ConstantFold.cpp revision 4ada00d166b1d265fa4e751b63d7f869c7541f1b
1//===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by the LLVM research group and is distributed under
6// the University of Illinois Open Source 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// template-based folder for simple primitive constants like ConstantInt, and
16// the special case hackery that we use to symbolically evaluate expressions
17// that use ConstantExprs.
18//
19//===----------------------------------------------------------------------===//
20
21#include "ConstantFold.h"
22#include "llvm/Constants.h"
23#include "llvm/Instructions.h"
24#include "llvm/DerivedTypes.h"
25#include "llvm/Function.h"
26#include "llvm/ADT/SmallVector.h"
27#include "llvm/Support/Compiler.h"
28#include "llvm/Support/GetElementPtrTypeIterator.h"
29#include "llvm/Support/ManagedStatic.h"
30#include "llvm/Support/MathExtras.h"
31#include <limits>
32using namespace llvm;
33
34//===----------------------------------------------------------------------===//
35//                ConstantFold*Instruction Implementations
36//===----------------------------------------------------------------------===//
37
38/// CastConstantVector - Convert the specified ConstantVector node to the
39/// specified vector type.  At this point, we know that the elements of the
40/// input packed constant are all simple integer or FP values.
41static Constant *CastConstantVector(ConstantVector *CV,
42                                    const VectorType *DstTy) {
43  unsigned SrcNumElts = CV->getType()->getNumElements();
44  unsigned DstNumElts = DstTy->getNumElements();
45  const Type *SrcEltTy = CV->getType()->getElementType();
46  const Type *DstEltTy = DstTy->getElementType();
47
48  // If both vectors have the same number of elements (thus, the elements
49  // are the same size), perform the conversion now.
50  if (SrcNumElts == DstNumElts) {
51    std::vector<Constant*> Result;
52
53    // If the src and dest elements are both integers, or both floats, we can
54    // just BitCast each element because the elements are the same size.
55    if ((SrcEltTy->isInteger() && DstEltTy->isInteger()) ||
56        (SrcEltTy->isFloatingPoint() && DstEltTy->isFloatingPoint())) {
57      for (unsigned i = 0; i != SrcNumElts; ++i)
58        Result.push_back(
59          ConstantExpr::getBitCast(CV->getOperand(i), DstEltTy));
60      return ConstantVector::get(Result);
61    }
62
63    // If this is an int-to-fp cast ..
64    if (SrcEltTy->isInteger()) {
65      // Ensure that it is int-to-fp cast
66      assert(DstEltTy->isFloatingPoint());
67      if (DstEltTy->getTypeID() == Type::DoubleTyID) {
68        for (unsigned i = 0; i != SrcNumElts; ++i) {
69          ConstantInt *CI = cast<ConstantInt>(CV->getOperand(i));
70          double V = CI->getValue().bitsToDouble();
71          Result.push_back(ConstantFP::get(Type::DoubleTy, V));
72        }
73        return ConstantVector::get(Result);
74      }
75      assert(DstEltTy == Type::FloatTy && "Unknown fp type!");
76      for (unsigned i = 0; i != SrcNumElts; ++i) {
77        ConstantInt *CI = cast<ConstantInt>(CV->getOperand(i));
78        float V = CI->getValue().bitsToFloat();
79        Result.push_back(ConstantFP::get(Type::FloatTy, V));
80      }
81      return ConstantVector::get(Result);
82    }
83
84    // Otherwise, this is an fp-to-int cast.
85    assert(SrcEltTy->isFloatingPoint() && DstEltTy->isInteger());
86
87    if (SrcEltTy->getTypeID() == Type::DoubleTyID) {
88      for (unsigned i = 0; i != SrcNumElts; ++i) {
89        uint64_t V =
90          DoubleToBits(cast<ConstantFP>(CV->getOperand(i))->getValue());
91        Constant *C = ConstantInt::get(Type::Int64Ty, V);
92        Result.push_back(ConstantExpr::getBitCast(C, DstEltTy ));
93      }
94      return ConstantVector::get(Result);
95    }
96
97    assert(SrcEltTy->getTypeID() == Type::FloatTyID);
98    for (unsigned i = 0; i != SrcNumElts; ++i) {
99      uint32_t V = FloatToBits(cast<ConstantFP>(CV->getOperand(i))->getValue());
100      Constant *C = ConstantInt::get(Type::Int32Ty, V);
101      Result.push_back(ConstantExpr::getBitCast(C, DstEltTy));
102    }
103    return ConstantVector::get(Result);
104  }
105
106  // Otherwise, this is a cast that changes element count and size.  Handle
107  // casts which shrink the elements here.
108
109  // FIXME: We need to know endianness to do this!
110
111  return 0;
112}
113
114/// This function determines which opcode to use to fold two constant cast
115/// expressions together. It uses CastInst::isEliminableCastPair to determine
116/// the opcode. Consequently its just a wrapper around that function.
117/// @Determine if it is valid to fold a cast of a cast
118static unsigned
119foldConstantCastPair(
120  unsigned opc,          ///< opcode of the second cast constant expression
121  const ConstantExpr*Op, ///< the first cast constant expression
122  const Type *DstTy      ///< desintation type of the first cast
123) {
124  assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
125  assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
126  assert(CastInst::isCast(opc) && "Invalid cast opcode");
127
128  // The the types and opcodes for the two Cast constant expressions
129  const Type *SrcTy = Op->getOperand(0)->getType();
130  const Type *MidTy = Op->getType();
131  Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
132  Instruction::CastOps secondOp = Instruction::CastOps(opc);
133
134  // Let CastInst::isEliminableCastPair do the heavy lifting.
135  return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
136                                        Type::Int64Ty);
137}
138
139Constant *llvm::ConstantFoldCastInstruction(unsigned opc, const Constant *V,
140                                            const Type *DestTy) {
141  const Type *SrcTy = V->getType();
142
143  if (isa<UndefValue>(V))
144    return UndefValue::get(DestTy);
145
146  // If the cast operand is a constant expression, there's a few things we can
147  // do to try to simplify it.
148  if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
149    if (CE->isCast()) {
150      // Try hard to fold cast of cast because they are often eliminable.
151      if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
152        return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
153    } else if (CE->getOpcode() == Instruction::GetElementPtr) {
154      // If all of the indexes in the GEP are null values, there is no pointer
155      // adjustment going on.  We might as well cast the source pointer.
156      bool isAllNull = true;
157      for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
158        if (!CE->getOperand(i)->isNullValue()) {
159          isAllNull = false;
160          break;
161        }
162      if (isAllNull)
163        // This is casting one pointer type to another, always BitCast
164        return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
165    }
166  }
167
168  // We actually have to do a cast now. Perform the cast according to the
169  // opcode specified.
170  switch (opc) {
171  case Instruction::FPTrunc:
172  case Instruction::FPExt:
173    if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V))
174      return ConstantFP::get(DestTy, FPC->getValue());
175    return 0; // Can't fold.
176  case Instruction::FPToUI:
177    if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
178      uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
179      APInt Val(APIntOps::RoundDoubleToAPInt(FPC->getValue(), DestBitWidth));
180      return ConstantInt::get(Val);
181    }
182    return 0; // Can't fold.
183  case Instruction::FPToSI:
184    if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
185      uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
186      APInt Val(APIntOps::RoundDoubleToAPInt(FPC->getValue(), DestBitWidth));
187      return ConstantInt::get(Val);
188    }
189    return 0; // Can't fold.
190  case Instruction::IntToPtr:   //always treated as unsigned
191    if (V->isNullValue())       // Is it an integral null value?
192      return ConstantPointerNull::get(cast<PointerType>(DestTy));
193    return 0;                   // Other pointer types cannot be casted
194  case Instruction::PtrToInt:   // always treated as unsigned
195    if (V->isNullValue())       // is it a null pointer value?
196      return ConstantInt::get(DestTy, 0);
197    return 0;                   // Other pointer types cannot be casted
198  case Instruction::UIToFP:
199    if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
200      return ConstantFP::get(DestTy, CI->getValue().roundToDouble());
201    return 0;
202  case Instruction::SIToFP:
203    if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
204      return ConstantFP::get(DestTy, CI->getValue().signedRoundToDouble());
205    return 0;
206  case Instruction::ZExt:
207    if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
208      uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
209      APInt Result(CI->getValue());
210      Result.zext(BitWidth);
211      return ConstantInt::get(Result);
212    }
213    return 0;
214  case Instruction::SExt:
215    if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
216      uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
217      APInt Result(CI->getValue());
218      Result.sext(BitWidth);
219      return ConstantInt::get(Result);
220    }
221    return 0;
222  case Instruction::Trunc:
223    if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
224      uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
225      APInt Result(CI->getValue());
226      Result.trunc(BitWidth);
227      return ConstantInt::get(Result);
228    }
229    return 0;
230  case Instruction::BitCast:
231    if (SrcTy == DestTy)
232      return (Constant*)V; // no-op cast
233
234    // Check to see if we are casting a pointer to an aggregate to a pointer to
235    // the first element.  If so, return the appropriate GEP instruction.
236    if (const PointerType *PTy = dyn_cast<PointerType>(V->getType()))
237      if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy)) {
238        SmallVector<Value*, 8> IdxList;
239        IdxList.push_back(Constant::getNullValue(Type::Int32Ty));
240        const Type *ElTy = PTy->getElementType();
241        while (ElTy != DPTy->getElementType()) {
242          if (const StructType *STy = dyn_cast<StructType>(ElTy)) {
243            if (STy->getNumElements() == 0) break;
244            ElTy = STy->getElementType(0);
245            IdxList.push_back(Constant::getNullValue(Type::Int32Ty));
246          } else if (const SequentialType *STy =
247                     dyn_cast<SequentialType>(ElTy)) {
248            if (isa<PointerType>(ElTy)) break;  // Can't index into pointers!
249            ElTy = STy->getElementType();
250            IdxList.push_back(IdxList[0]);
251          } else {
252            break;
253          }
254        }
255
256        if (ElTy == DPTy->getElementType())
257          return ConstantExpr::getGetElementPtr(
258              const_cast<Constant*>(V), &IdxList[0], IdxList.size());
259      }
260
261    // Handle casts from one packed constant to another.  We know that the src
262    // and dest type have the same size (otherwise its an illegal cast).
263    if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
264      if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
265        assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
266               "Not cast between same sized vectors!");
267        // First, check for null and undef
268        if (isa<ConstantAggregateZero>(V))
269          return Constant::getNullValue(DestTy);
270        if (isa<UndefValue>(V))
271          return UndefValue::get(DestTy);
272
273        if (const ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
274          // This is a cast from a ConstantVector of one type to a
275          // ConstantVector of another type.  Check to see if all elements of
276          // the input are simple.
277          bool AllSimpleConstants = true;
278          for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
279            if (!isa<ConstantInt>(CV->getOperand(i)) &&
280                !isa<ConstantFP>(CV->getOperand(i))) {
281              AllSimpleConstants = false;
282              break;
283            }
284          }
285
286          // If all of the elements are simple constants, we can fold this.
287          if (AllSimpleConstants)
288            return CastConstantVector(const_cast<ConstantVector*>(CV), DestPTy);
289        }
290      }
291    }
292
293    // Finally, implement bitcast folding now.   The code below doesn't handle
294    // bitcast right.
295    if (isa<ConstantPointerNull>(V))  // ptr->ptr cast.
296      return ConstantPointerNull::get(cast<PointerType>(DestTy));
297
298    // Handle integral constant input.
299    if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
300      if (DestTy->isInteger())
301        // Integral -> Integral. This is a no-op because the bit widths must
302        // be the same. Consequently, we just fold to V.
303        return const_cast<Constant*>(V);
304
305      if (DestTy->isFloatingPoint()) {
306        if (DestTy == Type::FloatTy)
307          return ConstantFP::get(DestTy, CI->getValue().bitsToFloat());
308        assert(DestTy == Type::DoubleTy && "Unknown FP type!");
309        return ConstantFP::get(DestTy, CI->getValue().bitsToDouble());
310      }
311      // Otherwise, can't fold this (packed?)
312      return 0;
313    }
314
315    // Handle ConstantFP input.
316    if (const ConstantFP *FP = dyn_cast<ConstantFP>(V)) {
317      // FP -> Integral.
318      if (DestTy == Type::Int32Ty) {
319        APInt Val(32, 0);
320        return ConstantInt::get(Val.floatToBits(FP->getValue()));
321      } else {
322        assert(DestTy == Type::Int64Ty && "only support f32/f64 for now!");
323        APInt Val(64, 0);
324        return ConstantInt::get(Val.doubleToBits(FP->getValue()));
325      }
326    }
327    return 0;
328  default:
329    assert(!"Invalid CE CastInst opcode");
330    break;
331  }
332
333  assert(0 && "Failed to cast constant expression");
334  return 0;
335}
336
337Constant *llvm::ConstantFoldSelectInstruction(const Constant *Cond,
338                                              const Constant *V1,
339                                              const Constant *V2) {
340  if (const ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
341    return const_cast<Constant*>(CB->getZExtValue() ? V1 : V2);
342
343  if (isa<UndefValue>(V1)) return const_cast<Constant*>(V2);
344  if (isa<UndefValue>(V2)) return const_cast<Constant*>(V1);
345  if (isa<UndefValue>(Cond)) return const_cast<Constant*>(V1);
346  if (V1 == V2) return const_cast<Constant*>(V1);
347  return 0;
348}
349
350Constant *llvm::ConstantFoldExtractElementInstruction(const Constant *Val,
351                                                      const Constant *Idx) {
352  if (isa<UndefValue>(Val))  // ee(undef, x) -> undef
353    return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
354  if (Val->isNullValue())  // ee(zero, x) -> zero
355    return Constant::getNullValue(
356                          cast<VectorType>(Val->getType())->getElementType());
357
358  if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
359    if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
360      return const_cast<Constant*>(CVal->getOperand(CIdx->getZExtValue()));
361    } else if (isa<UndefValue>(Idx)) {
362      // ee({w,x,y,z}, undef) -> w (an arbitrary value).
363      return const_cast<Constant*>(CVal->getOperand(0));
364    }
365  }
366  return 0;
367}
368
369Constant *llvm::ConstantFoldInsertElementInstruction(const Constant *Val,
370                                                     const Constant *Elt,
371                                                     const Constant *Idx) {
372  const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
373  if (!CIdx) return 0;
374  APInt idxVal = CIdx->getValue();
375  if (isa<UndefValue>(Val)) {
376    // Insertion of scalar constant into packed undef
377    // Optimize away insertion of undef
378    if (isa<UndefValue>(Elt))
379      return const_cast<Constant*>(Val);
380    // Otherwise break the aggregate undef into multiple undefs and do
381    // the insertion
382    unsigned numOps =
383      cast<VectorType>(Val->getType())->getNumElements();
384    std::vector<Constant*> Ops;
385    Ops.reserve(numOps);
386    for (unsigned i = 0; i < numOps; ++i) {
387      const Constant *Op =
388        (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
389      Ops.push_back(const_cast<Constant*>(Op));
390    }
391    return ConstantVector::get(Ops);
392  }
393  if (isa<ConstantAggregateZero>(Val)) {
394    // Insertion of scalar constant into packed aggregate zero
395    // Optimize away insertion of zero
396    if (Elt->isNullValue())
397      return const_cast<Constant*>(Val);
398    // Otherwise break the aggregate zero into multiple zeros and do
399    // the insertion
400    unsigned numOps =
401      cast<VectorType>(Val->getType())->getNumElements();
402    std::vector<Constant*> Ops;
403    Ops.reserve(numOps);
404    for (unsigned i = 0; i < numOps; ++i) {
405      const Constant *Op =
406        (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
407      Ops.push_back(const_cast<Constant*>(Op));
408    }
409    return ConstantVector::get(Ops);
410  }
411  if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
412    // Insertion of scalar constant into packed constant
413    std::vector<Constant*> Ops;
414    Ops.reserve(CVal->getNumOperands());
415    for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
416      const Constant *Op =
417        (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
418      Ops.push_back(const_cast<Constant*>(Op));
419    }
420    return ConstantVector::get(Ops);
421  }
422  return 0;
423}
424
425Constant *llvm::ConstantFoldShuffleVectorInstruction(const Constant *V1,
426                                                     const Constant *V2,
427                                                     const Constant *Mask) {
428  // TODO:
429  return 0;
430}
431
432/// EvalVectorOp - Given two packed constants and a function pointer, apply the
433/// function pointer to each element pair, producing a new ConstantVector
434/// constant.
435static Constant *EvalVectorOp(const ConstantVector *V1,
436                              const ConstantVector *V2,
437                              Constant *(*FP)(Constant*, Constant*)) {
438  std::vector<Constant*> Res;
439  for (unsigned i = 0, e = V1->getNumOperands(); i != e; ++i)
440    Res.push_back(FP(const_cast<Constant*>(V1->getOperand(i)),
441                     const_cast<Constant*>(V2->getOperand(i))));
442  return ConstantVector::get(Res);
443}
444
445Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
446                                              const Constant *C1,
447                                              const Constant *C2) {
448  // Handle UndefValue up front
449  if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
450    switch (Opcode) {
451    case Instruction::Add:
452    case Instruction::Sub:
453    case Instruction::Xor:
454      return UndefValue::get(C1->getType());
455    case Instruction::Mul:
456    case Instruction::And:
457      return Constant::getNullValue(C1->getType());
458    case Instruction::UDiv:
459    case Instruction::SDiv:
460    case Instruction::FDiv:
461    case Instruction::URem:
462    case Instruction::SRem:
463    case Instruction::FRem:
464      if (!isa<UndefValue>(C2))                    // undef / X -> 0
465        return Constant::getNullValue(C1->getType());
466      return const_cast<Constant*>(C2);            // X / undef -> undef
467    case Instruction::Or:                          // X | undef -> -1
468      if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType()))
469        return ConstantVector::getAllOnesValue(PTy);
470      return ConstantInt::getAllOnesValue(C1->getType());
471    case Instruction::LShr:
472      if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
473        return const_cast<Constant*>(C1);           // undef lshr undef -> undef
474      return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
475                                                    // undef lshr X -> 0
476    case Instruction::AShr:
477      if (!isa<UndefValue>(C2))
478        return const_cast<Constant*>(C1);           // undef ashr X --> undef
479      else if (isa<UndefValue>(C1))
480        return const_cast<Constant*>(C1);           // undef ashr undef -> undef
481      else
482        return const_cast<Constant*>(C1);           // X ashr undef --> X
483    case Instruction::Shl:
484      // undef << X -> 0   or   X << undef -> 0
485      return Constant::getNullValue(C1->getType());
486    }
487  }
488
489  if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
490    if (isa<ConstantExpr>(C2)) {
491      // There are many possible foldings we could do here.  We should probably
492      // at least fold add of a pointer with an integer into the appropriate
493      // getelementptr.  This will improve alias analysis a bit.
494    } else {
495      // Just implement a couple of simple identities.
496      switch (Opcode) {
497      case Instruction::Add:
498        if (C2->isNullValue()) return const_cast<Constant*>(C1);  // X + 0 == X
499        break;
500      case Instruction::Sub:
501        if (C2->isNullValue()) return const_cast<Constant*>(C1);  // X - 0 == X
502        break;
503      case Instruction::Mul:
504        if (C2->isNullValue()) return const_cast<Constant*>(C2);  // X * 0 == 0
505        if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
506          if (CI->equalsInt(1))
507            return const_cast<Constant*>(C1);                     // X * 1 == X
508        break;
509      case Instruction::UDiv:
510      case Instruction::SDiv:
511        if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
512          if (CI->equalsInt(1))
513            return const_cast<Constant*>(C1);                     // X / 1 == X
514        break;
515      case Instruction::URem:
516      case Instruction::SRem:
517        if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
518          if (CI->equalsInt(1))
519            return Constant::getNullValue(CI->getType());         // X % 1 == 0
520        break;
521      case Instruction::And:
522        if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2)) {
523          if (CI->isZero()) return const_cast<Constant*>(C2);     // X & 0 == 0
524          if (CI->isAllOnesValue())
525            return const_cast<Constant*>(C1);                     // X & -1 == X
526
527          // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
528          if (CE1->getOpcode() == Instruction::ZExt) {
529            APInt PossiblySetBits
530              = cast<IntegerType>(CE1->getOperand(0)->getType())->getMask();
531            PossiblySetBits.zext(C1->getType()->getPrimitiveSizeInBits());
532            if ((PossiblySetBits & CI->getValue()) == PossiblySetBits)
533              return const_cast<Constant*>(C1);
534          }
535        }
536        if (CE1->isCast() && isa<GlobalValue>(CE1->getOperand(0))) {
537          GlobalValue *CPR = cast<GlobalValue>(CE1->getOperand(0));
538
539          // Functions are at least 4-byte aligned.  If and'ing the address of a
540          // function with a constant < 4, fold it to zero.
541          if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
542            if (CI->getValue().ult(APInt(CI->getType()->getBitWidth(),4)) &&
543                isa<Function>(CPR))
544              return Constant::getNullValue(CI->getType());
545        }
546        break;
547      case Instruction::Or:
548        if (C2->isNullValue()) return const_cast<Constant*>(C1);  // X | 0 == X
549        if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
550          if (CI->isAllOnesValue())
551            return const_cast<Constant*>(C2);  // X | -1 == -1
552        break;
553      case Instruction::Xor:
554        if (C2->isNullValue()) return const_cast<Constant*>(C1);  // X ^ 0 == X
555        break;
556      case Instruction::AShr:
557        // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
558        if (CE1->getOpcode() == Instruction::ZExt)  // Top bits known zero.
559          return ConstantExpr::getLShr(const_cast<Constant*>(C1),
560                                       const_cast<Constant*>(C2));
561        break;
562      }
563    }
564  } else if (isa<ConstantExpr>(C2)) {
565    // If C2 is a constant expr and C1 isn't, flop them around and fold the
566    // other way if possible.
567    switch (Opcode) {
568    case Instruction::Add:
569    case Instruction::Mul:
570    case Instruction::And:
571    case Instruction::Or:
572    case Instruction::Xor:
573      // No change of opcode required.
574      return ConstantFoldBinaryInstruction(Opcode, C2, C1);
575
576    case Instruction::Shl:
577    case Instruction::LShr:
578    case Instruction::AShr:
579    case Instruction::Sub:
580    case Instruction::SDiv:
581    case Instruction::UDiv:
582    case Instruction::FDiv:
583    case Instruction::URem:
584    case Instruction::SRem:
585    case Instruction::FRem:
586    default:  // These instructions cannot be flopped around.
587      return 0;
588    }
589  }
590
591  // At this point we know neither constant is an UndefValue nor a ConstantExpr
592  // so look at directly computing the value.
593  if (const ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
594    if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
595      using namespace APIntOps;
596      APInt C1V = CI1->getValue();
597      APInt C2V = CI2->getValue();
598      switch (Opcode) {
599      default:
600        break;
601      case Instruction::Add:
602        return ConstantInt::get(C1V + C2V);
603      case Instruction::Sub:
604        return ConstantInt::get(C1V - C2V);
605      case Instruction::Mul:
606        return ConstantInt::get(C1V * C2V);
607      case Instruction::UDiv:
608        if (CI2->isNullValue())
609          return 0;        // X / 0 -> can't fold
610        return ConstantInt::get(C1V.udiv(C2V));
611      case Instruction::SDiv:
612        if (CI2->isNullValue())
613          return 0;        // X / 0 -> can't fold
614        if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
615          return 0;        // MIN_INT / -1 -> overflow
616        return ConstantInt::get(C1V.sdiv(C2V));
617      case Instruction::URem:
618        if (C2->isNullValue())
619          return 0;        // X / 0 -> can't fold
620        return ConstantInt::get(C1V.urem(C2V));
621      case Instruction::SRem:
622        if (CI2->isNullValue())
623          return 0;        // X % 0 -> can't fold
624        if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
625          return 0;        // MIN_INT % -1 -> overflow
626        return ConstantInt::get(C1V.srem(C2V));
627      case Instruction::And:
628        return ConstantInt::get(C1V & C2V);
629      case Instruction::Or:
630        return ConstantInt::get(C1V | C2V);
631      case Instruction::Xor:
632        return ConstantInt::get(C1V ^ C2V);
633      case Instruction::Shl:
634        if (uint32_t shiftAmt = C2V.getZExtValue())
635          if (shiftAmt < C1V.getBitWidth())
636            return ConstantInt::get(C1V.shl(shiftAmt));
637          else
638            return UndefValue::get(C1->getType()); // too big shift is undef
639        return const_cast<ConstantInt*>(CI1); // Zero shift is identity
640      case Instruction::LShr:
641        if (uint32_t shiftAmt = C2V.getZExtValue())
642          if (shiftAmt < C1V.getBitWidth())
643            return ConstantInt::get(C1V.lshr(shiftAmt));
644          else
645            return UndefValue::get(C1->getType()); // too big shift is undef
646        return const_cast<ConstantInt*>(CI1); // Zero shift is identity
647      case Instruction::AShr:
648        if (uint32_t shiftAmt = C2V.getZExtValue())
649          if (shiftAmt < C1V.getBitWidth())
650            return ConstantInt::get(C1V.ashr(shiftAmt));
651          else
652            return UndefValue::get(C1->getType()); // too big shift is undef
653        return const_cast<ConstantInt*>(CI1); // Zero shift is identity
654      }
655    }
656  } else if (const ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
657    if (const ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
658      double C1Val = CFP1->getValue();
659      double C2Val = CFP2->getValue();
660      switch (Opcode) {
661      default:
662        break;
663      case Instruction::Add:
664        return ConstantFP::get(CFP1->getType(), C1Val + C2Val);
665      case Instruction::Sub:
666        return ConstantFP::get(CFP1->getType(), C1Val - C2Val);
667      case Instruction::Mul:
668        return ConstantFP::get(CFP1->getType(), C1Val * C2Val);
669      case Instruction::FDiv:
670        if (CFP2->isExactlyValue(0.0) || CFP2->isExactlyValue(-0.0))
671          if (CFP1->isExactlyValue(0.0) || CFP1->isExactlyValue(-0.0))
672            // IEEE 754, Section 7.1, #4
673            return ConstantFP::get(CFP1->getType(),
674                                   std::numeric_limits<double>::quiet_NaN());
675          else if (CFP2->isExactlyValue(-0.0) || C1Val < 0.0)
676            // IEEE 754, Section 7.2, negative infinity case
677            return ConstantFP::get(CFP1->getType(),
678                                   -std::numeric_limits<double>::infinity());
679          else
680            // IEEE 754, Section 7.2, positive infinity case
681            return ConstantFP::get(CFP1->getType(),
682                                   std::numeric_limits<double>::infinity());
683        return ConstantFP::get(CFP1->getType(), C1Val / C2Val);
684      case Instruction::FRem:
685        if (CFP2->isExactlyValue(0.0) || CFP2->isExactlyValue(-0.0))
686          // IEEE 754, Section 7.1, #5
687          return ConstantFP::get(CFP1->getType(),
688                                 std::numeric_limits<double>::quiet_NaN());
689        return ConstantFP::get(CFP1->getType(), std::fmod(C1Val, C2Val));
690
691      }
692    }
693  } else if (const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1)) {
694    if (const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2)) {
695      switch (Opcode) {
696        default:
697          break;
698        case Instruction::Add:
699          return EvalVectorOp(CP1, CP2, ConstantExpr::getAdd);
700        case Instruction::Sub:
701          return EvalVectorOp(CP1, CP2, ConstantExpr::getSub);
702        case Instruction::Mul:
703          return EvalVectorOp(CP1, CP2, ConstantExpr::getMul);
704        case Instruction::UDiv:
705          return EvalVectorOp(CP1, CP2, ConstantExpr::getUDiv);
706        case Instruction::SDiv:
707          return EvalVectorOp(CP1, CP2, ConstantExpr::getSDiv);
708        case Instruction::FDiv:
709          return EvalVectorOp(CP1, CP2, ConstantExpr::getFDiv);
710        case Instruction::URem:
711          return EvalVectorOp(CP1, CP2, ConstantExpr::getURem);
712        case Instruction::SRem:
713          return EvalVectorOp(CP1, CP2, ConstantExpr::getSRem);
714        case Instruction::FRem:
715          return EvalVectorOp(CP1, CP2, ConstantExpr::getFRem);
716        case Instruction::And:
717          return EvalVectorOp(CP1, CP2, ConstantExpr::getAnd);
718        case Instruction::Or:
719          return EvalVectorOp(CP1, CP2, ConstantExpr::getOr);
720        case Instruction::Xor:
721          return EvalVectorOp(CP1, CP2, ConstantExpr::getXor);
722      }
723    }
724  }
725
726  // We don't know how to fold this
727  return 0;
728}
729
730/// isZeroSizedType - This type is zero sized if its an array or structure of
731/// zero sized types.  The only leaf zero sized type is an empty structure.
732static bool isMaybeZeroSizedType(const Type *Ty) {
733  if (isa<OpaqueType>(Ty)) return true;  // Can't say.
734  if (const StructType *STy = dyn_cast<StructType>(Ty)) {
735
736    // If all of elements have zero size, this does too.
737    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
738      if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
739    return true;
740
741  } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
742    return isMaybeZeroSizedType(ATy->getElementType());
743  }
744  return false;
745}
746
747/// IdxCompare - Compare the two constants as though they were getelementptr
748/// indices.  This allows coersion of the types to be the same thing.
749///
750/// If the two constants are the "same" (after coersion), return 0.  If the
751/// first is less than the second, return -1, if the second is less than the
752/// first, return 1.  If the constants are not integral, return -2.
753///
754static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) {
755  if (C1 == C2) return 0;
756
757  // Ok, we found a different index.  If they are not ConstantInt, we can't do
758  // anything with them.
759  if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
760    return -2; // don't know!
761
762  // Ok, we have two differing integer indices.  Sign extend them to be the same
763  // type.  Long is always big enough, so we use it.
764  if (C1->getType() != Type::Int64Ty)
765    C1 = ConstantExpr::getSExt(C1, Type::Int64Ty);
766
767  if (C2->getType() != Type::Int64Ty)
768    C2 = ConstantExpr::getSExt(C2, Type::Int64Ty);
769
770  if (C1 == C2) return 0;  // They are equal
771
772  // If the type being indexed over is really just a zero sized type, there is
773  // no pointer difference being made here.
774  if (isMaybeZeroSizedType(ElTy))
775    return -2; // dunno.
776
777  // If they are really different, now that they are the same type, then we
778  // found a difference!
779  if (cast<ConstantInt>(C1)->getSExtValue() <
780      cast<ConstantInt>(C2)->getSExtValue())
781    return -1;
782  else
783    return 1;
784}
785
786/// evaluateFCmpRelation - This function determines if there is anything we can
787/// decide about the two constants provided.  This doesn't need to handle simple
788/// things like ConstantFP comparisons, but should instead handle ConstantExprs.
789/// If we can determine that the two constants have a particular relation to
790/// each other, we should return the corresponding FCmpInst predicate,
791/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
792/// ConstantFoldCompareInstruction.
793///
794/// To simplify this code we canonicalize the relation so that the first
795/// operand is always the most "complex" of the two.  We consider ConstantFP
796/// to be the simplest, and ConstantExprs to be the most complex.
797static FCmpInst::Predicate evaluateFCmpRelation(const Constant *V1,
798                                                const Constant *V2) {
799  assert(V1->getType() == V2->getType() &&
800         "Cannot compare values of different types!");
801  // Handle degenerate case quickly
802  if (V1 == V2) return FCmpInst::FCMP_OEQ;
803
804  if (!isa<ConstantExpr>(V1)) {
805    if (!isa<ConstantExpr>(V2)) {
806      // We distilled thisUse the standard constant folder for a few cases
807      ConstantInt *R = 0;
808      Constant *C1 = const_cast<Constant*>(V1);
809      Constant *C2 = const_cast<Constant*>(V2);
810      R = dyn_cast<ConstantInt>(
811                             ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, C1, C2));
812      if (R && !R->isZero())
813        return FCmpInst::FCMP_OEQ;
814      R = dyn_cast<ConstantInt>(
815                             ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, C1, C2));
816      if (R && !R->isZero())
817        return FCmpInst::FCMP_OLT;
818      R = dyn_cast<ConstantInt>(
819                             ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, C1, C2));
820      if (R && !R->isZero())
821        return FCmpInst::FCMP_OGT;
822
823      // Nothing more we can do
824      return FCmpInst::BAD_FCMP_PREDICATE;
825    }
826
827    // If the first operand is simple and second is ConstantExpr, swap operands.
828    FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
829    if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
830      return FCmpInst::getSwappedPredicate(SwappedRelation);
831  } else {
832    // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
833    // constantexpr or a simple constant.
834    const ConstantExpr *CE1 = cast<ConstantExpr>(V1);
835    switch (CE1->getOpcode()) {
836    case Instruction::FPTrunc:
837    case Instruction::FPExt:
838    case Instruction::UIToFP:
839    case Instruction::SIToFP:
840      // We might be able to do something with these but we don't right now.
841      break;
842    default:
843      break;
844    }
845  }
846  // There are MANY other foldings that we could perform here.  They will
847  // probably be added on demand, as they seem needed.
848  return FCmpInst::BAD_FCMP_PREDICATE;
849}
850
851/// evaluateICmpRelation - This function determines if there is anything we can
852/// decide about the two constants provided.  This doesn't need to handle simple
853/// things like integer comparisons, but should instead handle ConstantExprs
854/// and GlobalValues.  If we can determine that the two constants have a
855/// particular relation to each other, we should return the corresponding ICmp
856/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
857///
858/// To simplify this code we canonicalize the relation so that the first
859/// operand is always the most "complex" of the two.  We consider simple
860/// constants (like ConstantInt) to be the simplest, followed by
861/// GlobalValues, followed by ConstantExpr's (the most complex).
862///
863static ICmpInst::Predicate evaluateICmpRelation(const Constant *V1,
864                                                const Constant *V2,
865                                                bool isSigned) {
866  assert(V1->getType() == V2->getType() &&
867         "Cannot compare different types of values!");
868  if (V1 == V2) return ICmpInst::ICMP_EQ;
869
870  if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1)) {
871    if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2)) {
872      // We distilled this down to a simple case, use the standard constant
873      // folder.
874      ConstantInt *R = 0;
875      Constant *C1 = const_cast<Constant*>(V1);
876      Constant *C2 = const_cast<Constant*>(V2);
877      ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
878      R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
879      if (R && !R->isZero())
880        return pred;
881      pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
882      R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
883      if (R && !R->isZero())
884        return pred;
885      pred = isSigned ?  ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
886      R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
887      if (R && !R->isZero())
888        return pred;
889
890      // If we couldn't figure it out, bail.
891      return ICmpInst::BAD_ICMP_PREDICATE;
892    }
893
894    // If the first operand is simple, swap operands.
895    ICmpInst::Predicate SwappedRelation =
896      evaluateICmpRelation(V2, V1, isSigned);
897    if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
898      return ICmpInst::getSwappedPredicate(SwappedRelation);
899
900  } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(V1)) {
901    if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
902      ICmpInst::Predicate SwappedRelation =
903        evaluateICmpRelation(V2, V1, isSigned);
904      if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
905        return ICmpInst::getSwappedPredicate(SwappedRelation);
906      else
907        return ICmpInst::BAD_ICMP_PREDICATE;
908    }
909
910    // Now we know that the RHS is a GlobalValue or simple constant,
911    // which (since the types must match) means that it's a ConstantPointerNull.
912    if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) {
913      if (!CPR1->hasExternalWeakLinkage() || !CPR2->hasExternalWeakLinkage())
914        return ICmpInst::ICMP_NE;
915    } else {
916      // GlobalVals can never be null.
917      assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
918      if (!CPR1->hasExternalWeakLinkage())
919        return ICmpInst::ICMP_NE;
920    }
921  } else {
922    // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
923    // constantexpr, a CPR, or a simple constant.
924    const ConstantExpr *CE1 = cast<ConstantExpr>(V1);
925    const Constant *CE1Op0 = CE1->getOperand(0);
926
927    switch (CE1->getOpcode()) {
928    case Instruction::Trunc:
929    case Instruction::FPTrunc:
930    case Instruction::FPExt:
931    case Instruction::FPToUI:
932    case Instruction::FPToSI:
933      break; // We can't evaluate floating point casts or truncations.
934
935    case Instruction::UIToFP:
936    case Instruction::SIToFP:
937    case Instruction::IntToPtr:
938    case Instruction::BitCast:
939    case Instruction::ZExt:
940    case Instruction::SExt:
941    case Instruction::PtrToInt:
942      // If the cast is not actually changing bits, and the second operand is a
943      // null pointer, do the comparison with the pre-casted value.
944      if (V2->isNullValue() &&
945          (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) {
946        bool sgnd = CE1->getOpcode() == Instruction::ZExt ? false :
947          (CE1->getOpcode() == Instruction::SExt ? true :
948           (CE1->getOpcode() == Instruction::PtrToInt ? false : isSigned));
949        return evaluateICmpRelation(
950            CE1Op0, Constant::getNullValue(CE1Op0->getType()), sgnd);
951      }
952
953      // If the dest type is a pointer type, and the RHS is a constantexpr cast
954      // from the same type as the src of the LHS, evaluate the inputs.  This is
955      // important for things like "icmp eq (cast 4 to int*), (cast 5 to int*)",
956      // which happens a lot in compilers with tagged integers.
957      if (const ConstantExpr *CE2 = dyn_cast<ConstantExpr>(V2))
958        if (CE2->isCast() && isa<PointerType>(CE1->getType()) &&
959            CE1->getOperand(0)->getType() == CE2->getOperand(0)->getType() &&
960            CE1->getOperand(0)->getType()->isInteger()) {
961          bool sgnd = CE1->getOpcode() == Instruction::ZExt ? false :
962            (CE1->getOpcode() == Instruction::SExt ? true :
963             (CE1->getOpcode() == Instruction::PtrToInt ? false : isSigned));
964          return evaluateICmpRelation(CE1->getOperand(0), CE2->getOperand(0),
965              sgnd);
966        }
967      break;
968
969    case Instruction::GetElementPtr:
970      // Ok, since this is a getelementptr, we know that the constant has a
971      // pointer type.  Check the various cases.
972      if (isa<ConstantPointerNull>(V2)) {
973        // If we are comparing a GEP to a null pointer, check to see if the base
974        // of the GEP equals the null pointer.
975        if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
976          if (GV->hasExternalWeakLinkage())
977            // Weak linkage GVals could be zero or not. We're comparing that
978            // to null pointer so its greater-or-equal
979            return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
980          else
981            // If its not weak linkage, the GVal must have a non-zero address
982            // so the result is greater-than
983            return isSigned ? ICmpInst::ICMP_SGT :  ICmpInst::ICMP_UGT;
984        } else if (isa<ConstantPointerNull>(CE1Op0)) {
985          // If we are indexing from a null pointer, check to see if we have any
986          // non-zero indices.
987          for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
988            if (!CE1->getOperand(i)->isNullValue())
989              // Offsetting from null, must not be equal.
990              return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
991          // Only zero indexes from null, must still be zero.
992          return ICmpInst::ICMP_EQ;
993        }
994        // Otherwise, we can't really say if the first operand is null or not.
995      } else if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) {
996        if (isa<ConstantPointerNull>(CE1Op0)) {
997          if (CPR2->hasExternalWeakLinkage())
998            // Weak linkage GVals could be zero or not. We're comparing it to
999            // a null pointer, so its less-or-equal
1000            return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1001          else
1002            // If its not weak linkage, the GVal must have a non-zero address
1003            // so the result is less-than
1004            return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1005        } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(CE1Op0)) {
1006          if (CPR1 == CPR2) {
1007            // If this is a getelementptr of the same global, then it must be
1008            // different.  Because the types must match, the getelementptr could
1009            // only have at most one index, and because we fold getelementptr's
1010            // with a single zero index, it must be nonzero.
1011            assert(CE1->getNumOperands() == 2 &&
1012                   !CE1->getOperand(1)->isNullValue() &&
1013                   "Suprising getelementptr!");
1014            return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1015          } else {
1016            // If they are different globals, we don't know what the value is,
1017            // but they can't be equal.
1018            return ICmpInst::ICMP_NE;
1019          }
1020        }
1021      } else {
1022        const ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1023        const Constant *CE2Op0 = CE2->getOperand(0);
1024
1025        // There are MANY other foldings that we could perform here.  They will
1026        // probably be added on demand, as they seem needed.
1027        switch (CE2->getOpcode()) {
1028        default: break;
1029        case Instruction::GetElementPtr:
1030          // By far the most common case to handle is when the base pointers are
1031          // obviously to the same or different globals.
1032          if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1033            if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1034              return ICmpInst::ICMP_NE;
1035            // Ok, we know that both getelementptr instructions are based on the
1036            // same global.  From this, we can precisely determine the relative
1037            // ordering of the resultant pointers.
1038            unsigned i = 1;
1039
1040            // Compare all of the operands the GEP's have in common.
1041            gep_type_iterator GTI = gep_type_begin(CE1);
1042            for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1043                 ++i, ++GTI)
1044              switch (IdxCompare(CE1->getOperand(i), CE2->getOperand(i),
1045                                 GTI.getIndexedType())) {
1046              case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1047              case 1:  return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1048              case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1049              }
1050
1051            // Ok, we ran out of things they have in common.  If any leftovers
1052            // are non-zero then we have a difference, otherwise we are equal.
1053            for (; i < CE1->getNumOperands(); ++i)
1054              if (!CE1->getOperand(i)->isNullValue())
1055                if (isa<ConstantInt>(CE1->getOperand(i)))
1056                  return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1057                else
1058                  return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1059
1060            for (; i < CE2->getNumOperands(); ++i)
1061              if (!CE2->getOperand(i)->isNullValue())
1062                if (isa<ConstantInt>(CE2->getOperand(i)))
1063                  return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1064                else
1065                  return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1066            return ICmpInst::ICMP_EQ;
1067          }
1068        }
1069      }
1070    default:
1071      break;
1072    }
1073  }
1074
1075  return ICmpInst::BAD_ICMP_PREDICATE;
1076}
1077
1078Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1079                                               const Constant *C1,
1080                                               const Constant *C2) {
1081
1082  // Handle some degenerate cases first
1083  if (isa<UndefValue>(C1) || isa<UndefValue>(C2))
1084    return UndefValue::get(Type::Int1Ty);
1085
1086  // icmp eq/ne(null,GV) -> false/true
1087  if (C1->isNullValue()) {
1088    if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1089      if (!GV->hasExternalWeakLinkage()) // External weak GV can be null
1090        if (pred == ICmpInst::ICMP_EQ)
1091          return ConstantInt::getFalse();
1092        else if (pred == ICmpInst::ICMP_NE)
1093          return ConstantInt::getTrue();
1094  // icmp eq/ne(GV,null) -> false/true
1095  } else if (C2->isNullValue()) {
1096    if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1097      if (!GV->hasExternalWeakLinkage()) // External weak GV can be null
1098        if (pred == ICmpInst::ICMP_EQ)
1099          return ConstantInt::getFalse();
1100        else if (pred == ICmpInst::ICMP_NE)
1101          return ConstantInt::getTrue();
1102  }
1103
1104  if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1105    APInt V1 = cast<ConstantInt>(C1)->getValue();
1106    APInt V2 = cast<ConstantInt>(C2)->getValue();
1107    switch (pred) {
1108    default: assert(0 && "Invalid ICmp Predicate"); return 0;
1109    case ICmpInst::ICMP_EQ: return ConstantInt::get(Type::Int1Ty, V1 == V2);
1110    case ICmpInst::ICMP_NE: return ConstantInt::get(Type::Int1Ty, V1 != V2);
1111    case ICmpInst::ICMP_SLT:return ConstantInt::get(Type::Int1Ty, V1.slt(V2));
1112    case ICmpInst::ICMP_SGT:return ConstantInt::get(Type::Int1Ty, V1.sgt(V2));
1113    case ICmpInst::ICMP_SLE:return ConstantInt::get(Type::Int1Ty, V1.sle(V2));
1114    case ICmpInst::ICMP_SGE:return ConstantInt::get(Type::Int1Ty, V1.sge(V2));
1115    case ICmpInst::ICMP_ULT:return ConstantInt::get(Type::Int1Ty, V1.ult(V2));
1116    case ICmpInst::ICMP_UGT:return ConstantInt::get(Type::Int1Ty, V1.ugt(V2));
1117    case ICmpInst::ICMP_ULE:return ConstantInt::get(Type::Int1Ty, V1.ule(V2));
1118    case ICmpInst::ICMP_UGE:return ConstantInt::get(Type::Int1Ty, V1.uge(V2));
1119    }
1120  } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1121    double C1Val = cast<ConstantFP>(C1)->getValue();
1122    double C2Val = cast<ConstantFP>(C2)->getValue();
1123    switch (pred) {
1124    default: assert(0 && "Invalid FCmp Predicate"); return 0;
1125    case FCmpInst::FCMP_FALSE: return ConstantInt::getFalse();
1126    case FCmpInst::FCMP_TRUE:  return ConstantInt::getTrue();
1127    case FCmpInst::FCMP_UNO:
1128      return ConstantInt::get(Type::Int1Ty, C1Val != C1Val || C2Val != C2Val);
1129    case FCmpInst::FCMP_ORD:
1130      return ConstantInt::get(Type::Int1Ty, C1Val == C1Val && C2Val == C2Val);
1131    case FCmpInst::FCMP_UEQ:
1132      if (C1Val != C1Val || C2Val != C2Val)
1133        return ConstantInt::getTrue();
1134      /* FALL THROUGH */
1135    case FCmpInst::FCMP_OEQ:
1136      return ConstantInt::get(Type::Int1Ty, C1Val == C2Val);
1137    case FCmpInst::FCMP_UNE:
1138      if (C1Val != C1Val || C2Val != C2Val)
1139        return ConstantInt::getTrue();
1140      /* FALL THROUGH */
1141    case FCmpInst::FCMP_ONE:
1142      return ConstantInt::get(Type::Int1Ty, C1Val != C2Val);
1143    case FCmpInst::FCMP_ULT:
1144      if (C1Val != C1Val || C2Val != C2Val)
1145        return ConstantInt::getTrue();
1146      /* FALL THROUGH */
1147    case FCmpInst::FCMP_OLT:
1148      return ConstantInt::get(Type::Int1Ty, C1Val < C2Val);
1149    case FCmpInst::FCMP_UGT:
1150      if (C1Val != C1Val || C2Val != C2Val)
1151        return ConstantInt::getTrue();
1152      /* FALL THROUGH */
1153    case FCmpInst::FCMP_OGT:
1154      return ConstantInt::get(Type::Int1Ty, C1Val > C2Val);
1155    case FCmpInst::FCMP_ULE:
1156      if (C1Val != C1Val || C2Val != C2Val)
1157        return ConstantInt::getTrue();
1158      /* FALL THROUGH */
1159    case FCmpInst::FCMP_OLE:
1160      return ConstantInt::get(Type::Int1Ty, C1Val <= C2Val);
1161    case FCmpInst::FCMP_UGE:
1162      if (C1Val != C1Val || C2Val != C2Val)
1163        return ConstantInt::getTrue();
1164      /* FALL THROUGH */
1165    case FCmpInst::FCMP_OGE:
1166      return ConstantInt::get(Type::Int1Ty, C1Val >= C2Val);
1167    }
1168  } else if (const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1)) {
1169    if (const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2)) {
1170      if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) {
1171        for (unsigned i = 0, e = CP1->getNumOperands(); i != e; ++i) {
1172          Constant *C= ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ,
1173              const_cast<Constant*>(CP1->getOperand(i)),
1174              const_cast<Constant*>(CP2->getOperand(i)));
1175          if (ConstantInt *CB = dyn_cast<ConstantInt>(C))
1176            return CB;
1177        }
1178        // Otherwise, could not decide from any element pairs.
1179        return 0;
1180      } else if (pred == ICmpInst::ICMP_EQ) {
1181        for (unsigned i = 0, e = CP1->getNumOperands(); i != e; ++i) {
1182          Constant *C = ConstantExpr::getICmp(ICmpInst::ICMP_EQ,
1183              const_cast<Constant*>(CP1->getOperand(i)),
1184              const_cast<Constant*>(CP2->getOperand(i)));
1185          if (ConstantInt *CB = dyn_cast<ConstantInt>(C))
1186            return CB;
1187        }
1188        // Otherwise, could not decide from any element pairs.
1189        return 0;
1190      }
1191    }
1192  }
1193
1194  if (C1->getType()->isFloatingPoint()) {
1195    switch (evaluateFCmpRelation(C1, C2)) {
1196    default: assert(0 && "Unknown relation!");
1197    case FCmpInst::FCMP_UNO:
1198    case FCmpInst::FCMP_ORD:
1199    case FCmpInst::FCMP_UEQ:
1200    case FCmpInst::FCMP_UNE:
1201    case FCmpInst::FCMP_ULT:
1202    case FCmpInst::FCMP_UGT:
1203    case FCmpInst::FCMP_ULE:
1204    case FCmpInst::FCMP_UGE:
1205    case FCmpInst::FCMP_TRUE:
1206    case FCmpInst::FCMP_FALSE:
1207    case FCmpInst::BAD_FCMP_PREDICATE:
1208      break; // Couldn't determine anything about these constants.
1209    case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1210      return ConstantInt::get(Type::Int1Ty,
1211          pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1212          pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1213          pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1214    case FCmpInst::FCMP_OLT: // We know that C1 < C2
1215      return ConstantInt::get(Type::Int1Ty,
1216          pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1217          pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1218          pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1219    case FCmpInst::FCMP_OGT: // We know that C1 > C2
1220      return ConstantInt::get(Type::Int1Ty,
1221          pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1222          pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1223          pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1224    case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1225      // We can only partially decide this relation.
1226      if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1227        return ConstantInt::getFalse();
1228      if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1229        return ConstantInt::getTrue();
1230      break;
1231    case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1232      // We can only partially decide this relation.
1233      if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1234        return ConstantInt::getFalse();
1235      if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1236        return ConstantInt::getTrue();
1237      break;
1238    case ICmpInst::ICMP_NE: // We know that C1 != C2
1239      // We can only partially decide this relation.
1240      if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1241        return ConstantInt::getFalse();
1242      if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1243        return ConstantInt::getTrue();
1244      break;
1245    }
1246  } else {
1247    // Evaluate the relation between the two constants, per the predicate.
1248    switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1249    default: assert(0 && "Unknown relational!");
1250    case ICmpInst::BAD_ICMP_PREDICATE:
1251      break;  // Couldn't determine anything about these constants.
1252    case ICmpInst::ICMP_EQ:   // We know the constants are equal!
1253      // If we know the constants are equal, we can decide the result of this
1254      // computation precisely.
1255      return ConstantInt::get(Type::Int1Ty,
1256                              pred == ICmpInst::ICMP_EQ  ||
1257                              pred == ICmpInst::ICMP_ULE ||
1258                              pred == ICmpInst::ICMP_SLE ||
1259                              pred == ICmpInst::ICMP_UGE ||
1260                              pred == ICmpInst::ICMP_SGE);
1261    case ICmpInst::ICMP_ULT:
1262      // If we know that C1 < C2, we can decide the result of this computation
1263      // precisely.
1264      return ConstantInt::get(Type::Int1Ty,
1265                              pred == ICmpInst::ICMP_ULT ||
1266                              pred == ICmpInst::ICMP_NE  ||
1267                              pred == ICmpInst::ICMP_ULE);
1268    case ICmpInst::ICMP_SLT:
1269      // If we know that C1 < C2, we can decide the result of this computation
1270      // precisely.
1271      return ConstantInt::get(Type::Int1Ty,
1272                              pred == ICmpInst::ICMP_SLT ||
1273                              pred == ICmpInst::ICMP_NE  ||
1274                              pred == ICmpInst::ICMP_SLE);
1275    case ICmpInst::ICMP_UGT:
1276      // If we know that C1 > C2, we can decide the result of this computation
1277      // precisely.
1278      return ConstantInt::get(Type::Int1Ty,
1279                              pred == ICmpInst::ICMP_UGT ||
1280                              pred == ICmpInst::ICMP_NE  ||
1281                              pred == ICmpInst::ICMP_UGE);
1282    case ICmpInst::ICMP_SGT:
1283      // If we know that C1 > C2, we can decide the result of this computation
1284      // precisely.
1285      return ConstantInt::get(Type::Int1Ty,
1286                              pred == ICmpInst::ICMP_SGT ||
1287                              pred == ICmpInst::ICMP_NE  ||
1288                              pred == ICmpInst::ICMP_SGE);
1289    case ICmpInst::ICMP_ULE:
1290      // If we know that C1 <= C2, we can only partially decide this relation.
1291      if (pred == ICmpInst::ICMP_UGT) return ConstantInt::getFalse();
1292      if (pred == ICmpInst::ICMP_ULT) return ConstantInt::getTrue();
1293      break;
1294    case ICmpInst::ICMP_SLE:
1295      // If we know that C1 <= C2, we can only partially decide this relation.
1296      if (pred == ICmpInst::ICMP_SGT) return ConstantInt::getFalse();
1297      if (pred == ICmpInst::ICMP_SLT) return ConstantInt::getTrue();
1298      break;
1299
1300    case ICmpInst::ICMP_UGE:
1301      // If we know that C1 >= C2, we can only partially decide this relation.
1302      if (pred == ICmpInst::ICMP_ULT) return ConstantInt::getFalse();
1303      if (pred == ICmpInst::ICMP_UGT) return ConstantInt::getTrue();
1304      break;
1305    case ICmpInst::ICMP_SGE:
1306      // If we know that C1 >= C2, we can only partially decide this relation.
1307      if (pred == ICmpInst::ICMP_SLT) return ConstantInt::getFalse();
1308      if (pred == ICmpInst::ICMP_SGT) return ConstantInt::getTrue();
1309      break;
1310
1311    case ICmpInst::ICMP_NE:
1312      // If we know that C1 != C2, we can only partially decide this relation.
1313      if (pred == ICmpInst::ICMP_EQ) return ConstantInt::getFalse();
1314      if (pred == ICmpInst::ICMP_NE) return ConstantInt::getTrue();
1315      break;
1316    }
1317
1318    if (!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) {
1319      // If C2 is a constant expr and C1 isn't, flop them around and fold the
1320      // other way if possible.
1321      switch (pred) {
1322      case ICmpInst::ICMP_EQ:
1323      case ICmpInst::ICMP_NE:
1324        // No change of predicate required.
1325        return ConstantFoldCompareInstruction(pred, C2, C1);
1326
1327      case ICmpInst::ICMP_ULT:
1328      case ICmpInst::ICMP_SLT:
1329      case ICmpInst::ICMP_UGT:
1330      case ICmpInst::ICMP_SGT:
1331      case ICmpInst::ICMP_ULE:
1332      case ICmpInst::ICMP_SLE:
1333      case ICmpInst::ICMP_UGE:
1334      case ICmpInst::ICMP_SGE:
1335        // Change the predicate as necessary to swap the operands.
1336        pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1337        return ConstantFoldCompareInstruction(pred, C2, C1);
1338
1339      default:  // These predicates cannot be flopped around.
1340        break;
1341      }
1342    }
1343  }
1344  return 0;
1345}
1346
1347Constant *llvm::ConstantFoldGetElementPtr(const Constant *C,
1348                                          Constant* const *Idxs,
1349                                          unsigned NumIdx) {
1350  if (NumIdx == 0 ||
1351      (NumIdx == 1 && Idxs[0]->isNullValue()))
1352    return const_cast<Constant*>(C);
1353
1354  if (isa<UndefValue>(C)) {
1355    const Type *Ty = GetElementPtrInst::getIndexedType(C->getType(),
1356                                                       (Value**)Idxs, NumIdx,
1357                                                       true);
1358    assert(Ty != 0 && "Invalid indices for GEP!");
1359    return UndefValue::get(PointerType::get(Ty));
1360  }
1361
1362  Constant *Idx0 = Idxs[0];
1363  if (C->isNullValue()) {
1364    bool isNull = true;
1365    for (unsigned i = 0, e = NumIdx; i != e; ++i)
1366      if (!Idxs[i]->isNullValue()) {
1367        isNull = false;
1368        break;
1369      }
1370    if (isNull) {
1371      const Type *Ty = GetElementPtrInst::getIndexedType(C->getType(),
1372                                                         (Value**)Idxs, NumIdx,
1373                                                         true);
1374      assert(Ty != 0 && "Invalid indices for GEP!");
1375      return ConstantPointerNull::get(PointerType::get(Ty));
1376    }
1377  }
1378
1379  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(const_cast<Constant*>(C))) {
1380    // Combine Indices - If the source pointer to this getelementptr instruction
1381    // is a getelementptr instruction, combine the indices of the two
1382    // getelementptr instructions into a single instruction.
1383    //
1384    if (CE->getOpcode() == Instruction::GetElementPtr) {
1385      const Type *LastTy = 0;
1386      for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
1387           I != E; ++I)
1388        LastTy = *I;
1389
1390      if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) {
1391        SmallVector<Value*, 16> NewIndices;
1392        NewIndices.reserve(NumIdx + CE->getNumOperands());
1393        for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
1394          NewIndices.push_back(CE->getOperand(i));
1395
1396        // Add the last index of the source with the first index of the new GEP.
1397        // Make sure to handle the case when they are actually different types.
1398        Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
1399        // Otherwise it must be an array.
1400        if (!Idx0->isNullValue()) {
1401          const Type *IdxTy = Combined->getType();
1402          if (IdxTy != Idx0->getType()) {
1403            Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Type::Int64Ty);
1404            Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined,
1405                                                          Type::Int64Ty);
1406            Combined = ConstantExpr::get(Instruction::Add, C1, C2);
1407          } else {
1408            Combined =
1409              ConstantExpr::get(Instruction::Add, Idx0, Combined);
1410          }
1411        }
1412
1413        NewIndices.push_back(Combined);
1414        NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx);
1415        return ConstantExpr::getGetElementPtr(CE->getOperand(0), &NewIndices[0],
1416                                              NewIndices.size());
1417      }
1418    }
1419
1420    // Implement folding of:
1421    //    int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*),
1422    //                        long 0, long 0)
1423    // To: int* getelementptr ([3 x int]* %X, long 0, long 0)
1424    //
1425    if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue())
1426      if (const PointerType *SPT =
1427          dyn_cast<PointerType>(CE->getOperand(0)->getType()))
1428        if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
1429          if (const ArrayType *CAT =
1430        dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
1431            if (CAT->getElementType() == SAT->getElementType())
1432              return ConstantExpr::getGetElementPtr(
1433                      (Constant*)CE->getOperand(0), Idxs, NumIdx);
1434  }
1435  return 0;
1436}
1437
1438