which we know is safe.  This is good for things like
3037    // turning: (splat(splat)) -> splat, or
3038    // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
3039    ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
3040    ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
3041    if (LHSShuffle)
3042      if (!match(LHSShuffle->getOperand(1), m_Poison()) &&
3043          !match(RHS, m_Poison()))
3044        LHSShuffle = nullptr;
3045    if (RHSShuffle)
3046      if (!match(RHSShuffle->getOperand(1), m_Poison()))
3047        RHSShuffle = nullptr;
3048    if (!LHSShuffle && !RHSShuffle)
3049      return MadeChange ? &SVI : nullptr;
3050  
3051    Value* LHSOp0 = nullptr;
3052    Value* LHSOp1 = nullptr;
3053    Value* RHSOp0 = nullptr;
3054    unsigned LHSOp0Width = 0;
3055    unsigned RHSOp0Width = 0;
3056    if (LHSShuffle) {
3057      LHSOp0 = LHSShuffle->getOperand(0);
3058      LHSOp1 = LHSShuffle->getOperand(1);
3059      LHSOp0Width = cast<FixedVectorType>(LHSOp0->getType())->getNumElements();
3060    }
3061    if (RHSShuffle) {
3062      RHSOp0 = RHSShuffle->getOperand(0);
3063      RHSOp0Width = cast<FixedVectorType>(RHSOp0->getType())->getNumElements();
3064    }
3065    Value* newLHS = LHS;
3066    Value* newRHS = RHS;
3067    if (LHSShuffle) {
3068      // case 1
3069      if (match(RHS, m_Poison())) {
3070        newLHS = LHSOp0;
3071        newRHS = LHSOp1;
3072      }
3073      // case 2 or 4
3074      else if (LHSOp0Width == LHSWidth) {
3075        newLHS = LHSOp0;
3076      }
3077    }
3078    // case 3 or 4
3079    if (RHSShuffle && RHSOp0Width == LHSWidth) {
3080      newRHS = RHSOp0;
3081    }
3082    // case 4
3083    if (LHSOp0 == RHSOp0) {
3084      newLHS = LHSOp0;
3085      newRHS = nullptr;
3086    }
3087  
3088    if (newLHS == LHS && newRHS == RHS)
3089      return MadeChange ? &SVI : nullptr;
3090  
3091    ArrayRef<int> LHSMask;
3092    ArrayRef<int> RHSMask;
3093    if (newLHS != LHS)
3094      LHSMask = LHSShuffle->getShuffleMask();
3095    if (RHSShuffle && newRHS != RHS)
3096      RHSMask = RHSShuffle->getShuffleMask();
3097  
3098    unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
3099    SmallVector<int, 16> newMask;
3100    bool isSplat = true;
3101    int SplatElt = -1;
3102    // Create a new mask for the new ShuffleVectorInst so that the new
3103    // ShuffleVectorInst is equivalent to the original one.
3104    for (unsigned i = 0; i < VWidth; ++i) {
3105      int eltMask;
3106      if (Mask[i] < 0) {
3107        // This element is a poison value.
3108        eltMask = -1;
3109      } else if (Mask[i] < (int)LHSWidth) {
3110        // This element is from left hand side vector operand.
3111        //
3112        // If LHS is going to be replaced (case 1, 2, or 4), calculate the
3113        // new mask value for the element.
3114        if (newLHS != LHS) {
3115          eltMask = LHSMask[Mask[i]];
3116          // If the value selected is an poison value, explicitly specify it
3117          // with a -1 mask value.
3118          if (eltMask >= (int)LHSOp0Width && isa<PoisonValue>(LHSOp1))
3119            eltMask = -1;
3120        } else
3121          eltMask = Mask[i];
3122      } else {
3123        // This element is from right hand side vector operand
3124        //
3125        // If the value selected is a poison value, explicitly specify it
3126        // with a -1 mask value. (case 1)
3127        if (match(RHS, m_Poison()))
3128          eltMask = -1;
3129        // If RHS is going to be replaced (case 3 or 4), calculate the
3130        // new mask value for the element.
3131        else if (newRHS != RHS) {
3132          eltMask = RHSMask[Mask[i]-LHSWidth];
3133          // If the value selected is an poison value, explicitly specify it
3134          // with a -1 mask value.
3135          if (eltMask >= (int)RHSOp0Width) {
3136            assert(match(RHSShuffle->getOperand(1), m_Poison()) &&
3137                   "should have been check above");
3138            eltMask = -1;
3139          }
3140        } else
3141          eltMask = Mask[i]-LHSWidth;
3142  
3143        // If LHS's width is changed, shift the mask value accordingly.
3144        // If newRHS == nullptr, i.e. LHSOp0 == RHSOp0, we want to remap any
3145        // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
3146        // If newRHS == newLHS, we want to remap any references from newRHS to
3147        // newLHS so that we can properly identify splats that may occur due to
3148        // obfuscation across the two vectors.
3149        if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
3150          eltMask += newLHSWidth;
3151      }
3152  
3153      // Check if this could still be a splat.
3154      if (eltMask >= 0) {
3155        if (SplatElt >= 0 && SplatElt != eltMask)
3156          isSplat = false;
3157        SplatElt = eltMask;
3158      }
3159  
3160      newMask.push_back(eltMask);
3161    }
3162  
3163    // If the result mask is equal to one of the original shuffle masks,
3164    // or is a splat, do the replacement.
3165    if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
3166      if (!newRHS)
3167        newRHS = PoisonValue::get(newLHS->getType());
3168      return new ShuffleVectorInst(newLHS, newRHS, newMask);
3169    }
3170  
3171    return MadeChange ? &SVI : nullptr;
3172  }
3173