xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp (revision 9f23cbd6cae82fd77edfad7173432fa8dccd0a95)
1 //===- InstCombineCasts.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 the visit functions for cast operations.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/SetVector.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/IR/DataLayout.h"
17 #include "llvm/IR/DebugInfo.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Support/KnownBits.h"
20 #include "llvm/Transforms/InstCombine/InstCombiner.h"
21 #include <optional>
22 
23 using namespace llvm;
24 using namespace PatternMatch;
25 
26 #define DEBUG_TYPE "instcombine"
27 
28 /// Analyze 'Val', seeing if it is a simple linear expression.
29 /// If so, decompose it, returning some value X, such that Val is
30 /// X*Scale+Offset.
31 ///
32 static Value *decomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
33                                         uint64_t &Offset) {
34   if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
35     Offset = CI->getZExtValue();
36     Scale  = 0;
37     return ConstantInt::get(Val->getType(), 0);
38   }
39 
40   if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
41     // Cannot look past anything that might overflow.
42     // We specifically require nuw because we store the Scale in an unsigned
43     // and perform an unsigned divide on it.
44     OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
45     if (OBI && !OBI->hasNoUnsignedWrap()) {
46       Scale = 1;
47       Offset = 0;
48       return Val;
49     }
50 
51     if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
52       if (I->getOpcode() == Instruction::Shl) {
53         // This is a value scaled by '1 << the shift amt'.
54         Scale = UINT64_C(1) << RHS->getZExtValue();
55         Offset = 0;
56         return I->getOperand(0);
57       }
58 
59       if (I->getOpcode() == Instruction::Mul) {
60         // This value is scaled by 'RHS'.
61         Scale = RHS->getZExtValue();
62         Offset = 0;
63         return I->getOperand(0);
64       }
65 
66       if (I->getOpcode() == Instruction::Add) {
67         // We have X+C.  Check to see if we really have (X*C2)+C1,
68         // where C1 is divisible by C2.
69         unsigned SubScale;
70         Value *SubVal =
71           decomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
72         Offset += RHS->getZExtValue();
73         Scale = SubScale;
74         return SubVal;
75       }
76     }
77   }
78 
79   // Otherwise, we can't look past this.
80   Scale = 1;
81   Offset = 0;
82   return Val;
83 }
84 
85 /// If we find a cast of an allocation instruction, try to eliminate the cast by
86 /// moving the type information into the alloc.
87 Instruction *InstCombinerImpl::PromoteCastOfAllocation(BitCastInst &CI,
88                                                        AllocaInst &AI) {
89   PointerType *PTy = cast<PointerType>(CI.getType());
90   // Opaque pointers don't have an element type we could replace with.
91   if (PTy->isOpaque())
92     return nullptr;
93 
94   IRBuilderBase::InsertPointGuard Guard(Builder);
95   Builder.SetInsertPoint(&AI);
96 
97   // Get the type really allocated and the type casted to.
98   Type *AllocElTy = AI.getAllocatedType();
99   Type *CastElTy = PTy->getNonOpaquePointerElementType();
100   if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr;
101 
102   // This optimisation does not work for cases where the cast type
103   // is scalable and the allocated type is not. This because we need to
104   // know how many times the casted type fits into the allocated type.
105   // For the opposite case where the allocated type is scalable and the
106   // cast type is not this leads to poor code quality due to the
107   // introduction of 'vscale' into the calculations. It seems better to
108   // bail out for this case too until we've done a proper cost-benefit
109   // analysis.
110   bool AllocIsScalable = isa<ScalableVectorType>(AllocElTy);
111   bool CastIsScalable = isa<ScalableVectorType>(CastElTy);
112   if (AllocIsScalable != CastIsScalable) return nullptr;
113 
114   Align AllocElTyAlign = DL.getABITypeAlign(AllocElTy);
115   Align CastElTyAlign = DL.getABITypeAlign(CastElTy);
116   if (CastElTyAlign < AllocElTyAlign) return nullptr;
117 
118   // If the allocation has multiple uses, only promote it if we are strictly
119   // increasing the alignment of the resultant allocation.  If we keep it the
120   // same, we open the door to infinite loops of various kinds.
121   if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr;
122 
123   // The alloc and cast types should be either both fixed or both scalable.
124   uint64_t AllocElTySize = DL.getTypeAllocSize(AllocElTy).getKnownMinValue();
125   uint64_t CastElTySize = DL.getTypeAllocSize(CastElTy).getKnownMinValue();
126   if (CastElTySize == 0 || AllocElTySize == 0) return nullptr;
127 
128   // If the allocation has multiple uses, only promote it if we're not
129   // shrinking the amount of memory being allocated.
130   uint64_t AllocElTyStoreSize =
131       DL.getTypeStoreSize(AllocElTy).getKnownMinValue();
132   uint64_t CastElTyStoreSize = DL.getTypeStoreSize(CastElTy).getKnownMinValue();
133   if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr;
134 
135   // See if we can satisfy the modulus by pulling a scale out of the array
136   // size argument.
137   unsigned ArraySizeScale;
138   uint64_t ArrayOffset;
139   Value *NumElements = // See if the array size is a decomposable linear expr.
140     decomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
141 
142   // If we can now satisfy the modulus, by using a non-1 scale, we really can
143   // do the xform.
144   if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
145       (AllocElTySize*ArrayOffset   ) % CastElTySize != 0) return nullptr;
146 
147   // We don't currently support arrays of scalable types.
148   assert(!AllocIsScalable || (ArrayOffset == 1 && ArraySizeScale == 0));
149 
150   unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
151   Value *Amt = nullptr;
152   if (Scale == 1) {
153     Amt = NumElements;
154   } else {
155     Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
156     // Insert before the alloca, not before the cast.
157     Amt = Builder.CreateMul(Amt, NumElements);
158   }
159 
160   if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
161     Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
162                                   Offset, true);
163     Amt = Builder.CreateAdd(Amt, Off);
164   }
165 
166   AllocaInst *New = Builder.CreateAlloca(CastElTy, AI.getAddressSpace(), Amt);
167   New->setAlignment(AI.getAlign());
168   New->takeName(&AI);
169   New->setUsedWithInAlloca(AI.isUsedWithInAlloca());
170   New->setMetadata(LLVMContext::MD_DIAssignID,
171                    AI.getMetadata(LLVMContext::MD_DIAssignID));
172 
173   replaceAllDbgUsesWith(AI, *New, *New, DT);
174 
175   // If the allocation has multiple real uses, insert a cast and change all
176   // things that used it to use the new cast.  This will also hack on CI, but it
177   // will die soon.
178   if (!AI.hasOneUse()) {
179     // New is the allocation instruction, pointer typed. AI is the original
180     // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
181     Value *NewCast = Builder.CreateBitCast(New, AI.getType(), "tmpcast");
182     replaceInstUsesWith(AI, NewCast);
183     eraseInstFromFunction(AI);
184   }
185   return replaceInstUsesWith(CI, New);
186 }
187 
188 /// Given an expression that CanEvaluateTruncated or CanEvaluateSExtd returns
189 /// true for, actually insert the code to evaluate the expression.
190 Value *InstCombinerImpl::EvaluateInDifferentType(Value *V, Type *Ty,
191                                                  bool isSigned) {
192   if (Constant *C = dyn_cast<Constant>(V)) {
193     C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
194     // If we got a constantexpr back, try to simplify it with DL info.
195     return ConstantFoldConstant(C, DL, &TLI);
196   }
197 
198   // Otherwise, it must be an instruction.
199   Instruction *I = cast<Instruction>(V);
200   Instruction *Res = nullptr;
201   unsigned Opc = I->getOpcode();
202   switch (Opc) {
203   case Instruction::Add:
204   case Instruction::Sub:
205   case Instruction::Mul:
206   case Instruction::And:
207   case Instruction::Or:
208   case Instruction::Xor:
209   case Instruction::AShr:
210   case Instruction::LShr:
211   case Instruction::Shl:
212   case Instruction::UDiv:
213   case Instruction::URem: {
214     Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
215     Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
216     Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
217     break;
218   }
219   case Instruction::Trunc:
220   case Instruction::ZExt:
221   case Instruction::SExt:
222     // If the source type of the cast is the type we're trying for then we can
223     // just return the source.  There's no need to insert it because it is not
224     // new.
225     if (I->getOperand(0)->getType() == Ty)
226       return I->getOperand(0);
227 
228     // Otherwise, must be the same type of cast, so just reinsert a new one.
229     // This also handles the case of zext(trunc(x)) -> zext(x).
230     Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
231                                       Opc == Instruction::SExt);
232     break;
233   case Instruction::Select: {
234     Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
235     Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
236     Res = SelectInst::Create(I->getOperand(0), True, False);
237     break;
238   }
239   case Instruction::PHI: {
240     PHINode *OPN = cast<PHINode>(I);
241     PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
242     for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
243       Value *V =
244           EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
245       NPN->addIncoming(V, OPN->getIncomingBlock(i));
246     }
247     Res = NPN;
248     break;
249   }
250   case Instruction::FPToUI:
251   case Instruction::FPToSI:
252     Res = CastInst::Create(
253       static_cast<Instruction::CastOps>(Opc), I->getOperand(0), Ty);
254     break;
255   default:
256     // TODO: Can handle more cases here.
257     llvm_unreachable("Unreachable!");
258   }
259 
260   Res->takeName(I);
261   return InsertNewInstWith(Res, *I);
262 }
263 
264 Instruction::CastOps
265 InstCombinerImpl::isEliminableCastPair(const CastInst *CI1,
266                                        const CastInst *CI2) {
267   Type *SrcTy = CI1->getSrcTy();
268   Type *MidTy = CI1->getDestTy();
269   Type *DstTy = CI2->getDestTy();
270 
271   Instruction::CastOps firstOp = CI1->getOpcode();
272   Instruction::CastOps secondOp = CI2->getOpcode();
273   Type *SrcIntPtrTy =
274       SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr;
275   Type *MidIntPtrTy =
276       MidTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(MidTy) : nullptr;
277   Type *DstIntPtrTy =
278       DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr;
279   unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
280                                                 DstTy, SrcIntPtrTy, MidIntPtrTy,
281                                                 DstIntPtrTy);
282 
283   // We don't want to form an inttoptr or ptrtoint that converts to an integer
284   // type that differs from the pointer size.
285   if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
286       (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
287     Res = 0;
288 
289   return Instruction::CastOps(Res);
290 }
291 
292 /// Implement the transforms common to all CastInst visitors.
293 Instruction *InstCombinerImpl::commonCastTransforms(CastInst &CI) {
294   Value *Src = CI.getOperand(0);
295   Type *Ty = CI.getType();
296 
297   // Try to eliminate a cast of a cast.
298   if (auto *CSrc = dyn_cast<CastInst>(Src)) {   // A->B->C cast
299     if (Instruction::CastOps NewOpc = isEliminableCastPair(CSrc, &CI)) {
300       // The first cast (CSrc) is eliminable so we need to fix up or replace
301       // the second cast (CI). CSrc will then have a good chance of being dead.
302       auto *Res = CastInst::Create(NewOpc, CSrc->getOperand(0), Ty);
303       // Point debug users of the dying cast to the new one.
304       if (CSrc->hasOneUse())
305         replaceAllDbgUsesWith(*CSrc, *Res, CI, DT);
306       return Res;
307     }
308   }
309 
310   if (auto *Sel = dyn_cast<SelectInst>(Src)) {
311     // We are casting a select. Try to fold the cast into the select if the
312     // select does not have a compare instruction with matching operand types
313     // or the select is likely better done in a narrow type.
314     // Creating a select with operands that are different sizes than its
315     // condition may inhibit other folds and lead to worse codegen.
316     auto *Cmp = dyn_cast<CmpInst>(Sel->getCondition());
317     if (!Cmp || Cmp->getOperand(0)->getType() != Sel->getType() ||
318         (CI.getOpcode() == Instruction::Trunc &&
319          shouldChangeType(CI.getSrcTy(), CI.getType()))) {
320       if (Instruction *NV = FoldOpIntoSelect(CI, Sel)) {
321         replaceAllDbgUsesWith(*Sel, *NV, CI, DT);
322         return NV;
323       }
324     }
325   }
326 
327   // If we are casting a PHI, then fold the cast into the PHI.
328   if (auto *PN = dyn_cast<PHINode>(Src)) {
329     // Don't do this if it would create a PHI node with an illegal type from a
330     // legal type.
331     if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||
332         shouldChangeType(CI.getSrcTy(), CI.getType()))
333       if (Instruction *NV = foldOpIntoPhi(CI, PN))
334         return NV;
335   }
336 
337   // Canonicalize a unary shuffle after the cast if neither operation changes
338   // the size or element size of the input vector.
339   // TODO: We could allow size-changing ops if that doesn't harm codegen.
340   // cast (shuffle X, Mask) --> shuffle (cast X), Mask
341   Value *X;
342   ArrayRef<int> Mask;
343   if (match(Src, m_OneUse(m_Shuffle(m_Value(X), m_Undef(), m_Mask(Mask))))) {
344     // TODO: Allow scalable vectors?
345     auto *SrcTy = dyn_cast<FixedVectorType>(X->getType());
346     auto *DestTy = dyn_cast<FixedVectorType>(Ty);
347     if (SrcTy && DestTy &&
348         SrcTy->getNumElements() == DestTy->getNumElements() &&
349         SrcTy->getPrimitiveSizeInBits() == DestTy->getPrimitiveSizeInBits()) {
350       Value *CastX = Builder.CreateCast(CI.getOpcode(), X, DestTy);
351       return new ShuffleVectorInst(CastX, Mask);
352     }
353   }
354 
355   return nullptr;
356 }
357 
358 /// Constants and extensions/truncates from the destination type are always
359 /// free to be evaluated in that type. This is a helper for canEvaluate*.
360 static bool canAlwaysEvaluateInType(Value *V, Type *Ty) {
361   if (isa<Constant>(V))
362     return true;
363   Value *X;
364   if ((match(V, m_ZExtOrSExt(m_Value(X))) || match(V, m_Trunc(m_Value(X)))) &&
365       X->getType() == Ty)
366     return true;
367 
368   return false;
369 }
370 
371 /// Filter out values that we can not evaluate in the destination type for free.
372 /// This is a helper for canEvaluate*.
373 static bool canNotEvaluateInType(Value *V, Type *Ty) {
374   assert(!isa<Constant>(V) && "Constant should already be handled.");
375   if (!isa<Instruction>(V))
376     return true;
377   // We don't extend or shrink something that has multiple uses --  doing so
378   // would require duplicating the instruction which isn't profitable.
379   if (!V->hasOneUse())
380     return true;
381 
382   return false;
383 }
384 
385 /// Return true if we can evaluate the specified expression tree as type Ty
386 /// instead of its larger type, and arrive with the same value.
387 /// This is used by code that tries to eliminate truncates.
388 ///
389 /// Ty will always be a type smaller than V.  We should return true if trunc(V)
390 /// can be computed by computing V in the smaller type.  If V is an instruction,
391 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
392 /// makes sense if x and y can be efficiently truncated.
393 ///
394 /// This function works on both vectors and scalars.
395 ///
396 static bool canEvaluateTruncated(Value *V, Type *Ty, InstCombinerImpl &IC,
397                                  Instruction *CxtI) {
398   if (canAlwaysEvaluateInType(V, Ty))
399     return true;
400   if (canNotEvaluateInType(V, Ty))
401     return false;
402 
403   auto *I = cast<Instruction>(V);
404   Type *OrigTy = V->getType();
405   switch (I->getOpcode()) {
406   case Instruction::Add:
407   case Instruction::Sub:
408   case Instruction::Mul:
409   case Instruction::And:
410   case Instruction::Or:
411   case Instruction::Xor:
412     // These operators can all arbitrarily be extended or truncated.
413     return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
414            canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
415 
416   case Instruction::UDiv:
417   case Instruction::URem: {
418     // UDiv and URem can be truncated if all the truncated bits are zero.
419     uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
420     uint32_t BitWidth = Ty->getScalarSizeInBits();
421     assert(BitWidth < OrigBitWidth && "Unexpected bitwidths!");
422     APInt Mask = APInt::getBitsSetFrom(OrigBitWidth, BitWidth);
423     if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) &&
424         IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) {
425       return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
426              canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
427     }
428     break;
429   }
430   case Instruction::Shl: {
431     // If we are truncating the result of this SHL, and if it's a shift of an
432     // inrange amount, we can always perform a SHL in a smaller type.
433     uint32_t BitWidth = Ty->getScalarSizeInBits();
434     KnownBits AmtKnownBits =
435         llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());
436     if (AmtKnownBits.getMaxValue().ult(BitWidth))
437       return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
438              canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
439     break;
440   }
441   case Instruction::LShr: {
442     // If this is a truncate of a logical shr, we can truncate it to a smaller
443     // lshr iff we know that the bits we would otherwise be shifting in are
444     // already zeros.
445     // TODO: It is enough to check that the bits we would be shifting in are
446     //       zero - use AmtKnownBits.getMaxValue().
447     uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
448     uint32_t BitWidth = Ty->getScalarSizeInBits();
449     KnownBits AmtKnownBits =
450         llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());
451     APInt ShiftedBits = APInt::getBitsSetFrom(OrigBitWidth, BitWidth);
452     if (AmtKnownBits.getMaxValue().ult(BitWidth) &&
453         IC.MaskedValueIsZero(I->getOperand(0), ShiftedBits, 0, CxtI)) {
454       return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
455              canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
456     }
457     break;
458   }
459   case Instruction::AShr: {
460     // If this is a truncate of an arithmetic shr, we can truncate it to a
461     // smaller ashr iff we know that all the bits from the sign bit of the
462     // original type and the sign bit of the truncate type are similar.
463     // TODO: It is enough to check that the bits we would be shifting in are
464     //       similar to sign bit of the truncate type.
465     uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
466     uint32_t BitWidth = Ty->getScalarSizeInBits();
467     KnownBits AmtKnownBits =
468         llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());
469     unsigned ShiftedBits = OrigBitWidth - BitWidth;
470     if (AmtKnownBits.getMaxValue().ult(BitWidth) &&
471         ShiftedBits < IC.ComputeNumSignBits(I->getOperand(0), 0, CxtI))
472       return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
473              canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
474     break;
475   }
476   case Instruction::Trunc:
477     // trunc(trunc(x)) -> trunc(x)
478     return true;
479   case Instruction::ZExt:
480   case Instruction::SExt:
481     // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
482     // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
483     return true;
484   case Instruction::Select: {
485     SelectInst *SI = cast<SelectInst>(I);
486     return canEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
487            canEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
488   }
489   case Instruction::PHI: {
490     // We can change a phi if we can change all operands.  Note that we never
491     // get into trouble with cyclic PHIs here because we only consider
492     // instructions with a single use.
493     PHINode *PN = cast<PHINode>(I);
494     for (Value *IncValue : PN->incoming_values())
495       if (!canEvaluateTruncated(IncValue, Ty, IC, CxtI))
496         return false;
497     return true;
498   }
499   case Instruction::FPToUI:
500   case Instruction::FPToSI: {
501     // If the integer type can hold the max FP value, it is safe to cast
502     // directly to that type. Otherwise, we may create poison via overflow
503     // that did not exist in the original code.
504     //
505     // The max FP value is pow(2, MaxExponent) * (1 + MaxFraction), so we need
506     // at least one more bit than the MaxExponent to hold the max FP value.
507     Type *InputTy = I->getOperand(0)->getType()->getScalarType();
508     const fltSemantics &Semantics = InputTy->getFltSemantics();
509     uint32_t MinBitWidth = APFloatBase::semanticsMaxExponent(Semantics);
510     // Extra sign bit needed.
511     if (I->getOpcode() == Instruction::FPToSI)
512       ++MinBitWidth;
513     return Ty->getScalarSizeInBits() > MinBitWidth;
514   }
515   default:
516     // TODO: Can handle more cases here.
517     break;
518   }
519 
520   return false;
521 }
522 
523 /// Given a vector that is bitcast to an integer, optionally logically
524 /// right-shifted, and truncated, convert it to an extractelement.
525 /// Example (big endian):
526 ///   trunc (lshr (bitcast <4 x i32> %X to i128), 32) to i32
527 ///   --->
528 ///   extractelement <4 x i32> %X, 1
529 static Instruction *foldVecTruncToExtElt(TruncInst &Trunc,
530                                          InstCombinerImpl &IC) {
531   Value *TruncOp = Trunc.getOperand(0);
532   Type *DestType = Trunc.getType();
533   if (!TruncOp->hasOneUse() || !isa<IntegerType>(DestType))
534     return nullptr;
535 
536   Value *VecInput = nullptr;
537   ConstantInt *ShiftVal = nullptr;
538   if (!match(TruncOp, m_CombineOr(m_BitCast(m_Value(VecInput)),
539                                   m_LShr(m_BitCast(m_Value(VecInput)),
540                                          m_ConstantInt(ShiftVal)))) ||
541       !isa<VectorType>(VecInput->getType()))
542     return nullptr;
543 
544   VectorType *VecType = cast<VectorType>(VecInput->getType());
545   unsigned VecWidth = VecType->getPrimitiveSizeInBits();
546   unsigned DestWidth = DestType->getPrimitiveSizeInBits();
547   unsigned ShiftAmount = ShiftVal ? ShiftVal->getZExtValue() : 0;
548 
549   if ((VecWidth % DestWidth != 0) || (ShiftAmount % DestWidth != 0))
550     return nullptr;
551 
552   // If the element type of the vector doesn't match the result type,
553   // bitcast it to a vector type that we can extract from.
554   unsigned NumVecElts = VecWidth / DestWidth;
555   if (VecType->getElementType() != DestType) {
556     VecType = FixedVectorType::get(DestType, NumVecElts);
557     VecInput = IC.Builder.CreateBitCast(VecInput, VecType, "bc");
558   }
559 
560   unsigned Elt = ShiftAmount / DestWidth;
561   if (IC.getDataLayout().isBigEndian())
562     Elt = NumVecElts - 1 - Elt;
563 
564   return ExtractElementInst::Create(VecInput, IC.Builder.getInt32(Elt));
565 }
566 
567 /// Funnel/Rotate left/right may occur in a wider type than necessary because of
568 /// type promotion rules. Try to narrow the inputs and convert to funnel shift.
569 Instruction *InstCombinerImpl::narrowFunnelShift(TruncInst &Trunc) {
570   assert((isa<VectorType>(Trunc.getSrcTy()) ||
571           shouldChangeType(Trunc.getSrcTy(), Trunc.getType())) &&
572          "Don't narrow to an illegal scalar type");
573 
574   // Bail out on strange types. It is possible to handle some of these patterns
575   // even with non-power-of-2 sizes, but it is not a likely scenario.
576   Type *DestTy = Trunc.getType();
577   unsigned NarrowWidth = DestTy->getScalarSizeInBits();
578   unsigned WideWidth = Trunc.getSrcTy()->getScalarSizeInBits();
579   if (!isPowerOf2_32(NarrowWidth))
580     return nullptr;
581 
582   // First, find an or'd pair of opposite shifts:
583   // trunc (or (lshr ShVal0, ShAmt0), (shl ShVal1, ShAmt1))
584   BinaryOperator *Or0, *Or1;
585   if (!match(Trunc.getOperand(0), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1)))))
586     return nullptr;
587 
588   Value *ShVal0, *ShVal1, *ShAmt0, *ShAmt1;
589   if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(ShVal0), m_Value(ShAmt0)))) ||
590       !match(Or1, m_OneUse(m_LogicalShift(m_Value(ShVal1), m_Value(ShAmt1)))) ||
591       Or0->getOpcode() == Or1->getOpcode())
592     return nullptr;
593 
594   // Canonicalize to or(shl(ShVal0, ShAmt0), lshr(ShVal1, ShAmt1)).
595   if (Or0->getOpcode() == BinaryOperator::LShr) {
596     std::swap(Or0, Or1);
597     std::swap(ShVal0, ShVal1);
598     std::swap(ShAmt0, ShAmt1);
599   }
600   assert(Or0->getOpcode() == BinaryOperator::Shl &&
601          Or1->getOpcode() == BinaryOperator::LShr &&
602          "Illegal or(shift,shift) pair");
603 
604   // Match the shift amount operands for a funnel/rotate pattern. This always
605   // matches a subtraction on the R operand.
606   auto matchShiftAmount = [&](Value *L, Value *R, unsigned Width) -> Value * {
607     // The shift amounts may add up to the narrow bit width:
608     // (shl ShVal0, L) | (lshr ShVal1, Width - L)
609     // If this is a funnel shift (different operands are shifted), then the
610     // shift amount can not over-shift (create poison) in the narrow type.
611     unsigned MaxShiftAmountWidth = Log2_32(NarrowWidth);
612     APInt HiBitMask = ~APInt::getLowBitsSet(WideWidth, MaxShiftAmountWidth);
613     if (ShVal0 == ShVal1 || MaskedValueIsZero(L, HiBitMask))
614       if (match(R, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(L)))))
615         return L;
616 
617     // The following patterns currently only work for rotation patterns.
618     // TODO: Add more general funnel-shift compatible patterns.
619     if (ShVal0 != ShVal1)
620       return nullptr;
621 
622     // The shift amount may be masked with negation:
623     // (shl ShVal0, (X & (Width - 1))) | (lshr ShVal1, ((-X) & (Width - 1)))
624     Value *X;
625     unsigned Mask = Width - 1;
626     if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) &&
627         match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))
628       return X;
629 
630     // Same as above, but the shift amount may be extended after masking:
631     if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&
632         match(R, m_ZExt(m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask)))))
633       return X;
634 
635     return nullptr;
636   };
637 
638   Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, NarrowWidth);
639   bool IsFshl = true; // Sub on LSHR.
640   if (!ShAmt) {
641     ShAmt = matchShiftAmount(ShAmt1, ShAmt0, NarrowWidth);
642     IsFshl = false; // Sub on SHL.
643   }
644   if (!ShAmt)
645     return nullptr;
646 
647   // The right-shifted value must have high zeros in the wide type (for example
648   // from 'zext', 'and' or 'shift'). High bits of the left-shifted value are
649   // truncated, so those do not matter.
650   APInt HiBitMask = APInt::getHighBitsSet(WideWidth, WideWidth - NarrowWidth);
651   if (!MaskedValueIsZero(ShVal1, HiBitMask, 0, &Trunc))
652     return nullptr;
653 
654   // We have an unnecessarily wide rotate!
655   // trunc (or (shl ShVal0, ShAmt), (lshr ShVal1, BitWidth - ShAmt))
656   // Narrow the inputs and convert to funnel shift intrinsic:
657   // llvm.fshl.i8(trunc(ShVal), trunc(ShVal), trunc(ShAmt))
658   Value *NarrowShAmt = Builder.CreateTrunc(ShAmt, DestTy);
659   Value *X, *Y;
660   X = Y = Builder.CreateTrunc(ShVal0, DestTy);
661   if (ShVal0 != ShVal1)
662     Y = Builder.CreateTrunc(ShVal1, DestTy);
663   Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
664   Function *F = Intrinsic::getDeclaration(Trunc.getModule(), IID, DestTy);
665   return CallInst::Create(F, {X, Y, NarrowShAmt});
666 }
667 
668 /// Try to narrow the width of math or bitwise logic instructions by pulling a
669 /// truncate ahead of binary operators.
670 Instruction *InstCombinerImpl::narrowBinOp(TruncInst &Trunc) {
671   Type *SrcTy = Trunc.getSrcTy();
672   Type *DestTy = Trunc.getType();
673   unsigned SrcWidth = SrcTy->getScalarSizeInBits();
674   unsigned DestWidth = DestTy->getScalarSizeInBits();
675 
676   if (!isa<VectorType>(SrcTy) && !shouldChangeType(SrcTy, DestTy))
677     return nullptr;
678 
679   BinaryOperator *BinOp;
680   if (!match(Trunc.getOperand(0), m_OneUse(m_BinOp(BinOp))))
681     return nullptr;
682 
683   Value *BinOp0 = BinOp->getOperand(0);
684   Value *BinOp1 = BinOp->getOperand(1);
685   switch (BinOp->getOpcode()) {
686   case Instruction::And:
687   case Instruction::Or:
688   case Instruction::Xor:
689   case Instruction::Add:
690   case Instruction::Sub:
691   case Instruction::Mul: {
692     Constant *C;
693     if (match(BinOp0, m_Constant(C))) {
694       // trunc (binop C, X) --> binop (trunc C', X)
695       Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);
696       Value *TruncX = Builder.CreateTrunc(BinOp1, DestTy);
697       return BinaryOperator::Create(BinOp->getOpcode(), NarrowC, TruncX);
698     }
699     if (match(BinOp1, m_Constant(C))) {
700       // trunc (binop X, C) --> binop (trunc X, C')
701       Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);
702       Value *TruncX = Builder.CreateTrunc(BinOp0, DestTy);
703       return BinaryOperator::Create(BinOp->getOpcode(), TruncX, NarrowC);
704     }
705     Value *X;
706     if (match(BinOp0, m_ZExtOrSExt(m_Value(X))) && X->getType() == DestTy) {
707       // trunc (binop (ext X), Y) --> binop X, (trunc Y)
708       Value *NarrowOp1 = Builder.CreateTrunc(BinOp1, DestTy);
709       return BinaryOperator::Create(BinOp->getOpcode(), X, NarrowOp1);
710     }
711     if (match(BinOp1, m_ZExtOrSExt(m_Value(X))) && X->getType() == DestTy) {
712       // trunc (binop Y, (ext X)) --> binop (trunc Y), X
713       Value *NarrowOp0 = Builder.CreateTrunc(BinOp0, DestTy);
714       return BinaryOperator::Create(BinOp->getOpcode(), NarrowOp0, X);
715     }
716     break;
717   }
718   case Instruction::LShr:
719   case Instruction::AShr: {
720     // trunc (*shr (trunc A), C) --> trunc(*shr A, C)
721     Value *A;
722     Constant *C;
723     if (match(BinOp0, m_Trunc(m_Value(A))) && match(BinOp1, m_Constant(C))) {
724       unsigned MaxShiftAmt = SrcWidth - DestWidth;
725       // If the shift is small enough, all zero/sign bits created by the shift
726       // are removed by the trunc.
727       if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULE,
728                                       APInt(SrcWidth, MaxShiftAmt)))) {
729         auto *OldShift = cast<Instruction>(Trunc.getOperand(0));
730         bool IsExact = OldShift->isExact();
731         auto *ShAmt = ConstantExpr::getIntegerCast(C, A->getType(), true);
732         ShAmt = Constant::mergeUndefsWith(ShAmt, C);
733         Value *Shift =
734             OldShift->getOpcode() == Instruction::AShr
735                 ? Builder.CreateAShr(A, ShAmt, OldShift->getName(), IsExact)
736                 : Builder.CreateLShr(A, ShAmt, OldShift->getName(), IsExact);
737         return CastInst::CreateTruncOrBitCast(Shift, DestTy);
738       }
739     }
740     break;
741   }
742   default: break;
743   }
744 
745   if (Instruction *NarrowOr = narrowFunnelShift(Trunc))
746     return NarrowOr;
747 
748   return nullptr;
749 }
750 
751 /// Try to narrow the width of a splat shuffle. This could be generalized to any
752 /// shuffle with a constant operand, but we limit the transform to avoid
753 /// creating a shuffle type that targets may not be able to lower effectively.
754 static Instruction *shrinkSplatShuffle(TruncInst &Trunc,
755                                        InstCombiner::BuilderTy &Builder) {
756   auto *Shuf = dyn_cast<ShuffleVectorInst>(Trunc.getOperand(0));
757   if (Shuf && Shuf->hasOneUse() && match(Shuf->getOperand(1), m_Undef()) &&
758       all_equal(Shuf->getShuffleMask()) &&
759       Shuf->getType() == Shuf->getOperand(0)->getType()) {
760     // trunc (shuf X, Undef, SplatMask) --> shuf (trunc X), Poison, SplatMask
761     // trunc (shuf X, Poison, SplatMask) --> shuf (trunc X), Poison, SplatMask
762     Value *NarrowOp = Builder.CreateTrunc(Shuf->getOperand(0), Trunc.getType());
763     return new ShuffleVectorInst(NarrowOp, Shuf->getShuffleMask());
764   }
765 
766   return nullptr;
767 }
768 
769 /// Try to narrow the width of an insert element. This could be generalized for
770 /// any vector constant, but we limit the transform to insertion into undef to
771 /// avoid potential backend problems from unsupported insertion widths. This
772 /// could also be extended to handle the case of inserting a scalar constant
773 /// into a vector variable.
774 static Instruction *shrinkInsertElt(CastInst &Trunc,
775                                     InstCombiner::BuilderTy &Builder) {
776   Instruction::CastOps Opcode = Trunc.getOpcode();
777   assert((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
778          "Unexpected instruction for shrinking");
779 
780   auto *InsElt = dyn_cast<InsertElementInst>(Trunc.getOperand(0));
781   if (!InsElt || !InsElt->hasOneUse())
782     return nullptr;
783 
784   Type *DestTy = Trunc.getType();
785   Type *DestScalarTy = DestTy->getScalarType();
786   Value *VecOp = InsElt->getOperand(0);
787   Value *ScalarOp = InsElt->getOperand(1);
788   Value *Index = InsElt->getOperand(2);
789 
790   if (match(VecOp, m_Undef())) {
791     // trunc   (inselt undef, X, Index) --> inselt undef,   (trunc X), Index
792     // fptrunc (inselt undef, X, Index) --> inselt undef, (fptrunc X), Index
793     UndefValue *NarrowUndef = UndefValue::get(DestTy);
794     Value *NarrowOp = Builder.CreateCast(Opcode, ScalarOp, DestScalarTy);
795     return InsertElementInst::Create(NarrowUndef, NarrowOp, Index);
796   }
797 
798   return nullptr;
799 }
800 
801 Instruction *InstCombinerImpl::visitTrunc(TruncInst &Trunc) {
802   if (Instruction *Result = commonCastTransforms(Trunc))
803     return Result;
804 
805   Value *Src = Trunc.getOperand(0);
806   Type *DestTy = Trunc.getType(), *SrcTy = Src->getType();
807   unsigned DestWidth = DestTy->getScalarSizeInBits();
808   unsigned SrcWidth = SrcTy->getScalarSizeInBits();
809 
810   // Attempt to truncate the entire input expression tree to the destination
811   // type.   Only do this if the dest type is a simple type, don't convert the
812   // expression tree to something weird like i93 unless the source is also
813   // strange.
814   if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
815       canEvaluateTruncated(Src, DestTy, *this, &Trunc)) {
816 
817     // If this cast is a truncate, evaluting in a different type always
818     // eliminates the cast, so it is always a win.
819     LLVM_DEBUG(
820         dbgs() << "ICE: EvaluateInDifferentType converting expression type"
821                   " to avoid cast: "
822                << Trunc << '\n');
823     Value *Res = EvaluateInDifferentType(Src, DestTy, false);
824     assert(Res->getType() == DestTy);
825     return replaceInstUsesWith(Trunc, Res);
826   }
827 
828   // For integer types, check if we can shorten the entire input expression to
829   // DestWidth * 2, which won't allow removing the truncate, but reducing the
830   // width may enable further optimizations, e.g. allowing for larger
831   // vectorization factors.
832   if (auto *DestITy = dyn_cast<IntegerType>(DestTy)) {
833     if (DestWidth * 2 < SrcWidth) {
834       auto *NewDestTy = DestITy->getExtendedType();
835       if (shouldChangeType(SrcTy, NewDestTy) &&
836           canEvaluateTruncated(Src, NewDestTy, *this, &Trunc)) {
837         LLVM_DEBUG(
838             dbgs() << "ICE: EvaluateInDifferentType converting expression type"
839                       " to reduce the width of operand of"
840                    << Trunc << '\n');
841         Value *Res = EvaluateInDifferentType(Src, NewDestTy, false);
842         return new TruncInst(Res, DestTy);
843       }
844     }
845   }
846 
847   // Test if the trunc is the user of a select which is part of a
848   // minimum or maximum operation. If so, don't do any more simplification.
849   // Even simplifying demanded bits can break the canonical form of a
850   // min/max.
851   Value *LHS, *RHS;
852   if (SelectInst *Sel = dyn_cast<SelectInst>(Src))
853     if (matchSelectPattern(Sel, LHS, RHS).Flavor != SPF_UNKNOWN)
854       return nullptr;
855 
856   // See if we can simplify any instructions used by the input whose sole
857   // purpose is to compute bits we don't care about.
858   if (SimplifyDemandedInstructionBits(Trunc))
859     return &Trunc;
860 
861   if (DestWidth == 1) {
862     Value *Zero = Constant::getNullValue(SrcTy);
863     if (DestTy->isIntegerTy()) {
864       // Canonicalize trunc x to i1 -> icmp ne (and x, 1), 0 (scalar only).
865       // TODO: We canonicalize to more instructions here because we are probably
866       // lacking equivalent analysis for trunc relative to icmp. There may also
867       // be codegen concerns. If those trunc limitations were removed, we could
868       // remove this transform.
869       Value *And = Builder.CreateAnd(Src, ConstantInt::get(SrcTy, 1));
870       return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
871     }
872 
873     // For vectors, we do not canonicalize all truncs to icmp, so optimize
874     // patterns that would be covered within visitICmpInst.
875     Value *X;
876     Constant *C;
877     if (match(Src, m_OneUse(m_LShr(m_Value(X), m_Constant(C))))) {
878       // trunc (lshr X, C) to i1 --> icmp ne (and X, C'), 0
879       Constant *One = ConstantInt::get(SrcTy, APInt(SrcWidth, 1));
880       Constant *MaskC = ConstantExpr::getShl(One, C);
881       Value *And = Builder.CreateAnd(X, MaskC);
882       return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
883     }
884     if (match(Src, m_OneUse(m_c_Or(m_LShr(m_Value(X), m_Constant(C)),
885                                    m_Deferred(X))))) {
886       // trunc (or (lshr X, C), X) to i1 --> icmp ne (and X, C'), 0
887       Constant *One = ConstantInt::get(SrcTy, APInt(SrcWidth, 1));
888       Constant *MaskC = ConstantExpr::getShl(One, C);
889       MaskC = ConstantExpr::getOr(MaskC, One);
890       Value *And = Builder.CreateAnd(X, MaskC);
891       return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
892     }
893   }
894 
895   Value *A, *B;
896   Constant *C;
897   if (match(Src, m_LShr(m_SExt(m_Value(A)), m_Constant(C)))) {
898     unsigned AWidth = A->getType()->getScalarSizeInBits();
899     unsigned MaxShiftAmt = SrcWidth - std::max(DestWidth, AWidth);
900     auto *OldSh = cast<Instruction>(Src);
901     bool IsExact = OldSh->isExact();
902 
903     // If the shift is small enough, all zero bits created by the shift are
904     // removed by the trunc.
905     if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULE,
906                                     APInt(SrcWidth, MaxShiftAmt)))) {
907       // trunc (lshr (sext A), C) --> ashr A, C
908       if (A->getType() == DestTy) {
909         Constant *MaxAmt = ConstantInt::get(SrcTy, DestWidth - 1, false);
910         Constant *ShAmt = ConstantExpr::getUMin(C, MaxAmt);
911         ShAmt = ConstantExpr::getTrunc(ShAmt, A->getType());
912         ShAmt = Constant::mergeUndefsWith(ShAmt, C);
913         return IsExact ? BinaryOperator::CreateExactAShr(A, ShAmt)
914                        : BinaryOperator::CreateAShr(A, ShAmt);
915       }
916       // The types are mismatched, so create a cast after shifting:
917       // trunc (lshr (sext A), C) --> sext/trunc (ashr A, C)
918       if (Src->hasOneUse()) {
919         Constant *MaxAmt = ConstantInt::get(SrcTy, AWidth - 1, false);
920         Constant *ShAmt = ConstantExpr::getUMin(C, MaxAmt);
921         ShAmt = ConstantExpr::getTrunc(ShAmt, A->getType());
922         Value *Shift = Builder.CreateAShr(A, ShAmt, "", IsExact);
923         return CastInst::CreateIntegerCast(Shift, DestTy, true);
924       }
925     }
926     // TODO: Mask high bits with 'and'.
927   }
928 
929   if (Instruction *I = narrowBinOp(Trunc))
930     return I;
931 
932   if (Instruction *I = shrinkSplatShuffle(Trunc, Builder))
933     return I;
934 
935   if (Instruction *I = shrinkInsertElt(Trunc, Builder))
936     return I;
937 
938   if (Src->hasOneUse() &&
939       (isa<VectorType>(SrcTy) || shouldChangeType(SrcTy, DestTy))) {
940     // Transform "trunc (shl X, cst)" -> "shl (trunc X), cst" so long as the
941     // dest type is native and cst < dest size.
942     if (match(Src, m_Shl(m_Value(A), m_Constant(C))) &&
943         !match(A, m_Shr(m_Value(), m_Constant()))) {
944       // Skip shifts of shift by constants. It undoes a combine in
945       // FoldShiftByConstant and is the extend in reg pattern.
946       APInt Threshold = APInt(C->getType()->getScalarSizeInBits(), DestWidth);
947       if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold))) {
948         Value *NewTrunc = Builder.CreateTrunc(A, DestTy, A->getName() + ".tr");
949         return BinaryOperator::Create(Instruction::Shl, NewTrunc,
950                                       ConstantExpr::getTrunc(C, DestTy));
951       }
952     }
953   }
954 
955   if (Instruction *I = foldVecTruncToExtElt(Trunc, *this))
956     return I;
957 
958   // Whenever an element is extracted from a vector, and then truncated,
959   // canonicalize by converting it to a bitcast followed by an
960   // extractelement.
961   //
962   // Example (little endian):
963   //   trunc (extractelement <4 x i64> %X, 0) to i32
964   //   --->
965   //   extractelement <8 x i32> (bitcast <4 x i64> %X to <8 x i32>), i32 0
966   Value *VecOp;
967   ConstantInt *Cst;
968   if (match(Src, m_OneUse(m_ExtractElt(m_Value(VecOp), m_ConstantInt(Cst))))) {
969     auto *VecOpTy = cast<VectorType>(VecOp->getType());
970     auto VecElts = VecOpTy->getElementCount();
971 
972     // A badly fit destination size would result in an invalid cast.
973     if (SrcWidth % DestWidth == 0) {
974       uint64_t TruncRatio = SrcWidth / DestWidth;
975       uint64_t BitCastNumElts = VecElts.getKnownMinValue() * TruncRatio;
976       uint64_t VecOpIdx = Cst->getZExtValue();
977       uint64_t NewIdx = DL.isBigEndian() ? (VecOpIdx + 1) * TruncRatio - 1
978                                          : VecOpIdx * TruncRatio;
979       assert(BitCastNumElts <= std::numeric_limits<uint32_t>::max() &&
980              "overflow 32-bits");
981 
982       auto *BitCastTo =
983           VectorType::get(DestTy, BitCastNumElts, VecElts.isScalable());
984       Value *BitCast = Builder.CreateBitCast(VecOp, BitCastTo);
985       return ExtractElementInst::Create(BitCast, Builder.getInt32(NewIdx));
986     }
987   }
988 
989   // trunc (ctlz_i32(zext(A), B) --> add(ctlz_i16(A, B), C)
990   if (match(Src, m_OneUse(m_Intrinsic<Intrinsic::ctlz>(m_ZExt(m_Value(A)),
991                                                        m_Value(B))))) {
992     unsigned AWidth = A->getType()->getScalarSizeInBits();
993     if (AWidth == DestWidth && AWidth > Log2_32(SrcWidth)) {
994       Value *WidthDiff = ConstantInt::get(A->getType(), SrcWidth - AWidth);
995       Value *NarrowCtlz =
996           Builder.CreateIntrinsic(Intrinsic::ctlz, {Trunc.getType()}, {A, B});
997       return BinaryOperator::CreateAdd(NarrowCtlz, WidthDiff);
998     }
999   }
1000 
1001   if (match(Src, m_VScale(DL))) {
1002     if (Trunc.getFunction() &&
1003         Trunc.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
1004       Attribute Attr =
1005           Trunc.getFunction()->getFnAttribute(Attribute::VScaleRange);
1006       if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax()) {
1007         if (Log2_32(*MaxVScale) < DestWidth) {
1008           Value *VScale = Builder.CreateVScale(ConstantInt::get(DestTy, 1));
1009           return replaceInstUsesWith(Trunc, VScale);
1010         }
1011       }
1012     }
1013   }
1014 
1015   return nullptr;
1016 }
1017 
1018 Instruction *InstCombinerImpl::transformZExtICmp(ICmpInst *Cmp,
1019                                                  ZExtInst &Zext) {
1020   // If we are just checking for a icmp eq of a single bit and zext'ing it
1021   // to an integer, then shift the bit to the appropriate place and then
1022   // cast to integer to avoid the comparison.
1023 
1024   // FIXME: This set of transforms does not check for extra uses and/or creates
1025   //        an extra instruction (an optional final cast is not included
1026   //        in the transform comments). We may also want to favor icmp over
1027   //        shifts in cases of equal instructions because icmp has better
1028   //        analysis in general (invert the transform).
1029 
1030   const APInt *Op1CV;
1031   if (match(Cmp->getOperand(1), m_APInt(Op1CV))) {
1032 
1033     // zext (x <s  0) to i32 --> x>>u31      true if signbit set.
1034     if (Cmp->getPredicate() == ICmpInst::ICMP_SLT && Op1CV->isZero()) {
1035       Value *In = Cmp->getOperand(0);
1036       Value *Sh = ConstantInt::get(In->getType(),
1037                                    In->getType()->getScalarSizeInBits() - 1);
1038       In = Builder.CreateLShr(In, Sh, In->getName() + ".lobit");
1039       if (In->getType() != Zext.getType())
1040         In = Builder.CreateIntCast(In, Zext.getType(), false /*ZExt*/);
1041 
1042       return replaceInstUsesWith(Zext, In);
1043     }
1044 
1045     // zext (X == 0) to i32 --> X^1      iff X has only the low bit set.
1046     // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
1047     // zext (X != 0) to i32 --> X        iff X has only the low bit set.
1048     // zext (X != 0) to i32 --> X>>1     iff X has only the 2nd bit set.
1049     if (Op1CV->isZero() && Cmp->isEquality() &&
1050         (Cmp->getOperand(0)->getType() == Zext.getType() ||
1051          Cmp->getPredicate() == ICmpInst::ICMP_NE)) {
1052       // If Op1C some other power of two, convert:
1053       KnownBits Known = computeKnownBits(Cmp->getOperand(0), 0, &Zext);
1054 
1055       // Exactly 1 possible 1? But not the high-bit because that is
1056       // canonicalized to this form.
1057       APInt KnownZeroMask(~Known.Zero);
1058       if (KnownZeroMask.isPowerOf2() &&
1059           (Zext.getType()->getScalarSizeInBits() !=
1060            KnownZeroMask.logBase2() + 1)) {
1061         uint32_t ShAmt = KnownZeroMask.logBase2();
1062         Value *In = Cmp->getOperand(0);
1063         if (ShAmt) {
1064           // Perform a logical shr by shiftamt.
1065           // Insert the shift to put the result in the low bit.
1066           In = Builder.CreateLShr(In, ConstantInt::get(In->getType(), ShAmt),
1067                                   In->getName() + ".lobit");
1068         }
1069 
1070         // Toggle the low bit for "X == 0".
1071         if (Cmp->getPredicate() == ICmpInst::ICMP_EQ)
1072           In = Builder.CreateXor(In, ConstantInt::get(In->getType(), 1));
1073 
1074         if (Zext.getType() == In->getType())
1075           return replaceInstUsesWith(Zext, In);
1076 
1077         Value *IntCast = Builder.CreateIntCast(In, Zext.getType(), false);
1078         return replaceInstUsesWith(Zext, IntCast);
1079       }
1080     }
1081   }
1082 
1083   if (Cmp->isEquality() && Zext.getType() == Cmp->getOperand(0)->getType()) {
1084     // Test if a bit is clear/set using a shifted-one mask:
1085     // zext (icmp eq (and X, (1 << ShAmt)), 0) --> and (lshr (not X), ShAmt), 1
1086     // zext (icmp ne (and X, (1 << ShAmt)), 0) --> and (lshr X, ShAmt), 1
1087     Value *X, *ShAmt;
1088     if (Cmp->hasOneUse() && match(Cmp->getOperand(1), m_ZeroInt()) &&
1089         match(Cmp->getOperand(0),
1090               m_OneUse(m_c_And(m_Shl(m_One(), m_Value(ShAmt)), m_Value(X))))) {
1091       if (Cmp->getPredicate() == ICmpInst::ICMP_EQ)
1092         X = Builder.CreateNot(X);
1093       Value *Lshr = Builder.CreateLShr(X, ShAmt);
1094       Value *And1 = Builder.CreateAnd(Lshr, ConstantInt::get(X->getType(), 1));
1095       return replaceInstUsesWith(Zext, And1);
1096     }
1097   }
1098 
1099   return nullptr;
1100 }
1101 
1102 /// Determine if the specified value can be computed in the specified wider type
1103 /// and produce the same low bits. If not, return false.
1104 ///
1105 /// If this function returns true, it can also return a non-zero number of bits
1106 /// (in BitsToClear) which indicates that the value it computes is correct for
1107 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
1108 /// out.  For example, to promote something like:
1109 ///
1110 ///   %B = trunc i64 %A to i32
1111 ///   %C = lshr i32 %B, 8
1112 ///   %E = zext i32 %C to i64
1113 ///
1114 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
1115 /// set to 8 to indicate that the promoted value needs to have bits 24-31
1116 /// cleared in addition to bits 32-63.  Since an 'and' will be generated to
1117 /// clear the top bits anyway, doing this has no extra cost.
1118 ///
1119 /// This function works on both vectors and scalars.
1120 static bool canEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
1121                              InstCombinerImpl &IC, Instruction *CxtI) {
1122   BitsToClear = 0;
1123   if (canAlwaysEvaluateInType(V, Ty))
1124     return true;
1125   if (canNotEvaluateInType(V, Ty))
1126     return false;
1127 
1128   auto *I = cast<Instruction>(V);
1129   unsigned Tmp;
1130   switch (I->getOpcode()) {
1131   case Instruction::ZExt:  // zext(zext(x)) -> zext(x).
1132   case Instruction::SExt:  // zext(sext(x)) -> sext(x).
1133   case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
1134     return true;
1135   case Instruction::And:
1136   case Instruction::Or:
1137   case Instruction::Xor:
1138   case Instruction::Add:
1139   case Instruction::Sub:
1140   case Instruction::Mul:
1141     if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
1142         !canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
1143       return false;
1144     // These can all be promoted if neither operand has 'bits to clear'.
1145     if (BitsToClear == 0 && Tmp == 0)
1146       return true;
1147 
1148     // If the operation is an AND/OR/XOR and the bits to clear are zero in the
1149     // other side, BitsToClear is ok.
1150     if (Tmp == 0 && I->isBitwiseLogicOp()) {
1151       // We use MaskedValueIsZero here for generality, but the case we care
1152       // about the most is constant RHS.
1153       unsigned VSize = V->getType()->getScalarSizeInBits();
1154       if (IC.MaskedValueIsZero(I->getOperand(1),
1155                                APInt::getHighBitsSet(VSize, BitsToClear),
1156                                0, CxtI)) {
1157         // If this is an And instruction and all of the BitsToClear are
1158         // known to be zero we can reset BitsToClear.
1159         if (I->getOpcode() == Instruction::And)
1160           BitsToClear = 0;
1161         return true;
1162       }
1163     }
1164 
1165     // Otherwise, we don't know how to analyze this BitsToClear case yet.
1166     return false;
1167 
1168   case Instruction::Shl: {
1169     // We can promote shl(x, cst) if we can promote x.  Since shl overwrites the
1170     // upper bits we can reduce BitsToClear by the shift amount.
1171     const APInt *Amt;
1172     if (match(I->getOperand(1), m_APInt(Amt))) {
1173       if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
1174         return false;
1175       uint64_t ShiftAmt = Amt->getZExtValue();
1176       BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
1177       return true;
1178     }
1179     return false;
1180   }
1181   case Instruction::LShr: {
1182     // We can promote lshr(x, cst) if we can promote x.  This requires the
1183     // ultimate 'and' to clear out the high zero bits we're clearing out though.
1184     const APInt *Amt;
1185     if (match(I->getOperand(1), m_APInt(Amt))) {
1186       if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
1187         return false;
1188       BitsToClear += Amt->getZExtValue();
1189       if (BitsToClear > V->getType()->getScalarSizeInBits())
1190         BitsToClear = V->getType()->getScalarSizeInBits();
1191       return true;
1192     }
1193     // Cannot promote variable LSHR.
1194     return false;
1195   }
1196   case Instruction::Select:
1197     if (!canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
1198         !canEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
1199         // TODO: If important, we could handle the case when the BitsToClear are
1200         // known zero in the disagreeing side.
1201         Tmp != BitsToClear)
1202       return false;
1203     return true;
1204 
1205   case Instruction::PHI: {
1206     // We can change a phi if we can change all operands.  Note that we never
1207     // get into trouble with cyclic PHIs here because we only consider
1208     // instructions with a single use.
1209     PHINode *PN = cast<PHINode>(I);
1210     if (!canEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
1211       return false;
1212     for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
1213       if (!canEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
1214           // TODO: If important, we could handle the case when the BitsToClear
1215           // are known zero in the disagreeing input.
1216           Tmp != BitsToClear)
1217         return false;
1218     return true;
1219   }
1220   default:
1221     // TODO: Can handle more cases here.
1222     return false;
1223   }
1224 }
1225 
1226 Instruction *InstCombinerImpl::visitZExt(ZExtInst &Zext) {
1227   // If this zero extend is only used by a truncate, let the truncate be
1228   // eliminated before we try to optimize this zext.
1229   if (Zext.hasOneUse() && isa<TruncInst>(Zext.user_back()))
1230     return nullptr;
1231 
1232   // If one of the common conversion will work, do it.
1233   if (Instruction *Result = commonCastTransforms(Zext))
1234     return Result;
1235 
1236   Value *Src = Zext.getOperand(0);
1237   Type *SrcTy = Src->getType(), *DestTy = Zext.getType();
1238 
1239   // Try to extend the entire expression tree to the wide destination type.
1240   unsigned BitsToClear;
1241   if (shouldChangeType(SrcTy, DestTy) &&
1242       canEvaluateZExtd(Src, DestTy, BitsToClear, *this, &Zext)) {
1243     assert(BitsToClear <= SrcTy->getScalarSizeInBits() &&
1244            "Can't clear more bits than in SrcTy");
1245 
1246     // Okay, we can transform this!  Insert the new expression now.
1247     LLVM_DEBUG(
1248         dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1249                   " to avoid zero extend: "
1250                << Zext << '\n');
1251     Value *Res = EvaluateInDifferentType(Src, DestTy, false);
1252     assert(Res->getType() == DestTy);
1253 
1254     // Preserve debug values referring to Src if the zext is its last use.
1255     if (auto *SrcOp = dyn_cast<Instruction>(Src))
1256       if (SrcOp->hasOneUse())
1257         replaceAllDbgUsesWith(*SrcOp, *Res, Zext, DT);
1258 
1259     uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits() - BitsToClear;
1260     uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1261 
1262     // If the high bits are already filled with zeros, just replace this
1263     // cast with the result.
1264     if (MaskedValueIsZero(Res,
1265                           APInt::getHighBitsSet(DestBitSize,
1266                                                 DestBitSize - SrcBitsKept),
1267                              0, &Zext))
1268       return replaceInstUsesWith(Zext, Res);
1269 
1270     // We need to emit an AND to clear the high bits.
1271     Constant *C = ConstantInt::get(Res->getType(),
1272                                APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
1273     return BinaryOperator::CreateAnd(Res, C);
1274   }
1275 
1276   // If this is a TRUNC followed by a ZEXT then we are dealing with integral
1277   // types and if the sizes are just right we can convert this into a logical
1278   // 'and' which will be much cheaper than the pair of casts.
1279   if (auto *CSrc = dyn_cast<TruncInst>(Src)) {   // A->B->C cast
1280     // TODO: Subsume this into EvaluateInDifferentType.
1281 
1282     // Get the sizes of the types involved.  We know that the intermediate type
1283     // will be smaller than A or C, but don't know the relation between A and C.
1284     Value *A = CSrc->getOperand(0);
1285     unsigned SrcSize = A->getType()->getScalarSizeInBits();
1286     unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
1287     unsigned DstSize = DestTy->getScalarSizeInBits();
1288     // If we're actually extending zero bits, then if
1289     // SrcSize <  DstSize: zext(a & mask)
1290     // SrcSize == DstSize: a & mask
1291     // SrcSize  > DstSize: trunc(a) & mask
1292     if (SrcSize < DstSize) {
1293       APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
1294       Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
1295       Value *And = Builder.CreateAnd(A, AndConst, CSrc->getName() + ".mask");
1296       return new ZExtInst(And, DestTy);
1297     }
1298 
1299     if (SrcSize == DstSize) {
1300       APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
1301       return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
1302                                                            AndValue));
1303     }
1304     if (SrcSize > DstSize) {
1305       Value *Trunc = Builder.CreateTrunc(A, DestTy);
1306       APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
1307       return BinaryOperator::CreateAnd(Trunc,
1308                                        ConstantInt::get(Trunc->getType(),
1309                                                         AndValue));
1310     }
1311   }
1312 
1313   if (auto *Cmp = dyn_cast<ICmpInst>(Src))
1314     return transformZExtICmp(Cmp, Zext);
1315 
1316   // zext(trunc(X) & C) -> (X & zext(C)).
1317   Constant *C;
1318   Value *X;
1319   if (match(Src, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
1320       X->getType() == DestTy)
1321     return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, DestTy));
1322 
1323   // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
1324   Value *And;
1325   if (match(Src, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
1326       match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
1327       X->getType() == DestTy) {
1328     Constant *ZC = ConstantExpr::getZExt(C, DestTy);
1329     return BinaryOperator::CreateXor(Builder.CreateAnd(X, ZC), ZC);
1330   }
1331 
1332   // If we are truncating, masking, and then zexting back to the original type,
1333   // that's just a mask. This is not handled by canEvaluateZextd if the
1334   // intermediate values have extra uses. This could be generalized further for
1335   // a non-constant mask operand.
1336   // zext (and (trunc X), C) --> and X, (zext C)
1337   if (match(Src, m_And(m_Trunc(m_Value(X)), m_Constant(C))) &&
1338       X->getType() == DestTy) {
1339     Constant *ZextC = ConstantExpr::getZExt(C, DestTy);
1340     return BinaryOperator::CreateAnd(X, ZextC);
1341   }
1342 
1343   if (match(Src, m_VScale(DL))) {
1344     if (Zext.getFunction() &&
1345         Zext.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
1346       Attribute Attr =
1347           Zext.getFunction()->getFnAttribute(Attribute::VScaleRange);
1348       if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax()) {
1349         unsigned TypeWidth = Src->getType()->getScalarSizeInBits();
1350         if (Log2_32(*MaxVScale) < TypeWidth) {
1351           Value *VScale = Builder.CreateVScale(ConstantInt::get(DestTy, 1));
1352           return replaceInstUsesWith(Zext, VScale);
1353         }
1354       }
1355     }
1356   }
1357 
1358   return nullptr;
1359 }
1360 
1361 /// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp.
1362 Instruction *InstCombinerImpl::transformSExtICmp(ICmpInst *Cmp,
1363                                                  SExtInst &Sext) {
1364   Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1365   ICmpInst::Predicate Pred = Cmp->getPredicate();
1366 
1367   // Don't bother if Op1 isn't of vector or integer type.
1368   if (!Op1->getType()->isIntOrIntVectorTy())
1369     return nullptr;
1370 
1371   if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_ZeroInt())) {
1372     // sext (x <s 0) --> ashr x, 31 (all ones if negative)
1373     Value *Sh = ConstantInt::get(Op0->getType(),
1374                                  Op0->getType()->getScalarSizeInBits() - 1);
1375     Value *In = Builder.CreateAShr(Op0, Sh, Op0->getName() + ".lobit");
1376     if (In->getType() != Sext.getType())
1377       In = Builder.CreateIntCast(In, Sext.getType(), true /*SExt*/);
1378 
1379     return replaceInstUsesWith(Sext, In);
1380   }
1381 
1382   if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1383     // If we know that only one bit of the LHS of the icmp can be set and we
1384     // have an equality comparison with zero or a power of 2, we can transform
1385     // the icmp and sext into bitwise/integer operations.
1386     if (Cmp->hasOneUse() &&
1387         Cmp->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
1388       KnownBits Known = computeKnownBits(Op0, 0, &Sext);
1389 
1390       APInt KnownZeroMask(~Known.Zero);
1391       if (KnownZeroMask.isPowerOf2()) {
1392         Value *In = Cmp->getOperand(0);
1393 
1394         // If the icmp tests for a known zero bit we can constant fold it.
1395         if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
1396           Value *V = Pred == ICmpInst::ICMP_NE ?
1397                        ConstantInt::getAllOnesValue(Sext.getType()) :
1398                        ConstantInt::getNullValue(Sext.getType());
1399           return replaceInstUsesWith(Sext, V);
1400         }
1401 
1402         if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
1403           // sext ((x & 2^n) == 0)   -> (x >> n) - 1
1404           // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
1405           unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
1406           // Perform a right shift to place the desired bit in the LSB.
1407           if (ShiftAmt)
1408             In = Builder.CreateLShr(In,
1409                                     ConstantInt::get(In->getType(), ShiftAmt));
1410 
1411           // At this point "In" is either 1 or 0. Subtract 1 to turn
1412           // {1, 0} -> {0, -1}.
1413           In = Builder.CreateAdd(In,
1414                                  ConstantInt::getAllOnesValue(In->getType()),
1415                                  "sext");
1416         } else {
1417           // sext ((x & 2^n) != 0)   -> (x << bitwidth-n) a>> bitwidth-1
1418           // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
1419           unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
1420           // Perform a left shift to place the desired bit in the MSB.
1421           if (ShiftAmt)
1422             In = Builder.CreateShl(In,
1423                                    ConstantInt::get(In->getType(), ShiftAmt));
1424 
1425           // Distribute the bit over the whole bit width.
1426           In = Builder.CreateAShr(In, ConstantInt::get(In->getType(),
1427                                   KnownZeroMask.getBitWidth() - 1), "sext");
1428         }
1429 
1430         if (Sext.getType() == In->getType())
1431           return replaceInstUsesWith(Sext, In);
1432         return CastInst::CreateIntegerCast(In, Sext.getType(), true/*SExt*/);
1433       }
1434     }
1435   }
1436 
1437   return nullptr;
1438 }
1439 
1440 /// Return true if we can take the specified value and return it as type Ty
1441 /// without inserting any new casts and without changing the value of the common
1442 /// low bits.  This is used by code that tries to promote integer operations to
1443 /// a wider types will allow us to eliminate the extension.
1444 ///
1445 /// This function works on both vectors and scalars.
1446 ///
1447 static bool canEvaluateSExtd(Value *V, Type *Ty) {
1448   assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
1449          "Can't sign extend type to a smaller type");
1450   if (canAlwaysEvaluateInType(V, Ty))
1451     return true;
1452   if (canNotEvaluateInType(V, Ty))
1453     return false;
1454 
1455   auto *I = cast<Instruction>(V);
1456   switch (I->getOpcode()) {
1457   case Instruction::SExt:  // sext(sext(x)) -> sext(x)
1458   case Instruction::ZExt:  // sext(zext(x)) -> zext(x)
1459   case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1460     return true;
1461   case Instruction::And:
1462   case Instruction::Or:
1463   case Instruction::Xor:
1464   case Instruction::Add:
1465   case Instruction::Sub:
1466   case Instruction::Mul:
1467     // These operators can all arbitrarily be extended if their inputs can.
1468     return canEvaluateSExtd(I->getOperand(0), Ty) &&
1469            canEvaluateSExtd(I->getOperand(1), Ty);
1470 
1471   //case Instruction::Shl:   TODO
1472   //case Instruction::LShr:  TODO
1473 
1474   case Instruction::Select:
1475     return canEvaluateSExtd(I->getOperand(1), Ty) &&
1476            canEvaluateSExtd(I->getOperand(2), Ty);
1477 
1478   case Instruction::PHI: {
1479     // We can change a phi if we can change all operands.  Note that we never
1480     // get into trouble with cyclic PHIs here because we only consider
1481     // instructions with a single use.
1482     PHINode *PN = cast<PHINode>(I);
1483     for (Value *IncValue : PN->incoming_values())
1484       if (!canEvaluateSExtd(IncValue, Ty)) return false;
1485     return true;
1486   }
1487   default:
1488     // TODO: Can handle more cases here.
1489     break;
1490   }
1491 
1492   return false;
1493 }
1494 
1495 Instruction *InstCombinerImpl::visitSExt(SExtInst &Sext) {
1496   // If this sign extend is only used by a truncate, let the truncate be
1497   // eliminated before we try to optimize this sext.
1498   if (Sext.hasOneUse() && isa<TruncInst>(Sext.user_back()))
1499     return nullptr;
1500 
1501   if (Instruction *I = commonCastTransforms(Sext))
1502     return I;
1503 
1504   Value *Src = Sext.getOperand(0);
1505   Type *SrcTy = Src->getType(), *DestTy = Sext.getType();
1506   unsigned SrcBitSize = SrcTy->getScalarSizeInBits();
1507   unsigned DestBitSize = DestTy->getScalarSizeInBits();
1508 
1509   // If the value being extended is zero or positive, use a zext instead.
1510   if (isKnownNonNegative(Src, DL, 0, &AC, &Sext, &DT))
1511     return CastInst::Create(Instruction::ZExt, Src, DestTy);
1512 
1513   // Try to extend the entire expression tree to the wide destination type.
1514   if (shouldChangeType(SrcTy, DestTy) && canEvaluateSExtd(Src, DestTy)) {
1515     // Okay, we can transform this!  Insert the new expression now.
1516     LLVM_DEBUG(
1517         dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1518                   " to avoid sign extend: "
1519                << Sext << '\n');
1520     Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1521     assert(Res->getType() == DestTy);
1522 
1523     // If the high bits are already filled with sign bit, just replace this
1524     // cast with the result.
1525     if (ComputeNumSignBits(Res, 0, &Sext) > DestBitSize - SrcBitSize)
1526       return replaceInstUsesWith(Sext, Res);
1527 
1528     // We need to emit a shl + ashr to do the sign extend.
1529     Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1530     return BinaryOperator::CreateAShr(Builder.CreateShl(Res, ShAmt, "sext"),
1531                                       ShAmt);
1532   }
1533 
1534   Value *X;
1535   if (match(Src, m_Trunc(m_Value(X)))) {
1536     // If the input has more sign bits than bits truncated, then convert
1537     // directly to final type.
1538     unsigned XBitSize = X->getType()->getScalarSizeInBits();
1539     if (ComputeNumSignBits(X, 0, &Sext) > XBitSize - SrcBitSize)
1540       return CastInst::CreateIntegerCast(X, DestTy, /* isSigned */ true);
1541 
1542     // If input is a trunc from the destination type, then convert into shifts.
1543     if (Src->hasOneUse() && X->getType() == DestTy) {
1544       // sext (trunc X) --> ashr (shl X, C), C
1545       Constant *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
1546       return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShAmt), ShAmt);
1547     }
1548 
1549     // If we are replacing shifted-in high zero bits with sign bits, convert
1550     // the logic shift to arithmetic shift and eliminate the cast to
1551     // intermediate type:
1552     // sext (trunc (lshr Y, C)) --> sext/trunc (ashr Y, C)
1553     Value *Y;
1554     if (Src->hasOneUse() &&
1555         match(X, m_LShr(m_Value(Y),
1556                         m_SpecificIntAllowUndef(XBitSize - SrcBitSize)))) {
1557       Value *Ashr = Builder.CreateAShr(Y, XBitSize - SrcBitSize);
1558       return CastInst::CreateIntegerCast(Ashr, DestTy, /* isSigned */ true);
1559     }
1560   }
1561 
1562   if (auto *Cmp = dyn_cast<ICmpInst>(Src))
1563     return transformSExtICmp(Cmp, Sext);
1564 
1565   // If the input is a shl/ashr pair of a same constant, then this is a sign
1566   // extension from a smaller value.  If we could trust arbitrary bitwidth
1567   // integers, we could turn this into a truncate to the smaller bit and then
1568   // use a sext for the whole extension.  Since we don't, look deeper and check
1569   // for a truncate.  If the source and dest are the same type, eliminate the
1570   // trunc and extend and just do shifts.  For example, turn:
1571   //   %a = trunc i32 %i to i8
1572   //   %b = shl i8 %a, C
1573   //   %c = ashr i8 %b, C
1574   //   %d = sext i8 %c to i32
1575   // into:
1576   //   %a = shl i32 %i, 32-(8-C)
1577   //   %d = ashr i32 %a, 32-(8-C)
1578   Value *A = nullptr;
1579   // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1580   Constant *BA = nullptr, *CA = nullptr;
1581   if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_Constant(BA)),
1582                         m_Constant(CA))) &&
1583       BA->isElementWiseEqual(CA) && A->getType() == DestTy) {
1584     Constant *WideCurrShAmt = ConstantExpr::getSExt(CA, DestTy);
1585     Constant *NumLowbitsLeft = ConstantExpr::getSub(
1586         ConstantInt::get(DestTy, SrcTy->getScalarSizeInBits()), WideCurrShAmt);
1587     Constant *NewShAmt = ConstantExpr::getSub(
1588         ConstantInt::get(DestTy, DestTy->getScalarSizeInBits()),
1589         NumLowbitsLeft);
1590     NewShAmt =
1591         Constant::mergeUndefsWith(Constant::mergeUndefsWith(NewShAmt, BA), CA);
1592     A = Builder.CreateShl(A, NewShAmt, Sext.getName());
1593     return BinaryOperator::CreateAShr(A, NewShAmt);
1594   }
1595 
1596   // Splatting a bit of constant-index across a value:
1597   // sext (ashr (trunc iN X to iM), M-1) to iN --> ashr (shl X, N-M), N-1
1598   // If the dest type is different, use a cast (adjust use check).
1599   if (match(Src, m_OneUse(m_AShr(m_Trunc(m_Value(X)),
1600                                  m_SpecificInt(SrcBitSize - 1))))) {
1601     Type *XTy = X->getType();
1602     unsigned XBitSize = XTy->getScalarSizeInBits();
1603     Constant *ShlAmtC = ConstantInt::get(XTy, XBitSize - SrcBitSize);
1604     Constant *AshrAmtC = ConstantInt::get(XTy, XBitSize - 1);
1605     if (XTy == DestTy)
1606       return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShlAmtC),
1607                                         AshrAmtC);
1608     if (cast<BinaryOperator>(Src)->getOperand(0)->hasOneUse()) {
1609       Value *Ashr = Builder.CreateAShr(Builder.CreateShl(X, ShlAmtC), AshrAmtC);
1610       return CastInst::CreateIntegerCast(Ashr, DestTy, /* isSigned */ true);
1611     }
1612   }
1613 
1614   if (match(Src, m_VScale(DL))) {
1615     if (Sext.getFunction() &&
1616         Sext.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
1617       Attribute Attr =
1618           Sext.getFunction()->getFnAttribute(Attribute::VScaleRange);
1619       if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax()) {
1620         if (Log2_32(*MaxVScale) < (SrcBitSize - 1)) {
1621           Value *VScale = Builder.CreateVScale(ConstantInt::get(DestTy, 1));
1622           return replaceInstUsesWith(Sext, VScale);
1623         }
1624       }
1625     }
1626   }
1627 
1628   return nullptr;
1629 }
1630 
1631 /// Return a Constant* for the specified floating-point constant if it fits
1632 /// in the specified FP type without changing its value.
1633 static bool fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1634   bool losesInfo;
1635   APFloat F = CFP->getValueAPF();
1636   (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1637   return !losesInfo;
1638 }
1639 
1640 static Type *shrinkFPConstant(ConstantFP *CFP) {
1641   if (CFP->getType() == Type::getPPC_FP128Ty(CFP->getContext()))
1642     return nullptr;  // No constant folding of this.
1643   // See if the value can be truncated to half and then reextended.
1644   if (fitsInFPType(CFP, APFloat::IEEEhalf()))
1645     return Type::getHalfTy(CFP->getContext());
1646   // See if the value can be truncated to float and then reextended.
1647   if (fitsInFPType(CFP, APFloat::IEEEsingle()))
1648     return Type::getFloatTy(CFP->getContext());
1649   if (CFP->getType()->isDoubleTy())
1650     return nullptr;  // Won't shrink.
1651   if (fitsInFPType(CFP, APFloat::IEEEdouble()))
1652     return Type::getDoubleTy(CFP->getContext());
1653   // Don't try to shrink to various long double types.
1654   return nullptr;
1655 }
1656 
1657 // Determine if this is a vector of ConstantFPs and if so, return the minimal
1658 // type we can safely truncate all elements to.
1659 static Type *shrinkFPConstantVector(Value *V) {
1660   auto *CV = dyn_cast<Constant>(V);
1661   auto *CVVTy = dyn_cast<FixedVectorType>(V->getType());
1662   if (!CV || !CVVTy)
1663     return nullptr;
1664 
1665   Type *MinType = nullptr;
1666 
1667   unsigned NumElts = CVVTy->getNumElements();
1668 
1669   // For fixed-width vectors we find the minimal type by looking
1670   // through the constant values of the vector.
1671   for (unsigned i = 0; i != NumElts; ++i) {
1672     if (isa<UndefValue>(CV->getAggregateElement(i)))
1673       continue;
1674 
1675     auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
1676     if (!CFP)
1677       return nullptr;
1678 
1679     Type *T = shrinkFPConstant(CFP);
1680     if (!T)
1681       return nullptr;
1682 
1683     // If we haven't found a type yet or this type has a larger mantissa than
1684     // our previous type, this is our new minimal type.
1685     if (!MinType || T->getFPMantissaWidth() > MinType->getFPMantissaWidth())
1686       MinType = T;
1687   }
1688 
1689   // Make a vector type from the minimal type.
1690   return MinType ? FixedVectorType::get(MinType, NumElts) : nullptr;
1691 }
1692 
1693 /// Find the minimum FP type we can safely truncate to.
1694 static Type *getMinimumFPType(Value *V) {
1695   if (auto *FPExt = dyn_cast<FPExtInst>(V))
1696     return FPExt->getOperand(0)->getType();
1697 
1698   // If this value is a constant, return the constant in the smallest FP type
1699   // that can accurately represent it.  This allows us to turn
1700   // (float)((double)X+2.0) into x+2.0f.
1701   if (auto *CFP = dyn_cast<ConstantFP>(V))
1702     if (Type *T = shrinkFPConstant(CFP))
1703       return T;
1704 
1705   // We can only correctly find a minimum type for a scalable vector when it is
1706   // a splat. For splats of constant values the fpext is wrapped up as a
1707   // ConstantExpr.
1708   if (auto *FPCExt = dyn_cast<ConstantExpr>(V))
1709     if (FPCExt->getOpcode() == Instruction::FPExt)
1710       return FPCExt->getOperand(0)->getType();
1711 
1712   // Try to shrink a vector of FP constants. This returns nullptr on scalable
1713   // vectors
1714   if (Type *T = shrinkFPConstantVector(V))
1715     return T;
1716 
1717   return V->getType();
1718 }
1719 
1720 /// Return true if the cast from integer to FP can be proven to be exact for all
1721 /// possible inputs (the conversion does not lose any precision).
1722 static bool isKnownExactCastIntToFP(CastInst &I, InstCombinerImpl &IC) {
1723   CastInst::CastOps Opcode = I.getOpcode();
1724   assert((Opcode == CastInst::SIToFP || Opcode == CastInst::UIToFP) &&
1725          "Unexpected cast");
1726   Value *Src = I.getOperand(0);
1727   Type *SrcTy = Src->getType();
1728   Type *FPTy = I.getType();
1729   bool IsSigned = Opcode == Instruction::SIToFP;
1730   int SrcSize = (int)SrcTy->getScalarSizeInBits() - IsSigned;
1731 
1732   // Easy case - if the source integer type has less bits than the FP mantissa,
1733   // then the cast must be exact.
1734   int DestNumSigBits = FPTy->getFPMantissaWidth();
1735   if (SrcSize <= DestNumSigBits)
1736     return true;
1737 
1738   // Cast from FP to integer and back to FP is independent of the intermediate
1739   // integer width because of poison on overflow.
1740   Value *F;
1741   if (match(Src, m_FPToSI(m_Value(F))) || match(Src, m_FPToUI(m_Value(F)))) {
1742     // If this is uitofp (fptosi F), the source needs an extra bit to avoid
1743     // potential rounding of negative FP input values.
1744     int SrcNumSigBits = F->getType()->getFPMantissaWidth();
1745     if (!IsSigned && match(Src, m_FPToSI(m_Value())))
1746       SrcNumSigBits++;
1747 
1748     // [su]itofp (fpto[su]i F) --> exact if the source type has less or equal
1749     // significant bits than the destination (and make sure neither type is
1750     // weird -- ppc_fp128).
1751     if (SrcNumSigBits > 0 && DestNumSigBits > 0 &&
1752         SrcNumSigBits <= DestNumSigBits)
1753       return true;
1754   }
1755 
1756   // TODO:
1757   // Try harder to find if the source integer type has less significant bits.
1758   // For example, compute number of sign bits.
1759   KnownBits SrcKnown = IC.computeKnownBits(Src, 0, &I);
1760   int SigBits = (int)SrcTy->getScalarSizeInBits() -
1761                 SrcKnown.countMinLeadingZeros() -
1762                 SrcKnown.countMinTrailingZeros();
1763   if (SigBits <= DestNumSigBits)
1764     return true;
1765 
1766   return false;
1767 }
1768 
1769 Instruction *InstCombinerImpl::visitFPTrunc(FPTruncInst &FPT) {
1770   if (Instruction *I = commonCastTransforms(FPT))
1771     return I;
1772 
1773   // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1774   // simplify this expression to avoid one or more of the trunc/extend
1775   // operations if we can do so without changing the numerical results.
1776   //
1777   // The exact manner in which the widths of the operands interact to limit
1778   // what we can and cannot do safely varies from operation to operation, and
1779   // is explained below in the various case statements.
1780   Type *Ty = FPT.getType();
1781   auto *BO = dyn_cast<BinaryOperator>(FPT.getOperand(0));
1782   if (BO && BO->hasOneUse()) {
1783     Type *LHSMinType = getMinimumFPType(BO->getOperand(0));
1784     Type *RHSMinType = getMinimumFPType(BO->getOperand(1));
1785     unsigned OpWidth = BO->getType()->getFPMantissaWidth();
1786     unsigned LHSWidth = LHSMinType->getFPMantissaWidth();
1787     unsigned RHSWidth = RHSMinType->getFPMantissaWidth();
1788     unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1789     unsigned DstWidth = Ty->getFPMantissaWidth();
1790     switch (BO->getOpcode()) {
1791       default: break;
1792       case Instruction::FAdd:
1793       case Instruction::FSub:
1794         // For addition and subtraction, the infinitely precise result can
1795         // essentially be arbitrarily wide; proving that double rounding
1796         // will not occur because the result of OpI is exact (as we will for
1797         // FMul, for example) is hopeless.  However, we *can* nonetheless
1798         // frequently know that double rounding cannot occur (or that it is
1799         // innocuous) by taking advantage of the specific structure of
1800         // infinitely-precise results that admit double rounding.
1801         //
1802         // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1803         // to represent both sources, we can guarantee that the double
1804         // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1805         // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1806         // for proof of this fact).
1807         //
1808         // Note: Figueroa does not consider the case where DstFormat !=
1809         // SrcFormat.  It's possible (likely even!) that this analysis
1810         // could be tightened for those cases, but they are rare (the main
1811         // case of interest here is (float)((double)float + float)).
1812         if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1813           Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);
1814           Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);
1815           Instruction *RI = BinaryOperator::Create(BO->getOpcode(), LHS, RHS);
1816           RI->copyFastMathFlags(BO);
1817           return RI;
1818         }
1819         break;
1820       case Instruction::FMul:
1821         // For multiplication, the infinitely precise result has at most
1822         // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1823         // that such a value can be exactly represented, then no double
1824         // rounding can possibly occur; we can safely perform the operation
1825         // in the destination format if it can represent both sources.
1826         if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1827           Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);
1828           Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);
1829           return BinaryOperator::CreateFMulFMF(LHS, RHS, BO);
1830         }
1831         break;
1832       case Instruction::FDiv:
1833         // For division, we use again use the bound from Figueroa's
1834         // dissertation.  I am entirely certain that this bound can be
1835         // tightened in the unbalanced operand case by an analysis based on
1836         // the diophantine rational approximation bound, but the well-known
1837         // condition used here is a good conservative first pass.
1838         // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1839         if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1840           Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);
1841           Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);
1842           return BinaryOperator::CreateFDivFMF(LHS, RHS, BO);
1843         }
1844         break;
1845       case Instruction::FRem: {
1846         // Remainder is straightforward.  Remainder is always exact, so the
1847         // type of OpI doesn't enter into things at all.  We simply evaluate
1848         // in whichever source type is larger, then convert to the
1849         // destination type.
1850         if (SrcWidth == OpWidth)
1851           break;
1852         Value *LHS, *RHS;
1853         if (LHSWidth == SrcWidth) {
1854            LHS = Builder.CreateFPTrunc(BO->getOperand(0), LHSMinType);
1855            RHS = Builder.CreateFPTrunc(BO->getOperand(1), LHSMinType);
1856         } else {
1857            LHS = Builder.CreateFPTrunc(BO->getOperand(0), RHSMinType);
1858            RHS = Builder.CreateFPTrunc(BO->getOperand(1), RHSMinType);
1859         }
1860 
1861         Value *ExactResult = Builder.CreateFRemFMF(LHS, RHS, BO);
1862         return CastInst::CreateFPCast(ExactResult, Ty);
1863       }
1864     }
1865   }
1866 
1867   // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1868   Value *X;
1869   Instruction *Op = dyn_cast<Instruction>(FPT.getOperand(0));
1870   if (Op && Op->hasOneUse()) {
1871     // FIXME: The FMF should propagate from the fptrunc, not the source op.
1872     IRBuilder<>::FastMathFlagGuard FMFG(Builder);
1873     if (isa<FPMathOperator>(Op))
1874       Builder.setFastMathFlags(Op->getFastMathFlags());
1875 
1876     if (match(Op, m_FNeg(m_Value(X)))) {
1877       Value *InnerTrunc = Builder.CreateFPTrunc(X, Ty);
1878 
1879       return UnaryOperator::CreateFNegFMF(InnerTrunc, Op);
1880     }
1881 
1882     // If we are truncating a select that has an extended operand, we can
1883     // narrow the other operand and do the select as a narrow op.
1884     Value *Cond, *X, *Y;
1885     if (match(Op, m_Select(m_Value(Cond), m_FPExt(m_Value(X)), m_Value(Y))) &&
1886         X->getType() == Ty) {
1887       // fptrunc (select Cond, (fpext X), Y --> select Cond, X, (fptrunc Y)
1888       Value *NarrowY = Builder.CreateFPTrunc(Y, Ty);
1889       Value *Sel = Builder.CreateSelect(Cond, X, NarrowY, "narrow.sel", Op);
1890       return replaceInstUsesWith(FPT, Sel);
1891     }
1892     if (match(Op, m_Select(m_Value(Cond), m_Value(Y), m_FPExt(m_Value(X)))) &&
1893         X->getType() == Ty) {
1894       // fptrunc (select Cond, Y, (fpext X) --> select Cond, (fptrunc Y), X
1895       Value *NarrowY = Builder.CreateFPTrunc(Y, Ty);
1896       Value *Sel = Builder.CreateSelect(Cond, NarrowY, X, "narrow.sel", Op);
1897       return replaceInstUsesWith(FPT, Sel);
1898     }
1899   }
1900 
1901   if (auto *II = dyn_cast<IntrinsicInst>(FPT.getOperand(0))) {
1902     switch (II->getIntrinsicID()) {
1903     default: break;
1904     case Intrinsic::ceil:
1905     case Intrinsic::fabs:
1906     case Intrinsic::floor:
1907     case Intrinsic::nearbyint:
1908     case Intrinsic::rint:
1909     case Intrinsic::round:
1910     case Intrinsic::roundeven:
1911     case Intrinsic::trunc: {
1912       Value *Src = II->getArgOperand(0);
1913       if (!Src->hasOneUse())
1914         break;
1915 
1916       // Except for fabs, this transformation requires the input of the unary FP
1917       // operation to be itself an fpext from the type to which we're
1918       // truncating.
1919       if (II->getIntrinsicID() != Intrinsic::fabs) {
1920         FPExtInst *FPExtSrc = dyn_cast<FPExtInst>(Src);
1921         if (!FPExtSrc || FPExtSrc->getSrcTy() != Ty)
1922           break;
1923       }
1924 
1925       // Do unary FP operation on smaller type.
1926       // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1927       Value *InnerTrunc = Builder.CreateFPTrunc(Src, Ty);
1928       Function *Overload = Intrinsic::getDeclaration(FPT.getModule(),
1929                                                      II->getIntrinsicID(), Ty);
1930       SmallVector<OperandBundleDef, 1> OpBundles;
1931       II->getOperandBundlesAsDefs(OpBundles);
1932       CallInst *NewCI =
1933           CallInst::Create(Overload, {InnerTrunc}, OpBundles, II->getName());
1934       NewCI->copyFastMathFlags(II);
1935       return NewCI;
1936     }
1937     }
1938   }
1939 
1940   if (Instruction *I = shrinkInsertElt(FPT, Builder))
1941     return I;
1942 
1943   Value *Src = FPT.getOperand(0);
1944   if (isa<SIToFPInst>(Src) || isa<UIToFPInst>(Src)) {
1945     auto *FPCast = cast<CastInst>(Src);
1946     if (isKnownExactCastIntToFP(*FPCast, *this))
1947       return CastInst::Create(FPCast->getOpcode(), FPCast->getOperand(0), Ty);
1948   }
1949 
1950   return nullptr;
1951 }
1952 
1953 Instruction *InstCombinerImpl::visitFPExt(CastInst &FPExt) {
1954   // If the source operand is a cast from integer to FP and known exact, then
1955   // cast the integer operand directly to the destination type.
1956   Type *Ty = FPExt.getType();
1957   Value *Src = FPExt.getOperand(0);
1958   if (isa<SIToFPInst>(Src) || isa<UIToFPInst>(Src)) {
1959     auto *FPCast = cast<CastInst>(Src);
1960     if (isKnownExactCastIntToFP(*FPCast, *this))
1961       return CastInst::Create(FPCast->getOpcode(), FPCast->getOperand(0), Ty);
1962   }
1963 
1964   return commonCastTransforms(FPExt);
1965 }
1966 
1967 /// fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)
1968 /// This is safe if the intermediate type has enough bits in its mantissa to
1969 /// accurately represent all values of X.  For example, this won't work with
1970 /// i64 -> float -> i64.
1971 Instruction *InstCombinerImpl::foldItoFPtoI(CastInst &FI) {
1972   if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))
1973     return nullptr;
1974 
1975   auto *OpI = cast<CastInst>(FI.getOperand(0));
1976   Value *X = OpI->getOperand(0);
1977   Type *XType = X->getType();
1978   Type *DestType = FI.getType();
1979   bool IsOutputSigned = isa<FPToSIInst>(FI);
1980 
1981   // Since we can assume the conversion won't overflow, our decision as to
1982   // whether the input will fit in the float should depend on the minimum
1983   // of the input range and output range.
1984 
1985   // This means this is also safe for a signed input and unsigned output, since
1986   // a negative input would lead to undefined behavior.
1987   if (!isKnownExactCastIntToFP(*OpI, *this)) {
1988     // The first cast may not round exactly based on the source integer width
1989     // and FP width, but the overflow UB rules can still allow this to fold.
1990     // If the destination type is narrow, that means the intermediate FP value
1991     // must be large enough to hold the source value exactly.
1992     // For example, (uint8_t)((float)(uint32_t 16777217) is undefined behavior.
1993     int OutputSize = (int)DestType->getScalarSizeInBits();
1994     if (OutputSize > OpI->getType()->getFPMantissaWidth())
1995       return nullptr;
1996   }
1997 
1998   if (DestType->getScalarSizeInBits() > XType->getScalarSizeInBits()) {
1999     bool IsInputSigned = isa<SIToFPInst>(OpI);
2000     if (IsInputSigned && IsOutputSigned)
2001       return new SExtInst(X, DestType);
2002     return new ZExtInst(X, DestType);
2003   }
2004   if (DestType->getScalarSizeInBits() < XType->getScalarSizeInBits())
2005     return new TruncInst(X, DestType);
2006 
2007   assert(XType == DestType && "Unexpected types for int to FP to int casts");
2008   return replaceInstUsesWith(FI, X);
2009 }
2010 
2011 Instruction *InstCombinerImpl::visitFPToUI(FPToUIInst &FI) {
2012   if (Instruction *I = foldItoFPtoI(FI))
2013     return I;
2014 
2015   return commonCastTransforms(FI);
2016 }
2017 
2018 Instruction *InstCombinerImpl::visitFPToSI(FPToSIInst &FI) {
2019   if (Instruction *I = foldItoFPtoI(FI))
2020     return I;
2021 
2022   return commonCastTransforms(FI);
2023 }
2024 
2025 Instruction *InstCombinerImpl::visitUIToFP(CastInst &CI) {
2026   return commonCastTransforms(CI);
2027 }
2028 
2029 Instruction *InstCombinerImpl::visitSIToFP(CastInst &CI) {
2030   return commonCastTransforms(CI);
2031 }
2032 
2033 Instruction *InstCombinerImpl::visitIntToPtr(IntToPtrInst &CI) {
2034   // If the source integer type is not the intptr_t type for this target, do a
2035   // trunc or zext to the intptr_t type, then inttoptr of it.  This allows the
2036   // cast to be exposed to other transforms.
2037   unsigned AS = CI.getAddressSpace();
2038   if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
2039       DL.getPointerSizeInBits(AS)) {
2040     Type *Ty = CI.getOperand(0)->getType()->getWithNewType(
2041         DL.getIntPtrType(CI.getContext(), AS));
2042     Value *P = Builder.CreateZExtOrTrunc(CI.getOperand(0), Ty);
2043     return new IntToPtrInst(P, CI.getType());
2044   }
2045 
2046   if (Instruction *I = commonCastTransforms(CI))
2047     return I;
2048 
2049   return nullptr;
2050 }
2051 
2052 /// Implement the transforms for cast of pointer (bitcast/ptrtoint)
2053 Instruction *InstCombinerImpl::commonPointerCastTransforms(CastInst &CI) {
2054   Value *Src = CI.getOperand(0);
2055 
2056   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
2057     // If casting the result of a getelementptr instruction with no offset, turn
2058     // this into a cast of the original pointer!
2059     if (GEP->hasAllZeroIndices() &&
2060         // If CI is an addrspacecast and GEP changes the poiner type, merging
2061         // GEP into CI would undo canonicalizing addrspacecast with different
2062         // pointer types, causing infinite loops.
2063         (!isa<AddrSpaceCastInst>(CI) ||
2064          GEP->getType() == GEP->getPointerOperandType())) {
2065       // Changing the cast operand is usually not a good idea but it is safe
2066       // here because the pointer operand is being replaced with another
2067       // pointer operand so the opcode doesn't need to change.
2068       return replaceOperand(CI, 0, GEP->getOperand(0));
2069     }
2070   }
2071 
2072   return commonCastTransforms(CI);
2073 }
2074 
2075 Instruction *InstCombinerImpl::visitPtrToInt(PtrToIntInst &CI) {
2076   // If the destination integer type is not the intptr_t type for this target,
2077   // do a ptrtoint to intptr_t then do a trunc or zext.  This allows the cast
2078   // to be exposed to other transforms.
2079   Value *SrcOp = CI.getPointerOperand();
2080   Type *SrcTy = SrcOp->getType();
2081   Type *Ty = CI.getType();
2082   unsigned AS = CI.getPointerAddressSpace();
2083   unsigned TySize = Ty->getScalarSizeInBits();
2084   unsigned PtrSize = DL.getPointerSizeInBits(AS);
2085   if (TySize != PtrSize) {
2086     Type *IntPtrTy =
2087         SrcTy->getWithNewType(DL.getIntPtrType(CI.getContext(), AS));
2088     Value *P = Builder.CreatePtrToInt(SrcOp, IntPtrTy);
2089     return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
2090   }
2091 
2092   if (auto *GEP = dyn_cast<GetElementPtrInst>(SrcOp)) {
2093     // Fold ptrtoint(gep null, x) to multiply + constant if the GEP has one use.
2094     // While this can increase the number of instructions it doesn't actually
2095     // increase the overall complexity since the arithmetic is just part of
2096     // the GEP otherwise.
2097     if (GEP->hasOneUse() &&
2098         isa<ConstantPointerNull>(GEP->getPointerOperand())) {
2099       return replaceInstUsesWith(CI,
2100                                  Builder.CreateIntCast(EmitGEPOffset(GEP), Ty,
2101                                                        /*isSigned=*/false));
2102     }
2103   }
2104 
2105   Value *Vec, *Scalar, *Index;
2106   if (match(SrcOp, m_OneUse(m_InsertElt(m_IntToPtr(m_Value(Vec)),
2107                                         m_Value(Scalar), m_Value(Index)))) &&
2108       Vec->getType() == Ty) {
2109     assert(Vec->getType()->getScalarSizeInBits() == PtrSize && "Wrong type");
2110     // Convert the scalar to int followed by insert to eliminate one cast:
2111     // p2i (ins (i2p Vec), Scalar, Index --> ins Vec, (p2i Scalar), Index
2112     Value *NewCast = Builder.CreatePtrToInt(Scalar, Ty->getScalarType());
2113     return InsertElementInst::Create(Vec, NewCast, Index);
2114   }
2115 
2116   return commonPointerCastTransforms(CI);
2117 }
2118 
2119 /// This input value (which is known to have vector type) is being zero extended
2120 /// or truncated to the specified vector type. Since the zext/trunc is done
2121 /// using an integer type, we have a (bitcast(cast(bitcast))) pattern,
2122 /// endianness will impact which end of the vector that is extended or
2123 /// truncated.
2124 ///
2125 /// A vector is always stored with index 0 at the lowest address, which
2126 /// corresponds to the most significant bits for a big endian stored integer and
2127 /// the least significant bits for little endian. A trunc/zext of an integer
2128 /// impacts the big end of the integer. Thus, we need to add/remove elements at
2129 /// the front of the vector for big endian targets, and the back of the vector
2130 /// for little endian targets.
2131 ///
2132 /// Try to replace it with a shuffle (and vector/vector bitcast) if possible.
2133 ///
2134 /// The source and destination vector types may have different element types.
2135 static Instruction *
2136 optimizeVectorResizeWithIntegerBitCasts(Value *InVal, VectorType *DestTy,
2137                                         InstCombinerImpl &IC) {
2138   // We can only do this optimization if the output is a multiple of the input
2139   // element size, or the input is a multiple of the output element size.
2140   // Convert the input type to have the same element type as the output.
2141   VectorType *SrcTy = cast<VectorType>(InVal->getType());
2142 
2143   if (SrcTy->getElementType() != DestTy->getElementType()) {
2144     // The input types don't need to be identical, but for now they must be the
2145     // same size.  There is no specific reason we couldn't handle things like
2146     // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
2147     // there yet.
2148     if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
2149         DestTy->getElementType()->getPrimitiveSizeInBits())
2150       return nullptr;
2151 
2152     SrcTy =
2153         FixedVectorType::get(DestTy->getElementType(),
2154                              cast<FixedVectorType>(SrcTy)->getNumElements());
2155     InVal = IC.Builder.CreateBitCast(InVal, SrcTy);
2156   }
2157 
2158   bool IsBigEndian = IC.getDataLayout().isBigEndian();
2159   unsigned SrcElts = cast<FixedVectorType>(SrcTy)->getNumElements();
2160   unsigned DestElts = cast<FixedVectorType>(DestTy)->getNumElements();
2161 
2162   assert(SrcElts != DestElts && "Element counts should be different.");
2163 
2164   // Now that the element types match, get the shuffle mask and RHS of the
2165   // shuffle to use, which depends on whether we're increasing or decreasing the
2166   // size of the input.
2167   auto ShuffleMaskStorage = llvm::to_vector<16>(llvm::seq<int>(0, SrcElts));
2168   ArrayRef<int> ShuffleMask;
2169   Value *V2;
2170 
2171   if (SrcElts > DestElts) {
2172     // If we're shrinking the number of elements (rewriting an integer
2173     // truncate), just shuffle in the elements corresponding to the least
2174     // significant bits from the input and use poison as the second shuffle
2175     // input.
2176     V2 = PoisonValue::get(SrcTy);
2177     // Make sure the shuffle mask selects the "least significant bits" by
2178     // keeping elements from back of the src vector for big endian, and from the
2179     // front for little endian.
2180     ShuffleMask = ShuffleMaskStorage;
2181     if (IsBigEndian)
2182       ShuffleMask = ShuffleMask.take_back(DestElts);
2183     else
2184       ShuffleMask = ShuffleMask.take_front(DestElts);
2185   } else {
2186     // If we're increasing the number of elements (rewriting an integer zext),
2187     // shuffle in all of the elements from InVal. Fill the rest of the result
2188     // elements with zeros from a constant zero.
2189     V2 = Constant::getNullValue(SrcTy);
2190     // Use first elt from V2 when indicating zero in the shuffle mask.
2191     uint32_t NullElt = SrcElts;
2192     // Extend with null values in the "most significant bits" by adding elements
2193     // in front of the src vector for big endian, and at the back for little
2194     // endian.
2195     unsigned DeltaElts = DestElts - SrcElts;
2196     if (IsBigEndian)
2197       ShuffleMaskStorage.insert(ShuffleMaskStorage.begin(), DeltaElts, NullElt);
2198     else
2199       ShuffleMaskStorage.append(DeltaElts, NullElt);
2200     ShuffleMask = ShuffleMaskStorage;
2201   }
2202 
2203   return new ShuffleVectorInst(InVal, V2, ShuffleMask);
2204 }
2205 
2206 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
2207   return Value % Ty->getPrimitiveSizeInBits() == 0;
2208 }
2209 
2210 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
2211   return Value / Ty->getPrimitiveSizeInBits();
2212 }
2213 
2214 /// V is a value which is inserted into a vector of VecEltTy.
2215 /// Look through the value to see if we can decompose it into
2216 /// insertions into the vector.  See the example in the comment for
2217 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
2218 /// The type of V is always a non-zero multiple of VecEltTy's size.
2219 /// Shift is the number of bits between the lsb of V and the lsb of
2220 /// the vector.
2221 ///
2222 /// This returns false if the pattern can't be matched or true if it can,
2223 /// filling in Elements with the elements found here.
2224 static bool collectInsertionElements(Value *V, unsigned Shift,
2225                                      SmallVectorImpl<Value *> &Elements,
2226                                      Type *VecEltTy, bool isBigEndian) {
2227   assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
2228          "Shift should be a multiple of the element type size");
2229 
2230   // Undef values never contribute useful bits to the result.
2231   if (isa<UndefValue>(V)) return true;
2232 
2233   // If we got down to a value of the right type, we win, try inserting into the
2234   // right element.
2235   if (V->getType() == VecEltTy) {
2236     // Inserting null doesn't actually insert any elements.
2237     if (Constant *C = dyn_cast<Constant>(V))
2238       if (C->isNullValue())
2239         return true;
2240 
2241     unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
2242     if (isBigEndian)
2243       ElementIndex = Elements.size() - ElementIndex - 1;
2244 
2245     // Fail if multiple elements are inserted into this slot.
2246     if (Elements[ElementIndex])
2247       return false;
2248 
2249     Elements[ElementIndex] = V;
2250     return true;
2251   }
2252 
2253   if (Constant *C = dyn_cast<Constant>(V)) {
2254     // Figure out the # elements this provides, and bitcast it or slice it up
2255     // as required.
2256     unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
2257                                         VecEltTy);
2258     // If the constant is the size of a vector element, we just need to bitcast
2259     // it to the right type so it gets properly inserted.
2260     if (NumElts == 1)
2261       return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
2262                                       Shift, Elements, VecEltTy, isBigEndian);
2263 
2264     // Okay, this is a constant that covers multiple elements.  Slice it up into
2265     // pieces and insert each element-sized piece into the vector.
2266     if (!isa<IntegerType>(C->getType()))
2267       C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
2268                                        C->getType()->getPrimitiveSizeInBits()));
2269     unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
2270     Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
2271 
2272     for (unsigned i = 0; i != NumElts; ++i) {
2273       unsigned ShiftI = Shift+i*ElementSize;
2274       Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
2275                                                                   ShiftI));
2276       Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
2277       if (!collectInsertionElements(Piece, ShiftI, Elements, VecEltTy,
2278                                     isBigEndian))
2279         return false;
2280     }
2281     return true;
2282   }
2283 
2284   if (!V->hasOneUse()) return false;
2285 
2286   Instruction *I = dyn_cast<Instruction>(V);
2287   if (!I) return false;
2288   switch (I->getOpcode()) {
2289   default: return false; // Unhandled case.
2290   case Instruction::BitCast:
2291     if (I->getOperand(0)->getType()->isVectorTy())
2292       return false;
2293     return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2294                                     isBigEndian);
2295   case Instruction::ZExt:
2296     if (!isMultipleOfTypeSize(
2297                           I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
2298                               VecEltTy))
2299       return false;
2300     return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2301                                     isBigEndian);
2302   case Instruction::Or:
2303     return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2304                                     isBigEndian) &&
2305            collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,
2306                                     isBigEndian);
2307   case Instruction::Shl: {
2308     // Must be shifting by a constant that is a multiple of the element size.
2309     ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
2310     if (!CI) return false;
2311     Shift += CI->getZExtValue();
2312     if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
2313     return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2314                                     isBigEndian);
2315   }
2316 
2317   }
2318 }
2319 
2320 
2321 /// If the input is an 'or' instruction, we may be doing shifts and ors to
2322 /// assemble the elements of the vector manually.
2323 /// Try to rip the code out and replace it with insertelements.  This is to
2324 /// optimize code like this:
2325 ///
2326 ///    %tmp37 = bitcast float %inc to i32
2327 ///    %tmp38 = zext i32 %tmp37 to i64
2328 ///    %tmp31 = bitcast float %inc5 to i32
2329 ///    %tmp32 = zext i32 %tmp31 to i64
2330 ///    %tmp33 = shl i64 %tmp32, 32
2331 ///    %ins35 = or i64 %tmp33, %tmp38
2332 ///    %tmp43 = bitcast i64 %ins35 to <2 x float>
2333 ///
2334 /// Into two insertelements that do "buildvector{%inc, %inc5}".
2335 static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,
2336                                                 InstCombinerImpl &IC) {
2337   auto *DestVecTy = cast<FixedVectorType>(CI.getType());
2338   Value *IntInput = CI.getOperand(0);
2339 
2340   SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
2341   if (!collectInsertionElements(IntInput, 0, Elements,
2342                                 DestVecTy->getElementType(),
2343                                 IC.getDataLayout().isBigEndian()))
2344     return nullptr;
2345 
2346   // If we succeeded, we know that all of the element are specified by Elements
2347   // or are zero if Elements has a null entry.  Recast this as a set of
2348   // insertions.
2349   Value *Result = Constant::getNullValue(CI.getType());
2350   for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
2351     if (!Elements[i]) continue;  // Unset element.
2352 
2353     Result = IC.Builder.CreateInsertElement(Result, Elements[i],
2354                                             IC.Builder.getInt32(i));
2355   }
2356 
2357   return Result;
2358 }
2359 
2360 /// Canonicalize scalar bitcasts of extracted elements into a bitcast of the
2361 /// vector followed by extract element. The backend tends to handle bitcasts of
2362 /// vectors better than bitcasts of scalars because vector registers are
2363 /// usually not type-specific like scalar integer or scalar floating-point.
2364 static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast,
2365                                               InstCombinerImpl &IC) {
2366   Value *VecOp, *Index;
2367   if (!match(BitCast.getOperand(0),
2368              m_OneUse(m_ExtractElt(m_Value(VecOp), m_Value(Index)))))
2369     return nullptr;
2370 
2371   // The bitcast must be to a vectorizable type, otherwise we can't make a new
2372   // type to extract from.
2373   Type *DestType = BitCast.getType();
2374   VectorType *VecType = cast<VectorType>(VecOp->getType());
2375   if (VectorType::isValidElementType(DestType)) {
2376     auto *NewVecType = VectorType::get(DestType, VecType);
2377     auto *NewBC = IC.Builder.CreateBitCast(VecOp, NewVecType, "bc");
2378     return ExtractElementInst::Create(NewBC, Index);
2379   }
2380 
2381   // Only solve DestType is vector to avoid inverse transform in visitBitCast.
2382   // bitcast (extractelement <1 x elt>, dest) -> bitcast(<1 x elt>, dest)
2383   auto *FixedVType = dyn_cast<FixedVectorType>(VecType);
2384   if (DestType->isVectorTy() && FixedVType && FixedVType->getNumElements() == 1)
2385     return CastInst::Create(Instruction::BitCast, VecOp, DestType);
2386 
2387   return nullptr;
2388 }
2389 
2390 /// Change the type of a bitwise logic operation if we can eliminate a bitcast.
2391 static Instruction *foldBitCastBitwiseLogic(BitCastInst &BitCast,
2392                                             InstCombiner::BuilderTy &Builder) {
2393   Type *DestTy = BitCast.getType();
2394   BinaryOperator *BO;
2395 
2396   if (!match(BitCast.getOperand(0), m_OneUse(m_BinOp(BO))) ||
2397       !BO->isBitwiseLogicOp())
2398     return nullptr;
2399 
2400   // FIXME: This transform is restricted to vector types to avoid backend
2401   // problems caused by creating potentially illegal operations. If a fix-up is
2402   // added to handle that situation, we can remove this check.
2403   if (!DestTy->isVectorTy() || !BO->getType()->isVectorTy())
2404     return nullptr;
2405 
2406   if (DestTy->isFPOrFPVectorTy()) {
2407     Value *X, *Y;
2408     // bitcast(logic(bitcast(X), bitcast(Y))) -> bitcast'(logic(bitcast'(X), Y))
2409     if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
2410         match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(Y))))) {
2411       if (X->getType()->isFPOrFPVectorTy() &&
2412           Y->getType()->isIntOrIntVectorTy()) {
2413         Value *CastedOp =
2414             Builder.CreateBitCast(BO->getOperand(0), Y->getType());
2415         Value *NewBO = Builder.CreateBinOp(BO->getOpcode(), CastedOp, Y);
2416         return CastInst::CreateBitOrPointerCast(NewBO, DestTy);
2417       }
2418       if (X->getType()->isIntOrIntVectorTy() &&
2419           Y->getType()->isFPOrFPVectorTy()) {
2420         Value *CastedOp =
2421             Builder.CreateBitCast(BO->getOperand(1), X->getType());
2422         Value *NewBO = Builder.CreateBinOp(BO->getOpcode(), CastedOp, X);
2423         return CastInst::CreateBitOrPointerCast(NewBO, DestTy);
2424       }
2425     }
2426     return nullptr;
2427   }
2428 
2429   if (!DestTy->isIntOrIntVectorTy())
2430     return nullptr;
2431 
2432   Value *X;
2433   if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
2434       X->getType() == DestTy && !isa<Constant>(X)) {
2435     // bitcast(logic(bitcast(X), Y)) --> logic'(X, bitcast(Y))
2436     Value *CastedOp1 = Builder.CreateBitCast(BO->getOperand(1), DestTy);
2437     return BinaryOperator::Create(BO->getOpcode(), X, CastedOp1);
2438   }
2439 
2440   if (match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(X)))) &&
2441       X->getType() == DestTy && !isa<Constant>(X)) {
2442     // bitcast(logic(Y, bitcast(X))) --> logic'(bitcast(Y), X)
2443     Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
2444     return BinaryOperator::Create(BO->getOpcode(), CastedOp0, X);
2445   }
2446 
2447   // Canonicalize vector bitcasts to come before vector bitwise logic with a
2448   // constant. This eases recognition of special constants for later ops.
2449   // Example:
2450   // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
2451   Constant *C;
2452   if (match(BO->getOperand(1), m_Constant(C))) {
2453     // bitcast (logic X, C) --> logic (bitcast X, C')
2454     Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
2455     Value *CastedC = Builder.CreateBitCast(C, DestTy);
2456     return BinaryOperator::Create(BO->getOpcode(), CastedOp0, CastedC);
2457   }
2458 
2459   return nullptr;
2460 }
2461 
2462 /// Change the type of a select if we can eliminate a bitcast.
2463 static Instruction *foldBitCastSelect(BitCastInst &BitCast,
2464                                       InstCombiner::BuilderTy &Builder) {
2465   Value *Cond, *TVal, *FVal;
2466   if (!match(BitCast.getOperand(0),
2467              m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
2468     return nullptr;
2469 
2470   // A vector select must maintain the same number of elements in its operands.
2471   Type *CondTy = Cond->getType();
2472   Type *DestTy = BitCast.getType();
2473   if (auto *CondVTy = dyn_cast<VectorType>(CondTy))
2474     if (!DestTy->isVectorTy() ||
2475         CondVTy->getElementCount() !=
2476             cast<VectorType>(DestTy)->getElementCount())
2477       return nullptr;
2478 
2479   // FIXME: This transform is restricted from changing the select between
2480   // scalars and vectors to avoid backend problems caused by creating
2481   // potentially illegal operations. If a fix-up is added to handle that
2482   // situation, we can remove this check.
2483   if (DestTy->isVectorTy() != TVal->getType()->isVectorTy())
2484     return nullptr;
2485 
2486   auto *Sel = cast<Instruction>(BitCast.getOperand(0));
2487   Value *X;
2488   if (match(TVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
2489       !isa<Constant>(X)) {
2490     // bitcast(select(Cond, bitcast(X), Y)) --> select'(Cond, X, bitcast(Y))
2491     Value *CastedVal = Builder.CreateBitCast(FVal, DestTy);
2492     return SelectInst::Create(Cond, X, CastedVal, "", nullptr, Sel);
2493   }
2494 
2495   if (match(FVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
2496       !isa<Constant>(X)) {
2497     // bitcast(select(Cond, Y, bitcast(X))) --> select'(Cond, bitcast(Y), X)
2498     Value *CastedVal = Builder.CreateBitCast(TVal, DestTy);
2499     return SelectInst::Create(Cond, CastedVal, X, "", nullptr, Sel);
2500   }
2501 
2502   return nullptr;
2503 }
2504 
2505 /// Check if all users of CI are StoreInsts.
2506 static bool hasStoreUsersOnly(CastInst &CI) {
2507   for (User *U : CI.users()) {
2508     if (!isa<StoreInst>(U))
2509       return false;
2510   }
2511   return true;
2512 }
2513 
2514 /// This function handles following case
2515 ///
2516 ///     A  ->  B    cast
2517 ///     PHI
2518 ///     B  ->  A    cast
2519 ///
2520 /// All the related PHI nodes can be replaced by new PHI nodes with type A.
2521 /// The uses of \p CI can be changed to the new PHI node corresponding to \p PN.
2522 Instruction *InstCombinerImpl::optimizeBitCastFromPhi(CastInst &CI,
2523                                                       PHINode *PN) {
2524   // BitCast used by Store can be handled in InstCombineLoadStoreAlloca.cpp.
2525   if (hasStoreUsersOnly(CI))
2526     return nullptr;
2527 
2528   Value *Src = CI.getOperand(0);
2529   Type *SrcTy = Src->getType();         // Type B
2530   Type *DestTy = CI.getType();          // Type A
2531 
2532   SmallVector<PHINode *, 4> PhiWorklist;
2533   SmallSetVector<PHINode *, 4> OldPhiNodes;
2534 
2535   // Find all of the A->B casts and PHI nodes.
2536   // We need to inspect all related PHI nodes, but PHIs can be cyclic, so
2537   // OldPhiNodes is used to track all known PHI nodes, before adding a new
2538   // PHI to PhiWorklist, it is checked against and added to OldPhiNodes first.
2539   PhiWorklist.push_back(PN);
2540   OldPhiNodes.insert(PN);
2541   while (!PhiWorklist.empty()) {
2542     auto *OldPN = PhiWorklist.pop_back_val();
2543     for (Value *IncValue : OldPN->incoming_values()) {
2544       if (isa<Constant>(IncValue))
2545         continue;
2546 
2547       if (auto *LI = dyn_cast<LoadInst>(IncValue)) {
2548         // If there is a sequence of one or more load instructions, each loaded
2549         // value is used as address of later load instruction, bitcast is
2550         // necessary to change the value type, don't optimize it. For
2551         // simplicity we give up if the load address comes from another load.
2552         Value *Addr = LI->getOperand(0);
2553         if (Addr == &CI || isa<LoadInst>(Addr))
2554           return nullptr;
2555         // Don't tranform "load <256 x i32>, <256 x i32>*" to
2556         // "load x86_amx, x86_amx*", because x86_amx* is invalid.
2557         // TODO: Remove this check when bitcast between vector and x86_amx
2558         // is replaced with a specific intrinsic.
2559         if (DestTy->isX86_AMXTy())
2560           return nullptr;
2561         if (LI->hasOneUse() && LI->isSimple())
2562           continue;
2563         // If a LoadInst has more than one use, changing the type of loaded
2564         // value may create another bitcast.
2565         return nullptr;
2566       }
2567 
2568       if (auto *PNode = dyn_cast<PHINode>(IncValue)) {
2569         if (OldPhiNodes.insert(PNode))
2570           PhiWorklist.push_back(PNode);
2571         continue;
2572       }
2573 
2574       auto *BCI = dyn_cast<BitCastInst>(IncValue);
2575       // We can't handle other instructions.
2576       if (!BCI)
2577         return nullptr;
2578 
2579       // Verify it's a A->B cast.
2580       Type *TyA = BCI->getOperand(0)->getType();
2581       Type *TyB = BCI->getType();
2582       if (TyA != DestTy || TyB != SrcTy)
2583         return nullptr;
2584     }
2585   }
2586 
2587   // Check that each user of each old PHI node is something that we can
2588   // rewrite, so that all of the old PHI nodes can be cleaned up afterwards.
2589   for (auto *OldPN : OldPhiNodes) {
2590     for (User *V : OldPN->users()) {
2591       if (auto *SI = dyn_cast<StoreInst>(V)) {
2592         if (!SI->isSimple() || SI->getOperand(0) != OldPN)
2593           return nullptr;
2594       } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2595         // Verify it's a B->A cast.
2596         Type *TyB = BCI->getOperand(0)->getType();
2597         Type *TyA = BCI->getType();
2598         if (TyA != DestTy || TyB != SrcTy)
2599           return nullptr;
2600       } else if (auto *PHI = dyn_cast<PHINode>(V)) {
2601         // As long as the user is another old PHI node, then even if we don't
2602         // rewrite it, the PHI web we're considering won't have any users
2603         // outside itself, so it'll be dead.
2604         if (!OldPhiNodes.contains(PHI))
2605           return nullptr;
2606       } else {
2607         return nullptr;
2608       }
2609     }
2610   }
2611 
2612   // For each old PHI node, create a corresponding new PHI node with a type A.
2613   SmallDenseMap<PHINode *, PHINode *> NewPNodes;
2614   for (auto *OldPN : OldPhiNodes) {
2615     Builder.SetInsertPoint(OldPN);
2616     PHINode *NewPN = Builder.CreatePHI(DestTy, OldPN->getNumOperands());
2617     NewPNodes[OldPN] = NewPN;
2618   }
2619 
2620   // Fill in the operands of new PHI nodes.
2621   for (auto *OldPN : OldPhiNodes) {
2622     PHINode *NewPN = NewPNodes[OldPN];
2623     for (unsigned j = 0, e = OldPN->getNumOperands(); j != e; ++j) {
2624       Value *V = OldPN->getOperand(j);
2625       Value *NewV = nullptr;
2626       if (auto *C = dyn_cast<Constant>(V)) {
2627         NewV = ConstantExpr::getBitCast(C, DestTy);
2628       } else if (auto *LI = dyn_cast<LoadInst>(V)) {
2629         // Explicitly perform load combine to make sure no opposing transform
2630         // can remove the bitcast in the meantime and trigger an infinite loop.
2631         Builder.SetInsertPoint(LI);
2632         NewV = combineLoadToNewType(*LI, DestTy);
2633         // Remove the old load and its use in the old phi, which itself becomes
2634         // dead once the whole transform finishes.
2635         replaceInstUsesWith(*LI, PoisonValue::get(LI->getType()));
2636         eraseInstFromFunction(*LI);
2637       } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2638         NewV = BCI->getOperand(0);
2639       } else if (auto *PrevPN = dyn_cast<PHINode>(V)) {
2640         NewV = NewPNodes[PrevPN];
2641       }
2642       assert(NewV);
2643       NewPN->addIncoming(NewV, OldPN->getIncomingBlock(j));
2644     }
2645   }
2646 
2647   // Traverse all accumulated PHI nodes and process its users,
2648   // which are Stores and BitcCasts. Without this processing
2649   // NewPHI nodes could be replicated and could lead to extra
2650   // moves generated after DeSSA.
2651   // If there is a store with type B, change it to type A.
2652 
2653 
2654   // Replace users of BitCast B->A with NewPHI. These will help
2655   // later to get rid off a closure formed by OldPHI nodes.
2656   Instruction *RetVal = nullptr;
2657   for (auto *OldPN : OldPhiNodes) {
2658     PHINode *NewPN = NewPNodes[OldPN];
2659     for (User *V : make_early_inc_range(OldPN->users())) {
2660       if (auto *SI = dyn_cast<StoreInst>(V)) {
2661         assert(SI->isSimple() && SI->getOperand(0) == OldPN);
2662         Builder.SetInsertPoint(SI);
2663         auto *NewBC =
2664           cast<BitCastInst>(Builder.CreateBitCast(NewPN, SrcTy));
2665         SI->setOperand(0, NewBC);
2666         Worklist.push(SI);
2667         assert(hasStoreUsersOnly(*NewBC));
2668       }
2669       else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2670         Type *TyB = BCI->getOperand(0)->getType();
2671         Type *TyA = BCI->getType();
2672         assert(TyA == DestTy && TyB == SrcTy);
2673         (void) TyA;
2674         (void) TyB;
2675         Instruction *I = replaceInstUsesWith(*BCI, NewPN);
2676         if (BCI == &CI)
2677           RetVal = I;
2678       } else if (auto *PHI = dyn_cast<PHINode>(V)) {
2679         assert(OldPhiNodes.contains(PHI));
2680         (void) PHI;
2681       } else {
2682         llvm_unreachable("all uses should be handled");
2683       }
2684     }
2685   }
2686 
2687   return RetVal;
2688 }
2689 
2690 static Instruction *convertBitCastToGEP(BitCastInst &CI, IRBuilderBase &Builder,
2691                                         const DataLayout &DL) {
2692   Value *Src = CI.getOperand(0);
2693   PointerType *SrcPTy = cast<PointerType>(Src->getType());
2694   PointerType *DstPTy = cast<PointerType>(CI.getType());
2695 
2696   // Bitcasts involving opaque pointers cannot be converted into a GEP.
2697   if (SrcPTy->isOpaque() || DstPTy->isOpaque())
2698     return nullptr;
2699 
2700   Type *DstElTy = DstPTy->getNonOpaquePointerElementType();
2701   Type *SrcElTy = SrcPTy->getNonOpaquePointerElementType();
2702 
2703   // When the type pointed to is not sized the cast cannot be
2704   // turned into a gep.
2705   if (!SrcElTy->isSized())
2706     return nullptr;
2707 
2708   // If the source and destination are pointers, and this cast is equivalent
2709   // to a getelementptr X, 0, 0, 0...  turn it into the appropriate gep.
2710   // This can enhance SROA and other transforms that want type-safe pointers.
2711   unsigned NumZeros = 0;
2712   while (SrcElTy && SrcElTy != DstElTy) {
2713     SrcElTy = GetElementPtrInst::getTypeAtIndex(SrcElTy, (uint64_t)0);
2714     ++NumZeros;
2715   }
2716 
2717   // If we found a path from the src to dest, create the getelementptr now.
2718   if (SrcElTy == DstElTy) {
2719     SmallVector<Value *, 8> Idxs(NumZeros + 1, Builder.getInt32(0));
2720     GetElementPtrInst *GEP = GetElementPtrInst::Create(
2721         SrcPTy->getNonOpaquePointerElementType(), Src, Idxs);
2722 
2723     // If the source pointer is dereferenceable, then assume it points to an
2724     // allocated object and apply "inbounds" to the GEP.
2725     bool CanBeNull, CanBeFreed;
2726     if (Src->getPointerDereferenceableBytes(DL, CanBeNull, CanBeFreed)) {
2727       // In a non-default address space (not 0), a null pointer can not be
2728       // assumed inbounds, so ignore that case (dereferenceable_or_null).
2729       // The reason is that 'null' is not treated differently in these address
2730       // spaces, and we consequently ignore the 'gep inbounds' special case
2731       // for 'null' which allows 'inbounds' on 'null' if the indices are
2732       // zeros.
2733       if (SrcPTy->getAddressSpace() == 0 || !CanBeNull)
2734         GEP->setIsInBounds();
2735     }
2736     return GEP;
2737   }
2738   return nullptr;
2739 }
2740 
2741 Instruction *InstCombinerImpl::visitBitCast(BitCastInst &CI) {
2742   // If the operands are integer typed then apply the integer transforms,
2743   // otherwise just apply the common ones.
2744   Value *Src = CI.getOperand(0);
2745   Type *SrcTy = Src->getType();
2746   Type *DestTy = CI.getType();
2747 
2748   // Get rid of casts from one type to the same type. These are useless and can
2749   // be replaced by the operand.
2750   if (DestTy == Src->getType())
2751     return replaceInstUsesWith(CI, Src);
2752 
2753   if (isa<PointerType>(SrcTy) && isa<PointerType>(DestTy)) {
2754     // If we are casting a alloca to a pointer to a type of the same
2755     // size, rewrite the allocation instruction to allocate the "right" type.
2756     // There is no need to modify malloc calls because it is their bitcast that
2757     // needs to be cleaned up.
2758     if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
2759       if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
2760         return V;
2761 
2762     if (Instruction *I = convertBitCastToGEP(CI, Builder, DL))
2763       return I;
2764   }
2765 
2766   if (FixedVectorType *DestVTy = dyn_cast<FixedVectorType>(DestTy)) {
2767     // Beware: messing with this target-specific oddity may cause trouble.
2768     if (DestVTy->getNumElements() == 1 && SrcTy->isX86_MMXTy()) {
2769       Value *Elem = Builder.CreateBitCast(Src, DestVTy->getElementType());
2770       return InsertElementInst::Create(PoisonValue::get(DestTy), Elem,
2771                      Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2772     }
2773 
2774     if (isa<IntegerType>(SrcTy)) {
2775       // If this is a cast from an integer to vector, check to see if the input
2776       // is a trunc or zext of a bitcast from vector.  If so, we can replace all
2777       // the casts with a shuffle and (potentially) a bitcast.
2778       if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
2779         CastInst *SrcCast = cast<CastInst>(Src);
2780         if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
2781           if (isa<VectorType>(BCIn->getOperand(0)->getType()))
2782             if (Instruction *I = optimizeVectorResizeWithIntegerBitCasts(
2783                     BCIn->getOperand(0), cast<VectorType>(DestTy), *this))
2784               return I;
2785       }
2786 
2787       // If the input is an 'or' instruction, we may be doing shifts and ors to
2788       // assemble the elements of the vector manually.  Try to rip the code out
2789       // and replace it with insertelements.
2790       if (Value *V = optimizeIntegerToVectorInsertions(CI, *this))
2791         return replaceInstUsesWith(CI, V);
2792     }
2793   }
2794 
2795   if (FixedVectorType *SrcVTy = dyn_cast<FixedVectorType>(SrcTy)) {
2796     if (SrcVTy->getNumElements() == 1) {
2797       // If our destination is not a vector, then make this a straight
2798       // scalar-scalar cast.
2799       if (!DestTy->isVectorTy()) {
2800         Value *Elem =
2801           Builder.CreateExtractElement(Src,
2802                      Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2803         return CastInst::Create(Instruction::BitCast, Elem, DestTy);
2804       }
2805 
2806       // Otherwise, see if our source is an insert. If so, then use the scalar
2807       // component directly:
2808       // bitcast (inselt <1 x elt> V, X, 0) to <n x m> --> bitcast X to <n x m>
2809       if (auto *InsElt = dyn_cast<InsertElementInst>(Src))
2810         return new BitCastInst(InsElt->getOperand(1), DestTy);
2811     }
2812 
2813     // Convert an artificial vector insert into more analyzable bitwise logic.
2814     unsigned BitWidth = DestTy->getScalarSizeInBits();
2815     Value *X, *Y;
2816     uint64_t IndexC;
2817     if (match(Src, m_OneUse(m_InsertElt(m_OneUse(m_BitCast(m_Value(X))),
2818                                         m_Value(Y), m_ConstantInt(IndexC)))) &&
2819         DestTy->isIntegerTy() && X->getType() == DestTy &&
2820         Y->getType()->isIntegerTy() && isDesirableIntType(BitWidth)) {
2821       // Adjust for big endian - the LSBs are at the high index.
2822       if (DL.isBigEndian())
2823         IndexC = SrcVTy->getNumElements() - 1 - IndexC;
2824 
2825       // We only handle (endian-normalized) insert to index 0. Any other insert
2826       // would require a left-shift, so that is an extra instruction.
2827       if (IndexC == 0) {
2828         // bitcast (inselt (bitcast X), Y, 0) --> or (and X, MaskC), (zext Y)
2829         unsigned EltWidth = Y->getType()->getScalarSizeInBits();
2830         APInt MaskC = APInt::getHighBitsSet(BitWidth, BitWidth - EltWidth);
2831         Value *AndX = Builder.CreateAnd(X, MaskC);
2832         Value *ZextY = Builder.CreateZExt(Y, DestTy);
2833         return BinaryOperator::CreateOr(AndX, ZextY);
2834       }
2835     }
2836   }
2837 
2838   if (auto *Shuf = dyn_cast<ShuffleVectorInst>(Src)) {
2839     // Okay, we have (bitcast (shuffle ..)).  Check to see if this is
2840     // a bitcast to a vector with the same # elts.
2841     Value *ShufOp0 = Shuf->getOperand(0);
2842     Value *ShufOp1 = Shuf->getOperand(1);
2843     auto ShufElts = cast<VectorType>(Shuf->getType())->getElementCount();
2844     auto SrcVecElts = cast<VectorType>(ShufOp0->getType())->getElementCount();
2845     if (Shuf->hasOneUse() && DestTy->isVectorTy() &&
2846         cast<VectorType>(DestTy)->getElementCount() == ShufElts &&
2847         ShufElts == SrcVecElts) {
2848       BitCastInst *Tmp;
2849       // If either of the operands is a cast from CI.getType(), then
2850       // evaluating the shuffle in the casted destination's type will allow
2851       // us to eliminate at least one cast.
2852       if (((Tmp = dyn_cast<BitCastInst>(ShufOp0)) &&
2853            Tmp->getOperand(0)->getType() == DestTy) ||
2854           ((Tmp = dyn_cast<BitCastInst>(ShufOp1)) &&
2855            Tmp->getOperand(0)->getType() == DestTy)) {
2856         Value *LHS = Builder.CreateBitCast(ShufOp0, DestTy);
2857         Value *RHS = Builder.CreateBitCast(ShufOp1, DestTy);
2858         // Return a new shuffle vector.  Use the same element ID's, as we
2859         // know the vector types match #elts.
2860         return new ShuffleVectorInst(LHS, RHS, Shuf->getShuffleMask());
2861       }
2862     }
2863 
2864     // A bitcasted-to-scalar and byte/bit reversing shuffle is better recognized
2865     // as a byte/bit swap:
2866     // bitcast <N x i8> (shuf X, undef, <N, N-1,...0>) -> bswap (bitcast X)
2867     // bitcast <N x i1> (shuf X, undef, <N, N-1,...0>) -> bitreverse (bitcast X)
2868     if (DestTy->isIntegerTy() && ShufElts.getKnownMinValue() % 2 == 0 &&
2869         Shuf->hasOneUse() && Shuf->isReverse()) {
2870       unsigned IntrinsicNum = 0;
2871       if (DL.isLegalInteger(DestTy->getScalarSizeInBits()) &&
2872           SrcTy->getScalarSizeInBits() == 8) {
2873         IntrinsicNum = Intrinsic::bswap;
2874       } else if (SrcTy->getScalarSizeInBits() == 1) {
2875         IntrinsicNum = Intrinsic::bitreverse;
2876       }
2877       if (IntrinsicNum != 0) {
2878         assert(ShufOp0->getType() == SrcTy && "Unexpected shuffle mask");
2879         assert(match(ShufOp1, m_Undef()) && "Unexpected shuffle op");
2880         Function *BswapOrBitreverse =
2881             Intrinsic::getDeclaration(CI.getModule(), IntrinsicNum, DestTy);
2882         Value *ScalarX = Builder.CreateBitCast(ShufOp0, DestTy);
2883         return CallInst::Create(BswapOrBitreverse, {ScalarX});
2884       }
2885     }
2886   }
2887 
2888   // Handle the A->B->A cast, and there is an intervening PHI node.
2889   if (PHINode *PN = dyn_cast<PHINode>(Src))
2890     if (Instruction *I = optimizeBitCastFromPhi(CI, PN))
2891       return I;
2892 
2893   if (Instruction *I = canonicalizeBitCastExtElt(CI, *this))
2894     return I;
2895 
2896   if (Instruction *I = foldBitCastBitwiseLogic(CI, Builder))
2897     return I;
2898 
2899   if (Instruction *I = foldBitCastSelect(CI, Builder))
2900     return I;
2901 
2902   if (SrcTy->isPointerTy())
2903     return commonPointerCastTransforms(CI);
2904   return commonCastTransforms(CI);
2905 }
2906 
2907 Instruction *InstCombinerImpl::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
2908   // If the destination pointer element type is not the same as the source's
2909   // first do a bitcast to the destination type, and then the addrspacecast.
2910   // This allows the cast to be exposed to other transforms.
2911   Value *Src = CI.getOperand(0);
2912   PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType());
2913   PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType());
2914 
2915   if (!SrcTy->hasSameElementTypeAs(DestTy)) {
2916     Type *MidTy =
2917         PointerType::getWithSamePointeeType(DestTy, SrcTy->getAddressSpace());
2918     // Handle vectors of pointers.
2919     if (VectorType *VT = dyn_cast<VectorType>(CI.getType()))
2920       MidTy = VectorType::get(MidTy, VT->getElementCount());
2921 
2922     Value *NewBitCast = Builder.CreateBitCast(Src, MidTy);
2923     return new AddrSpaceCastInst(NewBitCast, CI.getType());
2924   }
2925 
2926   return commonPointerCastTransforms(CI);
2927 }
2928