xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp (revision 700637cbb5e582861067a11aaca4d053546871d2)
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 /// Given an expression that CanEvaluateTruncated or CanEvaluateSExtd returns
29 /// true for, actually insert the code to evaluate the expression.
EvaluateInDifferentType(Value * V,Type * Ty,bool isSigned)30 Value *InstCombinerImpl::EvaluateInDifferentType(Value *V, Type *Ty,
31                                                  bool isSigned) {
32   if (Constant *C = dyn_cast<Constant>(V))
33     return ConstantFoldIntegerCast(C, Ty, isSigned, DL);
34 
35   // Otherwise, it must be an instruction.
36   Instruction *I = cast<Instruction>(V);
37   Instruction *Res = nullptr;
38   unsigned Opc = I->getOpcode();
39   switch (Opc) {
40   case Instruction::Add:
41   case Instruction::Sub:
42   case Instruction::Mul:
43   case Instruction::And:
44   case Instruction::Or:
45   case Instruction::Xor:
46   case Instruction::AShr:
47   case Instruction::LShr:
48   case Instruction::Shl:
49   case Instruction::UDiv:
50   case Instruction::URem: {
51     Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
52     Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
53     Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
54     if (Opc == Instruction::LShr || Opc == Instruction::AShr)
55       Res->setIsExact(I->isExact());
56     break;
57   }
58   case Instruction::Trunc:
59   case Instruction::ZExt:
60   case Instruction::SExt:
61     // If the source type of the cast is the type we're trying for then we can
62     // just return the source.  There's no need to insert it because it is not
63     // new.
64     if (I->getOperand(0)->getType() == Ty)
65       return I->getOperand(0);
66 
67     // Otherwise, must be the same type of cast, so just reinsert a new one.
68     // This also handles the case of zext(trunc(x)) -> zext(x).
69     Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
70                                       Opc == Instruction::SExt);
71     break;
72   case Instruction::Select: {
73     Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
74     Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
75     Res = SelectInst::Create(I->getOperand(0), True, False);
76     break;
77   }
78   case Instruction::PHI: {
79     PHINode *OPN = cast<PHINode>(I);
80     PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
81     for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
82       Value *V =
83           EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
84       NPN->addIncoming(V, OPN->getIncomingBlock(i));
85     }
86     Res = NPN;
87     break;
88   }
89   case Instruction::FPToUI:
90   case Instruction::FPToSI:
91     Res = CastInst::Create(
92       static_cast<Instruction::CastOps>(Opc), I->getOperand(0), Ty);
93     break;
94   case Instruction::Call:
95     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
96       switch (II->getIntrinsicID()) {
97       default:
98         llvm_unreachable("Unsupported call!");
99       case Intrinsic::vscale: {
100         Function *Fn = Intrinsic::getOrInsertDeclaration(
101             I->getModule(), Intrinsic::vscale, {Ty});
102         Res = CallInst::Create(Fn->getFunctionType(), Fn);
103         break;
104       }
105       }
106     }
107     break;
108   case Instruction::ShuffleVector: {
109     auto *ScalarTy = cast<VectorType>(Ty)->getElementType();
110     auto *VTy = cast<VectorType>(I->getOperand(0)->getType());
111     auto *FixedTy = VectorType::get(ScalarTy, VTy->getElementCount());
112     Value *Op0 = EvaluateInDifferentType(I->getOperand(0), FixedTy, isSigned);
113     Value *Op1 = EvaluateInDifferentType(I->getOperand(1), FixedTy, isSigned);
114     Res = new ShuffleVectorInst(Op0, Op1,
115                                 cast<ShuffleVectorInst>(I)->getShuffleMask());
116     break;
117   }
118   default:
119     // TODO: Can handle more cases here.
120     llvm_unreachable("Unreachable!");
121   }
122 
123   Res->takeName(I);
124   return InsertNewInstWith(Res, I->getIterator());
125 }
126 
127 Instruction::CastOps
isEliminableCastPair(const CastInst * CI1,const CastInst * CI2)128 InstCombinerImpl::isEliminableCastPair(const CastInst *CI1,
129                                        const CastInst *CI2) {
130   Type *SrcTy = CI1->getSrcTy();
131   Type *MidTy = CI1->getDestTy();
132   Type *DstTy = CI2->getDestTy();
133 
134   Instruction::CastOps firstOp = CI1->getOpcode();
135   Instruction::CastOps secondOp = CI2->getOpcode();
136   Type *SrcIntPtrTy =
137       SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr;
138   Type *MidIntPtrTy =
139       MidTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(MidTy) : nullptr;
140   Type *DstIntPtrTy =
141       DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr;
142   unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
143                                                 DstTy, SrcIntPtrTy, MidIntPtrTy,
144                                                 DstIntPtrTy);
145 
146   // We don't want to form an inttoptr or ptrtoint that converts to an integer
147   // type that differs from the pointer size.
148   if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
149       (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
150     Res = 0;
151 
152   return Instruction::CastOps(Res);
153 }
154 
155 /// Implement the transforms common to all CastInst visitors.
commonCastTransforms(CastInst & CI)156 Instruction *InstCombinerImpl::commonCastTransforms(CastInst &CI) {
157   Value *Src = CI.getOperand(0);
158   Type *Ty = CI.getType();
159 
160   if (auto *SrcC = dyn_cast<Constant>(Src))
161     if (Constant *Res = ConstantFoldCastOperand(CI.getOpcode(), SrcC, Ty, DL))
162       return replaceInstUsesWith(CI, Res);
163 
164   // Try to eliminate a cast of a cast.
165   if (auto *CSrc = dyn_cast<CastInst>(Src)) {   // A->B->C cast
166     if (Instruction::CastOps NewOpc = isEliminableCastPair(CSrc, &CI)) {
167       // The first cast (CSrc) is eliminable so we need to fix up or replace
168       // the second cast (CI). CSrc will then have a good chance of being dead.
169       auto *Res = CastInst::Create(NewOpc, CSrc->getOperand(0), Ty);
170       // Point debug users of the dying cast to the new one.
171       if (CSrc->hasOneUse())
172         replaceAllDbgUsesWith(*CSrc, *Res, CI, DT);
173       return Res;
174     }
175   }
176 
177   if (auto *Sel = dyn_cast<SelectInst>(Src)) {
178     // We are casting a select. Try to fold the cast into the select if the
179     // select does not have a compare instruction with matching operand types
180     // or the select is likely better done in a narrow type.
181     // Creating a select with operands that are different sizes than its
182     // condition may inhibit other folds and lead to worse codegen.
183     auto *Cmp = dyn_cast<CmpInst>(Sel->getCondition());
184     if (!Cmp || Cmp->getOperand(0)->getType() != Sel->getType() ||
185         (CI.getOpcode() == Instruction::Trunc &&
186          shouldChangeType(CI.getSrcTy(), CI.getType()))) {
187 
188       // If it's a bitcast involving vectors, make sure it has the same number
189       // of elements on both sides.
190       if (CI.getOpcode() != Instruction::BitCast ||
191           match(&CI, m_ElementWiseBitCast(m_Value()))) {
192         if (Instruction *NV = FoldOpIntoSelect(CI, Sel)) {
193           replaceAllDbgUsesWith(*Sel, *NV, CI, DT);
194           return NV;
195         }
196       }
197     }
198   }
199 
200   // If we are casting a PHI, then fold the cast into the PHI.
201   if (auto *PN = dyn_cast<PHINode>(Src)) {
202     // Don't do this if it would create a PHI node with an illegal type from a
203     // legal type.
204     if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||
205         shouldChangeType(CI.getSrcTy(), CI.getType()))
206       if (Instruction *NV = foldOpIntoPhi(CI, PN))
207         return NV;
208   }
209 
210   // Canonicalize a unary shuffle after the cast if neither operation changes
211   // the size or element size of the input vector.
212   // TODO: We could allow size-changing ops if that doesn't harm codegen.
213   // cast (shuffle X, Mask) --> shuffle (cast X), Mask
214   Value *X;
215   ArrayRef<int> Mask;
216   if (match(Src, m_OneUse(m_Shuffle(m_Value(X), m_Undef(), m_Mask(Mask))))) {
217     // TODO: Allow scalable vectors?
218     auto *SrcTy = dyn_cast<FixedVectorType>(X->getType());
219     auto *DestTy = dyn_cast<FixedVectorType>(Ty);
220     if (SrcTy && DestTy &&
221         SrcTy->getNumElements() == DestTy->getNumElements() &&
222         SrcTy->getPrimitiveSizeInBits() == DestTy->getPrimitiveSizeInBits()) {
223       Value *CastX = Builder.CreateCast(CI.getOpcode(), X, DestTy);
224       return new ShuffleVectorInst(CastX, Mask);
225     }
226   }
227 
228   return nullptr;
229 }
230 
231 /// Constants and extensions/truncates from the destination type are always
232 /// free to be evaluated in that type. This is a helper for canEvaluate*.
canAlwaysEvaluateInType(Value * V,Type * Ty)233 static bool canAlwaysEvaluateInType(Value *V, Type *Ty) {
234   if (isa<Constant>(V))
235     return match(V, m_ImmConstant());
236 
237   Value *X;
238   if ((match(V, m_ZExtOrSExt(m_Value(X))) || match(V, m_Trunc(m_Value(X)))) &&
239       X->getType() == Ty)
240     return true;
241 
242   return false;
243 }
244 
245 /// Filter out values that we can not evaluate in the destination type for free.
246 /// This is a helper for canEvaluate*.
canNotEvaluateInType(Value * V,Type * Ty)247 static bool canNotEvaluateInType(Value *V, Type *Ty) {
248   if (!isa<Instruction>(V))
249     return true;
250   // We don't extend or shrink something that has multiple uses --  doing so
251   // would require duplicating the instruction which isn't profitable.
252   if (!V->hasOneUse())
253     return true;
254 
255   return false;
256 }
257 
258 /// Return true if we can evaluate the specified expression tree as type Ty
259 /// instead of its larger type, and arrive with the same value.
260 /// This is used by code that tries to eliminate truncates.
261 ///
262 /// Ty will always be a type smaller than V.  We should return true if trunc(V)
263 /// can be computed by computing V in the smaller type.  If V is an instruction,
264 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
265 /// makes sense if x and y can be efficiently truncated.
266 ///
267 /// This function works on both vectors and scalars.
268 ///
canEvaluateTruncated(Value * V,Type * Ty,InstCombinerImpl & IC,Instruction * CxtI)269 static bool canEvaluateTruncated(Value *V, Type *Ty, InstCombinerImpl &IC,
270                                  Instruction *CxtI) {
271   if (canAlwaysEvaluateInType(V, Ty))
272     return true;
273   if (canNotEvaluateInType(V, Ty))
274     return false;
275 
276   auto *I = cast<Instruction>(V);
277   Type *OrigTy = V->getType();
278   switch (I->getOpcode()) {
279   case Instruction::Add:
280   case Instruction::Sub:
281   case Instruction::Mul:
282   case Instruction::And:
283   case Instruction::Or:
284   case Instruction::Xor:
285     // These operators can all arbitrarily be extended or truncated.
286     return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
287            canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
288 
289   case Instruction::UDiv:
290   case Instruction::URem: {
291     // UDiv and URem can be truncated if all the truncated bits are zero.
292     uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
293     uint32_t BitWidth = Ty->getScalarSizeInBits();
294     assert(BitWidth < OrigBitWidth && "Unexpected bitwidths!");
295     APInt Mask = APInt::getBitsSetFrom(OrigBitWidth, BitWidth);
296     // Do not preserve the original context instruction. Simplifying div/rem
297     // based on later context may introduce a trap.
298     if (IC.MaskedValueIsZero(I->getOperand(0), Mask, I) &&
299         IC.MaskedValueIsZero(I->getOperand(1), Mask, I)) {
300       return canEvaluateTruncated(I->getOperand(0), Ty, IC, I) &&
301              canEvaluateTruncated(I->getOperand(1), Ty, IC, I);
302     }
303     break;
304   }
305   case Instruction::Shl: {
306     // If we are truncating the result of this SHL, and if it's a shift of an
307     // inrange amount, we can always perform a SHL in a smaller type.
308     uint32_t BitWidth = Ty->getScalarSizeInBits();
309     KnownBits AmtKnownBits =
310         llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());
311     if (AmtKnownBits.getMaxValue().ult(BitWidth))
312       return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
313              canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
314     break;
315   }
316   case Instruction::LShr: {
317     // If this is a truncate of a logical shr, we can truncate it to a smaller
318     // lshr iff we know that the bits we would otherwise be shifting in are
319     // already zeros.
320     // TODO: It is enough to check that the bits we would be shifting in are
321     //       zero - use AmtKnownBits.getMaxValue().
322     uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
323     uint32_t BitWidth = Ty->getScalarSizeInBits();
324     KnownBits AmtKnownBits = IC.computeKnownBits(I->getOperand(1), CxtI);
325     APInt MaxShiftAmt = AmtKnownBits.getMaxValue();
326     APInt ShiftedBits = APInt::getBitsSetFrom(OrigBitWidth, BitWidth);
327     if (MaxShiftAmt.ult(BitWidth)) {
328       // If the only user is a trunc then we can narrow the shift if any new
329       // MSBs are not going to be used.
330       if (auto *Trunc = dyn_cast<TruncInst>(V->user_back())) {
331         auto DemandedBits = Trunc->getType()->getScalarSizeInBits();
332         if ((MaxShiftAmt + DemandedBits).ule(BitWidth))
333           return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
334                  canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
335       }
336       if (IC.MaskedValueIsZero(I->getOperand(0), ShiftedBits, CxtI))
337         return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
338                canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
339     }
340     break;
341   }
342   case Instruction::AShr: {
343     // If this is a truncate of an arithmetic shr, we can truncate it to a
344     // smaller ashr iff we know that all the bits from the sign bit of the
345     // original type and the sign bit of the truncate type are similar.
346     // TODO: It is enough to check that the bits we would be shifting in are
347     //       similar to sign bit of the truncate type.
348     uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
349     uint32_t BitWidth = Ty->getScalarSizeInBits();
350     KnownBits AmtKnownBits =
351         llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());
352     unsigned ShiftedBits = OrigBitWidth - BitWidth;
353     if (AmtKnownBits.getMaxValue().ult(BitWidth) &&
354         ShiftedBits < IC.ComputeNumSignBits(I->getOperand(0), CxtI))
355       return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
356              canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
357     break;
358   }
359   case Instruction::Trunc:
360     // trunc(trunc(x)) -> trunc(x)
361     return true;
362   case Instruction::ZExt:
363   case Instruction::SExt:
364     // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
365     // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
366     return true;
367   case Instruction::Select: {
368     SelectInst *SI = cast<SelectInst>(I);
369     return canEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
370            canEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
371   }
372   case Instruction::PHI: {
373     // We can change a phi if we can change all operands.  Note that we never
374     // get into trouble with cyclic PHIs here because we only consider
375     // instructions with a single use.
376     PHINode *PN = cast<PHINode>(I);
377     for (Value *IncValue : PN->incoming_values())
378       if (!canEvaluateTruncated(IncValue, Ty, IC, CxtI))
379         return false;
380     return true;
381   }
382   case Instruction::FPToUI:
383   case Instruction::FPToSI: {
384     // If the integer type can hold the max FP value, it is safe to cast
385     // directly to that type. Otherwise, we may create poison via overflow
386     // that did not exist in the original code.
387     Type *InputTy = I->getOperand(0)->getType()->getScalarType();
388     const fltSemantics &Semantics = InputTy->getFltSemantics();
389     uint32_t MinBitWidth =
390       APFloatBase::semanticsIntSizeInBits(Semantics,
391           I->getOpcode() == Instruction::FPToSI);
392     return Ty->getScalarSizeInBits() >= MinBitWidth;
393   }
394   case Instruction::ShuffleVector:
395     return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
396            canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
397   default:
398     // TODO: Can handle more cases here.
399     break;
400   }
401 
402   return false;
403 }
404 
405 /// Given a vector that is bitcast to an integer, optionally logically
406 /// right-shifted, and truncated, convert it to an extractelement.
407 /// Example (big endian):
408 ///   trunc (lshr (bitcast <4 x i32> %X to i128), 32) to i32
409 ///   --->
410 ///   extractelement <4 x i32> %X, 1
foldVecTruncToExtElt(TruncInst & Trunc,InstCombinerImpl & IC)411 static Instruction *foldVecTruncToExtElt(TruncInst &Trunc,
412                                          InstCombinerImpl &IC) {
413   Value *TruncOp = Trunc.getOperand(0);
414   Type *DestType = Trunc.getType();
415   if (!TruncOp->hasOneUse() || !isa<IntegerType>(DestType))
416     return nullptr;
417 
418   Value *VecInput = nullptr;
419   ConstantInt *ShiftVal = nullptr;
420   if (!match(TruncOp, m_CombineOr(m_BitCast(m_Value(VecInput)),
421                                   m_LShr(m_BitCast(m_Value(VecInput)),
422                                          m_ConstantInt(ShiftVal)))) ||
423       !isa<VectorType>(VecInput->getType()))
424     return nullptr;
425 
426   VectorType *VecType = cast<VectorType>(VecInput->getType());
427   unsigned VecWidth = VecType->getPrimitiveSizeInBits();
428   unsigned DestWidth = DestType->getPrimitiveSizeInBits();
429   unsigned ShiftAmount = ShiftVal ? ShiftVal->getZExtValue() : 0;
430 
431   if ((VecWidth % DestWidth != 0) || (ShiftAmount % DestWidth != 0))
432     return nullptr;
433 
434   // If the element type of the vector doesn't match the result type,
435   // bitcast it to a vector type that we can extract from.
436   unsigned NumVecElts = VecWidth / DestWidth;
437   if (VecType->getElementType() != DestType) {
438     VecType = FixedVectorType::get(DestType, NumVecElts);
439     VecInput = IC.Builder.CreateBitCast(VecInput, VecType, "bc");
440   }
441 
442   unsigned Elt = ShiftAmount / DestWidth;
443   if (IC.getDataLayout().isBigEndian())
444     Elt = NumVecElts - 1 - Elt;
445 
446   return ExtractElementInst::Create(VecInput, IC.Builder.getInt32(Elt));
447 }
448 
449 /// Whenever an element is extracted from a vector, optionally shifted down, and
450 /// then truncated, canonicalize by converting it to a bitcast followed by an
451 /// extractelement.
452 ///
453 /// Examples (little endian):
454 ///   trunc (extractelement <4 x i64> %X, 0) to i32
455 ///   --->
456 ///   extractelement <8 x i32> (bitcast <4 x i64> %X to <8 x i32>), i32 0
457 ///
458 ///   trunc (lshr (extractelement <4 x i32> %X, 0), 8) to i8
459 ///   --->
460 ///   extractelement <16 x i8> (bitcast <4 x i32> %X to <16 x i8>), i32 1
foldVecExtTruncToExtElt(TruncInst & Trunc,InstCombinerImpl & IC)461 static Instruction *foldVecExtTruncToExtElt(TruncInst &Trunc,
462                                             InstCombinerImpl &IC) {
463   Value *Src = Trunc.getOperand(0);
464   Type *SrcType = Src->getType();
465   Type *DstType = Trunc.getType();
466 
467   // Only attempt this if we have simple aliasing of the vector elements.
468   // A badly fit destination size would result in an invalid cast.
469   unsigned SrcBits = SrcType->getScalarSizeInBits();
470   unsigned DstBits = DstType->getScalarSizeInBits();
471   unsigned TruncRatio = SrcBits / DstBits;
472   if ((SrcBits % DstBits) != 0)
473     return nullptr;
474 
475   Value *VecOp;
476   ConstantInt *Cst;
477   const APInt *ShiftAmount = nullptr;
478   if (!match(Src, m_OneUse(m_ExtractElt(m_Value(VecOp), m_ConstantInt(Cst)))) &&
479       !match(Src,
480              m_OneUse(m_LShr(m_ExtractElt(m_Value(VecOp), m_ConstantInt(Cst)),
481                              m_APInt(ShiftAmount)))))
482     return nullptr;
483 
484   auto *VecOpTy = cast<VectorType>(VecOp->getType());
485   auto VecElts = VecOpTy->getElementCount();
486 
487   uint64_t BitCastNumElts = VecElts.getKnownMinValue() * TruncRatio;
488   uint64_t VecOpIdx = Cst->getZExtValue();
489   uint64_t NewIdx = IC.getDataLayout().isBigEndian()
490                         ? (VecOpIdx + 1) * TruncRatio - 1
491                         : VecOpIdx * TruncRatio;
492 
493   // Adjust index by the whole number of truncated elements.
494   if (ShiftAmount) {
495     // Check shift amount is in range and shifts a whole number of truncated
496     // elements.
497     if (ShiftAmount->uge(SrcBits) || ShiftAmount->urem(DstBits) != 0)
498       return nullptr;
499 
500     uint64_t IdxOfs = ShiftAmount->udiv(DstBits).getZExtValue();
501     NewIdx = IC.getDataLayout().isBigEndian() ? (NewIdx - IdxOfs)
502                                               : (NewIdx + IdxOfs);
503   }
504 
505   assert(BitCastNumElts <= std::numeric_limits<uint32_t>::max() &&
506          NewIdx <= std::numeric_limits<uint32_t>::max() && "overflow 32-bits");
507 
508   auto *BitCastTo =
509       VectorType::get(DstType, BitCastNumElts, VecElts.isScalable());
510   Value *BitCast = IC.Builder.CreateBitCast(VecOp, BitCastTo);
511   return ExtractElementInst::Create(BitCast, IC.Builder.getInt32(NewIdx));
512 }
513 
514 /// Funnel/Rotate left/right may occur in a wider type than necessary because of
515 /// type promotion rules. Try to narrow the inputs and convert to funnel shift.
narrowFunnelShift(TruncInst & Trunc)516 Instruction *InstCombinerImpl::narrowFunnelShift(TruncInst &Trunc) {
517   assert((isa<VectorType>(Trunc.getSrcTy()) ||
518           shouldChangeType(Trunc.getSrcTy(), Trunc.getType())) &&
519          "Don't narrow to an illegal scalar type");
520 
521   // Bail out on strange types. It is possible to handle some of these patterns
522   // even with non-power-of-2 sizes, but it is not a likely scenario.
523   Type *DestTy = Trunc.getType();
524   unsigned NarrowWidth = DestTy->getScalarSizeInBits();
525   unsigned WideWidth = Trunc.getSrcTy()->getScalarSizeInBits();
526   if (!isPowerOf2_32(NarrowWidth))
527     return nullptr;
528 
529   // First, find an or'd pair of opposite shifts:
530   // trunc (or (lshr ShVal0, ShAmt0), (shl ShVal1, ShAmt1))
531   BinaryOperator *Or0, *Or1;
532   if (!match(Trunc.getOperand(0), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1)))))
533     return nullptr;
534 
535   Value *ShVal0, *ShVal1, *ShAmt0, *ShAmt1;
536   if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(ShVal0), m_Value(ShAmt0)))) ||
537       !match(Or1, m_OneUse(m_LogicalShift(m_Value(ShVal1), m_Value(ShAmt1)))) ||
538       Or0->getOpcode() == Or1->getOpcode())
539     return nullptr;
540 
541   // Canonicalize to or(shl(ShVal0, ShAmt0), lshr(ShVal1, ShAmt1)).
542   if (Or0->getOpcode() == BinaryOperator::LShr) {
543     std::swap(Or0, Or1);
544     std::swap(ShVal0, ShVal1);
545     std::swap(ShAmt0, ShAmt1);
546   }
547   assert(Or0->getOpcode() == BinaryOperator::Shl &&
548          Or1->getOpcode() == BinaryOperator::LShr &&
549          "Illegal or(shift,shift) pair");
550 
551   // Match the shift amount operands for a funnel/rotate pattern. This always
552   // matches a subtraction on the R operand.
553   auto matchShiftAmount = [&](Value *L, Value *R, unsigned Width) -> Value * {
554     // The shift amounts may add up to the narrow bit width:
555     // (shl ShVal0, L) | (lshr ShVal1, Width - L)
556     // If this is a funnel shift (different operands are shifted), then the
557     // shift amount can not over-shift (create poison) in the narrow type.
558     unsigned MaxShiftAmountWidth = Log2_32(NarrowWidth);
559     APInt HiBitMask = ~APInt::getLowBitsSet(WideWidth, MaxShiftAmountWidth);
560     if (ShVal0 == ShVal1 || MaskedValueIsZero(L, HiBitMask))
561       if (match(R, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(L)))))
562         return L;
563 
564     // The following patterns currently only work for rotation patterns.
565     // TODO: Add more general funnel-shift compatible patterns.
566     if (ShVal0 != ShVal1)
567       return nullptr;
568 
569     // The shift amount may be masked with negation:
570     // (shl ShVal0, (X & (Width - 1))) | (lshr ShVal1, ((-X) & (Width - 1)))
571     Value *X;
572     unsigned Mask = Width - 1;
573     if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) &&
574         match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))
575       return X;
576 
577     // Same as above, but the shift amount may be extended after masking:
578     if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&
579         match(R, m_ZExt(m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask)))))
580       return X;
581 
582     return nullptr;
583   };
584 
585   Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, NarrowWidth);
586   bool IsFshl = true; // Sub on LSHR.
587   if (!ShAmt) {
588     ShAmt = matchShiftAmount(ShAmt1, ShAmt0, NarrowWidth);
589     IsFshl = false; // Sub on SHL.
590   }
591   if (!ShAmt)
592     return nullptr;
593 
594   // The right-shifted value must have high zeros in the wide type (for example
595   // from 'zext', 'and' or 'shift'). High bits of the left-shifted value are
596   // truncated, so those do not matter.
597   APInt HiBitMask = APInt::getHighBitsSet(WideWidth, WideWidth - NarrowWidth);
598   if (!MaskedValueIsZero(ShVal1, HiBitMask, &Trunc))
599     return nullptr;
600 
601   // Adjust the width of ShAmt for narrowed funnel shift operation:
602   // - Zero-extend if ShAmt is narrower than the destination type.
603   // - Truncate if ShAmt is wider, discarding non-significant high-order bits.
604   // This prepares ShAmt for llvm.fshl.i8(trunc(ShVal), trunc(ShVal),
605   // zext/trunc(ShAmt)).
606   Value *NarrowShAmt = Builder.CreateZExtOrTrunc(ShAmt, DestTy);
607 
608   Value *X, *Y;
609   X = Y = Builder.CreateTrunc(ShVal0, DestTy);
610   if (ShVal0 != ShVal1)
611     Y = Builder.CreateTrunc(ShVal1, DestTy);
612   Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
613   Function *F =
614       Intrinsic::getOrInsertDeclaration(Trunc.getModule(), IID, DestTy);
615   return CallInst::Create(F, {X, Y, NarrowShAmt});
616 }
617 
618 /// Try to narrow the width of math or bitwise logic instructions by pulling a
619 /// truncate ahead of binary operators.
narrowBinOp(TruncInst & Trunc)620 Instruction *InstCombinerImpl::narrowBinOp(TruncInst &Trunc) {
621   Type *SrcTy = Trunc.getSrcTy();
622   Type *DestTy = Trunc.getType();
623   unsigned SrcWidth = SrcTy->getScalarSizeInBits();
624   unsigned DestWidth = DestTy->getScalarSizeInBits();
625 
626   if (!isa<VectorType>(SrcTy) && !shouldChangeType(SrcTy, DestTy))
627     return nullptr;
628 
629   BinaryOperator *BinOp;
630   if (!match(Trunc.getOperand(0), m_OneUse(m_BinOp(BinOp))))
631     return nullptr;
632 
633   Value *BinOp0 = BinOp->getOperand(0);
634   Value *BinOp1 = BinOp->getOperand(1);
635   switch (BinOp->getOpcode()) {
636   case Instruction::And:
637   case Instruction::Or:
638   case Instruction::Xor:
639   case Instruction::Add:
640   case Instruction::Sub:
641   case Instruction::Mul: {
642     Constant *C;
643     if (match(BinOp0, m_Constant(C))) {
644       // trunc (binop C, X) --> binop (trunc C', X)
645       Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);
646       Value *TruncX = Builder.CreateTrunc(BinOp1, DestTy);
647       return BinaryOperator::Create(BinOp->getOpcode(), NarrowC, TruncX);
648     }
649     if (match(BinOp1, m_Constant(C))) {
650       // trunc (binop X, C) --> binop (trunc X, C')
651       Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);
652       Value *TruncX = Builder.CreateTrunc(BinOp0, DestTy);
653       return BinaryOperator::Create(BinOp->getOpcode(), TruncX, NarrowC);
654     }
655     Value *X;
656     if (match(BinOp0, m_ZExtOrSExt(m_Value(X))) && X->getType() == DestTy) {
657       // trunc (binop (ext X), Y) --> binop X, (trunc Y)
658       Value *NarrowOp1 = Builder.CreateTrunc(BinOp1, DestTy);
659       return BinaryOperator::Create(BinOp->getOpcode(), X, NarrowOp1);
660     }
661     if (match(BinOp1, m_ZExtOrSExt(m_Value(X))) && X->getType() == DestTy) {
662       // trunc (binop Y, (ext X)) --> binop (trunc Y), X
663       Value *NarrowOp0 = Builder.CreateTrunc(BinOp0, DestTy);
664       return BinaryOperator::Create(BinOp->getOpcode(), NarrowOp0, X);
665     }
666     break;
667   }
668   case Instruction::LShr:
669   case Instruction::AShr: {
670     // trunc (*shr (trunc A), C) --> trunc(*shr A, C)
671     Value *A;
672     Constant *C;
673     if (match(BinOp0, m_Trunc(m_Value(A))) && match(BinOp1, m_Constant(C))) {
674       unsigned MaxShiftAmt = SrcWidth - DestWidth;
675       // If the shift is small enough, all zero/sign bits created by the shift
676       // are removed by the trunc.
677       if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULE,
678                                       APInt(SrcWidth, MaxShiftAmt)))) {
679         auto *OldShift = cast<Instruction>(Trunc.getOperand(0));
680         bool IsExact = OldShift->isExact();
681         if (Constant *ShAmt = ConstantFoldIntegerCast(C, A->getType(),
682                                                       /*IsSigned*/ true, DL)) {
683           ShAmt = Constant::mergeUndefsWith(ShAmt, C);
684           Value *Shift =
685               OldShift->getOpcode() == Instruction::AShr
686                   ? Builder.CreateAShr(A, ShAmt, OldShift->getName(), IsExact)
687                   : Builder.CreateLShr(A, ShAmt, OldShift->getName(), IsExact);
688           return CastInst::CreateTruncOrBitCast(Shift, DestTy);
689         }
690       }
691     }
692     break;
693   }
694   default: break;
695   }
696 
697   if (Instruction *NarrowOr = narrowFunnelShift(Trunc))
698     return NarrowOr;
699 
700   return nullptr;
701 }
702 
703 /// Try to narrow the width of a splat shuffle. This could be generalized to any
704 /// shuffle with a constant operand, but we limit the transform to avoid
705 /// creating a shuffle type that targets may not be able to lower effectively.
shrinkSplatShuffle(TruncInst & Trunc,InstCombiner::BuilderTy & Builder)706 static Instruction *shrinkSplatShuffle(TruncInst &Trunc,
707                                        InstCombiner::BuilderTy &Builder) {
708   auto *Shuf = dyn_cast<ShuffleVectorInst>(Trunc.getOperand(0));
709   if (Shuf && Shuf->hasOneUse() && match(Shuf->getOperand(1), m_Undef()) &&
710       all_equal(Shuf->getShuffleMask()) &&
711       Shuf->getType() == Shuf->getOperand(0)->getType()) {
712     // trunc (shuf X, Undef, SplatMask) --> shuf (trunc X), Poison, SplatMask
713     // trunc (shuf X, Poison, SplatMask) --> shuf (trunc X), Poison, SplatMask
714     Value *NarrowOp = Builder.CreateTrunc(Shuf->getOperand(0), Trunc.getType());
715     return new ShuffleVectorInst(NarrowOp, Shuf->getShuffleMask());
716   }
717 
718   return nullptr;
719 }
720 
721 /// Try to narrow the width of an insert element. This could be generalized for
722 /// any vector constant, but we limit the transform to insertion into undef to
723 /// avoid potential backend problems from unsupported insertion widths. This
724 /// could also be extended to handle the case of inserting a scalar constant
725 /// into a vector variable.
shrinkInsertElt(CastInst & Trunc,InstCombiner::BuilderTy & Builder)726 static Instruction *shrinkInsertElt(CastInst &Trunc,
727                                     InstCombiner::BuilderTy &Builder) {
728   Instruction::CastOps Opcode = Trunc.getOpcode();
729   assert((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
730          "Unexpected instruction for shrinking");
731 
732   auto *InsElt = dyn_cast<InsertElementInst>(Trunc.getOperand(0));
733   if (!InsElt || !InsElt->hasOneUse())
734     return nullptr;
735 
736   Type *DestTy = Trunc.getType();
737   Type *DestScalarTy = DestTy->getScalarType();
738   Value *VecOp = InsElt->getOperand(0);
739   Value *ScalarOp = InsElt->getOperand(1);
740   Value *Index = InsElt->getOperand(2);
741 
742   if (match(VecOp, m_Undef())) {
743     // trunc   (inselt undef, X, Index) --> inselt undef,   (trunc X), Index
744     // fptrunc (inselt undef, X, Index) --> inselt undef, (fptrunc X), Index
745     UndefValue *NarrowUndef = UndefValue::get(DestTy);
746     Value *NarrowOp = Builder.CreateCast(Opcode, ScalarOp, DestScalarTy);
747     return InsertElementInst::Create(NarrowUndef, NarrowOp, Index);
748   }
749 
750   return nullptr;
751 }
752 
visitTrunc(TruncInst & Trunc)753 Instruction *InstCombinerImpl::visitTrunc(TruncInst &Trunc) {
754   if (Instruction *Result = commonCastTransforms(Trunc))
755     return Result;
756 
757   Value *Src = Trunc.getOperand(0);
758   Type *DestTy = Trunc.getType(), *SrcTy = Src->getType();
759   unsigned DestWidth = DestTy->getScalarSizeInBits();
760   unsigned SrcWidth = SrcTy->getScalarSizeInBits();
761 
762   // Attempt to truncate the entire input expression tree to the destination
763   // type.   Only do this if the dest type is a simple type, don't convert the
764   // expression tree to something weird like i93 unless the source is also
765   // strange.
766   if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
767       canEvaluateTruncated(Src, DestTy, *this, &Trunc)) {
768 
769     // If this cast is a truncate, evaluting in a different type always
770     // eliminates the cast, so it is always a win.
771     LLVM_DEBUG(
772         dbgs() << "ICE: EvaluateInDifferentType converting expression type"
773                   " to avoid cast: "
774                << Trunc << '\n');
775     Value *Res = EvaluateInDifferentType(Src, DestTy, false);
776     assert(Res->getType() == DestTy);
777     return replaceInstUsesWith(Trunc, Res);
778   }
779 
780   // For integer types, check if we can shorten the entire input expression to
781   // DestWidth * 2, which won't allow removing the truncate, but reducing the
782   // width may enable further optimizations, e.g. allowing for larger
783   // vectorization factors.
784   if (auto *DestITy = dyn_cast<IntegerType>(DestTy)) {
785     if (DestWidth * 2 < SrcWidth) {
786       auto *NewDestTy = DestITy->getExtendedType();
787       if (shouldChangeType(SrcTy, NewDestTy) &&
788           canEvaluateTruncated(Src, NewDestTy, *this, &Trunc)) {
789         LLVM_DEBUG(
790             dbgs() << "ICE: EvaluateInDifferentType converting expression type"
791                       " to reduce the width of operand of"
792                    << Trunc << '\n');
793         Value *Res = EvaluateInDifferentType(Src, NewDestTy, false);
794         return new TruncInst(Res, DestTy);
795       }
796     }
797   }
798 
799   // See if we can simplify any instructions used by the input whose sole
800   // purpose is to compute bits we don't care about.
801   if (SimplifyDemandedInstructionBits(Trunc))
802     return &Trunc;
803 
804   if (DestWidth == 1) {
805     Value *Zero = Constant::getNullValue(SrcTy);
806 
807     Value *X;
808     const APInt *C1;
809     Constant *C2;
810     if (match(Src, m_OneUse(m_Shr(m_Shl(m_Power2(C1), m_Value(X)),
811                                   m_ImmConstant(C2))))) {
812       // trunc ((C1 << X) >> C2) to i1 --> X == (C2-cttz(C1)), where C1 is pow2
813       Constant *Log2C1 = ConstantInt::get(SrcTy, C1->exactLogBase2());
814       Constant *CmpC = ConstantExpr::getSub(C2, Log2C1);
815       return new ICmpInst(ICmpInst::ICMP_EQ, X, CmpC);
816     }
817 
818     if (match(Src, m_Shr(m_Value(X), m_SpecificInt(SrcWidth - 1)))) {
819       // trunc (ashr X, BW-1) to i1 --> icmp slt X, 0
820       // trunc (lshr X, BW-1) to i1 --> icmp slt X, 0
821       return new ICmpInst(ICmpInst::ICMP_SLT, X, Zero);
822     }
823 
824     Constant *C;
825     if (match(Src, m_OneUse(m_LShr(m_Value(X), m_ImmConstant(C))))) {
826       // trunc (lshr X, C) to i1 --> icmp ne (and X, C'), 0
827       Constant *One = ConstantInt::get(SrcTy, APInt(SrcWidth, 1));
828       Value *MaskC = Builder.CreateShl(One, C);
829       Value *And = Builder.CreateAnd(X, MaskC);
830       return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
831     }
832     if (match(Src, m_OneUse(m_c_Or(m_LShr(m_Value(X), m_ImmConstant(C)),
833                                    m_Deferred(X))))) {
834       // trunc (or (lshr X, C), X) to i1 --> icmp ne (and X, C'), 0
835       Constant *One = ConstantInt::get(SrcTy, APInt(SrcWidth, 1));
836       Value *MaskC = Builder.CreateShl(One, C);
837       Value *And = Builder.CreateAnd(X, Builder.CreateOr(MaskC, One));
838       return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
839     }
840 
841     {
842       const APInt *C;
843       if (match(Src, m_Shl(m_APInt(C), m_Value(X))) && (*C)[0] == 1) {
844         // trunc (C << X) to i1 --> X == 0, where C is odd
845         return new ICmpInst(ICmpInst::Predicate::ICMP_EQ, X, Zero);
846       }
847     }
848 
849     if (Trunc.hasNoUnsignedWrap() || Trunc.hasNoSignedWrap()) {
850       Value *X, *Y;
851       if (match(Src, m_Xor(m_Value(X), m_Value(Y))))
852         return new ICmpInst(ICmpInst::ICMP_NE, X, Y);
853     }
854   }
855 
856   Value *A, *B;
857   Constant *C;
858   if (match(Src, m_LShr(m_SExt(m_Value(A)), m_Constant(C)))) {
859     unsigned AWidth = A->getType()->getScalarSizeInBits();
860     unsigned MaxShiftAmt = SrcWidth - std::max(DestWidth, AWidth);
861     auto *OldSh = cast<Instruction>(Src);
862     bool IsExact = OldSh->isExact();
863 
864     // If the shift is small enough, all zero bits created by the shift are
865     // removed by the trunc.
866     if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULE,
867                                     APInt(SrcWidth, MaxShiftAmt)))) {
868       auto GetNewShAmt = [&](unsigned Width) {
869         Constant *MaxAmt = ConstantInt::get(SrcTy, Width - 1, false);
870         Constant *Cmp =
871             ConstantFoldCompareInstOperands(ICmpInst::ICMP_ULT, C, MaxAmt, DL);
872         Constant *ShAmt = ConstantFoldSelectInstruction(Cmp, C, MaxAmt);
873         return ConstantFoldCastOperand(Instruction::Trunc, ShAmt, A->getType(),
874                                        DL);
875       };
876 
877       // trunc (lshr (sext A), C) --> ashr A, C
878       if (A->getType() == DestTy) {
879         Constant *ShAmt = GetNewShAmt(DestWidth);
880         ShAmt = Constant::mergeUndefsWith(ShAmt, C);
881         return IsExact ? BinaryOperator::CreateExactAShr(A, ShAmt)
882                        : BinaryOperator::CreateAShr(A, ShAmt);
883       }
884       // The types are mismatched, so create a cast after shifting:
885       // trunc (lshr (sext A), C) --> sext/trunc (ashr A, C)
886       if (Src->hasOneUse()) {
887         Constant *ShAmt = GetNewShAmt(AWidth);
888         Value *Shift = Builder.CreateAShr(A, ShAmt, "", IsExact);
889         return CastInst::CreateIntegerCast(Shift, DestTy, true);
890       }
891     }
892     // TODO: Mask high bits with 'and'.
893   }
894 
895   if (Instruction *I = narrowBinOp(Trunc))
896     return I;
897 
898   if (Instruction *I = shrinkSplatShuffle(Trunc, Builder))
899     return I;
900 
901   if (Instruction *I = shrinkInsertElt(Trunc, Builder))
902     return I;
903 
904   if (Src->hasOneUse() &&
905       (isa<VectorType>(SrcTy) || shouldChangeType(SrcTy, DestTy))) {
906     // Transform "trunc (shl X, cst)" -> "shl (trunc X), cst" so long as the
907     // dest type is native and cst < dest size.
908     if (match(Src, m_Shl(m_Value(A), m_Constant(C))) &&
909         !match(A, m_Shr(m_Value(), m_Constant()))) {
910       // Skip shifts of shift by constants. It undoes a combine in
911       // FoldShiftByConstant and is the extend in reg pattern.
912       APInt Threshold = APInt(C->getType()->getScalarSizeInBits(), DestWidth);
913       if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold))) {
914         Value *NewTrunc = Builder.CreateTrunc(A, DestTy, A->getName() + ".tr");
915         return BinaryOperator::Create(Instruction::Shl, NewTrunc,
916                                       ConstantExpr::getTrunc(C, DestTy));
917       }
918     }
919   }
920 
921   if (Instruction *I = foldVecTruncToExtElt(Trunc, *this))
922     return I;
923 
924   if (Instruction *I = foldVecExtTruncToExtElt(Trunc, *this))
925     return I;
926 
927   // trunc (ctlz_i32(zext(A), B) --> add(ctlz_i16(A, B), C)
928   if (match(Src, m_OneUse(m_Intrinsic<Intrinsic::ctlz>(m_ZExt(m_Value(A)),
929                                                        m_Value(B))))) {
930     unsigned AWidth = A->getType()->getScalarSizeInBits();
931     if (AWidth == DestWidth && AWidth > Log2_32(SrcWidth)) {
932       Value *WidthDiff = ConstantInt::get(A->getType(), SrcWidth - AWidth);
933       Value *NarrowCtlz =
934           Builder.CreateIntrinsic(Intrinsic::ctlz, {Trunc.getType()}, {A, B});
935       return BinaryOperator::CreateAdd(NarrowCtlz, WidthDiff);
936     }
937   }
938 
939   if (match(Src, m_VScale())) {
940     if (Trunc.getFunction() &&
941         Trunc.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
942       Attribute Attr =
943           Trunc.getFunction()->getFnAttribute(Attribute::VScaleRange);
944       if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax())
945         if (Log2_32(*MaxVScale) < DestWidth)
946           return replaceInstUsesWith(Trunc, Builder.CreateVScale(DestTy));
947     }
948   }
949 
950   if (DestWidth == 1 &&
951       (Trunc.hasNoUnsignedWrap() || Trunc.hasNoSignedWrap()) &&
952       isKnownNonZero(Src, SQ.getWithInstruction(&Trunc)))
953     return replaceInstUsesWith(Trunc, ConstantInt::getTrue(DestTy));
954 
955   bool Changed = false;
956   if (!Trunc.hasNoSignedWrap() &&
957       ComputeMaxSignificantBits(Src, &Trunc) <= DestWidth) {
958     Trunc.setHasNoSignedWrap(true);
959     Changed = true;
960   }
961   if (!Trunc.hasNoUnsignedWrap() &&
962       MaskedValueIsZero(Src, APInt::getBitsSetFrom(SrcWidth, DestWidth),
963                         &Trunc)) {
964     Trunc.setHasNoUnsignedWrap(true);
965     Changed = true;
966   }
967 
968   return Changed ? &Trunc : nullptr;
969 }
970 
transformZExtICmp(ICmpInst * Cmp,ZExtInst & Zext)971 Instruction *InstCombinerImpl::transformZExtICmp(ICmpInst *Cmp,
972                                                  ZExtInst &Zext) {
973   // If we are just checking for a icmp eq of a single bit and zext'ing it
974   // to an integer, then shift the bit to the appropriate place and then
975   // cast to integer to avoid the comparison.
976 
977   // FIXME: This set of transforms does not check for extra uses and/or creates
978   //        an extra instruction (an optional final cast is not included
979   //        in the transform comments). We may also want to favor icmp over
980   //        shifts in cases of equal instructions because icmp has better
981   //        analysis in general (invert the transform).
982 
983   const APInt *Op1CV;
984   if (match(Cmp->getOperand(1), m_APInt(Op1CV))) {
985 
986     // zext (x <s  0) to i32 --> x>>u31      true if signbit set.
987     if (Cmp->getPredicate() == ICmpInst::ICMP_SLT && Op1CV->isZero()) {
988       Value *In = Cmp->getOperand(0);
989       Value *Sh = ConstantInt::get(In->getType(),
990                                    In->getType()->getScalarSizeInBits() - 1);
991       In = Builder.CreateLShr(In, Sh, In->getName() + ".lobit");
992       if (In->getType() != Zext.getType())
993         In = Builder.CreateIntCast(In, Zext.getType(), false /*ZExt*/);
994 
995       return replaceInstUsesWith(Zext, In);
996     }
997 
998     // zext (X == 0) to i32 --> X^1      iff X has only the low bit set.
999     // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
1000     // zext (X != 0) to i32 --> X        iff X has only the low bit set.
1001     // zext (X != 0) to i32 --> X>>1     iff X has only the 2nd bit set.
1002 
1003     if (Op1CV->isZero() && Cmp->isEquality()) {
1004       // Exactly 1 possible 1? But not the high-bit because that is
1005       // canonicalized to this form.
1006       KnownBits Known = computeKnownBits(Cmp->getOperand(0), &Zext);
1007       APInt KnownZeroMask(~Known.Zero);
1008       uint32_t ShAmt = KnownZeroMask.logBase2();
1009       bool IsExpectShAmt = KnownZeroMask.isPowerOf2() &&
1010                            (Zext.getType()->getScalarSizeInBits() != ShAmt + 1);
1011       if (IsExpectShAmt &&
1012           (Cmp->getOperand(0)->getType() == Zext.getType() ||
1013            Cmp->getPredicate() == ICmpInst::ICMP_NE || ShAmt == 0)) {
1014         Value *In = Cmp->getOperand(0);
1015         if (ShAmt) {
1016           // Perform a logical shr by shiftamt.
1017           // Insert the shift to put the result in the low bit.
1018           In = Builder.CreateLShr(In, ConstantInt::get(In->getType(), ShAmt),
1019                                   In->getName() + ".lobit");
1020         }
1021 
1022         // Toggle the low bit for "X == 0".
1023         if (Cmp->getPredicate() == ICmpInst::ICMP_EQ)
1024           In = Builder.CreateXor(In, ConstantInt::get(In->getType(), 1));
1025 
1026         if (Zext.getType() == In->getType())
1027           return replaceInstUsesWith(Zext, In);
1028 
1029         Value *IntCast = Builder.CreateIntCast(In, Zext.getType(), false);
1030         return replaceInstUsesWith(Zext, IntCast);
1031       }
1032     }
1033   }
1034 
1035   if (Cmp->isEquality()) {
1036     // Test if a bit is clear/set using a shifted-one mask:
1037     // zext (icmp eq (and X, (1 << ShAmt)), 0) --> and (lshr (not X), ShAmt), 1
1038     // zext (icmp ne (and X, (1 << ShAmt)), 0) --> and (lshr X, ShAmt), 1
1039     Value *X, *ShAmt;
1040     if (Cmp->hasOneUse() && match(Cmp->getOperand(1), m_ZeroInt()) &&
1041         match(Cmp->getOperand(0),
1042               m_OneUse(m_c_And(m_Shl(m_One(), m_Value(ShAmt)), m_Value(X))))) {
1043       auto *And = cast<BinaryOperator>(Cmp->getOperand(0));
1044       Value *Shift = And->getOperand(X == And->getOperand(0) ? 1 : 0);
1045       if (Zext.getType() == And->getType() ||
1046           Cmp->getPredicate() != ICmpInst::ICMP_EQ || Shift->hasOneUse()) {
1047         if (Cmp->getPredicate() == ICmpInst::ICMP_EQ)
1048           X = Builder.CreateNot(X);
1049         Value *Lshr = Builder.CreateLShr(X, ShAmt);
1050         Value *And1 =
1051             Builder.CreateAnd(Lshr, ConstantInt::get(X->getType(), 1));
1052         return replaceInstUsesWith(
1053             Zext, Builder.CreateZExtOrTrunc(And1, Zext.getType()));
1054       }
1055     }
1056   }
1057 
1058   return nullptr;
1059 }
1060 
1061 /// Determine if the specified value can be computed in the specified wider type
1062 /// and produce the same low bits. If not, return false.
1063 ///
1064 /// If this function returns true, it can also return a non-zero number of bits
1065 /// (in BitsToClear) which indicates that the value it computes is correct for
1066 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
1067 /// out.  For example, to promote something like:
1068 ///
1069 ///   %B = trunc i64 %A to i32
1070 ///   %C = lshr i32 %B, 8
1071 ///   %E = zext i32 %C to i64
1072 ///
1073 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
1074 /// set to 8 to indicate that the promoted value needs to have bits 24-31
1075 /// cleared in addition to bits 32-63.  Since an 'and' will be generated to
1076 /// clear the top bits anyway, doing this has no extra cost.
1077 ///
1078 /// This function works on both vectors and scalars.
canEvaluateZExtd(Value * V,Type * Ty,unsigned & BitsToClear,InstCombinerImpl & IC,Instruction * CxtI)1079 static bool canEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
1080                              InstCombinerImpl &IC, Instruction *CxtI) {
1081   BitsToClear = 0;
1082   if (canAlwaysEvaluateInType(V, Ty))
1083     return true;
1084   if (canNotEvaluateInType(V, Ty))
1085     return false;
1086 
1087   auto *I = cast<Instruction>(V);
1088   unsigned Tmp;
1089   switch (I->getOpcode()) {
1090   case Instruction::ZExt:  // zext(zext(x)) -> zext(x).
1091   case Instruction::SExt:  // zext(sext(x)) -> sext(x).
1092   case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
1093     return true;
1094   case Instruction::And:
1095   case Instruction::Or:
1096   case Instruction::Xor:
1097   case Instruction::Add:
1098   case Instruction::Sub:
1099   case Instruction::Mul:
1100     if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
1101         !canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
1102       return false;
1103     // These can all be promoted if neither operand has 'bits to clear'.
1104     if (BitsToClear == 0 && Tmp == 0)
1105       return true;
1106 
1107     // If the operation is an AND/OR/XOR and the bits to clear are zero in the
1108     // other side, BitsToClear is ok.
1109     if (Tmp == 0 && I->isBitwiseLogicOp()) {
1110       // We use MaskedValueIsZero here for generality, but the case we care
1111       // about the most is constant RHS.
1112       unsigned VSize = V->getType()->getScalarSizeInBits();
1113       if (IC.MaskedValueIsZero(I->getOperand(1),
1114                                APInt::getHighBitsSet(VSize, BitsToClear),
1115                                CxtI)) {
1116         // If this is an And instruction and all of the BitsToClear are
1117         // known to be zero we can reset BitsToClear.
1118         if (I->getOpcode() == Instruction::And)
1119           BitsToClear = 0;
1120         return true;
1121       }
1122     }
1123 
1124     // Otherwise, we don't know how to analyze this BitsToClear case yet.
1125     return false;
1126 
1127   case Instruction::Shl: {
1128     // We can promote shl(x, cst) if we can promote x.  Since shl overwrites the
1129     // upper bits we can reduce BitsToClear by the shift amount.
1130     const APInt *Amt;
1131     if (match(I->getOperand(1), m_APInt(Amt))) {
1132       if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
1133         return false;
1134       uint64_t ShiftAmt = Amt->getZExtValue();
1135       BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
1136       return true;
1137     }
1138     return false;
1139   }
1140   case Instruction::LShr: {
1141     // We can promote lshr(x, cst) if we can promote x.  This requires the
1142     // ultimate 'and' to clear out the high zero bits we're clearing out though.
1143     const APInt *Amt;
1144     if (match(I->getOperand(1), m_APInt(Amt))) {
1145       if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
1146         return false;
1147       BitsToClear += Amt->getZExtValue();
1148       if (BitsToClear > V->getType()->getScalarSizeInBits())
1149         BitsToClear = V->getType()->getScalarSizeInBits();
1150       return true;
1151     }
1152     // Cannot promote variable LSHR.
1153     return false;
1154   }
1155   case Instruction::Select:
1156     if (!canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
1157         !canEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
1158         // TODO: If important, we could handle the case when the BitsToClear are
1159         // known zero in the disagreeing side.
1160         Tmp != BitsToClear)
1161       return false;
1162     return true;
1163 
1164   case Instruction::PHI: {
1165     // We can change a phi if we can change all operands.  Note that we never
1166     // get into trouble with cyclic PHIs here because we only consider
1167     // instructions with a single use.
1168     PHINode *PN = cast<PHINode>(I);
1169     if (!canEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
1170       return false;
1171     for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
1172       if (!canEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
1173           // TODO: If important, we could handle the case when the BitsToClear
1174           // are known zero in the disagreeing input.
1175           Tmp != BitsToClear)
1176         return false;
1177     return true;
1178   }
1179   case Instruction::Call:
1180     // llvm.vscale() can always be executed in larger type, because the
1181     // value is automatically zero-extended.
1182     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1183       if (II->getIntrinsicID() == Intrinsic::vscale)
1184         return true;
1185     return false;
1186   default:
1187     // TODO: Can handle more cases here.
1188     return false;
1189   }
1190 }
1191 
visitZExt(ZExtInst & Zext)1192 Instruction *InstCombinerImpl::visitZExt(ZExtInst &Zext) {
1193   // If this zero extend is only used by a truncate, let the truncate be
1194   // eliminated before we try to optimize this zext.
1195   if (Zext.hasOneUse() && isa<TruncInst>(Zext.user_back()) &&
1196       !isa<Constant>(Zext.getOperand(0)))
1197     return nullptr;
1198 
1199   // If one of the common conversion will work, do it.
1200   if (Instruction *Result = commonCastTransforms(Zext))
1201     return Result;
1202 
1203   Value *Src = Zext.getOperand(0);
1204   Type *SrcTy = Src->getType(), *DestTy = Zext.getType();
1205 
1206   // zext nneg bool x -> 0
1207   if (SrcTy->isIntOrIntVectorTy(1) && Zext.hasNonNeg())
1208     return replaceInstUsesWith(Zext, Constant::getNullValue(Zext.getType()));
1209 
1210   // Try to extend the entire expression tree to the wide destination type.
1211   unsigned BitsToClear;
1212   if (shouldChangeType(SrcTy, DestTy) &&
1213       canEvaluateZExtd(Src, DestTy, BitsToClear, *this, &Zext)) {
1214     assert(BitsToClear <= SrcTy->getScalarSizeInBits() &&
1215            "Can't clear more bits than in SrcTy");
1216 
1217     // Okay, we can transform this!  Insert the new expression now.
1218     LLVM_DEBUG(
1219         dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1220                   " to avoid zero extend: "
1221                << Zext << '\n');
1222     Value *Res = EvaluateInDifferentType(Src, DestTy, false);
1223     assert(Res->getType() == DestTy);
1224 
1225     // Preserve debug values referring to Src if the zext is its last use.
1226     if (auto *SrcOp = dyn_cast<Instruction>(Src))
1227       if (SrcOp->hasOneUse())
1228         replaceAllDbgUsesWith(*SrcOp, *Res, Zext, DT);
1229 
1230     uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits() - BitsToClear;
1231     uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1232 
1233     // If the high bits are already filled with zeros, just replace this
1234     // cast with the result.
1235     if (MaskedValueIsZero(
1236             Res, APInt::getHighBitsSet(DestBitSize, DestBitSize - SrcBitsKept),
1237             &Zext))
1238       return replaceInstUsesWith(Zext, Res);
1239 
1240     // We need to emit an AND to clear the high bits.
1241     Constant *C = ConstantInt::get(Res->getType(),
1242                                APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
1243     return BinaryOperator::CreateAnd(Res, C);
1244   }
1245 
1246   // If this is a TRUNC followed by a ZEXT then we are dealing with integral
1247   // types and if the sizes are just right we can convert this into a logical
1248   // 'and' which will be much cheaper than the pair of casts.
1249   if (auto *CSrc = dyn_cast<TruncInst>(Src)) {   // A->B->C cast
1250     // TODO: Subsume this into EvaluateInDifferentType.
1251 
1252     // Get the sizes of the types involved.  We know that the intermediate type
1253     // will be smaller than A or C, but don't know the relation between A and C.
1254     Value *A = CSrc->getOperand(0);
1255     unsigned SrcSize = A->getType()->getScalarSizeInBits();
1256     unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
1257     unsigned DstSize = DestTy->getScalarSizeInBits();
1258     // If we're actually extending zero bits, then if
1259     // SrcSize <  DstSize: zext(a & mask)
1260     // SrcSize == DstSize: a & mask
1261     // SrcSize  > DstSize: trunc(a) & mask
1262     if (SrcSize < DstSize) {
1263       APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
1264       Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
1265       Value *And = Builder.CreateAnd(A, AndConst, CSrc->getName() + ".mask");
1266       return new ZExtInst(And, DestTy);
1267     }
1268 
1269     if (SrcSize == DstSize) {
1270       APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
1271       return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
1272                                                            AndValue));
1273     }
1274     if (SrcSize > DstSize) {
1275       Value *Trunc = Builder.CreateTrunc(A, DestTy);
1276       APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
1277       return BinaryOperator::CreateAnd(Trunc,
1278                                        ConstantInt::get(Trunc->getType(),
1279                                                         AndValue));
1280     }
1281   }
1282 
1283   if (auto *Cmp = dyn_cast<ICmpInst>(Src))
1284     return transformZExtICmp(Cmp, Zext);
1285 
1286   // zext(trunc(X) & C) -> (X & zext(C)).
1287   Constant *C;
1288   Value *X;
1289   if (match(Src, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
1290       X->getType() == DestTy)
1291     return BinaryOperator::CreateAnd(X, Builder.CreateZExt(C, DestTy));
1292 
1293   // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
1294   Value *And;
1295   if (match(Src, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
1296       match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
1297       X->getType() == DestTy) {
1298     Value *ZC = Builder.CreateZExt(C, DestTy);
1299     return BinaryOperator::CreateXor(Builder.CreateAnd(X, ZC), ZC);
1300   }
1301 
1302   // If we are truncating, masking, and then zexting back to the original type,
1303   // that's just a mask. This is not handled by canEvaluateZextd if the
1304   // intermediate values have extra uses. This could be generalized further for
1305   // a non-constant mask operand.
1306   // zext (and (trunc X), C) --> and X, (zext C)
1307   if (match(Src, m_And(m_Trunc(m_Value(X)), m_Constant(C))) &&
1308       X->getType() == DestTy) {
1309     Value *ZextC = Builder.CreateZExt(C, DestTy);
1310     return BinaryOperator::CreateAnd(X, ZextC);
1311   }
1312 
1313   if (match(Src, m_VScale())) {
1314     if (Zext.getFunction() &&
1315         Zext.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
1316       Attribute Attr =
1317           Zext.getFunction()->getFnAttribute(Attribute::VScaleRange);
1318       if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax()) {
1319         unsigned TypeWidth = Src->getType()->getScalarSizeInBits();
1320         if (Log2_32(*MaxVScale) < TypeWidth)
1321           return replaceInstUsesWith(Zext, Builder.CreateVScale(DestTy));
1322       }
1323     }
1324   }
1325 
1326   if (!Zext.hasNonNeg()) {
1327     // If this zero extend is only used by a shift, add nneg flag.
1328     if (Zext.hasOneUse() &&
1329         SrcTy->getScalarSizeInBits() >
1330             Log2_64_Ceil(DestTy->getScalarSizeInBits()) &&
1331         match(Zext.user_back(), m_Shift(m_Value(), m_Specific(&Zext)))) {
1332       Zext.setNonNeg();
1333       return &Zext;
1334     }
1335 
1336     if (isKnownNonNegative(Src, SQ.getWithInstruction(&Zext))) {
1337       Zext.setNonNeg();
1338       return &Zext;
1339     }
1340   }
1341 
1342   return nullptr;
1343 }
1344 
1345 /// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp.
transformSExtICmp(ICmpInst * Cmp,SExtInst & Sext)1346 Instruction *InstCombinerImpl::transformSExtICmp(ICmpInst *Cmp,
1347                                                  SExtInst &Sext) {
1348   Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1349   ICmpInst::Predicate Pred = Cmp->getPredicate();
1350 
1351   // Don't bother if Op1 isn't of vector or integer type.
1352   if (!Op1->getType()->isIntOrIntVectorTy())
1353     return nullptr;
1354 
1355   if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_ZeroInt())) {
1356     // sext (x <s 0) --> ashr x, 31 (all ones if negative)
1357     Value *Sh = ConstantInt::get(Op0->getType(),
1358                                  Op0->getType()->getScalarSizeInBits() - 1);
1359     Value *In = Builder.CreateAShr(Op0, Sh, Op0->getName() + ".lobit");
1360     if (In->getType() != Sext.getType())
1361       In = Builder.CreateIntCast(In, Sext.getType(), true /*SExt*/);
1362 
1363     return replaceInstUsesWith(Sext, In);
1364   }
1365 
1366   if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1367     // If we know that only one bit of the LHS of the icmp can be set and we
1368     // have an equality comparison with zero or a power of 2, we can transform
1369     // the icmp and sext into bitwise/integer operations.
1370     if (Cmp->hasOneUse() &&
1371         Cmp->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
1372       KnownBits Known = computeKnownBits(Op0, &Sext);
1373 
1374       APInt KnownZeroMask(~Known.Zero);
1375       if (KnownZeroMask.isPowerOf2()) {
1376         Value *In = Cmp->getOperand(0);
1377 
1378         // If the icmp tests for a known zero bit we can constant fold it.
1379         if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
1380           Value *V = Pred == ICmpInst::ICMP_NE ?
1381                        ConstantInt::getAllOnesValue(Sext.getType()) :
1382                        ConstantInt::getNullValue(Sext.getType());
1383           return replaceInstUsesWith(Sext, V);
1384         }
1385 
1386         if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
1387           // sext ((x & 2^n) == 0)   -> (x >> n) - 1
1388           // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
1389           unsigned ShiftAmt = KnownZeroMask.countr_zero();
1390           // Perform a right shift to place the desired bit in the LSB.
1391           if (ShiftAmt)
1392             In = Builder.CreateLShr(In,
1393                                     ConstantInt::get(In->getType(), ShiftAmt));
1394 
1395           // At this point "In" is either 1 or 0. Subtract 1 to turn
1396           // {1, 0} -> {0, -1}.
1397           In = Builder.CreateAdd(In,
1398                                  ConstantInt::getAllOnesValue(In->getType()),
1399                                  "sext");
1400         } else {
1401           // sext ((x & 2^n) != 0)   -> (x << bitwidth-n) a>> bitwidth-1
1402           // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
1403           unsigned ShiftAmt = KnownZeroMask.countl_zero();
1404           // Perform a left shift to place the desired bit in the MSB.
1405           if (ShiftAmt)
1406             In = Builder.CreateShl(In,
1407                                    ConstantInt::get(In->getType(), ShiftAmt));
1408 
1409           // Distribute the bit over the whole bit width.
1410           In = Builder.CreateAShr(In, ConstantInt::get(In->getType(),
1411                                   KnownZeroMask.getBitWidth() - 1), "sext");
1412         }
1413 
1414         if (Sext.getType() == In->getType())
1415           return replaceInstUsesWith(Sext, In);
1416         return CastInst::CreateIntegerCast(In, Sext.getType(), true/*SExt*/);
1417       }
1418     }
1419   }
1420 
1421   return nullptr;
1422 }
1423 
1424 /// Return true if we can take the specified value and return it as type Ty
1425 /// without inserting any new casts and without changing the value of the common
1426 /// low bits.  This is used by code that tries to promote integer operations to
1427 /// a wider types will allow us to eliminate the extension.
1428 ///
1429 /// This function works on both vectors and scalars.
1430 ///
canEvaluateSExtd(Value * V,Type * Ty)1431 static bool canEvaluateSExtd(Value *V, Type *Ty) {
1432   assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
1433          "Can't sign extend type to a smaller type");
1434   if (canAlwaysEvaluateInType(V, Ty))
1435     return true;
1436   if (canNotEvaluateInType(V, Ty))
1437     return false;
1438 
1439   auto *I = cast<Instruction>(V);
1440   switch (I->getOpcode()) {
1441   case Instruction::SExt:  // sext(sext(x)) -> sext(x)
1442   case Instruction::ZExt:  // sext(zext(x)) -> zext(x)
1443   case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1444     return true;
1445   case Instruction::And:
1446   case Instruction::Or:
1447   case Instruction::Xor:
1448   case Instruction::Add:
1449   case Instruction::Sub:
1450   case Instruction::Mul:
1451     // These operators can all arbitrarily be extended if their inputs can.
1452     return canEvaluateSExtd(I->getOperand(0), Ty) &&
1453            canEvaluateSExtd(I->getOperand(1), Ty);
1454 
1455   //case Instruction::Shl:   TODO
1456   //case Instruction::LShr:  TODO
1457 
1458   case Instruction::Select:
1459     return canEvaluateSExtd(I->getOperand(1), Ty) &&
1460            canEvaluateSExtd(I->getOperand(2), Ty);
1461 
1462   case Instruction::PHI: {
1463     // We can change a phi if we can change all operands.  Note that we never
1464     // get into trouble with cyclic PHIs here because we only consider
1465     // instructions with a single use.
1466     PHINode *PN = cast<PHINode>(I);
1467     for (Value *IncValue : PN->incoming_values())
1468       if (!canEvaluateSExtd(IncValue, Ty)) return false;
1469     return true;
1470   }
1471   default:
1472     // TODO: Can handle more cases here.
1473     break;
1474   }
1475 
1476   return false;
1477 }
1478 
visitSExt(SExtInst & Sext)1479 Instruction *InstCombinerImpl::visitSExt(SExtInst &Sext) {
1480   // If this sign extend is only used by a truncate, let the truncate be
1481   // eliminated before we try to optimize this sext.
1482   if (Sext.hasOneUse() && isa<TruncInst>(Sext.user_back()))
1483     return nullptr;
1484 
1485   if (Instruction *I = commonCastTransforms(Sext))
1486     return I;
1487 
1488   Value *Src = Sext.getOperand(0);
1489   Type *SrcTy = Src->getType(), *DestTy = Sext.getType();
1490   unsigned SrcBitSize = SrcTy->getScalarSizeInBits();
1491   unsigned DestBitSize = DestTy->getScalarSizeInBits();
1492 
1493   // If the value being extended is zero or positive, use a zext instead.
1494   if (isKnownNonNegative(Src, SQ.getWithInstruction(&Sext))) {
1495     auto CI = CastInst::Create(Instruction::ZExt, Src, DestTy);
1496     CI->setNonNeg(true);
1497     return CI;
1498   }
1499 
1500   // Try to extend the entire expression tree to the wide destination type.
1501   if (shouldChangeType(SrcTy, DestTy) && canEvaluateSExtd(Src, DestTy)) {
1502     // Okay, we can transform this!  Insert the new expression now.
1503     LLVM_DEBUG(
1504         dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1505                   " to avoid sign extend: "
1506                << Sext << '\n');
1507     Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1508     assert(Res->getType() == DestTy);
1509 
1510     // If the high bits are already filled with sign bit, just replace this
1511     // cast with the result.
1512     if (ComputeNumSignBits(Res, &Sext) > DestBitSize - SrcBitSize)
1513       return replaceInstUsesWith(Sext, Res);
1514 
1515     // We need to emit a shl + ashr to do the sign extend.
1516     Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1517     return BinaryOperator::CreateAShr(Builder.CreateShl(Res, ShAmt, "sext"),
1518                                       ShAmt);
1519   }
1520 
1521   Value *X;
1522   if (match(Src, m_Trunc(m_Value(X)))) {
1523     // If the input has more sign bits than bits truncated, then convert
1524     // directly to final type.
1525     unsigned XBitSize = X->getType()->getScalarSizeInBits();
1526     if (ComputeNumSignBits(X, &Sext) > XBitSize - SrcBitSize)
1527       return CastInst::CreateIntegerCast(X, DestTy, /* isSigned */ true);
1528 
1529     // If input is a trunc from the destination type, then convert into shifts.
1530     if (Src->hasOneUse() && X->getType() == DestTy) {
1531       // sext (trunc X) --> ashr (shl X, C), C
1532       Constant *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
1533       return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShAmt), ShAmt);
1534     }
1535 
1536     // If we are replacing shifted-in high zero bits with sign bits, convert
1537     // the logic shift to arithmetic shift and eliminate the cast to
1538     // intermediate type:
1539     // sext (trunc (lshr Y, C)) --> sext/trunc (ashr Y, C)
1540     Value *Y;
1541     if (Src->hasOneUse() &&
1542         match(X, m_LShr(m_Value(Y),
1543                         m_SpecificIntAllowPoison(XBitSize - SrcBitSize)))) {
1544       Value *Ashr = Builder.CreateAShr(Y, XBitSize - SrcBitSize);
1545       return CastInst::CreateIntegerCast(Ashr, DestTy, /* isSigned */ true);
1546     }
1547   }
1548 
1549   if (auto *Cmp = dyn_cast<ICmpInst>(Src))
1550     return transformSExtICmp(Cmp, Sext);
1551 
1552   // If the input is a shl/ashr pair of a same constant, then this is a sign
1553   // extension from a smaller value.  If we could trust arbitrary bitwidth
1554   // integers, we could turn this into a truncate to the smaller bit and then
1555   // use a sext for the whole extension.  Since we don't, look deeper and check
1556   // for a truncate.  If the source and dest are the same type, eliminate the
1557   // trunc and extend and just do shifts.  For example, turn:
1558   //   %a = trunc i32 %i to i8
1559   //   %b = shl i8 %a, C
1560   //   %c = ashr i8 %b, C
1561   //   %d = sext i8 %c to i32
1562   // into:
1563   //   %a = shl i32 %i, 32-(8-C)
1564   //   %d = ashr i32 %a, 32-(8-C)
1565   Value *A = nullptr;
1566   // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1567   Constant *BA = nullptr, *CA = nullptr;
1568   if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_Constant(BA)),
1569                         m_ImmConstant(CA))) &&
1570       BA->isElementWiseEqual(CA) && A->getType() == DestTy) {
1571     Constant *WideCurrShAmt =
1572         ConstantFoldCastOperand(Instruction::SExt, CA, DestTy, DL);
1573     assert(WideCurrShAmt && "Constant folding of ImmConstant cannot fail");
1574     Constant *NumLowbitsLeft = ConstantExpr::getSub(
1575         ConstantInt::get(DestTy, SrcTy->getScalarSizeInBits()), WideCurrShAmt);
1576     Constant *NewShAmt = ConstantExpr::getSub(
1577         ConstantInt::get(DestTy, DestTy->getScalarSizeInBits()),
1578         NumLowbitsLeft);
1579     NewShAmt =
1580         Constant::mergeUndefsWith(Constant::mergeUndefsWith(NewShAmt, BA), CA);
1581     A = Builder.CreateShl(A, NewShAmt, Sext.getName());
1582     return BinaryOperator::CreateAShr(A, NewShAmt);
1583   }
1584 
1585   // Splatting a bit of constant-index across a value:
1586   // sext (ashr (trunc iN X to iM), M-1) to iN --> ashr (shl X, N-M), N-1
1587   // If the dest type is different, use a cast (adjust use check).
1588   if (match(Src, m_OneUse(m_AShr(m_Trunc(m_Value(X)),
1589                                  m_SpecificInt(SrcBitSize - 1))))) {
1590     Type *XTy = X->getType();
1591     unsigned XBitSize = XTy->getScalarSizeInBits();
1592     Constant *ShlAmtC = ConstantInt::get(XTy, XBitSize - SrcBitSize);
1593     Constant *AshrAmtC = ConstantInt::get(XTy, XBitSize - 1);
1594     if (XTy == DestTy)
1595       return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShlAmtC),
1596                                         AshrAmtC);
1597     if (cast<BinaryOperator>(Src)->getOperand(0)->hasOneUse()) {
1598       Value *Ashr = Builder.CreateAShr(Builder.CreateShl(X, ShlAmtC), AshrAmtC);
1599       return CastInst::CreateIntegerCast(Ashr, DestTy, /* isSigned */ true);
1600     }
1601   }
1602 
1603   if (match(Src, m_VScale())) {
1604     if (Sext.getFunction() &&
1605         Sext.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
1606       Attribute Attr =
1607           Sext.getFunction()->getFnAttribute(Attribute::VScaleRange);
1608       if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax())
1609         if (Log2_32(*MaxVScale) < (SrcBitSize - 1))
1610           return replaceInstUsesWith(Sext, Builder.CreateVScale(DestTy));
1611     }
1612   }
1613 
1614   return nullptr;
1615 }
1616 
1617 /// Return a Constant* for the specified floating-point constant if it fits
1618 /// in the specified FP type without changing its value.
fitsInFPType(ConstantFP * CFP,const fltSemantics & Sem)1619 static bool fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1620   bool losesInfo;
1621   APFloat F = CFP->getValueAPF();
1622   (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1623   return !losesInfo;
1624 }
1625 
shrinkFPConstant(ConstantFP * CFP,bool PreferBFloat)1626 static Type *shrinkFPConstant(ConstantFP *CFP, bool PreferBFloat) {
1627   if (CFP->getType() == Type::getPPC_FP128Ty(CFP->getContext()))
1628     return nullptr;  // No constant folding of this.
1629   // See if the value can be truncated to bfloat and then reextended.
1630   if (PreferBFloat && fitsInFPType(CFP, APFloat::BFloat()))
1631     return Type::getBFloatTy(CFP->getContext());
1632   // See if the value can be truncated to half and then reextended.
1633   if (!PreferBFloat && fitsInFPType(CFP, APFloat::IEEEhalf()))
1634     return Type::getHalfTy(CFP->getContext());
1635   // See if the value can be truncated to float and then reextended.
1636   if (fitsInFPType(CFP, APFloat::IEEEsingle()))
1637     return Type::getFloatTy(CFP->getContext());
1638   if (CFP->getType()->isDoubleTy())
1639     return nullptr;  // Won't shrink.
1640   if (fitsInFPType(CFP, APFloat::IEEEdouble()))
1641     return Type::getDoubleTy(CFP->getContext());
1642   // Don't try to shrink to various long double types.
1643   return nullptr;
1644 }
1645 
1646 // Determine if this is a vector of ConstantFPs and if so, return the minimal
1647 // type we can safely truncate all elements to.
shrinkFPConstantVector(Value * V,bool PreferBFloat)1648 static Type *shrinkFPConstantVector(Value *V, bool PreferBFloat) {
1649   auto *CV = dyn_cast<Constant>(V);
1650   auto *CVVTy = dyn_cast<FixedVectorType>(V->getType());
1651   if (!CV || !CVVTy)
1652     return nullptr;
1653 
1654   Type *MinType = nullptr;
1655 
1656   unsigned NumElts = CVVTy->getNumElements();
1657 
1658   // For fixed-width vectors we find the minimal type by looking
1659   // through the constant values of the vector.
1660   for (unsigned i = 0; i != NumElts; ++i) {
1661     if (isa<UndefValue>(CV->getAggregateElement(i)))
1662       continue;
1663 
1664     auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
1665     if (!CFP)
1666       return nullptr;
1667 
1668     Type *T = shrinkFPConstant(CFP, PreferBFloat);
1669     if (!T)
1670       return nullptr;
1671 
1672     // If we haven't found a type yet or this type has a larger mantissa than
1673     // our previous type, this is our new minimal type.
1674     if (!MinType || T->getFPMantissaWidth() > MinType->getFPMantissaWidth())
1675       MinType = T;
1676   }
1677 
1678   // Make a vector type from the minimal type.
1679   return MinType ? FixedVectorType::get(MinType, NumElts) : nullptr;
1680 }
1681 
1682 /// Find the minimum FP type we can safely truncate to.
getMinimumFPType(Value * V,bool PreferBFloat)1683 static Type *getMinimumFPType(Value *V, bool PreferBFloat) {
1684   if (auto *FPExt = dyn_cast<FPExtInst>(V))
1685     return FPExt->getOperand(0)->getType();
1686 
1687   // If this value is a constant, return the constant in the smallest FP type
1688   // that can accurately represent it.  This allows us to turn
1689   // (float)((double)X+2.0) into x+2.0f.
1690   if (auto *CFP = dyn_cast<ConstantFP>(V))
1691     if (Type *T = shrinkFPConstant(CFP, PreferBFloat))
1692       return T;
1693 
1694   // Try to shrink scalable and fixed splat vectors.
1695   if (auto *FPC = dyn_cast<Constant>(V))
1696     if (isa<VectorType>(V->getType()))
1697       if (auto *Splat = dyn_cast_or_null<ConstantFP>(FPC->getSplatValue()))
1698         if (Type *T = shrinkFPConstant(Splat, PreferBFloat))
1699           return T;
1700 
1701   // Try to shrink a vector of FP constants. This returns nullptr on scalable
1702   // vectors
1703   if (Type *T = shrinkFPConstantVector(V, PreferBFloat))
1704     return T;
1705 
1706   return V->getType();
1707 }
1708 
1709 /// Return true if the cast from integer to FP can be proven to be exact for all
1710 /// possible inputs (the conversion does not lose any precision).
isKnownExactCastIntToFP(CastInst & I,InstCombinerImpl & IC)1711 static bool isKnownExactCastIntToFP(CastInst &I, InstCombinerImpl &IC) {
1712   CastInst::CastOps Opcode = I.getOpcode();
1713   assert((Opcode == CastInst::SIToFP || Opcode == CastInst::UIToFP) &&
1714          "Unexpected cast");
1715   Value *Src = I.getOperand(0);
1716   Type *SrcTy = Src->getType();
1717   Type *FPTy = I.getType();
1718   bool IsSigned = Opcode == Instruction::SIToFP;
1719   int SrcSize = (int)SrcTy->getScalarSizeInBits() - IsSigned;
1720 
1721   // Easy case - if the source integer type has less bits than the FP mantissa,
1722   // then the cast must be exact.
1723   int DestNumSigBits = FPTy->getFPMantissaWidth();
1724   if (SrcSize <= DestNumSigBits)
1725     return true;
1726 
1727   // Cast from FP to integer and back to FP is independent of the intermediate
1728   // integer width because of poison on overflow.
1729   Value *F;
1730   if (match(Src, m_FPToSI(m_Value(F))) || match(Src, m_FPToUI(m_Value(F)))) {
1731     // If this is uitofp (fptosi F), the source needs an extra bit to avoid
1732     // potential rounding of negative FP input values.
1733     int SrcNumSigBits = F->getType()->getFPMantissaWidth();
1734     if (!IsSigned && match(Src, m_FPToSI(m_Value())))
1735       SrcNumSigBits++;
1736 
1737     // [su]itofp (fpto[su]i F) --> exact if the source type has less or equal
1738     // significant bits than the destination (and make sure neither type is
1739     // weird -- ppc_fp128).
1740     if (SrcNumSigBits > 0 && DestNumSigBits > 0 &&
1741         SrcNumSigBits <= DestNumSigBits)
1742       return true;
1743   }
1744 
1745   // TODO:
1746   // Try harder to find if the source integer type has less significant bits.
1747   // For example, compute number of sign bits.
1748   KnownBits SrcKnown = IC.computeKnownBits(Src, &I);
1749   int SigBits = (int)SrcTy->getScalarSizeInBits() -
1750                 SrcKnown.countMinLeadingZeros() -
1751                 SrcKnown.countMinTrailingZeros();
1752   if (SigBits <= DestNumSigBits)
1753     return true;
1754 
1755   return false;
1756 }
1757 
visitFPTrunc(FPTruncInst & FPT)1758 Instruction *InstCombinerImpl::visitFPTrunc(FPTruncInst &FPT) {
1759   if (Instruction *I = commonCastTransforms(FPT))
1760     return I;
1761 
1762   // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1763   // simplify this expression to avoid one or more of the trunc/extend
1764   // operations if we can do so without changing the numerical results.
1765   //
1766   // The exact manner in which the widths of the operands interact to limit
1767   // what we can and cannot do safely varies from operation to operation, and
1768   // is explained below in the various case statements.
1769   Type *Ty = FPT.getType();
1770   auto *BO = dyn_cast<BinaryOperator>(FPT.getOperand(0));
1771   if (BO && BO->hasOneUse()) {
1772     Type *LHSMinType =
1773         getMinimumFPType(BO->getOperand(0), /*PreferBFloat=*/Ty->isBFloatTy());
1774     Type *RHSMinType =
1775         getMinimumFPType(BO->getOperand(1), /*PreferBFloat=*/Ty->isBFloatTy());
1776     unsigned OpWidth = BO->getType()->getFPMantissaWidth();
1777     unsigned LHSWidth = LHSMinType->getFPMantissaWidth();
1778     unsigned RHSWidth = RHSMinType->getFPMantissaWidth();
1779     unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1780     unsigned DstWidth = Ty->getFPMantissaWidth();
1781     switch (BO->getOpcode()) {
1782       default: break;
1783       case Instruction::FAdd:
1784       case Instruction::FSub:
1785         // For addition and subtraction, the infinitely precise result can
1786         // essentially be arbitrarily wide; proving that double rounding
1787         // will not occur because the result of OpI is exact (as we will for
1788         // FMul, for example) is hopeless.  However, we *can* nonetheless
1789         // frequently know that double rounding cannot occur (or that it is
1790         // innocuous) by taking advantage of the specific structure of
1791         // infinitely-precise results that admit double rounding.
1792         //
1793         // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1794         // to represent both sources, we can guarantee that the double
1795         // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1796         // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1797         // for proof of this fact).
1798         //
1799         // Note: Figueroa does not consider the case where DstFormat !=
1800         // SrcFormat.  It's possible (likely even!) that this analysis
1801         // could be tightened for those cases, but they are rare (the main
1802         // case of interest here is (float)((double)float + float)).
1803         if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1804           Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);
1805           Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);
1806           Instruction *RI = BinaryOperator::Create(BO->getOpcode(), LHS, RHS);
1807           RI->copyFastMathFlags(BO);
1808           return RI;
1809         }
1810         break;
1811       case Instruction::FMul:
1812         // For multiplication, the infinitely precise result has at most
1813         // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1814         // that such a value can be exactly represented, then no double
1815         // rounding can possibly occur; we can safely perform the operation
1816         // in the destination format if it can represent both sources.
1817         if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1818           Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);
1819           Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);
1820           return BinaryOperator::CreateFMulFMF(LHS, RHS, BO);
1821         }
1822         break;
1823       case Instruction::FDiv:
1824         // For division, we use again use the bound from Figueroa's
1825         // dissertation.  I am entirely certain that this bound can be
1826         // tightened in the unbalanced operand case by an analysis based on
1827         // the diophantine rational approximation bound, but the well-known
1828         // condition used here is a good conservative first pass.
1829         // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1830         if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1831           Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);
1832           Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);
1833           return BinaryOperator::CreateFDivFMF(LHS, RHS, BO);
1834         }
1835         break;
1836       case Instruction::FRem: {
1837         // Remainder is straightforward.  Remainder is always exact, so the
1838         // type of OpI doesn't enter into things at all.  We simply evaluate
1839         // in whichever source type is larger, then convert to the
1840         // destination type.
1841         if (SrcWidth == OpWidth)
1842           break;
1843         Value *LHS, *RHS;
1844         if (LHSWidth == SrcWidth) {
1845            LHS = Builder.CreateFPTrunc(BO->getOperand(0), LHSMinType);
1846            RHS = Builder.CreateFPTrunc(BO->getOperand(1), LHSMinType);
1847         } else {
1848            LHS = Builder.CreateFPTrunc(BO->getOperand(0), RHSMinType);
1849            RHS = Builder.CreateFPTrunc(BO->getOperand(1), RHSMinType);
1850         }
1851 
1852         Value *ExactResult = Builder.CreateFRemFMF(LHS, RHS, BO);
1853         return CastInst::CreateFPCast(ExactResult, Ty);
1854       }
1855     }
1856   }
1857 
1858   // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1859   Value *X;
1860   Instruction *Op = dyn_cast<Instruction>(FPT.getOperand(0));
1861   if (Op && Op->hasOneUse()) {
1862     FastMathFlags FMF = FPT.getFastMathFlags();
1863     if (auto *FPMO = dyn_cast<FPMathOperator>(Op))
1864       FMF &= FPMO->getFastMathFlags();
1865 
1866     if (match(Op, m_FNeg(m_Value(X)))) {
1867       Value *InnerTrunc = Builder.CreateFPTruncFMF(X, Ty, FMF);
1868       Value *Neg = Builder.CreateFNegFMF(InnerTrunc, FMF);
1869       return replaceInstUsesWith(FPT, Neg);
1870     }
1871 
1872     // If we are truncating a select that has an extended operand, we can
1873     // narrow the other operand and do the select as a narrow op.
1874     Value *Cond, *X, *Y;
1875     if (match(Op, m_Select(m_Value(Cond), m_FPExt(m_Value(X)), m_Value(Y))) &&
1876         X->getType() == Ty) {
1877       // fptrunc (select Cond, (fpext X), Y --> select Cond, X, (fptrunc Y)
1878       Value *NarrowY = Builder.CreateFPTruncFMF(Y, Ty, FMF);
1879       Value *Sel =
1880           Builder.CreateSelectFMF(Cond, X, NarrowY, FMF, "narrow.sel", Op);
1881       return replaceInstUsesWith(FPT, Sel);
1882     }
1883     if (match(Op, m_Select(m_Value(Cond), m_Value(Y), m_FPExt(m_Value(X)))) &&
1884         X->getType() == Ty) {
1885       // fptrunc (select Cond, Y, (fpext X) --> select Cond, (fptrunc Y), X
1886       Value *NarrowY = Builder.CreateFPTruncFMF(Y, Ty, FMF);
1887       Value *Sel =
1888           Builder.CreateSelectFMF(Cond, NarrowY, X, FMF, "narrow.sel", Op);
1889       return replaceInstUsesWith(FPT, Sel);
1890     }
1891   }
1892 
1893   if (auto *II = dyn_cast<IntrinsicInst>(FPT.getOperand(0))) {
1894     switch (II->getIntrinsicID()) {
1895     default: break;
1896     case Intrinsic::ceil:
1897     case Intrinsic::fabs:
1898     case Intrinsic::floor:
1899     case Intrinsic::nearbyint:
1900     case Intrinsic::rint:
1901     case Intrinsic::round:
1902     case Intrinsic::roundeven:
1903     case Intrinsic::trunc: {
1904       Value *Src = II->getArgOperand(0);
1905       if (!Src->hasOneUse())
1906         break;
1907 
1908       // Except for fabs, this transformation requires the input of the unary FP
1909       // operation to be itself an fpext from the type to which we're
1910       // truncating.
1911       if (II->getIntrinsicID() != Intrinsic::fabs) {
1912         FPExtInst *FPExtSrc = dyn_cast<FPExtInst>(Src);
1913         if (!FPExtSrc || FPExtSrc->getSrcTy() != Ty)
1914           break;
1915       }
1916 
1917       // Do unary FP operation on smaller type.
1918       // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1919       Value *InnerTrunc = Builder.CreateFPTrunc(Src, Ty);
1920       Function *Overload = Intrinsic::getOrInsertDeclaration(
1921           FPT.getModule(), II->getIntrinsicID(), Ty);
1922       SmallVector<OperandBundleDef, 1> OpBundles;
1923       II->getOperandBundlesAsDefs(OpBundles);
1924       CallInst *NewCI =
1925           CallInst::Create(Overload, {InnerTrunc}, OpBundles, II->getName());
1926       // A normal value may be converted to an infinity. It means that we cannot
1927       // propagate ninf from the intrinsic. So we propagate FMF from fptrunc.
1928       NewCI->copyFastMathFlags(&FPT);
1929       return NewCI;
1930     }
1931     }
1932   }
1933 
1934   if (Instruction *I = shrinkInsertElt(FPT, Builder))
1935     return I;
1936 
1937   Value *Src = FPT.getOperand(0);
1938   if (isa<SIToFPInst>(Src) || isa<UIToFPInst>(Src)) {
1939     auto *FPCast = cast<CastInst>(Src);
1940     if (isKnownExactCastIntToFP(*FPCast, *this))
1941       return CastInst::Create(FPCast->getOpcode(), FPCast->getOperand(0), Ty);
1942   }
1943 
1944   return nullptr;
1945 }
1946 
visitFPExt(CastInst & FPExt)1947 Instruction *InstCombinerImpl::visitFPExt(CastInst &FPExt) {
1948   // If the source operand is a cast from integer to FP and known exact, then
1949   // cast the integer operand directly to the destination type.
1950   Type *Ty = FPExt.getType();
1951   Value *Src = FPExt.getOperand(0);
1952   if (isa<SIToFPInst>(Src) || isa<UIToFPInst>(Src)) {
1953     auto *FPCast = cast<CastInst>(Src);
1954     if (isKnownExactCastIntToFP(*FPCast, *this))
1955       return CastInst::Create(FPCast->getOpcode(), FPCast->getOperand(0), Ty);
1956   }
1957 
1958   return commonCastTransforms(FPExt);
1959 }
1960 
1961 /// fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)
1962 /// This is safe if the intermediate type has enough bits in its mantissa to
1963 /// accurately represent all values of X.  For example, this won't work with
1964 /// i64 -> float -> i64.
foldItoFPtoI(CastInst & FI)1965 Instruction *InstCombinerImpl::foldItoFPtoI(CastInst &FI) {
1966   if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))
1967     return nullptr;
1968 
1969   auto *OpI = cast<CastInst>(FI.getOperand(0));
1970   Value *X = OpI->getOperand(0);
1971   Type *XType = X->getType();
1972   Type *DestType = FI.getType();
1973   bool IsOutputSigned = isa<FPToSIInst>(FI);
1974 
1975   // Since we can assume the conversion won't overflow, our decision as to
1976   // whether the input will fit in the float should depend on the minimum
1977   // of the input range and output range.
1978 
1979   // This means this is also safe for a signed input and unsigned output, since
1980   // a negative input would lead to undefined behavior.
1981   if (!isKnownExactCastIntToFP(*OpI, *this)) {
1982     // The first cast may not round exactly based on the source integer width
1983     // and FP width, but the overflow UB rules can still allow this to fold.
1984     // If the destination type is narrow, that means the intermediate FP value
1985     // must be large enough to hold the source value exactly.
1986     // For example, (uint8_t)((float)(uint32_t 16777217) is undefined behavior.
1987     int OutputSize = (int)DestType->getScalarSizeInBits();
1988     if (OutputSize > OpI->getType()->getFPMantissaWidth())
1989       return nullptr;
1990   }
1991 
1992   if (DestType->getScalarSizeInBits() > XType->getScalarSizeInBits()) {
1993     bool IsInputSigned = isa<SIToFPInst>(OpI);
1994     if (IsInputSigned && IsOutputSigned)
1995       return new SExtInst(X, DestType);
1996     return new ZExtInst(X, DestType);
1997   }
1998   if (DestType->getScalarSizeInBits() < XType->getScalarSizeInBits())
1999     return new TruncInst(X, DestType);
2000 
2001   assert(XType == DestType && "Unexpected types for int to FP to int casts");
2002   return replaceInstUsesWith(FI, X);
2003 }
2004 
foldFPtoI(Instruction & FI,InstCombiner & IC)2005 static Instruction *foldFPtoI(Instruction &FI, InstCombiner &IC) {
2006   // fpto{u/s}i non-norm --> 0
2007   FPClassTest Mask =
2008       FI.getOpcode() == Instruction::FPToUI ? fcPosNormal : fcNormal;
2009   KnownFPClass FPClass = computeKnownFPClass(
2010       FI.getOperand(0), Mask, IC.getSimplifyQuery().getWithInstruction(&FI));
2011   if (FPClass.isKnownNever(Mask))
2012     return IC.replaceInstUsesWith(FI, ConstantInt::getNullValue(FI.getType()));
2013 
2014   return nullptr;
2015 }
2016 
visitFPToUI(FPToUIInst & FI)2017 Instruction *InstCombinerImpl::visitFPToUI(FPToUIInst &FI) {
2018   if (Instruction *I = foldItoFPtoI(FI))
2019     return I;
2020 
2021   if (Instruction *I = foldFPtoI(FI, *this))
2022     return I;
2023 
2024   return commonCastTransforms(FI);
2025 }
2026 
visitFPToSI(FPToSIInst & FI)2027 Instruction *InstCombinerImpl::visitFPToSI(FPToSIInst &FI) {
2028   if (Instruction *I = foldItoFPtoI(FI))
2029     return I;
2030 
2031   if (Instruction *I = foldFPtoI(FI, *this))
2032     return I;
2033 
2034   return commonCastTransforms(FI);
2035 }
2036 
visitUIToFP(CastInst & CI)2037 Instruction *InstCombinerImpl::visitUIToFP(CastInst &CI) {
2038   if (Instruction *R = commonCastTransforms(CI))
2039     return R;
2040   if (!CI.hasNonNeg() && isKnownNonNegative(CI.getOperand(0), SQ)) {
2041     CI.setNonNeg();
2042     return &CI;
2043   }
2044   return nullptr;
2045 }
2046 
visitSIToFP(CastInst & CI)2047 Instruction *InstCombinerImpl::visitSIToFP(CastInst &CI) {
2048   if (Instruction *R = commonCastTransforms(CI))
2049     return R;
2050   if (isKnownNonNegative(CI.getOperand(0), SQ)) {
2051     auto *UI =
2052         CastInst::Create(Instruction::UIToFP, CI.getOperand(0), CI.getType());
2053     UI->setNonNeg(true);
2054     return UI;
2055   }
2056   return nullptr;
2057 }
2058 
visitIntToPtr(IntToPtrInst & CI)2059 Instruction *InstCombinerImpl::visitIntToPtr(IntToPtrInst &CI) {
2060   // If the source integer type is not the intptr_t type for this target, do a
2061   // trunc or zext to the intptr_t type, then inttoptr of it.  This allows the
2062   // cast to be exposed to other transforms.
2063   unsigned AS = CI.getAddressSpace();
2064   if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
2065       DL.getPointerSizeInBits(AS)) {
2066     Type *Ty = CI.getOperand(0)->getType()->getWithNewType(
2067         DL.getIntPtrType(CI.getContext(), AS));
2068     Value *P = Builder.CreateZExtOrTrunc(CI.getOperand(0), Ty);
2069     return new IntToPtrInst(P, CI.getType());
2070   }
2071 
2072   if (Instruction *I = commonCastTransforms(CI))
2073     return I;
2074 
2075   return nullptr;
2076 }
2077 
foldPtrToIntOfGEP(Type * IntTy,Value * Ptr)2078 Value *InstCombinerImpl::foldPtrToIntOfGEP(Type *IntTy, Value *Ptr) {
2079   // Look through chain of one-use GEPs.
2080   Type *PtrTy = Ptr->getType();
2081   SmallVector<GEPOperator *> GEPs;
2082   while (true) {
2083     auto *GEP = dyn_cast<GEPOperator>(Ptr);
2084     if (!GEP || !GEP->hasOneUse())
2085       break;
2086     GEPs.push_back(GEP);
2087     Ptr = GEP->getPointerOperand();
2088   }
2089 
2090   // Don't handle case where GEP converts from pointer to vector.
2091   if (GEPs.empty() || PtrTy != Ptr->getType())
2092     return nullptr;
2093 
2094   // Check whether we know the integer value of the base pointer.
2095   Value *Res;
2096   Type *IdxTy = DL.getIndexType(PtrTy);
2097   if (match(Ptr, m_OneUse(m_IntToPtr(m_Value(Res)))) &&
2098       Res->getType() == IntTy && IntTy == IdxTy) {
2099     // pass
2100   } else if (isa<ConstantPointerNull>(Ptr)) {
2101     Res = Constant::getNullValue(IdxTy);
2102   } else {
2103     return nullptr;
2104   }
2105 
2106   // Perform the entire operation on integers instead.
2107   for (GEPOperator *GEP : reverse(GEPs)) {
2108     Value *Offset = EmitGEPOffset(GEP);
2109     Res = Builder.CreateAdd(Res, Offset, "", GEP->hasNoUnsignedWrap());
2110   }
2111   return Builder.CreateZExtOrTrunc(Res, IntTy);
2112 }
2113 
visitPtrToInt(PtrToIntInst & CI)2114 Instruction *InstCombinerImpl::visitPtrToInt(PtrToIntInst &CI) {
2115   // If the destination integer type is not the intptr_t type for this target,
2116   // do a ptrtoint to intptr_t then do a trunc or zext.  This allows the cast
2117   // to be exposed to other transforms.
2118   Value *SrcOp = CI.getPointerOperand();
2119   Type *SrcTy = SrcOp->getType();
2120   Type *Ty = CI.getType();
2121   unsigned AS = CI.getPointerAddressSpace();
2122   unsigned TySize = Ty->getScalarSizeInBits();
2123   unsigned PtrSize = DL.getPointerSizeInBits(AS);
2124   if (TySize != PtrSize) {
2125     Type *IntPtrTy =
2126         SrcTy->getWithNewType(DL.getIntPtrType(CI.getContext(), AS));
2127     Value *P = Builder.CreatePtrToInt(SrcOp, IntPtrTy);
2128     return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
2129   }
2130 
2131   // (ptrtoint (ptrmask P, M))
2132   //    -> (and (ptrtoint P), M)
2133   // This is generally beneficial as `and` is better supported than `ptrmask`.
2134   Value *Ptr, *Mask;
2135   if (match(SrcOp, m_OneUse(m_Intrinsic<Intrinsic::ptrmask>(m_Value(Ptr),
2136                                                             m_Value(Mask)))) &&
2137       Mask->getType() == Ty)
2138     return BinaryOperator::CreateAnd(Builder.CreatePtrToInt(Ptr, Ty), Mask);
2139 
2140   if (Value *V = foldPtrToIntOfGEP(Ty, SrcOp))
2141     return replaceInstUsesWith(CI, V);
2142 
2143   Value *Vec, *Scalar, *Index;
2144   if (match(SrcOp, m_OneUse(m_InsertElt(m_IntToPtr(m_Value(Vec)),
2145                                         m_Value(Scalar), m_Value(Index)))) &&
2146       Vec->getType() == Ty) {
2147     assert(Vec->getType()->getScalarSizeInBits() == PtrSize && "Wrong type");
2148     // Convert the scalar to int followed by insert to eliminate one cast:
2149     // p2i (ins (i2p Vec), Scalar, Index --> ins Vec, (p2i Scalar), Index
2150     Value *NewCast = Builder.CreatePtrToInt(Scalar, Ty->getScalarType());
2151     return InsertElementInst::Create(Vec, NewCast, Index);
2152   }
2153 
2154   return commonCastTransforms(CI);
2155 }
2156 
2157 /// This input value (which is known to have vector type) is being zero extended
2158 /// or truncated to the specified vector type. Since the zext/trunc is done
2159 /// using an integer type, we have a (bitcast(cast(bitcast))) pattern,
2160 /// endianness will impact which end of the vector that is extended or
2161 /// truncated.
2162 ///
2163 /// A vector is always stored with index 0 at the lowest address, which
2164 /// corresponds to the most significant bits for a big endian stored integer and
2165 /// the least significant bits for little endian. A trunc/zext of an integer
2166 /// impacts the big end of the integer. Thus, we need to add/remove elements at
2167 /// the front of the vector for big endian targets, and the back of the vector
2168 /// for little endian targets.
2169 ///
2170 /// Try to replace it with a shuffle (and vector/vector bitcast) if possible.
2171 ///
2172 /// The source and destination vector types may have different element types.
2173 static Instruction *
optimizeVectorResizeWithIntegerBitCasts(Value * InVal,VectorType * DestTy,InstCombinerImpl & IC)2174 optimizeVectorResizeWithIntegerBitCasts(Value *InVal, VectorType *DestTy,
2175                                         InstCombinerImpl &IC) {
2176   // We can only do this optimization if the output is a multiple of the input
2177   // element size, or the input is a multiple of the output element size.
2178   // Convert the input type to have the same element type as the output.
2179   VectorType *SrcTy = cast<VectorType>(InVal->getType());
2180 
2181   if (SrcTy->getElementType() != DestTy->getElementType()) {
2182     // The input types don't need to be identical, but for now they must be the
2183     // same size.  There is no specific reason we couldn't handle things like
2184     // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
2185     // there yet.
2186     if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
2187         DestTy->getElementType()->getPrimitiveSizeInBits())
2188       return nullptr;
2189 
2190     SrcTy =
2191         FixedVectorType::get(DestTy->getElementType(),
2192                              cast<FixedVectorType>(SrcTy)->getNumElements());
2193     InVal = IC.Builder.CreateBitCast(InVal, SrcTy);
2194   }
2195 
2196   bool IsBigEndian = IC.getDataLayout().isBigEndian();
2197   unsigned SrcElts = cast<FixedVectorType>(SrcTy)->getNumElements();
2198   unsigned DestElts = cast<FixedVectorType>(DestTy)->getNumElements();
2199 
2200   assert(SrcElts != DestElts && "Element counts should be different.");
2201 
2202   // Now that the element types match, get the shuffle mask and RHS of the
2203   // shuffle to use, which depends on whether we're increasing or decreasing the
2204   // size of the input.
2205   auto ShuffleMaskStorage = llvm::to_vector<16>(llvm::seq<int>(0, SrcElts));
2206   ArrayRef<int> ShuffleMask;
2207   Value *V2;
2208 
2209   if (SrcElts > DestElts) {
2210     // If we're shrinking the number of elements (rewriting an integer
2211     // truncate), just shuffle in the elements corresponding to the least
2212     // significant bits from the input and use poison as the second shuffle
2213     // input.
2214     V2 = PoisonValue::get(SrcTy);
2215     // Make sure the shuffle mask selects the "least significant bits" by
2216     // keeping elements from back of the src vector for big endian, and from the
2217     // front for little endian.
2218     ShuffleMask = ShuffleMaskStorage;
2219     if (IsBigEndian)
2220       ShuffleMask = ShuffleMask.take_back(DestElts);
2221     else
2222       ShuffleMask = ShuffleMask.take_front(DestElts);
2223   } else {
2224     // If we're increasing the number of elements (rewriting an integer zext),
2225     // shuffle in all of the elements from InVal. Fill the rest of the result
2226     // elements with zeros from a constant zero.
2227     V2 = Constant::getNullValue(SrcTy);
2228     // Use first elt from V2 when indicating zero in the shuffle mask.
2229     uint32_t NullElt = SrcElts;
2230     // Extend with null values in the "most significant bits" by adding elements
2231     // in front of the src vector for big endian, and at the back for little
2232     // endian.
2233     unsigned DeltaElts = DestElts - SrcElts;
2234     if (IsBigEndian)
2235       ShuffleMaskStorage.insert(ShuffleMaskStorage.begin(), DeltaElts, NullElt);
2236     else
2237       ShuffleMaskStorage.append(DeltaElts, NullElt);
2238     ShuffleMask = ShuffleMaskStorage;
2239   }
2240 
2241   return new ShuffleVectorInst(InVal, V2, ShuffleMask);
2242 }
2243 
isMultipleOfTypeSize(unsigned Value,Type * Ty)2244 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
2245   return Value % Ty->getPrimitiveSizeInBits() == 0;
2246 }
2247 
getTypeSizeIndex(unsigned Value,Type * Ty)2248 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
2249   return Value / Ty->getPrimitiveSizeInBits();
2250 }
2251 
2252 /// V is a value which is inserted into a vector of VecEltTy.
2253 /// Look through the value to see if we can decompose it into
2254 /// insertions into the vector.  See the example in the comment for
2255 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
2256 /// The type of V is always a non-zero multiple of VecEltTy's size.
2257 /// Shift is the number of bits between the lsb of V and the lsb of
2258 /// the vector.
2259 ///
2260 /// This returns false if the pattern can't be matched or true if it can,
2261 /// filling in Elements with the elements found here.
collectInsertionElements(Value * V,unsigned Shift,SmallVectorImpl<Value * > & Elements,Type * VecEltTy,bool isBigEndian)2262 static bool collectInsertionElements(Value *V, unsigned Shift,
2263                                      SmallVectorImpl<Value *> &Elements,
2264                                      Type *VecEltTy, bool isBigEndian) {
2265   assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
2266          "Shift should be a multiple of the element type size");
2267 
2268   // Undef values never contribute useful bits to the result.
2269   if (isa<UndefValue>(V)) return true;
2270 
2271   // If we got down to a value of the right type, we win, try inserting into the
2272   // right element.
2273   if (V->getType() == VecEltTy) {
2274     // Inserting null doesn't actually insert any elements.
2275     if (Constant *C = dyn_cast<Constant>(V))
2276       if (C->isNullValue())
2277         return true;
2278 
2279     unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
2280     if (isBigEndian)
2281       ElementIndex = Elements.size() - ElementIndex - 1;
2282 
2283     // Fail if multiple elements are inserted into this slot.
2284     if (Elements[ElementIndex])
2285       return false;
2286 
2287     Elements[ElementIndex] = V;
2288     return true;
2289   }
2290 
2291   if (Constant *C = dyn_cast<Constant>(V)) {
2292     // Figure out the # elements this provides, and bitcast it or slice it up
2293     // as required.
2294     unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
2295                                         VecEltTy);
2296     // If the constant is the size of a vector element, we just need to bitcast
2297     // it to the right type so it gets properly inserted.
2298     if (NumElts == 1)
2299       return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
2300                                       Shift, Elements, VecEltTy, isBigEndian);
2301 
2302     // Okay, this is a constant that covers multiple elements.  Slice it up into
2303     // pieces and insert each element-sized piece into the vector.
2304     if (!isa<IntegerType>(C->getType()))
2305       C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
2306                                        C->getType()->getPrimitiveSizeInBits()));
2307     unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
2308     Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
2309 
2310     for (unsigned i = 0; i != NumElts; ++i) {
2311       unsigned ShiftI = i * ElementSize;
2312       Constant *Piece = ConstantFoldBinaryInstruction(
2313           Instruction::LShr, C, ConstantInt::get(C->getType(), ShiftI));
2314       if (!Piece)
2315         return false;
2316 
2317       Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
2318       if (!collectInsertionElements(Piece, ShiftI + Shift, Elements, VecEltTy,
2319                                     isBigEndian))
2320         return false;
2321     }
2322     return true;
2323   }
2324 
2325   if (!V->hasOneUse()) return false;
2326 
2327   Instruction *I = dyn_cast<Instruction>(V);
2328   if (!I) return false;
2329   switch (I->getOpcode()) {
2330   default: return false; // Unhandled case.
2331   case Instruction::BitCast:
2332     if (I->getOperand(0)->getType()->isVectorTy())
2333       return false;
2334     return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2335                                     isBigEndian);
2336   case Instruction::ZExt:
2337     if (!isMultipleOfTypeSize(
2338                           I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
2339                               VecEltTy))
2340       return false;
2341     return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2342                                     isBigEndian);
2343   case Instruction::Or:
2344     return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2345                                     isBigEndian) &&
2346            collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,
2347                                     isBigEndian);
2348   case Instruction::Shl: {
2349     // Must be shifting by a constant that is a multiple of the element size.
2350     ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
2351     if (!CI) return false;
2352     Shift += CI->getZExtValue();
2353     if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
2354     return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2355                                     isBigEndian);
2356   }
2357 
2358   }
2359 }
2360 
2361 
2362 /// If the input is an 'or' instruction, we may be doing shifts and ors to
2363 /// assemble the elements of the vector manually.
2364 /// Try to rip the code out and replace it with insertelements.  This is to
2365 /// optimize code like this:
2366 ///
2367 ///    %tmp37 = bitcast float %inc to i32
2368 ///    %tmp38 = zext i32 %tmp37 to i64
2369 ///    %tmp31 = bitcast float %inc5 to i32
2370 ///    %tmp32 = zext i32 %tmp31 to i64
2371 ///    %tmp33 = shl i64 %tmp32, 32
2372 ///    %ins35 = or i64 %tmp33, %tmp38
2373 ///    %tmp43 = bitcast i64 %ins35 to <2 x float>
2374 ///
2375 /// Into two insertelements that do "buildvector{%inc, %inc5}".
optimizeIntegerToVectorInsertions(BitCastInst & CI,InstCombinerImpl & IC)2376 static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,
2377                                                 InstCombinerImpl &IC) {
2378   auto *DestVecTy = cast<FixedVectorType>(CI.getType());
2379   Value *IntInput = CI.getOperand(0);
2380 
2381   // if the int input is just an undef value do not try to optimize to vector
2382   // insertions as it will prevent undef propagation
2383   if (isa<UndefValue>(IntInput))
2384     return nullptr;
2385 
2386   SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
2387   if (!collectInsertionElements(IntInput, 0, Elements,
2388                                 DestVecTy->getElementType(),
2389                                 IC.getDataLayout().isBigEndian()))
2390     return nullptr;
2391 
2392   // If we succeeded, we know that all of the element are specified by Elements
2393   // or are zero if Elements has a null entry.  Recast this as a set of
2394   // insertions.
2395   Value *Result = Constant::getNullValue(CI.getType());
2396   for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
2397     if (!Elements[i]) continue;  // Unset element.
2398 
2399     Result = IC.Builder.CreateInsertElement(Result, Elements[i],
2400                                             IC.Builder.getInt32(i));
2401   }
2402 
2403   return Result;
2404 }
2405 
2406 /// Canonicalize scalar bitcasts of extracted elements into a bitcast of the
2407 /// vector followed by extract element. The backend tends to handle bitcasts of
2408 /// vectors better than bitcasts of scalars because vector registers are
2409 /// usually not type-specific like scalar integer or scalar floating-point.
canonicalizeBitCastExtElt(BitCastInst & BitCast,InstCombinerImpl & IC)2410 static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast,
2411                                               InstCombinerImpl &IC) {
2412   Value *VecOp, *Index;
2413   if (!match(BitCast.getOperand(0),
2414              m_OneUse(m_ExtractElt(m_Value(VecOp), m_Value(Index)))))
2415     return nullptr;
2416 
2417   // The bitcast must be to a vectorizable type, otherwise we can't make a new
2418   // type to extract from.
2419   Type *DestType = BitCast.getType();
2420   VectorType *VecType = cast<VectorType>(VecOp->getType());
2421   if (VectorType::isValidElementType(DestType)) {
2422     auto *NewVecType = VectorType::get(DestType, VecType);
2423     auto *NewBC = IC.Builder.CreateBitCast(VecOp, NewVecType, "bc");
2424     return ExtractElementInst::Create(NewBC, Index);
2425   }
2426 
2427   // Only solve DestType is vector to avoid inverse transform in visitBitCast.
2428   // bitcast (extractelement <1 x elt>, dest) -> bitcast(<1 x elt>, dest)
2429   auto *FixedVType = dyn_cast<FixedVectorType>(VecType);
2430   if (DestType->isVectorTy() && FixedVType && FixedVType->getNumElements() == 1)
2431     return CastInst::Create(Instruction::BitCast, VecOp, DestType);
2432 
2433   return nullptr;
2434 }
2435 
2436 /// Change the type of a bitwise logic operation if we can eliminate a bitcast.
foldBitCastBitwiseLogic(BitCastInst & BitCast,InstCombiner::BuilderTy & Builder)2437 static Instruction *foldBitCastBitwiseLogic(BitCastInst &BitCast,
2438                                             InstCombiner::BuilderTy &Builder) {
2439   Type *DestTy = BitCast.getType();
2440   BinaryOperator *BO;
2441 
2442   if (!match(BitCast.getOperand(0), m_OneUse(m_BinOp(BO))) ||
2443       !BO->isBitwiseLogicOp())
2444     return nullptr;
2445 
2446   // FIXME: This transform is restricted to vector types to avoid backend
2447   // problems caused by creating potentially illegal operations. If a fix-up is
2448   // added to handle that situation, we can remove this check.
2449   if (!DestTy->isVectorTy() || !BO->getType()->isVectorTy())
2450     return nullptr;
2451 
2452   if (DestTy->isFPOrFPVectorTy()) {
2453     Value *X, *Y;
2454     // bitcast(logic(bitcast(X), bitcast(Y))) -> bitcast'(logic(bitcast'(X), Y))
2455     if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
2456         match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(Y))))) {
2457       if (X->getType()->isFPOrFPVectorTy() &&
2458           Y->getType()->isIntOrIntVectorTy()) {
2459         Value *CastedOp =
2460             Builder.CreateBitCast(BO->getOperand(0), Y->getType());
2461         Value *NewBO = Builder.CreateBinOp(BO->getOpcode(), CastedOp, Y);
2462         return CastInst::CreateBitOrPointerCast(NewBO, DestTy);
2463       }
2464       if (X->getType()->isIntOrIntVectorTy() &&
2465           Y->getType()->isFPOrFPVectorTy()) {
2466         Value *CastedOp =
2467             Builder.CreateBitCast(BO->getOperand(1), X->getType());
2468         Value *NewBO = Builder.CreateBinOp(BO->getOpcode(), CastedOp, X);
2469         return CastInst::CreateBitOrPointerCast(NewBO, DestTy);
2470       }
2471     }
2472     return nullptr;
2473   }
2474 
2475   if (!DestTy->isIntOrIntVectorTy())
2476     return nullptr;
2477 
2478   Value *X;
2479   if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
2480       X->getType() == DestTy && !isa<Constant>(X)) {
2481     // bitcast(logic(bitcast(X), Y)) --> logic'(X, bitcast(Y))
2482     Value *CastedOp1 = Builder.CreateBitCast(BO->getOperand(1), DestTy);
2483     return BinaryOperator::Create(BO->getOpcode(), X, CastedOp1);
2484   }
2485 
2486   if (match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(X)))) &&
2487       X->getType() == DestTy && !isa<Constant>(X)) {
2488     // bitcast(logic(Y, bitcast(X))) --> logic'(bitcast(Y), X)
2489     Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
2490     return BinaryOperator::Create(BO->getOpcode(), CastedOp0, X);
2491   }
2492 
2493   // Canonicalize vector bitcasts to come before vector bitwise logic with a
2494   // constant. This eases recognition of special constants for later ops.
2495   // Example:
2496   // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
2497   Constant *C;
2498   if (match(BO->getOperand(1), m_Constant(C))) {
2499     // bitcast (logic X, C) --> logic (bitcast X, C')
2500     Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
2501     Value *CastedC = Builder.CreateBitCast(C, DestTy);
2502     return BinaryOperator::Create(BO->getOpcode(), CastedOp0, CastedC);
2503   }
2504 
2505   return nullptr;
2506 }
2507 
2508 /// Change the type of a select if we can eliminate a bitcast.
foldBitCastSelect(BitCastInst & BitCast,InstCombiner::BuilderTy & Builder)2509 static Instruction *foldBitCastSelect(BitCastInst &BitCast,
2510                                       InstCombiner::BuilderTy &Builder) {
2511   Value *Cond, *TVal, *FVal;
2512   if (!match(BitCast.getOperand(0),
2513              m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
2514     return nullptr;
2515 
2516   // A vector select must maintain the same number of elements in its operands.
2517   Type *CondTy = Cond->getType();
2518   Type *DestTy = BitCast.getType();
2519   if (auto *CondVTy = dyn_cast<VectorType>(CondTy))
2520     if (!DestTy->isVectorTy() ||
2521         CondVTy->getElementCount() !=
2522             cast<VectorType>(DestTy)->getElementCount())
2523       return nullptr;
2524 
2525   // FIXME: This transform is restricted from changing the select between
2526   // scalars and vectors to avoid backend problems caused by creating
2527   // potentially illegal operations. If a fix-up is added to handle that
2528   // situation, we can remove this check.
2529   if (DestTy->isVectorTy() != TVal->getType()->isVectorTy())
2530     return nullptr;
2531 
2532   auto *Sel = cast<Instruction>(BitCast.getOperand(0));
2533   Value *X;
2534   if (match(TVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
2535       !isa<Constant>(X)) {
2536     // bitcast(select(Cond, bitcast(X), Y)) --> select'(Cond, X, bitcast(Y))
2537     Value *CastedVal = Builder.CreateBitCast(FVal, DestTy);
2538     return SelectInst::Create(Cond, X, CastedVal, "", nullptr, Sel);
2539   }
2540 
2541   if (match(FVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
2542       !isa<Constant>(X)) {
2543     // bitcast(select(Cond, Y, bitcast(X))) --> select'(Cond, bitcast(Y), X)
2544     Value *CastedVal = Builder.CreateBitCast(TVal, DestTy);
2545     return SelectInst::Create(Cond, CastedVal, X, "", nullptr, Sel);
2546   }
2547 
2548   return nullptr;
2549 }
2550 
2551 /// Check if all users of CI are StoreInsts.
hasStoreUsersOnly(CastInst & CI)2552 static bool hasStoreUsersOnly(CastInst &CI) {
2553   for (User *U : CI.users()) {
2554     if (!isa<StoreInst>(U))
2555       return false;
2556   }
2557   return true;
2558 }
2559 
2560 /// This function handles following case
2561 ///
2562 ///     A  ->  B    cast
2563 ///     PHI
2564 ///     B  ->  A    cast
2565 ///
2566 /// All the related PHI nodes can be replaced by new PHI nodes with type A.
2567 /// The uses of \p CI can be changed to the new PHI node corresponding to \p PN.
optimizeBitCastFromPhi(CastInst & CI,PHINode * PN)2568 Instruction *InstCombinerImpl::optimizeBitCastFromPhi(CastInst &CI,
2569                                                       PHINode *PN) {
2570   // BitCast used by Store can be handled in InstCombineLoadStoreAlloca.cpp.
2571   if (hasStoreUsersOnly(CI))
2572     return nullptr;
2573 
2574   Value *Src = CI.getOperand(0);
2575   Type *SrcTy = Src->getType();         // Type B
2576   Type *DestTy = CI.getType();          // Type A
2577 
2578   SmallVector<PHINode *, 4> PhiWorklist;
2579   SmallSetVector<PHINode *, 4> OldPhiNodes;
2580 
2581   // Find all of the A->B casts and PHI nodes.
2582   // We need to inspect all related PHI nodes, but PHIs can be cyclic, so
2583   // OldPhiNodes is used to track all known PHI nodes, before adding a new
2584   // PHI to PhiWorklist, it is checked against and added to OldPhiNodes first.
2585   PhiWorklist.push_back(PN);
2586   OldPhiNodes.insert(PN);
2587   while (!PhiWorklist.empty()) {
2588     auto *OldPN = PhiWorklist.pop_back_val();
2589     for (Value *IncValue : OldPN->incoming_values()) {
2590       if (isa<Constant>(IncValue))
2591         continue;
2592 
2593       if (auto *LI = dyn_cast<LoadInst>(IncValue)) {
2594         // If there is a sequence of one or more load instructions, each loaded
2595         // value is used as address of later load instruction, bitcast is
2596         // necessary to change the value type, don't optimize it. For
2597         // simplicity we give up if the load address comes from another load.
2598         Value *Addr = LI->getOperand(0);
2599         if (Addr == &CI || isa<LoadInst>(Addr))
2600           return nullptr;
2601         // Don't tranform "load <256 x i32>, <256 x i32>*" to
2602         // "load x86_amx, x86_amx*", because x86_amx* is invalid.
2603         // TODO: Remove this check when bitcast between vector and x86_amx
2604         // is replaced with a specific intrinsic.
2605         if (DestTy->isX86_AMXTy())
2606           return nullptr;
2607         if (LI->hasOneUse() && LI->isSimple())
2608           continue;
2609         // If a LoadInst has more than one use, changing the type of loaded
2610         // value may create another bitcast.
2611         return nullptr;
2612       }
2613 
2614       if (auto *PNode = dyn_cast<PHINode>(IncValue)) {
2615         if (OldPhiNodes.insert(PNode))
2616           PhiWorklist.push_back(PNode);
2617         continue;
2618       }
2619 
2620       auto *BCI = dyn_cast<BitCastInst>(IncValue);
2621       // We can't handle other instructions.
2622       if (!BCI)
2623         return nullptr;
2624 
2625       // Verify it's a A->B cast.
2626       Type *TyA = BCI->getOperand(0)->getType();
2627       Type *TyB = BCI->getType();
2628       if (TyA != DestTy || TyB != SrcTy)
2629         return nullptr;
2630     }
2631   }
2632 
2633   // Check that each user of each old PHI node is something that we can
2634   // rewrite, so that all of the old PHI nodes can be cleaned up afterwards.
2635   for (auto *OldPN : OldPhiNodes) {
2636     for (User *V : OldPN->users()) {
2637       if (auto *SI = dyn_cast<StoreInst>(V)) {
2638         if (!SI->isSimple() || SI->getOperand(0) != OldPN)
2639           return nullptr;
2640       } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2641         // Verify it's a B->A cast.
2642         Type *TyB = BCI->getOperand(0)->getType();
2643         Type *TyA = BCI->getType();
2644         if (TyA != DestTy || TyB != SrcTy)
2645           return nullptr;
2646       } else if (auto *PHI = dyn_cast<PHINode>(V)) {
2647         // As long as the user is another old PHI node, then even if we don't
2648         // rewrite it, the PHI web we're considering won't have any users
2649         // outside itself, so it'll be dead.
2650         if (!OldPhiNodes.contains(PHI))
2651           return nullptr;
2652       } else {
2653         return nullptr;
2654       }
2655     }
2656   }
2657 
2658   // For each old PHI node, create a corresponding new PHI node with a type A.
2659   SmallDenseMap<PHINode *, PHINode *> NewPNodes;
2660   for (auto *OldPN : OldPhiNodes) {
2661     Builder.SetInsertPoint(OldPN);
2662     PHINode *NewPN = Builder.CreatePHI(DestTy, OldPN->getNumOperands());
2663     NewPNodes[OldPN] = NewPN;
2664   }
2665 
2666   // Fill in the operands of new PHI nodes.
2667   for (auto *OldPN : OldPhiNodes) {
2668     PHINode *NewPN = NewPNodes[OldPN];
2669     for (unsigned j = 0, e = OldPN->getNumOperands(); j != e; ++j) {
2670       Value *V = OldPN->getOperand(j);
2671       Value *NewV = nullptr;
2672       if (auto *C = dyn_cast<Constant>(V)) {
2673         NewV = ConstantExpr::getBitCast(C, DestTy);
2674       } else if (auto *LI = dyn_cast<LoadInst>(V)) {
2675         // Explicitly perform load combine to make sure no opposing transform
2676         // can remove the bitcast in the meantime and trigger an infinite loop.
2677         Builder.SetInsertPoint(LI);
2678         NewV = combineLoadToNewType(*LI, DestTy);
2679         // Remove the old load and its use in the old phi, which itself becomes
2680         // dead once the whole transform finishes.
2681         replaceInstUsesWith(*LI, PoisonValue::get(LI->getType()));
2682         eraseInstFromFunction(*LI);
2683       } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2684         NewV = BCI->getOperand(0);
2685       } else if (auto *PrevPN = dyn_cast<PHINode>(V)) {
2686         NewV = NewPNodes[PrevPN];
2687       }
2688       assert(NewV);
2689       NewPN->addIncoming(NewV, OldPN->getIncomingBlock(j));
2690     }
2691   }
2692 
2693   // Traverse all accumulated PHI nodes and process its users,
2694   // which are Stores and BitcCasts. Without this processing
2695   // NewPHI nodes could be replicated and could lead to extra
2696   // moves generated after DeSSA.
2697   // If there is a store with type B, change it to type A.
2698 
2699 
2700   // Replace users of BitCast B->A with NewPHI. These will help
2701   // later to get rid off a closure formed by OldPHI nodes.
2702   Instruction *RetVal = nullptr;
2703   for (auto *OldPN : OldPhiNodes) {
2704     PHINode *NewPN = NewPNodes[OldPN];
2705     for (User *V : make_early_inc_range(OldPN->users())) {
2706       if (auto *SI = dyn_cast<StoreInst>(V)) {
2707         assert(SI->isSimple() && SI->getOperand(0) == OldPN);
2708         Builder.SetInsertPoint(SI);
2709         auto *NewBC =
2710           cast<BitCastInst>(Builder.CreateBitCast(NewPN, SrcTy));
2711         SI->setOperand(0, NewBC);
2712         Worklist.push(SI);
2713         assert(hasStoreUsersOnly(*NewBC));
2714       }
2715       else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2716         Type *TyB = BCI->getOperand(0)->getType();
2717         Type *TyA = BCI->getType();
2718         assert(TyA == DestTy && TyB == SrcTy);
2719         (void) TyA;
2720         (void) TyB;
2721         Instruction *I = replaceInstUsesWith(*BCI, NewPN);
2722         if (BCI == &CI)
2723           RetVal = I;
2724       } else if (auto *PHI = dyn_cast<PHINode>(V)) {
2725         assert(OldPhiNodes.contains(PHI));
2726         (void) PHI;
2727       } else {
2728         llvm_unreachable("all uses should be handled");
2729       }
2730     }
2731   }
2732 
2733   return RetVal;
2734 }
2735 
2736 /// Fold (bitcast (or (and (bitcast X to int), signmask), nneg Y) to fp) to
2737 /// copysign((bitcast Y to fp), X)
foldCopySignIdioms(BitCastInst & CI,InstCombiner::BuilderTy & Builder,const SimplifyQuery & SQ)2738 static Value *foldCopySignIdioms(BitCastInst &CI,
2739                                  InstCombiner::BuilderTy &Builder,
2740                                  const SimplifyQuery &SQ) {
2741   Value *X, *Y;
2742   Type *FTy = CI.getType();
2743   if (!FTy->isFPOrFPVectorTy())
2744     return nullptr;
2745   if (!match(&CI, m_ElementWiseBitCast(m_c_Or(
2746                       m_And(m_ElementWiseBitCast(m_Value(X)), m_SignMask()),
2747                       m_Value(Y)))))
2748     return nullptr;
2749   if (X->getType() != FTy)
2750     return nullptr;
2751   if (!isKnownNonNegative(Y, SQ))
2752     return nullptr;
2753 
2754   return Builder.CreateCopySign(Builder.CreateBitCast(Y, FTy), X);
2755 }
2756 
visitBitCast(BitCastInst & CI)2757 Instruction *InstCombinerImpl::visitBitCast(BitCastInst &CI) {
2758   // If the operands are integer typed then apply the integer transforms,
2759   // otherwise just apply the common ones.
2760   Value *Src = CI.getOperand(0);
2761   Type *SrcTy = Src->getType();
2762   Type *DestTy = CI.getType();
2763 
2764   // Get rid of casts from one type to the same type. These are useless and can
2765   // be replaced by the operand.
2766   if (DestTy == Src->getType())
2767     return replaceInstUsesWith(CI, Src);
2768 
2769   if (isa<FixedVectorType>(DestTy)) {
2770     if (isa<IntegerType>(SrcTy)) {
2771       // If this is a cast from an integer to vector, check to see if the input
2772       // is a trunc or zext of a bitcast from vector.  If so, we can replace all
2773       // the casts with a shuffle and (potentially) a bitcast.
2774       if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
2775         CastInst *SrcCast = cast<CastInst>(Src);
2776         if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
2777           if (isa<VectorType>(BCIn->getOperand(0)->getType()))
2778             if (Instruction *I = optimizeVectorResizeWithIntegerBitCasts(
2779                     BCIn->getOperand(0), cast<VectorType>(DestTy), *this))
2780               return I;
2781       }
2782 
2783       // If the input is an 'or' instruction, we may be doing shifts and ors to
2784       // assemble the elements of the vector manually.  Try to rip the code out
2785       // and replace it with insertelements.
2786       if (Value *V = optimizeIntegerToVectorInsertions(CI, *this))
2787         return replaceInstUsesWith(CI, V);
2788     }
2789   }
2790 
2791   if (FixedVectorType *SrcVTy = dyn_cast<FixedVectorType>(SrcTy)) {
2792     if (SrcVTy->getNumElements() == 1) {
2793       // If our destination is not a vector, then make this a straight
2794       // scalar-scalar cast.
2795       if (!DestTy->isVectorTy()) {
2796         Value *Elem =
2797           Builder.CreateExtractElement(Src,
2798                      Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2799         return CastInst::Create(Instruction::BitCast, Elem, DestTy);
2800       }
2801 
2802       // Otherwise, see if our source is an insert. If so, then use the scalar
2803       // component directly:
2804       // bitcast (inselt <1 x elt> V, X, 0) to <n x m> --> bitcast X to <n x m>
2805       if (auto *InsElt = dyn_cast<InsertElementInst>(Src))
2806         return new BitCastInst(InsElt->getOperand(1), DestTy);
2807     }
2808 
2809     // Convert an artificial vector insert into more analyzable bitwise logic.
2810     unsigned BitWidth = DestTy->getScalarSizeInBits();
2811     Value *X, *Y;
2812     uint64_t IndexC;
2813     if (match(Src, m_OneUse(m_InsertElt(m_OneUse(m_BitCast(m_Value(X))),
2814                                         m_Value(Y), m_ConstantInt(IndexC)))) &&
2815         DestTy->isIntegerTy() && X->getType() == DestTy &&
2816         Y->getType()->isIntegerTy() && isDesirableIntType(BitWidth)) {
2817       // Adjust for big endian - the LSBs are at the high index.
2818       if (DL.isBigEndian())
2819         IndexC = SrcVTy->getNumElements() - 1 - IndexC;
2820 
2821       // We only handle (endian-normalized) insert to index 0. Any other insert
2822       // would require a left-shift, so that is an extra instruction.
2823       if (IndexC == 0) {
2824         // bitcast (inselt (bitcast X), Y, 0) --> or (and X, MaskC), (zext Y)
2825         unsigned EltWidth = Y->getType()->getScalarSizeInBits();
2826         APInt MaskC = APInt::getHighBitsSet(BitWidth, BitWidth - EltWidth);
2827         Value *AndX = Builder.CreateAnd(X, MaskC);
2828         Value *ZextY = Builder.CreateZExt(Y, DestTy);
2829         return BinaryOperator::CreateOr(AndX, ZextY);
2830       }
2831     }
2832   }
2833 
2834   if (auto *Shuf = dyn_cast<ShuffleVectorInst>(Src)) {
2835     // Okay, we have (bitcast (shuffle ..)).  Check to see if this is
2836     // a bitcast to a vector with the same # elts.
2837     Value *ShufOp0 = Shuf->getOperand(0);
2838     Value *ShufOp1 = Shuf->getOperand(1);
2839     auto ShufElts = cast<VectorType>(Shuf->getType())->getElementCount();
2840     auto SrcVecElts = cast<VectorType>(ShufOp0->getType())->getElementCount();
2841     if (Shuf->hasOneUse() && DestTy->isVectorTy() &&
2842         cast<VectorType>(DestTy)->getElementCount() == ShufElts &&
2843         ShufElts == SrcVecElts) {
2844       BitCastInst *Tmp;
2845       // If either of the operands is a cast from CI.getType(), then
2846       // evaluating the shuffle in the casted destination's type will allow
2847       // us to eliminate at least one cast.
2848       if (((Tmp = dyn_cast<BitCastInst>(ShufOp0)) &&
2849            Tmp->getOperand(0)->getType() == DestTy) ||
2850           ((Tmp = dyn_cast<BitCastInst>(ShufOp1)) &&
2851            Tmp->getOperand(0)->getType() == DestTy)) {
2852         Value *LHS = Builder.CreateBitCast(ShufOp0, DestTy);
2853         Value *RHS = Builder.CreateBitCast(ShufOp1, DestTy);
2854         // Return a new shuffle vector.  Use the same element ID's, as we
2855         // know the vector types match #elts.
2856         return new ShuffleVectorInst(LHS, RHS, Shuf->getShuffleMask());
2857       }
2858     }
2859 
2860     // A bitcasted-to-scalar and byte/bit reversing shuffle is better recognized
2861     // as a byte/bit swap:
2862     // bitcast <N x i8> (shuf X, undef, <N, N-1,...0>) -> bswap (bitcast X)
2863     // bitcast <N x i1> (shuf X, undef, <N, N-1,...0>) -> bitreverse (bitcast X)
2864     if (DestTy->isIntegerTy() && ShufElts.getKnownMinValue() % 2 == 0 &&
2865         Shuf->hasOneUse() && Shuf->isReverse()) {
2866       unsigned IntrinsicNum = 0;
2867       if (DL.isLegalInteger(DestTy->getScalarSizeInBits()) &&
2868           SrcTy->getScalarSizeInBits() == 8) {
2869         IntrinsicNum = Intrinsic::bswap;
2870       } else if (SrcTy->getScalarSizeInBits() == 1) {
2871         IntrinsicNum = Intrinsic::bitreverse;
2872       }
2873       if (IntrinsicNum != 0) {
2874         assert(ShufOp0->getType() == SrcTy && "Unexpected shuffle mask");
2875         assert(match(ShufOp1, m_Undef()) && "Unexpected shuffle op");
2876         Function *BswapOrBitreverse = Intrinsic::getOrInsertDeclaration(
2877             CI.getModule(), IntrinsicNum, DestTy);
2878         Value *ScalarX = Builder.CreateBitCast(ShufOp0, DestTy);
2879         return CallInst::Create(BswapOrBitreverse, {ScalarX});
2880       }
2881     }
2882   }
2883 
2884   // Handle the A->B->A cast, and there is an intervening PHI node.
2885   if (PHINode *PN = dyn_cast<PHINode>(Src))
2886     if (Instruction *I = optimizeBitCastFromPhi(CI, PN))
2887       return I;
2888 
2889   if (Instruction *I = canonicalizeBitCastExtElt(CI, *this))
2890     return I;
2891 
2892   if (Instruction *I = foldBitCastBitwiseLogic(CI, Builder))
2893     return I;
2894 
2895   if (Instruction *I = foldBitCastSelect(CI, Builder))
2896     return I;
2897 
2898   if (Value *V = foldCopySignIdioms(CI, Builder, SQ.getWithInstruction(&CI)))
2899     return replaceInstUsesWith(CI, V);
2900 
2901   return commonCastTransforms(CI);
2902 }
2903 
visitAddrSpaceCast(AddrSpaceCastInst & CI)2904 Instruction *InstCombinerImpl::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
2905   return commonCastTransforms(CI);
2906 }
2907