1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements folding of constants for LLVM.  This implements the
11 // (internal) ConstantFold.h interface, which is used by the
12 // ConstantExpr::get* methods to automatically fold constants when possible.
13 //
14 // The current constant folding implementation is implemented in two pieces: the
15 // pieces that don't need DataLayout, and the pieces that do. This is to avoid
16 // a dependence in IR on Target.
17 //
18 //===----------------------------------------------------------------------===//
19 
20 #include "ConstantFold.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/Operator.h"
30 #include "llvm/IR/PatternMatch.h"
31 #include "llvm/Support/Compiler.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/ManagedStatic.h"
34 #include "llvm/Support/MathExtras.h"
35 #include <limits>
36 using namespace llvm;
37 using namespace llvm::PatternMatch;
38 
39 //===----------------------------------------------------------------------===//
40 //                ConstantFold*Instruction Implementations
41 //===----------------------------------------------------------------------===//
42 
43 /// BitCastConstantVector - Convert the specified vector Constant node to the
44 /// specified vector type.  At this point, we know that the elements of the
45 /// input vector constant are all simple integer or FP values.
BitCastConstantVector(Constant * CV,VectorType * DstTy)46 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) {
47 
48   if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
49   if (CV->isNullValue()) return Constant::getNullValue(DstTy);
50 
51   // If this cast changes element count then we can't handle it here:
52   // doing so requires endianness information.  This should be handled by
53   // Analysis/ConstantFolding.cpp
54   unsigned NumElts = DstTy->getNumElements();
55   if (NumElts != CV->getType()->getVectorNumElements())
56     return nullptr;
57 
58   Type *DstEltTy = DstTy->getElementType();
59 
60   SmallVector<Constant*, 16> Result;
61   Type *Ty = IntegerType::get(CV->getContext(), 32);
62   for (unsigned i = 0; i != NumElts; ++i) {
63     Constant *C =
64       ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i));
65     C = ConstantExpr::getBitCast(C, DstEltTy);
66     Result.push_back(C);
67   }
68 
69   return ConstantVector::get(Result);
70 }
71 
72 /// This function determines which opcode to use to fold two constant cast
73 /// expressions together. It uses CastInst::isEliminableCastPair to determine
74 /// the opcode. Consequently its just a wrapper around that function.
75 /// @brief Determine if it is valid to fold a cast of a cast
76 static unsigned
foldConstantCastPair(unsigned opc,ConstantExpr * Op,Type * DstTy)77 foldConstantCastPair(
78   unsigned opc,          ///< opcode of the second cast constant expression
79   ConstantExpr *Op,      ///< the first cast constant expression
80   Type *DstTy            ///< destination type of the first cast
81 ) {
82   assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
83   assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
84   assert(CastInst::isCast(opc) && "Invalid cast opcode");
85 
86   // The types and opcodes for the two Cast constant expressions
87   Type *SrcTy = Op->getOperand(0)->getType();
88   Type *MidTy = Op->getType();
89   Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
90   Instruction::CastOps secondOp = Instruction::CastOps(opc);
91 
92   // Assume that pointers are never more than 64 bits wide, and only use this
93   // for the middle type. Otherwise we could end up folding away illegal
94   // bitcasts between address spaces with different sizes.
95   IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext());
96 
97   // Let CastInst::isEliminableCastPair do the heavy lifting.
98   return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
99                                         nullptr, FakeIntPtrTy, nullptr);
100 }
101 
FoldBitCast(Constant * V,Type * DestTy)102 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
103   Type *SrcTy = V->getType();
104   if (SrcTy == DestTy)
105     return V; // no-op cast
106 
107   // Check to see if we are casting a pointer to an aggregate to a pointer to
108   // the first element.  If so, return the appropriate GEP instruction.
109   if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
110     if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
111       if (PTy->getAddressSpace() == DPTy->getAddressSpace()
112           && PTy->getElementType()->isSized()) {
113         SmallVector<Value*, 8> IdxList;
114         Value *Zero =
115           Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
116         IdxList.push_back(Zero);
117         Type *ElTy = PTy->getElementType();
118         while (ElTy != DPTy->getElementType()) {
119           if (StructType *STy = dyn_cast<StructType>(ElTy)) {
120             if (STy->getNumElements() == 0) break;
121             ElTy = STy->getElementType(0);
122             IdxList.push_back(Zero);
123           } else if (SequentialType *STy =
124                      dyn_cast<SequentialType>(ElTy)) {
125             if (ElTy->isPointerTy()) break;  // Can't index into pointers!
126             ElTy = STy->getElementType();
127             IdxList.push_back(Zero);
128           } else {
129             break;
130           }
131         }
132 
133         if (ElTy == DPTy->getElementType())
134           // This GEP is inbounds because all indices are zero.
135           return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(),
136                                                         V, IdxList);
137       }
138 
139   // Handle casts from one vector constant to another.  We know that the src
140   // and dest type have the same size (otherwise its an illegal cast).
141   if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
142     if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
143       assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
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     return ConstantInt::get(FP->getContext(),
195                             FP->getValueAPF().bitcastToAPInt());
196   }
197 
198   return nullptr;
199 }
200 
201 
202 /// ExtractConstantBytes - V is an integer constant which only has a subset of
203 /// its bytes used.  The bytes used are indicated by ByteStart (which is the
204 /// first byte used, counting from the least significant byte) and ByteSize,
205 /// which is the number of bytes used.
206 ///
207 /// This function analyzes the specified constant to see if the specified byte
208 /// range can be returned as a simplified constant.  If so, the constant is
209 /// returned, otherwise null is returned.
210 ///
ExtractConstantBytes(Constant * C,unsigned ByteStart,unsigned ByteSize)211 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
212                                       unsigned ByteSize) {
213   assert(C->getType()->isIntegerTy() &&
214          (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
215          "Non-byte sized integer input");
216   unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
217   assert(ByteSize && "Must be accessing some piece");
218   assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
219   assert(ByteSize != CSize && "Should not extract everything");
220 
221   // Constant Integers are simple.
222   if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
223     APInt V = CI->getValue();
224     if (ByteStart)
225       V = V.lshr(ByteStart*8);
226     V = V.trunc(ByteSize*8);
227     return ConstantInt::get(CI->getContext(), V);
228   }
229 
230   // In the input is a constant expr, we might be able to recursively simplify.
231   // If not, we definitely can't do anything.
232   ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
233   if (!CE) return nullptr;
234 
235   switch (CE->getOpcode()) {
236   default: return nullptr;
237   case Instruction::Or: {
238     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
239     if (!RHS)
240       return nullptr;
241 
242     // X | -1 -> -1.
243     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
244       if (RHSC->isAllOnesValue())
245         return RHSC;
246 
247     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
248     if (!LHS)
249       return nullptr;
250     return ConstantExpr::getOr(LHS, RHS);
251   }
252   case Instruction::And: {
253     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
254     if (!RHS)
255       return nullptr;
256 
257     // X & 0 -> 0.
258     if (RHS->isNullValue())
259       return RHS;
260 
261     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
262     if (!LHS)
263       return nullptr;
264     return ConstantExpr::getAnd(LHS, RHS);
265   }
266   case Instruction::LShr: {
267     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
268     if (!Amt)
269       return nullptr;
270     unsigned ShAmt = Amt->getZExtValue();
271     // Cannot analyze non-byte shifts.
272     if ((ShAmt & 7) != 0)
273       return nullptr;
274     ShAmt >>= 3;
275 
276     // If the extract is known to be all zeros, return zero.
277     if (ByteStart >= CSize-ShAmt)
278       return Constant::getNullValue(IntegerType::get(CE->getContext(),
279                                                      ByteSize*8));
280     // If the extract is known to be fully in the input, extract it.
281     if (ByteStart+ByteSize+ShAmt <= CSize)
282       return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
283 
284     // TODO: Handle the 'partially zero' case.
285     return nullptr;
286   }
287 
288   case Instruction::Shl: {
289     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
290     if (!Amt)
291       return nullptr;
292     unsigned ShAmt = Amt->getZExtValue();
293     // Cannot analyze non-byte shifts.
294     if ((ShAmt & 7) != 0)
295       return nullptr;
296     ShAmt >>= 3;
297 
298     // If the extract is known to be all zeros, return zero.
299     if (ByteStart+ByteSize <= ShAmt)
300       return Constant::getNullValue(IntegerType::get(CE->getContext(),
301                                                      ByteSize*8));
302     // If the extract is known to be fully in the input, extract it.
303     if (ByteStart >= ShAmt)
304       return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
305 
306     // TODO: Handle the 'partially zero' case.
307     return nullptr;
308   }
309 
310   case Instruction::ZExt: {
311     unsigned SrcBitSize =
312       cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
313 
314     // If extracting something that is completely zero, return 0.
315     if (ByteStart*8 >= SrcBitSize)
316       return Constant::getNullValue(IntegerType::get(CE->getContext(),
317                                                      ByteSize*8));
318 
319     // If exactly extracting the input, return it.
320     if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
321       return CE->getOperand(0);
322 
323     // If extracting something completely in the input, if if the input is a
324     // multiple of 8 bits, recurse.
325     if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
326       return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
327 
328     // Otherwise, if extracting a subset of the input, which is not multiple of
329     // 8 bits, do a shift and trunc to get the bits.
330     if ((ByteStart+ByteSize)*8 < SrcBitSize) {
331       assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
332       Constant *Res = CE->getOperand(0);
333       if (ByteStart)
334         Res = ConstantExpr::getLShr(Res,
335                                  ConstantInt::get(Res->getType(), ByteStart*8));
336       return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
337                                                           ByteSize*8));
338     }
339 
340     // TODO: Handle the 'partially zero' case.
341     return nullptr;
342   }
343   }
344 }
345 
346 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
347 /// on Ty, with any known factors factored out. If Folded is false,
348 /// return null if no factoring was possible, to avoid endlessly
349 /// bouncing an unfoldable expression back into the top-level folder.
350 ///
getFoldedSizeOf(Type * Ty,Type * DestTy,bool Folded)351 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy,
352                                  bool Folded) {
353   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
354     Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
355     Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
356     return ConstantExpr::getNUWMul(E, N);
357   }
358 
359   if (StructType *STy = dyn_cast<StructType>(Ty))
360     if (!STy->isPacked()) {
361       unsigned NumElems = STy->getNumElements();
362       // An empty struct has size zero.
363       if (NumElems == 0)
364         return ConstantExpr::getNullValue(DestTy);
365       // Check for a struct with all members having the same size.
366       Constant *MemberSize =
367         getFoldedSizeOf(STy->getElementType(0), DestTy, true);
368       bool AllSame = true;
369       for (unsigned i = 1; i != NumElems; ++i)
370         if (MemberSize !=
371             getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
372           AllSame = false;
373           break;
374         }
375       if (AllSame) {
376         Constant *N = ConstantInt::get(DestTy, NumElems);
377         return ConstantExpr::getNUWMul(MemberSize, N);
378       }
379     }
380 
381   // Pointer size doesn't depend on the pointee type, so canonicalize them
382   // to an arbitrary pointee.
383   if (PointerType *PTy = dyn_cast<PointerType>(Ty))
384     if (!PTy->getElementType()->isIntegerTy(1))
385       return
386         getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
387                                          PTy->getAddressSpace()),
388                         DestTy, true);
389 
390   // If there's no interesting folding happening, bail so that we don't create
391   // a constant that looks like it needs folding but really doesn't.
392   if (!Folded)
393     return nullptr;
394 
395   // Base case: Get a regular sizeof expression.
396   Constant *C = ConstantExpr::getSizeOf(Ty);
397   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
398                                                     DestTy, false),
399                             C, DestTy);
400   return C;
401 }
402 
403 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
404 /// on Ty, with any known factors factored out. If Folded is false,
405 /// return null if no factoring was possible, to avoid endlessly
406 /// bouncing an unfoldable expression back into the top-level folder.
407 ///
getFoldedAlignOf(Type * Ty,Type * DestTy,bool Folded)408 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy,
409                                   bool Folded) {
410   // The alignment of an array is equal to the alignment of the
411   // array element. Note that this is not always true for vectors.
412   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
413     Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
414     C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
415                                                       DestTy,
416                                                       false),
417                               C, DestTy);
418     return C;
419   }
420 
421   if (StructType *STy = dyn_cast<StructType>(Ty)) {
422     // Packed structs always have an alignment of 1.
423     if (STy->isPacked())
424       return ConstantInt::get(DestTy, 1);
425 
426     // Otherwise, struct alignment is the maximum alignment of any member.
427     // Without target data, we can't compare much, but we can check to see
428     // if all the members have the same alignment.
429     unsigned NumElems = STy->getNumElements();
430     // An empty struct has minimal alignment.
431     if (NumElems == 0)
432       return ConstantInt::get(DestTy, 1);
433     // Check for a struct with all members having the same alignment.
434     Constant *MemberAlign =
435       getFoldedAlignOf(STy->getElementType(0), DestTy, true);
436     bool AllSame = true;
437     for (unsigned i = 1; i != NumElems; ++i)
438       if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
439         AllSame = false;
440         break;
441       }
442     if (AllSame)
443       return MemberAlign;
444   }
445 
446   // Pointer alignment doesn't depend on the pointee type, so canonicalize them
447   // to an arbitrary pointee.
448   if (PointerType *PTy = dyn_cast<PointerType>(Ty))
449     if (!PTy->getElementType()->isIntegerTy(1))
450       return
451         getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
452                                                            1),
453                                           PTy->getAddressSpace()),
454                          DestTy, true);
455 
456   // If there's no interesting folding happening, bail so that we don't create
457   // a constant that looks like it needs folding but really doesn't.
458   if (!Folded)
459     return nullptr;
460 
461   // Base case: Get a regular alignof expression.
462   Constant *C = ConstantExpr::getAlignOf(Ty);
463   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
464                                                     DestTy, false),
465                             C, DestTy);
466   return C;
467 }
468 
469 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
470 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
471 /// return null if no factoring was possible, to avoid endlessly
472 /// bouncing an unfoldable expression back into the top-level folder.
473 ///
getFoldedOffsetOf(Type * Ty,Constant * FieldNo,Type * DestTy,bool Folded)474 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo,
475                                    Type *DestTy,
476                                    bool Folded) {
477   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
478     Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
479                                                                 DestTy, false),
480                                         FieldNo, DestTy);
481     Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
482     return ConstantExpr::getNUWMul(E, N);
483   }
484 
485   if (StructType *STy = dyn_cast<StructType>(Ty))
486     if (!STy->isPacked()) {
487       unsigned NumElems = STy->getNumElements();
488       // An empty struct has no members.
489       if (NumElems == 0)
490         return nullptr;
491       // Check for a struct with all members having the same size.
492       Constant *MemberSize =
493         getFoldedSizeOf(STy->getElementType(0), DestTy, true);
494       bool AllSame = true;
495       for (unsigned i = 1; i != NumElems; ++i)
496         if (MemberSize !=
497             getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
498           AllSame = false;
499           break;
500         }
501       if (AllSame) {
502         Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
503                                                                     false,
504                                                                     DestTy,
505                                                                     false),
506                                             FieldNo, DestTy);
507         return ConstantExpr::getNUWMul(MemberSize, N);
508       }
509     }
510 
511   // If there's no interesting folding happening, bail so that we don't create
512   // a constant that looks like it needs folding but really doesn't.
513   if (!Folded)
514     return nullptr;
515 
516   // Base case: Get a regular offsetof expression.
517   Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
518   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
519                                                     DestTy, false),
520                             C, DestTy);
521   return C;
522 }
523 
ConstantFoldCastInstruction(unsigned opc,Constant * V,Type * DestTy)524 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
525                                             Type *DestTy) {
526   if (isa<UndefValue>(V)) {
527     // zext(undef) = 0, because the top bits will be zero.
528     // sext(undef) = 0, because the top bits will all be the same.
529     // [us]itofp(undef) = 0, because the result value is bounded.
530     if (opc == Instruction::ZExt || opc == Instruction::SExt ||
531         opc == Instruction::UIToFP || opc == Instruction::SIToFP)
532       return Constant::getNullValue(DestTy);
533     return UndefValue::get(DestTy);
534   }
535 
536   if (V->isNullValue() && !DestTy->isX86_MMXTy())
537     return Constant::getNullValue(DestTy);
538 
539   // If the cast operand is a constant expression, there's a few things we can
540   // do to try to simplify it.
541   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
542     if (CE->isCast()) {
543       // Try hard to fold cast of cast because they are often eliminable.
544       if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
545         return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
546     } else if (CE->getOpcode() == Instruction::GetElementPtr &&
547                // Do not fold addrspacecast (gep 0, .., 0). It might make the
548                // addrspacecast uncanonicalized.
549                opc != Instruction::AddrSpaceCast) {
550       // If all of the indexes in the GEP are null values, there is no pointer
551       // adjustment going on.  We might as well cast the source pointer.
552       bool isAllNull = true;
553       for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
554         if (!CE->getOperand(i)->isNullValue()) {
555           isAllNull = false;
556           break;
557         }
558       if (isAllNull)
559         // This is casting one pointer type to another, always BitCast
560         return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
561     }
562   }
563 
564   // If the cast operand is a constant vector, perform the cast by
565   // operating on each element. In the cast of bitcasts, the element
566   // count may be mismatched; don't attempt to handle that here.
567   if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
568       DestTy->isVectorTy() &&
569       DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) {
570     SmallVector<Constant*, 16> res;
571     VectorType *DestVecTy = cast<VectorType>(DestTy);
572     Type *DstEltTy = DestVecTy->getElementType();
573     Type *Ty = IntegerType::get(V->getContext(), 32);
574     for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) {
575       Constant *C =
576         ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
577       res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
578     }
579     return ConstantVector::get(res);
580   }
581 
582   // We actually have to do a cast now. Perform the cast according to the
583   // opcode specified.
584   switch (opc) {
585   default:
586     llvm_unreachable("Failed to cast constant expression");
587   case Instruction::FPTrunc:
588   case Instruction::FPExt:
589     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
590       bool ignored;
591       APFloat Val = FPC->getValueAPF();
592       Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf :
593                   DestTy->isFloatTy() ? APFloat::IEEEsingle :
594                   DestTy->isDoubleTy() ? APFloat::IEEEdouble :
595                   DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
596                   DestTy->isFP128Ty() ? APFloat::IEEEquad :
597                   DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble :
598                   APFloat::Bogus,
599                   APFloat::rmNearestTiesToEven, &ignored);
600       return ConstantFP::get(V->getContext(), Val);
601     }
602     return nullptr; // Can't fold.
603   case Instruction::FPToUI:
604   case Instruction::FPToSI:
605     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
606       const APFloat &V = FPC->getValueAPF();
607       bool ignored;
608       uint64_t x[2];
609       uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
610       if (APFloat::opInvalidOp ==
611           V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
612                              APFloat::rmTowardZero, &ignored)) {
613         // Undefined behavior invoked - the destination type can't represent
614         // the input constant.
615         return UndefValue::get(DestTy);
616       }
617       APInt Val(DestBitWidth, x);
618       return ConstantInt::get(FPC->getContext(), Val);
619     }
620     return nullptr; // Can't fold.
621   case Instruction::IntToPtr:   //always treated as unsigned
622     if (V->isNullValue())       // Is it an integral null value?
623       return ConstantPointerNull::get(cast<PointerType>(DestTy));
624     return nullptr;                   // Other pointer types cannot be casted
625   case Instruction::PtrToInt:   // always treated as unsigned
626     // Is it a null pointer value?
627     if (V->isNullValue())
628       return ConstantInt::get(DestTy, 0);
629     // If this is a sizeof-like expression, pull out multiplications by
630     // known factors to expose them to subsequent folding. If it's an
631     // alignof-like expression, factor out known factors.
632     if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
633       if (CE->getOpcode() == Instruction::GetElementPtr &&
634           CE->getOperand(0)->isNullValue()) {
635         GEPOperator *GEPO = cast<GEPOperator>(CE);
636         Type *Ty = GEPO->getSourceElementType();
637         if (CE->getNumOperands() == 2) {
638           // Handle a sizeof-like expression.
639           Constant *Idx = CE->getOperand(1);
640           bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
641           if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
642             Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
643                                                                 DestTy, false),
644                                         Idx, DestTy);
645             return ConstantExpr::getMul(C, Idx);
646           }
647         } else if (CE->getNumOperands() == 3 &&
648                    CE->getOperand(1)->isNullValue()) {
649           // Handle an alignof-like expression.
650           if (StructType *STy = dyn_cast<StructType>(Ty))
651             if (!STy->isPacked()) {
652               ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
653               if (CI->isOne() &&
654                   STy->getNumElements() == 2 &&
655                   STy->getElementType(0)->isIntegerTy(1)) {
656                 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
657               }
658             }
659           // Handle an offsetof-like expression.
660           if (Ty->isStructTy() || Ty->isArrayTy()) {
661             if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
662                                                 DestTy, false))
663               return C;
664           }
665         }
666       }
667     // Other pointer types cannot be casted
668     return nullptr;
669   case Instruction::UIToFP:
670   case Instruction::SIToFP:
671     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
672       APInt api = CI->getValue();
673       APFloat apf(DestTy->getFltSemantics(),
674                   APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
675       if (APFloat::opOverflow &
676           apf.convertFromAPInt(api, opc==Instruction::SIToFP,
677                               APFloat::rmNearestTiesToEven)) {
678         // Undefined behavior invoked - the destination type can't represent
679         // the input constant.
680         return UndefValue::get(DestTy);
681       }
682       return ConstantFP::get(V->getContext(), apf);
683     }
684     return nullptr;
685   case Instruction::ZExt:
686     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
687       uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
688       return ConstantInt::get(V->getContext(),
689                               CI->getValue().zext(BitWidth));
690     }
691     return nullptr;
692   case Instruction::SExt:
693     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
694       uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
695       return ConstantInt::get(V->getContext(),
696                               CI->getValue().sext(BitWidth));
697     }
698     return nullptr;
699   case Instruction::Trunc: {
700     if (V->getType()->isVectorTy())
701       return nullptr;
702 
703     uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
704     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
705       return ConstantInt::get(V->getContext(),
706                               CI->getValue().trunc(DestBitWidth));
707     }
708 
709     // The input must be a constantexpr.  See if we can simplify this based on
710     // the bytes we are demanding.  Only do this if the source and dest are an
711     // even multiple of a byte.
712     if ((DestBitWidth & 7) == 0 &&
713         (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
714       if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
715         return Res;
716 
717     return nullptr;
718   }
719   case Instruction::BitCast:
720     return FoldBitCast(V, DestTy);
721   case Instruction::AddrSpaceCast:
722     return nullptr;
723   }
724 }
725 
ConstantFoldSelectInstruction(Constant * Cond,Constant * V1,Constant * V2)726 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
727                                               Constant *V1, Constant *V2) {
728   // Check for i1 and vector true/false conditions.
729   if (Cond->isNullValue()) return V2;
730   if (Cond->isAllOnesValue()) return V1;
731 
732   // If the condition is a vector constant, fold the result elementwise.
733   if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
734     SmallVector<Constant*, 16> Result;
735     Type *Ty = IntegerType::get(CondV->getContext(), 32);
736     for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){
737       Constant *V;
738       Constant *V1Element = ConstantExpr::getExtractElement(V1,
739                                                     ConstantInt::get(Ty, i));
740       Constant *V2Element = ConstantExpr::getExtractElement(V2,
741                                                     ConstantInt::get(Ty, i));
742       Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i));
743       if (V1Element == V2Element) {
744         V = V1Element;
745       } else if (isa<UndefValue>(Cond)) {
746         V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
747       } else {
748         if (!isa<ConstantInt>(Cond)) break;
749         V = Cond->isNullValue() ? V2Element : V1Element;
750       }
751       Result.push_back(V);
752     }
753 
754     // If we were able to build the vector, return it.
755     if (Result.size() == V1->getType()->getVectorNumElements())
756       return ConstantVector::get(Result);
757   }
758 
759   if (isa<UndefValue>(Cond)) {
760     if (isa<UndefValue>(V1)) return V1;
761     return V2;
762   }
763   if (isa<UndefValue>(V1)) return V2;
764   if (isa<UndefValue>(V2)) return V1;
765   if (V1 == V2) return V1;
766 
767   if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
768     if (TrueVal->getOpcode() == Instruction::Select)
769       if (TrueVal->getOperand(0) == Cond)
770         return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
771   }
772   if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
773     if (FalseVal->getOpcode() == Instruction::Select)
774       if (FalseVal->getOperand(0) == Cond)
775         return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
776   }
777 
778   return nullptr;
779 }
780 
ConstantFoldExtractElementInstruction(Constant * Val,Constant * Idx)781 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
782                                                       Constant *Idx) {
783   if (isa<UndefValue>(Val))  // ee(undef, x) -> undef
784     return UndefValue::get(Val->getType()->getVectorElementType());
785   if (Val->isNullValue())  // ee(zero, x) -> zero
786     return Constant::getNullValue(Val->getType()->getVectorElementType());
787   // ee({w,x,y,z}, undef) -> undef
788   if (isa<UndefValue>(Idx))
789     return UndefValue::get(Val->getType()->getVectorElementType());
790 
791   if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
792     // ee({w,x,y,z}, wrong_value) -> undef
793     if (CIdx->uge(Val->getType()->getVectorNumElements()))
794       return UndefValue::get(Val->getType()->getVectorElementType());
795     return Val->getAggregateElement(CIdx->getZExtValue());
796   }
797   return nullptr;
798 }
799 
ConstantFoldInsertElementInstruction(Constant * Val,Constant * Elt,Constant * Idx)800 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
801                                                      Constant *Elt,
802                                                      Constant *Idx) {
803   if (isa<UndefValue>(Idx))
804     return UndefValue::get(Val->getType());
805 
806   ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
807   if (!CIdx) return nullptr;
808 
809   unsigned NumElts = Val->getType()->getVectorNumElements();
810   if (CIdx->uge(NumElts))
811     return UndefValue::get(Val->getType());
812 
813   SmallVector<Constant*, 16> Result;
814   Result.reserve(NumElts);
815   auto *Ty = Type::getInt32Ty(Val->getContext());
816   uint64_t IdxVal = CIdx->getZExtValue();
817   for (unsigned i = 0; i != NumElts; ++i) {
818     if (i == IdxVal) {
819       Result.push_back(Elt);
820       continue;
821     }
822 
823     Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
824     Result.push_back(C);
825   }
826 
827   return ConstantVector::get(Result);
828 }
829 
ConstantFoldShuffleVectorInstruction(Constant * V1,Constant * V2,Constant * Mask)830 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
831                                                      Constant *V2,
832                                                      Constant *Mask) {
833   unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
834   Type *EltTy = V1->getType()->getVectorElementType();
835 
836   // Undefined shuffle mask -> undefined value.
837   if (isa<UndefValue>(Mask))
838     return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
839 
840   // Don't break the bitcode reader hack.
841   if (isa<ConstantExpr>(Mask)) return nullptr;
842 
843   unsigned SrcNumElts = V1->getType()->getVectorNumElements();
844 
845   // Loop over the shuffle mask, evaluating each element.
846   SmallVector<Constant*, 32> Result;
847   for (unsigned i = 0; i != MaskNumElts; ++i) {
848     int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
849     if (Elt == -1) {
850       Result.push_back(UndefValue::get(EltTy));
851       continue;
852     }
853     Constant *InElt;
854     if (unsigned(Elt) >= SrcNumElts*2)
855       InElt = UndefValue::get(EltTy);
856     else if (unsigned(Elt) >= SrcNumElts) {
857       Type *Ty = IntegerType::get(V2->getContext(), 32);
858       InElt =
859         ConstantExpr::getExtractElement(V2,
860                                         ConstantInt::get(Ty, Elt - SrcNumElts));
861     } else {
862       Type *Ty = IntegerType::get(V1->getContext(), 32);
863       InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
864     }
865     Result.push_back(InElt);
866   }
867 
868   return ConstantVector::get(Result);
869 }
870 
ConstantFoldExtractValueInstruction(Constant * Agg,ArrayRef<unsigned> Idxs)871 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
872                                                     ArrayRef<unsigned> Idxs) {
873   // Base case: no indices, so return the entire value.
874   if (Idxs.empty())
875     return Agg;
876 
877   if (Constant *C = Agg->getAggregateElement(Idxs[0]))
878     return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
879 
880   return nullptr;
881 }
882 
ConstantFoldInsertValueInstruction(Constant * Agg,Constant * Val,ArrayRef<unsigned> Idxs)883 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
884                                                    Constant *Val,
885                                                    ArrayRef<unsigned> Idxs) {
886   // Base case: no indices, so replace the entire value.
887   if (Idxs.empty())
888     return Val;
889 
890   unsigned NumElts;
891   if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
892     NumElts = ST->getNumElements();
893   else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
894     NumElts = AT->getNumElements();
895   else
896     NumElts = Agg->getType()->getVectorNumElements();
897 
898   SmallVector<Constant*, 32> Result;
899   for (unsigned i = 0; i != NumElts; ++i) {
900     Constant *C = Agg->getAggregateElement(i);
901     if (!C) return nullptr;
902 
903     if (Idxs[0] == i)
904       C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
905 
906     Result.push_back(C);
907   }
908 
909   if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
910     return ConstantStruct::get(ST, Result);
911   if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
912     return ConstantArray::get(AT, Result);
913   return ConstantVector::get(Result);
914 }
915 
916 
ConstantFoldBinaryInstruction(unsigned Opcode,Constant * C1,Constant * C2)917 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
918                                               Constant *C1, Constant *C2) {
919   // Handle UndefValue up front.
920   if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
921     switch (Opcode) {
922     case Instruction::Xor:
923       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
924         // Handle undef ^ undef -> 0 special case. This is a common
925         // idiom (misuse).
926         return Constant::getNullValue(C1->getType());
927       // Fallthrough
928     case Instruction::Add:
929     case Instruction::Sub:
930       return UndefValue::get(C1->getType());
931     case Instruction::And:
932       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
933         return C1;
934       return Constant::getNullValue(C1->getType());   // undef & X -> 0
935     case Instruction::Mul: {
936       // undef * undef -> undef
937       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
938         return C1;
939       const APInt *CV;
940       // X * undef -> undef   if X is odd
941       if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV)))
942         if ((*CV)[0])
943           return UndefValue::get(C1->getType());
944 
945       // X * undef -> 0       otherwise
946       return Constant::getNullValue(C1->getType());
947     }
948     case Instruction::SDiv:
949     case Instruction::UDiv:
950       // X / undef -> undef
951       if (match(C1, m_Zero()))
952         return C2;
953       // undef / 0 -> undef
954       // undef / 1 -> undef
955       if (match(C2, m_Zero()) || match(C2, m_One()))
956         return C1;
957       // undef / X -> 0       otherwise
958       return Constant::getNullValue(C1->getType());
959     case Instruction::URem:
960     case Instruction::SRem:
961       // X % undef -> undef
962       if (match(C2, m_Undef()))
963         return C2;
964       // undef % 0 -> undef
965       if (match(C2, m_Zero()))
966         return C1;
967       // undef % X -> 0       otherwise
968       return Constant::getNullValue(C1->getType());
969     case Instruction::Or:                          // X | undef -> -1
970       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
971         return C1;
972       return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
973     case Instruction::LShr:
974       // X >>l undef -> undef
975       if (isa<UndefValue>(C2))
976         return C2;
977       // undef >>l 0 -> undef
978       if (match(C2, m_Zero()))
979         return C1;
980       // undef >>l X -> 0
981       return Constant::getNullValue(C1->getType());
982     case Instruction::AShr:
983       // X >>a undef -> undef
984       if (isa<UndefValue>(C2))
985         return C2;
986       // undef >>a 0 -> undef
987       if (match(C2, m_Zero()))
988         return C1;
989       // TODO: undef >>a X -> undef if the shift is exact
990       // undef >>a X -> 0
991       return Constant::getNullValue(C1->getType());
992     case Instruction::Shl:
993       // X << undef -> undef
994       if (isa<UndefValue>(C2))
995         return C2;
996       // undef << 0 -> undef
997       if (match(C2, m_Zero()))
998         return C1;
999       // undef << X -> 0
1000       return Constant::getNullValue(C1->getType());
1001     }
1002   }
1003 
1004   // Handle simplifications when the RHS is a constant int.
1005   if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1006     switch (Opcode) {
1007     case Instruction::Add:
1008       if (CI2->equalsInt(0)) return C1;                         // X + 0 == X
1009       break;
1010     case Instruction::Sub:
1011       if (CI2->equalsInt(0)) return C1;                         // X - 0 == X
1012       break;
1013     case Instruction::Mul:
1014       if (CI2->equalsInt(0)) return C2;                         // X * 0 == 0
1015       if (CI2->equalsInt(1))
1016         return C1;                                              // X * 1 == X
1017       break;
1018     case Instruction::UDiv:
1019     case Instruction::SDiv:
1020       if (CI2->equalsInt(1))
1021         return C1;                                            // X / 1 == X
1022       if (CI2->equalsInt(0))
1023         return UndefValue::get(CI2->getType());               // X / 0 == undef
1024       break;
1025     case Instruction::URem:
1026     case Instruction::SRem:
1027       if (CI2->equalsInt(1))
1028         return Constant::getNullValue(CI2->getType());        // X % 1 == 0
1029       if (CI2->equalsInt(0))
1030         return UndefValue::get(CI2->getType());               // X % 0 == undef
1031       break;
1032     case Instruction::And:
1033       if (CI2->isZero()) return C2;                           // X & 0 == 0
1034       if (CI2->isAllOnesValue())
1035         return C1;                                            // X & -1 == X
1036 
1037       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1038         // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1039         if (CE1->getOpcode() == Instruction::ZExt) {
1040           unsigned DstWidth = CI2->getType()->getBitWidth();
1041           unsigned SrcWidth =
1042             CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1043           APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1044           if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1045             return C1;
1046         }
1047 
1048         // If and'ing the address of a global with a constant, fold it.
1049         if (CE1->getOpcode() == Instruction::PtrToInt &&
1050             isa<GlobalValue>(CE1->getOperand(0))) {
1051           GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1052 
1053           // Functions are at least 4-byte aligned.
1054           unsigned GVAlign = GV->getAlignment();
1055           if (isa<Function>(GV))
1056             GVAlign = std::max(GVAlign, 4U);
1057 
1058           if (GVAlign > 1) {
1059             unsigned DstWidth = CI2->getType()->getBitWidth();
1060             unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1061             APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1062 
1063             // If checking bits we know are clear, return zero.
1064             if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1065               return Constant::getNullValue(CI2->getType());
1066           }
1067         }
1068       }
1069       break;
1070     case Instruction::Or:
1071       if (CI2->equalsInt(0)) return C1;    // X | 0 == X
1072       if (CI2->isAllOnesValue())
1073         return C2;                         // X | -1 == -1
1074       break;
1075     case Instruction::Xor:
1076       if (CI2->equalsInt(0)) return C1;    // X ^ 0 == X
1077 
1078       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1079         switch (CE1->getOpcode()) {
1080         default: break;
1081         case Instruction::ICmp:
1082         case Instruction::FCmp:
1083           // cmp pred ^ true -> cmp !pred
1084           assert(CI2->equalsInt(1));
1085           CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1086           pred = CmpInst::getInversePredicate(pred);
1087           return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1088                                           CE1->getOperand(1));
1089         }
1090       }
1091       break;
1092     case Instruction::AShr:
1093       // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1094       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1095         if (CE1->getOpcode() == Instruction::ZExt)  // Top bits known zero.
1096           return ConstantExpr::getLShr(C1, C2);
1097       break;
1098     }
1099   } else if (isa<ConstantInt>(C1)) {
1100     // If C1 is a ConstantInt and C2 is not, swap the operands.
1101     if (Instruction::isCommutative(Opcode))
1102       return ConstantExpr::get(Opcode, C2, C1);
1103   }
1104 
1105   // At this point we know neither constant is an UndefValue.
1106   if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1107     if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1108       const APInt &C1V = CI1->getValue();
1109       const APInt &C2V = CI2->getValue();
1110       switch (Opcode) {
1111       default:
1112         break;
1113       case Instruction::Add:
1114         return ConstantInt::get(CI1->getContext(), C1V + C2V);
1115       case Instruction::Sub:
1116         return ConstantInt::get(CI1->getContext(), C1V - C2V);
1117       case Instruction::Mul:
1118         return ConstantInt::get(CI1->getContext(), C1V * C2V);
1119       case Instruction::UDiv:
1120         assert(!CI2->isNullValue() && "Div by zero handled above");
1121         return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1122       case Instruction::SDiv:
1123         assert(!CI2->isNullValue() && "Div by zero handled above");
1124         if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1125           return UndefValue::get(CI1->getType());   // MIN_INT / -1 -> undef
1126         return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1127       case Instruction::URem:
1128         assert(!CI2->isNullValue() && "Div by zero handled above");
1129         return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1130       case Instruction::SRem:
1131         assert(!CI2->isNullValue() && "Div by zero handled above");
1132         if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1133           return UndefValue::get(CI1->getType());   // MIN_INT % -1 -> undef
1134         return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1135       case Instruction::And:
1136         return ConstantInt::get(CI1->getContext(), C1V & C2V);
1137       case Instruction::Or:
1138         return ConstantInt::get(CI1->getContext(), C1V | C2V);
1139       case Instruction::Xor:
1140         return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1141       case Instruction::Shl:
1142         if (C2V.ult(C1V.getBitWidth()))
1143           return ConstantInt::get(CI1->getContext(), C1V.shl(C2V));
1144         return UndefValue::get(C1->getType()); // too big shift is undef
1145       case Instruction::LShr:
1146         if (C2V.ult(C1V.getBitWidth()))
1147           return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V));
1148         return UndefValue::get(C1->getType()); // too big shift is undef
1149       case Instruction::AShr:
1150         if (C2V.ult(C1V.getBitWidth()))
1151           return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V));
1152         return UndefValue::get(C1->getType()); // too big shift is undef
1153       }
1154     }
1155 
1156     switch (Opcode) {
1157     case Instruction::SDiv:
1158     case Instruction::UDiv:
1159     case Instruction::URem:
1160     case Instruction::SRem:
1161     case Instruction::LShr:
1162     case Instruction::AShr:
1163     case Instruction::Shl:
1164       if (CI1->equalsInt(0)) return C1;
1165       break;
1166     default:
1167       break;
1168     }
1169   } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1170     if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1171       APFloat C1V = CFP1->getValueAPF();
1172       APFloat C2V = CFP2->getValueAPF();
1173       APFloat C3V = C1V;  // copy for modification
1174       switch (Opcode) {
1175       default:
1176         break;
1177       case Instruction::FAdd:
1178         (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1179         return ConstantFP::get(C1->getContext(), C3V);
1180       case Instruction::FSub:
1181         (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1182         return ConstantFP::get(C1->getContext(), C3V);
1183       case Instruction::FMul:
1184         (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1185         return ConstantFP::get(C1->getContext(), C3V);
1186       case Instruction::FDiv:
1187         (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1188         return ConstantFP::get(C1->getContext(), C3V);
1189       case Instruction::FRem:
1190         (void)C3V.mod(C2V);
1191         return ConstantFP::get(C1->getContext(), C3V);
1192       }
1193     }
1194   } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1195     // Perform elementwise folding.
1196     SmallVector<Constant*, 16> Result;
1197     Type *Ty = IntegerType::get(VTy->getContext(), 32);
1198     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1199       Constant *LHS =
1200         ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1201       Constant *RHS =
1202         ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1203 
1204       Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1205     }
1206 
1207     return ConstantVector::get(Result);
1208   }
1209 
1210   if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1211     // There are many possible foldings we could do here.  We should probably
1212     // at least fold add of a pointer with an integer into the appropriate
1213     // getelementptr.  This will improve alias analysis a bit.
1214 
1215     // Given ((a + b) + c), if (b + c) folds to something interesting, return
1216     // (a + (b + c)).
1217     if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1218       Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1219       if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1220         return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1221     }
1222   } else if (isa<ConstantExpr>(C2)) {
1223     // If C2 is a constant expr and C1 isn't, flop them around and fold the
1224     // other way if possible.
1225     if (Instruction::isCommutative(Opcode))
1226       return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1227   }
1228 
1229   // i1 can be simplified in many cases.
1230   if (C1->getType()->isIntegerTy(1)) {
1231     switch (Opcode) {
1232     case Instruction::Add:
1233     case Instruction::Sub:
1234       return ConstantExpr::getXor(C1, C2);
1235     case Instruction::Mul:
1236       return ConstantExpr::getAnd(C1, C2);
1237     case Instruction::Shl:
1238     case Instruction::LShr:
1239     case Instruction::AShr:
1240       // We can assume that C2 == 0.  If it were one the result would be
1241       // undefined because the shift value is as large as the bitwidth.
1242       return C1;
1243     case Instruction::SDiv:
1244     case Instruction::UDiv:
1245       // We can assume that C2 == 1.  If it were zero the result would be
1246       // undefined through division by zero.
1247       return C1;
1248     case Instruction::URem:
1249     case Instruction::SRem:
1250       // We can assume that C2 == 1.  If it were zero the result would be
1251       // undefined through division by zero.
1252       return ConstantInt::getFalse(C1->getContext());
1253     default:
1254       break;
1255     }
1256   }
1257 
1258   // We don't know how to fold this.
1259   return nullptr;
1260 }
1261 
1262 /// isZeroSizedType - This type is zero sized if its an array or structure of
1263 /// zero sized types.  The only leaf zero sized type is an empty structure.
isMaybeZeroSizedType(Type * Ty)1264 static bool isMaybeZeroSizedType(Type *Ty) {
1265   if (StructType *STy = dyn_cast<StructType>(Ty)) {
1266     if (STy->isOpaque()) return true;  // Can't say.
1267 
1268     // If all of elements have zero size, this does too.
1269     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1270       if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1271     return true;
1272 
1273   } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1274     return isMaybeZeroSizedType(ATy->getElementType());
1275   }
1276   return false;
1277 }
1278 
1279 /// IdxCompare - Compare the two constants as though they were getelementptr
1280 /// indices.  This allows coercion of the types to be the same thing.
1281 ///
1282 /// If the two constants are the "same" (after coercion), return 0.  If the
1283 /// first is less than the second, return -1, if the second is less than the
1284 /// first, return 1.  If the constants are not integral, return -2.
1285 ///
IdxCompare(Constant * C1,Constant * C2,Type * ElTy)1286 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1287   if (C1 == C2) return 0;
1288 
1289   // Ok, we found a different index.  If they are not ConstantInt, we can't do
1290   // anything with them.
1291   if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1292     return -2; // don't know!
1293 
1294   // We cannot compare the indices if they don't fit in an int64_t.
1295   if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 ||
1296       cast<ConstantInt>(C2)->getValue().getActiveBits() > 64)
1297     return -2; // don't know!
1298 
1299   // Ok, we have two differing integer indices.  Sign extend them to be the same
1300   // type.
1301   int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue();
1302   int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue();
1303 
1304   if (C1Val == C2Val) return 0;  // They are equal
1305 
1306   // If the type being indexed over is really just a zero sized type, there is
1307   // no pointer difference being made here.
1308   if (isMaybeZeroSizedType(ElTy))
1309     return -2; // dunno.
1310 
1311   // If they are really different, now that they are the same type, then we
1312   // found a difference!
1313   if (C1Val < C2Val)
1314     return -1;
1315   else
1316     return 1;
1317 }
1318 
1319 /// evaluateFCmpRelation - This function determines if there is anything we can
1320 /// decide about the two constants provided.  This doesn't need to handle simple
1321 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1322 /// If we can determine that the two constants have a particular relation to
1323 /// each other, we should return the corresponding FCmpInst predicate,
1324 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1325 /// ConstantFoldCompareInstruction.
1326 ///
1327 /// To simplify this code we canonicalize the relation so that the first
1328 /// operand is always the most "complex" of the two.  We consider ConstantFP
1329 /// to be the simplest, and ConstantExprs to be the most complex.
evaluateFCmpRelation(Constant * V1,Constant * V2)1330 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1331   assert(V1->getType() == V2->getType() &&
1332          "Cannot compare values of different types!");
1333 
1334   // Handle degenerate case quickly
1335   if (V1 == V2) return FCmpInst::FCMP_OEQ;
1336 
1337   if (!isa<ConstantExpr>(V1)) {
1338     if (!isa<ConstantExpr>(V2)) {
1339       // Simple case, use the standard constant folder.
1340       ConstantInt *R = nullptr;
1341       R = dyn_cast<ConstantInt>(
1342                       ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1343       if (R && !R->isZero())
1344         return FCmpInst::FCMP_OEQ;
1345       R = dyn_cast<ConstantInt>(
1346                       ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1347       if (R && !R->isZero())
1348         return FCmpInst::FCMP_OLT;
1349       R = dyn_cast<ConstantInt>(
1350                       ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1351       if (R && !R->isZero())
1352         return FCmpInst::FCMP_OGT;
1353 
1354       // Nothing more we can do
1355       return FCmpInst::BAD_FCMP_PREDICATE;
1356     }
1357 
1358     // If the first operand is simple and second is ConstantExpr, swap operands.
1359     FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1360     if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1361       return FCmpInst::getSwappedPredicate(SwappedRelation);
1362   } else {
1363     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1364     // constantexpr or a simple constant.
1365     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1366     switch (CE1->getOpcode()) {
1367     case Instruction::FPTrunc:
1368     case Instruction::FPExt:
1369     case Instruction::UIToFP:
1370     case Instruction::SIToFP:
1371       // We might be able to do something with these but we don't right now.
1372       break;
1373     default:
1374       break;
1375     }
1376   }
1377   // There are MANY other foldings that we could perform here.  They will
1378   // probably be added on demand, as they seem needed.
1379   return FCmpInst::BAD_FCMP_PREDICATE;
1380 }
1381 
areGlobalsPotentiallyEqual(const GlobalValue * GV1,const GlobalValue * GV2)1382 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1383                                                       const GlobalValue *GV2) {
1384   auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) {
1385     if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage())
1386       return true;
1387     if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) {
1388       Type *Ty = GVar->getValueType();
1389       // A global with opaque type might end up being zero sized.
1390       if (!Ty->isSized())
1391         return true;
1392       // A global with an empty type might lie at the address of any other
1393       // global.
1394       if (Ty->isEmptyTy())
1395         return true;
1396     }
1397     return false;
1398   };
1399   // Don't try to decide equality of aliases.
1400   if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1401     if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2))
1402       return ICmpInst::ICMP_NE;
1403   return ICmpInst::BAD_ICMP_PREDICATE;
1404 }
1405 
1406 /// evaluateICmpRelation - This function determines if there is anything we can
1407 /// decide about the two constants provided.  This doesn't need to handle simple
1408 /// things like integer comparisons, but should instead handle ConstantExprs
1409 /// and GlobalValues.  If we can determine that the two constants have a
1410 /// particular relation to each other, we should return the corresponding ICmp
1411 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1412 ///
1413 /// To simplify this code we canonicalize the relation so that the first
1414 /// operand is always the most "complex" of the two.  We consider simple
1415 /// constants (like ConstantInt) to be the simplest, followed by
1416 /// GlobalValues, followed by ConstantExpr's (the most complex).
1417 ///
evaluateICmpRelation(Constant * V1,Constant * V2,bool isSigned)1418 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1419                                                 bool isSigned) {
1420   assert(V1->getType() == V2->getType() &&
1421          "Cannot compare different types of values!");
1422   if (V1 == V2) return ICmpInst::ICMP_EQ;
1423 
1424   if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1425       !isa<BlockAddress>(V1)) {
1426     if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1427         !isa<BlockAddress>(V2)) {
1428       // We distilled this down to a simple case, use the standard constant
1429       // folder.
1430       ConstantInt *R = nullptr;
1431       ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1432       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1433       if (R && !R->isZero())
1434         return pred;
1435       pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1436       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1437       if (R && !R->isZero())
1438         return pred;
1439       pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1440       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1441       if (R && !R->isZero())
1442         return pred;
1443 
1444       // If we couldn't figure it out, bail.
1445       return ICmpInst::BAD_ICMP_PREDICATE;
1446     }
1447 
1448     // If the first operand is simple, swap operands.
1449     ICmpInst::Predicate SwappedRelation =
1450       evaluateICmpRelation(V2, V1, isSigned);
1451     if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1452       return ICmpInst::getSwappedPredicate(SwappedRelation);
1453 
1454   } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1455     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1456       ICmpInst::Predicate SwappedRelation =
1457         evaluateICmpRelation(V2, V1, isSigned);
1458       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1459         return ICmpInst::getSwappedPredicate(SwappedRelation);
1460       return ICmpInst::BAD_ICMP_PREDICATE;
1461     }
1462 
1463     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1464     // constant (which, since the types must match, means that it's a
1465     // ConstantPointerNull).
1466     if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1467       return areGlobalsPotentiallyEqual(GV, GV2);
1468     } else if (isa<BlockAddress>(V2)) {
1469       return ICmpInst::ICMP_NE; // Globals never equal labels.
1470     } else {
1471       assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1472       // GlobalVals can never be null unless they have external weak linkage.
1473       // We don't try to evaluate aliases here.
1474       if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1475         return ICmpInst::ICMP_NE;
1476     }
1477   } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1478     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1479       ICmpInst::Predicate SwappedRelation =
1480         evaluateICmpRelation(V2, V1, isSigned);
1481       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1482         return ICmpInst::getSwappedPredicate(SwappedRelation);
1483       return ICmpInst::BAD_ICMP_PREDICATE;
1484     }
1485 
1486     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1487     // constant (which, since the types must match, means that it is a
1488     // ConstantPointerNull).
1489     if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1490       // Block address in another function can't equal this one, but block
1491       // addresses in the current function might be the same if blocks are
1492       // empty.
1493       if (BA2->getFunction() != BA->getFunction())
1494         return ICmpInst::ICMP_NE;
1495     } else {
1496       // Block addresses aren't null, don't equal the address of globals.
1497       assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1498              "Canonicalization guarantee!");
1499       return ICmpInst::ICMP_NE;
1500     }
1501   } else {
1502     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1503     // constantexpr, a global, block address, or a simple constant.
1504     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1505     Constant *CE1Op0 = CE1->getOperand(0);
1506 
1507     switch (CE1->getOpcode()) {
1508     case Instruction::Trunc:
1509     case Instruction::FPTrunc:
1510     case Instruction::FPExt:
1511     case Instruction::FPToUI:
1512     case Instruction::FPToSI:
1513       break; // We can't evaluate floating point casts or truncations.
1514 
1515     case Instruction::UIToFP:
1516     case Instruction::SIToFP:
1517     case Instruction::BitCast:
1518     case Instruction::ZExt:
1519     case Instruction::SExt:
1520       // If the cast is not actually changing bits, and the second operand is a
1521       // null pointer, do the comparison with the pre-casted value.
1522       if (V2->isNullValue() &&
1523           (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1524         if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1525         if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1526         return evaluateICmpRelation(CE1Op0,
1527                                     Constant::getNullValue(CE1Op0->getType()),
1528                                     isSigned);
1529       }
1530       break;
1531 
1532     case Instruction::GetElementPtr: {
1533       GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1534       // Ok, since this is a getelementptr, we know that the constant has a
1535       // pointer type.  Check the various cases.
1536       if (isa<ConstantPointerNull>(V2)) {
1537         // If we are comparing a GEP to a null pointer, check to see if the base
1538         // of the GEP equals the null pointer.
1539         if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1540           if (GV->hasExternalWeakLinkage())
1541             // Weak linkage GVals could be zero or not. We're comparing that
1542             // to null pointer so its greater-or-equal
1543             return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1544           else
1545             // If its not weak linkage, the GVal must have a non-zero address
1546             // so the result is greater-than
1547             return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1548         } else if (isa<ConstantPointerNull>(CE1Op0)) {
1549           // If we are indexing from a null pointer, check to see if we have any
1550           // non-zero indices.
1551           for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1552             if (!CE1->getOperand(i)->isNullValue())
1553               // Offsetting from null, must not be equal.
1554               return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1555           // Only zero indexes from null, must still be zero.
1556           return ICmpInst::ICMP_EQ;
1557         }
1558         // Otherwise, we can't really say if the first operand is null or not.
1559       } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1560         if (isa<ConstantPointerNull>(CE1Op0)) {
1561           if (GV2->hasExternalWeakLinkage())
1562             // Weak linkage GVals could be zero or not. We're comparing it to
1563             // a null pointer, so its less-or-equal
1564             return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1565           else
1566             // If its not weak linkage, the GVal must have a non-zero address
1567             // so the result is less-than
1568             return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1569         } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1570           if (GV == GV2) {
1571             // If this is a getelementptr of the same global, then it must be
1572             // different.  Because the types must match, the getelementptr could
1573             // only have at most one index, and because we fold getelementptr's
1574             // with a single zero index, it must be nonzero.
1575             assert(CE1->getNumOperands() == 2 &&
1576                    !CE1->getOperand(1)->isNullValue() &&
1577                    "Surprising getelementptr!");
1578             return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1579           } else {
1580             if (CE1GEP->hasAllZeroIndices())
1581               return areGlobalsPotentiallyEqual(GV, GV2);
1582             return ICmpInst::BAD_ICMP_PREDICATE;
1583           }
1584         }
1585       } else {
1586         ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1587         Constant *CE2Op0 = CE2->getOperand(0);
1588 
1589         // There are MANY other foldings that we could perform here.  They will
1590         // probably be added on demand, as they seem needed.
1591         switch (CE2->getOpcode()) {
1592         default: break;
1593         case Instruction::GetElementPtr:
1594           // By far the most common case to handle is when the base pointers are
1595           // obviously to the same global.
1596           if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1597             // Don't know relative ordering, but check for inequality.
1598             if (CE1Op0 != CE2Op0) {
1599               GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1600               if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1601                 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1602                                                   cast<GlobalValue>(CE2Op0));
1603               return ICmpInst::BAD_ICMP_PREDICATE;
1604             }
1605             // Ok, we know that both getelementptr instructions are based on the
1606             // same global.  From this, we can precisely determine the relative
1607             // ordering of the resultant pointers.
1608             unsigned i = 1;
1609 
1610             // The logic below assumes that the result of the comparison
1611             // can be determined by finding the first index that differs.
1612             // This doesn't work if there is over-indexing in any
1613             // subsequent indices, so check for that case first.
1614             if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1615                 !CE2->isGEPWithNoNotionalOverIndexing())
1616                return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1617 
1618             // Compare all of the operands the GEP's have in common.
1619             gep_type_iterator GTI = gep_type_begin(CE1);
1620             for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1621                  ++i, ++GTI)
1622               switch (IdxCompare(CE1->getOperand(i),
1623                                  CE2->getOperand(i), GTI.getIndexedType())) {
1624               case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1625               case 1:  return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1626               case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1627               }
1628 
1629             // Ok, we ran out of things they have in common.  If any leftovers
1630             // are non-zero then we have a difference, otherwise we are equal.
1631             for (; i < CE1->getNumOperands(); ++i)
1632               if (!CE1->getOperand(i)->isNullValue()) {
1633                 if (isa<ConstantInt>(CE1->getOperand(i)))
1634                   return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1635                 else
1636                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1637               }
1638 
1639             for (; i < CE2->getNumOperands(); ++i)
1640               if (!CE2->getOperand(i)->isNullValue()) {
1641                 if (isa<ConstantInt>(CE2->getOperand(i)))
1642                   return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1643                 else
1644                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1645               }
1646             return ICmpInst::ICMP_EQ;
1647           }
1648         }
1649       }
1650     }
1651     default:
1652       break;
1653     }
1654   }
1655 
1656   return ICmpInst::BAD_ICMP_PREDICATE;
1657 }
1658 
ConstantFoldCompareInstruction(unsigned short pred,Constant * C1,Constant * C2)1659 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1660                                                Constant *C1, Constant *C2) {
1661   Type *ResultTy;
1662   if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1663     ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1664                                VT->getNumElements());
1665   else
1666     ResultTy = Type::getInt1Ty(C1->getContext());
1667 
1668   // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1669   if (pred == FCmpInst::FCMP_FALSE)
1670     return Constant::getNullValue(ResultTy);
1671 
1672   if (pred == FCmpInst::FCMP_TRUE)
1673     return Constant::getAllOnesValue(ResultTy);
1674 
1675   // Handle some degenerate cases first
1676   if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1677     CmpInst::Predicate Predicate = CmpInst::Predicate(pred);
1678     bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate);
1679     // For EQ and NE, we can always pick a value for the undef to make the
1680     // predicate pass or fail, so we can return undef.
1681     // Also, if both operands are undef, we can return undef for int comparison.
1682     if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2))
1683       return UndefValue::get(ResultTy);
1684 
1685     // Otherwise, for integer compare, pick the same value as the non-undef
1686     // operand, and fold it to true or false.
1687     if (isIntegerPredicate)
1688       return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate));
1689 
1690     // Choosing NaN for the undef will always make unordered comparison succeed
1691     // and ordered comparison fails.
1692     return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate));
1693   }
1694 
1695   // icmp eq/ne(null,GV) -> false/true
1696   if (C1->isNullValue()) {
1697     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1698       // Don't try to evaluate aliases.  External weak GV can be null.
1699       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1700         if (pred == ICmpInst::ICMP_EQ)
1701           return ConstantInt::getFalse(C1->getContext());
1702         else if (pred == ICmpInst::ICMP_NE)
1703           return ConstantInt::getTrue(C1->getContext());
1704       }
1705   // icmp eq/ne(GV,null) -> false/true
1706   } else if (C2->isNullValue()) {
1707     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1708       // Don't try to evaluate aliases.  External weak GV can be null.
1709       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1710         if (pred == ICmpInst::ICMP_EQ)
1711           return ConstantInt::getFalse(C1->getContext());
1712         else if (pred == ICmpInst::ICMP_NE)
1713           return ConstantInt::getTrue(C1->getContext());
1714       }
1715   }
1716 
1717   // If the comparison is a comparison between two i1's, simplify it.
1718   if (C1->getType()->isIntegerTy(1)) {
1719     switch(pred) {
1720     case ICmpInst::ICMP_EQ:
1721       if (isa<ConstantInt>(C2))
1722         return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1723       return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1724     case ICmpInst::ICMP_NE:
1725       return ConstantExpr::getXor(C1, C2);
1726     default:
1727       break;
1728     }
1729   }
1730 
1731   if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1732     APInt V1 = cast<ConstantInt>(C1)->getValue();
1733     APInt V2 = cast<ConstantInt>(C2)->getValue();
1734     switch (pred) {
1735     default: llvm_unreachable("Invalid ICmp Predicate");
1736     case ICmpInst::ICMP_EQ:  return ConstantInt::get(ResultTy, V1 == V2);
1737     case ICmpInst::ICMP_NE:  return ConstantInt::get(ResultTy, V1 != V2);
1738     case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1739     case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1740     case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1741     case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1742     case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1743     case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1744     case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1745     case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1746     }
1747   } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1748     APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1749     APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1750     APFloat::cmpResult R = C1V.compare(C2V);
1751     switch (pred) {
1752     default: llvm_unreachable("Invalid FCmp Predicate");
1753     case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1754     case FCmpInst::FCMP_TRUE:  return Constant::getAllOnesValue(ResultTy);
1755     case FCmpInst::FCMP_UNO:
1756       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1757     case FCmpInst::FCMP_ORD:
1758       return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1759     case FCmpInst::FCMP_UEQ:
1760       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1761                                         R==APFloat::cmpEqual);
1762     case FCmpInst::FCMP_OEQ:
1763       return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1764     case FCmpInst::FCMP_UNE:
1765       return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1766     case FCmpInst::FCMP_ONE:
1767       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1768                                         R==APFloat::cmpGreaterThan);
1769     case FCmpInst::FCMP_ULT:
1770       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1771                                         R==APFloat::cmpLessThan);
1772     case FCmpInst::FCMP_OLT:
1773       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1774     case FCmpInst::FCMP_UGT:
1775       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1776                                         R==APFloat::cmpGreaterThan);
1777     case FCmpInst::FCMP_OGT:
1778       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1779     case FCmpInst::FCMP_ULE:
1780       return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1781     case FCmpInst::FCMP_OLE:
1782       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1783                                         R==APFloat::cmpEqual);
1784     case FCmpInst::FCMP_UGE:
1785       return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1786     case FCmpInst::FCMP_OGE:
1787       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1788                                         R==APFloat::cmpEqual);
1789     }
1790   } else if (C1->getType()->isVectorTy()) {
1791     // If we can constant fold the comparison of each element, constant fold
1792     // the whole vector comparison.
1793     SmallVector<Constant*, 4> ResElts;
1794     Type *Ty = IntegerType::get(C1->getContext(), 32);
1795     // Compare the elements, producing an i1 result or constant expr.
1796     for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1797       Constant *C1E =
1798         ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1799       Constant *C2E =
1800         ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1801 
1802       ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1803     }
1804 
1805     return ConstantVector::get(ResElts);
1806   }
1807 
1808   if (C1->getType()->isFloatingPointTy() &&
1809       // Only call evaluateFCmpRelation if we have a constant expr to avoid
1810       // infinite recursive loop
1811       (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) {
1812     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
1813     switch (evaluateFCmpRelation(C1, C2)) {
1814     default: llvm_unreachable("Unknown relation!");
1815     case FCmpInst::FCMP_UNO:
1816     case FCmpInst::FCMP_ORD:
1817     case FCmpInst::FCMP_UEQ:
1818     case FCmpInst::FCMP_UNE:
1819     case FCmpInst::FCMP_ULT:
1820     case FCmpInst::FCMP_UGT:
1821     case FCmpInst::FCMP_ULE:
1822     case FCmpInst::FCMP_UGE:
1823     case FCmpInst::FCMP_TRUE:
1824     case FCmpInst::FCMP_FALSE:
1825     case FCmpInst::BAD_FCMP_PREDICATE:
1826       break; // Couldn't determine anything about these constants.
1827     case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1828       Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1829                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1830                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1831       break;
1832     case FCmpInst::FCMP_OLT: // We know that C1 < C2
1833       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1834                 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1835                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1836       break;
1837     case FCmpInst::FCMP_OGT: // We know that C1 > C2
1838       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1839                 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1840                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1841       break;
1842     case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1843       // We can only partially decide this relation.
1844       if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1845         Result = 0;
1846       else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1847         Result = 1;
1848       break;
1849     case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1850       // We can only partially decide this relation.
1851       if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1852         Result = 0;
1853       else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1854         Result = 1;
1855       break;
1856     case FCmpInst::FCMP_ONE: // We know that C1 != C2
1857       // We can only partially decide this relation.
1858       if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1859         Result = 0;
1860       else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1861         Result = 1;
1862       break;
1863     }
1864 
1865     // If we evaluated the result, return it now.
1866     if (Result != -1)
1867       return ConstantInt::get(ResultTy, Result);
1868 
1869   } else {
1870     // Evaluate the relation between the two constants, per the predicate.
1871     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
1872     switch (evaluateICmpRelation(C1, C2,
1873                                  CmpInst::isSigned((CmpInst::Predicate)pred))) {
1874     default: llvm_unreachable("Unknown relational!");
1875     case ICmpInst::BAD_ICMP_PREDICATE:
1876       break;  // Couldn't determine anything about these constants.
1877     case ICmpInst::ICMP_EQ:   // We know the constants are equal!
1878       // If we know the constants are equal, we can decide the result of this
1879       // computation precisely.
1880       Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1881       break;
1882     case ICmpInst::ICMP_ULT:
1883       switch (pred) {
1884       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1885         Result = 1; break;
1886       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1887         Result = 0; break;
1888       }
1889       break;
1890     case ICmpInst::ICMP_SLT:
1891       switch (pred) {
1892       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1893         Result = 1; break;
1894       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1895         Result = 0; break;
1896       }
1897       break;
1898     case ICmpInst::ICMP_UGT:
1899       switch (pred) {
1900       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1901         Result = 1; break;
1902       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1903         Result = 0; break;
1904       }
1905       break;
1906     case ICmpInst::ICMP_SGT:
1907       switch (pred) {
1908       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1909         Result = 1; break;
1910       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1911         Result = 0; break;
1912       }
1913       break;
1914     case ICmpInst::ICMP_ULE:
1915       if (pred == ICmpInst::ICMP_UGT) Result = 0;
1916       if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1917       break;
1918     case ICmpInst::ICMP_SLE:
1919       if (pred == ICmpInst::ICMP_SGT) Result = 0;
1920       if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1921       break;
1922     case ICmpInst::ICMP_UGE:
1923       if (pred == ICmpInst::ICMP_ULT) Result = 0;
1924       if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
1925       break;
1926     case ICmpInst::ICMP_SGE:
1927       if (pred == ICmpInst::ICMP_SLT) Result = 0;
1928       if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
1929       break;
1930     case ICmpInst::ICMP_NE:
1931       if (pred == ICmpInst::ICMP_EQ) Result = 0;
1932       if (pred == ICmpInst::ICMP_NE) Result = 1;
1933       break;
1934     }
1935 
1936     // If we evaluated the result, return it now.
1937     if (Result != -1)
1938       return ConstantInt::get(ResultTy, Result);
1939 
1940     // If the right hand side is a bitcast, try using its inverse to simplify
1941     // it by moving it to the left hand side.  We can't do this if it would turn
1942     // a vector compare into a scalar compare or visa versa.
1943     if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
1944       Constant *CE2Op0 = CE2->getOperand(0);
1945       if (CE2->getOpcode() == Instruction::BitCast &&
1946           CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
1947         Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
1948         return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
1949       }
1950     }
1951 
1952     // If the left hand side is an extension, try eliminating it.
1953     if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1954       if ((CE1->getOpcode() == Instruction::SExt &&
1955            ICmpInst::isSigned((ICmpInst::Predicate)pred)) ||
1956           (CE1->getOpcode() == Instruction::ZExt &&
1957            !ICmpInst::isSigned((ICmpInst::Predicate)pred))){
1958         Constant *CE1Op0 = CE1->getOperand(0);
1959         Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
1960         if (CE1Inverse == CE1Op0) {
1961           // Check whether we can safely truncate the right hand side.
1962           Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
1963           if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
1964                                     C2->getType()) == C2)
1965             return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
1966         }
1967       }
1968     }
1969 
1970     if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
1971         (C1->isNullValue() && !C2->isNullValue())) {
1972       // If C2 is a constant expr and C1 isn't, flip them around and fold the
1973       // other way if possible.
1974       // Also, if C1 is null and C2 isn't, flip them around.
1975       pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1976       return ConstantExpr::getICmp(pred, C2, C1);
1977     }
1978   }
1979   return nullptr;
1980 }
1981 
1982 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
1983 /// is "inbounds".
1984 template<typename IndexTy>
isInBoundsIndices(ArrayRef<IndexTy> Idxs)1985 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
1986   // No indices means nothing that could be out of bounds.
1987   if (Idxs.empty()) return true;
1988 
1989   // If the first index is zero, it's in bounds.
1990   if (cast<Constant>(Idxs[0])->isNullValue()) return true;
1991 
1992   // If the first index is one and all the rest are zero, it's in bounds,
1993   // by the one-past-the-end rule.
1994   if (!cast<ConstantInt>(Idxs[0])->isOne())
1995     return false;
1996   for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
1997     if (!cast<Constant>(Idxs[i])->isNullValue())
1998       return false;
1999   return true;
2000 }
2001 
2002 /// \brief Test whether a given ConstantInt is in-range for a SequentialType.
isIndexInRangeOfSequentialType(SequentialType * STy,const ConstantInt * CI)2003 static bool isIndexInRangeOfSequentialType(SequentialType *STy,
2004                                            const ConstantInt *CI) {
2005   // And indices are valid when indexing along a pointer
2006   if (isa<PointerType>(STy))
2007     return true;
2008 
2009   uint64_t NumElements = 0;
2010   // Determine the number of elements in our sequential type.
2011   if (auto *ATy = dyn_cast<ArrayType>(STy))
2012     NumElements = ATy->getNumElements();
2013   else if (auto *VTy = dyn_cast<VectorType>(STy))
2014     NumElements = VTy->getNumElements();
2015 
2016   assert((isa<ArrayType>(STy) || NumElements > 0) &&
2017          "didn't expect non-array type to have zero elements!");
2018 
2019   // We cannot bounds check the index if it doesn't fit in an int64_t.
2020   if (CI->getValue().getActiveBits() > 64)
2021     return false;
2022 
2023   // A negative index or an index past the end of our sequential type is
2024   // considered out-of-range.
2025   int64_t IndexVal = CI->getSExtValue();
2026   if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
2027     return false;
2028 
2029   // Otherwise, it is in-range.
2030   return true;
2031 }
2032 
2033 template<typename IndexTy>
ConstantFoldGetElementPtrImpl(Type * PointeeTy,Constant * C,bool inBounds,ArrayRef<IndexTy> Idxs)2034 static Constant *ConstantFoldGetElementPtrImpl(Type *PointeeTy, Constant *C,
2035                                                bool inBounds,
2036                                                ArrayRef<IndexTy> Idxs) {
2037   if (Idxs.empty()) return C;
2038   Constant *Idx0 = cast<Constant>(Idxs[0]);
2039   if ((Idxs.size() == 1 && Idx0->isNullValue()))
2040     return C;
2041 
2042   if (isa<UndefValue>(C)) {
2043     PointerType *Ptr = cast<PointerType>(C->getType());
2044     Type *Ty = GetElementPtrInst::getIndexedType(
2045         cast<PointerType>(Ptr->getScalarType())->getElementType(), Idxs);
2046     assert(Ty && "Invalid indices for GEP!");
2047     return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
2048   }
2049 
2050   if (C->isNullValue()) {
2051     bool isNull = true;
2052     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2053       if (!cast<Constant>(Idxs[i])->isNullValue()) {
2054         isNull = false;
2055         break;
2056       }
2057     if (isNull) {
2058       PointerType *Ptr = cast<PointerType>(C->getType());
2059       Type *Ty = GetElementPtrInst::getIndexedType(
2060           cast<PointerType>(Ptr->getScalarType())->getElementType(), Idxs);
2061       assert(Ty && "Invalid indices for GEP!");
2062       return ConstantPointerNull::get(PointerType::get(Ty,
2063                                                        Ptr->getAddressSpace()));
2064     }
2065   }
2066 
2067   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2068     // Combine Indices - If the source pointer to this getelementptr instruction
2069     // is a getelementptr instruction, combine the indices of the two
2070     // getelementptr instructions into a single instruction.
2071     //
2072     if (CE->getOpcode() == Instruction::GetElementPtr) {
2073       Type *LastTy = nullptr;
2074       for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2075            I != E; ++I)
2076         LastTy = *I;
2077 
2078       // We cannot combine indices if doing so would take us outside of an
2079       // array or vector.  Doing otherwise could trick us if we evaluated such a
2080       // GEP as part of a load.
2081       //
2082       // e.g. Consider if the original GEP was:
2083       // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2084       //                    i32 0, i32 0, i64 0)
2085       //
2086       // If we then tried to offset it by '8' to get to the third element,
2087       // an i8, we should *not* get:
2088       // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2089       //                    i32 0, i32 0, i64 8)
2090       //
2091       // This GEP tries to index array element '8  which runs out-of-bounds.
2092       // Subsequent evaluation would get confused and produce erroneous results.
2093       //
2094       // The following prohibits such a GEP from being formed by checking to see
2095       // if the index is in-range with respect to an array or vector.
2096       bool PerformFold = false;
2097       if (Idx0->isNullValue())
2098         PerformFold = true;
2099       else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy))
2100         if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0))
2101           PerformFold = isIndexInRangeOfSequentialType(STy, CI);
2102 
2103       if (PerformFold) {
2104         SmallVector<Value*, 16> NewIndices;
2105         NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2106         NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1);
2107 
2108         // Add the last index of the source with the first index of the new GEP.
2109         // Make sure to handle the case when they are actually different types.
2110         Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2111         // Otherwise it must be an array.
2112         if (!Idx0->isNullValue()) {
2113           Type *IdxTy = Combined->getType();
2114           if (IdxTy != Idx0->getType()) {
2115             unsigned CommonExtendedWidth =
2116                 std::max(IdxTy->getIntegerBitWidth(),
2117                          Idx0->getType()->getIntegerBitWidth());
2118             CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2119 
2120             Type *CommonTy =
2121                 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth);
2122             Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy);
2123             Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy);
2124             Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2125           } else {
2126             Combined =
2127               ConstantExpr::get(Instruction::Add, Idx0, Combined);
2128           }
2129         }
2130 
2131         NewIndices.push_back(Combined);
2132         NewIndices.append(Idxs.begin() + 1, Idxs.end());
2133         return ConstantExpr::getGetElementPtr(
2134             cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0),
2135             NewIndices, inBounds && cast<GEPOperator>(CE)->isInBounds());
2136       }
2137     }
2138 
2139     // Attempt to fold casts to the same type away.  For example, folding:
2140     //
2141     //   i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2142     //                       i64 0, i64 0)
2143     // into:
2144     //
2145     //   i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2146     //
2147     // Don't fold if the cast is changing address spaces.
2148     if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2149       PointerType *SrcPtrTy =
2150         dyn_cast<PointerType>(CE->getOperand(0)->getType());
2151       PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2152       if (SrcPtrTy && DstPtrTy) {
2153         ArrayType *SrcArrayTy =
2154           dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2155         ArrayType *DstArrayTy =
2156           dyn_cast<ArrayType>(DstPtrTy->getElementType());
2157         if (SrcArrayTy && DstArrayTy
2158             && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2159             && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2160           return ConstantExpr::getGetElementPtr(
2161               SrcArrayTy, (Constant *)CE->getOperand(0), Idxs, inBounds);
2162       }
2163     }
2164   }
2165 
2166   // Check to see if any array indices are not within the corresponding
2167   // notional array or vector bounds. If so, try to determine if they can be
2168   // factored out into preceding dimensions.
2169   SmallVector<Constant *, 8> NewIdxs;
2170   Type *Ty = PointeeTy;
2171   Type *Prev = C->getType();
2172   bool Unknown = !isa<ConstantInt>(Idxs[0]);
2173   for (unsigned i = 1, e = Idxs.size(); i != e;
2174        Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2175     if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2176       if (isa<ArrayType>(Ty) || isa<VectorType>(Ty))
2177         if (CI->getSExtValue() > 0 &&
2178             !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) {
2179           if (isa<SequentialType>(Prev)) {
2180             // It's out of range, but we can factor it into the prior
2181             // dimension.
2182             NewIdxs.resize(Idxs.size());
2183             uint64_t NumElements = 0;
2184             if (auto *ATy = dyn_cast<ArrayType>(Ty))
2185               NumElements = ATy->getNumElements();
2186             else
2187               NumElements = cast<VectorType>(Ty)->getNumElements();
2188 
2189             ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements);
2190             NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2191 
2192             Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2193             Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2194 
2195             unsigned CommonExtendedWidth =
2196                 std::max(PrevIdx->getType()->getIntegerBitWidth(),
2197                          Div->getType()->getIntegerBitWidth());
2198             CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2199 
2200             // Before adding, extend both operands to i64 to avoid
2201             // overflow trouble.
2202             if (!PrevIdx->getType()->isIntegerTy(CommonExtendedWidth))
2203               PrevIdx = ConstantExpr::getSExt(
2204                   PrevIdx,
2205                   Type::getIntNTy(Div->getContext(), CommonExtendedWidth));
2206             if (!Div->getType()->isIntegerTy(CommonExtendedWidth))
2207               Div = ConstantExpr::getSExt(
2208                   Div, Type::getIntNTy(Div->getContext(), CommonExtendedWidth));
2209 
2210             NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2211           } else {
2212             // It's out of range, but the prior dimension is a struct
2213             // so we can't do anything about it.
2214             Unknown = true;
2215           }
2216         }
2217     } else {
2218       // We don't know if it's in range or not.
2219       Unknown = true;
2220     }
2221   }
2222 
2223   // If we did any factoring, start over with the adjusted indices.
2224   if (!NewIdxs.empty()) {
2225     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2226       if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2227     return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, inBounds);
2228   }
2229 
2230   // If all indices are known integers and normalized, we can do a simple
2231   // check for the "inbounds" property.
2232   if (!Unknown && !inBounds)
2233     if (auto *GV = dyn_cast<GlobalVariable>(C))
2234       if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs))
2235         return ConstantExpr::getInBoundsGetElementPtr(PointeeTy, C, Idxs);
2236 
2237   return nullptr;
2238 }
2239 
ConstantFoldGetElementPtr(Constant * C,bool inBounds,ArrayRef<Constant * > Idxs)2240 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2241                                           bool inBounds,
2242                                           ArrayRef<Constant *> Idxs) {
2243   return ConstantFoldGetElementPtrImpl(
2244       cast<PointerType>(C->getType()->getScalarType())->getElementType(), C,
2245       inBounds, Idxs);
2246 }
2247 
ConstantFoldGetElementPtr(Constant * C,bool inBounds,ArrayRef<Value * > Idxs)2248 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2249                                           bool inBounds,
2250                                           ArrayRef<Value *> Idxs) {
2251   return ConstantFoldGetElementPtrImpl(
2252       cast<PointerType>(C->getType()->getScalarType())->getElementType(), C,
2253       inBounds, Idxs);
2254 }
2255 
ConstantFoldGetElementPtr(Type * Ty,Constant * C,bool inBounds,ArrayRef<Constant * > Idxs)2256 Constant *llvm::ConstantFoldGetElementPtr(Type *Ty, Constant *C,
2257                                           bool inBounds,
2258                                           ArrayRef<Constant *> Idxs) {
2259   return ConstantFoldGetElementPtrImpl(Ty, C, inBounds, Idxs);
2260 }
2261 
ConstantFoldGetElementPtr(Type * Ty,Constant * C,bool inBounds,ArrayRef<Value * > Idxs)2262 Constant *llvm::ConstantFoldGetElementPtr(Type *Ty, Constant *C,
2263                                           bool inBounds,
2264                                           ArrayRef<Value *> Idxs) {
2265   return ConstantFoldGetElementPtrImpl(Ty, C, inBounds, Idxs);
2266 }
2267