xref: /netbsd-src/external/apache2/llvm/dist/llvm/lib/IR/ConstantFold.cpp (revision 82d56013d7b633d116a93943de88e08335357a7c)
1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
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 folding of constants for LLVM.  This implements the
10 // (internal) ConstantFold.h interface, which is used by the
11 // ConstantExpr::get* methods to automatically fold constants when possible.
12 //
13 // The current constant folding implementation is implemented in two pieces: the
14 // pieces that don't need DataLayout, and the pieces that do. This is to avoid
15 // a dependence in IR on Target.
16 //
17 //===----------------------------------------------------------------------===//
18 
19 #include "ConstantFold.h"
20 #include "llvm/ADT/APSInt.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DerivedTypes.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/Operator.h"
31 #include "llvm/IR/PatternMatch.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/ManagedStatic.h"
34 #include "llvm/Support/MathExtras.h"
35 using namespace llvm;
36 using namespace llvm::PatternMatch;
37 
38 //===----------------------------------------------------------------------===//
39 //                ConstantFold*Instruction Implementations
40 //===----------------------------------------------------------------------===//
41 
42 /// Convert the specified vector Constant node to the specified vector type.
43 /// At this point, we know that the elements of the input vector constant are
44 /// all simple integer or FP values.
BitCastConstantVector(Constant * CV,VectorType * DstTy)45 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) {
46 
47   if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
48   if (CV->isNullValue()) return Constant::getNullValue(DstTy);
49 
50   // Do not iterate on scalable vector. The num of elements is unknown at
51   // compile-time.
52   if (isa<ScalableVectorType>(DstTy))
53     return nullptr;
54 
55   // If this cast changes element count then we can't handle it here:
56   // doing so requires endianness information.  This should be handled by
57   // Analysis/ConstantFolding.cpp
58   unsigned NumElts = cast<FixedVectorType>(DstTy)->getNumElements();
59   if (NumElts != cast<FixedVectorType>(CV->getType())->getNumElements())
60     return nullptr;
61 
62   Type *DstEltTy = DstTy->getElementType();
63   // Fast path for splatted constants.
64   if (Constant *Splat = CV->getSplatValue()) {
65     return ConstantVector::getSplat(DstTy->getElementCount(),
66                                     ConstantExpr::getBitCast(Splat, DstEltTy));
67   }
68 
69   SmallVector<Constant*, 16> Result;
70   Type *Ty = IntegerType::get(CV->getContext(), 32);
71   for (unsigned i = 0; i != NumElts; ++i) {
72     Constant *C =
73       ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i));
74     C = ConstantExpr::getBitCast(C, DstEltTy);
75     Result.push_back(C);
76   }
77 
78   return ConstantVector::get(Result);
79 }
80 
81 /// This function determines which opcode to use to fold two constant cast
82 /// expressions together. It uses CastInst::isEliminableCastPair to determine
83 /// the opcode. Consequently its just a wrapper around that function.
84 /// Determine if it is valid to fold a cast of a cast
85 static unsigned
foldConstantCastPair(unsigned opc,ConstantExpr * Op,Type * DstTy)86 foldConstantCastPair(
87   unsigned opc,          ///< opcode of the second cast constant expression
88   ConstantExpr *Op,      ///< the first cast constant expression
89   Type *DstTy            ///< destination type of the first cast
90 ) {
91   assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
92   assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
93   assert(CastInst::isCast(opc) && "Invalid cast opcode");
94 
95   // The types and opcodes for the two Cast constant expressions
96   Type *SrcTy = Op->getOperand(0)->getType();
97   Type *MidTy = Op->getType();
98   Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
99   Instruction::CastOps secondOp = Instruction::CastOps(opc);
100 
101   // Assume that pointers are never more than 64 bits wide, and only use this
102   // for the middle type. Otherwise we could end up folding away illegal
103   // bitcasts between address spaces with different sizes.
104   IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext());
105 
106   // Let CastInst::isEliminableCastPair do the heavy lifting.
107   return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
108                                         nullptr, FakeIntPtrTy, nullptr);
109 }
110 
FoldBitCast(Constant * V,Type * DestTy)111 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
112   Type *SrcTy = V->getType();
113   if (SrcTy == DestTy)
114     return V; // no-op cast
115 
116   // Check to see if we are casting a pointer to an aggregate to a pointer to
117   // the first element.  If so, return the appropriate GEP instruction.
118   if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
119     if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
120       if (PTy->getAddressSpace() == DPTy->getAddressSpace()
121           && PTy->getElementType()->isSized()) {
122         SmallVector<Value*, 8> IdxList;
123         Value *Zero =
124           Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
125         IdxList.push_back(Zero);
126         Type *ElTy = PTy->getElementType();
127         while (ElTy && ElTy != DPTy->getElementType()) {
128           ElTy = GetElementPtrInst::getTypeAtIndex(ElTy, (uint64_t)0);
129           IdxList.push_back(Zero);
130         }
131 
132         if (ElTy == DPTy->getElementType())
133           // This GEP is inbounds because all indices are zero.
134           return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(),
135                                                         V, IdxList);
136       }
137 
138   // Handle casts from one vector constant to another.  We know that the src
139   // and dest type have the same size (otherwise its an illegal cast).
140   if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
141     if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
142       assert(DestPTy->getPrimitiveSizeInBits() ==
143                  SrcTy->getPrimitiveSizeInBits() &&
144              "Not cast between same sized vectors!");
145       SrcTy = nullptr;
146       // First, check for null.  Undef is already handled.
147       if (isa<ConstantAggregateZero>(V))
148         return Constant::getNullValue(DestTy);
149 
150       // Handle ConstantVector and ConstantAggregateVector.
151       return BitCastConstantVector(V, DestPTy);
152     }
153 
154     // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
155     // This allows for other simplifications (although some of them
156     // can only be handled by Analysis/ConstantFolding.cpp).
157     if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
158       return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
159   }
160 
161   // Finally, implement bitcast folding now.   The code below doesn't handle
162   // bitcast right.
163   if (isa<ConstantPointerNull>(V))  // ptr->ptr cast.
164     return ConstantPointerNull::get(cast<PointerType>(DestTy));
165 
166   // Handle integral constant input.
167   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
168     if (DestTy->isIntegerTy())
169       // Integral -> Integral. This is a no-op because the bit widths must
170       // be the same. Consequently, we just fold to V.
171       return V;
172 
173     // See note below regarding the PPC_FP128 restriction.
174     if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty())
175       return ConstantFP::get(DestTy->getContext(),
176                              APFloat(DestTy->getFltSemantics(),
177                                      CI->getValue()));
178 
179     // Otherwise, can't fold this (vector?)
180     return nullptr;
181   }
182 
183   // Handle ConstantFP input: FP -> Integral.
184   if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) {
185     // PPC_FP128 is really the sum of two consecutive doubles, where the first
186     // double is always stored first in memory, regardless of the target
187     // endianness. The memory layout of i128, however, depends on the target
188     // endianness, and so we can't fold this without target endianness
189     // information. This should instead be handled by
190     // Analysis/ConstantFolding.cpp
191     if (FP->getType()->isPPC_FP128Ty())
192       return nullptr;
193 
194     // Make sure dest type is compatible with the folded integer constant.
195     if (!DestTy->isIntegerTy())
196       return nullptr;
197 
198     return ConstantInt::get(FP->getContext(),
199                             FP->getValueAPF().bitcastToAPInt());
200   }
201 
202   return nullptr;
203 }
204 
205 
206 /// V is an integer constant which only has a subset of its bytes used.
207 /// The bytes used are indicated by ByteStart (which is the first byte used,
208 /// counting from the least significant byte) and ByteSize, which is the number
209 /// of bytes used.
210 ///
211 /// This function analyzes the specified constant to see if the specified byte
212 /// range can be returned as a simplified constant.  If so, the constant is
213 /// returned, otherwise null is returned.
ExtractConstantBytes(Constant * C,unsigned ByteStart,unsigned ByteSize)214 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
215                                       unsigned ByteSize) {
216   assert(C->getType()->isIntegerTy() &&
217          (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
218          "Non-byte sized integer input");
219   unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
220   assert(ByteSize && "Must be accessing some piece");
221   assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
222   assert(ByteSize != CSize && "Should not extract everything");
223 
224   // Constant Integers are simple.
225   if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
226     APInt V = CI->getValue();
227     if (ByteStart)
228       V.lshrInPlace(ByteStart*8);
229     V = V.trunc(ByteSize*8);
230     return ConstantInt::get(CI->getContext(), V);
231   }
232 
233   // In the input is a constant expr, we might be able to recursively simplify.
234   // If not, we definitely can't do anything.
235   ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
236   if (!CE) return nullptr;
237 
238   switch (CE->getOpcode()) {
239   default: return nullptr;
240   case Instruction::Or: {
241     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
242     if (!RHS)
243       return nullptr;
244 
245     // X | -1 -> -1.
246     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
247       if (RHSC->isMinusOne())
248         return RHSC;
249 
250     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
251     if (!LHS)
252       return nullptr;
253     return ConstantExpr::getOr(LHS, RHS);
254   }
255   case Instruction::And: {
256     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
257     if (!RHS)
258       return nullptr;
259 
260     // X & 0 -> 0.
261     if (RHS->isNullValue())
262       return RHS;
263 
264     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
265     if (!LHS)
266       return nullptr;
267     return ConstantExpr::getAnd(LHS, RHS);
268   }
269   case Instruction::LShr: {
270     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
271     if (!Amt)
272       return nullptr;
273     APInt ShAmt = Amt->getValue();
274     // Cannot analyze non-byte shifts.
275     if ((ShAmt & 7) != 0)
276       return nullptr;
277     ShAmt.lshrInPlace(3);
278 
279     // If the extract is known to be all zeros, return zero.
280     if (ShAmt.uge(CSize - ByteStart))
281       return Constant::getNullValue(
282           IntegerType::get(CE->getContext(), ByteSize * 8));
283     // If the extract is known to be fully in the input, extract it.
284     if (ShAmt.ule(CSize - (ByteStart + ByteSize)))
285       return ExtractConstantBytes(CE->getOperand(0),
286                                   ByteStart + ShAmt.getZExtValue(), ByteSize);
287 
288     // TODO: Handle the 'partially zero' case.
289     return nullptr;
290   }
291 
292   case Instruction::Shl: {
293     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
294     if (!Amt)
295       return nullptr;
296     APInt ShAmt = Amt->getValue();
297     // Cannot analyze non-byte shifts.
298     if ((ShAmt & 7) != 0)
299       return nullptr;
300     ShAmt.lshrInPlace(3);
301 
302     // If the extract is known to be all zeros, return zero.
303     if (ShAmt.uge(ByteStart + ByteSize))
304       return Constant::getNullValue(
305           IntegerType::get(CE->getContext(), ByteSize * 8));
306     // If the extract is known to be fully in the input, extract it.
307     if (ShAmt.ule(ByteStart))
308       return ExtractConstantBytes(CE->getOperand(0),
309                                   ByteStart - ShAmt.getZExtValue(), ByteSize);
310 
311     // TODO: Handle the 'partially zero' case.
312     return nullptr;
313   }
314 
315   case Instruction::ZExt: {
316     unsigned SrcBitSize =
317       cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
318 
319     // If extracting something that is completely zero, return 0.
320     if (ByteStart*8 >= SrcBitSize)
321       return Constant::getNullValue(IntegerType::get(CE->getContext(),
322                                                      ByteSize*8));
323 
324     // If exactly extracting the input, return it.
325     if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
326       return CE->getOperand(0);
327 
328     // If extracting something completely in the input, if the input is a
329     // multiple of 8 bits, recurse.
330     if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
331       return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
332 
333     // Otherwise, if extracting a subset of the input, which is not multiple of
334     // 8 bits, do a shift and trunc to get the bits.
335     if ((ByteStart+ByteSize)*8 < SrcBitSize) {
336       assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
337       Constant *Res = CE->getOperand(0);
338       if (ByteStart)
339         Res = ConstantExpr::getLShr(Res,
340                                  ConstantInt::get(Res->getType(), ByteStart*8));
341       return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
342                                                           ByteSize*8));
343     }
344 
345     // TODO: Handle the 'partially zero' case.
346     return nullptr;
347   }
348   }
349 }
350 
351 /// Wrapper around getFoldedSizeOfImpl() that adds caching.
352 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, bool Folded,
353                                  DenseMap<Type *, Constant *> &Cache);
354 
355 /// Return a ConstantExpr with type DestTy for sizeof on Ty, with any known
356 /// factors factored out. If Folded is false, return null if no factoring was
357 /// possible, to avoid endlessly bouncing an unfoldable expression back into the
358 /// top-level folder.
getFoldedSizeOfImpl(Type * Ty,Type * DestTy,bool Folded,DenseMap<Type *,Constant * > & Cache)359 static Constant *getFoldedSizeOfImpl(Type *Ty, Type *DestTy, bool Folded,
360                                      DenseMap<Type *, Constant *> &Cache) {
361   // This is the actual implementation of getFoldedSizeOf(). To get the caching
362   // behavior, we need to call getFoldedSizeOf() when we recurse.
363 
364   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
365     Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
366     Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true, Cache);
367     return ConstantExpr::getNUWMul(E, N);
368   }
369 
370   if (StructType *STy = dyn_cast<StructType>(Ty))
371     if (!STy->isPacked()) {
372       unsigned NumElems = STy->getNumElements();
373       // An empty struct has size zero.
374       if (NumElems == 0)
375         return ConstantExpr::getNullValue(DestTy);
376       // Check for a struct with all members having the same size.
377       Constant *MemberSize =
378           getFoldedSizeOf(STy->getElementType(0), DestTy, true, Cache);
379       bool AllSame = true;
380       for (unsigned i = 1; i != NumElems; ++i)
381         if (MemberSize !=
382             getFoldedSizeOf(STy->getElementType(i), DestTy, true, Cache)) {
383           AllSame = false;
384           break;
385         }
386       if (AllSame) {
387         Constant *N = ConstantInt::get(DestTy, NumElems);
388         return ConstantExpr::getNUWMul(MemberSize, N);
389       }
390     }
391 
392   // Pointer size doesn't depend on the pointee type, so canonicalize them
393   // to an arbitrary pointee.
394   if (PointerType *PTy = dyn_cast<PointerType>(Ty))
395     if (!PTy->getElementType()->isIntegerTy(1))
396       return getFoldedSizeOf(
397           PointerType::get(IntegerType::get(PTy->getContext(), 1),
398                            PTy->getAddressSpace()),
399           DestTy, true, Cache);
400 
401   // If there's no interesting folding happening, bail so that we don't create
402   // a constant that looks like it needs folding but really doesn't.
403   if (!Folded)
404     return nullptr;
405 
406   // Base case: Get a regular sizeof expression.
407   Constant *C = ConstantExpr::getSizeOf(Ty);
408   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
409                                                     DestTy, false),
410                             C, DestTy);
411   return C;
412 }
413 
getFoldedSizeOf(Type * Ty,Type * DestTy,bool Folded,DenseMap<Type *,Constant * > & Cache)414 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, bool Folded,
415                                  DenseMap<Type *, Constant *> &Cache) {
416   // Check for previously generated folded size constant.
417   auto It = Cache.find(Ty);
418   if (It != Cache.end())
419     return It->second;
420   return Cache[Ty] = getFoldedSizeOfImpl(Ty, DestTy, Folded, Cache);
421 }
422 
getFoldedSizeOf(Type * Ty,Type * DestTy,bool Folded)423 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, bool Folded) {
424   DenseMap<Type *, Constant *> Cache;
425   return getFoldedSizeOf(Ty, DestTy, Folded, Cache);
426 }
427 
428 /// Return a ConstantExpr with type DestTy for alignof on Ty, with any known
429 /// factors factored out. If Folded is false, return null if no factoring was
430 /// possible, to avoid endlessly bouncing an unfoldable expression back into the
431 /// top-level folder.
getFoldedAlignOf(Type * Ty,Type * DestTy,bool Folded)432 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, bool Folded) {
433   // The alignment of an array is equal to the alignment of the
434   // array element. Note that this is not always true for vectors.
435   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
436     Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
437     C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
438                                                       DestTy,
439                                                       false),
440                               C, DestTy);
441     return C;
442   }
443 
444   if (StructType *STy = dyn_cast<StructType>(Ty)) {
445     // Packed structs always have an alignment of 1.
446     if (STy->isPacked())
447       return ConstantInt::get(DestTy, 1);
448 
449     // Otherwise, struct alignment is the maximum alignment of any member.
450     // Without target data, we can't compare much, but we can check to see
451     // if all the members have the same alignment.
452     unsigned NumElems = STy->getNumElements();
453     // An empty struct has minimal alignment.
454     if (NumElems == 0)
455       return ConstantInt::get(DestTy, 1);
456     // Check for a struct with all members having the same alignment.
457     Constant *MemberAlign =
458       getFoldedAlignOf(STy->getElementType(0), DestTy, true);
459     bool AllSame = true;
460     for (unsigned i = 1; i != NumElems; ++i)
461       if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
462         AllSame = false;
463         break;
464       }
465     if (AllSame)
466       return MemberAlign;
467   }
468 
469   // Pointer alignment doesn't depend on the pointee type, so canonicalize them
470   // to an arbitrary pointee.
471   if (PointerType *PTy = dyn_cast<PointerType>(Ty))
472     if (!PTy->getElementType()->isIntegerTy(1))
473       return
474         getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
475                                                            1),
476                                           PTy->getAddressSpace()),
477                          DestTy, true);
478 
479   // If there's no interesting folding happening, bail so that we don't create
480   // a constant that looks like it needs folding but really doesn't.
481   if (!Folded)
482     return nullptr;
483 
484   // Base case: Get a regular alignof expression.
485   Constant *C = ConstantExpr::getAlignOf(Ty);
486   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
487                                                     DestTy, false),
488                             C, DestTy);
489   return C;
490 }
491 
492 /// Return a ConstantExpr with type DestTy for offsetof on Ty and FieldNo, with
493 /// any known factors factored out. If Folded is false, return null if no
494 /// factoring was possible, to avoid endlessly bouncing an unfoldable expression
495 /// back into the top-level folder.
getFoldedOffsetOf(Type * Ty,Constant * FieldNo,Type * DestTy,bool Folded)496 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, Type *DestTy,
497                                    bool Folded) {
498   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
499     Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
500                                                                 DestTy, false),
501                                         FieldNo, DestTy);
502     Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
503     return ConstantExpr::getNUWMul(E, N);
504   }
505 
506   if (StructType *STy = dyn_cast<StructType>(Ty))
507     if (!STy->isPacked()) {
508       unsigned NumElems = STy->getNumElements();
509       // An empty struct has no members.
510       if (NumElems == 0)
511         return nullptr;
512       // Check for a struct with all members having the same size.
513       Constant *MemberSize =
514         getFoldedSizeOf(STy->getElementType(0), DestTy, true);
515       bool AllSame = true;
516       for (unsigned i = 1; i != NumElems; ++i)
517         if (MemberSize !=
518             getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
519           AllSame = false;
520           break;
521         }
522       if (AllSame) {
523         Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
524                                                                     false,
525                                                                     DestTy,
526                                                                     false),
527                                             FieldNo, DestTy);
528         return ConstantExpr::getNUWMul(MemberSize, N);
529       }
530     }
531 
532   // If there's no interesting folding happening, bail so that we don't create
533   // a constant that looks like it needs folding but really doesn't.
534   if (!Folded)
535     return nullptr;
536 
537   // Base case: Get a regular offsetof expression.
538   Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
539   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
540                                                     DestTy, false),
541                             C, DestTy);
542   return C;
543 }
544 
ConstantFoldCastInstruction(unsigned opc,Constant * V,Type * DestTy)545 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
546                                             Type *DestTy) {
547   if (isa<PoisonValue>(V))
548     return PoisonValue::get(DestTy);
549 
550   if (isa<UndefValue>(V)) {
551     // zext(undef) = 0, because the top bits will be zero.
552     // sext(undef) = 0, because the top bits will all be the same.
553     // [us]itofp(undef) = 0, because the result value is bounded.
554     if (opc == Instruction::ZExt || opc == Instruction::SExt ||
555         opc == Instruction::UIToFP || opc == Instruction::SIToFP)
556       return Constant::getNullValue(DestTy);
557     return UndefValue::get(DestTy);
558   }
559 
560   if (V->isNullValue() && !DestTy->isX86_MMXTy() && !DestTy->isX86_AMXTy() &&
561       opc != Instruction::AddrSpaceCast)
562     return Constant::getNullValue(DestTy);
563 
564   // If the cast operand is a constant expression, there's a few things we can
565   // do to try to simplify it.
566   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
567     if (CE->isCast()) {
568       // Try hard to fold cast of cast because they are often eliminable.
569       if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
570         return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
571     } else if (CE->getOpcode() == Instruction::GetElementPtr &&
572                // Do not fold addrspacecast (gep 0, .., 0). It might make the
573                // addrspacecast uncanonicalized.
574                opc != Instruction::AddrSpaceCast &&
575                // Do not fold bitcast (gep) with inrange index, as this loses
576                // information.
577                !cast<GEPOperator>(CE)->getInRangeIndex().hasValue() &&
578                // Do not fold if the gep type is a vector, as bitcasting
579                // operand 0 of a vector gep will result in a bitcast between
580                // different sizes.
581                !CE->getType()->isVectorTy()) {
582       // If all of the indexes in the GEP are null values, there is no pointer
583       // adjustment going on.  We might as well cast the source pointer.
584       bool isAllNull = true;
585       for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
586         if (!CE->getOperand(i)->isNullValue()) {
587           isAllNull = false;
588           break;
589         }
590       if (isAllNull)
591         // This is casting one pointer type to another, always BitCast
592         return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
593     }
594   }
595 
596   // If the cast operand is a constant vector, perform the cast by
597   // operating on each element. In the cast of bitcasts, the element
598   // count may be mismatched; don't attempt to handle that here.
599   if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
600       DestTy->isVectorTy() &&
601       cast<FixedVectorType>(DestTy)->getNumElements() ==
602           cast<FixedVectorType>(V->getType())->getNumElements()) {
603     VectorType *DestVecTy = cast<VectorType>(DestTy);
604     Type *DstEltTy = DestVecTy->getElementType();
605     // Fast path for splatted constants.
606     if (Constant *Splat = V->getSplatValue()) {
607       return ConstantVector::getSplat(
608           cast<VectorType>(DestTy)->getElementCount(),
609           ConstantExpr::getCast(opc, Splat, DstEltTy));
610     }
611     SmallVector<Constant *, 16> res;
612     Type *Ty = IntegerType::get(V->getContext(), 32);
613     for (unsigned i = 0,
614                   e = cast<FixedVectorType>(V->getType())->getNumElements();
615          i != e; ++i) {
616       Constant *C =
617         ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
618       res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
619     }
620     return ConstantVector::get(res);
621   }
622 
623   // We actually have to do a cast now. Perform the cast according to the
624   // opcode specified.
625   switch (opc) {
626   default:
627     llvm_unreachable("Failed to cast constant expression");
628   case Instruction::FPTrunc:
629   case Instruction::FPExt:
630     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
631       bool ignored;
632       APFloat Val = FPC->getValueAPF();
633       Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() :
634                   DestTy->isFloatTy() ? APFloat::IEEEsingle() :
635                   DestTy->isDoubleTy() ? APFloat::IEEEdouble() :
636                   DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() :
637                   DestTy->isFP128Ty() ? APFloat::IEEEquad() :
638                   DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() :
639                   APFloat::Bogus(),
640                   APFloat::rmNearestTiesToEven, &ignored);
641       return ConstantFP::get(V->getContext(), Val);
642     }
643     return nullptr; // Can't fold.
644   case Instruction::FPToUI:
645   case Instruction::FPToSI:
646     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
647       const APFloat &V = FPC->getValueAPF();
648       bool ignored;
649       uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
650       APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI);
651       if (APFloat::opInvalidOp ==
652           V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) {
653         // Undefined behavior invoked - the destination type can't represent
654         // the input constant.
655         return PoisonValue::get(DestTy);
656       }
657       return ConstantInt::get(FPC->getContext(), IntVal);
658     }
659     return nullptr; // Can't fold.
660   case Instruction::IntToPtr:   //always treated as unsigned
661     if (V->isNullValue())       // Is it an integral null value?
662       return ConstantPointerNull::get(cast<PointerType>(DestTy));
663     return nullptr;                   // Other pointer types cannot be casted
664   case Instruction::PtrToInt:   // always treated as unsigned
665     // Is it a null pointer value?
666     if (V->isNullValue())
667       return ConstantInt::get(DestTy, 0);
668     // If this is a sizeof-like expression, pull out multiplications by
669     // known factors to expose them to subsequent folding. If it's an
670     // alignof-like expression, factor out known factors.
671     if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
672       if (CE->getOpcode() == Instruction::GetElementPtr &&
673           CE->getOperand(0)->isNullValue()) {
674         // FIXME: Looks like getFoldedSizeOf(), getFoldedOffsetOf() and
675         // getFoldedAlignOf() don't handle the case when DestTy is a vector of
676         // pointers yet. We end up in asserts in CastInst::getCastOpcode (see
677         // test/Analysis/ConstantFolding/cast-vector.ll). I've only seen this
678         // happen in one "real" C-code test case, so it does not seem to be an
679         // important optimization to handle vectors here. For now, simply bail
680         // out.
681         if (DestTy->isVectorTy())
682           return nullptr;
683         GEPOperator *GEPO = cast<GEPOperator>(CE);
684         Type *Ty = GEPO->getSourceElementType();
685         if (CE->getNumOperands() == 2) {
686           // Handle a sizeof-like expression.
687           Constant *Idx = CE->getOperand(1);
688           bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
689           if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
690             Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
691                                                                 DestTy, false),
692                                         Idx, DestTy);
693             return ConstantExpr::getMul(C, Idx);
694           }
695         } else if (CE->getNumOperands() == 3 &&
696                    CE->getOperand(1)->isNullValue()) {
697           // Handle an alignof-like expression.
698           if (StructType *STy = dyn_cast<StructType>(Ty))
699             if (!STy->isPacked()) {
700               ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
701               if (CI->isOne() &&
702                   STy->getNumElements() == 2 &&
703                   STy->getElementType(0)->isIntegerTy(1)) {
704                 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
705               }
706             }
707           // Handle an offsetof-like expression.
708           if (Ty->isStructTy() || Ty->isArrayTy()) {
709             if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
710                                                 DestTy, false))
711               return C;
712           }
713         }
714       }
715     // Other pointer types cannot be casted
716     return nullptr;
717   case Instruction::UIToFP:
718   case Instruction::SIToFP:
719     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
720       const APInt &api = CI->getValue();
721       APFloat apf(DestTy->getFltSemantics(),
722                   APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
723       apf.convertFromAPInt(api, opc==Instruction::SIToFP,
724                            APFloat::rmNearestTiesToEven);
725       return ConstantFP::get(V->getContext(), apf);
726     }
727     return nullptr;
728   case Instruction::ZExt:
729     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
730       uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
731       return ConstantInt::get(V->getContext(),
732                               CI->getValue().zext(BitWidth));
733     }
734     return nullptr;
735   case Instruction::SExt:
736     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
737       uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
738       return ConstantInt::get(V->getContext(),
739                               CI->getValue().sext(BitWidth));
740     }
741     return nullptr;
742   case Instruction::Trunc: {
743     if (V->getType()->isVectorTy())
744       return nullptr;
745 
746     uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
747     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
748       return ConstantInt::get(V->getContext(),
749                               CI->getValue().trunc(DestBitWidth));
750     }
751 
752     // The input must be a constantexpr.  See if we can simplify this based on
753     // the bytes we are demanding.  Only do this if the source and dest are an
754     // even multiple of a byte.
755     if ((DestBitWidth & 7) == 0 &&
756         (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
757       if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
758         return Res;
759 
760     return nullptr;
761   }
762   case Instruction::BitCast:
763     return FoldBitCast(V, DestTy);
764   case Instruction::AddrSpaceCast:
765     return nullptr;
766   }
767 }
768 
ConstantFoldSelectInstruction(Constant * Cond,Constant * V1,Constant * V2)769 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
770                                               Constant *V1, Constant *V2) {
771   // Check for i1 and vector true/false conditions.
772   if (Cond->isNullValue()) return V2;
773   if (Cond->isAllOnesValue()) return V1;
774 
775   // If the condition is a vector constant, fold the result elementwise.
776   if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
777     auto *V1VTy = CondV->getType();
778     SmallVector<Constant*, 16> Result;
779     Type *Ty = IntegerType::get(CondV->getContext(), 32);
780     for (unsigned i = 0, e = V1VTy->getNumElements(); i != e; ++i) {
781       Constant *V;
782       Constant *V1Element = ConstantExpr::getExtractElement(V1,
783                                                     ConstantInt::get(Ty, i));
784       Constant *V2Element = ConstantExpr::getExtractElement(V2,
785                                                     ConstantInt::get(Ty, i));
786       auto *Cond = cast<Constant>(CondV->getOperand(i));
787       if (isa<PoisonValue>(Cond)) {
788         V = PoisonValue::get(V1Element->getType());
789       } else if (V1Element == V2Element) {
790         V = V1Element;
791       } else if (isa<UndefValue>(Cond)) {
792         V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
793       } else {
794         if (!isa<ConstantInt>(Cond)) break;
795         V = Cond->isNullValue() ? V2Element : V1Element;
796       }
797       Result.push_back(V);
798     }
799 
800     // If we were able to build the vector, return it.
801     if (Result.size() == V1VTy->getNumElements())
802       return ConstantVector::get(Result);
803   }
804 
805   if (isa<PoisonValue>(Cond))
806     return PoisonValue::get(V1->getType());
807 
808   if (isa<UndefValue>(Cond)) {
809     if (isa<UndefValue>(V1)) return V1;
810     return V2;
811   }
812 
813   if (V1 == V2) return V1;
814 
815   if (isa<PoisonValue>(V1))
816     return V2;
817   if (isa<PoisonValue>(V2))
818     return V1;
819 
820   // If the true or false value is undef, we can fold to the other value as
821   // long as the other value isn't poison.
822   auto NotPoison = [](Constant *C) {
823     if (isa<PoisonValue>(C))
824       return false;
825 
826     // TODO: We can analyze ConstExpr by opcode to determine if there is any
827     //       possibility of poison.
828     if (isa<ConstantExpr>(C))
829       return false;
830 
831     if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(C) ||
832         isa<ConstantPointerNull>(C) || isa<Function>(C))
833       return true;
834 
835     if (C->getType()->isVectorTy())
836       return !C->containsPoisonElement() && !C->containsConstantExpression();
837 
838     // TODO: Recursively analyze aggregates or other constants.
839     return false;
840   };
841   if (isa<UndefValue>(V1) && NotPoison(V2)) return V2;
842   if (isa<UndefValue>(V2) && NotPoison(V1)) return V1;
843 
844   if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
845     if (TrueVal->getOpcode() == Instruction::Select)
846       if (TrueVal->getOperand(0) == Cond)
847         return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
848   }
849   if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
850     if (FalseVal->getOpcode() == Instruction::Select)
851       if (FalseVal->getOperand(0) == Cond)
852         return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
853   }
854 
855   return nullptr;
856 }
857 
ConstantFoldExtractElementInstruction(Constant * Val,Constant * Idx)858 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
859                                                       Constant *Idx) {
860   auto *ValVTy = cast<VectorType>(Val->getType());
861 
862   // extractelt poison, C -> poison
863   // extractelt C, undef -> poison
864   if (isa<PoisonValue>(Val) || isa<UndefValue>(Idx))
865     return PoisonValue::get(ValVTy->getElementType());
866 
867   // extractelt undef, C -> undef
868   if (isa<UndefValue>(Val))
869     return UndefValue::get(ValVTy->getElementType());
870 
871   auto *CIdx = dyn_cast<ConstantInt>(Idx);
872   if (!CIdx)
873     return nullptr;
874 
875   if (auto *ValFVTy = dyn_cast<FixedVectorType>(Val->getType())) {
876     // ee({w,x,y,z}, wrong_value) -> poison
877     if (CIdx->uge(ValFVTy->getNumElements()))
878       return PoisonValue::get(ValFVTy->getElementType());
879   }
880 
881   // ee (gep (ptr, idx0, ...), idx) -> gep (ee (ptr, idx), ee (idx0, idx), ...)
882   if (auto *CE = dyn_cast<ConstantExpr>(Val)) {
883     if (CE->getOpcode() == Instruction::GetElementPtr) {
884       SmallVector<Constant *, 8> Ops;
885       Ops.reserve(CE->getNumOperands());
886       for (unsigned i = 0, e = CE->getNumOperands(); i != e; ++i) {
887         Constant *Op = CE->getOperand(i);
888         if (Op->getType()->isVectorTy()) {
889           Constant *ScalarOp = ConstantExpr::getExtractElement(Op, Idx);
890           if (!ScalarOp)
891             return nullptr;
892           Ops.push_back(ScalarOp);
893         } else
894           Ops.push_back(Op);
895       }
896       return CE->getWithOperands(Ops, ValVTy->getElementType(), false,
897                                  Ops[0]->getType()->getPointerElementType());
898     } else if (CE->getOpcode() == Instruction::InsertElement) {
899       if (const auto *IEIdx = dyn_cast<ConstantInt>(CE->getOperand(2))) {
900         if (APSInt::isSameValue(APSInt(IEIdx->getValue()),
901                                 APSInt(CIdx->getValue()))) {
902           return CE->getOperand(1);
903         } else {
904           return ConstantExpr::getExtractElement(CE->getOperand(0), CIdx);
905         }
906       }
907     }
908   }
909 
910   // CAZ of type ScalableVectorType and n < CAZ->getMinNumElements() =>
911   //   extractelt CAZ, n -> 0
912   if (auto *ValSVTy = dyn_cast<ScalableVectorType>(Val->getType())) {
913     if (!CIdx->uge(ValSVTy->getMinNumElements())) {
914       if (auto *CAZ = dyn_cast<ConstantAggregateZero>(Val))
915         return CAZ->getElementValue(CIdx->getZExtValue());
916     }
917     return nullptr;
918   }
919 
920   return Val->getAggregateElement(CIdx);
921 }
922 
ConstantFoldInsertElementInstruction(Constant * Val,Constant * Elt,Constant * Idx)923 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
924                                                      Constant *Elt,
925                                                      Constant *Idx) {
926   if (isa<UndefValue>(Idx))
927     return PoisonValue::get(Val->getType());
928 
929   ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
930   if (!CIdx) return nullptr;
931 
932   // Do not iterate on scalable vector. The num of elements is unknown at
933   // compile-time.
934   if (isa<ScalableVectorType>(Val->getType()))
935     return nullptr;
936 
937   auto *ValTy = cast<FixedVectorType>(Val->getType());
938 
939   unsigned NumElts = ValTy->getNumElements();
940   if (CIdx->uge(NumElts))
941     return PoisonValue::get(Val->getType());
942 
943   SmallVector<Constant*, 16> Result;
944   Result.reserve(NumElts);
945   auto *Ty = Type::getInt32Ty(Val->getContext());
946   uint64_t IdxVal = CIdx->getZExtValue();
947   for (unsigned i = 0; i != NumElts; ++i) {
948     if (i == IdxVal) {
949       Result.push_back(Elt);
950       continue;
951     }
952 
953     Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
954     Result.push_back(C);
955   }
956 
957   return ConstantVector::get(Result);
958 }
959 
ConstantFoldShuffleVectorInstruction(Constant * V1,Constant * V2,ArrayRef<int> Mask)960 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, Constant *V2,
961                                                      ArrayRef<int> Mask) {
962   auto *V1VTy = cast<VectorType>(V1->getType());
963   unsigned MaskNumElts = Mask.size();
964   auto MaskEltCount =
965       ElementCount::get(MaskNumElts, isa<ScalableVectorType>(V1VTy));
966   Type *EltTy = V1VTy->getElementType();
967 
968   // Undefined shuffle mask -> undefined value.
969   if (all_of(Mask, [](int Elt) { return Elt == UndefMaskElem; })) {
970     return UndefValue::get(FixedVectorType::get(EltTy, MaskNumElts));
971   }
972 
973   // If the mask is all zeros this is a splat, no need to go through all
974   // elements.
975   if (all_of(Mask, [](int Elt) { return Elt == 0; }) &&
976       !MaskEltCount.isScalable()) {
977     Type *Ty = IntegerType::get(V1->getContext(), 32);
978     Constant *Elt =
979         ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, 0));
980     return ConstantVector::getSplat(MaskEltCount, Elt);
981   }
982   // Do not iterate on scalable vector. The num of elements is unknown at
983   // compile-time.
984   if (isa<ScalableVectorType>(V1VTy))
985     return nullptr;
986 
987   unsigned SrcNumElts = V1VTy->getElementCount().getKnownMinValue();
988 
989   // Loop over the shuffle mask, evaluating each element.
990   SmallVector<Constant*, 32> Result;
991   for (unsigned i = 0; i != MaskNumElts; ++i) {
992     int Elt = Mask[i];
993     if (Elt == -1) {
994       Result.push_back(UndefValue::get(EltTy));
995       continue;
996     }
997     Constant *InElt;
998     if (unsigned(Elt) >= SrcNumElts*2)
999       InElt = UndefValue::get(EltTy);
1000     else if (unsigned(Elt) >= SrcNumElts) {
1001       Type *Ty = IntegerType::get(V2->getContext(), 32);
1002       InElt =
1003         ConstantExpr::getExtractElement(V2,
1004                                         ConstantInt::get(Ty, Elt - SrcNumElts));
1005     } else {
1006       Type *Ty = IntegerType::get(V1->getContext(), 32);
1007       InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
1008     }
1009     Result.push_back(InElt);
1010   }
1011 
1012   return ConstantVector::get(Result);
1013 }
1014 
ConstantFoldExtractValueInstruction(Constant * Agg,ArrayRef<unsigned> Idxs)1015 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
1016                                                     ArrayRef<unsigned> Idxs) {
1017   // Base case: no indices, so return the entire value.
1018   if (Idxs.empty())
1019     return Agg;
1020 
1021   if (Constant *C = Agg->getAggregateElement(Idxs[0]))
1022     return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
1023 
1024   return nullptr;
1025 }
1026 
ConstantFoldInsertValueInstruction(Constant * Agg,Constant * Val,ArrayRef<unsigned> Idxs)1027 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
1028                                                    Constant *Val,
1029                                                    ArrayRef<unsigned> Idxs) {
1030   // Base case: no indices, so replace the entire value.
1031   if (Idxs.empty())
1032     return Val;
1033 
1034   unsigned NumElts;
1035   if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
1036     NumElts = ST->getNumElements();
1037   else
1038     NumElts = cast<ArrayType>(Agg->getType())->getNumElements();
1039 
1040   SmallVector<Constant*, 32> Result;
1041   for (unsigned i = 0; i != NumElts; ++i) {
1042     Constant *C = Agg->getAggregateElement(i);
1043     if (!C) return nullptr;
1044 
1045     if (Idxs[0] == i)
1046       C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
1047 
1048     Result.push_back(C);
1049   }
1050 
1051   if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
1052     return ConstantStruct::get(ST, Result);
1053   return ConstantArray::get(cast<ArrayType>(Agg->getType()), Result);
1054 }
1055 
ConstantFoldUnaryInstruction(unsigned Opcode,Constant * C)1056 Constant *llvm::ConstantFoldUnaryInstruction(unsigned Opcode, Constant *C) {
1057   assert(Instruction::isUnaryOp(Opcode) && "Non-unary instruction detected");
1058 
1059   // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length
1060   // vectors are always evaluated per element.
1061   bool IsScalableVector = isa<ScalableVectorType>(C->getType());
1062   bool HasScalarUndefOrScalableVectorUndef =
1063       (!C->getType()->isVectorTy() || IsScalableVector) && isa<UndefValue>(C);
1064 
1065   if (HasScalarUndefOrScalableVectorUndef) {
1066     switch (static_cast<Instruction::UnaryOps>(Opcode)) {
1067     case Instruction::FNeg:
1068       return C; // -undef -> undef
1069     case Instruction::UnaryOpsEnd:
1070       llvm_unreachable("Invalid UnaryOp");
1071     }
1072   }
1073 
1074   // Constant should not be UndefValue, unless these are vector constants.
1075   assert(!HasScalarUndefOrScalableVectorUndef && "Unexpected UndefValue");
1076   // We only have FP UnaryOps right now.
1077   assert(!isa<ConstantInt>(C) && "Unexpected Integer UnaryOp");
1078 
1079   if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
1080     const APFloat &CV = CFP->getValueAPF();
1081     switch (Opcode) {
1082     default:
1083       break;
1084     case Instruction::FNeg:
1085       return ConstantFP::get(C->getContext(), neg(CV));
1086     }
1087   } else if (auto *VTy = dyn_cast<FixedVectorType>(C->getType())) {
1088 
1089     Type *Ty = IntegerType::get(VTy->getContext(), 32);
1090     // Fast path for splatted constants.
1091     if (Constant *Splat = C->getSplatValue()) {
1092       Constant *Elt = ConstantExpr::get(Opcode, Splat);
1093       return ConstantVector::getSplat(VTy->getElementCount(), Elt);
1094     }
1095 
1096     // Fold each element and create a vector constant from those constants.
1097     SmallVector<Constant *, 16> Result;
1098     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1099       Constant *ExtractIdx = ConstantInt::get(Ty, i);
1100       Constant *Elt = ConstantExpr::getExtractElement(C, ExtractIdx);
1101 
1102       Result.push_back(ConstantExpr::get(Opcode, Elt));
1103     }
1104 
1105     return ConstantVector::get(Result);
1106   }
1107 
1108   // We don't know how to fold this.
1109   return nullptr;
1110 }
1111 
ConstantFoldBinaryInstruction(unsigned Opcode,Constant * C1,Constant * C2)1112 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1,
1113                                               Constant *C2) {
1114   assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected");
1115 
1116   // Simplify BinOps with their identity values first. They are no-ops and we
1117   // can always return the other value, including undef or poison values.
1118   // FIXME: remove unnecessary duplicated identity patterns below.
1119   // FIXME: Use AllowRHSConstant with getBinOpIdentity to handle additional ops,
1120   //        like X << 0 = X.
1121   Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, C1->getType());
1122   if (Identity) {
1123     if (C1 == Identity)
1124       return C2;
1125     if (C2 == Identity)
1126       return C1;
1127   }
1128 
1129   // Binary operations propagate poison.
1130   // FIXME: Currently, or/and i1 poison aren't folded into poison because
1131   // it causes miscompilation when combined with another optimization in
1132   // InstCombine (select i1 -> and/or). The select fold is wrong, but
1133   // fixing it requires an effort, so temporarily disable this until it is
1134   // fixed.
1135   bool PoisonFold = !C1->getType()->isIntegerTy(1) ||
1136                     (Opcode != Instruction::Or && Opcode != Instruction::And);
1137   if (PoisonFold && (isa<PoisonValue>(C1) || isa<PoisonValue>(C2)))
1138     return PoisonValue::get(C1->getType());
1139 
1140   // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length
1141   // vectors are always evaluated per element.
1142   bool IsScalableVector = isa<ScalableVectorType>(C1->getType());
1143   bool HasScalarUndefOrScalableVectorUndef =
1144       (!C1->getType()->isVectorTy() || IsScalableVector) &&
1145       (isa<UndefValue>(C1) || isa<UndefValue>(C2));
1146   if (HasScalarUndefOrScalableVectorUndef) {
1147     switch (static_cast<Instruction::BinaryOps>(Opcode)) {
1148     case Instruction::Xor:
1149       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
1150         // Handle undef ^ undef -> 0 special case. This is a common
1151         // idiom (misuse).
1152         return Constant::getNullValue(C1->getType());
1153       LLVM_FALLTHROUGH;
1154     case Instruction::Add:
1155     case Instruction::Sub:
1156       return UndefValue::get(C1->getType());
1157     case Instruction::And:
1158       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
1159         return C1;
1160       return Constant::getNullValue(C1->getType());   // undef & X -> 0
1161     case Instruction::Mul: {
1162       // undef * undef -> undef
1163       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
1164         return C1;
1165       const APInt *CV;
1166       // X * undef -> undef   if X is odd
1167       if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV)))
1168         if ((*CV)[0])
1169           return UndefValue::get(C1->getType());
1170 
1171       // X * undef -> 0       otherwise
1172       return Constant::getNullValue(C1->getType());
1173     }
1174     case Instruction::SDiv:
1175     case Instruction::UDiv:
1176       // X / undef -> poison
1177       // X / 0 -> poison
1178       if (match(C2, m_CombineOr(m_Undef(), m_Zero())))
1179         return PoisonValue::get(C2->getType());
1180       // undef / 1 -> undef
1181       if (match(C2, m_One()))
1182         return C1;
1183       // undef / X -> 0       otherwise
1184       return Constant::getNullValue(C1->getType());
1185     case Instruction::URem:
1186     case Instruction::SRem:
1187       // X % undef -> poison
1188       // X % 0 -> poison
1189       if (match(C2, m_CombineOr(m_Undef(), m_Zero())))
1190         return PoisonValue::get(C2->getType());
1191       // undef % X -> 0       otherwise
1192       return Constant::getNullValue(C1->getType());
1193     case Instruction::Or:                          // X | undef -> -1
1194       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
1195         return C1;
1196       return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
1197     case Instruction::LShr:
1198       // X >>l undef -> poison
1199       if (isa<UndefValue>(C2))
1200         return PoisonValue::get(C2->getType());
1201       // undef >>l 0 -> undef
1202       if (match(C2, m_Zero()))
1203         return C1;
1204       // undef >>l X -> 0
1205       return Constant::getNullValue(C1->getType());
1206     case Instruction::AShr:
1207       // X >>a undef -> poison
1208       if (isa<UndefValue>(C2))
1209         return PoisonValue::get(C2->getType());
1210       // undef >>a 0 -> undef
1211       if (match(C2, m_Zero()))
1212         return C1;
1213       // TODO: undef >>a X -> poison if the shift is exact
1214       // undef >>a X -> 0
1215       return Constant::getNullValue(C1->getType());
1216     case Instruction::Shl:
1217       // X << undef -> undef
1218       if (isa<UndefValue>(C2))
1219         return PoisonValue::get(C2->getType());
1220       // undef << 0 -> undef
1221       if (match(C2, m_Zero()))
1222         return C1;
1223       // undef << X -> 0
1224       return Constant::getNullValue(C1->getType());
1225     case Instruction::FSub:
1226       // -0.0 - undef --> undef (consistent with "fneg undef")
1227       if (match(C1, m_NegZeroFP()) && isa<UndefValue>(C2))
1228         return C2;
1229       LLVM_FALLTHROUGH;
1230     case Instruction::FAdd:
1231     case Instruction::FMul:
1232     case Instruction::FDiv:
1233     case Instruction::FRem:
1234       // [any flop] undef, undef -> undef
1235       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
1236         return C1;
1237       // [any flop] C, undef -> NaN
1238       // [any flop] undef, C -> NaN
1239       // We could potentially specialize NaN/Inf constants vs. 'normal'
1240       // constants (possibly differently depending on opcode and operand). This
1241       // would allow returning undef sometimes. But it is always safe to fold to
1242       // NaN because we can choose the undef operand as NaN, and any FP opcode
1243       // with a NaN operand will propagate NaN.
1244       return ConstantFP::getNaN(C1->getType());
1245     case Instruction::BinaryOpsEnd:
1246       llvm_unreachable("Invalid BinaryOp");
1247     }
1248   }
1249 
1250   // Neither constant should be UndefValue, unless these are vector constants.
1251   assert((!HasScalarUndefOrScalableVectorUndef) && "Unexpected UndefValue");
1252 
1253   // Handle simplifications when the RHS is a constant int.
1254   if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1255     switch (Opcode) {
1256     case Instruction::Add:
1257       if (CI2->isZero()) return C1;                             // X + 0 == X
1258       break;
1259     case Instruction::Sub:
1260       if (CI2->isZero()) return C1;                             // X - 0 == X
1261       break;
1262     case Instruction::Mul:
1263       if (CI2->isZero()) return C2;                             // X * 0 == 0
1264       if (CI2->isOne())
1265         return C1;                                              // X * 1 == X
1266       break;
1267     case Instruction::UDiv:
1268     case Instruction::SDiv:
1269       if (CI2->isOne())
1270         return C1;                                            // X / 1 == X
1271       if (CI2->isZero())
1272         return PoisonValue::get(CI2->getType());              // X / 0 == poison
1273       break;
1274     case Instruction::URem:
1275     case Instruction::SRem:
1276       if (CI2->isOne())
1277         return Constant::getNullValue(CI2->getType());        // X % 1 == 0
1278       if (CI2->isZero())
1279         return PoisonValue::get(CI2->getType());              // X % 0 == poison
1280       break;
1281     case Instruction::And:
1282       if (CI2->isZero()) return C2;                           // X & 0 == 0
1283       if (CI2->isMinusOne())
1284         return C1;                                            // X & -1 == X
1285 
1286       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1287         // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1288         if (CE1->getOpcode() == Instruction::ZExt) {
1289           unsigned DstWidth = CI2->getType()->getBitWidth();
1290           unsigned SrcWidth =
1291             CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1292           APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1293           if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1294             return C1;
1295         }
1296 
1297         // If and'ing the address of a global with a constant, fold it.
1298         if (CE1->getOpcode() == Instruction::PtrToInt &&
1299             isa<GlobalValue>(CE1->getOperand(0))) {
1300           GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1301 
1302           MaybeAlign GVAlign;
1303 
1304           if (Module *TheModule = GV->getParent()) {
1305             const DataLayout &DL = TheModule->getDataLayout();
1306             GVAlign = GV->getPointerAlignment(DL);
1307 
1308             // If the function alignment is not specified then assume that it
1309             // is 4.
1310             // This is dangerous; on x86, the alignment of the pointer
1311             // corresponds to the alignment of the function, but might be less
1312             // than 4 if it isn't explicitly specified.
1313             // However, a fix for this behaviour was reverted because it
1314             // increased code size (see https://reviews.llvm.org/D55115)
1315             // FIXME: This code should be deleted once existing targets have
1316             // appropriate defaults
1317             if (isa<Function>(GV) && !DL.getFunctionPtrAlign())
1318               GVAlign = Align(4);
1319           } else if (isa<Function>(GV)) {
1320             // Without a datalayout we have to assume the worst case: that the
1321             // function pointer isn't aligned at all.
1322             GVAlign = llvm::None;
1323           } else if (isa<GlobalVariable>(GV)) {
1324             GVAlign = cast<GlobalVariable>(GV)->getAlign();
1325           }
1326 
1327           if (GVAlign && *GVAlign > 1) {
1328             unsigned DstWidth = CI2->getType()->getBitWidth();
1329             unsigned SrcWidth = std::min(DstWidth, Log2(*GVAlign));
1330             APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1331 
1332             // If checking bits we know are clear, return zero.
1333             if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1334               return Constant::getNullValue(CI2->getType());
1335           }
1336         }
1337       }
1338       break;
1339     case Instruction::Or:
1340       if (CI2->isZero()) return C1;        // X | 0 == X
1341       if (CI2->isMinusOne())
1342         return C2;                         // X | -1 == -1
1343       break;
1344     case Instruction::Xor:
1345       if (CI2->isZero()) return C1;        // X ^ 0 == X
1346 
1347       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1348         switch (CE1->getOpcode()) {
1349         default: break;
1350         case Instruction::ICmp:
1351         case Instruction::FCmp:
1352           // cmp pred ^ true -> cmp !pred
1353           assert(CI2->isOne());
1354           CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1355           pred = CmpInst::getInversePredicate(pred);
1356           return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1357                                           CE1->getOperand(1));
1358         }
1359       }
1360       break;
1361     case Instruction::AShr:
1362       // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1363       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1364         if (CE1->getOpcode() == Instruction::ZExt)  // Top bits known zero.
1365           return ConstantExpr::getLShr(C1, C2);
1366       break;
1367     }
1368   } else if (isa<ConstantInt>(C1)) {
1369     // If C1 is a ConstantInt and C2 is not, swap the operands.
1370     if (Instruction::isCommutative(Opcode))
1371       return ConstantExpr::get(Opcode, C2, C1);
1372   }
1373 
1374   if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1375     if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1376       const APInt &C1V = CI1->getValue();
1377       const APInt &C2V = CI2->getValue();
1378       switch (Opcode) {
1379       default:
1380         break;
1381       case Instruction::Add:
1382         return ConstantInt::get(CI1->getContext(), C1V + C2V);
1383       case Instruction::Sub:
1384         return ConstantInt::get(CI1->getContext(), C1V - C2V);
1385       case Instruction::Mul:
1386         return ConstantInt::get(CI1->getContext(), C1V * C2V);
1387       case Instruction::UDiv:
1388         assert(!CI2->isZero() && "Div by zero handled above");
1389         return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1390       case Instruction::SDiv:
1391         assert(!CI2->isZero() && "Div by zero handled above");
1392         if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1393           return PoisonValue::get(CI1->getType());   // MIN_INT / -1 -> poison
1394         return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1395       case Instruction::URem:
1396         assert(!CI2->isZero() && "Div by zero handled above");
1397         return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1398       case Instruction::SRem:
1399         assert(!CI2->isZero() && "Div by zero handled above");
1400         if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1401           return PoisonValue::get(CI1->getType());   // MIN_INT % -1 -> poison
1402         return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1403       case Instruction::And:
1404         return ConstantInt::get(CI1->getContext(), C1V & C2V);
1405       case Instruction::Or:
1406         return ConstantInt::get(CI1->getContext(), C1V | C2V);
1407       case Instruction::Xor:
1408         return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1409       case Instruction::Shl:
1410         if (C2V.ult(C1V.getBitWidth()))
1411           return ConstantInt::get(CI1->getContext(), C1V.shl(C2V));
1412         return PoisonValue::get(C1->getType()); // too big shift is poison
1413       case Instruction::LShr:
1414         if (C2V.ult(C1V.getBitWidth()))
1415           return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V));
1416         return PoisonValue::get(C1->getType()); // too big shift is poison
1417       case Instruction::AShr:
1418         if (C2V.ult(C1V.getBitWidth()))
1419           return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V));
1420         return PoisonValue::get(C1->getType()); // too big shift is poison
1421       }
1422     }
1423 
1424     switch (Opcode) {
1425     case Instruction::SDiv:
1426     case Instruction::UDiv:
1427     case Instruction::URem:
1428     case Instruction::SRem:
1429     case Instruction::LShr:
1430     case Instruction::AShr:
1431     case Instruction::Shl:
1432       if (CI1->isZero()) return C1;
1433       break;
1434     default:
1435       break;
1436     }
1437   } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1438     if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1439       const APFloat &C1V = CFP1->getValueAPF();
1440       const APFloat &C2V = CFP2->getValueAPF();
1441       APFloat C3V = C1V;  // copy for modification
1442       switch (Opcode) {
1443       default:
1444         break;
1445       case Instruction::FAdd:
1446         (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1447         return ConstantFP::get(C1->getContext(), C3V);
1448       case Instruction::FSub:
1449         (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1450         return ConstantFP::get(C1->getContext(), C3V);
1451       case Instruction::FMul:
1452         (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1453         return ConstantFP::get(C1->getContext(), C3V);
1454       case Instruction::FDiv:
1455         (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1456         return ConstantFP::get(C1->getContext(), C3V);
1457       case Instruction::FRem:
1458         (void)C3V.mod(C2V);
1459         return ConstantFP::get(C1->getContext(), C3V);
1460       }
1461     }
1462   } else if (auto *VTy = dyn_cast<VectorType>(C1->getType())) {
1463     // Fast path for splatted constants.
1464     if (Constant *C2Splat = C2->getSplatValue()) {
1465       if (Instruction::isIntDivRem(Opcode) && C2Splat->isNullValue())
1466         return PoisonValue::get(VTy);
1467       if (Constant *C1Splat = C1->getSplatValue()) {
1468         return ConstantVector::getSplat(
1469             VTy->getElementCount(),
1470             ConstantExpr::get(Opcode, C1Splat, C2Splat));
1471       }
1472     }
1473 
1474     if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) {
1475       // Fold each element and create a vector constant from those constants.
1476       SmallVector<Constant*, 16> Result;
1477       Type *Ty = IntegerType::get(FVTy->getContext(), 32);
1478       for (unsigned i = 0, e = FVTy->getNumElements(); i != e; ++i) {
1479         Constant *ExtractIdx = ConstantInt::get(Ty, i);
1480         Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx);
1481         Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx);
1482 
1483         // If any element of a divisor vector is zero, the whole op is poison.
1484         if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue())
1485           return PoisonValue::get(VTy);
1486 
1487         Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1488       }
1489 
1490       return ConstantVector::get(Result);
1491     }
1492   }
1493 
1494   if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1495     // There are many possible foldings we could do here.  We should probably
1496     // at least fold add of a pointer with an integer into the appropriate
1497     // getelementptr.  This will improve alias analysis a bit.
1498 
1499     // Given ((a + b) + c), if (b + c) folds to something interesting, return
1500     // (a + (b + c)).
1501     if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1502       Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1503       if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1504         return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1505     }
1506   } else if (isa<ConstantExpr>(C2)) {
1507     // If C2 is a constant expr and C1 isn't, flop them around and fold the
1508     // other way if possible.
1509     if (Instruction::isCommutative(Opcode))
1510       return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1511   }
1512 
1513   // i1 can be simplified in many cases.
1514   if (C1->getType()->isIntegerTy(1)) {
1515     switch (Opcode) {
1516     case Instruction::Add:
1517     case Instruction::Sub:
1518       return ConstantExpr::getXor(C1, C2);
1519     case Instruction::Mul:
1520       return ConstantExpr::getAnd(C1, C2);
1521     case Instruction::Shl:
1522     case Instruction::LShr:
1523     case Instruction::AShr:
1524       // We can assume that C2 == 0.  If it were one the result would be
1525       // undefined because the shift value is as large as the bitwidth.
1526       return C1;
1527     case Instruction::SDiv:
1528     case Instruction::UDiv:
1529       // We can assume that C2 == 1.  If it were zero the result would be
1530       // undefined through division by zero.
1531       return C1;
1532     case Instruction::URem:
1533     case Instruction::SRem:
1534       // We can assume that C2 == 1.  If it were zero the result would be
1535       // undefined through division by zero.
1536       return ConstantInt::getFalse(C1->getContext());
1537     default:
1538       break;
1539     }
1540   }
1541 
1542   // We don't know how to fold this.
1543   return nullptr;
1544 }
1545 
1546 /// This type is zero-sized if it's an array or structure of zero-sized types.
1547 /// The only leaf zero-sized type is an empty structure.
isMaybeZeroSizedType(Type * Ty)1548 static bool isMaybeZeroSizedType(Type *Ty) {
1549   if (StructType *STy = dyn_cast<StructType>(Ty)) {
1550     if (STy->isOpaque()) return true;  // Can't say.
1551 
1552     // If all of elements have zero size, this does too.
1553     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1554       if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1555     return true;
1556 
1557   } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1558     return isMaybeZeroSizedType(ATy->getElementType());
1559   }
1560   return false;
1561 }
1562 
1563 /// Compare the two constants as though they were getelementptr indices.
1564 /// This allows coercion of the types to be the same thing.
1565 ///
1566 /// If the two constants are the "same" (after coercion), return 0.  If the
1567 /// first is less than the second, return -1, if the second is less than the
1568 /// first, return 1.  If the constants are not integral, return -2.
1569 ///
IdxCompare(Constant * C1,Constant * C2,Type * ElTy)1570 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1571   if (C1 == C2) return 0;
1572 
1573   // Ok, we found a different index.  If they are not ConstantInt, we can't do
1574   // anything with them.
1575   if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1576     return -2; // don't know!
1577 
1578   // We cannot compare the indices if they don't fit in an int64_t.
1579   if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 ||
1580       cast<ConstantInt>(C2)->getValue().getActiveBits() > 64)
1581     return -2; // don't know!
1582 
1583   // Ok, we have two differing integer indices.  Sign extend them to be the same
1584   // type.
1585   int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue();
1586   int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue();
1587 
1588   if (C1Val == C2Val) return 0;  // They are equal
1589 
1590   // If the type being indexed over is really just a zero sized type, there is
1591   // no pointer difference being made here.
1592   if (isMaybeZeroSizedType(ElTy))
1593     return -2; // dunno.
1594 
1595   // If they are really different, now that they are the same type, then we
1596   // found a difference!
1597   if (C1Val < C2Val)
1598     return -1;
1599   else
1600     return 1;
1601 }
1602 
1603 /// This function determines if there is anything we can decide about the two
1604 /// constants provided. This doesn't need to handle simple things like
1605 /// ConstantFP comparisons, but should instead handle ConstantExprs.
1606 /// If we can determine that the two constants have a particular relation to
1607 /// each other, we should return the corresponding FCmpInst predicate,
1608 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1609 /// ConstantFoldCompareInstruction.
1610 ///
1611 /// To simplify this code we canonicalize the relation so that the first
1612 /// operand is always the most "complex" of the two.  We consider ConstantFP
1613 /// to be the simplest, and ConstantExprs to be the most complex.
evaluateFCmpRelation(Constant * V1,Constant * V2)1614 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1615   assert(V1->getType() == V2->getType() &&
1616          "Cannot compare values of different types!");
1617 
1618   // We do not know if a constant expression will evaluate to a number or NaN.
1619   // Therefore, we can only say that the relation is unordered or equal.
1620   if (V1 == V2) return FCmpInst::FCMP_UEQ;
1621 
1622   if (!isa<ConstantExpr>(V1)) {
1623     if (!isa<ConstantExpr>(V2)) {
1624       // Simple case, use the standard constant folder.
1625       ConstantInt *R = nullptr;
1626       R = dyn_cast<ConstantInt>(
1627                       ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1628       if (R && !R->isZero())
1629         return FCmpInst::FCMP_OEQ;
1630       R = dyn_cast<ConstantInt>(
1631                       ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1632       if (R && !R->isZero())
1633         return FCmpInst::FCMP_OLT;
1634       R = dyn_cast<ConstantInt>(
1635                       ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1636       if (R && !R->isZero())
1637         return FCmpInst::FCMP_OGT;
1638 
1639       // Nothing more we can do
1640       return FCmpInst::BAD_FCMP_PREDICATE;
1641     }
1642 
1643     // If the first operand is simple and second is ConstantExpr, swap operands.
1644     FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1645     if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1646       return FCmpInst::getSwappedPredicate(SwappedRelation);
1647   } else {
1648     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1649     // constantexpr or a simple constant.
1650     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1651     switch (CE1->getOpcode()) {
1652     case Instruction::FPTrunc:
1653     case Instruction::FPExt:
1654     case Instruction::UIToFP:
1655     case Instruction::SIToFP:
1656       // We might be able to do something with these but we don't right now.
1657       break;
1658     default:
1659       break;
1660     }
1661   }
1662   // There are MANY other foldings that we could perform here.  They will
1663   // probably be added on demand, as they seem needed.
1664   return FCmpInst::BAD_FCMP_PREDICATE;
1665 }
1666 
areGlobalsPotentiallyEqual(const GlobalValue * GV1,const GlobalValue * GV2)1667 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1668                                                       const GlobalValue *GV2) {
1669   auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) {
1670     if (GV->isInterposable() || GV->hasGlobalUnnamedAddr())
1671       return true;
1672     if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) {
1673       Type *Ty = GVar->getValueType();
1674       // A global with opaque type might end up being zero sized.
1675       if (!Ty->isSized())
1676         return true;
1677       // A global with an empty type might lie at the address of any other
1678       // global.
1679       if (Ty->isEmptyTy())
1680         return true;
1681     }
1682     return false;
1683   };
1684   // Don't try to decide equality of aliases.
1685   if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1686     if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2))
1687       return ICmpInst::ICMP_NE;
1688   return ICmpInst::BAD_ICMP_PREDICATE;
1689 }
1690 
1691 /// This function determines if there is anything we can decide about the two
1692 /// constants provided. This doesn't need to handle simple things like integer
1693 /// comparisons, but should instead handle ConstantExprs and GlobalValues.
1694 /// If we can determine that the two constants have a particular relation to
1695 /// each other, we should return the corresponding ICmp predicate, otherwise
1696 /// return ICmpInst::BAD_ICMP_PREDICATE.
1697 ///
1698 /// To simplify this code we canonicalize the relation so that the first
1699 /// operand is always the most "complex" of the two.  We consider simple
1700 /// constants (like ConstantInt) to be the simplest, followed by
1701 /// GlobalValues, followed by ConstantExpr's (the most complex).
1702 ///
evaluateICmpRelation(Constant * V1,Constant * V2,bool isSigned)1703 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1704                                                 bool isSigned) {
1705   assert(V1->getType() == V2->getType() &&
1706          "Cannot compare different types of values!");
1707   if (V1 == V2) return ICmpInst::ICMP_EQ;
1708 
1709   if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1710       !isa<BlockAddress>(V1)) {
1711     if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1712         !isa<BlockAddress>(V2)) {
1713       // We distilled this down to a simple case, use the standard constant
1714       // folder.
1715       ConstantInt *R = nullptr;
1716       ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1717       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1718       if (R && !R->isZero())
1719         return pred;
1720       pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1721       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1722       if (R && !R->isZero())
1723         return pred;
1724       pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1725       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1726       if (R && !R->isZero())
1727         return pred;
1728 
1729       // If we couldn't figure it out, bail.
1730       return ICmpInst::BAD_ICMP_PREDICATE;
1731     }
1732 
1733     // If the first operand is simple, swap operands.
1734     ICmpInst::Predicate SwappedRelation =
1735       evaluateICmpRelation(V2, V1, isSigned);
1736     if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1737       return ICmpInst::getSwappedPredicate(SwappedRelation);
1738 
1739   } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1740     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1741       ICmpInst::Predicate SwappedRelation =
1742         evaluateICmpRelation(V2, V1, isSigned);
1743       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1744         return ICmpInst::getSwappedPredicate(SwappedRelation);
1745       return ICmpInst::BAD_ICMP_PREDICATE;
1746     }
1747 
1748     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1749     // constant (which, since the types must match, means that it's a
1750     // ConstantPointerNull).
1751     if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1752       return areGlobalsPotentiallyEqual(GV, GV2);
1753     } else if (isa<BlockAddress>(V2)) {
1754       return ICmpInst::ICMP_NE; // Globals never equal labels.
1755     } else {
1756       assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1757       // GlobalVals can never be null unless they have external weak linkage.
1758       // We don't try to evaluate aliases here.
1759       // NOTE: We should not be doing this constant folding if null pointer
1760       // is considered valid for the function. But currently there is no way to
1761       // query it from the Constant type.
1762       if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) &&
1763           !NullPointerIsDefined(nullptr /* F */,
1764                                 GV->getType()->getAddressSpace()))
1765         return ICmpInst::ICMP_UGT;
1766     }
1767   } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1768     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1769       ICmpInst::Predicate SwappedRelation =
1770         evaluateICmpRelation(V2, V1, isSigned);
1771       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1772         return ICmpInst::getSwappedPredicate(SwappedRelation);
1773       return ICmpInst::BAD_ICMP_PREDICATE;
1774     }
1775 
1776     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1777     // constant (which, since the types must match, means that it is a
1778     // ConstantPointerNull).
1779     if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1780       // Block address in another function can't equal this one, but block
1781       // addresses in the current function might be the same if blocks are
1782       // empty.
1783       if (BA2->getFunction() != BA->getFunction())
1784         return ICmpInst::ICMP_NE;
1785     } else {
1786       // Block addresses aren't null, don't equal the address of globals.
1787       assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1788              "Canonicalization guarantee!");
1789       return ICmpInst::ICMP_NE;
1790     }
1791   } else {
1792     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1793     // constantexpr, a global, block address, or a simple constant.
1794     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1795     Constant *CE1Op0 = CE1->getOperand(0);
1796 
1797     switch (CE1->getOpcode()) {
1798     case Instruction::Trunc:
1799     case Instruction::FPTrunc:
1800     case Instruction::FPExt:
1801     case Instruction::FPToUI:
1802     case Instruction::FPToSI:
1803       break; // We can't evaluate floating point casts or truncations.
1804 
1805     case Instruction::BitCast:
1806       // If this is a global value cast, check to see if the RHS is also a
1807       // GlobalValue.
1808       if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0))
1809         if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2))
1810           return areGlobalsPotentiallyEqual(GV, GV2);
1811       LLVM_FALLTHROUGH;
1812     case Instruction::UIToFP:
1813     case Instruction::SIToFP:
1814     case Instruction::ZExt:
1815     case Instruction::SExt:
1816       // We can't evaluate floating point casts or truncations.
1817       if (CE1Op0->getType()->isFPOrFPVectorTy())
1818         break;
1819 
1820       // If the cast is not actually changing bits, and the second operand is a
1821       // null pointer, do the comparison with the pre-casted value.
1822       if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) {
1823         if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1824         if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1825         return evaluateICmpRelation(CE1Op0,
1826                                     Constant::getNullValue(CE1Op0->getType()),
1827                                     isSigned);
1828       }
1829       break;
1830 
1831     case Instruction::GetElementPtr: {
1832       GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1833       // Ok, since this is a getelementptr, we know that the constant has a
1834       // pointer type.  Check the various cases.
1835       if (isa<ConstantPointerNull>(V2)) {
1836         // If we are comparing a GEP to a null pointer, check to see if the base
1837         // of the GEP equals the null pointer.
1838         if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1839           // If its not weak linkage, the GVal must have a non-zero address
1840           // so the result is greater-than
1841           if (!GV->hasExternalWeakLinkage())
1842             return ICmpInst::ICMP_UGT;
1843         } else if (isa<ConstantPointerNull>(CE1Op0)) {
1844           // If we are indexing from a null pointer, check to see if we have any
1845           // non-zero indices.
1846           for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1847             if (!CE1->getOperand(i)->isNullValue())
1848               // Offsetting from null, must not be equal.
1849               return ICmpInst::ICMP_UGT;
1850           // Only zero indexes from null, must still be zero.
1851           return ICmpInst::ICMP_EQ;
1852         }
1853         // Otherwise, we can't really say if the first operand is null or not.
1854       } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1855         if (isa<ConstantPointerNull>(CE1Op0)) {
1856           // If its not weak linkage, the GVal must have a non-zero address
1857           // so the result is less-than
1858           if (!GV2->hasExternalWeakLinkage())
1859             return ICmpInst::ICMP_ULT;
1860         } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1861           if (GV == GV2) {
1862             // If this is a getelementptr of the same global, then it must be
1863             // different.  Because the types must match, the getelementptr could
1864             // only have at most one index, and because we fold getelementptr's
1865             // with a single zero index, it must be nonzero.
1866             assert(CE1->getNumOperands() == 2 &&
1867                    !CE1->getOperand(1)->isNullValue() &&
1868                    "Surprising getelementptr!");
1869             return ICmpInst::ICMP_UGT;
1870           } else {
1871             if (CE1GEP->hasAllZeroIndices())
1872               return areGlobalsPotentiallyEqual(GV, GV2);
1873             return ICmpInst::BAD_ICMP_PREDICATE;
1874           }
1875         }
1876       } else {
1877         ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1878         Constant *CE2Op0 = CE2->getOperand(0);
1879 
1880         // There are MANY other foldings that we could perform here.  They will
1881         // probably be added on demand, as they seem needed.
1882         switch (CE2->getOpcode()) {
1883         default: break;
1884         case Instruction::GetElementPtr:
1885           // By far the most common case to handle is when the base pointers are
1886           // obviously to the same global.
1887           if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1888             // Don't know relative ordering, but check for inequality.
1889             if (CE1Op0 != CE2Op0) {
1890               GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1891               if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1892                 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1893                                                   cast<GlobalValue>(CE2Op0));
1894               return ICmpInst::BAD_ICMP_PREDICATE;
1895             }
1896             // Ok, we know that both getelementptr instructions are based on the
1897             // same global.  From this, we can precisely determine the relative
1898             // ordering of the resultant pointers.
1899             unsigned i = 1;
1900 
1901             // The logic below assumes that the result of the comparison
1902             // can be determined by finding the first index that differs.
1903             // This doesn't work if there is over-indexing in any
1904             // subsequent indices, so check for that case first.
1905             if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1906                 !CE2->isGEPWithNoNotionalOverIndexing())
1907                return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1908 
1909             // Compare all of the operands the GEP's have in common.
1910             gep_type_iterator GTI = gep_type_begin(CE1);
1911             for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1912                  ++i, ++GTI)
1913               switch (IdxCompare(CE1->getOperand(i),
1914                                  CE2->getOperand(i), GTI.getIndexedType())) {
1915               case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1916               case 1:  return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1917               case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1918               }
1919 
1920             // Ok, we ran out of things they have in common.  If any leftovers
1921             // are non-zero then we have a difference, otherwise we are equal.
1922             for (; i < CE1->getNumOperands(); ++i)
1923               if (!CE1->getOperand(i)->isNullValue()) {
1924                 if (isa<ConstantInt>(CE1->getOperand(i)))
1925                   return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1926                 else
1927                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1928               }
1929 
1930             for (; i < CE2->getNumOperands(); ++i)
1931               if (!CE2->getOperand(i)->isNullValue()) {
1932                 if (isa<ConstantInt>(CE2->getOperand(i)))
1933                   return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1934                 else
1935                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1936               }
1937             return ICmpInst::ICMP_EQ;
1938           }
1939         }
1940       }
1941       break;
1942     }
1943     default:
1944       break;
1945     }
1946   }
1947 
1948   return ICmpInst::BAD_ICMP_PREDICATE;
1949 }
1950 
ConstantFoldCompareInstruction(unsigned short pred,Constant * C1,Constant * C2)1951 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1952                                                Constant *C1, Constant *C2) {
1953   Type *ResultTy;
1954   if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1955     ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1956                                VT->getElementCount());
1957   else
1958     ResultTy = Type::getInt1Ty(C1->getContext());
1959 
1960   // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1961   if (pred == FCmpInst::FCMP_FALSE)
1962     return Constant::getNullValue(ResultTy);
1963 
1964   if (pred == FCmpInst::FCMP_TRUE)
1965     return Constant::getAllOnesValue(ResultTy);
1966 
1967   // Handle some degenerate cases first
1968   if (isa<PoisonValue>(C1) || isa<PoisonValue>(C2))
1969     return PoisonValue::get(ResultTy);
1970 
1971   if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1972     CmpInst::Predicate Predicate = CmpInst::Predicate(pred);
1973     bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate);
1974     // For EQ and NE, we can always pick a value for the undef to make the
1975     // predicate pass or fail, so we can return undef.
1976     // Also, if both operands are undef, we can return undef for int comparison.
1977     if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2))
1978       return UndefValue::get(ResultTy);
1979 
1980     // Otherwise, for integer compare, pick the same value as the non-undef
1981     // operand, and fold it to true or false.
1982     if (isIntegerPredicate)
1983       return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate));
1984 
1985     // Choosing NaN for the undef will always make unordered comparison succeed
1986     // and ordered comparison fails.
1987     return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate));
1988   }
1989 
1990   // icmp eq/ne(null,GV) -> false/true
1991   if (C1->isNullValue()) {
1992     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1993       // Don't try to evaluate aliases.  External weak GV can be null.
1994       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1995           !NullPointerIsDefined(nullptr /* F */,
1996                                 GV->getType()->getAddressSpace())) {
1997         if (pred == ICmpInst::ICMP_EQ)
1998           return ConstantInt::getFalse(C1->getContext());
1999         else if (pred == ICmpInst::ICMP_NE)
2000           return ConstantInt::getTrue(C1->getContext());
2001       }
2002   // icmp eq/ne(GV,null) -> false/true
2003   } else if (C2->isNullValue()) {
2004     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) {
2005       // Don't try to evaluate aliases.  External weak GV can be null.
2006       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
2007           !NullPointerIsDefined(nullptr /* F */,
2008                                 GV->getType()->getAddressSpace())) {
2009         if (pred == ICmpInst::ICMP_EQ)
2010           return ConstantInt::getFalse(C1->getContext());
2011         else if (pred == ICmpInst::ICMP_NE)
2012           return ConstantInt::getTrue(C1->getContext());
2013       }
2014     }
2015 
2016     // The caller is expected to commute the operands if the constant expression
2017     // is C2.
2018     // C1 >= 0 --> true
2019     if (pred == ICmpInst::ICMP_UGE)
2020       return Constant::getAllOnesValue(ResultTy);
2021     // C1 < 0 --> false
2022     if (pred == ICmpInst::ICMP_ULT)
2023       return Constant::getNullValue(ResultTy);
2024   }
2025 
2026   // If the comparison is a comparison between two i1's, simplify it.
2027   if (C1->getType()->isIntegerTy(1)) {
2028     switch(pred) {
2029     case ICmpInst::ICMP_EQ:
2030       if (isa<ConstantInt>(C2))
2031         return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
2032       return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
2033     case ICmpInst::ICMP_NE:
2034       return ConstantExpr::getXor(C1, C2);
2035     default:
2036       break;
2037     }
2038   }
2039 
2040   if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
2041     const APInt &V1 = cast<ConstantInt>(C1)->getValue();
2042     const APInt &V2 = cast<ConstantInt>(C2)->getValue();
2043     switch (pred) {
2044     default: llvm_unreachable("Invalid ICmp Predicate");
2045     case ICmpInst::ICMP_EQ:  return ConstantInt::get(ResultTy, V1 == V2);
2046     case ICmpInst::ICMP_NE:  return ConstantInt::get(ResultTy, V1 != V2);
2047     case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
2048     case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
2049     case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
2050     case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
2051     case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
2052     case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
2053     case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
2054     case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
2055     }
2056   } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
2057     const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF();
2058     const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF();
2059     APFloat::cmpResult R = C1V.compare(C2V);
2060     switch (pred) {
2061     default: llvm_unreachable("Invalid FCmp Predicate");
2062     case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
2063     case FCmpInst::FCMP_TRUE:  return Constant::getAllOnesValue(ResultTy);
2064     case FCmpInst::FCMP_UNO:
2065       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
2066     case FCmpInst::FCMP_ORD:
2067       return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
2068     case FCmpInst::FCMP_UEQ:
2069       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
2070                                         R==APFloat::cmpEqual);
2071     case FCmpInst::FCMP_OEQ:
2072       return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
2073     case FCmpInst::FCMP_UNE:
2074       return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
2075     case FCmpInst::FCMP_ONE:
2076       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
2077                                         R==APFloat::cmpGreaterThan);
2078     case FCmpInst::FCMP_ULT:
2079       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
2080                                         R==APFloat::cmpLessThan);
2081     case FCmpInst::FCMP_OLT:
2082       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
2083     case FCmpInst::FCMP_UGT:
2084       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
2085                                         R==APFloat::cmpGreaterThan);
2086     case FCmpInst::FCMP_OGT:
2087       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
2088     case FCmpInst::FCMP_ULE:
2089       return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
2090     case FCmpInst::FCMP_OLE:
2091       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
2092                                         R==APFloat::cmpEqual);
2093     case FCmpInst::FCMP_UGE:
2094       return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
2095     case FCmpInst::FCMP_OGE:
2096       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
2097                                         R==APFloat::cmpEqual);
2098     }
2099   } else if (auto *C1VTy = dyn_cast<VectorType>(C1->getType())) {
2100 
2101     // Fast path for splatted constants.
2102     if (Constant *C1Splat = C1->getSplatValue())
2103       if (Constant *C2Splat = C2->getSplatValue())
2104         return ConstantVector::getSplat(
2105             C1VTy->getElementCount(),
2106             ConstantExpr::getCompare(pred, C1Splat, C2Splat));
2107 
2108     // Do not iterate on scalable vector. The number of elements is unknown at
2109     // compile-time.
2110     if (isa<ScalableVectorType>(C1VTy))
2111       return nullptr;
2112 
2113     // If we can constant fold the comparison of each element, constant fold
2114     // the whole vector comparison.
2115     SmallVector<Constant*, 4> ResElts;
2116     Type *Ty = IntegerType::get(C1->getContext(), 32);
2117     // Compare the elements, producing an i1 result or constant expr.
2118     for (unsigned I = 0, E = C1VTy->getElementCount().getKnownMinValue();
2119          I != E; ++I) {
2120       Constant *C1E =
2121           ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, I));
2122       Constant *C2E =
2123           ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, I));
2124 
2125       ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
2126     }
2127 
2128     return ConstantVector::get(ResElts);
2129   }
2130 
2131   if (C1->getType()->isFloatingPointTy() &&
2132       // Only call evaluateFCmpRelation if we have a constant expr to avoid
2133       // infinite recursive loop
2134       (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) {
2135     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
2136     switch (evaluateFCmpRelation(C1, C2)) {
2137     default: llvm_unreachable("Unknown relation!");
2138     case FCmpInst::FCMP_UNO:
2139     case FCmpInst::FCMP_ORD:
2140     case FCmpInst::FCMP_UNE:
2141     case FCmpInst::FCMP_ULT:
2142     case FCmpInst::FCMP_UGT:
2143     case FCmpInst::FCMP_ULE:
2144     case FCmpInst::FCMP_UGE:
2145     case FCmpInst::FCMP_TRUE:
2146     case FCmpInst::FCMP_FALSE:
2147     case FCmpInst::BAD_FCMP_PREDICATE:
2148       break; // Couldn't determine anything about these constants.
2149     case FCmpInst::FCMP_OEQ: // We know that C1 == C2
2150       Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
2151                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
2152                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
2153       break;
2154     case FCmpInst::FCMP_OLT: // We know that C1 < C2
2155       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
2156                 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
2157                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
2158       break;
2159     case FCmpInst::FCMP_OGT: // We know that C1 > C2
2160       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
2161                 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
2162                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
2163       break;
2164     case FCmpInst::FCMP_OLE: // We know that C1 <= C2
2165       // We can only partially decide this relation.
2166       if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
2167         Result = 0;
2168       else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
2169         Result = 1;
2170       break;
2171     case FCmpInst::FCMP_OGE: // We known that C1 >= C2
2172       // We can only partially decide this relation.
2173       if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
2174         Result = 0;
2175       else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
2176         Result = 1;
2177       break;
2178     case FCmpInst::FCMP_ONE: // We know that C1 != C2
2179       // We can only partially decide this relation.
2180       if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
2181         Result = 0;
2182       else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
2183         Result = 1;
2184       break;
2185     case FCmpInst::FCMP_UEQ: // We know that C1 == C2 || isUnordered(C1, C2).
2186       // We can only partially decide this relation.
2187       if (pred == FCmpInst::FCMP_ONE)
2188         Result = 0;
2189       else if (pred == FCmpInst::FCMP_UEQ)
2190         Result = 1;
2191       break;
2192     }
2193 
2194     // If we evaluated the result, return it now.
2195     if (Result != -1)
2196       return ConstantInt::get(ResultTy, Result);
2197 
2198   } else {
2199     // Evaluate the relation between the two constants, per the predicate.
2200     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
2201     switch (evaluateICmpRelation(C1, C2,
2202                                  CmpInst::isSigned((CmpInst::Predicate)pred))) {
2203     default: llvm_unreachable("Unknown relational!");
2204     case ICmpInst::BAD_ICMP_PREDICATE:
2205       break;  // Couldn't determine anything about these constants.
2206     case ICmpInst::ICMP_EQ:   // We know the constants are equal!
2207       // If we know the constants are equal, we can decide the result of this
2208       // computation precisely.
2209       Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
2210       break;
2211     case ICmpInst::ICMP_ULT:
2212       switch (pred) {
2213       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
2214         Result = 1; break;
2215       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
2216         Result = 0; break;
2217       }
2218       break;
2219     case ICmpInst::ICMP_SLT:
2220       switch (pred) {
2221       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
2222         Result = 1; break;
2223       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
2224         Result = 0; break;
2225       }
2226       break;
2227     case ICmpInst::ICMP_UGT:
2228       switch (pred) {
2229       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
2230         Result = 1; break;
2231       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
2232         Result = 0; break;
2233       }
2234       break;
2235     case ICmpInst::ICMP_SGT:
2236       switch (pred) {
2237       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
2238         Result = 1; break;
2239       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
2240         Result = 0; break;
2241       }
2242       break;
2243     case ICmpInst::ICMP_ULE:
2244       if (pred == ICmpInst::ICMP_UGT) Result = 0;
2245       if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
2246       break;
2247     case ICmpInst::ICMP_SLE:
2248       if (pred == ICmpInst::ICMP_SGT) Result = 0;
2249       if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
2250       break;
2251     case ICmpInst::ICMP_UGE:
2252       if (pred == ICmpInst::ICMP_ULT) Result = 0;
2253       if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2254       break;
2255     case ICmpInst::ICMP_SGE:
2256       if (pred == ICmpInst::ICMP_SLT) Result = 0;
2257       if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2258       break;
2259     case ICmpInst::ICMP_NE:
2260       if (pred == ICmpInst::ICMP_EQ) Result = 0;
2261       if (pred == ICmpInst::ICMP_NE) Result = 1;
2262       break;
2263     }
2264 
2265     // If we evaluated the result, return it now.
2266     if (Result != -1)
2267       return ConstantInt::get(ResultTy, Result);
2268 
2269     // If the right hand side is a bitcast, try using its inverse to simplify
2270     // it by moving it to the left hand side.  We can't do this if it would turn
2271     // a vector compare into a scalar compare or visa versa, or if it would turn
2272     // the operands into FP values.
2273     if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2274       Constant *CE2Op0 = CE2->getOperand(0);
2275       if (CE2->getOpcode() == Instruction::BitCast &&
2276           CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy() &&
2277           !CE2Op0->getType()->isFPOrFPVectorTy()) {
2278         Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2279         return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2280       }
2281     }
2282 
2283     // If the left hand side is an extension, try eliminating it.
2284     if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2285       if ((CE1->getOpcode() == Instruction::SExt &&
2286            ICmpInst::isSigned((ICmpInst::Predicate)pred)) ||
2287           (CE1->getOpcode() == Instruction::ZExt &&
2288            !ICmpInst::isSigned((ICmpInst::Predicate)pred))){
2289         Constant *CE1Op0 = CE1->getOperand(0);
2290         Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2291         if (CE1Inverse == CE1Op0) {
2292           // Check whether we can safely truncate the right hand side.
2293           Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2294           if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
2295                                     C2->getType()) == C2)
2296             return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2297         }
2298       }
2299     }
2300 
2301     if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2302         (C1->isNullValue() && !C2->isNullValue())) {
2303       // If C2 is a constant expr and C1 isn't, flip them around and fold the
2304       // other way if possible.
2305       // Also, if C1 is null and C2 isn't, flip them around.
2306       pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2307       return ConstantExpr::getICmp(pred, C2, C1);
2308     }
2309   }
2310   return nullptr;
2311 }
2312 
2313 /// Test whether the given sequence of *normalized* indices is "inbounds".
2314 template<typename IndexTy>
isInBoundsIndices(ArrayRef<IndexTy> Idxs)2315 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
2316   // No indices means nothing that could be out of bounds.
2317   if (Idxs.empty()) return true;
2318 
2319   // If the first index is zero, it's in bounds.
2320   if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2321 
2322   // If the first index is one and all the rest are zero, it's in bounds,
2323   // by the one-past-the-end rule.
2324   if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) {
2325     if (!CI->isOne())
2326       return false;
2327   } else {
2328     auto *CV = cast<ConstantDataVector>(Idxs[0]);
2329     CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
2330     if (!CI || !CI->isOne())
2331       return false;
2332   }
2333 
2334   for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
2335     if (!cast<Constant>(Idxs[i])->isNullValue())
2336       return false;
2337   return true;
2338 }
2339 
2340 /// Test whether a given ConstantInt is in-range for a SequentialType.
isIndexInRangeOfArrayType(uint64_t NumElements,const ConstantInt * CI)2341 static bool isIndexInRangeOfArrayType(uint64_t NumElements,
2342                                       const ConstantInt *CI) {
2343   // We cannot bounds check the index if it doesn't fit in an int64_t.
2344   if (CI->getValue().getMinSignedBits() > 64)
2345     return false;
2346 
2347   // A negative index or an index past the end of our sequential type is
2348   // considered out-of-range.
2349   int64_t IndexVal = CI->getSExtValue();
2350   if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
2351     return false;
2352 
2353   // Otherwise, it is in-range.
2354   return true;
2355 }
2356 
ConstantFoldGetElementPtr(Type * PointeeTy,Constant * C,bool InBounds,Optional<unsigned> InRangeIndex,ArrayRef<Value * > Idxs)2357 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C,
2358                                           bool InBounds,
2359                                           Optional<unsigned> InRangeIndex,
2360                                           ArrayRef<Value *> Idxs) {
2361   if (Idxs.empty()) return C;
2362 
2363   Type *GEPTy = GetElementPtrInst::getGEPReturnType(
2364       PointeeTy, C, makeArrayRef((Value *const *)Idxs.data(), Idxs.size()));
2365 
2366   if (isa<PoisonValue>(C))
2367     return PoisonValue::get(GEPTy);
2368 
2369   if (isa<UndefValue>(C))
2370     // If inbounds, we can choose an out-of-bounds pointer as a base pointer.
2371     return InBounds ? PoisonValue::get(GEPTy) : UndefValue::get(GEPTy);
2372 
2373   Constant *Idx0 = cast<Constant>(Idxs[0]);
2374   if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0)))
2375     return GEPTy->isVectorTy() && !C->getType()->isVectorTy()
2376                ? ConstantVector::getSplat(
2377                      cast<VectorType>(GEPTy)->getElementCount(), C)
2378                : C;
2379 
2380   if (C->isNullValue()) {
2381     bool isNull = true;
2382     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2383       if (!isa<UndefValue>(Idxs[i]) &&
2384           !cast<Constant>(Idxs[i])->isNullValue()) {
2385         isNull = false;
2386         break;
2387       }
2388     if (isNull) {
2389       PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType());
2390       Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs);
2391 
2392       assert(Ty && "Invalid indices for GEP!");
2393       Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2394       Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2395       if (VectorType *VT = dyn_cast<VectorType>(C->getType()))
2396         GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount());
2397 
2398       // The GEP returns a vector of pointers when one of more of
2399       // its arguments is a vector.
2400       for (unsigned i = 0, e = Idxs.size(); i != e; ++i) {
2401         if (auto *VT = dyn_cast<VectorType>(Idxs[i]->getType())) {
2402           assert((!isa<VectorType>(GEPTy) || isa<ScalableVectorType>(GEPTy) ==
2403                                                  isa<ScalableVectorType>(VT)) &&
2404                  "Mismatched GEPTy vector types");
2405           GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount());
2406           break;
2407         }
2408       }
2409 
2410       return Constant::getNullValue(GEPTy);
2411     }
2412   }
2413 
2414   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2415     // Combine Indices - If the source pointer to this getelementptr instruction
2416     // is a getelementptr instruction, combine the indices of the two
2417     // getelementptr instructions into a single instruction.
2418     //
2419     if (CE->getOpcode() == Instruction::GetElementPtr) {
2420       gep_type_iterator LastI = gep_type_end(CE);
2421       for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2422            I != E; ++I)
2423         LastI = I;
2424 
2425       // We cannot combine indices if doing so would take us outside of an
2426       // array or vector.  Doing otherwise could trick us if we evaluated such a
2427       // GEP as part of a load.
2428       //
2429       // e.g. Consider if the original GEP was:
2430       // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2431       //                    i32 0, i32 0, i64 0)
2432       //
2433       // If we then tried to offset it by '8' to get to the third element,
2434       // an i8, we should *not* get:
2435       // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2436       //                    i32 0, i32 0, i64 8)
2437       //
2438       // This GEP tries to index array element '8  which runs out-of-bounds.
2439       // Subsequent evaluation would get confused and produce erroneous results.
2440       //
2441       // The following prohibits such a GEP from being formed by checking to see
2442       // if the index is in-range with respect to an array.
2443       // TODO: This code may be extended to handle vectors as well.
2444       bool PerformFold = false;
2445       if (Idx0->isNullValue())
2446         PerformFold = true;
2447       else if (LastI.isSequential())
2448         if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0))
2449           PerformFold = (!LastI.isBoundedSequential() ||
2450                          isIndexInRangeOfArrayType(
2451                              LastI.getSequentialNumElements(), CI)) &&
2452                         !CE->getOperand(CE->getNumOperands() - 1)
2453                              ->getType()
2454                              ->isVectorTy();
2455 
2456       if (PerformFold) {
2457         SmallVector<Value*, 16> NewIndices;
2458         NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2459         NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1);
2460 
2461         // Add the last index of the source with the first index of the new GEP.
2462         // Make sure to handle the case when they are actually different types.
2463         Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2464         // Otherwise it must be an array.
2465         if (!Idx0->isNullValue()) {
2466           Type *IdxTy = Combined->getType();
2467           if (IdxTy != Idx0->getType()) {
2468             unsigned CommonExtendedWidth =
2469                 std::max(IdxTy->getIntegerBitWidth(),
2470                          Idx0->getType()->getIntegerBitWidth());
2471             CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2472 
2473             Type *CommonTy =
2474                 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth);
2475             Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy);
2476             Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy);
2477             Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2478           } else {
2479             Combined =
2480               ConstantExpr::get(Instruction::Add, Idx0, Combined);
2481           }
2482         }
2483 
2484         NewIndices.push_back(Combined);
2485         NewIndices.append(Idxs.begin() + 1, Idxs.end());
2486 
2487         // The combined GEP normally inherits its index inrange attribute from
2488         // the inner GEP, but if the inner GEP's last index was adjusted by the
2489         // outer GEP, any inbounds attribute on that index is invalidated.
2490         Optional<unsigned> IRIndex = cast<GEPOperator>(CE)->getInRangeIndex();
2491         if (IRIndex && *IRIndex == CE->getNumOperands() - 2 && !Idx0->isNullValue())
2492           IRIndex = None;
2493 
2494         return ConstantExpr::getGetElementPtr(
2495             cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0),
2496             NewIndices, InBounds && cast<GEPOperator>(CE)->isInBounds(),
2497             IRIndex);
2498       }
2499     }
2500 
2501     // Attempt to fold casts to the same type away.  For example, folding:
2502     //
2503     //   i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2504     //                       i64 0, i64 0)
2505     // into:
2506     //
2507     //   i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2508     //
2509     // Don't fold if the cast is changing address spaces.
2510     if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2511       PointerType *SrcPtrTy =
2512         dyn_cast<PointerType>(CE->getOperand(0)->getType());
2513       PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2514       if (SrcPtrTy && DstPtrTy) {
2515         ArrayType *SrcArrayTy =
2516           dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2517         ArrayType *DstArrayTy =
2518           dyn_cast<ArrayType>(DstPtrTy->getElementType());
2519         if (SrcArrayTy && DstArrayTy
2520             && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2521             && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2522           return ConstantExpr::getGetElementPtr(SrcArrayTy,
2523                                                 (Constant *)CE->getOperand(0),
2524                                                 Idxs, InBounds, InRangeIndex);
2525       }
2526     }
2527   }
2528 
2529   // Check to see if any array indices are not within the corresponding
2530   // notional array or vector bounds. If so, try to determine if they can be
2531   // factored out into preceding dimensions.
2532   SmallVector<Constant *, 8> NewIdxs;
2533   Type *Ty = PointeeTy;
2534   Type *Prev = C->getType();
2535   auto GEPIter = gep_type_begin(PointeeTy, Idxs);
2536   bool Unknown =
2537       !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]);
2538   for (unsigned i = 1, e = Idxs.size(); i != e;
2539        Prev = Ty, Ty = (++GEPIter).getIndexedType(), ++i) {
2540     if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) {
2541       // We don't know if it's in range or not.
2542       Unknown = true;
2543       continue;
2544     }
2545     if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1]))
2546       // Skip if the type of the previous index is not supported.
2547       continue;
2548     if (InRangeIndex && i == *InRangeIndex + 1) {
2549       // If an index is marked inrange, we cannot apply this canonicalization to
2550       // the following index, as that will cause the inrange index to point to
2551       // the wrong element.
2552       continue;
2553     }
2554     if (isa<StructType>(Ty)) {
2555       // The verify makes sure that GEPs into a struct are in range.
2556       continue;
2557     }
2558     if (isa<VectorType>(Ty)) {
2559       // There can be awkward padding in after a non-power of two vector.
2560       Unknown = true;
2561       continue;
2562     }
2563     auto *STy = cast<ArrayType>(Ty);
2564     if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2565       if (isIndexInRangeOfArrayType(STy->getNumElements(), CI))
2566         // It's in range, skip to the next index.
2567         continue;
2568       if (CI->getSExtValue() < 0) {
2569         // It's out of range and negative, don't try to factor it.
2570         Unknown = true;
2571         continue;
2572       }
2573     } else {
2574       auto *CV = cast<ConstantDataVector>(Idxs[i]);
2575       bool InRange = true;
2576       for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
2577         auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I));
2578         InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI);
2579         if (CI->getSExtValue() < 0) {
2580           Unknown = true;
2581           break;
2582         }
2583       }
2584       if (InRange || Unknown)
2585         // It's in range, skip to the next index.
2586         // It's out of range and negative, don't try to factor it.
2587         continue;
2588     }
2589     if (isa<StructType>(Prev)) {
2590       // It's out of range, but the prior dimension is a struct
2591       // so we can't do anything about it.
2592       Unknown = true;
2593       continue;
2594     }
2595     // It's out of range, but we can factor it into the prior
2596     // dimension.
2597     NewIdxs.resize(Idxs.size());
2598     // Determine the number of elements in our sequential type.
2599     uint64_t NumElements = STy->getArrayNumElements();
2600 
2601     // Expand the current index or the previous index to a vector from a scalar
2602     // if necessary.
2603     Constant *CurrIdx = cast<Constant>(Idxs[i]);
2604     auto *PrevIdx =
2605         NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]);
2606     bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy();
2607     bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy();
2608     bool UseVector = IsCurrIdxVector || IsPrevIdxVector;
2609 
2610     if (!IsCurrIdxVector && IsPrevIdxVector)
2611       CurrIdx = ConstantDataVector::getSplat(
2612           cast<FixedVectorType>(PrevIdx->getType())->getNumElements(), CurrIdx);
2613 
2614     if (!IsPrevIdxVector && IsCurrIdxVector)
2615       PrevIdx = ConstantDataVector::getSplat(
2616           cast<FixedVectorType>(CurrIdx->getType())->getNumElements(), PrevIdx);
2617 
2618     Constant *Factor =
2619         ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements);
2620     if (UseVector)
2621       Factor = ConstantDataVector::getSplat(
2622           IsPrevIdxVector
2623               ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements()
2624               : cast<FixedVectorType>(CurrIdx->getType())->getNumElements(),
2625           Factor);
2626 
2627     NewIdxs[i] = ConstantExpr::getSRem(CurrIdx, Factor);
2628 
2629     Constant *Div = ConstantExpr::getSDiv(CurrIdx, Factor);
2630 
2631     unsigned CommonExtendedWidth =
2632         std::max(PrevIdx->getType()->getScalarSizeInBits(),
2633                  Div->getType()->getScalarSizeInBits());
2634     CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2635 
2636     // Before adding, extend both operands to i64 to avoid
2637     // overflow trouble.
2638     Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth);
2639     if (UseVector)
2640       ExtendedTy = FixedVectorType::get(
2641           ExtendedTy,
2642           IsPrevIdxVector
2643               ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements()
2644               : cast<FixedVectorType>(CurrIdx->getType())->getNumElements());
2645 
2646     if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2647       PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy);
2648 
2649     if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2650       Div = ConstantExpr::getSExt(Div, ExtendedTy);
2651 
2652     NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div);
2653   }
2654 
2655   // If we did any factoring, start over with the adjusted indices.
2656   if (!NewIdxs.empty()) {
2657     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2658       if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2659     return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds,
2660                                           InRangeIndex);
2661   }
2662 
2663   // If all indices are known integers and normalized, we can do a simple
2664   // check for the "inbounds" property.
2665   if (!Unknown && !InBounds)
2666     if (auto *GV = dyn_cast<GlobalVariable>(C))
2667       if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs))
2668         return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs,
2669                                               /*InBounds=*/true, InRangeIndex);
2670 
2671   return nullptr;
2672 }
2673