• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 TargetData, and the pieces that do. This is to avoid
16 // a dependence in VMCore on Target.
17 //
18 //===----------------------------------------------------------------------===//
19 
20 #include "ConstantFold.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Operator.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/Support/Compiler.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/GetElementPtrTypeIterator.h"
32 #include "llvm/Support/ManagedStatic.h"
33 #include "llvm/Support/MathExtras.h"
34 #include <limits>
35 using namespace llvm;
36 
37 //===----------------------------------------------------------------------===//
38 //                ConstantFold*Instruction Implementations
39 //===----------------------------------------------------------------------===//
40 
41 /// BitCastConstantVector - Convert the specified ConstantVector node to the
42 /// specified vector type.  At this point, we know that the elements of the
43 /// input vector constant are all simple integer or FP values.
BitCastConstantVector(ConstantVector * CV,VectorType * DstTy)44 static Constant *BitCastConstantVector(ConstantVector *CV,
45                                        VectorType *DstTy) {
46 
47   if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
48   if (CV->isNullValue()) return Constant::getNullValue(DstTy);
49 
50   // If this cast changes element count then we can't handle it here:
51   // doing so requires endianness information.  This should be handled by
52   // Analysis/ConstantFolding.cpp
53   unsigned NumElts = DstTy->getNumElements();
54   if (NumElts != CV->getNumOperands())
55     return 0;
56 
57   // Check to verify that all elements of the input are simple.
58   for (unsigned i = 0; i != NumElts; ++i) {
59     if (!isa<ConstantInt>(CV->getOperand(i)) &&
60         !isa<ConstantFP>(CV->getOperand(i)))
61       return 0;
62   }
63 
64   // Bitcast each element now.
65   std::vector<Constant*> Result;
66   Type *DstEltTy = DstTy->getElementType();
67   for (unsigned i = 0; i != NumElts; ++i)
68     Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i),
69                                                     DstEltTy));
70   return ConstantVector::get(Result);
71 }
72 
73 /// This function determines which opcode to use to fold two constant cast
74 /// expressions together. It uses CastInst::isEliminableCastPair to determine
75 /// the opcode. Consequently its just a wrapper around that function.
76 /// @brief Determine if it is valid to fold a cast of a cast
77 static unsigned
foldConstantCastPair(unsigned opc,ConstantExpr * Op,Type * DstTy)78 foldConstantCastPair(
79   unsigned opc,          ///< opcode of the second cast constant expression
80   ConstantExpr *Op,      ///< the first cast constant expression
81   Type *DstTy      ///< desintation type of the first cast
82 ) {
83   assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
84   assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
85   assert(CastInst::isCast(opc) && "Invalid cast opcode");
86 
87   // The the types and opcodes for the two Cast constant expressions
88   Type *SrcTy = Op->getOperand(0)->getType();
89   Type *MidTy = Op->getType();
90   Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
91   Instruction::CastOps secondOp = Instruction::CastOps(opc);
92 
93   // Let CastInst::isEliminableCastPair do the heavy lifting.
94   return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
95                                         Type::getInt64Ty(DstTy->getContext()));
96 }
97 
FoldBitCast(Constant * V,Type * DestTy)98 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
99   Type *SrcTy = V->getType();
100   if (SrcTy == DestTy)
101     return V; // no-op cast
102 
103   // Check to see if we are casting a pointer to an aggregate to a pointer to
104   // the first element.  If so, return the appropriate GEP instruction.
105   if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
106     if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
107       if (PTy->getAddressSpace() == DPTy->getAddressSpace()) {
108         SmallVector<Value*, 8> IdxList;
109         Value *Zero =
110           Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
111         IdxList.push_back(Zero);
112         Type *ElTy = PTy->getElementType();
113         while (ElTy != DPTy->getElementType()) {
114           if (StructType *STy = dyn_cast<StructType>(ElTy)) {
115             if (STy->getNumElements() == 0) break;
116             ElTy = STy->getElementType(0);
117             IdxList.push_back(Zero);
118           } else if (SequentialType *STy =
119                      dyn_cast<SequentialType>(ElTy)) {
120             if (ElTy->isPointerTy()) break;  // Can't index into pointers!
121             ElTy = STy->getElementType();
122             IdxList.push_back(Zero);
123           } else {
124             break;
125           }
126         }
127 
128         if (ElTy == DPTy->getElementType())
129           // This GEP is inbounds because all indices are zero.
130           return ConstantExpr::getInBoundsGetElementPtr(V, IdxList);
131       }
132 
133   // Handle casts from one vector constant to another.  We know that the src
134   // and dest type have the same size (otherwise its an illegal cast).
135   if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
136     if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
137       assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
138              "Not cast between same sized vectors!");
139       SrcTy = NULL;
140       // First, check for null.  Undef is already handled.
141       if (isa<ConstantAggregateZero>(V))
142         return Constant::getNullValue(DestTy);
143 
144       if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
145         return BitCastConstantVector(CV, DestPTy);
146     }
147 
148     // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
149     // This allows for other simplifications (although some of them
150     // can only be handled by Analysis/ConstantFolding.cpp).
151     if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
152       return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
153   }
154 
155   // Finally, implement bitcast folding now.   The code below doesn't handle
156   // bitcast right.
157   if (isa<ConstantPointerNull>(V))  // ptr->ptr cast.
158     return ConstantPointerNull::get(cast<PointerType>(DestTy));
159 
160   // Handle integral constant input.
161   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
162     if (DestTy->isIntegerTy())
163       // Integral -> Integral. This is a no-op because the bit widths must
164       // be the same. Consequently, we just fold to V.
165       return V;
166 
167     if (DestTy->isFloatingPointTy())
168       return ConstantFP::get(DestTy->getContext(),
169                              APFloat(CI->getValue(),
170                                      !DestTy->isPPC_FP128Ty()));
171 
172     // Otherwise, can't fold this (vector?)
173     return 0;
174   }
175 
176   // Handle ConstantFP input: FP -> Integral.
177   if (ConstantFP *FP = dyn_cast<ConstantFP>(V))
178     return ConstantInt::get(FP->getContext(),
179                             FP->getValueAPF().bitcastToAPInt());
180 
181   return 0;
182 }
183 
184 
185 /// ExtractConstantBytes - V is an integer constant which only has a subset of
186 /// its bytes used.  The bytes used are indicated by ByteStart (which is the
187 /// first byte used, counting from the least significant byte) and ByteSize,
188 /// which is the number of bytes used.
189 ///
190 /// This function analyzes the specified constant to see if the specified byte
191 /// range can be returned as a simplified constant.  If so, the constant is
192 /// returned, otherwise null is returned.
193 ///
ExtractConstantBytes(Constant * C,unsigned ByteStart,unsigned ByteSize)194 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
195                                       unsigned ByteSize) {
196   assert(C->getType()->isIntegerTy() &&
197          (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
198          "Non-byte sized integer input");
199   unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
200   assert(ByteSize && "Must be accessing some piece");
201   assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
202   assert(ByteSize != CSize && "Should not extract everything");
203 
204   // Constant Integers are simple.
205   if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
206     APInt V = CI->getValue();
207     if (ByteStart)
208       V = V.lshr(ByteStart*8);
209     V = V.trunc(ByteSize*8);
210     return ConstantInt::get(CI->getContext(), V);
211   }
212 
213   // In the input is a constant expr, we might be able to recursively simplify.
214   // If not, we definitely can't do anything.
215   ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
216   if (CE == 0) return 0;
217 
218   switch (CE->getOpcode()) {
219   default: return 0;
220   case Instruction::Or: {
221     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
222     if (RHS == 0)
223       return 0;
224 
225     // X | -1 -> -1.
226     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
227       if (RHSC->isAllOnesValue())
228         return RHSC;
229 
230     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
231     if (LHS == 0)
232       return 0;
233     return ConstantExpr::getOr(LHS, RHS);
234   }
235   case Instruction::And: {
236     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
237     if (RHS == 0)
238       return 0;
239 
240     // X & 0 -> 0.
241     if (RHS->isNullValue())
242       return RHS;
243 
244     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
245     if (LHS == 0)
246       return 0;
247     return ConstantExpr::getAnd(LHS, RHS);
248   }
249   case Instruction::LShr: {
250     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
251     if (Amt == 0)
252       return 0;
253     unsigned ShAmt = Amt->getZExtValue();
254     // Cannot analyze non-byte shifts.
255     if ((ShAmt & 7) != 0)
256       return 0;
257     ShAmt >>= 3;
258 
259     // If the extract is known to be all zeros, return zero.
260     if (ByteStart >= CSize-ShAmt)
261       return Constant::getNullValue(IntegerType::get(CE->getContext(),
262                                                      ByteSize*8));
263     // If the extract is known to be fully in the input, extract it.
264     if (ByteStart+ByteSize+ShAmt <= CSize)
265       return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
266 
267     // TODO: Handle the 'partially zero' case.
268     return 0;
269   }
270 
271   case Instruction::Shl: {
272     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
273     if (Amt == 0)
274       return 0;
275     unsigned ShAmt = Amt->getZExtValue();
276     // Cannot analyze non-byte shifts.
277     if ((ShAmt & 7) != 0)
278       return 0;
279     ShAmt >>= 3;
280 
281     // If the extract is known to be all zeros, return zero.
282     if (ByteStart+ByteSize <= ShAmt)
283       return Constant::getNullValue(IntegerType::get(CE->getContext(),
284                                                      ByteSize*8));
285     // If the extract is known to be fully in the input, extract it.
286     if (ByteStart >= ShAmt)
287       return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
288 
289     // TODO: Handle the 'partially zero' case.
290     return 0;
291   }
292 
293   case Instruction::ZExt: {
294     unsigned SrcBitSize =
295       cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
296 
297     // If extracting something that is completely zero, return 0.
298     if (ByteStart*8 >= SrcBitSize)
299       return Constant::getNullValue(IntegerType::get(CE->getContext(),
300                                                      ByteSize*8));
301 
302     // If exactly extracting the input, return it.
303     if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
304       return CE->getOperand(0);
305 
306     // If extracting something completely in the input, if if the input is a
307     // multiple of 8 bits, recurse.
308     if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
309       return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
310 
311     // Otherwise, if extracting a subset of the input, which is not multiple of
312     // 8 bits, do a shift and trunc to get the bits.
313     if ((ByteStart+ByteSize)*8 < SrcBitSize) {
314       assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
315       Constant *Res = CE->getOperand(0);
316       if (ByteStart)
317         Res = ConstantExpr::getLShr(Res,
318                                  ConstantInt::get(Res->getType(), ByteStart*8));
319       return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
320                                                           ByteSize*8));
321     }
322 
323     // TODO: Handle the 'partially zero' case.
324     return 0;
325   }
326   }
327 }
328 
329 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
330 /// on Ty, with any known factors factored out. If Folded is false,
331 /// return null if no factoring was possible, to avoid endlessly
332 /// bouncing an unfoldable expression back into the top-level folder.
333 ///
getFoldedSizeOf(Type * Ty,Type * DestTy,bool Folded)334 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy,
335                                  bool Folded) {
336   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
337     Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
338     Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
339     return ConstantExpr::getNUWMul(E, N);
340   }
341 
342   if (StructType *STy = dyn_cast<StructType>(Ty))
343     if (!STy->isPacked()) {
344       unsigned NumElems = STy->getNumElements();
345       // An empty struct has size zero.
346       if (NumElems == 0)
347         return ConstantExpr::getNullValue(DestTy);
348       // Check for a struct with all members having the same size.
349       Constant *MemberSize =
350         getFoldedSizeOf(STy->getElementType(0), DestTy, true);
351       bool AllSame = true;
352       for (unsigned i = 1; i != NumElems; ++i)
353         if (MemberSize !=
354             getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
355           AllSame = false;
356           break;
357         }
358       if (AllSame) {
359         Constant *N = ConstantInt::get(DestTy, NumElems);
360         return ConstantExpr::getNUWMul(MemberSize, N);
361       }
362     }
363 
364   // Pointer size doesn't depend on the pointee type, so canonicalize them
365   // to an arbitrary pointee.
366   if (PointerType *PTy = dyn_cast<PointerType>(Ty))
367     if (!PTy->getElementType()->isIntegerTy(1))
368       return
369         getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
370                                          PTy->getAddressSpace()),
371                         DestTy, true);
372 
373   // If there's no interesting folding happening, bail so that we don't create
374   // a constant that looks like it needs folding but really doesn't.
375   if (!Folded)
376     return 0;
377 
378   // Base case: Get a regular sizeof expression.
379   Constant *C = ConstantExpr::getSizeOf(Ty);
380   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
381                                                     DestTy, false),
382                             C, DestTy);
383   return C;
384 }
385 
386 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
387 /// on Ty, with any known factors factored out. If Folded is false,
388 /// return null if no factoring was possible, to avoid endlessly
389 /// bouncing an unfoldable expression back into the top-level folder.
390 ///
getFoldedAlignOf(Type * Ty,Type * DestTy,bool Folded)391 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy,
392                                   bool Folded) {
393   // The alignment of an array is equal to the alignment of the
394   // array element. Note that this is not always true for vectors.
395   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
396     Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
397     C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
398                                                       DestTy,
399                                                       false),
400                               C, DestTy);
401     return C;
402   }
403 
404   if (StructType *STy = dyn_cast<StructType>(Ty)) {
405     // Packed structs always have an alignment of 1.
406     if (STy->isPacked())
407       return ConstantInt::get(DestTy, 1);
408 
409     // Otherwise, struct alignment is the maximum alignment of any member.
410     // Without target data, we can't compare much, but we can check to see
411     // if all the members have the same alignment.
412     unsigned NumElems = STy->getNumElements();
413     // An empty struct has minimal alignment.
414     if (NumElems == 0)
415       return ConstantInt::get(DestTy, 1);
416     // Check for a struct with all members having the same alignment.
417     Constant *MemberAlign =
418       getFoldedAlignOf(STy->getElementType(0), DestTy, true);
419     bool AllSame = true;
420     for (unsigned i = 1; i != NumElems; ++i)
421       if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
422         AllSame = false;
423         break;
424       }
425     if (AllSame)
426       return MemberAlign;
427   }
428 
429   // Pointer alignment doesn't depend on the pointee type, so canonicalize them
430   // to an arbitrary pointee.
431   if (PointerType *PTy = dyn_cast<PointerType>(Ty))
432     if (!PTy->getElementType()->isIntegerTy(1))
433       return
434         getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
435                                                            1),
436                                           PTy->getAddressSpace()),
437                          DestTy, true);
438 
439   // If there's no interesting folding happening, bail so that we don't create
440   // a constant that looks like it needs folding but really doesn't.
441   if (!Folded)
442     return 0;
443 
444   // Base case: Get a regular alignof expression.
445   Constant *C = ConstantExpr::getAlignOf(Ty);
446   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
447                                                     DestTy, false),
448                             C, DestTy);
449   return C;
450 }
451 
452 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
453 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
454 /// return null if no factoring was possible, to avoid endlessly
455 /// bouncing an unfoldable expression back into the top-level folder.
456 ///
getFoldedOffsetOf(Type * Ty,Constant * FieldNo,Type * DestTy,bool Folded)457 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo,
458                                    Type *DestTy,
459                                    bool Folded) {
460   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
461     Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
462                                                                 DestTy, false),
463                                         FieldNo, DestTy);
464     Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
465     return ConstantExpr::getNUWMul(E, N);
466   }
467 
468   if (StructType *STy = dyn_cast<StructType>(Ty))
469     if (!STy->isPacked()) {
470       unsigned NumElems = STy->getNumElements();
471       // An empty struct has no members.
472       if (NumElems == 0)
473         return 0;
474       // Check for a struct with all members having the same size.
475       Constant *MemberSize =
476         getFoldedSizeOf(STy->getElementType(0), DestTy, true);
477       bool AllSame = true;
478       for (unsigned i = 1; i != NumElems; ++i)
479         if (MemberSize !=
480             getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
481           AllSame = false;
482           break;
483         }
484       if (AllSame) {
485         Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
486                                                                     false,
487                                                                     DestTy,
488                                                                     false),
489                                             FieldNo, DestTy);
490         return ConstantExpr::getNUWMul(MemberSize, N);
491       }
492     }
493 
494   // If there's no interesting folding happening, bail so that we don't create
495   // a constant that looks like it needs folding but really doesn't.
496   if (!Folded)
497     return 0;
498 
499   // Base case: Get a regular offsetof expression.
500   Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
501   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
502                                                     DestTy, false),
503                             C, DestTy);
504   return C;
505 }
506 
ConstantFoldCastInstruction(unsigned opc,Constant * V,Type * DestTy)507 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
508                                             Type *DestTy) {
509   if (isa<UndefValue>(V)) {
510     // zext(undef) = 0, because the top bits will be zero.
511     // sext(undef) = 0, because the top bits will all be the same.
512     // [us]itofp(undef) = 0, because the result value is bounded.
513     if (opc == Instruction::ZExt || opc == Instruction::SExt ||
514         opc == Instruction::UIToFP || opc == Instruction::SIToFP)
515       return Constant::getNullValue(DestTy);
516     return UndefValue::get(DestTy);
517   }
518 
519   // No compile-time operations on this type yet.
520   if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty())
521     return 0;
522 
523   if (V->isNullValue() && !DestTy->isX86_MMXTy())
524     return Constant::getNullValue(DestTy);
525 
526   // If the cast operand is a constant expression, there's a few things we can
527   // do to try to simplify it.
528   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
529     if (CE->isCast()) {
530       // Try hard to fold cast of cast because they are often eliminable.
531       if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
532         return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
533     } else if (CE->getOpcode() == Instruction::GetElementPtr) {
534       // If all of the indexes in the GEP are null values, there is no pointer
535       // adjustment going on.  We might as well cast the source pointer.
536       bool isAllNull = true;
537       for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
538         if (!CE->getOperand(i)->isNullValue()) {
539           isAllNull = false;
540           break;
541         }
542       if (isAllNull)
543         // This is casting one pointer type to another, always BitCast
544         return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
545     }
546   }
547 
548   // If the cast operand is a constant vector, perform the cast by
549   // operating on each element. In the cast of bitcasts, the element
550   // count may be mismatched; don't attempt to handle that here.
551   if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
552     if (DestTy->isVectorTy() &&
553         cast<VectorType>(DestTy)->getNumElements() ==
554         CV->getType()->getNumElements()) {
555       std::vector<Constant*> res;
556       VectorType *DestVecTy = cast<VectorType>(DestTy);
557       Type *DstEltTy = DestVecTy->getElementType();
558       for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i)
559         res.push_back(ConstantExpr::getCast(opc,
560                                             CV->getOperand(i), DstEltTy));
561       return ConstantVector::get(res);
562     }
563 
564   // We actually have to do a cast now. Perform the cast according to the
565   // opcode specified.
566   switch (opc) {
567   default:
568     llvm_unreachable("Failed to cast constant expression");
569   case Instruction::FPTrunc:
570   case Instruction::FPExt:
571     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
572       bool ignored;
573       APFloat Val = FPC->getValueAPF();
574       Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle :
575                   DestTy->isDoubleTy() ? APFloat::IEEEdouble :
576                   DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
577                   DestTy->isFP128Ty() ? APFloat::IEEEquad :
578                   APFloat::Bogus,
579                   APFloat::rmNearestTiesToEven, &ignored);
580       return ConstantFP::get(V->getContext(), Val);
581     }
582     return 0; // Can't fold.
583   case Instruction::FPToUI:
584   case Instruction::FPToSI:
585     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
586       const APFloat &V = FPC->getValueAPF();
587       bool ignored;
588       uint64_t x[2];
589       uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
590       (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
591                                 APFloat::rmTowardZero, &ignored);
592       APInt Val(DestBitWidth, x);
593       return ConstantInt::get(FPC->getContext(), Val);
594     }
595     return 0; // Can't fold.
596   case Instruction::IntToPtr:   //always treated as unsigned
597     if (V->isNullValue())       // Is it an integral null value?
598       return ConstantPointerNull::get(cast<PointerType>(DestTy));
599     return 0;                   // Other pointer types cannot be casted
600   case Instruction::PtrToInt:   // always treated as unsigned
601     // Is it a null pointer value?
602     if (V->isNullValue())
603       return ConstantInt::get(DestTy, 0);
604     // If this is a sizeof-like expression, pull out multiplications by
605     // known factors to expose them to subsequent folding. If it's an
606     // alignof-like expression, factor out known factors.
607     if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
608       if (CE->getOpcode() == Instruction::GetElementPtr &&
609           CE->getOperand(0)->isNullValue()) {
610         Type *Ty =
611           cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
612         if (CE->getNumOperands() == 2) {
613           // Handle a sizeof-like expression.
614           Constant *Idx = CE->getOperand(1);
615           bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
616           if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
617             Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
618                                                                 DestTy, false),
619                                         Idx, DestTy);
620             return ConstantExpr::getMul(C, Idx);
621           }
622         } else if (CE->getNumOperands() == 3 &&
623                    CE->getOperand(1)->isNullValue()) {
624           // Handle an alignof-like expression.
625           if (StructType *STy = dyn_cast<StructType>(Ty))
626             if (!STy->isPacked()) {
627               ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
628               if (CI->isOne() &&
629                   STy->getNumElements() == 2 &&
630                   STy->getElementType(0)->isIntegerTy(1)) {
631                 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
632               }
633             }
634           // Handle an offsetof-like expression.
635           if (Ty->isStructTy() || Ty->isArrayTy()) {
636             if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
637                                                 DestTy, false))
638               return C;
639           }
640         }
641       }
642     // Other pointer types cannot be casted
643     return 0;
644   case Instruction::UIToFP:
645   case Instruction::SIToFP:
646     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
647       APInt api = CI->getValue();
648       APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true);
649       (void)apf.convertFromAPInt(api,
650                                  opc==Instruction::SIToFP,
651                                  APFloat::rmNearestTiesToEven);
652       return ConstantFP::get(V->getContext(), apf);
653     }
654     return 0;
655   case Instruction::ZExt:
656     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
657       uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
658       return ConstantInt::get(V->getContext(),
659                               CI->getValue().zext(BitWidth));
660     }
661     return 0;
662   case Instruction::SExt:
663     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
664       uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
665       return ConstantInt::get(V->getContext(),
666                               CI->getValue().sext(BitWidth));
667     }
668     return 0;
669   case Instruction::Trunc: {
670     uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
671     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
672       return ConstantInt::get(V->getContext(),
673                               CI->getValue().trunc(DestBitWidth));
674     }
675 
676     // The input must be a constantexpr.  See if we can simplify this based on
677     // the bytes we are demanding.  Only do this if the source and dest are an
678     // even multiple of a byte.
679     if ((DestBitWidth & 7) == 0 &&
680         (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
681       if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
682         return Res;
683 
684     return 0;
685   }
686   case Instruction::BitCast:
687     return FoldBitCast(V, DestTy);
688   }
689 }
690 
ConstantFoldSelectInstruction(Constant * Cond,Constant * V1,Constant * V2)691 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
692                                               Constant *V1, Constant *V2) {
693   if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
694     return CB->getZExtValue() ? V1 : V2;
695 
696   // Check for zero aggregate and ConstantVector of zeros
697   if (Cond->isNullValue()) return V2;
698 
699   if (ConstantVector* CondV = dyn_cast<ConstantVector>(Cond)) {
700 
701     if (CondV->isAllOnesValue()) return V1;
702 
703     VectorType *VTy = cast<VectorType>(V1->getType());
704     ConstantVector *CP1 = dyn_cast<ConstantVector>(V1);
705     ConstantVector *CP2 = dyn_cast<ConstantVector>(V2);
706 
707     if ((CP1 || isa<ConstantAggregateZero>(V1)) &&
708         (CP2 || isa<ConstantAggregateZero>(V2))) {
709 
710       // Find the element type of the returned vector
711       Type *EltTy = VTy->getElementType();
712       unsigned NumElem = VTy->getNumElements();
713       std::vector<Constant*> Res(NumElem);
714 
715       bool Valid = true;
716       for (unsigned i = 0; i < NumElem; ++i) {
717         ConstantInt* c = dyn_cast<ConstantInt>(CondV->getOperand(i));
718         if (!c) {
719           Valid = false;
720           break;
721         }
722         Constant *C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
723         Constant *C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
724         Res[i] = c->getZExtValue() ? C1 : C2;
725       }
726       // If we were able to build the vector, return it
727       if (Valid) return ConstantVector::get(Res);
728     }
729   }
730 
731 
732   if (isa<UndefValue>(Cond)) {
733     if (isa<UndefValue>(V1)) return V1;
734     return V2;
735   }
736   if (isa<UndefValue>(V1)) return V2;
737   if (isa<UndefValue>(V2)) return V1;
738   if (V1 == V2) return V1;
739 
740   if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
741     if (TrueVal->getOpcode() == Instruction::Select)
742       if (TrueVal->getOperand(0) == Cond)
743 	return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
744   }
745   if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
746     if (FalseVal->getOpcode() == Instruction::Select)
747       if (FalseVal->getOperand(0) == Cond)
748 	return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
749   }
750 
751   return 0;
752 }
753 
ConstantFoldExtractElementInstruction(Constant * Val,Constant * Idx)754 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
755                                                       Constant *Idx) {
756   if (isa<UndefValue>(Val))  // ee(undef, x) -> undef
757     return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
758   if (Val->isNullValue())  // ee(zero, x) -> zero
759     return Constant::getNullValue(
760                           cast<VectorType>(Val->getType())->getElementType());
761 
762   if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
763     if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
764       uint64_t Index = CIdx->getZExtValue();
765       if (Index >= CVal->getNumOperands())
766         // ee({w,x,y,z}, wrong_value) -> undef
767         return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
768       return CVal->getOperand(CIdx->getZExtValue());
769     } else if (isa<UndefValue>(Idx)) {
770       // ee({w,x,y,z}, undef) -> undef
771       return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
772     }
773   }
774   return 0;
775 }
776 
ConstantFoldInsertElementInstruction(Constant * Val,Constant * Elt,Constant * Idx)777 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
778                                                      Constant *Elt,
779                                                      Constant *Idx) {
780   ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
781   if (!CIdx) return 0;
782   APInt idxVal = CIdx->getValue();
783   if (isa<UndefValue>(Val)) {
784     // Insertion of scalar constant into vector undef
785     // Optimize away insertion of undef
786     if (isa<UndefValue>(Elt))
787       return Val;
788     // Otherwise break the aggregate undef into multiple undefs and do
789     // the insertion
790     unsigned numOps =
791       cast<VectorType>(Val->getType())->getNumElements();
792     std::vector<Constant*> Ops;
793     Ops.reserve(numOps);
794     for (unsigned i = 0; i < numOps; ++i) {
795       Constant *Op =
796         (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
797       Ops.push_back(Op);
798     }
799     return ConstantVector::get(Ops);
800   }
801   if (isa<ConstantAggregateZero>(Val)) {
802     // Insertion of scalar constant into vector aggregate zero
803     // Optimize away insertion of zero
804     if (Elt->isNullValue())
805       return Val;
806     // Otherwise break the aggregate zero into multiple zeros and do
807     // the insertion
808     unsigned numOps =
809       cast<VectorType>(Val->getType())->getNumElements();
810     std::vector<Constant*> Ops;
811     Ops.reserve(numOps);
812     for (unsigned i = 0; i < numOps; ++i) {
813       Constant *Op =
814         (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
815       Ops.push_back(Op);
816     }
817     return ConstantVector::get(Ops);
818   }
819   if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
820     // Insertion of scalar constant into vector constant
821     std::vector<Constant*> Ops;
822     Ops.reserve(CVal->getNumOperands());
823     for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
824       Constant *Op =
825         (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
826       Ops.push_back(Op);
827     }
828     return ConstantVector::get(Ops);
829   }
830 
831   return 0;
832 }
833 
834 /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef
835 /// return the specified element value.  Otherwise return null.
GetVectorElement(Constant * C,unsigned EltNo)836 static Constant *GetVectorElement(Constant *C, unsigned EltNo) {
837   if (ConstantVector *CV = dyn_cast<ConstantVector>(C))
838     return CV->getOperand(EltNo);
839 
840   Type *EltTy = cast<VectorType>(C->getType())->getElementType();
841   if (isa<ConstantAggregateZero>(C))
842     return Constant::getNullValue(EltTy);
843   if (isa<UndefValue>(C))
844     return UndefValue::get(EltTy);
845   return 0;
846 }
847 
ConstantFoldShuffleVectorInstruction(Constant * V1,Constant * V2,Constant * Mask)848 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
849                                                      Constant *V2,
850                                                      Constant *Mask) {
851   // Undefined shuffle mask -> undefined value.
852   if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType());
853 
854   unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements();
855   unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements();
856   Type *EltTy = cast<VectorType>(V1->getType())->getElementType();
857 
858   // Loop over the shuffle mask, evaluating each element.
859   SmallVector<Constant*, 32> Result;
860   for (unsigned i = 0; i != MaskNumElts; ++i) {
861     Constant *InElt = GetVectorElement(Mask, i);
862     if (InElt == 0) return 0;
863 
864     if (isa<UndefValue>(InElt))
865       InElt = UndefValue::get(EltTy);
866     else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) {
867       unsigned Elt = CI->getZExtValue();
868       if (Elt >= SrcNumElts*2)
869         InElt = UndefValue::get(EltTy);
870       else if (Elt >= SrcNumElts)
871         InElt = GetVectorElement(V2, Elt - SrcNumElts);
872       else
873         InElt = GetVectorElement(V1, Elt);
874       if (InElt == 0) return 0;
875     } else {
876       // Unknown value.
877       return 0;
878     }
879     Result.push_back(InElt);
880   }
881 
882   return ConstantVector::get(Result);
883 }
884 
ConstantFoldExtractValueInstruction(Constant * Agg,ArrayRef<unsigned> Idxs)885 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
886                                                     ArrayRef<unsigned> Idxs) {
887   // Base case: no indices, so return the entire value.
888   if (Idxs.empty())
889     return Agg;
890 
891   if (isa<UndefValue>(Agg))  // ev(undef, x) -> undef
892     return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(),
893                                                             Idxs));
894 
895   if (isa<ConstantAggregateZero>(Agg))  // ev(0, x) -> 0
896     return
897       Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(),
898                                                               Idxs));
899 
900   // Otherwise recurse.
901   if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg))
902     return ConstantFoldExtractValueInstruction(CS->getOperand(Idxs[0]),
903                                                Idxs.slice(1));
904 
905   if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg))
906     return ConstantFoldExtractValueInstruction(CA->getOperand(Idxs[0]),
907                                                Idxs.slice(1));
908   ConstantVector *CV = cast<ConstantVector>(Agg);
909   return ConstantFoldExtractValueInstruction(CV->getOperand(Idxs[0]),
910                                              Idxs.slice(1));
911 }
912 
ConstantFoldInsertValueInstruction(Constant * Agg,Constant * Val,ArrayRef<unsigned> Idxs)913 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
914                                                    Constant *Val,
915                                                    ArrayRef<unsigned> Idxs) {
916   // Base case: no indices, so replace the entire value.
917   if (Idxs.empty())
918     return Val;
919 
920   if (isa<UndefValue>(Agg)) {
921     // Insertion of constant into aggregate undef
922     // Optimize away insertion of undef.
923     if (isa<UndefValue>(Val))
924       return Agg;
925 
926     // Otherwise break the aggregate undef into multiple undefs and do
927     // the insertion.
928     CompositeType *AggTy = cast<CompositeType>(Agg->getType());
929     unsigned numOps;
930     if (ArrayType *AR = dyn_cast<ArrayType>(AggTy))
931       numOps = AR->getNumElements();
932     else
933       numOps = cast<StructType>(AggTy)->getNumElements();
934 
935     std::vector<Constant*> Ops(numOps);
936     for (unsigned i = 0; i < numOps; ++i) {
937       Type *MemberTy = AggTy->getTypeAtIndex(i);
938       Constant *Op =
939         (Idxs[0] == i) ?
940         ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy),
941                                            Val, Idxs.slice(1)) :
942         UndefValue::get(MemberTy);
943       Ops[i] = Op;
944     }
945 
946     if (StructType* ST = dyn_cast<StructType>(AggTy))
947       return ConstantStruct::get(ST, Ops);
948     return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
949   }
950 
951   if (isa<ConstantAggregateZero>(Agg)) {
952     // Insertion of constant into aggregate zero
953     // Optimize away insertion of zero.
954     if (Val->isNullValue())
955       return Agg;
956 
957     // Otherwise break the aggregate zero into multiple zeros and do
958     // the insertion.
959     CompositeType *AggTy = cast<CompositeType>(Agg->getType());
960     unsigned numOps;
961     if (ArrayType *AR = dyn_cast<ArrayType>(AggTy))
962       numOps = AR->getNumElements();
963     else
964       numOps = cast<StructType>(AggTy)->getNumElements();
965 
966     std::vector<Constant*> Ops(numOps);
967     for (unsigned i = 0; i < numOps; ++i) {
968       Type *MemberTy = AggTy->getTypeAtIndex(i);
969       Constant *Op =
970         (Idxs[0] == i) ?
971         ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy),
972                                            Val, Idxs.slice(1)) :
973         Constant::getNullValue(MemberTy);
974       Ops[i] = Op;
975     }
976 
977     if (StructType *ST = dyn_cast<StructType>(AggTy))
978       return ConstantStruct::get(ST, Ops);
979     return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
980   }
981 
982   if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) {
983     // Insertion of constant into aggregate constant.
984     std::vector<Constant*> Ops(Agg->getNumOperands());
985     for (unsigned i = 0; i < Agg->getNumOperands(); ++i) {
986       Constant *Op = cast<Constant>(Agg->getOperand(i));
987       if (Idxs[0] == i)
988         Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs.slice(1));
989       Ops[i] = Op;
990     }
991 
992     if (StructType* ST = dyn_cast<StructType>(Agg->getType()))
993       return ConstantStruct::get(ST, Ops);
994     return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops);
995   }
996 
997   return 0;
998 }
999 
1000 
ConstantFoldBinaryInstruction(unsigned Opcode,Constant * C1,Constant * C2)1001 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
1002                                               Constant *C1, Constant *C2) {
1003   // No compile-time operations on this type yet.
1004   if (C1->getType()->isPPC_FP128Ty())
1005     return 0;
1006 
1007   // Handle UndefValue up front.
1008   if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1009     switch (Opcode) {
1010     case Instruction::Xor:
1011       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
1012         // Handle undef ^ undef -> 0 special case. This is a common
1013         // idiom (misuse).
1014         return Constant::getNullValue(C1->getType());
1015       // Fallthrough
1016     case Instruction::Add:
1017     case Instruction::Sub:
1018       return UndefValue::get(C1->getType());
1019     case Instruction::And:
1020       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
1021         return C1;
1022       return Constant::getNullValue(C1->getType());   // undef & X -> 0
1023     case Instruction::Mul: {
1024       ConstantInt *CI;
1025       // X * undef -> undef   if X is odd or undef
1026       if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) ||
1027           ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) ||
1028           (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
1029         return UndefValue::get(C1->getType());
1030 
1031       // X * undef -> 0       otherwise
1032       return Constant::getNullValue(C1->getType());
1033     }
1034     case Instruction::UDiv:
1035     case Instruction::SDiv:
1036       // undef / 1 -> undef
1037       if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv)
1038         if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2))
1039           if (CI2->isOne())
1040             return C1;
1041       // FALL THROUGH
1042     case Instruction::URem:
1043     case Instruction::SRem:
1044       if (!isa<UndefValue>(C2))                    // undef / X -> 0
1045         return Constant::getNullValue(C1->getType());
1046       return C2;                                   // X / undef -> undef
1047     case Instruction::Or:                          // X | undef -> -1
1048       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
1049         return C1;
1050       return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
1051     case Instruction::LShr:
1052       if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
1053         return C1;                                  // undef lshr undef -> undef
1054       return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
1055                                                     // undef lshr X -> 0
1056     case Instruction::AShr:
1057       if (!isa<UndefValue>(C2))                     // undef ashr X --> all ones
1058         return Constant::getAllOnesValue(C1->getType());
1059       else if (isa<UndefValue>(C1))
1060         return C1;                                  // undef ashr undef -> undef
1061       else
1062         return C1;                                  // X ashr undef --> X
1063     case Instruction::Shl:
1064       if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
1065         return C1;                                  // undef shl undef -> undef
1066       // undef << X -> 0   or   X << undef -> 0
1067       return Constant::getNullValue(C1->getType());
1068     }
1069   }
1070 
1071   // Handle simplifications when the RHS is a constant int.
1072   if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1073     switch (Opcode) {
1074     case Instruction::Add:
1075       if (CI2->equalsInt(0)) return C1;                         // X + 0 == X
1076       break;
1077     case Instruction::Sub:
1078       if (CI2->equalsInt(0)) return C1;                         // X - 0 == X
1079       break;
1080     case Instruction::Mul:
1081       if (CI2->equalsInt(0)) return C2;                         // X * 0 == 0
1082       if (CI2->equalsInt(1))
1083         return C1;                                              // X * 1 == X
1084       break;
1085     case Instruction::UDiv:
1086     case Instruction::SDiv:
1087       if (CI2->equalsInt(1))
1088         return C1;                                            // X / 1 == X
1089       if (CI2->equalsInt(0))
1090         return UndefValue::get(CI2->getType());               // X / 0 == undef
1091       break;
1092     case Instruction::URem:
1093     case Instruction::SRem:
1094       if (CI2->equalsInt(1))
1095         return Constant::getNullValue(CI2->getType());        // X % 1 == 0
1096       if (CI2->equalsInt(0))
1097         return UndefValue::get(CI2->getType());               // X % 0 == undef
1098       break;
1099     case Instruction::And:
1100       if (CI2->isZero()) return C2;                           // X & 0 == 0
1101       if (CI2->isAllOnesValue())
1102         return C1;                                            // X & -1 == X
1103 
1104       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1105         // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1106         if (CE1->getOpcode() == Instruction::ZExt) {
1107           unsigned DstWidth = CI2->getType()->getBitWidth();
1108           unsigned SrcWidth =
1109             CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1110           APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1111           if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1112             return C1;
1113         }
1114 
1115         // If and'ing the address of a global with a constant, fold it.
1116         if (CE1->getOpcode() == Instruction::PtrToInt &&
1117             isa<GlobalValue>(CE1->getOperand(0))) {
1118           GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1119 
1120           // Functions are at least 4-byte aligned.
1121           unsigned GVAlign = GV->getAlignment();
1122           if (isa<Function>(GV))
1123             GVAlign = std::max(GVAlign, 4U);
1124 
1125           if (GVAlign > 1) {
1126             unsigned DstWidth = CI2->getType()->getBitWidth();
1127             unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1128             APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1129 
1130             // If checking bits we know are clear, return zero.
1131             if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1132               return Constant::getNullValue(CI2->getType());
1133           }
1134         }
1135       }
1136       break;
1137     case Instruction::Or:
1138       if (CI2->equalsInt(0)) return C1;    // X | 0 == X
1139       if (CI2->isAllOnesValue())
1140         return C2;                         // X | -1 == -1
1141       break;
1142     case Instruction::Xor:
1143       if (CI2->equalsInt(0)) return C1;    // X ^ 0 == X
1144 
1145       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1146         switch (CE1->getOpcode()) {
1147         default: break;
1148         case Instruction::ICmp:
1149         case Instruction::FCmp:
1150           // cmp pred ^ true -> cmp !pred
1151           assert(CI2->equalsInt(1));
1152           CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1153           pred = CmpInst::getInversePredicate(pred);
1154           return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1155                                           CE1->getOperand(1));
1156         }
1157       }
1158       break;
1159     case Instruction::AShr:
1160       // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1161       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1162         if (CE1->getOpcode() == Instruction::ZExt)  // Top bits known zero.
1163           return ConstantExpr::getLShr(C1, C2);
1164       break;
1165     }
1166   } else if (isa<ConstantInt>(C1)) {
1167     // If C1 is a ConstantInt and C2 is not, swap the operands.
1168     if (Instruction::isCommutative(Opcode))
1169       return ConstantExpr::get(Opcode, C2, C1);
1170   }
1171 
1172   // At this point we know neither constant is an UndefValue.
1173   if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1174     if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1175       using namespace APIntOps;
1176       const APInt &C1V = CI1->getValue();
1177       const APInt &C2V = CI2->getValue();
1178       switch (Opcode) {
1179       default:
1180         break;
1181       case Instruction::Add:
1182         return ConstantInt::get(CI1->getContext(), C1V + C2V);
1183       case Instruction::Sub:
1184         return ConstantInt::get(CI1->getContext(), C1V - C2V);
1185       case Instruction::Mul:
1186         return ConstantInt::get(CI1->getContext(), C1V * C2V);
1187       case Instruction::UDiv:
1188         assert(!CI2->isNullValue() && "Div by zero handled above");
1189         return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1190       case Instruction::SDiv:
1191         assert(!CI2->isNullValue() && "Div by zero handled above");
1192         if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1193           return UndefValue::get(CI1->getType());   // MIN_INT / -1 -> undef
1194         return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1195       case Instruction::URem:
1196         assert(!CI2->isNullValue() && "Div by zero handled above");
1197         return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1198       case Instruction::SRem:
1199         assert(!CI2->isNullValue() && "Div by zero handled above");
1200         if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1201           return UndefValue::get(CI1->getType());   // MIN_INT % -1 -> undef
1202         return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1203       case Instruction::And:
1204         return ConstantInt::get(CI1->getContext(), C1V & C2V);
1205       case Instruction::Or:
1206         return ConstantInt::get(CI1->getContext(), C1V | C2V);
1207       case Instruction::Xor:
1208         return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1209       case Instruction::Shl: {
1210         uint32_t shiftAmt = C2V.getZExtValue();
1211         if (shiftAmt < C1V.getBitWidth())
1212           return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1213         else
1214           return UndefValue::get(C1->getType()); // too big shift is undef
1215       }
1216       case Instruction::LShr: {
1217         uint32_t shiftAmt = C2V.getZExtValue();
1218         if (shiftAmt < C1V.getBitWidth())
1219           return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1220         else
1221           return UndefValue::get(C1->getType()); // too big shift is undef
1222       }
1223       case Instruction::AShr: {
1224         uint32_t shiftAmt = C2V.getZExtValue();
1225         if (shiftAmt < C1V.getBitWidth())
1226           return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1227         else
1228           return UndefValue::get(C1->getType()); // too big shift is undef
1229       }
1230       }
1231     }
1232 
1233     switch (Opcode) {
1234     case Instruction::SDiv:
1235     case Instruction::UDiv:
1236     case Instruction::URem:
1237     case Instruction::SRem:
1238     case Instruction::LShr:
1239     case Instruction::AShr:
1240     case Instruction::Shl:
1241       if (CI1->equalsInt(0)) return C1;
1242       break;
1243     default:
1244       break;
1245     }
1246   } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1247     if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1248       APFloat C1V = CFP1->getValueAPF();
1249       APFloat C2V = CFP2->getValueAPF();
1250       APFloat C3V = C1V;  // copy for modification
1251       switch (Opcode) {
1252       default:
1253         break;
1254       case Instruction::FAdd:
1255         (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1256         return ConstantFP::get(C1->getContext(), C3V);
1257       case Instruction::FSub:
1258         (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1259         return ConstantFP::get(C1->getContext(), C3V);
1260       case Instruction::FMul:
1261         (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1262         return ConstantFP::get(C1->getContext(), C3V);
1263       case Instruction::FDiv:
1264         (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1265         return ConstantFP::get(C1->getContext(), C3V);
1266       case Instruction::FRem:
1267         (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1268         return ConstantFP::get(C1->getContext(), C3V);
1269       }
1270     }
1271   } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1272     ConstantVector *CP1 = dyn_cast<ConstantVector>(C1);
1273     ConstantVector *CP2 = dyn_cast<ConstantVector>(C2);
1274     if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) &&
1275         (CP2 != NULL || isa<ConstantAggregateZero>(C2))) {
1276       std::vector<Constant*> Res;
1277       Type* EltTy = VTy->getElementType();
1278       Constant *C1 = 0;
1279       Constant *C2 = 0;
1280       switch (Opcode) {
1281       default:
1282         break;
1283       case Instruction::Add:
1284         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1285           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1286           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1287           Res.push_back(ConstantExpr::getAdd(C1, C2));
1288         }
1289         return ConstantVector::get(Res);
1290       case Instruction::FAdd:
1291         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1292           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1293           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1294           Res.push_back(ConstantExpr::getFAdd(C1, C2));
1295         }
1296         return ConstantVector::get(Res);
1297       case Instruction::Sub:
1298         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1299           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1300           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1301           Res.push_back(ConstantExpr::getSub(C1, C2));
1302         }
1303         return ConstantVector::get(Res);
1304       case Instruction::FSub:
1305         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1306           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1307           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1308           Res.push_back(ConstantExpr::getFSub(C1, C2));
1309         }
1310         return ConstantVector::get(Res);
1311       case Instruction::Mul:
1312         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1313           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1314           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1315           Res.push_back(ConstantExpr::getMul(C1, C2));
1316         }
1317         return ConstantVector::get(Res);
1318       case Instruction::FMul:
1319         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1320           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1321           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1322           Res.push_back(ConstantExpr::getFMul(C1, C2));
1323         }
1324         return ConstantVector::get(Res);
1325       case Instruction::UDiv:
1326         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1327           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1328           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1329           Res.push_back(ConstantExpr::getUDiv(C1, C2));
1330         }
1331         return ConstantVector::get(Res);
1332       case Instruction::SDiv:
1333         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1334           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1335           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1336           Res.push_back(ConstantExpr::getSDiv(C1, C2));
1337         }
1338         return ConstantVector::get(Res);
1339       case Instruction::FDiv:
1340         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1341           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1342           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1343           Res.push_back(ConstantExpr::getFDiv(C1, C2));
1344         }
1345         return ConstantVector::get(Res);
1346       case Instruction::URem:
1347         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1348           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1349           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1350           Res.push_back(ConstantExpr::getURem(C1, C2));
1351         }
1352         return ConstantVector::get(Res);
1353       case Instruction::SRem:
1354         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1355           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1356           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1357           Res.push_back(ConstantExpr::getSRem(C1, C2));
1358         }
1359         return ConstantVector::get(Res);
1360       case Instruction::FRem:
1361         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1362           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1363           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1364           Res.push_back(ConstantExpr::getFRem(C1, C2));
1365         }
1366         return ConstantVector::get(Res);
1367       case Instruction::And:
1368         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1369           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1370           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1371           Res.push_back(ConstantExpr::getAnd(C1, C2));
1372         }
1373         return ConstantVector::get(Res);
1374       case Instruction::Or:
1375         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1376           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1377           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1378           Res.push_back(ConstantExpr::getOr(C1, C2));
1379         }
1380         return ConstantVector::get(Res);
1381       case Instruction::Xor:
1382         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1383           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1384           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1385           Res.push_back(ConstantExpr::getXor(C1, C2));
1386         }
1387         return ConstantVector::get(Res);
1388       case Instruction::LShr:
1389         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1390           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1391           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1392           Res.push_back(ConstantExpr::getLShr(C1, C2));
1393         }
1394         return ConstantVector::get(Res);
1395       case Instruction::AShr:
1396         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1397           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1398           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1399           Res.push_back(ConstantExpr::getAShr(C1, C2));
1400         }
1401         return ConstantVector::get(Res);
1402       case Instruction::Shl:
1403         for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1404           C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1405           C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1406           Res.push_back(ConstantExpr::getShl(C1, C2));
1407         }
1408         return ConstantVector::get(Res);
1409       }
1410     }
1411   }
1412 
1413   if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1414     // There are many possible foldings we could do here.  We should probably
1415     // at least fold add of a pointer with an integer into the appropriate
1416     // getelementptr.  This will improve alias analysis a bit.
1417 
1418     // Given ((a + b) + c), if (b + c) folds to something interesting, return
1419     // (a + (b + c)).
1420     if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1421       Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1422       if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1423         return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1424     }
1425   } else if (isa<ConstantExpr>(C2)) {
1426     // If C2 is a constant expr and C1 isn't, flop them around and fold the
1427     // other way if possible.
1428     if (Instruction::isCommutative(Opcode))
1429       return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1430   }
1431 
1432   // i1 can be simplified in many cases.
1433   if (C1->getType()->isIntegerTy(1)) {
1434     switch (Opcode) {
1435     case Instruction::Add:
1436     case Instruction::Sub:
1437       return ConstantExpr::getXor(C1, C2);
1438     case Instruction::Mul:
1439       return ConstantExpr::getAnd(C1, C2);
1440     case Instruction::Shl:
1441     case Instruction::LShr:
1442     case Instruction::AShr:
1443       // We can assume that C2 == 0.  If it were one the result would be
1444       // undefined because the shift value is as large as the bitwidth.
1445       return C1;
1446     case Instruction::SDiv:
1447     case Instruction::UDiv:
1448       // We can assume that C2 == 1.  If it were zero the result would be
1449       // undefined through division by zero.
1450       return C1;
1451     case Instruction::URem:
1452     case Instruction::SRem:
1453       // We can assume that C2 == 1.  If it were zero the result would be
1454       // undefined through division by zero.
1455       return ConstantInt::getFalse(C1->getContext());
1456     default:
1457       break;
1458     }
1459   }
1460 
1461   // We don't know how to fold this.
1462   return 0;
1463 }
1464 
1465 /// isZeroSizedType - This type is zero sized if its an array or structure of
1466 /// zero sized types.  The only leaf zero sized type is an empty structure.
isMaybeZeroSizedType(Type * Ty)1467 static bool isMaybeZeroSizedType(Type *Ty) {
1468   if (StructType *STy = dyn_cast<StructType>(Ty)) {
1469     if (STy->isOpaque()) return true;  // Can't say.
1470 
1471     // If all of elements have zero size, this does too.
1472     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1473       if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1474     return true;
1475 
1476   } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1477     return isMaybeZeroSizedType(ATy->getElementType());
1478   }
1479   return false;
1480 }
1481 
1482 /// IdxCompare - Compare the two constants as though they were getelementptr
1483 /// indices.  This allows coersion of the types to be the same thing.
1484 ///
1485 /// If the two constants are the "same" (after coersion), return 0.  If the
1486 /// first is less than the second, return -1, if the second is less than the
1487 /// first, return 1.  If the constants are not integral, return -2.
1488 ///
IdxCompare(Constant * C1,Constant * C2,Type * ElTy)1489 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1490   if (C1 == C2) return 0;
1491 
1492   // Ok, we found a different index.  If they are not ConstantInt, we can't do
1493   // anything with them.
1494   if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1495     return -2; // don't know!
1496 
1497   // Ok, we have two differing integer indices.  Sign extend them to be the same
1498   // type.  Long is always big enough, so we use it.
1499   if (!C1->getType()->isIntegerTy(64))
1500     C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1501 
1502   if (!C2->getType()->isIntegerTy(64))
1503     C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1504 
1505   if (C1 == C2) return 0;  // They are equal
1506 
1507   // If the type being indexed over is really just a zero sized type, there is
1508   // no pointer difference being made here.
1509   if (isMaybeZeroSizedType(ElTy))
1510     return -2; // dunno.
1511 
1512   // If they are really different, now that they are the same type, then we
1513   // found a difference!
1514   if (cast<ConstantInt>(C1)->getSExtValue() <
1515       cast<ConstantInt>(C2)->getSExtValue())
1516     return -1;
1517   else
1518     return 1;
1519 }
1520 
1521 /// evaluateFCmpRelation - This function determines if there is anything we can
1522 /// decide about the two constants provided.  This doesn't need to handle simple
1523 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1524 /// If we can determine that the two constants have a particular relation to
1525 /// each other, we should return the corresponding FCmpInst predicate,
1526 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1527 /// ConstantFoldCompareInstruction.
1528 ///
1529 /// To simplify this code we canonicalize the relation so that the first
1530 /// operand is always the most "complex" of the two.  We consider ConstantFP
1531 /// to be the simplest, and ConstantExprs to be the most complex.
evaluateFCmpRelation(Constant * V1,Constant * V2)1532 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1533   assert(V1->getType() == V2->getType() &&
1534          "Cannot compare values of different types!");
1535 
1536   // No compile-time operations on this type yet.
1537   if (V1->getType()->isPPC_FP128Ty())
1538     return FCmpInst::BAD_FCMP_PREDICATE;
1539 
1540   // Handle degenerate case quickly
1541   if (V1 == V2) return FCmpInst::FCMP_OEQ;
1542 
1543   if (!isa<ConstantExpr>(V1)) {
1544     if (!isa<ConstantExpr>(V2)) {
1545       // We distilled thisUse the standard constant folder for a few cases
1546       ConstantInt *R = 0;
1547       R = dyn_cast<ConstantInt>(
1548                       ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1549       if (R && !R->isZero())
1550         return FCmpInst::FCMP_OEQ;
1551       R = dyn_cast<ConstantInt>(
1552                       ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1553       if (R && !R->isZero())
1554         return FCmpInst::FCMP_OLT;
1555       R = dyn_cast<ConstantInt>(
1556                       ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1557       if (R && !R->isZero())
1558         return FCmpInst::FCMP_OGT;
1559 
1560       // Nothing more we can do
1561       return FCmpInst::BAD_FCMP_PREDICATE;
1562     }
1563 
1564     // If the first operand is simple and second is ConstantExpr, swap operands.
1565     FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1566     if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1567       return FCmpInst::getSwappedPredicate(SwappedRelation);
1568   } else {
1569     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1570     // constantexpr or a simple constant.
1571     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1572     switch (CE1->getOpcode()) {
1573     case Instruction::FPTrunc:
1574     case Instruction::FPExt:
1575     case Instruction::UIToFP:
1576     case Instruction::SIToFP:
1577       // We might be able to do something with these but we don't right now.
1578       break;
1579     default:
1580       break;
1581     }
1582   }
1583   // There are MANY other foldings that we could perform here.  They will
1584   // probably be added on demand, as they seem needed.
1585   return FCmpInst::BAD_FCMP_PREDICATE;
1586 }
1587 
1588 /// evaluateICmpRelation - This function determines if there is anything we can
1589 /// decide about the two constants provided.  This doesn't need to handle simple
1590 /// things like integer comparisons, but should instead handle ConstantExprs
1591 /// and GlobalValues.  If we can determine that the two constants have a
1592 /// particular relation to each other, we should return the corresponding ICmp
1593 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1594 ///
1595 /// To simplify this code we canonicalize the relation so that the first
1596 /// operand is always the most "complex" of the two.  We consider simple
1597 /// constants (like ConstantInt) to be the simplest, followed by
1598 /// GlobalValues, followed by ConstantExpr's (the most complex).
1599 ///
evaluateICmpRelation(Constant * V1,Constant * V2,bool isSigned)1600 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1601                                                 bool isSigned) {
1602   assert(V1->getType() == V2->getType() &&
1603          "Cannot compare different types of values!");
1604   if (V1 == V2) return ICmpInst::ICMP_EQ;
1605 
1606   if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1607       !isa<BlockAddress>(V1)) {
1608     if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1609         !isa<BlockAddress>(V2)) {
1610       // We distilled this down to a simple case, use the standard constant
1611       // folder.
1612       ConstantInt *R = 0;
1613       ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1614       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1615       if (R && !R->isZero())
1616         return pred;
1617       pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1618       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1619       if (R && !R->isZero())
1620         return pred;
1621       pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1622       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1623       if (R && !R->isZero())
1624         return pred;
1625 
1626       // If we couldn't figure it out, bail.
1627       return ICmpInst::BAD_ICMP_PREDICATE;
1628     }
1629 
1630     // If the first operand is simple, swap operands.
1631     ICmpInst::Predicate SwappedRelation =
1632       evaluateICmpRelation(V2, V1, isSigned);
1633     if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1634       return ICmpInst::getSwappedPredicate(SwappedRelation);
1635 
1636   } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1637     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1638       ICmpInst::Predicate SwappedRelation =
1639         evaluateICmpRelation(V2, V1, isSigned);
1640       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1641         return ICmpInst::getSwappedPredicate(SwappedRelation);
1642       return ICmpInst::BAD_ICMP_PREDICATE;
1643     }
1644 
1645     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1646     // constant (which, since the types must match, means that it's a
1647     // ConstantPointerNull).
1648     if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1649       // Don't try to decide equality of aliases.
1650       if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2))
1651         if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1652           return ICmpInst::ICMP_NE;
1653     } else if (isa<BlockAddress>(V2)) {
1654       return ICmpInst::ICMP_NE; // Globals never equal labels.
1655     } else {
1656       assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1657       // GlobalVals can never be null unless they have external weak linkage.
1658       // We don't try to evaluate aliases here.
1659       if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1660         return ICmpInst::ICMP_NE;
1661     }
1662   } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1663     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1664       ICmpInst::Predicate SwappedRelation =
1665         evaluateICmpRelation(V2, V1, isSigned);
1666       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1667         return ICmpInst::getSwappedPredicate(SwappedRelation);
1668       return ICmpInst::BAD_ICMP_PREDICATE;
1669     }
1670 
1671     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1672     // constant (which, since the types must match, means that it is a
1673     // ConstantPointerNull).
1674     if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1675       // Block address in another function can't equal this one, but block
1676       // addresses in the current function might be the same if blocks are
1677       // empty.
1678       if (BA2->getFunction() != BA->getFunction())
1679         return ICmpInst::ICMP_NE;
1680     } else {
1681       // Block addresses aren't null, don't equal the address of globals.
1682       assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1683              "Canonicalization guarantee!");
1684       return ICmpInst::ICMP_NE;
1685     }
1686   } else {
1687     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1688     // constantexpr, a global, block address, or a simple constant.
1689     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1690     Constant *CE1Op0 = CE1->getOperand(0);
1691 
1692     switch (CE1->getOpcode()) {
1693     case Instruction::Trunc:
1694     case Instruction::FPTrunc:
1695     case Instruction::FPExt:
1696     case Instruction::FPToUI:
1697     case Instruction::FPToSI:
1698       break; // We can't evaluate floating point casts or truncations.
1699 
1700     case Instruction::UIToFP:
1701     case Instruction::SIToFP:
1702     case Instruction::BitCast:
1703     case Instruction::ZExt:
1704     case Instruction::SExt:
1705       // If the cast is not actually changing bits, and the second operand is a
1706       // null pointer, do the comparison with the pre-casted value.
1707       if (V2->isNullValue() &&
1708           (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1709         if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1710         if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1711         return evaluateICmpRelation(CE1Op0,
1712                                     Constant::getNullValue(CE1Op0->getType()),
1713                                     isSigned);
1714       }
1715       break;
1716 
1717     case Instruction::GetElementPtr:
1718       // Ok, since this is a getelementptr, we know that the constant has a
1719       // pointer type.  Check the various cases.
1720       if (isa<ConstantPointerNull>(V2)) {
1721         // If we are comparing a GEP to a null pointer, check to see if the base
1722         // of the GEP equals the null pointer.
1723         if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1724           if (GV->hasExternalWeakLinkage())
1725             // Weak linkage GVals could be zero or not. We're comparing that
1726             // to null pointer so its greater-or-equal
1727             return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1728           else
1729             // If its not weak linkage, the GVal must have a non-zero address
1730             // so the result is greater-than
1731             return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1732         } else if (isa<ConstantPointerNull>(CE1Op0)) {
1733           // If we are indexing from a null pointer, check to see if we have any
1734           // non-zero indices.
1735           for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1736             if (!CE1->getOperand(i)->isNullValue())
1737               // Offsetting from null, must not be equal.
1738               return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1739           // Only zero indexes from null, must still be zero.
1740           return ICmpInst::ICMP_EQ;
1741         }
1742         // Otherwise, we can't really say if the first operand is null or not.
1743       } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1744         if (isa<ConstantPointerNull>(CE1Op0)) {
1745           if (GV2->hasExternalWeakLinkage())
1746             // Weak linkage GVals could be zero or not. We're comparing it to
1747             // a null pointer, so its less-or-equal
1748             return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1749           else
1750             // If its not weak linkage, the GVal must have a non-zero address
1751             // so the result is less-than
1752             return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1753         } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1754           if (GV == GV2) {
1755             // If this is a getelementptr of the same global, then it must be
1756             // different.  Because the types must match, the getelementptr could
1757             // only have at most one index, and because we fold getelementptr's
1758             // with a single zero index, it must be nonzero.
1759             assert(CE1->getNumOperands() == 2 &&
1760                    !CE1->getOperand(1)->isNullValue() &&
1761                    "Surprising getelementptr!");
1762             return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1763           } else {
1764             // If they are different globals, we don't know what the value is,
1765             // but they can't be equal.
1766             return ICmpInst::ICMP_NE;
1767           }
1768         }
1769       } else {
1770         ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1771         Constant *CE2Op0 = CE2->getOperand(0);
1772 
1773         // There are MANY other foldings that we could perform here.  They will
1774         // probably be added on demand, as they seem needed.
1775         switch (CE2->getOpcode()) {
1776         default: break;
1777         case Instruction::GetElementPtr:
1778           // By far the most common case to handle is when the base pointers are
1779           // obviously to the same or different globals.
1780           if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1781             if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1782               return ICmpInst::ICMP_NE;
1783             // Ok, we know that both getelementptr instructions are based on the
1784             // same global.  From this, we can precisely determine the relative
1785             // ordering of the resultant pointers.
1786             unsigned i = 1;
1787 
1788             // The logic below assumes that the result of the comparison
1789             // can be determined by finding the first index that differs.
1790             // This doesn't work if there is over-indexing in any
1791             // subsequent indices, so check for that case first.
1792             if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1793                 !CE2->isGEPWithNoNotionalOverIndexing())
1794                return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1795 
1796             // Compare all of the operands the GEP's have in common.
1797             gep_type_iterator GTI = gep_type_begin(CE1);
1798             for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1799                  ++i, ++GTI)
1800               switch (IdxCompare(CE1->getOperand(i),
1801                                  CE2->getOperand(i), GTI.getIndexedType())) {
1802               case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1803               case 1:  return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1804               case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1805               }
1806 
1807             // Ok, we ran out of things they have in common.  If any leftovers
1808             // are non-zero then we have a difference, otherwise we are equal.
1809             for (; i < CE1->getNumOperands(); ++i)
1810               if (!CE1->getOperand(i)->isNullValue()) {
1811                 if (isa<ConstantInt>(CE1->getOperand(i)))
1812                   return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1813                 else
1814                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1815               }
1816 
1817             for (; i < CE2->getNumOperands(); ++i)
1818               if (!CE2->getOperand(i)->isNullValue()) {
1819                 if (isa<ConstantInt>(CE2->getOperand(i)))
1820                   return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1821                 else
1822                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1823               }
1824             return ICmpInst::ICMP_EQ;
1825           }
1826         }
1827       }
1828     default:
1829       break;
1830     }
1831   }
1832 
1833   return ICmpInst::BAD_ICMP_PREDICATE;
1834 }
1835 
ConstantFoldCompareInstruction(unsigned short pred,Constant * C1,Constant * C2)1836 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1837                                                Constant *C1, Constant *C2) {
1838   Type *ResultTy;
1839   if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1840     ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1841                                VT->getNumElements());
1842   else
1843     ResultTy = Type::getInt1Ty(C1->getContext());
1844 
1845   // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1846   if (pred == FCmpInst::FCMP_FALSE)
1847     return Constant::getNullValue(ResultTy);
1848 
1849   if (pred == FCmpInst::FCMP_TRUE)
1850     return Constant::getAllOnesValue(ResultTy);
1851 
1852   // Handle some degenerate cases first
1853   if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1854     // For EQ and NE, we can always pick a value for the undef to make the
1855     // predicate pass or fail, so we can return undef.
1856     // Also, if both operands are undef, we can return undef.
1857     if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) ||
1858         (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
1859       return UndefValue::get(ResultTy);
1860     // Otherwise, pick the same value as the non-undef operand, and fold
1861     // it to true or false.
1862     return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred));
1863   }
1864 
1865   // No compile-time operations on this type yet.
1866   if (C1->getType()->isPPC_FP128Ty())
1867     return 0;
1868 
1869   // icmp eq/ne(null,GV) -> false/true
1870   if (C1->isNullValue()) {
1871     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1872       // Don't try to evaluate aliases.  External weak GV can be null.
1873       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1874         if (pred == ICmpInst::ICMP_EQ)
1875           return ConstantInt::getFalse(C1->getContext());
1876         else if (pred == ICmpInst::ICMP_NE)
1877           return ConstantInt::getTrue(C1->getContext());
1878       }
1879   // icmp eq/ne(GV,null) -> false/true
1880   } else if (C2->isNullValue()) {
1881     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1882       // Don't try to evaluate aliases.  External weak GV can be null.
1883       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1884         if (pred == ICmpInst::ICMP_EQ)
1885           return ConstantInt::getFalse(C1->getContext());
1886         else if (pred == ICmpInst::ICMP_NE)
1887           return ConstantInt::getTrue(C1->getContext());
1888       }
1889   }
1890 
1891   // If the comparison is a comparison between two i1's, simplify it.
1892   if (C1->getType()->isIntegerTy(1)) {
1893     switch(pred) {
1894     case ICmpInst::ICMP_EQ:
1895       if (isa<ConstantInt>(C2))
1896         return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1897       return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1898     case ICmpInst::ICMP_NE:
1899       return ConstantExpr::getXor(C1, C2);
1900     default:
1901       break;
1902     }
1903   }
1904 
1905   if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1906     APInt V1 = cast<ConstantInt>(C1)->getValue();
1907     APInt V2 = cast<ConstantInt>(C2)->getValue();
1908     switch (pred) {
1909     default: llvm_unreachable("Invalid ICmp Predicate"); return 0;
1910     case ICmpInst::ICMP_EQ:  return ConstantInt::get(ResultTy, V1 == V2);
1911     case ICmpInst::ICMP_NE:  return ConstantInt::get(ResultTy, V1 != V2);
1912     case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1913     case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1914     case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1915     case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1916     case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1917     case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1918     case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1919     case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1920     }
1921   } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1922     APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1923     APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1924     APFloat::cmpResult R = C1V.compare(C2V);
1925     switch (pred) {
1926     default: llvm_unreachable("Invalid FCmp Predicate"); return 0;
1927     case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1928     case FCmpInst::FCMP_TRUE:  return Constant::getAllOnesValue(ResultTy);
1929     case FCmpInst::FCMP_UNO:
1930       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1931     case FCmpInst::FCMP_ORD:
1932       return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1933     case FCmpInst::FCMP_UEQ:
1934       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1935                                         R==APFloat::cmpEqual);
1936     case FCmpInst::FCMP_OEQ:
1937       return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1938     case FCmpInst::FCMP_UNE:
1939       return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1940     case FCmpInst::FCMP_ONE:
1941       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1942                                         R==APFloat::cmpGreaterThan);
1943     case FCmpInst::FCMP_ULT:
1944       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1945                                         R==APFloat::cmpLessThan);
1946     case FCmpInst::FCMP_OLT:
1947       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1948     case FCmpInst::FCMP_UGT:
1949       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1950                                         R==APFloat::cmpGreaterThan);
1951     case FCmpInst::FCMP_OGT:
1952       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1953     case FCmpInst::FCMP_ULE:
1954       return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1955     case FCmpInst::FCMP_OLE:
1956       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1957                                         R==APFloat::cmpEqual);
1958     case FCmpInst::FCMP_UGE:
1959       return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1960     case FCmpInst::FCMP_OGE:
1961       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1962                                         R==APFloat::cmpEqual);
1963     }
1964   } else if (C1->getType()->isVectorTy()) {
1965     SmallVector<Constant*, 16> C1Elts, C2Elts;
1966     C1->getVectorElements(C1Elts);
1967     C2->getVectorElements(C2Elts);
1968     if (C1Elts.empty() || C2Elts.empty())
1969       return 0;
1970 
1971     // If we can constant fold the comparison of each element, constant fold
1972     // the whole vector comparison.
1973     SmallVector<Constant*, 4> ResElts;
1974     // Compare the elements, producing an i1 result or constant expr.
1975     for (unsigned i = 0, e = C1Elts.size(); i != e; ++i)
1976       ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i]));
1977 
1978     return ConstantVector::get(ResElts);
1979   }
1980 
1981   if (C1->getType()->isFloatingPointTy()) {
1982     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
1983     switch (evaluateFCmpRelation(C1, C2)) {
1984     default: llvm_unreachable("Unknown relation!");
1985     case FCmpInst::FCMP_UNO:
1986     case FCmpInst::FCMP_ORD:
1987     case FCmpInst::FCMP_UEQ:
1988     case FCmpInst::FCMP_UNE:
1989     case FCmpInst::FCMP_ULT:
1990     case FCmpInst::FCMP_UGT:
1991     case FCmpInst::FCMP_ULE:
1992     case FCmpInst::FCMP_UGE:
1993     case FCmpInst::FCMP_TRUE:
1994     case FCmpInst::FCMP_FALSE:
1995     case FCmpInst::BAD_FCMP_PREDICATE:
1996       break; // Couldn't determine anything about these constants.
1997     case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1998       Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1999                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
2000                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
2001       break;
2002     case FCmpInst::FCMP_OLT: // We know that C1 < C2
2003       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
2004                 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
2005                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
2006       break;
2007     case FCmpInst::FCMP_OGT: // We know that C1 > C2
2008       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
2009                 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
2010                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
2011       break;
2012     case FCmpInst::FCMP_OLE: // We know that C1 <= C2
2013       // We can only partially decide this relation.
2014       if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
2015         Result = 0;
2016       else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
2017         Result = 1;
2018       break;
2019     case FCmpInst::FCMP_OGE: // We known that C1 >= C2
2020       // We can only partially decide this relation.
2021       if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
2022         Result = 0;
2023       else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
2024         Result = 1;
2025       break;
2026     case FCmpInst::FCMP_ONE: // We know that C1 != C2
2027       // We can only partially decide this relation.
2028       if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
2029         Result = 0;
2030       else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
2031         Result = 1;
2032       break;
2033     }
2034 
2035     // If we evaluated the result, return it now.
2036     if (Result != -1)
2037       return ConstantInt::get(ResultTy, Result);
2038 
2039   } else {
2040     // Evaluate the relation between the two constants, per the predicate.
2041     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
2042     switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
2043     default: llvm_unreachable("Unknown relational!");
2044     case ICmpInst::BAD_ICMP_PREDICATE:
2045       break;  // Couldn't determine anything about these constants.
2046     case ICmpInst::ICMP_EQ:   // We know the constants are equal!
2047       // If we know the constants are equal, we can decide the result of this
2048       // computation precisely.
2049       Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
2050       break;
2051     case ICmpInst::ICMP_ULT:
2052       switch (pred) {
2053       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
2054         Result = 1; break;
2055       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
2056         Result = 0; break;
2057       }
2058       break;
2059     case ICmpInst::ICMP_SLT:
2060       switch (pred) {
2061       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
2062         Result = 1; break;
2063       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
2064         Result = 0; break;
2065       }
2066       break;
2067     case ICmpInst::ICMP_UGT:
2068       switch (pred) {
2069       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
2070         Result = 1; break;
2071       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
2072         Result = 0; break;
2073       }
2074       break;
2075     case ICmpInst::ICMP_SGT:
2076       switch (pred) {
2077       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
2078         Result = 1; break;
2079       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
2080         Result = 0; break;
2081       }
2082       break;
2083     case ICmpInst::ICMP_ULE:
2084       if (pred == ICmpInst::ICMP_UGT) Result = 0;
2085       if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
2086       break;
2087     case ICmpInst::ICMP_SLE:
2088       if (pred == ICmpInst::ICMP_SGT) Result = 0;
2089       if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
2090       break;
2091     case ICmpInst::ICMP_UGE:
2092       if (pred == ICmpInst::ICMP_ULT) Result = 0;
2093       if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2094       break;
2095     case ICmpInst::ICMP_SGE:
2096       if (pred == ICmpInst::ICMP_SLT) Result = 0;
2097       if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2098       break;
2099     case ICmpInst::ICMP_NE:
2100       if (pred == ICmpInst::ICMP_EQ) Result = 0;
2101       if (pred == ICmpInst::ICMP_NE) Result = 1;
2102       break;
2103     }
2104 
2105     // If we evaluated the result, return it now.
2106     if (Result != -1)
2107       return ConstantInt::get(ResultTy, Result);
2108 
2109     // If the right hand side is a bitcast, try using its inverse to simplify
2110     // it by moving it to the left hand side.  We can't do this if it would turn
2111     // a vector compare into a scalar compare or visa versa.
2112     if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2113       Constant *CE2Op0 = CE2->getOperand(0);
2114       if (CE2->getOpcode() == Instruction::BitCast &&
2115           CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
2116         Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2117         return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2118       }
2119     }
2120 
2121     // If the left hand side is an extension, try eliminating it.
2122     if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2123       if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) ||
2124           (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){
2125         Constant *CE1Op0 = CE1->getOperand(0);
2126         Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2127         if (CE1Inverse == CE1Op0) {
2128           // Check whether we can safely truncate the right hand side.
2129           Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2130           if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) {
2131             return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2132           }
2133         }
2134       }
2135     }
2136 
2137     if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2138         (C1->isNullValue() && !C2->isNullValue())) {
2139       // If C2 is a constant expr and C1 isn't, flip them around and fold the
2140       // other way if possible.
2141       // Also, if C1 is null and C2 isn't, flip them around.
2142       pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2143       return ConstantExpr::getICmp(pred, C2, C1);
2144     }
2145   }
2146   return 0;
2147 }
2148 
2149 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
2150 /// is "inbounds".
2151 template<typename IndexTy>
isInBoundsIndices(ArrayRef<IndexTy> Idxs)2152 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
2153   // No indices means nothing that could be out of bounds.
2154   if (Idxs.empty()) return true;
2155 
2156   // If the first index is zero, it's in bounds.
2157   if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2158 
2159   // If the first index is one and all the rest are zero, it's in bounds,
2160   // by the one-past-the-end rule.
2161   if (!cast<ConstantInt>(Idxs[0])->isOne())
2162     return false;
2163   for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
2164     if (!cast<Constant>(Idxs[i])->isNullValue())
2165       return false;
2166   return true;
2167 }
2168 
2169 template<typename IndexTy>
ConstantFoldGetElementPtrImpl(Constant * C,bool inBounds,ArrayRef<IndexTy> Idxs)2170 static Constant *ConstantFoldGetElementPtrImpl(Constant *C,
2171                                                bool inBounds,
2172                                                ArrayRef<IndexTy> Idxs) {
2173   if (Idxs.empty()) return C;
2174   Constant *Idx0 = cast<Constant>(Idxs[0]);
2175   if ((Idxs.size() == 1 && Idx0->isNullValue()))
2176     return C;
2177 
2178   if (isa<UndefValue>(C)) {
2179     PointerType *Ptr = cast<PointerType>(C->getType());
2180     Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
2181     assert(Ty != 0 && "Invalid indices for GEP!");
2182     return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
2183   }
2184 
2185   if (C->isNullValue()) {
2186     bool isNull = true;
2187     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2188       if (!cast<Constant>(Idxs[i])->isNullValue()) {
2189         isNull = false;
2190         break;
2191       }
2192     if (isNull) {
2193       PointerType *Ptr = cast<PointerType>(C->getType());
2194       Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
2195       assert(Ty != 0 && "Invalid indices for GEP!");
2196       return ConstantPointerNull::get(PointerType::get(Ty,
2197                                                        Ptr->getAddressSpace()));
2198     }
2199   }
2200 
2201   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2202     // Combine Indices - If the source pointer to this getelementptr instruction
2203     // is a getelementptr instruction, combine the indices of the two
2204     // getelementptr instructions into a single instruction.
2205     //
2206     if (CE->getOpcode() == Instruction::GetElementPtr) {
2207       Type *LastTy = 0;
2208       for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2209            I != E; ++I)
2210         LastTy = *I;
2211 
2212       if ((LastTy && LastTy->isArrayTy()) || Idx0->isNullValue()) {
2213         SmallVector<Value*, 16> NewIndices;
2214         NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2215         for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
2216           NewIndices.push_back(CE->getOperand(i));
2217 
2218         // Add the last index of the source with the first index of the new GEP.
2219         // Make sure to handle the case when they are actually different types.
2220         Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2221         // Otherwise it must be an array.
2222         if (!Idx0->isNullValue()) {
2223           Type *IdxTy = Combined->getType();
2224           if (IdxTy != Idx0->getType()) {
2225             Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
2226             Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
2227             Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
2228             Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2229           } else {
2230             Combined =
2231               ConstantExpr::get(Instruction::Add, Idx0, Combined);
2232           }
2233         }
2234 
2235         NewIndices.push_back(Combined);
2236         NewIndices.append(Idxs.begin() + 1, Idxs.end());
2237         return
2238           ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices,
2239                                          inBounds &&
2240                                            cast<GEPOperator>(CE)->isInBounds());
2241       }
2242     }
2243 
2244     // Implement folding of:
2245     //    i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2246     //                        i64 0, i64 0)
2247     // To: i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2248     //
2249     if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2250       if (PointerType *SPT =
2251           dyn_cast<PointerType>(CE->getOperand(0)->getType()))
2252         if (ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
2253           if (ArrayType *CAT =
2254         dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
2255             if (CAT->getElementType() == SAT->getElementType())
2256               return
2257                 ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0),
2258                                                Idxs, inBounds);
2259     }
2260   }
2261 
2262   // Check to see if any array indices are not within the corresponding
2263   // notional array bounds. If so, try to determine if they can be factored
2264   // out into preceding dimensions.
2265   bool Unknown = false;
2266   SmallVector<Constant *, 8> NewIdxs;
2267   Type *Ty = C->getType();
2268   Type *Prev = 0;
2269   for (unsigned i = 0, e = Idxs.size(); i != e;
2270        Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2271     if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2272       if (ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2273         if (ATy->getNumElements() <= INT64_MAX &&
2274             ATy->getNumElements() != 0 &&
2275             CI->getSExtValue() >= (int64_t)ATy->getNumElements()) {
2276           if (isa<SequentialType>(Prev)) {
2277             // It's out of range, but we can factor it into the prior
2278             // dimension.
2279             NewIdxs.resize(Idxs.size());
2280             ConstantInt *Factor = ConstantInt::get(CI->getType(),
2281                                                    ATy->getNumElements());
2282             NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2283 
2284             Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2285             Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2286 
2287             // Before adding, extend both operands to i64 to avoid
2288             // overflow trouble.
2289             if (!PrevIdx->getType()->isIntegerTy(64))
2290               PrevIdx = ConstantExpr::getSExt(PrevIdx,
2291                                            Type::getInt64Ty(Div->getContext()));
2292             if (!Div->getType()->isIntegerTy(64))
2293               Div = ConstantExpr::getSExt(Div,
2294                                           Type::getInt64Ty(Div->getContext()));
2295 
2296             NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2297           } else {
2298             // It's out of range, but the prior dimension is a struct
2299             // so we can't do anything about it.
2300             Unknown = true;
2301           }
2302         }
2303     } else {
2304       // We don't know if it's in range or not.
2305       Unknown = true;
2306     }
2307   }
2308 
2309   // If we did any factoring, start over with the adjusted indices.
2310   if (!NewIdxs.empty()) {
2311     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2312       if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2313     return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds);
2314   }
2315 
2316   // If all indices are known integers and normalized, we can do a simple
2317   // check for the "inbounds" property.
2318   if (!Unknown && !inBounds &&
2319       isa<GlobalVariable>(C) && isInBoundsIndices(Idxs))
2320     return ConstantExpr::getInBoundsGetElementPtr(C, Idxs);
2321 
2322   return 0;
2323 }
2324 
ConstantFoldGetElementPtr(Constant * C,bool inBounds,ArrayRef<Constant * > Idxs)2325 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2326                                           bool inBounds,
2327                                           ArrayRef<Constant *> Idxs) {
2328   return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);
2329 }
2330 
ConstantFoldGetElementPtr(Constant * C,bool inBounds,ArrayRef<Value * > Idxs)2331 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2332                                           bool inBounds,
2333                                           ArrayRef<Value *> Idxs) {
2334   return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);
2335 }
2336