LLVM API Documentation
00001 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This file implements folding of constants for LLVM. This implements the 00011 // (internal) ConstantFold.h interface, which is used by the 00012 // ConstantExpr::get* methods to automatically fold constants when possible. 00013 // 00014 // The current constant folding implementation is implemented in two pieces: the 00015 // pieces that don't need DataLayout, and the pieces that do. This is to avoid 00016 // a dependence in IR on Target. 00017 // 00018 //===----------------------------------------------------------------------===// 00019 00020 #include "ConstantFold.h" 00021 #include "llvm/ADT/SmallVector.h" 00022 #include "llvm/IR/Constants.h" 00023 #include "llvm/IR/DerivedTypes.h" 00024 #include "llvm/IR/Function.h" 00025 #include "llvm/IR/GetElementPtrTypeIterator.h" 00026 #include "llvm/IR/GlobalAlias.h" 00027 #include "llvm/IR/GlobalVariable.h" 00028 #include "llvm/IR/Instructions.h" 00029 #include "llvm/IR/Operator.h" 00030 #include "llvm/Support/Compiler.h" 00031 #include "llvm/Support/ErrorHandling.h" 00032 #include "llvm/Support/ManagedStatic.h" 00033 #include "llvm/Support/MathExtras.h" 00034 #include <limits> 00035 using namespace llvm; 00036 00037 //===----------------------------------------------------------------------===// 00038 // ConstantFold*Instruction Implementations 00039 //===----------------------------------------------------------------------===// 00040 00041 /// BitCastConstantVector - Convert the specified vector Constant node to the 00042 /// specified vector type. At this point, we know that the elements of the 00043 /// input vector constant are all simple integer or FP values. 00044 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 00045 00046 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 00047 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 00048 00049 // If this cast changes element count then we can't handle it here: 00050 // doing so requires endianness information. This should be handled by 00051 // Analysis/ConstantFolding.cpp 00052 unsigned NumElts = DstTy->getNumElements(); 00053 if (NumElts != CV->getType()->getVectorNumElements()) 00054 return nullptr; 00055 00056 Type *DstEltTy = DstTy->getElementType(); 00057 00058 SmallVector<Constant*, 16> Result; 00059 Type *Ty = IntegerType::get(CV->getContext(), 32); 00060 for (unsigned i = 0; i != NumElts; ++i) { 00061 Constant *C = 00062 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 00063 C = ConstantExpr::getBitCast(C, DstEltTy); 00064 Result.push_back(C); 00065 } 00066 00067 return ConstantVector::get(Result); 00068 } 00069 00070 /// This function determines which opcode to use to fold two constant cast 00071 /// expressions together. It uses CastInst::isEliminableCastPair to determine 00072 /// the opcode. Consequently its just a wrapper around that function. 00073 /// @brief Determine if it is valid to fold a cast of a cast 00074 static unsigned 00075 foldConstantCastPair( 00076 unsigned opc, ///< opcode of the second cast constant expression 00077 ConstantExpr *Op, ///< the first cast constant expression 00078 Type *DstTy ///< destination type of the first cast 00079 ) { 00080 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 00081 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 00082 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 00083 00084 // The the types and opcodes for the two Cast constant expressions 00085 Type *SrcTy = Op->getOperand(0)->getType(); 00086 Type *MidTy = Op->getType(); 00087 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 00088 Instruction::CastOps secondOp = Instruction::CastOps(opc); 00089 00090 // Assume that pointers are never more than 64 bits wide, and only use this 00091 // for the middle type. Otherwise we could end up folding away illegal 00092 // bitcasts between address spaces with different sizes. 00093 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 00094 00095 // Let CastInst::isEliminableCastPair do the heavy lifting. 00096 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 00097 nullptr, FakeIntPtrTy, nullptr); 00098 } 00099 00100 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 00101 Type *SrcTy = V->getType(); 00102 if (SrcTy == DestTy) 00103 return V; // no-op cast 00104 00105 // Check to see if we are casting a pointer to an aggregate to a pointer to 00106 // the first element. If so, return the appropriate GEP instruction. 00107 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 00108 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 00109 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 00110 && DPTy->getElementType()->isSized()) { 00111 SmallVector<Value*, 8> IdxList; 00112 Value *Zero = 00113 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 00114 IdxList.push_back(Zero); 00115 Type *ElTy = PTy->getElementType(); 00116 while (ElTy != DPTy->getElementType()) { 00117 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 00118 if (STy->getNumElements() == 0) break; 00119 ElTy = STy->getElementType(0); 00120 IdxList.push_back(Zero); 00121 } else if (SequentialType *STy = 00122 dyn_cast<SequentialType>(ElTy)) { 00123 if (ElTy->isPointerTy()) break; // Can't index into pointers! 00124 ElTy = STy->getElementType(); 00125 IdxList.push_back(Zero); 00126 } else { 00127 break; 00128 } 00129 } 00130 00131 if (ElTy == DPTy->getElementType()) 00132 // This GEP is inbounds because all indices are zero. 00133 return ConstantExpr::getInBoundsGetElementPtr(V, IdxList); 00134 } 00135 00136 // Handle casts from one vector constant to another. We know that the src 00137 // and dest type have the same size (otherwise its an illegal cast). 00138 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 00139 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 00140 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 00141 "Not cast between same sized vectors!"); 00142 SrcTy = nullptr; 00143 // First, check for null. Undef is already handled. 00144 if (isa<ConstantAggregateZero>(V)) 00145 return Constant::getNullValue(DestTy); 00146 00147 // Handle ConstantVector and ConstantAggregateVector. 00148 return BitCastConstantVector(V, DestPTy); 00149 } 00150 00151 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 00152 // This allows for other simplifications (although some of them 00153 // can only be handled by Analysis/ConstantFolding.cpp). 00154 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 00155 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 00156 } 00157 00158 // Finally, implement bitcast folding now. The code below doesn't handle 00159 // bitcast right. 00160 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 00161 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 00162 00163 // Handle integral constant input. 00164 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 00165 if (DestTy->isIntegerTy()) 00166 // Integral -> Integral. This is a no-op because the bit widths must 00167 // be the same. Consequently, we just fold to V. 00168 return V; 00169 00170 if (DestTy->isFloatingPointTy()) 00171 return ConstantFP::get(DestTy->getContext(), 00172 APFloat(DestTy->getFltSemantics(), 00173 CI->getValue())); 00174 00175 // Otherwise, can't fold this (vector?) 00176 return nullptr; 00177 } 00178 00179 // Handle ConstantFP input: FP -> Integral. 00180 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) 00181 return ConstantInt::get(FP->getContext(), 00182 FP->getValueAPF().bitcastToAPInt()); 00183 00184 return nullptr; 00185 } 00186 00187 00188 /// ExtractConstantBytes - V is an integer constant which only has a subset of 00189 /// its bytes used. The bytes used are indicated by ByteStart (which is the 00190 /// first byte used, counting from the least significant byte) and ByteSize, 00191 /// which is the number of bytes used. 00192 /// 00193 /// This function analyzes the specified constant to see if the specified byte 00194 /// range can be returned as a simplified constant. If so, the constant is 00195 /// returned, otherwise null is returned. 00196 /// 00197 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 00198 unsigned ByteSize) { 00199 assert(C->getType()->isIntegerTy() && 00200 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 00201 "Non-byte sized integer input"); 00202 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 00203 assert(ByteSize && "Must be accessing some piece"); 00204 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 00205 assert(ByteSize != CSize && "Should not extract everything"); 00206 00207 // Constant Integers are simple. 00208 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 00209 APInt V = CI->getValue(); 00210 if (ByteStart) 00211 V = V.lshr(ByteStart*8); 00212 V = V.trunc(ByteSize*8); 00213 return ConstantInt::get(CI->getContext(), V); 00214 } 00215 00216 // In the input is a constant expr, we might be able to recursively simplify. 00217 // If not, we definitely can't do anything. 00218 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 00219 if (!CE) return nullptr; 00220 00221 switch (CE->getOpcode()) { 00222 default: return nullptr; 00223 case Instruction::Or: { 00224 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 00225 if (!RHS) 00226 return nullptr; 00227 00228 // X | -1 -> -1. 00229 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 00230 if (RHSC->isAllOnesValue()) 00231 return RHSC; 00232 00233 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 00234 if (!LHS) 00235 return nullptr; 00236 return ConstantExpr::getOr(LHS, RHS); 00237 } 00238 case Instruction::And: { 00239 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 00240 if (!RHS) 00241 return nullptr; 00242 00243 // X & 0 -> 0. 00244 if (RHS->isNullValue()) 00245 return RHS; 00246 00247 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 00248 if (!LHS) 00249 return nullptr; 00250 return ConstantExpr::getAnd(LHS, RHS); 00251 } 00252 case Instruction::LShr: { 00253 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 00254 if (!Amt) 00255 return nullptr; 00256 unsigned ShAmt = Amt->getZExtValue(); 00257 // Cannot analyze non-byte shifts. 00258 if ((ShAmt & 7) != 0) 00259 return nullptr; 00260 ShAmt >>= 3; 00261 00262 // If the extract is known to be all zeros, return zero. 00263 if (ByteStart >= CSize-ShAmt) 00264 return Constant::getNullValue(IntegerType::get(CE->getContext(), 00265 ByteSize*8)); 00266 // If the extract is known to be fully in the input, extract it. 00267 if (ByteStart+ByteSize+ShAmt <= CSize) 00268 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 00269 00270 // TODO: Handle the 'partially zero' case. 00271 return nullptr; 00272 } 00273 00274 case Instruction::Shl: { 00275 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 00276 if (!Amt) 00277 return nullptr; 00278 unsigned ShAmt = Amt->getZExtValue(); 00279 // Cannot analyze non-byte shifts. 00280 if ((ShAmt & 7) != 0) 00281 return nullptr; 00282 ShAmt >>= 3; 00283 00284 // If the extract is known to be all zeros, return zero. 00285 if (ByteStart+ByteSize <= ShAmt) 00286 return Constant::getNullValue(IntegerType::get(CE->getContext(), 00287 ByteSize*8)); 00288 // If the extract is known to be fully in the input, extract it. 00289 if (ByteStart >= ShAmt) 00290 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 00291 00292 // TODO: Handle the 'partially zero' case. 00293 return nullptr; 00294 } 00295 00296 case Instruction::ZExt: { 00297 unsigned SrcBitSize = 00298 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 00299 00300 // If extracting something that is completely zero, return 0. 00301 if (ByteStart*8 >= SrcBitSize) 00302 return Constant::getNullValue(IntegerType::get(CE->getContext(), 00303 ByteSize*8)); 00304 00305 // If exactly extracting the input, return it. 00306 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 00307 return CE->getOperand(0); 00308 00309 // If extracting something completely in the input, if if the input is a 00310 // multiple of 8 bits, recurse. 00311 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 00312 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 00313 00314 // Otherwise, if extracting a subset of the input, which is not multiple of 00315 // 8 bits, do a shift and trunc to get the bits. 00316 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 00317 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 00318 Constant *Res = CE->getOperand(0); 00319 if (ByteStart) 00320 Res = ConstantExpr::getLShr(Res, 00321 ConstantInt::get(Res->getType(), ByteStart*8)); 00322 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 00323 ByteSize*8)); 00324 } 00325 00326 // TODO: Handle the 'partially zero' case. 00327 return nullptr; 00328 } 00329 } 00330 } 00331 00332 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 00333 /// on Ty, with any known factors factored out. If Folded is false, 00334 /// return null if no factoring was possible, to avoid endlessly 00335 /// bouncing an unfoldable expression back into the top-level folder. 00336 /// 00337 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, 00338 bool Folded) { 00339 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 00340 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 00341 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 00342 return ConstantExpr::getNUWMul(E, N); 00343 } 00344 00345 if (StructType *STy = dyn_cast<StructType>(Ty)) 00346 if (!STy->isPacked()) { 00347 unsigned NumElems = STy->getNumElements(); 00348 // An empty struct has size zero. 00349 if (NumElems == 0) 00350 return ConstantExpr::getNullValue(DestTy); 00351 // Check for a struct with all members having the same size. 00352 Constant *MemberSize = 00353 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 00354 bool AllSame = true; 00355 for (unsigned i = 1; i != NumElems; ++i) 00356 if (MemberSize != 00357 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 00358 AllSame = false; 00359 break; 00360 } 00361 if (AllSame) { 00362 Constant *N = ConstantInt::get(DestTy, NumElems); 00363 return ConstantExpr::getNUWMul(MemberSize, N); 00364 } 00365 } 00366 00367 // Pointer size doesn't depend on the pointee type, so canonicalize them 00368 // to an arbitrary pointee. 00369 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 00370 if (!PTy->getElementType()->isIntegerTy(1)) 00371 return 00372 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 00373 PTy->getAddressSpace()), 00374 DestTy, true); 00375 00376 // If there's no interesting folding happening, bail so that we don't create 00377 // a constant that looks like it needs folding but really doesn't. 00378 if (!Folded) 00379 return nullptr; 00380 00381 // Base case: Get a regular sizeof expression. 00382 Constant *C = ConstantExpr::getSizeOf(Ty); 00383 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 00384 DestTy, false), 00385 C, DestTy); 00386 return C; 00387 } 00388 00389 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 00390 /// on Ty, with any known factors factored out. If Folded is false, 00391 /// return null if no factoring was possible, to avoid endlessly 00392 /// bouncing an unfoldable expression back into the top-level folder. 00393 /// 00394 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, 00395 bool Folded) { 00396 // The alignment of an array is equal to the alignment of the 00397 // array element. Note that this is not always true for vectors. 00398 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 00399 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 00400 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 00401 DestTy, 00402 false), 00403 C, DestTy); 00404 return C; 00405 } 00406 00407 if (StructType *STy = dyn_cast<StructType>(Ty)) { 00408 // Packed structs always have an alignment of 1. 00409 if (STy->isPacked()) 00410 return ConstantInt::get(DestTy, 1); 00411 00412 // Otherwise, struct alignment is the maximum alignment of any member. 00413 // Without target data, we can't compare much, but we can check to see 00414 // if all the members have the same alignment. 00415 unsigned NumElems = STy->getNumElements(); 00416 // An empty struct has minimal alignment. 00417 if (NumElems == 0) 00418 return ConstantInt::get(DestTy, 1); 00419 // Check for a struct with all members having the same alignment. 00420 Constant *MemberAlign = 00421 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 00422 bool AllSame = true; 00423 for (unsigned i = 1; i != NumElems; ++i) 00424 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 00425 AllSame = false; 00426 break; 00427 } 00428 if (AllSame) 00429 return MemberAlign; 00430 } 00431 00432 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 00433 // to an arbitrary pointee. 00434 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 00435 if (!PTy->getElementType()->isIntegerTy(1)) 00436 return 00437 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 00438 1), 00439 PTy->getAddressSpace()), 00440 DestTy, true); 00441 00442 // If there's no interesting folding happening, bail so that we don't create 00443 // a constant that looks like it needs folding but really doesn't. 00444 if (!Folded) 00445 return nullptr; 00446 00447 // Base case: Get a regular alignof expression. 00448 Constant *C = ConstantExpr::getAlignOf(Ty); 00449 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 00450 DestTy, false), 00451 C, DestTy); 00452 return C; 00453 } 00454 00455 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 00456 /// on Ty and FieldNo, with any known factors factored out. If Folded is false, 00457 /// return null if no factoring was possible, to avoid endlessly 00458 /// bouncing an unfoldable expression back into the top-level folder. 00459 /// 00460 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, 00461 Type *DestTy, 00462 bool Folded) { 00463 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 00464 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 00465 DestTy, false), 00466 FieldNo, DestTy); 00467 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 00468 return ConstantExpr::getNUWMul(E, N); 00469 } 00470 00471 if (StructType *STy = dyn_cast<StructType>(Ty)) 00472 if (!STy->isPacked()) { 00473 unsigned NumElems = STy->getNumElements(); 00474 // An empty struct has no members. 00475 if (NumElems == 0) 00476 return nullptr; 00477 // Check for a struct with all members having the same size. 00478 Constant *MemberSize = 00479 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 00480 bool AllSame = true; 00481 for (unsigned i = 1; i != NumElems; ++i) 00482 if (MemberSize != 00483 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 00484 AllSame = false; 00485 break; 00486 } 00487 if (AllSame) { 00488 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 00489 false, 00490 DestTy, 00491 false), 00492 FieldNo, DestTy); 00493 return ConstantExpr::getNUWMul(MemberSize, N); 00494 } 00495 } 00496 00497 // If there's no interesting folding happening, bail so that we don't create 00498 // a constant that looks like it needs folding but really doesn't. 00499 if (!Folded) 00500 return nullptr; 00501 00502 // Base case: Get a regular offsetof expression. 00503 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 00504 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 00505 DestTy, false), 00506 C, DestTy); 00507 return C; 00508 } 00509 00510 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 00511 Type *DestTy) { 00512 if (isa<UndefValue>(V)) { 00513 // zext(undef) = 0, because the top bits will be zero. 00514 // sext(undef) = 0, because the top bits will all be the same. 00515 // [us]itofp(undef) = 0, because the result value is bounded. 00516 if (opc == Instruction::ZExt || opc == Instruction::SExt || 00517 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 00518 return Constant::getNullValue(DestTy); 00519 return UndefValue::get(DestTy); 00520 } 00521 00522 if (V->isNullValue() && !DestTy->isX86_MMXTy()) 00523 return Constant::getNullValue(DestTy); 00524 00525 // If the cast operand is a constant expression, there's a few things we can 00526 // do to try to simplify it. 00527 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 00528 if (CE->isCast()) { 00529 // Try hard to fold cast of cast because they are often eliminable. 00530 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 00531 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 00532 } else if (CE->getOpcode() == Instruction::GetElementPtr && 00533 // Do not fold addrspacecast (gep 0, .., 0). It might make the 00534 // addrspacecast uncanonicalized. 00535 opc != Instruction::AddrSpaceCast) { 00536 // If all of the indexes in the GEP are null values, there is no pointer 00537 // adjustment going on. We might as well cast the source pointer. 00538 bool isAllNull = true; 00539 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 00540 if (!CE->getOperand(i)->isNullValue()) { 00541 isAllNull = false; 00542 break; 00543 } 00544 if (isAllNull) 00545 // This is casting one pointer type to another, always BitCast 00546 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 00547 } 00548 } 00549 00550 // If the cast operand is a constant vector, perform the cast by 00551 // operating on each element. In the cast of bitcasts, the element 00552 // count may be mismatched; don't attempt to handle that here. 00553 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 00554 DestTy->isVectorTy() && 00555 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 00556 SmallVector<Constant*, 16> res; 00557 VectorType *DestVecTy = cast<VectorType>(DestTy); 00558 Type *DstEltTy = DestVecTy->getElementType(); 00559 Type *Ty = IntegerType::get(V->getContext(), 32); 00560 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 00561 Constant *C = 00562 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 00563 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 00564 } 00565 return ConstantVector::get(res); 00566 } 00567 00568 // We actually have to do a cast now. Perform the cast according to the 00569 // opcode specified. 00570 switch (opc) { 00571 default: 00572 llvm_unreachable("Failed to cast constant expression"); 00573 case Instruction::FPTrunc: 00574 case Instruction::FPExt: 00575 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 00576 bool ignored; 00577 APFloat Val = FPC->getValueAPF(); 00578 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf : 00579 DestTy->isFloatTy() ? APFloat::IEEEsingle : 00580 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 00581 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 00582 DestTy->isFP128Ty() ? APFloat::IEEEquad : 00583 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble : 00584 APFloat::Bogus, 00585 APFloat::rmNearestTiesToEven, &ignored); 00586 return ConstantFP::get(V->getContext(), Val); 00587 } 00588 return nullptr; // Can't fold. 00589 case Instruction::FPToUI: 00590 case Instruction::FPToSI: 00591 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 00592 const APFloat &V = FPC->getValueAPF(); 00593 bool ignored; 00594 uint64_t x[2]; 00595 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 00596 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 00597 APFloat::rmTowardZero, &ignored); 00598 APInt Val(DestBitWidth, x); 00599 return ConstantInt::get(FPC->getContext(), Val); 00600 } 00601 return nullptr; // Can't fold. 00602 case Instruction::IntToPtr: //always treated as unsigned 00603 if (V->isNullValue()) // Is it an integral null value? 00604 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 00605 return nullptr; // Other pointer types cannot be casted 00606 case Instruction::PtrToInt: // always treated as unsigned 00607 // Is it a null pointer value? 00608 if (V->isNullValue()) 00609 return ConstantInt::get(DestTy, 0); 00610 // If this is a sizeof-like expression, pull out multiplications by 00611 // known factors to expose them to subsequent folding. If it's an 00612 // alignof-like expression, factor out known factors. 00613 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 00614 if (CE->getOpcode() == Instruction::GetElementPtr && 00615 CE->getOperand(0)->isNullValue()) { 00616 Type *Ty = 00617 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 00618 if (CE->getNumOperands() == 2) { 00619 // Handle a sizeof-like expression. 00620 Constant *Idx = CE->getOperand(1); 00621 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 00622 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 00623 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 00624 DestTy, false), 00625 Idx, DestTy); 00626 return ConstantExpr::getMul(C, Idx); 00627 } 00628 } else if (CE->getNumOperands() == 3 && 00629 CE->getOperand(1)->isNullValue()) { 00630 // Handle an alignof-like expression. 00631 if (StructType *STy = dyn_cast<StructType>(Ty)) 00632 if (!STy->isPacked()) { 00633 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 00634 if (CI->isOne() && 00635 STy->getNumElements() == 2 && 00636 STy->getElementType(0)->isIntegerTy(1)) { 00637 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 00638 } 00639 } 00640 // Handle an offsetof-like expression. 00641 if (Ty->isStructTy() || Ty->isArrayTy()) { 00642 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 00643 DestTy, false)) 00644 return C; 00645 } 00646 } 00647 } 00648 // Other pointer types cannot be casted 00649 return nullptr; 00650 case Instruction::UIToFP: 00651 case Instruction::SIToFP: 00652 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 00653 APInt api = CI->getValue(); 00654 APFloat apf(DestTy->getFltSemantics(), 00655 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 00656 (void)apf.convertFromAPInt(api, 00657 opc==Instruction::SIToFP, 00658 APFloat::rmNearestTiesToEven); 00659 return ConstantFP::get(V->getContext(), apf); 00660 } 00661 return nullptr; 00662 case Instruction::ZExt: 00663 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 00664 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 00665 return ConstantInt::get(V->getContext(), 00666 CI->getValue().zext(BitWidth)); 00667 } 00668 return nullptr; 00669 case Instruction::SExt: 00670 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 00671 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 00672 return ConstantInt::get(V->getContext(), 00673 CI->getValue().sext(BitWidth)); 00674 } 00675 return nullptr; 00676 case Instruction::Trunc: { 00677 if (V->getType()->isVectorTy()) 00678 return nullptr; 00679 00680 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 00681 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 00682 return ConstantInt::get(V->getContext(), 00683 CI->getValue().trunc(DestBitWidth)); 00684 } 00685 00686 // The input must be a constantexpr. See if we can simplify this based on 00687 // the bytes we are demanding. Only do this if the source and dest are an 00688 // even multiple of a byte. 00689 if ((DestBitWidth & 7) == 0 && 00690 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 00691 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 00692 return Res; 00693 00694 return nullptr; 00695 } 00696 case Instruction::BitCast: 00697 return FoldBitCast(V, DestTy); 00698 case Instruction::AddrSpaceCast: 00699 return nullptr; 00700 } 00701 } 00702 00703 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 00704 Constant *V1, Constant *V2) { 00705 // Check for i1 and vector true/false conditions. 00706 if (Cond->isNullValue()) return V2; 00707 if (Cond->isAllOnesValue()) return V1; 00708 00709 // If the condition is a vector constant, fold the result elementwise. 00710 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 00711 SmallVector<Constant*, 16> Result; 00712 Type *Ty = IntegerType::get(CondV->getContext(), 32); 00713 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 00714 Constant *V; 00715 Constant *V1Element = ConstantExpr::getExtractElement(V1, 00716 ConstantInt::get(Ty, i)); 00717 Constant *V2Element = ConstantExpr::getExtractElement(V2, 00718 ConstantInt::get(Ty, i)); 00719 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i)); 00720 if (V1Element == V2Element) { 00721 V = V1Element; 00722 } else if (isa<UndefValue>(Cond)) { 00723 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 00724 } else { 00725 if (!isa<ConstantInt>(Cond)) break; 00726 V = Cond->isNullValue() ? V2Element : V1Element; 00727 } 00728 Result.push_back(V); 00729 } 00730 00731 // If we were able to build the vector, return it. 00732 if (Result.size() == V1->getType()->getVectorNumElements()) 00733 return ConstantVector::get(Result); 00734 } 00735 00736 if (isa<UndefValue>(Cond)) { 00737 if (isa<UndefValue>(V1)) return V1; 00738 return V2; 00739 } 00740 if (isa<UndefValue>(V1)) return V2; 00741 if (isa<UndefValue>(V2)) return V1; 00742 if (V1 == V2) return V1; 00743 00744 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 00745 if (TrueVal->getOpcode() == Instruction::Select) 00746 if (TrueVal->getOperand(0) == Cond) 00747 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 00748 } 00749 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 00750 if (FalseVal->getOpcode() == Instruction::Select) 00751 if (FalseVal->getOperand(0) == Cond) 00752 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 00753 } 00754 00755 return nullptr; 00756 } 00757 00758 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 00759 Constant *Idx) { 00760 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 00761 return UndefValue::get(Val->getType()->getVectorElementType()); 00762 if (Val->isNullValue()) // ee(zero, x) -> zero 00763 return Constant::getNullValue(Val->getType()->getVectorElementType()); 00764 // ee({w,x,y,z}, undef) -> undef 00765 if (isa<UndefValue>(Idx)) 00766 return UndefValue::get(Val->getType()->getVectorElementType()); 00767 00768 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 00769 uint64_t Index = CIdx->getZExtValue(); 00770 // ee({w,x,y,z}, wrong_value) -> undef 00771 if (Index >= Val->getType()->getVectorNumElements()) 00772 return UndefValue::get(Val->getType()->getVectorElementType()); 00773 return Val->getAggregateElement(Index); 00774 } 00775 return nullptr; 00776 } 00777 00778 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 00779 Constant *Elt, 00780 Constant *Idx) { 00781 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 00782 if (!CIdx) return nullptr; 00783 const APInt &IdxVal = CIdx->getValue(); 00784 00785 SmallVector<Constant*, 16> Result; 00786 Type *Ty = IntegerType::get(Val->getContext(), 32); 00787 for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){ 00788 if (i == IdxVal) { 00789 Result.push_back(Elt); 00790 continue; 00791 } 00792 00793 Constant *C = 00794 ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 00795 Result.push_back(C); 00796 } 00797 00798 return ConstantVector::get(Result); 00799 } 00800 00801 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 00802 Constant *V2, 00803 Constant *Mask) { 00804 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 00805 Type *EltTy = V1->getType()->getVectorElementType(); 00806 00807 // Undefined shuffle mask -> undefined value. 00808 if (isa<UndefValue>(Mask)) 00809 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 00810 00811 // Don't break the bitcode reader hack. 00812 if (isa<ConstantExpr>(Mask)) return nullptr; 00813 00814 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 00815 00816 // Loop over the shuffle mask, evaluating each element. 00817 SmallVector<Constant*, 32> Result; 00818 for (unsigned i = 0; i != MaskNumElts; ++i) { 00819 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 00820 if (Elt == -1) { 00821 Result.push_back(UndefValue::get(EltTy)); 00822 continue; 00823 } 00824 Constant *InElt; 00825 if (unsigned(Elt) >= SrcNumElts*2) 00826 InElt = UndefValue::get(EltTy); 00827 else if (unsigned(Elt) >= SrcNumElts) { 00828 Type *Ty = IntegerType::get(V2->getContext(), 32); 00829 InElt = 00830 ConstantExpr::getExtractElement(V2, 00831 ConstantInt::get(Ty, Elt - SrcNumElts)); 00832 } else { 00833 Type *Ty = IntegerType::get(V1->getContext(), 32); 00834 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 00835 } 00836 Result.push_back(InElt); 00837 } 00838 00839 return ConstantVector::get(Result); 00840 } 00841 00842 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 00843 ArrayRef<unsigned> Idxs) { 00844 // Base case: no indices, so return the entire value. 00845 if (Idxs.empty()) 00846 return Agg; 00847 00848 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 00849 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 00850 00851 return nullptr; 00852 } 00853 00854 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 00855 Constant *Val, 00856 ArrayRef<unsigned> Idxs) { 00857 // Base case: no indices, so replace the entire value. 00858 if (Idxs.empty()) 00859 return Val; 00860 00861 unsigned NumElts; 00862 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 00863 NumElts = ST->getNumElements(); 00864 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 00865 NumElts = AT->getNumElements(); 00866 else 00867 NumElts = Agg->getType()->getVectorNumElements(); 00868 00869 SmallVector<Constant*, 32> Result; 00870 for (unsigned i = 0; i != NumElts; ++i) { 00871 Constant *C = Agg->getAggregateElement(i); 00872 if (!C) return nullptr; 00873 00874 if (Idxs[0] == i) 00875 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 00876 00877 Result.push_back(C); 00878 } 00879 00880 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 00881 return ConstantStruct::get(ST, Result); 00882 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 00883 return ConstantArray::get(AT, Result); 00884 return ConstantVector::get(Result); 00885 } 00886 00887 00888 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 00889 Constant *C1, Constant *C2) { 00890 // Handle UndefValue up front. 00891 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 00892 switch (Opcode) { 00893 case Instruction::Xor: 00894 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 00895 // Handle undef ^ undef -> 0 special case. This is a common 00896 // idiom (misuse). 00897 return Constant::getNullValue(C1->getType()); 00898 // Fallthrough 00899 case Instruction::Add: 00900 case Instruction::Sub: 00901 return UndefValue::get(C1->getType()); 00902 case Instruction::And: 00903 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 00904 return C1; 00905 return Constant::getNullValue(C1->getType()); // undef & X -> 0 00906 case Instruction::Mul: { 00907 ConstantInt *CI; 00908 // X * undef -> undef if X is odd or undef 00909 if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) || 00910 ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) || 00911 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 00912 return UndefValue::get(C1->getType()); 00913 00914 // X * undef -> 0 otherwise 00915 return Constant::getNullValue(C1->getType()); 00916 } 00917 case Instruction::UDiv: 00918 case Instruction::SDiv: 00919 // undef / 1 -> undef 00920 if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv) 00921 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) 00922 if (CI2->isOne()) 00923 return C1; 00924 // FALL THROUGH 00925 case Instruction::URem: 00926 case Instruction::SRem: 00927 if (!isa<UndefValue>(C2)) // undef / X -> 0 00928 return Constant::getNullValue(C1->getType()); 00929 return C2; // X / undef -> undef 00930 case Instruction::Or: // X | undef -> -1 00931 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 00932 return C1; 00933 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 00934 case Instruction::LShr: 00935 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 00936 return C1; // undef lshr undef -> undef 00937 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 00938 // undef lshr X -> 0 00939 case Instruction::AShr: 00940 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones 00941 return Constant::getAllOnesValue(C1->getType()); 00942 else if (isa<UndefValue>(C1)) 00943 return C1; // undef ashr undef -> undef 00944 else 00945 return C1; // X ashr undef --> X 00946 case Instruction::Shl: 00947 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 00948 return C1; // undef shl undef -> undef 00949 // undef << X -> 0 or X << undef -> 0 00950 return Constant::getNullValue(C1->getType()); 00951 } 00952 } 00953 00954 // Handle simplifications when the RHS is a constant int. 00955 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 00956 switch (Opcode) { 00957 case Instruction::Add: 00958 if (CI2->equalsInt(0)) return C1; // X + 0 == X 00959 break; 00960 case Instruction::Sub: 00961 if (CI2->equalsInt(0)) return C1; // X - 0 == X 00962 break; 00963 case Instruction::Mul: 00964 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 00965 if (CI2->equalsInt(1)) 00966 return C1; // X * 1 == X 00967 break; 00968 case Instruction::UDiv: 00969 case Instruction::SDiv: 00970 if (CI2->equalsInt(1)) 00971 return C1; // X / 1 == X 00972 if (CI2->equalsInt(0)) 00973 return UndefValue::get(CI2->getType()); // X / 0 == undef 00974 break; 00975 case Instruction::URem: 00976 case Instruction::SRem: 00977 if (CI2->equalsInt(1)) 00978 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 00979 if (CI2->equalsInt(0)) 00980 return UndefValue::get(CI2->getType()); // X % 0 == undef 00981 break; 00982 case Instruction::And: 00983 if (CI2->isZero()) return C2; // X & 0 == 0 00984 if (CI2->isAllOnesValue()) 00985 return C1; // X & -1 == X 00986 00987 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 00988 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 00989 if (CE1->getOpcode() == Instruction::ZExt) { 00990 unsigned DstWidth = CI2->getType()->getBitWidth(); 00991 unsigned SrcWidth = 00992 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 00993 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 00994 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 00995 return C1; 00996 } 00997 00998 // If and'ing the address of a global with a constant, fold it. 00999 if (CE1->getOpcode() == Instruction::PtrToInt && 01000 isa<GlobalValue>(CE1->getOperand(0))) { 01001 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 01002 01003 // Functions are at least 4-byte aligned. 01004 unsigned GVAlign = GV->getAlignment(); 01005 if (isa<Function>(GV)) 01006 GVAlign = std::max(GVAlign, 4U); 01007 01008 if (GVAlign > 1) { 01009 unsigned DstWidth = CI2->getType()->getBitWidth(); 01010 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 01011 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 01012 01013 // If checking bits we know are clear, return zero. 01014 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 01015 return Constant::getNullValue(CI2->getType()); 01016 } 01017 } 01018 } 01019 break; 01020 case Instruction::Or: 01021 if (CI2->equalsInt(0)) return C1; // X | 0 == X 01022 if (CI2->isAllOnesValue()) 01023 return C2; // X | -1 == -1 01024 break; 01025 case Instruction::Xor: 01026 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 01027 01028 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 01029 switch (CE1->getOpcode()) { 01030 default: break; 01031 case Instruction::ICmp: 01032 case Instruction::FCmp: 01033 // cmp pred ^ true -> cmp !pred 01034 assert(CI2->equalsInt(1)); 01035 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 01036 pred = CmpInst::getInversePredicate(pred); 01037 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 01038 CE1->getOperand(1)); 01039 } 01040 } 01041 break; 01042 case Instruction::AShr: 01043 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 01044 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 01045 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 01046 return ConstantExpr::getLShr(C1, C2); 01047 break; 01048 } 01049 } else if (isa<ConstantInt>(C1)) { 01050 // If C1 is a ConstantInt and C2 is not, swap the operands. 01051 if (Instruction::isCommutative(Opcode)) 01052 return ConstantExpr::get(Opcode, C2, C1); 01053 } 01054 01055 // At this point we know neither constant is an UndefValue. 01056 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 01057 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 01058 const APInt &C1V = CI1->getValue(); 01059 const APInt &C2V = CI2->getValue(); 01060 switch (Opcode) { 01061 default: 01062 break; 01063 case Instruction::Add: 01064 return ConstantInt::get(CI1->getContext(), C1V + C2V); 01065 case Instruction::Sub: 01066 return ConstantInt::get(CI1->getContext(), C1V - C2V); 01067 case Instruction::Mul: 01068 return ConstantInt::get(CI1->getContext(), C1V * C2V); 01069 case Instruction::UDiv: 01070 assert(!CI2->isNullValue() && "Div by zero handled above"); 01071 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 01072 case Instruction::SDiv: 01073 assert(!CI2->isNullValue() && "Div by zero handled above"); 01074 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 01075 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 01076 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 01077 case Instruction::URem: 01078 assert(!CI2->isNullValue() && "Div by zero handled above"); 01079 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 01080 case Instruction::SRem: 01081 assert(!CI2->isNullValue() && "Div by zero handled above"); 01082 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 01083 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 01084 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 01085 case Instruction::And: 01086 return ConstantInt::get(CI1->getContext(), C1V & C2V); 01087 case Instruction::Or: 01088 return ConstantInt::get(CI1->getContext(), C1V | C2V); 01089 case Instruction::Xor: 01090 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 01091 case Instruction::Shl: { 01092 uint32_t shiftAmt = C2V.getZExtValue(); 01093 if (shiftAmt < C1V.getBitWidth()) 01094 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt)); 01095 else 01096 return UndefValue::get(C1->getType()); // too big shift is undef 01097 } 01098 case Instruction::LShr: { 01099 uint32_t shiftAmt = C2V.getZExtValue(); 01100 if (shiftAmt < C1V.getBitWidth()) 01101 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt)); 01102 else 01103 return UndefValue::get(C1->getType()); // too big shift is undef 01104 } 01105 case Instruction::AShr: { 01106 uint32_t shiftAmt = C2V.getZExtValue(); 01107 if (shiftAmt < C1V.getBitWidth()) 01108 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); 01109 else 01110 return UndefValue::get(C1->getType()); // too big shift is undef 01111 } 01112 } 01113 } 01114 01115 switch (Opcode) { 01116 case Instruction::SDiv: 01117 case Instruction::UDiv: 01118 case Instruction::URem: 01119 case Instruction::SRem: 01120 case Instruction::LShr: 01121 case Instruction::AShr: 01122 case Instruction::Shl: 01123 if (CI1->equalsInt(0)) return C1; 01124 break; 01125 default: 01126 break; 01127 } 01128 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 01129 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 01130 APFloat C1V = CFP1->getValueAPF(); 01131 APFloat C2V = CFP2->getValueAPF(); 01132 APFloat C3V = C1V; // copy for modification 01133 switch (Opcode) { 01134 default: 01135 break; 01136 case Instruction::FAdd: 01137 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 01138 return ConstantFP::get(C1->getContext(), C3V); 01139 case Instruction::FSub: 01140 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 01141 return ConstantFP::get(C1->getContext(), C3V); 01142 case Instruction::FMul: 01143 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 01144 return ConstantFP::get(C1->getContext(), C3V); 01145 case Instruction::FDiv: 01146 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 01147 return ConstantFP::get(C1->getContext(), C3V); 01148 case Instruction::FRem: 01149 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 01150 return ConstantFP::get(C1->getContext(), C3V); 01151 } 01152 } 01153 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 01154 // Perform elementwise folding. 01155 SmallVector<Constant*, 16> Result; 01156 Type *Ty = IntegerType::get(VTy->getContext(), 32); 01157 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 01158 Constant *LHS = 01159 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 01160 Constant *RHS = 01161 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 01162 01163 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 01164 } 01165 01166 return ConstantVector::get(Result); 01167 } 01168 01169 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 01170 // There are many possible foldings we could do here. We should probably 01171 // at least fold add of a pointer with an integer into the appropriate 01172 // getelementptr. This will improve alias analysis a bit. 01173 01174 // Given ((a + b) + c), if (b + c) folds to something interesting, return 01175 // (a + (b + c)). 01176 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 01177 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 01178 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 01179 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 01180 } 01181 } else if (isa<ConstantExpr>(C2)) { 01182 // If C2 is a constant expr and C1 isn't, flop them around and fold the 01183 // other way if possible. 01184 if (Instruction::isCommutative(Opcode)) 01185 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 01186 } 01187 01188 // i1 can be simplified in many cases. 01189 if (C1->getType()->isIntegerTy(1)) { 01190 switch (Opcode) { 01191 case Instruction::Add: 01192 case Instruction::Sub: 01193 return ConstantExpr::getXor(C1, C2); 01194 case Instruction::Mul: 01195 return ConstantExpr::getAnd(C1, C2); 01196 case Instruction::Shl: 01197 case Instruction::LShr: 01198 case Instruction::AShr: 01199 // We can assume that C2 == 0. If it were one the result would be 01200 // undefined because the shift value is as large as the bitwidth. 01201 return C1; 01202 case Instruction::SDiv: 01203 case Instruction::UDiv: 01204 // We can assume that C2 == 1. If it were zero the result would be 01205 // undefined through division by zero. 01206 return C1; 01207 case Instruction::URem: 01208 case Instruction::SRem: 01209 // We can assume that C2 == 1. If it were zero the result would be 01210 // undefined through division by zero. 01211 return ConstantInt::getFalse(C1->getContext()); 01212 default: 01213 break; 01214 } 01215 } 01216 01217 // We don't know how to fold this. 01218 return nullptr; 01219 } 01220 01221 /// isZeroSizedType - This type is zero sized if its an array or structure of 01222 /// zero sized types. The only leaf zero sized type is an empty structure. 01223 static bool isMaybeZeroSizedType(Type *Ty) { 01224 if (StructType *STy = dyn_cast<StructType>(Ty)) { 01225 if (STy->isOpaque()) return true; // Can't say. 01226 01227 // If all of elements have zero size, this does too. 01228 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 01229 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 01230 return true; 01231 01232 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 01233 return isMaybeZeroSizedType(ATy->getElementType()); 01234 } 01235 return false; 01236 } 01237 01238 /// IdxCompare - Compare the two constants as though they were getelementptr 01239 /// indices. This allows coersion of the types to be the same thing. 01240 /// 01241 /// If the two constants are the "same" (after coersion), return 0. If the 01242 /// first is less than the second, return -1, if the second is less than the 01243 /// first, return 1. If the constants are not integral, return -2. 01244 /// 01245 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 01246 if (C1 == C2) return 0; 01247 01248 // Ok, we found a different index. If they are not ConstantInt, we can't do 01249 // anything with them. 01250 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 01251 return -2; // don't know! 01252 01253 // Ok, we have two differing integer indices. Sign extend them to be the same 01254 // type. Long is always big enough, so we use it. 01255 if (!C1->getType()->isIntegerTy(64)) 01256 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); 01257 01258 if (!C2->getType()->isIntegerTy(64)) 01259 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); 01260 01261 if (C1 == C2) return 0; // They are equal 01262 01263 // If the type being indexed over is really just a zero sized type, there is 01264 // no pointer difference being made here. 01265 if (isMaybeZeroSizedType(ElTy)) 01266 return -2; // dunno. 01267 01268 // If they are really different, now that they are the same type, then we 01269 // found a difference! 01270 if (cast<ConstantInt>(C1)->getSExtValue() < 01271 cast<ConstantInt>(C2)->getSExtValue()) 01272 return -1; 01273 else 01274 return 1; 01275 } 01276 01277 /// evaluateFCmpRelation - This function determines if there is anything we can 01278 /// decide about the two constants provided. This doesn't need to handle simple 01279 /// things like ConstantFP comparisons, but should instead handle ConstantExprs. 01280 /// If we can determine that the two constants have a particular relation to 01281 /// each other, we should return the corresponding FCmpInst predicate, 01282 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 01283 /// ConstantFoldCompareInstruction. 01284 /// 01285 /// To simplify this code we canonicalize the relation so that the first 01286 /// operand is always the most "complex" of the two. We consider ConstantFP 01287 /// to be the simplest, and ConstantExprs to be the most complex. 01288 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 01289 assert(V1->getType() == V2->getType() && 01290 "Cannot compare values of different types!"); 01291 01292 // Handle degenerate case quickly 01293 if (V1 == V2) return FCmpInst::FCMP_OEQ; 01294 01295 if (!isa<ConstantExpr>(V1)) { 01296 if (!isa<ConstantExpr>(V2)) { 01297 // We distilled thisUse the standard constant folder for a few cases 01298 ConstantInt *R = nullptr; 01299 R = dyn_cast<ConstantInt>( 01300 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 01301 if (R && !R->isZero()) 01302 return FCmpInst::FCMP_OEQ; 01303 R = dyn_cast<ConstantInt>( 01304 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 01305 if (R && !R->isZero()) 01306 return FCmpInst::FCMP_OLT; 01307 R = dyn_cast<ConstantInt>( 01308 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 01309 if (R && !R->isZero()) 01310 return FCmpInst::FCMP_OGT; 01311 01312 // Nothing more we can do 01313 return FCmpInst::BAD_FCMP_PREDICATE; 01314 } 01315 01316 // If the first operand is simple and second is ConstantExpr, swap operands. 01317 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 01318 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 01319 return FCmpInst::getSwappedPredicate(SwappedRelation); 01320 } else { 01321 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 01322 // constantexpr or a simple constant. 01323 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 01324 switch (CE1->getOpcode()) { 01325 case Instruction::FPTrunc: 01326 case Instruction::FPExt: 01327 case Instruction::UIToFP: 01328 case Instruction::SIToFP: 01329 // We might be able to do something with these but we don't right now. 01330 break; 01331 default: 01332 break; 01333 } 01334 } 01335 // There are MANY other foldings that we could perform here. They will 01336 // probably be added on demand, as they seem needed. 01337 return FCmpInst::BAD_FCMP_PREDICATE; 01338 } 01339 01340 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 01341 const GlobalValue *GV2) { 01342 // Don't try to decide equality of aliases. 01343 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 01344 if (!GV1->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage()) 01345 return ICmpInst::ICMP_NE; 01346 return ICmpInst::BAD_ICMP_PREDICATE; 01347 } 01348 01349 /// evaluateICmpRelation - This function determines if there is anything we can 01350 /// decide about the two constants provided. This doesn't need to handle simple 01351 /// things like integer comparisons, but should instead handle ConstantExprs 01352 /// and GlobalValues. If we can determine that the two constants have a 01353 /// particular relation to each other, we should return the corresponding ICmp 01354 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 01355 /// 01356 /// To simplify this code we canonicalize the relation so that the first 01357 /// operand is always the most "complex" of the two. We consider simple 01358 /// constants (like ConstantInt) to be the simplest, followed by 01359 /// GlobalValues, followed by ConstantExpr's (the most complex). 01360 /// 01361 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 01362 bool isSigned) { 01363 assert(V1->getType() == V2->getType() && 01364 "Cannot compare different types of values!"); 01365 if (V1 == V2) return ICmpInst::ICMP_EQ; 01366 01367 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 01368 !isa<BlockAddress>(V1)) { 01369 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 01370 !isa<BlockAddress>(V2)) { 01371 // We distilled this down to a simple case, use the standard constant 01372 // folder. 01373 ConstantInt *R = nullptr; 01374 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 01375 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 01376 if (R && !R->isZero()) 01377 return pred; 01378 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 01379 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 01380 if (R && !R->isZero()) 01381 return pred; 01382 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 01383 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 01384 if (R && !R->isZero()) 01385 return pred; 01386 01387 // If we couldn't figure it out, bail. 01388 return ICmpInst::BAD_ICMP_PREDICATE; 01389 } 01390 01391 // If the first operand is simple, swap operands. 01392 ICmpInst::Predicate SwappedRelation = 01393 evaluateICmpRelation(V2, V1, isSigned); 01394 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 01395 return ICmpInst::getSwappedPredicate(SwappedRelation); 01396 01397 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 01398 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 01399 ICmpInst::Predicate SwappedRelation = 01400 evaluateICmpRelation(V2, V1, isSigned); 01401 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 01402 return ICmpInst::getSwappedPredicate(SwappedRelation); 01403 return ICmpInst::BAD_ICMP_PREDICATE; 01404 } 01405 01406 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 01407 // constant (which, since the types must match, means that it's a 01408 // ConstantPointerNull). 01409 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 01410 return areGlobalsPotentiallyEqual(GV, GV2); 01411 } else if (isa<BlockAddress>(V2)) { 01412 return ICmpInst::ICMP_NE; // Globals never equal labels. 01413 } else { 01414 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 01415 // GlobalVals can never be null unless they have external weak linkage. 01416 // We don't try to evaluate aliases here. 01417 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 01418 return ICmpInst::ICMP_NE; 01419 } 01420 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 01421 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 01422 ICmpInst::Predicate SwappedRelation = 01423 evaluateICmpRelation(V2, V1, isSigned); 01424 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 01425 return ICmpInst::getSwappedPredicate(SwappedRelation); 01426 return ICmpInst::BAD_ICMP_PREDICATE; 01427 } 01428 01429 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 01430 // constant (which, since the types must match, means that it is a 01431 // ConstantPointerNull). 01432 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 01433 // Block address in another function can't equal this one, but block 01434 // addresses in the current function might be the same if blocks are 01435 // empty. 01436 if (BA2->getFunction() != BA->getFunction()) 01437 return ICmpInst::ICMP_NE; 01438 } else { 01439 // Block addresses aren't null, don't equal the address of globals. 01440 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 01441 "Canonicalization guarantee!"); 01442 return ICmpInst::ICMP_NE; 01443 } 01444 } else { 01445 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 01446 // constantexpr, a global, block address, or a simple constant. 01447 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 01448 Constant *CE1Op0 = CE1->getOperand(0); 01449 01450 switch (CE1->getOpcode()) { 01451 case Instruction::Trunc: 01452 case Instruction::FPTrunc: 01453 case Instruction::FPExt: 01454 case Instruction::FPToUI: 01455 case Instruction::FPToSI: 01456 break; // We can't evaluate floating point casts or truncations. 01457 01458 case Instruction::UIToFP: 01459 case Instruction::SIToFP: 01460 case Instruction::BitCast: 01461 case Instruction::ZExt: 01462 case Instruction::SExt: 01463 // If the cast is not actually changing bits, and the second operand is a 01464 // null pointer, do the comparison with the pre-casted value. 01465 if (V2->isNullValue() && 01466 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 01467 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 01468 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 01469 return evaluateICmpRelation(CE1Op0, 01470 Constant::getNullValue(CE1Op0->getType()), 01471 isSigned); 01472 } 01473 break; 01474 01475 case Instruction::GetElementPtr: { 01476 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 01477 // Ok, since this is a getelementptr, we know that the constant has a 01478 // pointer type. Check the various cases. 01479 if (isa<ConstantPointerNull>(V2)) { 01480 // If we are comparing a GEP to a null pointer, check to see if the base 01481 // of the GEP equals the null pointer. 01482 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 01483 if (GV->hasExternalWeakLinkage()) 01484 // Weak linkage GVals could be zero or not. We're comparing that 01485 // to null pointer so its greater-or-equal 01486 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 01487 else 01488 // If its not weak linkage, the GVal must have a non-zero address 01489 // so the result is greater-than 01490 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 01491 } else if (isa<ConstantPointerNull>(CE1Op0)) { 01492 // If we are indexing from a null pointer, check to see if we have any 01493 // non-zero indices. 01494 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 01495 if (!CE1->getOperand(i)->isNullValue()) 01496 // Offsetting from null, must not be equal. 01497 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 01498 // Only zero indexes from null, must still be zero. 01499 return ICmpInst::ICMP_EQ; 01500 } 01501 // Otherwise, we can't really say if the first operand is null or not. 01502 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 01503 if (isa<ConstantPointerNull>(CE1Op0)) { 01504 if (GV2->hasExternalWeakLinkage()) 01505 // Weak linkage GVals could be zero or not. We're comparing it to 01506 // a null pointer, so its less-or-equal 01507 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 01508 else 01509 // If its not weak linkage, the GVal must have a non-zero address 01510 // so the result is less-than 01511 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 01512 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 01513 if (GV == GV2) { 01514 // If this is a getelementptr of the same global, then it must be 01515 // different. Because the types must match, the getelementptr could 01516 // only have at most one index, and because we fold getelementptr's 01517 // with a single zero index, it must be nonzero. 01518 assert(CE1->getNumOperands() == 2 && 01519 !CE1->getOperand(1)->isNullValue() && 01520 "Surprising getelementptr!"); 01521 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 01522 } else { 01523 if (CE1GEP->hasAllZeroIndices()) 01524 return areGlobalsPotentiallyEqual(GV, GV2); 01525 return ICmpInst::BAD_ICMP_PREDICATE; 01526 } 01527 } 01528 } else { 01529 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 01530 Constant *CE2Op0 = CE2->getOperand(0); 01531 01532 // There are MANY other foldings that we could perform here. They will 01533 // probably be added on demand, as they seem needed. 01534 switch (CE2->getOpcode()) { 01535 default: break; 01536 case Instruction::GetElementPtr: 01537 // By far the most common case to handle is when the base pointers are 01538 // obviously to the same global. 01539 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 01540 // Don't know relative ordering, but check for inequality. 01541 if (CE1Op0 != CE2Op0) { 01542 GEPOperator *CE2GEP = cast<GEPOperator>(CE2); 01543 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 01544 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 01545 cast<GlobalValue>(CE2Op0)); 01546 return ICmpInst::BAD_ICMP_PREDICATE; 01547 } 01548 // Ok, we know that both getelementptr instructions are based on the 01549 // same global. From this, we can precisely determine the relative 01550 // ordering of the resultant pointers. 01551 unsigned i = 1; 01552 01553 // The logic below assumes that the result of the comparison 01554 // can be determined by finding the first index that differs. 01555 // This doesn't work if there is over-indexing in any 01556 // subsequent indices, so check for that case first. 01557 if (!CE1->isGEPWithNoNotionalOverIndexing() || 01558 !CE2->isGEPWithNoNotionalOverIndexing()) 01559 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 01560 01561 // Compare all of the operands the GEP's have in common. 01562 gep_type_iterator GTI = gep_type_begin(CE1); 01563 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 01564 ++i, ++GTI) 01565 switch (IdxCompare(CE1->getOperand(i), 01566 CE2->getOperand(i), GTI.getIndexedType())) { 01567 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 01568 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 01569 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 01570 } 01571 01572 // Ok, we ran out of things they have in common. If any leftovers 01573 // are non-zero then we have a difference, otherwise we are equal. 01574 for (; i < CE1->getNumOperands(); ++i) 01575 if (!CE1->getOperand(i)->isNullValue()) { 01576 if (isa<ConstantInt>(CE1->getOperand(i))) 01577 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 01578 else 01579 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 01580 } 01581 01582 for (; i < CE2->getNumOperands(); ++i) 01583 if (!CE2->getOperand(i)->isNullValue()) { 01584 if (isa<ConstantInt>(CE2->getOperand(i))) 01585 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 01586 else 01587 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 01588 } 01589 return ICmpInst::ICMP_EQ; 01590 } 01591 } 01592 } 01593 } 01594 default: 01595 break; 01596 } 01597 } 01598 01599 return ICmpInst::BAD_ICMP_PREDICATE; 01600 } 01601 01602 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 01603 Constant *C1, Constant *C2) { 01604 Type *ResultTy; 01605 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 01606 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 01607 VT->getNumElements()); 01608 else 01609 ResultTy = Type::getInt1Ty(C1->getContext()); 01610 01611 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 01612 if (pred == FCmpInst::FCMP_FALSE) 01613 return Constant::getNullValue(ResultTy); 01614 01615 if (pred == FCmpInst::FCMP_TRUE) 01616 return Constant::getAllOnesValue(ResultTy); 01617 01618 // Handle some degenerate cases first 01619 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 01620 // For EQ and NE, we can always pick a value for the undef to make the 01621 // predicate pass or fail, so we can return undef. 01622 // Also, if both operands are undef, we can return undef. 01623 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) || 01624 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 01625 return UndefValue::get(ResultTy); 01626 // Otherwise, pick the same value as the non-undef operand, and fold 01627 // it to true or false. 01628 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); 01629 } 01630 01631 // icmp eq/ne(null,GV) -> false/true 01632 if (C1->isNullValue()) { 01633 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 01634 // Don't try to evaluate aliases. External weak GV can be null. 01635 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 01636 if (pred == ICmpInst::ICMP_EQ) 01637 return ConstantInt::getFalse(C1->getContext()); 01638 else if (pred == ICmpInst::ICMP_NE) 01639 return ConstantInt::getTrue(C1->getContext()); 01640 } 01641 // icmp eq/ne(GV,null) -> false/true 01642 } else if (C2->isNullValue()) { 01643 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 01644 // Don't try to evaluate aliases. External weak GV can be null. 01645 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 01646 if (pred == ICmpInst::ICMP_EQ) 01647 return ConstantInt::getFalse(C1->getContext()); 01648 else if (pred == ICmpInst::ICMP_NE) 01649 return ConstantInt::getTrue(C1->getContext()); 01650 } 01651 } 01652 01653 // If the comparison is a comparison between two i1's, simplify it. 01654 if (C1->getType()->isIntegerTy(1)) { 01655 switch(pred) { 01656 case ICmpInst::ICMP_EQ: 01657 if (isa<ConstantInt>(C2)) 01658 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 01659 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 01660 case ICmpInst::ICMP_NE: 01661 return ConstantExpr::getXor(C1, C2); 01662 default: 01663 break; 01664 } 01665 } 01666 01667 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 01668 APInt V1 = cast<ConstantInt>(C1)->getValue(); 01669 APInt V2 = cast<ConstantInt>(C2)->getValue(); 01670 switch (pred) { 01671 default: llvm_unreachable("Invalid ICmp Predicate"); 01672 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 01673 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 01674 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 01675 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 01676 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 01677 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 01678 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 01679 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 01680 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 01681 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 01682 } 01683 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 01684 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 01685 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 01686 APFloat::cmpResult R = C1V.compare(C2V); 01687 switch (pred) { 01688 default: llvm_unreachable("Invalid FCmp Predicate"); 01689 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 01690 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 01691 case FCmpInst::FCMP_UNO: 01692 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 01693 case FCmpInst::FCMP_ORD: 01694 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 01695 case FCmpInst::FCMP_UEQ: 01696 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 01697 R==APFloat::cmpEqual); 01698 case FCmpInst::FCMP_OEQ: 01699 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 01700 case FCmpInst::FCMP_UNE: 01701 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 01702 case FCmpInst::FCMP_ONE: 01703 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 01704 R==APFloat::cmpGreaterThan); 01705 case FCmpInst::FCMP_ULT: 01706 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 01707 R==APFloat::cmpLessThan); 01708 case FCmpInst::FCMP_OLT: 01709 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 01710 case FCmpInst::FCMP_UGT: 01711 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 01712 R==APFloat::cmpGreaterThan); 01713 case FCmpInst::FCMP_OGT: 01714 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 01715 case FCmpInst::FCMP_ULE: 01716 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 01717 case FCmpInst::FCMP_OLE: 01718 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 01719 R==APFloat::cmpEqual); 01720 case FCmpInst::FCMP_UGE: 01721 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 01722 case FCmpInst::FCMP_OGE: 01723 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 01724 R==APFloat::cmpEqual); 01725 } 01726 } else if (C1->getType()->isVectorTy()) { 01727 // If we can constant fold the comparison of each element, constant fold 01728 // the whole vector comparison. 01729 SmallVector<Constant*, 4> ResElts; 01730 Type *Ty = IntegerType::get(C1->getContext(), 32); 01731 // Compare the elements, producing an i1 result or constant expr. 01732 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 01733 Constant *C1E = 01734 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 01735 Constant *C2E = 01736 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 01737 01738 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 01739 } 01740 01741 return ConstantVector::get(ResElts); 01742 } 01743 01744 if (C1->getType()->isFloatingPointTy()) { 01745 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 01746 switch (evaluateFCmpRelation(C1, C2)) { 01747 default: llvm_unreachable("Unknown relation!"); 01748 case FCmpInst::FCMP_UNO: 01749 case FCmpInst::FCMP_ORD: 01750 case FCmpInst::FCMP_UEQ: 01751 case FCmpInst::FCMP_UNE: 01752 case FCmpInst::FCMP_ULT: 01753 case FCmpInst::FCMP_UGT: 01754 case FCmpInst::FCMP_ULE: 01755 case FCmpInst::FCMP_UGE: 01756 case FCmpInst::FCMP_TRUE: 01757 case FCmpInst::FCMP_FALSE: 01758 case FCmpInst::BAD_FCMP_PREDICATE: 01759 break; // Couldn't determine anything about these constants. 01760 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 01761 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 01762 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 01763 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 01764 break; 01765 case FCmpInst::FCMP_OLT: // We know that C1 < C2 01766 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 01767 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 01768 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 01769 break; 01770 case FCmpInst::FCMP_OGT: // We know that C1 > C2 01771 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 01772 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 01773 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 01774 break; 01775 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 01776 // We can only partially decide this relation. 01777 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 01778 Result = 0; 01779 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 01780 Result = 1; 01781 break; 01782 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 01783 // We can only partially decide this relation. 01784 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 01785 Result = 0; 01786 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 01787 Result = 1; 01788 break; 01789 case FCmpInst::FCMP_ONE: // We know that C1 != C2 01790 // We can only partially decide this relation. 01791 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 01792 Result = 0; 01793 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 01794 Result = 1; 01795 break; 01796 } 01797 01798 // If we evaluated the result, return it now. 01799 if (Result != -1) 01800 return ConstantInt::get(ResultTy, Result); 01801 01802 } else { 01803 // Evaluate the relation between the two constants, per the predicate. 01804 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 01805 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { 01806 default: llvm_unreachable("Unknown relational!"); 01807 case ICmpInst::BAD_ICMP_PREDICATE: 01808 break; // Couldn't determine anything about these constants. 01809 case ICmpInst::ICMP_EQ: // We know the constants are equal! 01810 // If we know the constants are equal, we can decide the result of this 01811 // computation precisely. 01812 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 01813 break; 01814 case ICmpInst::ICMP_ULT: 01815 switch (pred) { 01816 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 01817 Result = 1; break; 01818 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 01819 Result = 0; break; 01820 } 01821 break; 01822 case ICmpInst::ICMP_SLT: 01823 switch (pred) { 01824 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 01825 Result = 1; break; 01826 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 01827 Result = 0; break; 01828 } 01829 break; 01830 case ICmpInst::ICMP_UGT: 01831 switch (pred) { 01832 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 01833 Result = 1; break; 01834 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 01835 Result = 0; break; 01836 } 01837 break; 01838 case ICmpInst::ICMP_SGT: 01839 switch (pred) { 01840 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 01841 Result = 1; break; 01842 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 01843 Result = 0; break; 01844 } 01845 break; 01846 case ICmpInst::ICMP_ULE: 01847 if (pred == ICmpInst::ICMP_UGT) Result = 0; 01848 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 01849 break; 01850 case ICmpInst::ICMP_SLE: 01851 if (pred == ICmpInst::ICMP_SGT) Result = 0; 01852 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 01853 break; 01854 case ICmpInst::ICMP_UGE: 01855 if (pred == ICmpInst::ICMP_ULT) Result = 0; 01856 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 01857 break; 01858 case ICmpInst::ICMP_SGE: 01859 if (pred == ICmpInst::ICMP_SLT) Result = 0; 01860 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 01861 break; 01862 case ICmpInst::ICMP_NE: 01863 if (pred == ICmpInst::ICMP_EQ) Result = 0; 01864 if (pred == ICmpInst::ICMP_NE) Result = 1; 01865 break; 01866 } 01867 01868 // If we evaluated the result, return it now. 01869 if (Result != -1) 01870 return ConstantInt::get(ResultTy, Result); 01871 01872 // If the right hand side is a bitcast, try using its inverse to simplify 01873 // it by moving it to the left hand side. We can't do this if it would turn 01874 // a vector compare into a scalar compare or visa versa. 01875 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 01876 Constant *CE2Op0 = CE2->getOperand(0); 01877 if (CE2->getOpcode() == Instruction::BitCast && 01878 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 01879 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 01880 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 01881 } 01882 } 01883 01884 // If the left hand side is an extension, try eliminating it. 01885 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 01886 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || 01887 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ 01888 Constant *CE1Op0 = CE1->getOperand(0); 01889 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 01890 if (CE1Inverse == CE1Op0) { 01891 // Check whether we can safely truncate the right hand side. 01892 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 01893 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 01894 C2->getType()) == C2) 01895 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 01896 } 01897 } 01898 } 01899 01900 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 01901 (C1->isNullValue() && !C2->isNullValue())) { 01902 // If C2 is a constant expr and C1 isn't, flip them around and fold the 01903 // other way if possible. 01904 // Also, if C1 is null and C2 isn't, flip them around. 01905 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 01906 return ConstantExpr::getICmp(pred, C2, C1); 01907 } 01908 } 01909 return nullptr; 01910 } 01911 01912 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices 01913 /// is "inbounds". 01914 template<typename IndexTy> 01915 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 01916 // No indices means nothing that could be out of bounds. 01917 if (Idxs.empty()) return true; 01918 01919 // If the first index is zero, it's in bounds. 01920 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 01921 01922 // If the first index is one and all the rest are zero, it's in bounds, 01923 // by the one-past-the-end rule. 01924 if (!cast<ConstantInt>(Idxs[0])->isOne()) 01925 return false; 01926 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 01927 if (!cast<Constant>(Idxs[i])->isNullValue()) 01928 return false; 01929 return true; 01930 } 01931 01932 /// \brief Test whether a given ConstantInt is in-range for a SequentialType. 01933 static bool isIndexInRangeOfSequentialType(const SequentialType *STy, 01934 const ConstantInt *CI) { 01935 if (const PointerType *PTy = dyn_cast<PointerType>(STy)) 01936 // Only handle pointers to sized types, not pointers to functions. 01937 return PTy->getElementType()->isSized(); 01938 01939 uint64_t NumElements = 0; 01940 // Determine the number of elements in our sequential type. 01941 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 01942 NumElements = ATy->getNumElements(); 01943 else if (const VectorType *VTy = dyn_cast<VectorType>(STy)) 01944 NumElements = VTy->getNumElements(); 01945 01946 assert((isa<ArrayType>(STy) || NumElements > 0) && 01947 "didn't expect non-array type to have zero elements!"); 01948 01949 // We cannot bounds check the index if it doesn't fit in an int64_t. 01950 if (CI->getValue().getActiveBits() > 64) 01951 return false; 01952 01953 // A negative index or an index past the end of our sequential type is 01954 // considered out-of-range. 01955 int64_t IndexVal = CI->getSExtValue(); 01956 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 01957 return false; 01958 01959 // Otherwise, it is in-range. 01960 return true; 01961 } 01962 01963 template<typename IndexTy> 01964 static Constant *ConstantFoldGetElementPtrImpl(Constant *C, 01965 bool inBounds, 01966 ArrayRef<IndexTy> Idxs) { 01967 if (Idxs.empty()) return C; 01968 Constant *Idx0 = cast<Constant>(Idxs[0]); 01969 if ((Idxs.size() == 1 && Idx0->isNullValue())) 01970 return C; 01971 01972 if (isa<UndefValue>(C)) { 01973 PointerType *Ptr = cast<PointerType>(C->getType()); 01974 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); 01975 assert(Ty && "Invalid indices for GEP!"); 01976 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 01977 } 01978 01979 if (C->isNullValue()) { 01980 bool isNull = true; 01981 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 01982 if (!cast<Constant>(Idxs[i])->isNullValue()) { 01983 isNull = false; 01984 break; 01985 } 01986 if (isNull) { 01987 PointerType *Ptr = cast<PointerType>(C->getType()); 01988 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); 01989 assert(Ty && "Invalid indices for GEP!"); 01990 return ConstantPointerNull::get(PointerType::get(Ty, 01991 Ptr->getAddressSpace())); 01992 } 01993 } 01994 01995 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 01996 // Combine Indices - If the source pointer to this getelementptr instruction 01997 // is a getelementptr instruction, combine the indices of the two 01998 // getelementptr instructions into a single instruction. 01999 // 02000 if (CE->getOpcode() == Instruction::GetElementPtr) { 02001 Type *LastTy = nullptr; 02002 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 02003 I != E; ++I) 02004 LastTy = *I; 02005 02006 // We cannot combine indices if doing so would take us outside of an 02007 // array or vector. Doing otherwise could trick us if we evaluated such a 02008 // GEP as part of a load. 02009 // 02010 // e.g. Consider if the original GEP was: 02011 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 02012 // i32 0, i32 0, i64 0) 02013 // 02014 // If we then tried to offset it by '8' to get to the third element, 02015 // an i8, we should *not* get: 02016 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 02017 // i32 0, i32 0, i64 8) 02018 // 02019 // This GEP tries to index array element '8 which runs out-of-bounds. 02020 // Subsequent evaluation would get confused and produce erroneous results. 02021 // 02022 // The following prohibits such a GEP from being formed by checking to see 02023 // if the index is in-range with respect to an array or vector. 02024 bool PerformFold = false; 02025 if (Idx0->isNullValue()) 02026 PerformFold = true; 02027 else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy)) 02028 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 02029 PerformFold = isIndexInRangeOfSequentialType(STy, CI); 02030 02031 if (PerformFold) { 02032 SmallVector<Value*, 16> NewIndices; 02033 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 02034 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 02035 NewIndices.push_back(CE->getOperand(i)); 02036 02037 // Add the last index of the source with the first index of the new GEP. 02038 // Make sure to handle the case when they are actually different types. 02039 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 02040 // Otherwise it must be an array. 02041 if (!Idx0->isNullValue()) { 02042 Type *IdxTy = Combined->getType(); 02043 if (IdxTy != Idx0->getType()) { 02044 Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); 02045 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); 02046 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); 02047 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 02048 } else { 02049 Combined = 02050 ConstantExpr::get(Instruction::Add, Idx0, Combined); 02051 } 02052 } 02053 02054 NewIndices.push_back(Combined); 02055 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 02056 return 02057 ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices, 02058 inBounds && 02059 cast<GEPOperator>(CE)->isInBounds()); 02060 } 02061 } 02062 02063 // Attempt to fold casts to the same type away. For example, folding: 02064 // 02065 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 02066 // i64 0, i64 0) 02067 // into: 02068 // 02069 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 02070 // 02071 // Don't fold if the cast is changing address spaces. 02072 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 02073 PointerType *SrcPtrTy = 02074 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 02075 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 02076 if (SrcPtrTy && DstPtrTy) { 02077 ArrayType *SrcArrayTy = 02078 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 02079 ArrayType *DstArrayTy = 02080 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 02081 if (SrcArrayTy && DstArrayTy 02082 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 02083 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 02084 return ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0), 02085 Idxs, inBounds); 02086 } 02087 } 02088 } 02089 02090 // Check to see if any array indices are not within the corresponding 02091 // notional array or vector bounds. If so, try to determine if they can be 02092 // factored out into preceding dimensions. 02093 bool Unknown = false; 02094 SmallVector<Constant *, 8> NewIdxs; 02095 Type *Ty = C->getType(); 02096 Type *Prev = nullptr; 02097 for (unsigned i = 0, e = Idxs.size(); i != e; 02098 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 02099 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 02100 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) 02101 if (CI->getSExtValue() > 0 && 02102 !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) { 02103 if (isa<SequentialType>(Prev)) { 02104 // It's out of range, but we can factor it into the prior 02105 // dimension. 02106 NewIdxs.resize(Idxs.size()); 02107 uint64_t NumElements = 0; 02108 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 02109 NumElements = ATy->getNumElements(); 02110 else 02111 NumElements = cast<VectorType>(Ty)->getNumElements(); 02112 02113 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements); 02114 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 02115 02116 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 02117 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 02118 02119 // Before adding, extend both operands to i64 to avoid 02120 // overflow trouble. 02121 if (!PrevIdx->getType()->isIntegerTy(64)) 02122 PrevIdx = ConstantExpr::getSExt(PrevIdx, 02123 Type::getInt64Ty(Div->getContext())); 02124 if (!Div->getType()->isIntegerTy(64)) 02125 Div = ConstantExpr::getSExt(Div, 02126 Type::getInt64Ty(Div->getContext())); 02127 02128 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 02129 } else { 02130 // It's out of range, but the prior dimension is a struct 02131 // so we can't do anything about it. 02132 Unknown = true; 02133 } 02134 } 02135 } else { 02136 // We don't know if it's in range or not. 02137 Unknown = true; 02138 } 02139 } 02140 02141 // If we did any factoring, start over with the adjusted indices. 02142 if (!NewIdxs.empty()) { 02143 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 02144 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 02145 return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds); 02146 } 02147 02148 // If all indices are known integers and normalized, we can do a simple 02149 // check for the "inbounds" property. 02150 if (!Unknown && !inBounds) 02151 if (auto *GV = dyn_cast<GlobalVariable>(C)) 02152 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 02153 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs); 02154 02155 return nullptr; 02156 } 02157 02158 Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 02159 bool inBounds, 02160 ArrayRef<Constant *> Idxs) { 02161 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 02162 } 02163 02164 Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 02165 bool inBounds, 02166 ArrayRef<Value *> Idxs) { 02167 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 02168 }