• 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 DataLayout, and the pieces that do. This is to avoid
16 // a dependence in IR on Target.
17 //
18 //===----------------------------------------------------------------------===//
19 
20 #include "ConstantFold.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DerivedTypes.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Operator.h"
30 #include "llvm/Support/Compiler.h"
31 #include "llvm/Support/ErrorHandling.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 vector Constant 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(Constant * CV,VectorType * DstTy)44 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) {
45 
46   if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
47   if (CV->isNullValue()) return Constant::getNullValue(DstTy);
48 
49   // If this cast changes element count then we can't handle it here:
50   // doing so requires endianness information.  This should be handled by
51   // Analysis/ConstantFolding.cpp
52   unsigned NumElts = DstTy->getNumElements();
53   if (NumElts != CV->getType()->getVectorNumElements())
54     return nullptr;
55 
56   Type *DstEltTy = DstTy->getElementType();
57 
58   SmallVector<Constant*, 16> Result;
59   Type *Ty = IntegerType::get(CV->getContext(), 32);
60   for (unsigned i = 0; i != NumElts; ++i) {
61     Constant *C =
62       ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i));
63     C = ConstantExpr::getBitCast(C, DstEltTy);
64     Result.push_back(C);
65   }
66 
67   return ConstantVector::get(Result);
68 }
69 
70 /// This function determines which opcode to use to fold two constant cast
71 /// expressions together. It uses CastInst::isEliminableCastPair to determine
72 /// the opcode. Consequently its just a wrapper around that function.
73 /// @brief Determine if it is valid to fold a cast of a cast
74 static unsigned
foldConstantCastPair(unsigned opc,ConstantExpr * Op,Type * DstTy)75 foldConstantCastPair(
76   unsigned opc,          ///< opcode of the second cast constant expression
77   ConstantExpr *Op,      ///< the first cast constant expression
78   Type *DstTy            ///< destination type of the first cast
79 ) {
80   assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
81   assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
82   assert(CastInst::isCast(opc) && "Invalid cast opcode");
83 
84   // The the types and opcodes for the two Cast constant expressions
85   Type *SrcTy = Op->getOperand(0)->getType();
86   Type *MidTy = Op->getType();
87   Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
88   Instruction::CastOps secondOp = Instruction::CastOps(opc);
89 
90   // Assume that pointers are never more than 64 bits wide, and only use this
91   // for the middle type. Otherwise we could end up folding away illegal
92   // bitcasts between address spaces with different sizes.
93   IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext());
94 
95   // Let CastInst::isEliminableCastPair do the heavy lifting.
96   return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
97                                         nullptr, FakeIntPtrTy, nullptr);
98 }
99 
FoldBitCast(Constant * V,Type * DestTy)100 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
101   Type *SrcTy = V->getType();
102   if (SrcTy == DestTy)
103     return V; // no-op cast
104 
105   // Check to see if we are casting a pointer to an aggregate to a pointer to
106   // the first element.  If so, return the appropriate GEP instruction.
107   if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
108     if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
109       if (PTy->getAddressSpace() == DPTy->getAddressSpace()
110           && DPTy->getElementType()->isSized()) {
111         SmallVector<Value*, 8> IdxList;
112         Value *Zero =
113           Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
114         IdxList.push_back(Zero);
115         Type *ElTy = PTy->getElementType();
116         while (ElTy != DPTy->getElementType()) {
117           if (StructType *STy = dyn_cast<StructType>(ElTy)) {
118             if (STy->getNumElements() == 0) break;
119             ElTy = STy->getElementType(0);
120             IdxList.push_back(Zero);
121           } else if (SequentialType *STy =
122                      dyn_cast<SequentialType>(ElTy)) {
123             if (ElTy->isPointerTy()) break;  // Can't index into pointers!
124             ElTy = STy->getElementType();
125             IdxList.push_back(Zero);
126           } else {
127             break;
128           }
129         }
130 
131         if (ElTy == DPTy->getElementType())
132           // This GEP is inbounds because all indices are zero.
133           return ConstantExpr::getInBoundsGetElementPtr(V, IdxList);
134       }
135 
136   // Handle casts from one vector constant to another.  We know that the src
137   // and dest type have the same size (otherwise its an illegal cast).
138   if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
139     if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
140       assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
141              "Not cast between same sized vectors!");
142       SrcTy = nullptr;
143       // First, check for null.  Undef is already handled.
144       if (isa<ConstantAggregateZero>(V))
145         return Constant::getNullValue(DestTy);
146 
147       // Handle ConstantVector and ConstantAggregateVector.
148       return BitCastConstantVector(V, DestPTy);
149     }
150 
151     // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
152     // This allows for other simplifications (although some of them
153     // can only be handled by Analysis/ConstantFolding.cpp).
154     if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
155       return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
156   }
157 
158   // Finally, implement bitcast folding now.   The code below doesn't handle
159   // bitcast right.
160   if (isa<ConstantPointerNull>(V))  // ptr->ptr cast.
161     return ConstantPointerNull::get(cast<PointerType>(DestTy));
162 
163   // Handle integral constant input.
164   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
165     if (DestTy->isIntegerTy())
166       // Integral -> Integral. This is a no-op because the bit widths must
167       // be the same. Consequently, we just fold to V.
168       return V;
169 
170     if (DestTy->isFloatingPointTy())
171       return ConstantFP::get(DestTy->getContext(),
172                              APFloat(DestTy->getFltSemantics(),
173                                      CI->getValue()));
174 
175     // Otherwise, can't fold this (vector?)
176     return nullptr;
177   }
178 
179   // Handle ConstantFP input: FP -> Integral.
180   if (ConstantFP *FP = dyn_cast<ConstantFP>(V))
181     return ConstantInt::get(FP->getContext(),
182                             FP->getValueAPF().bitcastToAPInt());
183 
184   return nullptr;
185 }
186 
187 
188 /// ExtractConstantBytes - V is an integer constant which only has a subset of
189 /// its bytes used.  The bytes used are indicated by ByteStart (which is the
190 /// first byte used, counting from the least significant byte) and ByteSize,
191 /// which is the number of bytes used.
192 ///
193 /// This function analyzes the specified constant to see if the specified byte
194 /// range can be returned as a simplified constant.  If so, the constant is
195 /// returned, otherwise null is returned.
196 ///
ExtractConstantBytes(Constant * C,unsigned ByteStart,unsigned ByteSize)197 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
198                                       unsigned ByteSize) {
199   assert(C->getType()->isIntegerTy() &&
200          (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
201          "Non-byte sized integer input");
202   unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
203   assert(ByteSize && "Must be accessing some piece");
204   assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
205   assert(ByteSize != CSize && "Should not extract everything");
206 
207   // Constant Integers are simple.
208   if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
209     APInt V = CI->getValue();
210     if (ByteStart)
211       V = V.lshr(ByteStart*8);
212     V = V.trunc(ByteSize*8);
213     return ConstantInt::get(CI->getContext(), V);
214   }
215 
216   // In the input is a constant expr, we might be able to recursively simplify.
217   // If not, we definitely can't do anything.
218   ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
219   if (!CE) return nullptr;
220 
221   switch (CE->getOpcode()) {
222   default: return nullptr;
223   case Instruction::Or: {
224     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
225     if (!RHS)
226       return nullptr;
227 
228     // X | -1 -> -1.
229     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
230       if (RHSC->isAllOnesValue())
231         return RHSC;
232 
233     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
234     if (!LHS)
235       return nullptr;
236     return ConstantExpr::getOr(LHS, RHS);
237   }
238   case Instruction::And: {
239     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
240     if (!RHS)
241       return nullptr;
242 
243     // X & 0 -> 0.
244     if (RHS->isNullValue())
245       return RHS;
246 
247     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
248     if (!LHS)
249       return nullptr;
250     return ConstantExpr::getAnd(LHS, RHS);
251   }
252   case Instruction::LShr: {
253     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
254     if (!Amt)
255       return nullptr;
256     unsigned ShAmt = Amt->getZExtValue();
257     // Cannot analyze non-byte shifts.
258     if ((ShAmt & 7) != 0)
259       return nullptr;
260     ShAmt >>= 3;
261 
262     // If the extract is known to be all zeros, return zero.
263     if (ByteStart >= CSize-ShAmt)
264       return Constant::getNullValue(IntegerType::get(CE->getContext(),
265                                                      ByteSize*8));
266     // If the extract is known to be fully in the input, extract it.
267     if (ByteStart+ByteSize+ShAmt <= CSize)
268       return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
269 
270     // TODO: Handle the 'partially zero' case.
271     return nullptr;
272   }
273 
274   case Instruction::Shl: {
275     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
276     if (!Amt)
277       return nullptr;
278     unsigned ShAmt = Amt->getZExtValue();
279     // Cannot analyze non-byte shifts.
280     if ((ShAmt & 7) != 0)
281       return nullptr;
282     ShAmt >>= 3;
283 
284     // If the extract is known to be all zeros, return zero.
285     if (ByteStart+ByteSize <= ShAmt)
286       return Constant::getNullValue(IntegerType::get(CE->getContext(),
287                                                      ByteSize*8));
288     // If the extract is known to be fully in the input, extract it.
289     if (ByteStart >= ShAmt)
290       return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
291 
292     // TODO: Handle the 'partially zero' case.
293     return nullptr;
294   }
295 
296   case Instruction::ZExt: {
297     unsigned SrcBitSize =
298       cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
299 
300     // If extracting something that is completely zero, return 0.
301     if (ByteStart*8 >= SrcBitSize)
302       return Constant::getNullValue(IntegerType::get(CE->getContext(),
303                                                      ByteSize*8));
304 
305     // If exactly extracting the input, return it.
306     if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
307       return CE->getOperand(0);
308 
309     // If extracting something completely in the input, if if the input is a
310     // multiple of 8 bits, recurse.
311     if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
312       return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
313 
314     // Otherwise, if extracting a subset of the input, which is not multiple of
315     // 8 bits, do a shift and trunc to get the bits.
316     if ((ByteStart+ByteSize)*8 < SrcBitSize) {
317       assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
318       Constant *Res = CE->getOperand(0);
319       if (ByteStart)
320         Res = ConstantExpr::getLShr(Res,
321                                  ConstantInt::get(Res->getType(), ByteStart*8));
322       return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
323                                                           ByteSize*8));
324     }
325 
326     // TODO: Handle the 'partially zero' case.
327     return nullptr;
328   }
329   }
330 }
331 
332 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
333 /// on Ty, with any known factors factored out. If Folded is false,
334 /// return null if no factoring was possible, to avoid endlessly
335 /// bouncing an unfoldable expression back into the top-level folder.
336 ///
getFoldedSizeOf(Type * Ty,Type * DestTy,bool Folded)337 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy,
338                                  bool Folded) {
339   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
340     Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
341     Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
342     return ConstantExpr::getNUWMul(E, N);
343   }
344 
345   if (StructType *STy = dyn_cast<StructType>(Ty))
346     if (!STy->isPacked()) {
347       unsigned NumElems = STy->getNumElements();
348       // An empty struct has size zero.
349       if (NumElems == 0)
350         return ConstantExpr::getNullValue(DestTy);
351       // Check for a struct with all members having the same size.
352       Constant *MemberSize =
353         getFoldedSizeOf(STy->getElementType(0), DestTy, true);
354       bool AllSame = true;
355       for (unsigned i = 1; i != NumElems; ++i)
356         if (MemberSize !=
357             getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
358           AllSame = false;
359           break;
360         }
361       if (AllSame) {
362         Constant *N = ConstantInt::get(DestTy, NumElems);
363         return ConstantExpr::getNUWMul(MemberSize, N);
364       }
365     }
366 
367   // Pointer size doesn't depend on the pointee type, so canonicalize them
368   // to an arbitrary pointee.
369   if (PointerType *PTy = dyn_cast<PointerType>(Ty))
370     if (!PTy->getElementType()->isIntegerTy(1))
371       return
372         getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
373                                          PTy->getAddressSpace()),
374                         DestTy, true);
375 
376   // If there's no interesting folding happening, bail so that we don't create
377   // a constant that looks like it needs folding but really doesn't.
378   if (!Folded)
379     return nullptr;
380 
381   // Base case: Get a regular sizeof expression.
382   Constant *C = ConstantExpr::getSizeOf(Ty);
383   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
384                                                     DestTy, false),
385                             C, DestTy);
386   return C;
387 }
388 
389 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
390 /// on Ty, with any known factors factored out. If Folded is false,
391 /// return null if no factoring was possible, to avoid endlessly
392 /// bouncing an unfoldable expression back into the top-level folder.
393 ///
getFoldedAlignOf(Type * Ty,Type * DestTy,bool Folded)394 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy,
395                                   bool Folded) {
396   // The alignment of an array is equal to the alignment of the
397   // array element. Note that this is not always true for vectors.
398   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
399     Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
400     C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
401                                                       DestTy,
402                                                       false),
403                               C, DestTy);
404     return C;
405   }
406 
407   if (StructType *STy = dyn_cast<StructType>(Ty)) {
408     // Packed structs always have an alignment of 1.
409     if (STy->isPacked())
410       return ConstantInt::get(DestTy, 1);
411 
412     // Otherwise, struct alignment is the maximum alignment of any member.
413     // Without target data, we can't compare much, but we can check to see
414     // if all the members have the same alignment.
415     unsigned NumElems = STy->getNumElements();
416     // An empty struct has minimal alignment.
417     if (NumElems == 0)
418       return ConstantInt::get(DestTy, 1);
419     // Check for a struct with all members having the same alignment.
420     Constant *MemberAlign =
421       getFoldedAlignOf(STy->getElementType(0), DestTy, true);
422     bool AllSame = true;
423     for (unsigned i = 1; i != NumElems; ++i)
424       if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
425         AllSame = false;
426         break;
427       }
428     if (AllSame)
429       return MemberAlign;
430   }
431 
432   // Pointer alignment doesn't depend on the pointee type, so canonicalize them
433   // to an arbitrary pointee.
434   if (PointerType *PTy = dyn_cast<PointerType>(Ty))
435     if (!PTy->getElementType()->isIntegerTy(1))
436       return
437         getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
438                                                            1),
439                                           PTy->getAddressSpace()),
440                          DestTy, true);
441 
442   // If there's no interesting folding happening, bail so that we don't create
443   // a constant that looks like it needs folding but really doesn't.
444   if (!Folded)
445     return nullptr;
446 
447   // Base case: Get a regular alignof expression.
448   Constant *C = ConstantExpr::getAlignOf(Ty);
449   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
450                                                     DestTy, false),
451                             C, DestTy);
452   return C;
453 }
454 
455 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
456 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
457 /// return null if no factoring was possible, to avoid endlessly
458 /// bouncing an unfoldable expression back into the top-level folder.
459 ///
getFoldedOffsetOf(Type * Ty,Constant * FieldNo,Type * DestTy,bool Folded)460 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo,
461                                    Type *DestTy,
462                                    bool Folded) {
463   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
464     Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
465                                                                 DestTy, false),
466                                         FieldNo, DestTy);
467     Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
468     return ConstantExpr::getNUWMul(E, N);
469   }
470 
471   if (StructType *STy = dyn_cast<StructType>(Ty))
472     if (!STy->isPacked()) {
473       unsigned NumElems = STy->getNumElements();
474       // An empty struct has no members.
475       if (NumElems == 0)
476         return nullptr;
477       // Check for a struct with all members having the same size.
478       Constant *MemberSize =
479         getFoldedSizeOf(STy->getElementType(0), DestTy, true);
480       bool AllSame = true;
481       for (unsigned i = 1; i != NumElems; ++i)
482         if (MemberSize !=
483             getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
484           AllSame = false;
485           break;
486         }
487       if (AllSame) {
488         Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
489                                                                     false,
490                                                                     DestTy,
491                                                                     false),
492                                             FieldNo, DestTy);
493         return ConstantExpr::getNUWMul(MemberSize, N);
494       }
495     }
496 
497   // If there's no interesting folding happening, bail so that we don't create
498   // a constant that looks like it needs folding but really doesn't.
499   if (!Folded)
500     return nullptr;
501 
502   // Base case: Get a regular offsetof expression.
503   Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
504   C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
505                                                     DestTy, false),
506                             C, DestTy);
507   return C;
508 }
509 
ConstantFoldCastInstruction(unsigned opc,Constant * V,Type * DestTy)510 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
511                                             Type *DestTy) {
512   if (isa<UndefValue>(V)) {
513     // zext(undef) = 0, because the top bits will be zero.
514     // sext(undef) = 0, because the top bits will all be the same.
515     // [us]itofp(undef) = 0, because the result value is bounded.
516     if (opc == Instruction::ZExt || opc == Instruction::SExt ||
517         opc == Instruction::UIToFP || opc == Instruction::SIToFP)
518       return Constant::getNullValue(DestTy);
519     return UndefValue::get(DestTy);
520   }
521 
522   if (V->isNullValue() && !DestTy->isX86_MMXTy())
523     return Constant::getNullValue(DestTy);
524 
525   // If the cast operand is a constant expression, there's a few things we can
526   // do to try to simplify it.
527   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
528     if (CE->isCast()) {
529       // Try hard to fold cast of cast because they are often eliminable.
530       if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
531         return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
532     } else if (CE->getOpcode() == Instruction::GetElementPtr &&
533                // Do not fold addrspacecast (gep 0, .., 0). It might make the
534                // addrspacecast uncanonicalized.
535                opc != Instruction::AddrSpaceCast) {
536       // If all of the indexes in the GEP are null values, there is no pointer
537       // adjustment going on.  We might as well cast the source pointer.
538       bool isAllNull = true;
539       for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
540         if (!CE->getOperand(i)->isNullValue()) {
541           isAllNull = false;
542           break;
543         }
544       if (isAllNull)
545         // This is casting one pointer type to another, always BitCast
546         return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
547     }
548   }
549 
550   // If the cast operand is a constant vector, perform the cast by
551   // operating on each element. In the cast of bitcasts, the element
552   // count may be mismatched; don't attempt to handle that here.
553   if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
554       DestTy->isVectorTy() &&
555       DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) {
556     SmallVector<Constant*, 16> res;
557     VectorType *DestVecTy = cast<VectorType>(DestTy);
558     Type *DstEltTy = DestVecTy->getElementType();
559     Type *Ty = IntegerType::get(V->getContext(), 32);
560     for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) {
561       Constant *C =
562         ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
563       res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
564     }
565     return ConstantVector::get(res);
566   }
567 
568   // We actually have to do a cast now. Perform the cast according to the
569   // opcode specified.
570   switch (opc) {
571   default:
572     llvm_unreachable("Failed to cast constant expression");
573   case Instruction::FPTrunc:
574   case Instruction::FPExt:
575     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
576       bool ignored;
577       APFloat Val = FPC->getValueAPF();
578       Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf :
579                   DestTy->isFloatTy() ? APFloat::IEEEsingle :
580                   DestTy->isDoubleTy() ? APFloat::IEEEdouble :
581                   DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
582                   DestTy->isFP128Ty() ? APFloat::IEEEquad :
583                   DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble :
584                   APFloat::Bogus,
585                   APFloat::rmNearestTiesToEven, &ignored);
586       return ConstantFP::get(V->getContext(), Val);
587     }
588     return nullptr; // Can't fold.
589   case Instruction::FPToUI:
590   case Instruction::FPToSI:
591     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
592       const APFloat &V = FPC->getValueAPF();
593       bool ignored;
594       uint64_t x[2];
595       uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
596       (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
597                                 APFloat::rmTowardZero, &ignored);
598       APInt Val(DestBitWidth, x);
599       return ConstantInt::get(FPC->getContext(), Val);
600     }
601     return nullptr; // Can't fold.
602   case Instruction::IntToPtr:   //always treated as unsigned
603     if (V->isNullValue())       // Is it an integral null value?
604       return ConstantPointerNull::get(cast<PointerType>(DestTy));
605     return nullptr;                   // Other pointer types cannot be casted
606   case Instruction::PtrToInt:   // always treated as unsigned
607     // Is it a null pointer value?
608     if (V->isNullValue())
609       return ConstantInt::get(DestTy, 0);
610     // If this is a sizeof-like expression, pull out multiplications by
611     // known factors to expose them to subsequent folding. If it's an
612     // alignof-like expression, factor out known factors.
613     if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
614       if (CE->getOpcode() == Instruction::GetElementPtr &&
615           CE->getOperand(0)->isNullValue()) {
616         Type *Ty =
617           cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
618         if (CE->getNumOperands() == 2) {
619           // Handle a sizeof-like expression.
620           Constant *Idx = CE->getOperand(1);
621           bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
622           if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
623             Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
624                                                                 DestTy, false),
625                                         Idx, DestTy);
626             return ConstantExpr::getMul(C, Idx);
627           }
628         } else if (CE->getNumOperands() == 3 &&
629                    CE->getOperand(1)->isNullValue()) {
630           // Handle an alignof-like expression.
631           if (StructType *STy = dyn_cast<StructType>(Ty))
632             if (!STy->isPacked()) {
633               ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
634               if (CI->isOne() &&
635                   STy->getNumElements() == 2 &&
636                   STy->getElementType(0)->isIntegerTy(1)) {
637                 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
638               }
639             }
640           // Handle an offsetof-like expression.
641           if (Ty->isStructTy() || Ty->isArrayTy()) {
642             if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
643                                                 DestTy, false))
644               return C;
645           }
646         }
647       }
648     // Other pointer types cannot be casted
649     return nullptr;
650   case Instruction::UIToFP:
651   case Instruction::SIToFP:
652     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
653       APInt api = CI->getValue();
654       APFloat apf(DestTy->getFltSemantics(),
655                   APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
656       (void)apf.convertFromAPInt(api,
657                                  opc==Instruction::SIToFP,
658                                  APFloat::rmNearestTiesToEven);
659       return ConstantFP::get(V->getContext(), apf);
660     }
661     return nullptr;
662   case Instruction::ZExt:
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().zext(BitWidth));
667     }
668     return nullptr;
669   case Instruction::SExt:
670     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
671       uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
672       return ConstantInt::get(V->getContext(),
673                               CI->getValue().sext(BitWidth));
674     }
675     return nullptr;
676   case Instruction::Trunc: {
677     uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
678     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
679       return ConstantInt::get(V->getContext(),
680                               CI->getValue().trunc(DestBitWidth));
681     }
682 
683     // The input must be a constantexpr.  See if we can simplify this based on
684     // the bytes we are demanding.  Only do this if the source and dest are an
685     // even multiple of a byte.
686     if ((DestBitWidth & 7) == 0 &&
687         (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
688       if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
689         return Res;
690 
691     return nullptr;
692   }
693   case Instruction::BitCast:
694     return FoldBitCast(V, DestTy);
695   case Instruction::AddrSpaceCast:
696     return nullptr;
697   }
698 }
699 
ConstantFoldSelectInstruction(Constant * Cond,Constant * V1,Constant * V2)700 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
701                                               Constant *V1, Constant *V2) {
702   // Check for i1 and vector true/false conditions.
703   if (Cond->isNullValue()) return V2;
704   if (Cond->isAllOnesValue()) return V1;
705 
706   // If the condition is a vector constant, fold the result elementwise.
707   if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
708     SmallVector<Constant*, 16> Result;
709     Type *Ty = IntegerType::get(CondV->getContext(), 32);
710     for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){
711       Constant *V;
712       Constant *V1Element = ConstantExpr::getExtractElement(V1,
713                                                     ConstantInt::get(Ty, i));
714       Constant *V2Element = ConstantExpr::getExtractElement(V2,
715                                                     ConstantInt::get(Ty, i));
716       Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i));
717       if (V1Element == V2Element) {
718         V = V1Element;
719       } else if (isa<UndefValue>(Cond)) {
720         V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
721       } else {
722         if (!isa<ConstantInt>(Cond)) break;
723         V = Cond->isNullValue() ? V2Element : V1Element;
724       }
725       Result.push_back(V);
726     }
727 
728     // If we were able to build the vector, return it.
729     if (Result.size() == V1->getType()->getVectorNumElements())
730       return ConstantVector::get(Result);
731   }
732 
733   if (isa<UndefValue>(Cond)) {
734     if (isa<UndefValue>(V1)) return V1;
735     return V2;
736   }
737   if (isa<UndefValue>(V1)) return V2;
738   if (isa<UndefValue>(V2)) return V1;
739   if (V1 == V2) return V1;
740 
741   if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
742     if (TrueVal->getOpcode() == Instruction::Select)
743       if (TrueVal->getOperand(0) == Cond)
744         return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
745   }
746   if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
747     if (FalseVal->getOpcode() == Instruction::Select)
748       if (FalseVal->getOperand(0) == Cond)
749         return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
750   }
751 
752   return nullptr;
753 }
754 
ConstantFoldExtractElementInstruction(Constant * Val,Constant * Idx)755 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
756                                                       Constant *Idx) {
757   if (isa<UndefValue>(Val))  // ee(undef, x) -> undef
758     return UndefValue::get(Val->getType()->getVectorElementType());
759   if (Val->isNullValue())  // ee(zero, x) -> zero
760     return Constant::getNullValue(Val->getType()->getVectorElementType());
761   // ee({w,x,y,z}, undef) -> undef
762   if (isa<UndefValue>(Idx))
763     return UndefValue::get(Val->getType()->getVectorElementType());
764 
765   if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
766     uint64_t Index = CIdx->getZExtValue();
767     // ee({w,x,y,z}, wrong_value) -> undef
768     if (Index >= Val->getType()->getVectorNumElements())
769       return UndefValue::get(Val->getType()->getVectorElementType());
770     return Val->getAggregateElement(Index);
771   }
772   return nullptr;
773 }
774 
ConstantFoldInsertElementInstruction(Constant * Val,Constant * Elt,Constant * Idx)775 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
776                                                      Constant *Elt,
777                                                      Constant *Idx) {
778   ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
779   if (!CIdx) return nullptr;
780   const APInt &IdxVal = CIdx->getValue();
781 
782   SmallVector<Constant*, 16> Result;
783   Type *Ty = IntegerType::get(Val->getContext(), 32);
784   for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){
785     if (i == IdxVal) {
786       Result.push_back(Elt);
787       continue;
788     }
789 
790     Constant *C =
791       ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
792     Result.push_back(C);
793   }
794 
795   return ConstantVector::get(Result);
796 }
797 
ConstantFoldShuffleVectorInstruction(Constant * V1,Constant * V2,Constant * Mask)798 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
799                                                      Constant *V2,
800                                                      Constant *Mask) {
801   unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
802   Type *EltTy = V1->getType()->getVectorElementType();
803 
804   // Undefined shuffle mask -> undefined value.
805   if (isa<UndefValue>(Mask))
806     return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
807 
808   // Don't break the bitcode reader hack.
809   if (isa<ConstantExpr>(Mask)) return nullptr;
810 
811   unsigned SrcNumElts = V1->getType()->getVectorNumElements();
812 
813   // Loop over the shuffle mask, evaluating each element.
814   SmallVector<Constant*, 32> Result;
815   for (unsigned i = 0; i != MaskNumElts; ++i) {
816     int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
817     if (Elt == -1) {
818       Result.push_back(UndefValue::get(EltTy));
819       continue;
820     }
821     Constant *InElt;
822     if (unsigned(Elt) >= SrcNumElts*2)
823       InElt = UndefValue::get(EltTy);
824     else if (unsigned(Elt) >= SrcNumElts) {
825       Type *Ty = IntegerType::get(V2->getContext(), 32);
826       InElt =
827         ConstantExpr::getExtractElement(V2,
828                                         ConstantInt::get(Ty, Elt - SrcNumElts));
829     } else {
830       Type *Ty = IntegerType::get(V1->getContext(), 32);
831       InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
832     }
833     Result.push_back(InElt);
834   }
835 
836   return ConstantVector::get(Result);
837 }
838 
ConstantFoldExtractValueInstruction(Constant * Agg,ArrayRef<unsigned> Idxs)839 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
840                                                     ArrayRef<unsigned> Idxs) {
841   // Base case: no indices, so return the entire value.
842   if (Idxs.empty())
843     return Agg;
844 
845   if (Constant *C = Agg->getAggregateElement(Idxs[0]))
846     return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
847 
848   return nullptr;
849 }
850 
ConstantFoldInsertValueInstruction(Constant * Agg,Constant * Val,ArrayRef<unsigned> Idxs)851 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
852                                                    Constant *Val,
853                                                    ArrayRef<unsigned> Idxs) {
854   // Base case: no indices, so replace the entire value.
855   if (Idxs.empty())
856     return Val;
857 
858   unsigned NumElts;
859   if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
860     NumElts = ST->getNumElements();
861   else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
862     NumElts = AT->getNumElements();
863   else
864     NumElts = Agg->getType()->getVectorNumElements();
865 
866   SmallVector<Constant*, 32> Result;
867   for (unsigned i = 0; i != NumElts; ++i) {
868     Constant *C = Agg->getAggregateElement(i);
869     if (!C) return nullptr;
870 
871     if (Idxs[0] == i)
872       C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
873 
874     Result.push_back(C);
875   }
876 
877   if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
878     return ConstantStruct::get(ST, Result);
879   if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
880     return ConstantArray::get(AT, Result);
881   return ConstantVector::get(Result);
882 }
883 
884 
ConstantFoldBinaryInstruction(unsigned Opcode,Constant * C1,Constant * C2)885 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
886                                               Constant *C1, Constant *C2) {
887   // Handle UndefValue up front.
888   if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
889     switch (Opcode) {
890     case Instruction::Xor:
891       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
892         // Handle undef ^ undef -> 0 special case. This is a common
893         // idiom (misuse).
894         return Constant::getNullValue(C1->getType());
895       // Fallthrough
896     case Instruction::Add:
897     case Instruction::Sub:
898       return UndefValue::get(C1->getType());
899     case Instruction::And:
900       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
901         return C1;
902       return Constant::getNullValue(C1->getType());   // undef & X -> 0
903     case Instruction::Mul: {
904       ConstantInt *CI;
905       // X * undef -> undef   if X is odd or undef
906       if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) ||
907           ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) ||
908           (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
909         return UndefValue::get(C1->getType());
910 
911       // X * undef -> 0       otherwise
912       return Constant::getNullValue(C1->getType());
913     }
914     case Instruction::UDiv:
915     case Instruction::SDiv:
916       // undef / 1 -> undef
917       if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv)
918         if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2))
919           if (CI2->isOne())
920             return C1;
921       // FALL THROUGH
922     case Instruction::URem:
923     case Instruction::SRem:
924       if (!isa<UndefValue>(C2))                    // undef / X -> 0
925         return Constant::getNullValue(C1->getType());
926       return C2;                                   // X / undef -> undef
927     case Instruction::Or:                          // X | undef -> -1
928       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
929         return C1;
930       return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
931     case Instruction::LShr:
932       if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
933         return C1;                                  // undef lshr undef -> undef
934       return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
935                                                     // undef lshr X -> 0
936     case Instruction::AShr:
937       if (!isa<UndefValue>(C2))                     // undef ashr X --> all ones
938         return Constant::getAllOnesValue(C1->getType());
939       else if (isa<UndefValue>(C1))
940         return C1;                                  // undef ashr undef -> undef
941       else
942         return C1;                                  // X ashr undef --> X
943     case Instruction::Shl:
944       if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
945         return C1;                                  // undef shl undef -> undef
946       // undef << X -> 0   or   X << undef -> 0
947       return Constant::getNullValue(C1->getType());
948     }
949   }
950 
951   // Handle simplifications when the RHS is a constant int.
952   if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
953     switch (Opcode) {
954     case Instruction::Add:
955       if (CI2->equalsInt(0)) return C1;                         // X + 0 == X
956       break;
957     case Instruction::Sub:
958       if (CI2->equalsInt(0)) return C1;                         // X - 0 == X
959       break;
960     case Instruction::Mul:
961       if (CI2->equalsInt(0)) return C2;                         // X * 0 == 0
962       if (CI2->equalsInt(1))
963         return C1;                                              // X * 1 == X
964       break;
965     case Instruction::UDiv:
966     case Instruction::SDiv:
967       if (CI2->equalsInt(1))
968         return C1;                                            // X / 1 == X
969       if (CI2->equalsInt(0))
970         return UndefValue::get(CI2->getType());               // X / 0 == undef
971       break;
972     case Instruction::URem:
973     case Instruction::SRem:
974       if (CI2->equalsInt(1))
975         return Constant::getNullValue(CI2->getType());        // X % 1 == 0
976       if (CI2->equalsInt(0))
977         return UndefValue::get(CI2->getType());               // X % 0 == undef
978       break;
979     case Instruction::And:
980       if (CI2->isZero()) return C2;                           // X & 0 == 0
981       if (CI2->isAllOnesValue())
982         return C1;                                            // X & -1 == X
983 
984       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
985         // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
986         if (CE1->getOpcode() == Instruction::ZExt) {
987           unsigned DstWidth = CI2->getType()->getBitWidth();
988           unsigned SrcWidth =
989             CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
990           APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
991           if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
992             return C1;
993         }
994 
995         // If and'ing the address of a global with a constant, fold it.
996         if (CE1->getOpcode() == Instruction::PtrToInt &&
997             isa<GlobalValue>(CE1->getOperand(0))) {
998           GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
999 
1000           // Functions are at least 4-byte aligned.
1001           unsigned GVAlign = GV->getAlignment();
1002           if (isa<Function>(GV))
1003             GVAlign = std::max(GVAlign, 4U);
1004 
1005           if (GVAlign > 1) {
1006             unsigned DstWidth = CI2->getType()->getBitWidth();
1007             unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1008             APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1009 
1010             // If checking bits we know are clear, return zero.
1011             if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1012               return Constant::getNullValue(CI2->getType());
1013           }
1014         }
1015       }
1016       break;
1017     case Instruction::Or:
1018       if (CI2->equalsInt(0)) return C1;    // X | 0 == X
1019       if (CI2->isAllOnesValue())
1020         return C2;                         // X | -1 == -1
1021       break;
1022     case Instruction::Xor:
1023       if (CI2->equalsInt(0)) return C1;    // X ^ 0 == X
1024 
1025       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1026         switch (CE1->getOpcode()) {
1027         default: break;
1028         case Instruction::ICmp:
1029         case Instruction::FCmp:
1030           // cmp pred ^ true -> cmp !pred
1031           assert(CI2->equalsInt(1));
1032           CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1033           pred = CmpInst::getInversePredicate(pred);
1034           return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1035                                           CE1->getOperand(1));
1036         }
1037       }
1038       break;
1039     case Instruction::AShr:
1040       // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1041       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1042         if (CE1->getOpcode() == Instruction::ZExt)  // Top bits known zero.
1043           return ConstantExpr::getLShr(C1, C2);
1044       break;
1045     }
1046   } else if (isa<ConstantInt>(C1)) {
1047     // If C1 is a ConstantInt and C2 is not, swap the operands.
1048     if (Instruction::isCommutative(Opcode))
1049       return ConstantExpr::get(Opcode, C2, C1);
1050   }
1051 
1052   // At this point we know neither constant is an UndefValue.
1053   if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1054     if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1055       const APInt &C1V = CI1->getValue();
1056       const APInt &C2V = CI2->getValue();
1057       switch (Opcode) {
1058       default:
1059         break;
1060       case Instruction::Add:
1061         return ConstantInt::get(CI1->getContext(), C1V + C2V);
1062       case Instruction::Sub:
1063         return ConstantInt::get(CI1->getContext(), C1V - C2V);
1064       case Instruction::Mul:
1065         return ConstantInt::get(CI1->getContext(), C1V * C2V);
1066       case Instruction::UDiv:
1067         assert(!CI2->isNullValue() && "Div by zero handled above");
1068         return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1069       case Instruction::SDiv:
1070         assert(!CI2->isNullValue() && "Div by zero handled above");
1071         if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1072           return UndefValue::get(CI1->getType());   // MIN_INT / -1 -> undef
1073         return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1074       case Instruction::URem:
1075         assert(!CI2->isNullValue() && "Div by zero handled above");
1076         return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1077       case Instruction::SRem:
1078         assert(!CI2->isNullValue() && "Div by zero handled above");
1079         if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1080           return UndefValue::get(CI1->getType());   // MIN_INT % -1 -> undef
1081         return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1082       case Instruction::And:
1083         return ConstantInt::get(CI1->getContext(), C1V & C2V);
1084       case Instruction::Or:
1085         return ConstantInt::get(CI1->getContext(), C1V | C2V);
1086       case Instruction::Xor:
1087         return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1088       case Instruction::Shl: {
1089         uint32_t shiftAmt = C2V.getZExtValue();
1090         if (shiftAmt < C1V.getBitWidth())
1091           return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1092         else
1093           return UndefValue::get(C1->getType()); // too big shift is undef
1094       }
1095       case Instruction::LShr: {
1096         uint32_t shiftAmt = C2V.getZExtValue();
1097         if (shiftAmt < C1V.getBitWidth())
1098           return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1099         else
1100           return UndefValue::get(C1->getType()); // too big shift is undef
1101       }
1102       case Instruction::AShr: {
1103         uint32_t shiftAmt = C2V.getZExtValue();
1104         if (shiftAmt < C1V.getBitWidth())
1105           return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1106         else
1107           return UndefValue::get(C1->getType()); // too big shift is undef
1108       }
1109       }
1110     }
1111 
1112     switch (Opcode) {
1113     case Instruction::SDiv:
1114     case Instruction::UDiv:
1115     case Instruction::URem:
1116     case Instruction::SRem:
1117     case Instruction::LShr:
1118     case Instruction::AShr:
1119     case Instruction::Shl:
1120       if (CI1->equalsInt(0)) return C1;
1121       break;
1122     default:
1123       break;
1124     }
1125   } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1126     if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1127       APFloat C1V = CFP1->getValueAPF();
1128       APFloat C2V = CFP2->getValueAPF();
1129       APFloat C3V = C1V;  // copy for modification
1130       switch (Opcode) {
1131       default:
1132         break;
1133       case Instruction::FAdd:
1134         (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1135         return ConstantFP::get(C1->getContext(), C3V);
1136       case Instruction::FSub:
1137         (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1138         return ConstantFP::get(C1->getContext(), C3V);
1139       case Instruction::FMul:
1140         (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1141         return ConstantFP::get(C1->getContext(), C3V);
1142       case Instruction::FDiv:
1143         (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1144         return ConstantFP::get(C1->getContext(), C3V);
1145       case Instruction::FRem:
1146         (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1147         return ConstantFP::get(C1->getContext(), C3V);
1148       }
1149     }
1150   } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1151     // Perform elementwise folding.
1152     SmallVector<Constant*, 16> Result;
1153     Type *Ty = IntegerType::get(VTy->getContext(), 32);
1154     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1155       Constant *LHS =
1156         ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1157       Constant *RHS =
1158         ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1159 
1160       Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1161     }
1162 
1163     return ConstantVector::get(Result);
1164   }
1165 
1166   if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1167     // There are many possible foldings we could do here.  We should probably
1168     // at least fold add of a pointer with an integer into the appropriate
1169     // getelementptr.  This will improve alias analysis a bit.
1170 
1171     // Given ((a + b) + c), if (b + c) folds to something interesting, return
1172     // (a + (b + c)).
1173     if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1174       Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1175       if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1176         return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1177     }
1178   } else if (isa<ConstantExpr>(C2)) {
1179     // If C2 is a constant expr and C1 isn't, flop them around and fold the
1180     // other way if possible.
1181     if (Instruction::isCommutative(Opcode))
1182       return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1183   }
1184 
1185   // i1 can be simplified in many cases.
1186   if (C1->getType()->isIntegerTy(1)) {
1187     switch (Opcode) {
1188     case Instruction::Add:
1189     case Instruction::Sub:
1190       return ConstantExpr::getXor(C1, C2);
1191     case Instruction::Mul:
1192       return ConstantExpr::getAnd(C1, C2);
1193     case Instruction::Shl:
1194     case Instruction::LShr:
1195     case Instruction::AShr:
1196       // We can assume that C2 == 0.  If it were one the result would be
1197       // undefined because the shift value is as large as the bitwidth.
1198       return C1;
1199     case Instruction::SDiv:
1200     case Instruction::UDiv:
1201       // We can assume that C2 == 1.  If it were zero the result would be
1202       // undefined through division by zero.
1203       return C1;
1204     case Instruction::URem:
1205     case Instruction::SRem:
1206       // We can assume that C2 == 1.  If it were zero the result would be
1207       // undefined through division by zero.
1208       return ConstantInt::getFalse(C1->getContext());
1209     default:
1210       break;
1211     }
1212   }
1213 
1214   // We don't know how to fold this.
1215   return nullptr;
1216 }
1217 
1218 /// isZeroSizedType - This type is zero sized if its an array or structure of
1219 /// zero sized types.  The only leaf zero sized type is an empty structure.
isMaybeZeroSizedType(Type * Ty)1220 static bool isMaybeZeroSizedType(Type *Ty) {
1221   if (StructType *STy = dyn_cast<StructType>(Ty)) {
1222     if (STy->isOpaque()) return true;  // Can't say.
1223 
1224     // If all of elements have zero size, this does too.
1225     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1226       if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1227     return true;
1228 
1229   } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1230     return isMaybeZeroSizedType(ATy->getElementType());
1231   }
1232   return false;
1233 }
1234 
1235 /// IdxCompare - Compare the two constants as though they were getelementptr
1236 /// indices.  This allows coersion of the types to be the same thing.
1237 ///
1238 /// If the two constants are the "same" (after coersion), return 0.  If the
1239 /// first is less than the second, return -1, if the second is less than the
1240 /// first, return 1.  If the constants are not integral, return -2.
1241 ///
IdxCompare(Constant * C1,Constant * C2,Type * ElTy)1242 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1243   if (C1 == C2) return 0;
1244 
1245   // Ok, we found a different index.  If they are not ConstantInt, we can't do
1246   // anything with them.
1247   if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1248     return -2; // don't know!
1249 
1250   // Ok, we have two differing integer indices.  Sign extend them to be the same
1251   // type.  Long is always big enough, so we use it.
1252   if (!C1->getType()->isIntegerTy(64))
1253     C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1254 
1255   if (!C2->getType()->isIntegerTy(64))
1256     C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1257 
1258   if (C1 == C2) return 0;  // They are equal
1259 
1260   // If the type being indexed over is really just a zero sized type, there is
1261   // no pointer difference being made here.
1262   if (isMaybeZeroSizedType(ElTy))
1263     return -2; // dunno.
1264 
1265   // If they are really different, now that they are the same type, then we
1266   // found a difference!
1267   if (cast<ConstantInt>(C1)->getSExtValue() <
1268       cast<ConstantInt>(C2)->getSExtValue())
1269     return -1;
1270   else
1271     return 1;
1272 }
1273 
1274 /// evaluateFCmpRelation - This function determines if there is anything we can
1275 /// decide about the two constants provided.  This doesn't need to handle simple
1276 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1277 /// If we can determine that the two constants have a particular relation to
1278 /// each other, we should return the corresponding FCmpInst predicate,
1279 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1280 /// ConstantFoldCompareInstruction.
1281 ///
1282 /// To simplify this code we canonicalize the relation so that the first
1283 /// operand is always the most "complex" of the two.  We consider ConstantFP
1284 /// to be the simplest, and ConstantExprs to be the most complex.
evaluateFCmpRelation(Constant * V1,Constant * V2)1285 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1286   assert(V1->getType() == V2->getType() &&
1287          "Cannot compare values of different types!");
1288 
1289   // Handle degenerate case quickly
1290   if (V1 == V2) return FCmpInst::FCMP_OEQ;
1291 
1292   if (!isa<ConstantExpr>(V1)) {
1293     if (!isa<ConstantExpr>(V2)) {
1294       // We distilled thisUse the standard constant folder for a few cases
1295       ConstantInt *R = nullptr;
1296       R = dyn_cast<ConstantInt>(
1297                       ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1298       if (R && !R->isZero())
1299         return FCmpInst::FCMP_OEQ;
1300       R = dyn_cast<ConstantInt>(
1301                       ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1302       if (R && !R->isZero())
1303         return FCmpInst::FCMP_OLT;
1304       R = dyn_cast<ConstantInt>(
1305                       ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1306       if (R && !R->isZero())
1307         return FCmpInst::FCMP_OGT;
1308 
1309       // Nothing more we can do
1310       return FCmpInst::BAD_FCMP_PREDICATE;
1311     }
1312 
1313     // If the first operand is simple and second is ConstantExpr, swap operands.
1314     FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1315     if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1316       return FCmpInst::getSwappedPredicate(SwappedRelation);
1317   } else {
1318     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1319     // constantexpr or a simple constant.
1320     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1321     switch (CE1->getOpcode()) {
1322     case Instruction::FPTrunc:
1323     case Instruction::FPExt:
1324     case Instruction::UIToFP:
1325     case Instruction::SIToFP:
1326       // We might be able to do something with these but we don't right now.
1327       break;
1328     default:
1329       break;
1330     }
1331   }
1332   // There are MANY other foldings that we could perform here.  They will
1333   // probably be added on demand, as they seem needed.
1334   return FCmpInst::BAD_FCMP_PREDICATE;
1335 }
1336 
areGlobalsPotentiallyEqual(const GlobalValue * GV1,const GlobalValue * GV2)1337 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1338                                                       const GlobalValue *GV2) {
1339   // Don't try to decide equality of aliases.
1340   if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1341     if (!GV1->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1342       return ICmpInst::ICMP_NE;
1343   return ICmpInst::BAD_ICMP_PREDICATE;
1344 }
1345 
1346 /// evaluateICmpRelation - This function determines if there is anything we can
1347 /// decide about the two constants provided.  This doesn't need to handle simple
1348 /// things like integer comparisons, but should instead handle ConstantExprs
1349 /// and GlobalValues.  If we can determine that the two constants have a
1350 /// particular relation to each other, we should return the corresponding ICmp
1351 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1352 ///
1353 /// To simplify this code we canonicalize the relation so that the first
1354 /// operand is always the most "complex" of the two.  We consider simple
1355 /// constants (like ConstantInt) to be the simplest, followed by
1356 /// GlobalValues, followed by ConstantExpr's (the most complex).
1357 ///
evaluateICmpRelation(Constant * V1,Constant * V2,bool isSigned)1358 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1359                                                 bool isSigned) {
1360   assert(V1->getType() == V2->getType() &&
1361          "Cannot compare different types of values!");
1362   if (V1 == V2) return ICmpInst::ICMP_EQ;
1363 
1364   if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1365       !isa<BlockAddress>(V1)) {
1366     if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1367         !isa<BlockAddress>(V2)) {
1368       // We distilled this down to a simple case, use the standard constant
1369       // folder.
1370       ConstantInt *R = nullptr;
1371       ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1372       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1373       if (R && !R->isZero())
1374         return pred;
1375       pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1376       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1377       if (R && !R->isZero())
1378         return pred;
1379       pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1380       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1381       if (R && !R->isZero())
1382         return pred;
1383 
1384       // If we couldn't figure it out, bail.
1385       return ICmpInst::BAD_ICMP_PREDICATE;
1386     }
1387 
1388     // If the first operand is simple, swap operands.
1389     ICmpInst::Predicate SwappedRelation =
1390       evaluateICmpRelation(V2, V1, isSigned);
1391     if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1392       return ICmpInst::getSwappedPredicate(SwappedRelation);
1393 
1394   } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1395     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1396       ICmpInst::Predicate SwappedRelation =
1397         evaluateICmpRelation(V2, V1, isSigned);
1398       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1399         return ICmpInst::getSwappedPredicate(SwappedRelation);
1400       return ICmpInst::BAD_ICMP_PREDICATE;
1401     }
1402 
1403     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1404     // constant (which, since the types must match, means that it's a
1405     // ConstantPointerNull).
1406     if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1407       return areGlobalsPotentiallyEqual(GV, GV2);
1408     } else if (isa<BlockAddress>(V2)) {
1409       return ICmpInst::ICMP_NE; // Globals never equal labels.
1410     } else {
1411       assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1412       // GlobalVals can never be null unless they have external weak linkage.
1413       // We don't try to evaluate aliases here.
1414       if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1415         return ICmpInst::ICMP_NE;
1416     }
1417   } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1418     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1419       ICmpInst::Predicate SwappedRelation =
1420         evaluateICmpRelation(V2, V1, isSigned);
1421       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1422         return ICmpInst::getSwappedPredicate(SwappedRelation);
1423       return ICmpInst::BAD_ICMP_PREDICATE;
1424     }
1425 
1426     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1427     // constant (which, since the types must match, means that it is a
1428     // ConstantPointerNull).
1429     if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1430       // Block address in another function can't equal this one, but block
1431       // addresses in the current function might be the same if blocks are
1432       // empty.
1433       if (BA2->getFunction() != BA->getFunction())
1434         return ICmpInst::ICMP_NE;
1435     } else {
1436       // Block addresses aren't null, don't equal the address of globals.
1437       assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1438              "Canonicalization guarantee!");
1439       return ICmpInst::ICMP_NE;
1440     }
1441   } else {
1442     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1443     // constantexpr, a global, block address, or a simple constant.
1444     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1445     Constant *CE1Op0 = CE1->getOperand(0);
1446 
1447     switch (CE1->getOpcode()) {
1448     case Instruction::Trunc:
1449     case Instruction::FPTrunc:
1450     case Instruction::FPExt:
1451     case Instruction::FPToUI:
1452     case Instruction::FPToSI:
1453       break; // We can't evaluate floating point casts or truncations.
1454 
1455     case Instruction::UIToFP:
1456     case Instruction::SIToFP:
1457     case Instruction::BitCast:
1458     case Instruction::ZExt:
1459     case Instruction::SExt:
1460       // If the cast is not actually changing bits, and the second operand is a
1461       // null pointer, do the comparison with the pre-casted value.
1462       if (V2->isNullValue() &&
1463           (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1464         if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1465         if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1466         return evaluateICmpRelation(CE1Op0,
1467                                     Constant::getNullValue(CE1Op0->getType()),
1468                                     isSigned);
1469       }
1470       break;
1471 
1472     case Instruction::GetElementPtr: {
1473       GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1474       // Ok, since this is a getelementptr, we know that the constant has a
1475       // pointer type.  Check the various cases.
1476       if (isa<ConstantPointerNull>(V2)) {
1477         // If we are comparing a GEP to a null pointer, check to see if the base
1478         // of the GEP equals the null pointer.
1479         if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1480           if (GV->hasExternalWeakLinkage())
1481             // Weak linkage GVals could be zero or not. We're comparing that
1482             // to null pointer so its greater-or-equal
1483             return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1484           else
1485             // If its not weak linkage, the GVal must have a non-zero address
1486             // so the result is greater-than
1487             return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1488         } else if (isa<ConstantPointerNull>(CE1Op0)) {
1489           // If we are indexing from a null pointer, check to see if we have any
1490           // non-zero indices.
1491           for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1492             if (!CE1->getOperand(i)->isNullValue())
1493               // Offsetting from null, must not be equal.
1494               return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1495           // Only zero indexes from null, must still be zero.
1496           return ICmpInst::ICMP_EQ;
1497         }
1498         // Otherwise, we can't really say if the first operand is null or not.
1499       } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1500         if (isa<ConstantPointerNull>(CE1Op0)) {
1501           if (GV2->hasExternalWeakLinkage())
1502             // Weak linkage GVals could be zero or not. We're comparing it to
1503             // a null pointer, so its less-or-equal
1504             return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1505           else
1506             // If its not weak linkage, the GVal must have a non-zero address
1507             // so the result is less-than
1508             return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1509         } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1510           if (GV == GV2) {
1511             // If this is a getelementptr of the same global, then it must be
1512             // different.  Because the types must match, the getelementptr could
1513             // only have at most one index, and because we fold getelementptr's
1514             // with a single zero index, it must be nonzero.
1515             assert(CE1->getNumOperands() == 2 &&
1516                    !CE1->getOperand(1)->isNullValue() &&
1517                    "Surprising getelementptr!");
1518             return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1519           } else {
1520             if (CE1GEP->hasAllZeroIndices())
1521               return areGlobalsPotentiallyEqual(GV, GV2);
1522             return ICmpInst::BAD_ICMP_PREDICATE;
1523           }
1524         }
1525       } else {
1526         ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1527         Constant *CE2Op0 = CE2->getOperand(0);
1528 
1529         // There are MANY other foldings that we could perform here.  They will
1530         // probably be added on demand, as they seem needed.
1531         switch (CE2->getOpcode()) {
1532         default: break;
1533         case Instruction::GetElementPtr:
1534           // By far the most common case to handle is when the base pointers are
1535           // obviously to the same global.
1536           if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1537             // Don't know relative ordering, but check for inequality.
1538             if (CE1Op0 != CE2Op0) {
1539               GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1540               if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1541                 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1542                                                   cast<GlobalValue>(CE2Op0));
1543               return ICmpInst::BAD_ICMP_PREDICATE;
1544             }
1545             // Ok, we know that both getelementptr instructions are based on the
1546             // same global.  From this, we can precisely determine the relative
1547             // ordering of the resultant pointers.
1548             unsigned i = 1;
1549 
1550             // The logic below assumes that the result of the comparison
1551             // can be determined by finding the first index that differs.
1552             // This doesn't work if there is over-indexing in any
1553             // subsequent indices, so check for that case first.
1554             if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1555                 !CE2->isGEPWithNoNotionalOverIndexing())
1556                return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1557 
1558             // Compare all of the operands the GEP's have in common.
1559             gep_type_iterator GTI = gep_type_begin(CE1);
1560             for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1561                  ++i, ++GTI)
1562               switch (IdxCompare(CE1->getOperand(i),
1563                                  CE2->getOperand(i), GTI.getIndexedType())) {
1564               case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1565               case 1:  return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1566               case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1567               }
1568 
1569             // Ok, we ran out of things they have in common.  If any leftovers
1570             // are non-zero then we have a difference, otherwise we are equal.
1571             for (; i < CE1->getNumOperands(); ++i)
1572               if (!CE1->getOperand(i)->isNullValue()) {
1573                 if (isa<ConstantInt>(CE1->getOperand(i)))
1574                   return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1575                 else
1576                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1577               }
1578 
1579             for (; i < CE2->getNumOperands(); ++i)
1580               if (!CE2->getOperand(i)->isNullValue()) {
1581                 if (isa<ConstantInt>(CE2->getOperand(i)))
1582                   return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1583                 else
1584                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1585               }
1586             return ICmpInst::ICMP_EQ;
1587           }
1588         }
1589       }
1590     }
1591     default:
1592       break;
1593     }
1594   }
1595 
1596   return ICmpInst::BAD_ICMP_PREDICATE;
1597 }
1598 
ConstantFoldCompareInstruction(unsigned short pred,Constant * C1,Constant * C2)1599 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1600                                                Constant *C1, Constant *C2) {
1601   Type *ResultTy;
1602   if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1603     ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1604                                VT->getNumElements());
1605   else
1606     ResultTy = Type::getInt1Ty(C1->getContext());
1607 
1608   // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1609   if (pred == FCmpInst::FCMP_FALSE)
1610     return Constant::getNullValue(ResultTy);
1611 
1612   if (pred == FCmpInst::FCMP_TRUE)
1613     return Constant::getAllOnesValue(ResultTy);
1614 
1615   // Handle some degenerate cases first
1616   if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1617     // For EQ and NE, we can always pick a value for the undef to make the
1618     // predicate pass or fail, so we can return undef.
1619     // Also, if both operands are undef, we can return undef.
1620     if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) ||
1621         (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
1622       return UndefValue::get(ResultTy);
1623     // Otherwise, pick the same value as the non-undef operand, and fold
1624     // it to true or false.
1625     return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred));
1626   }
1627 
1628   // icmp eq/ne(null,GV) -> false/true
1629   if (C1->isNullValue()) {
1630     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1631       // Don't try to evaluate aliases.  External weak GV can be null.
1632       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1633         if (pred == ICmpInst::ICMP_EQ)
1634           return ConstantInt::getFalse(C1->getContext());
1635         else if (pred == ICmpInst::ICMP_NE)
1636           return ConstantInt::getTrue(C1->getContext());
1637       }
1638   // icmp eq/ne(GV,null) -> false/true
1639   } else if (C2->isNullValue()) {
1640     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1641       // Don't try to evaluate aliases.  External weak GV can be null.
1642       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1643         if (pred == ICmpInst::ICMP_EQ)
1644           return ConstantInt::getFalse(C1->getContext());
1645         else if (pred == ICmpInst::ICMP_NE)
1646           return ConstantInt::getTrue(C1->getContext());
1647       }
1648   }
1649 
1650   // If the comparison is a comparison between two i1's, simplify it.
1651   if (C1->getType()->isIntegerTy(1)) {
1652     switch(pred) {
1653     case ICmpInst::ICMP_EQ:
1654       if (isa<ConstantInt>(C2))
1655         return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1656       return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1657     case ICmpInst::ICMP_NE:
1658       return ConstantExpr::getXor(C1, C2);
1659     default:
1660       break;
1661     }
1662   }
1663 
1664   if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1665     APInt V1 = cast<ConstantInt>(C1)->getValue();
1666     APInt V2 = cast<ConstantInt>(C2)->getValue();
1667     switch (pred) {
1668     default: llvm_unreachable("Invalid ICmp Predicate");
1669     case ICmpInst::ICMP_EQ:  return ConstantInt::get(ResultTy, V1 == V2);
1670     case ICmpInst::ICMP_NE:  return ConstantInt::get(ResultTy, V1 != V2);
1671     case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1672     case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1673     case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1674     case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1675     case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1676     case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1677     case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1678     case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1679     }
1680   } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1681     APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1682     APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1683     APFloat::cmpResult R = C1V.compare(C2V);
1684     switch (pred) {
1685     default: llvm_unreachable("Invalid FCmp Predicate");
1686     case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1687     case FCmpInst::FCMP_TRUE:  return Constant::getAllOnesValue(ResultTy);
1688     case FCmpInst::FCMP_UNO:
1689       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1690     case FCmpInst::FCMP_ORD:
1691       return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1692     case FCmpInst::FCMP_UEQ:
1693       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1694                                         R==APFloat::cmpEqual);
1695     case FCmpInst::FCMP_OEQ:
1696       return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1697     case FCmpInst::FCMP_UNE:
1698       return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1699     case FCmpInst::FCMP_ONE:
1700       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1701                                         R==APFloat::cmpGreaterThan);
1702     case FCmpInst::FCMP_ULT:
1703       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1704                                         R==APFloat::cmpLessThan);
1705     case FCmpInst::FCMP_OLT:
1706       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1707     case FCmpInst::FCMP_UGT:
1708       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1709                                         R==APFloat::cmpGreaterThan);
1710     case FCmpInst::FCMP_OGT:
1711       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1712     case FCmpInst::FCMP_ULE:
1713       return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1714     case FCmpInst::FCMP_OLE:
1715       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1716                                         R==APFloat::cmpEqual);
1717     case FCmpInst::FCMP_UGE:
1718       return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1719     case FCmpInst::FCMP_OGE:
1720       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1721                                         R==APFloat::cmpEqual);
1722     }
1723   } else if (C1->getType()->isVectorTy()) {
1724     // If we can constant fold the comparison of each element, constant fold
1725     // the whole vector comparison.
1726     SmallVector<Constant*, 4> ResElts;
1727     Type *Ty = IntegerType::get(C1->getContext(), 32);
1728     // Compare the elements, producing an i1 result or constant expr.
1729     for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1730       Constant *C1E =
1731         ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1732       Constant *C2E =
1733         ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1734 
1735       ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1736     }
1737 
1738     return ConstantVector::get(ResElts);
1739   }
1740 
1741   if (C1->getType()->isFloatingPointTy()) {
1742     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
1743     switch (evaluateFCmpRelation(C1, C2)) {
1744     default: llvm_unreachable("Unknown relation!");
1745     case FCmpInst::FCMP_UNO:
1746     case FCmpInst::FCMP_ORD:
1747     case FCmpInst::FCMP_UEQ:
1748     case FCmpInst::FCMP_UNE:
1749     case FCmpInst::FCMP_ULT:
1750     case FCmpInst::FCMP_UGT:
1751     case FCmpInst::FCMP_ULE:
1752     case FCmpInst::FCMP_UGE:
1753     case FCmpInst::FCMP_TRUE:
1754     case FCmpInst::FCMP_FALSE:
1755     case FCmpInst::BAD_FCMP_PREDICATE:
1756       break; // Couldn't determine anything about these constants.
1757     case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1758       Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1759                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1760                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1761       break;
1762     case FCmpInst::FCMP_OLT: // We know that C1 < C2
1763       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1764                 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1765                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1766       break;
1767     case FCmpInst::FCMP_OGT: // We know that C1 > C2
1768       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1769                 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1770                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1771       break;
1772     case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1773       // We can only partially decide this relation.
1774       if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1775         Result = 0;
1776       else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1777         Result = 1;
1778       break;
1779     case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1780       // We can only partially decide this relation.
1781       if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1782         Result = 0;
1783       else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1784         Result = 1;
1785       break;
1786     case FCmpInst::FCMP_ONE: // We know that C1 != C2
1787       // We can only partially decide this relation.
1788       if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1789         Result = 0;
1790       else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1791         Result = 1;
1792       break;
1793     }
1794 
1795     // If we evaluated the result, return it now.
1796     if (Result != -1)
1797       return ConstantInt::get(ResultTy, Result);
1798 
1799   } else {
1800     // Evaluate the relation between the two constants, per the predicate.
1801     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
1802     switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1803     default: llvm_unreachable("Unknown relational!");
1804     case ICmpInst::BAD_ICMP_PREDICATE:
1805       break;  // Couldn't determine anything about these constants.
1806     case ICmpInst::ICMP_EQ:   // We know the constants are equal!
1807       // If we know the constants are equal, we can decide the result of this
1808       // computation precisely.
1809       Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1810       break;
1811     case ICmpInst::ICMP_ULT:
1812       switch (pred) {
1813       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1814         Result = 1; break;
1815       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1816         Result = 0; break;
1817       }
1818       break;
1819     case ICmpInst::ICMP_SLT:
1820       switch (pred) {
1821       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1822         Result = 1; break;
1823       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1824         Result = 0; break;
1825       }
1826       break;
1827     case ICmpInst::ICMP_UGT:
1828       switch (pred) {
1829       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1830         Result = 1; break;
1831       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1832         Result = 0; break;
1833       }
1834       break;
1835     case ICmpInst::ICMP_SGT:
1836       switch (pred) {
1837       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1838         Result = 1; break;
1839       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1840         Result = 0; break;
1841       }
1842       break;
1843     case ICmpInst::ICMP_ULE:
1844       if (pred == ICmpInst::ICMP_UGT) Result = 0;
1845       if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1846       break;
1847     case ICmpInst::ICMP_SLE:
1848       if (pred == ICmpInst::ICMP_SGT) Result = 0;
1849       if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1850       break;
1851     case ICmpInst::ICMP_UGE:
1852       if (pred == ICmpInst::ICMP_ULT) Result = 0;
1853       if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
1854       break;
1855     case ICmpInst::ICMP_SGE:
1856       if (pred == ICmpInst::ICMP_SLT) Result = 0;
1857       if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
1858       break;
1859     case ICmpInst::ICMP_NE:
1860       if (pred == ICmpInst::ICMP_EQ) Result = 0;
1861       if (pred == ICmpInst::ICMP_NE) Result = 1;
1862       break;
1863     }
1864 
1865     // If we evaluated the result, return it now.
1866     if (Result != -1)
1867       return ConstantInt::get(ResultTy, Result);
1868 
1869     // If the right hand side is a bitcast, try using its inverse to simplify
1870     // it by moving it to the left hand side.  We can't do this if it would turn
1871     // a vector compare into a scalar compare or visa versa.
1872     if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
1873       Constant *CE2Op0 = CE2->getOperand(0);
1874       if (CE2->getOpcode() == Instruction::BitCast &&
1875           CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
1876         Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
1877         return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
1878       }
1879     }
1880 
1881     // If the left hand side is an extension, try eliminating it.
1882     if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1883       if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) ||
1884           (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){
1885         Constant *CE1Op0 = CE1->getOperand(0);
1886         Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
1887         if (CE1Inverse == CE1Op0) {
1888           // Check whether we can safely truncate the right hand side.
1889           Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
1890           if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
1891                                     C2->getType()) == C2)
1892             return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
1893         }
1894       }
1895     }
1896 
1897     if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
1898         (C1->isNullValue() && !C2->isNullValue())) {
1899       // If C2 is a constant expr and C1 isn't, flip them around and fold the
1900       // other way if possible.
1901       // Also, if C1 is null and C2 isn't, flip them around.
1902       pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1903       return ConstantExpr::getICmp(pred, C2, C1);
1904     }
1905   }
1906   return nullptr;
1907 }
1908 
1909 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
1910 /// is "inbounds".
1911 template<typename IndexTy>
isInBoundsIndices(ArrayRef<IndexTy> Idxs)1912 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
1913   // No indices means nothing that could be out of bounds.
1914   if (Idxs.empty()) return true;
1915 
1916   // If the first index is zero, it's in bounds.
1917   if (cast<Constant>(Idxs[0])->isNullValue()) return true;
1918 
1919   // If the first index is one and all the rest are zero, it's in bounds,
1920   // by the one-past-the-end rule.
1921   if (!cast<ConstantInt>(Idxs[0])->isOne())
1922     return false;
1923   for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
1924     if (!cast<Constant>(Idxs[i])->isNullValue())
1925       return false;
1926   return true;
1927 }
1928 
1929 /// \brief Test whether a given ConstantInt is in-range for a SequentialType.
isIndexInRangeOfSequentialType(const SequentialType * STy,const ConstantInt * CI)1930 static bool isIndexInRangeOfSequentialType(const SequentialType *STy,
1931                                            const ConstantInt *CI) {
1932   if (const PointerType *PTy = dyn_cast<PointerType>(STy))
1933     // Only handle pointers to sized types, not pointers to functions.
1934     return PTy->getElementType()->isSized();
1935 
1936   uint64_t NumElements = 0;
1937   // Determine the number of elements in our sequential type.
1938   if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
1939     NumElements = ATy->getNumElements();
1940   else if (const VectorType *VTy = dyn_cast<VectorType>(STy))
1941     NumElements = VTy->getNumElements();
1942 
1943   assert((isa<ArrayType>(STy) || NumElements > 0) &&
1944          "didn't expect non-array type to have zero elements!");
1945 
1946   // We cannot bounds check the index if it doesn't fit in an int64_t.
1947   if (CI->getValue().getActiveBits() > 64)
1948     return false;
1949 
1950   // A negative index or an index past the end of our sequential type is
1951   // considered out-of-range.
1952   int64_t IndexVal = CI->getSExtValue();
1953   if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
1954     return false;
1955 
1956   // Otherwise, it is in-range.
1957   return true;
1958 }
1959 
1960 template<typename IndexTy>
ConstantFoldGetElementPtrImpl(Constant * C,bool inBounds,ArrayRef<IndexTy> Idxs)1961 static Constant *ConstantFoldGetElementPtrImpl(Constant *C,
1962                                                bool inBounds,
1963                                                ArrayRef<IndexTy> Idxs) {
1964   if (Idxs.empty()) return C;
1965   Constant *Idx0 = cast<Constant>(Idxs[0]);
1966   if ((Idxs.size() == 1 && Idx0->isNullValue()))
1967     return C;
1968 
1969   if (isa<UndefValue>(C)) {
1970     PointerType *Ptr = cast<PointerType>(C->getType());
1971     Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
1972     assert(Ty && "Invalid indices for GEP!");
1973     return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
1974   }
1975 
1976   if (C->isNullValue()) {
1977     bool isNull = true;
1978     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
1979       if (!cast<Constant>(Idxs[i])->isNullValue()) {
1980         isNull = false;
1981         break;
1982       }
1983     if (isNull) {
1984       PointerType *Ptr = cast<PointerType>(C->getType());
1985       Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
1986       assert(Ty && "Invalid indices for GEP!");
1987       return ConstantPointerNull::get(PointerType::get(Ty,
1988                                                        Ptr->getAddressSpace()));
1989     }
1990   }
1991 
1992   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1993     // Combine Indices - If the source pointer to this getelementptr instruction
1994     // is a getelementptr instruction, combine the indices of the two
1995     // getelementptr instructions into a single instruction.
1996     //
1997     if (CE->getOpcode() == Instruction::GetElementPtr) {
1998       Type *LastTy = nullptr;
1999       for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2000            I != E; ++I)
2001         LastTy = *I;
2002 
2003       // We cannot combine indices if doing so would take us outside of an
2004       // array or vector.  Doing otherwise could trick us if we evaluated such a
2005       // GEP as part of a load.
2006       //
2007       // e.g. Consider if the original GEP was:
2008       // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2009       //                    i32 0, i32 0, i64 0)
2010       //
2011       // If we then tried to offset it by '8' to get to the third element,
2012       // an i8, we should *not* get:
2013       // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2014       //                    i32 0, i32 0, i64 8)
2015       //
2016       // This GEP tries to index array element '8  which runs out-of-bounds.
2017       // Subsequent evaluation would get confused and produce erroneous results.
2018       //
2019       // The following prohibits such a GEP from being formed by checking to see
2020       // if the index is in-range with respect to an array or vector.
2021       bool PerformFold = false;
2022       if (Idx0->isNullValue())
2023         PerformFold = true;
2024       else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy))
2025         if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0))
2026           PerformFold = isIndexInRangeOfSequentialType(STy, CI);
2027 
2028       if (PerformFold) {
2029         SmallVector<Value*, 16> NewIndices;
2030         NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2031         for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
2032           NewIndices.push_back(CE->getOperand(i));
2033 
2034         // Add the last index of the source with the first index of the new GEP.
2035         // Make sure to handle the case when they are actually different types.
2036         Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2037         // Otherwise it must be an array.
2038         if (!Idx0->isNullValue()) {
2039           Type *IdxTy = Combined->getType();
2040           if (IdxTy != Idx0->getType()) {
2041             Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
2042             Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
2043             Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
2044             Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2045           } else {
2046             Combined =
2047               ConstantExpr::get(Instruction::Add, Idx0, Combined);
2048           }
2049         }
2050 
2051         NewIndices.push_back(Combined);
2052         NewIndices.append(Idxs.begin() + 1, Idxs.end());
2053         return
2054           ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices,
2055                                          inBounds &&
2056                                            cast<GEPOperator>(CE)->isInBounds());
2057       }
2058     }
2059 
2060     // Attempt to fold casts to the same type away.  For example, folding:
2061     //
2062     //   i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2063     //                       i64 0, i64 0)
2064     // into:
2065     //
2066     //   i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2067     //
2068     // Don't fold if the cast is changing address spaces.
2069     if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2070       PointerType *SrcPtrTy =
2071         dyn_cast<PointerType>(CE->getOperand(0)->getType());
2072       PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2073       if (SrcPtrTy && DstPtrTy) {
2074         ArrayType *SrcArrayTy =
2075           dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2076         ArrayType *DstArrayTy =
2077           dyn_cast<ArrayType>(DstPtrTy->getElementType());
2078         if (SrcArrayTy && DstArrayTy
2079             && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2080             && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2081           return ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0),
2082                                                 Idxs, inBounds);
2083       }
2084     }
2085   }
2086 
2087   // Check to see if any array indices are not within the corresponding
2088   // notional array or vector bounds. If so, try to determine if they can be
2089   // factored out into preceding dimensions.
2090   bool Unknown = false;
2091   SmallVector<Constant *, 8> NewIdxs;
2092   Type *Ty = C->getType();
2093   Type *Prev = nullptr;
2094   for (unsigned i = 0, e = Idxs.size(); i != e;
2095        Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2096     if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2097       if (isa<ArrayType>(Ty) || isa<VectorType>(Ty))
2098         if (CI->getSExtValue() > 0 &&
2099             !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) {
2100           if (isa<SequentialType>(Prev)) {
2101             // It's out of range, but we can factor it into the prior
2102             // dimension.
2103             NewIdxs.resize(Idxs.size());
2104             uint64_t NumElements = 0;
2105             if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2106               NumElements = ATy->getNumElements();
2107             else
2108               NumElements = cast<VectorType>(Ty)->getNumElements();
2109 
2110             ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements);
2111             NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2112 
2113             Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2114             Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2115 
2116             // Before adding, extend both operands to i64 to avoid
2117             // overflow trouble.
2118             if (!PrevIdx->getType()->isIntegerTy(64))
2119               PrevIdx = ConstantExpr::getSExt(PrevIdx,
2120                                            Type::getInt64Ty(Div->getContext()));
2121             if (!Div->getType()->isIntegerTy(64))
2122               Div = ConstantExpr::getSExt(Div,
2123                                           Type::getInt64Ty(Div->getContext()));
2124 
2125             NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2126           } else {
2127             // It's out of range, but the prior dimension is a struct
2128             // so we can't do anything about it.
2129             Unknown = true;
2130           }
2131         }
2132     } else {
2133       // We don't know if it's in range or not.
2134       Unknown = true;
2135     }
2136   }
2137 
2138   // If we did any factoring, start over with the adjusted indices.
2139   if (!NewIdxs.empty()) {
2140     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2141       if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2142     return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds);
2143   }
2144 
2145   // If all indices are known integers and normalized, we can do a simple
2146   // check for the "inbounds" property.
2147   if (!Unknown && !inBounds &&
2148       isa<GlobalVariable>(C) && isInBoundsIndices(Idxs))
2149     return ConstantExpr::getInBoundsGetElementPtr(C, Idxs);
2150 
2151   return nullptr;
2152 }
2153 
ConstantFoldGetElementPtr(Constant * C,bool inBounds,ArrayRef<Constant * > Idxs)2154 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2155                                           bool inBounds,
2156                                           ArrayRef<Constant *> Idxs) {
2157   return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);
2158 }
2159 
ConstantFoldGetElementPtr(Constant * C,bool inBounds,ArrayRef<Value * > Idxs)2160 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2161                                           bool inBounds,
2162                                           ArrayRef<Value *> Idxs) {
2163   return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);
2164 }
2165