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