ConstantFold.cpp revision 3bfbc4587a7e79f08f8c126a9e62c3475fb90f8b
1//===- ConstantFold.cpp - LLVM constant folder ----------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements folding of constants for LLVM. This implements the 11// (internal) ConstantFold.h interface, which is used by the 12// ConstantExpr::get* methods to automatically fold constants when possible. 13// 14// The current constant folding implementation is implemented in two pieces: the 15// pieces that don't need TargetData, and the pieces that do. This is to avoid 16// a dependence in VMCore on Target. 17// 18//===----------------------------------------------------------------------===// 19 20#include "ConstantFold.h" 21#include "llvm/Constants.h" 22#include "llvm/Instructions.h" 23#include "llvm/DerivedTypes.h" 24#include "llvm/Function.h" 25#include "llvm/GlobalAlias.h" 26#include "llvm/GlobalVariable.h" 27#include "llvm/LLVMContext.h" 28#include "llvm/ADT/SmallVector.h" 29#include "llvm/Support/Compiler.h" 30#include "llvm/Support/ErrorHandling.h" 31#include "llvm/Support/GetElementPtrTypeIterator.h" 32#include "llvm/Support/ManagedStatic.h" 33#include "llvm/Support/MathExtras.h" 34#include <limits> 35using namespace llvm; 36 37//===----------------------------------------------------------------------===// 38// ConstantFold*Instruction Implementations 39//===----------------------------------------------------------------------===// 40 41/// BitCastConstantVector - Convert the specified ConstantVector node to the 42/// specified vector type. At this point, we know that the elements of the 43/// input vector constant are all simple integer or FP values. 44static Constant *BitCastConstantVector(LLVMContext &Context, ConstantVector *CV, 45 const VectorType *DstTy) { 46 // If this cast changes element count then we can't handle it here: 47 // doing so requires endianness information. This should be handled by 48 // Analysis/ConstantFolding.cpp 49 unsigned NumElts = DstTy->getNumElements(); 50 if (NumElts != CV->getNumOperands()) 51 return 0; 52 53 // Check to verify that all elements of the input are simple. 54 for (unsigned i = 0; i != NumElts; ++i) { 55 if (!isa<ConstantInt>(CV->getOperand(i)) && 56 !isa<ConstantFP>(CV->getOperand(i))) 57 return 0; 58 } 59 60 // Bitcast each element now. 61 std::vector<Constant*> Result; 62 const Type *DstEltTy = DstTy->getElementType(); 63 for (unsigned i = 0; i != NumElts; ++i) 64 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i), 65 DstEltTy)); 66 return ConstantVector::get(Result); 67} 68 69/// This function determines which opcode to use to fold two constant cast 70/// expressions together. It uses CastInst::isEliminableCastPair to determine 71/// the opcode. Consequently its just a wrapper around that function. 72/// @brief Determine if it is valid to fold a cast of a cast 73static unsigned 74foldConstantCastPair( 75 unsigned opc, ///< opcode of the second cast constant expression 76 const ConstantExpr*Op, ///< the first cast constant expression 77 const Type *DstTy ///< desintation type of the first cast 78) { 79 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 80 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 81 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 82 83 // The the types and opcodes for the two Cast constant expressions 84 const Type *SrcTy = Op->getOperand(0)->getType(); 85 const Type *MidTy = Op->getType(); 86 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 87 Instruction::CastOps secondOp = Instruction::CastOps(opc); 88 89 // Let CastInst::isEliminableCastPair do the heavy lifting. 90 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 91 Type::getInt64Ty(DstTy->getContext())); 92} 93 94static Constant *FoldBitCast(LLVMContext &Context, 95 Constant *V, const Type *DestTy) { 96 const Type *SrcTy = V->getType(); 97 if (SrcTy == DestTy) 98 return V; // no-op cast 99 100 // Check to see if we are casting a pointer to an aggregate to a pointer to 101 // the first element. If so, return the appropriate GEP instruction. 102 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) 103 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 104 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) { 105 SmallVector<Value*, 8> IdxList; 106 Value *Zero = Constant::getNullValue(Type::getInt32Ty(Context)); 107 IdxList.push_back(Zero); 108 const Type *ElTy = PTy->getElementType(); 109 while (ElTy != DPTy->getElementType()) { 110 if (const StructType *STy = dyn_cast<StructType>(ElTy)) { 111 if (STy->getNumElements() == 0) break; 112 ElTy = STy->getElementType(0); 113 IdxList.push_back(Zero); 114 } else if (const SequentialType *STy = 115 dyn_cast<SequentialType>(ElTy)) { 116 if (isa<PointerType>(ElTy)) break; // Can't index into pointers! 117 ElTy = STy->getElementType(); 118 IdxList.push_back(Zero); 119 } else { 120 break; 121 } 122 } 123 124 if (ElTy == DPTy->getElementType()) 125 // This GEP is inbounds because all indices are zero. 126 return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0], 127 IdxList.size()); 128 } 129 130 // Handle casts from one vector constant to another. We know that the src 131 // and dest type have the same size (otherwise its an illegal cast). 132 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 133 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 134 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 135 "Not cast between same sized vectors!"); 136 SrcTy = NULL; 137 // First, check for null. Undef is already handled. 138 if (isa<ConstantAggregateZero>(V)) 139 return Constant::getNullValue(DestTy); 140 141 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 142 return BitCastConstantVector(Context, CV, DestPTy); 143 } 144 145 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 146 // This allows for other simplifications (although some of them 147 // can only be handled by Analysis/ConstantFolding.cpp). 148 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 149 return ConstantExpr::getBitCast( 150 ConstantVector::get(&V, 1), DestPTy); 151 } 152 153 // Finally, implement bitcast folding now. The code below doesn't handle 154 // bitcast right. 155 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 156 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 157 158 // Handle integral constant input. 159 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 160 if (DestTy->isInteger()) 161 // Integral -> Integral. This is a no-op because the bit widths must 162 // be the same. Consequently, we just fold to V. 163 return V; 164 165 if (DestTy->isFloatingPoint()) 166 return ConstantFP::get(Context, APFloat(CI->getValue(), 167 DestTy != Type::getPPC_FP128Ty(Context))); 168 169 // Otherwise, can't fold this (vector?) 170 return 0; 171 } 172 173 // Handle ConstantFP input. 174 if (const ConstantFP *FP = dyn_cast<ConstantFP>(V)) 175 // FP -> Integral. 176 return ConstantInt::get(Context, FP->getValueAPF().bitcastToAPInt()); 177 178 return 0; 179} 180 181 182Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, 183 unsigned opc, const Constant *V, 184 const Type *DestTy) { 185 if (isa<UndefValue>(V)) { 186 // zext(undef) = 0, because the top bits will be zero. 187 // sext(undef) = 0, because the top bits will all be the same. 188 // [us]itofp(undef) = 0, because the result value is bounded. 189 if (opc == Instruction::ZExt || opc == Instruction::SExt || 190 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 191 return Constant::getNullValue(DestTy); 192 return UndefValue::get(DestTy); 193 } 194 // No compile-time operations on this type yet. 195 if (V->getType() == Type::getPPC_FP128Ty(Context) || DestTy == Type::getPPC_FP128Ty(Context)) 196 return 0; 197 198 // If the cast operand is a constant expression, there's a few things we can 199 // do to try to simplify it. 200 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 201 if (CE->isCast()) { 202 // Try hard to fold cast of cast because they are often eliminable. 203 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 204 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 205 } else if (CE->getOpcode() == Instruction::GetElementPtr) { 206 // If all of the indexes in the GEP are null values, there is no pointer 207 // adjustment going on. We might as well cast the source pointer. 208 bool isAllNull = true; 209 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 210 if (!CE->getOperand(i)->isNullValue()) { 211 isAllNull = false; 212 break; 213 } 214 if (isAllNull) 215 // This is casting one pointer type to another, always BitCast 216 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 217 } 218 } 219 220 // If the cast operand is a constant vector, perform the cast by 221 // operating on each element. In the cast of bitcasts, the element 222 // count may be mismatched; don't attempt to handle that here. 223 if (const ConstantVector *CV = dyn_cast<ConstantVector>(V)) 224 if (isa<VectorType>(DestTy) && 225 cast<VectorType>(DestTy)->getNumElements() == 226 CV->getType()->getNumElements()) { 227 std::vector<Constant*> res; 228 const VectorType *DestVecTy = cast<VectorType>(DestTy); 229 const Type *DstEltTy = DestVecTy->getElementType(); 230 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) 231 res.push_back(ConstantExpr::getCast(opc, 232 CV->getOperand(i), DstEltTy)); 233 return ConstantVector::get(DestVecTy, res); 234 } 235 236 // We actually have to do a cast now. Perform the cast according to the 237 // opcode specified. 238 switch (opc) { 239 case Instruction::FPTrunc: 240 case Instruction::FPExt: 241 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 242 bool ignored; 243 APFloat Val = FPC->getValueAPF(); 244 Val.convert(DestTy == Type::getFloatTy(Context) ? APFloat::IEEEsingle : 245 DestTy == Type::getDoubleTy(Context) ? APFloat::IEEEdouble : 246 DestTy == Type::getX86_FP80Ty(Context) ? APFloat::x87DoubleExtended : 247 DestTy == Type::getFP128Ty(Context) ? APFloat::IEEEquad : 248 APFloat::Bogus, 249 APFloat::rmNearestTiesToEven, &ignored); 250 return ConstantFP::get(Context, Val); 251 } 252 return 0; // Can't fold. 253 case Instruction::FPToUI: 254 case Instruction::FPToSI: 255 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 256 const APFloat &V = FPC->getValueAPF(); 257 bool ignored; 258 uint64_t x[2]; 259 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 260 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 261 APFloat::rmTowardZero, &ignored); 262 APInt Val(DestBitWidth, 2, x); 263 return ConstantInt::get(Context, Val); 264 } 265 return 0; // Can't fold. 266 case Instruction::IntToPtr: //always treated as unsigned 267 if (V->isNullValue()) // Is it an integral null value? 268 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 269 return 0; // Other pointer types cannot be casted 270 case Instruction::PtrToInt: // always treated as unsigned 271 if (V->isNullValue()) // is it a null pointer value? 272 return ConstantInt::get(DestTy, 0); 273 return 0; // Other pointer types cannot be casted 274 case Instruction::UIToFP: 275 case Instruction::SIToFP: 276 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 277 APInt api = CI->getValue(); 278 const uint64_t zero[] = {0, 0}; 279 APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(), 280 2, zero)); 281 (void)apf.convertFromAPInt(api, 282 opc==Instruction::SIToFP, 283 APFloat::rmNearestTiesToEven); 284 return ConstantFP::get(Context, apf); 285 } 286 return 0; 287 case Instruction::ZExt: 288 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 289 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 290 APInt Result(CI->getValue()); 291 Result.zext(BitWidth); 292 return ConstantInt::get(Context, Result); 293 } 294 return 0; 295 case Instruction::SExt: 296 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 297 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 298 APInt Result(CI->getValue()); 299 Result.sext(BitWidth); 300 return ConstantInt::get(Context, Result); 301 } 302 return 0; 303 case Instruction::Trunc: 304 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 305 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 306 APInt Result(CI->getValue()); 307 Result.trunc(BitWidth); 308 return ConstantInt::get(Context, Result); 309 } 310 return 0; 311 case Instruction::BitCast: 312 return FoldBitCast(Context, const_cast<Constant*>(V), DestTy); 313 default: 314 assert(!"Invalid CE CastInst opcode"); 315 break; 316 } 317 318 llvm_unreachable("Failed to cast constant expression"); 319 return 0; 320} 321 322Constant *llvm::ConstantFoldSelectInstruction(LLVMContext&, 323 const Constant *Cond, 324 const Constant *V1, 325 const Constant *V2) { 326 if (const ConstantInt *CB = dyn_cast<ConstantInt>(Cond)) 327 return const_cast<Constant*>(CB->getZExtValue() ? V1 : V2); 328 329 if (isa<UndefValue>(V1)) return const_cast<Constant*>(V2); 330 if (isa<UndefValue>(V2)) return const_cast<Constant*>(V1); 331 if (isa<UndefValue>(Cond)) return const_cast<Constant*>(V1); 332 if (V1 == V2) return const_cast<Constant*>(V1); 333 return 0; 334} 335 336Constant *llvm::ConstantFoldExtractElementInstruction(LLVMContext &Context, 337 const Constant *Val, 338 const Constant *Idx) { 339 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 340 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType()); 341 if (Val->isNullValue()) // ee(zero, x) -> zero 342 return Constant::getNullValue( 343 cast<VectorType>(Val->getType())->getElementType()); 344 345 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 346 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 347 return CVal->getOperand(CIdx->getZExtValue()); 348 } else if (isa<UndefValue>(Idx)) { 349 // ee({w,x,y,z}, undef) -> w (an arbitrary value). 350 return CVal->getOperand(0); 351 } 352 } 353 return 0; 354} 355 356Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, 357 const Constant *Val, 358 const Constant *Elt, 359 const Constant *Idx) { 360 const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 361 if (!CIdx) return 0; 362 APInt idxVal = CIdx->getValue(); 363 if (isa<UndefValue>(Val)) { 364 // Insertion of scalar constant into vector undef 365 // Optimize away insertion of undef 366 if (isa<UndefValue>(Elt)) 367 return const_cast<Constant*>(Val); 368 // Otherwise break the aggregate undef into multiple undefs and do 369 // the insertion 370 unsigned numOps = 371 cast<VectorType>(Val->getType())->getNumElements(); 372 std::vector<Constant*> Ops; 373 Ops.reserve(numOps); 374 for (unsigned i = 0; i < numOps; ++i) { 375 const Constant *Op = 376 (idxVal == i) ? Elt : UndefValue::get(Elt->getType()); 377 Ops.push_back(const_cast<Constant*>(Op)); 378 } 379 return ConstantVector::get(Ops); 380 } 381 if (isa<ConstantAggregateZero>(Val)) { 382 // Insertion of scalar constant into vector aggregate zero 383 // Optimize away insertion of zero 384 if (Elt->isNullValue()) 385 return const_cast<Constant*>(Val); 386 // Otherwise break the aggregate zero into multiple zeros and do 387 // the insertion 388 unsigned numOps = 389 cast<VectorType>(Val->getType())->getNumElements(); 390 std::vector<Constant*> Ops; 391 Ops.reserve(numOps); 392 for (unsigned i = 0; i < numOps; ++i) { 393 const Constant *Op = 394 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType()); 395 Ops.push_back(const_cast<Constant*>(Op)); 396 } 397 return ConstantVector::get(Ops); 398 } 399 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 400 // Insertion of scalar constant into vector constant 401 std::vector<Constant*> Ops; 402 Ops.reserve(CVal->getNumOperands()); 403 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) { 404 const Constant *Op = 405 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i)); 406 Ops.push_back(const_cast<Constant*>(Op)); 407 } 408 return ConstantVector::get(Ops); 409 } 410 411 return 0; 412} 413 414/// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef 415/// return the specified element value. Otherwise return null. 416static Constant *GetVectorElement(LLVMContext &Context, const Constant *C, 417 unsigned EltNo) { 418 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) 419 return CV->getOperand(EltNo); 420 421 const Type *EltTy = cast<VectorType>(C->getType())->getElementType(); 422 if (isa<ConstantAggregateZero>(C)) 423 return Constant::getNullValue(EltTy); 424 if (isa<UndefValue>(C)) 425 return UndefValue::get(EltTy); 426 return 0; 427} 428 429Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, 430 const Constant *V1, 431 const Constant *V2, 432 const Constant *Mask) { 433 // Undefined shuffle mask -> undefined value. 434 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType()); 435 436 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements(); 437 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements(); 438 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType(); 439 440 // Loop over the shuffle mask, evaluating each element. 441 SmallVector<Constant*, 32> Result; 442 for (unsigned i = 0; i != MaskNumElts; ++i) { 443 Constant *InElt = GetVectorElement(Context, Mask, i); 444 if (InElt == 0) return 0; 445 446 if (isa<UndefValue>(InElt)) 447 InElt = UndefValue::get(EltTy); 448 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) { 449 unsigned Elt = CI->getZExtValue(); 450 if (Elt >= SrcNumElts*2) 451 InElt = UndefValue::get(EltTy); 452 else if (Elt >= SrcNumElts) 453 InElt = GetVectorElement(Context, V2, Elt - SrcNumElts); 454 else 455 InElt = GetVectorElement(Context, V1, Elt); 456 if (InElt == 0) return 0; 457 } else { 458 // Unknown value. 459 return 0; 460 } 461 Result.push_back(InElt); 462 } 463 464 return ConstantVector::get(&Result[0], Result.size()); 465} 466 467Constant *llvm::ConstantFoldExtractValueInstruction(LLVMContext &Context, 468 const Constant *Agg, 469 const unsigned *Idxs, 470 unsigned NumIdx) { 471 // Base case: no indices, so return the entire value. 472 if (NumIdx == 0) 473 return const_cast<Constant *>(Agg); 474 475 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef 476 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(), 477 Idxs, 478 Idxs + NumIdx)); 479 480 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0 481 return 482 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), 483 Idxs, 484 Idxs + NumIdx)); 485 486 // Otherwise recurse. 487 return ConstantFoldExtractValueInstruction(Context, Agg->getOperand(*Idxs), 488 Idxs+1, NumIdx-1); 489} 490 491Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, 492 const Constant *Agg, 493 const Constant *Val, 494 const unsigned *Idxs, 495 unsigned NumIdx) { 496 // Base case: no indices, so replace the entire value. 497 if (NumIdx == 0) 498 return const_cast<Constant *>(Val); 499 500 if (isa<UndefValue>(Agg)) { 501 // Insertion of constant into aggregate undef 502 // Optimize away insertion of undef 503 if (isa<UndefValue>(Val)) 504 return const_cast<Constant*>(Agg); 505 // Otherwise break the aggregate undef into multiple undefs and do 506 // the insertion 507 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 508 unsigned numOps; 509 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 510 numOps = AR->getNumElements(); 511 else 512 numOps = cast<StructType>(AggTy)->getNumElements(); 513 std::vector<Constant*> Ops(numOps); 514 for (unsigned i = 0; i < numOps; ++i) { 515 const Type *MemberTy = AggTy->getTypeAtIndex(i); 516 const Constant *Op = 517 (*Idxs == i) ? 518 ConstantFoldInsertValueInstruction(Context, UndefValue::get(MemberTy), 519 Val, Idxs+1, NumIdx-1) : 520 UndefValue::get(MemberTy); 521 Ops[i] = const_cast<Constant*>(Op); 522 } 523 if (isa<StructType>(AggTy)) 524 return ConstantStruct::get(Context, Ops); 525 else 526 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 527 } 528 if (isa<ConstantAggregateZero>(Agg)) { 529 // Insertion of constant into aggregate zero 530 // Optimize away insertion of zero 531 if (Val->isNullValue()) 532 return const_cast<Constant*>(Agg); 533 // Otherwise break the aggregate zero into multiple zeros and do 534 // the insertion 535 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 536 unsigned numOps; 537 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 538 numOps = AR->getNumElements(); 539 else 540 numOps = cast<StructType>(AggTy)->getNumElements(); 541 std::vector<Constant*> Ops(numOps); 542 for (unsigned i = 0; i < numOps; ++i) { 543 const Type *MemberTy = AggTy->getTypeAtIndex(i); 544 const Constant *Op = 545 (*Idxs == i) ? 546 ConstantFoldInsertValueInstruction(Context, 547 Constant::getNullValue(MemberTy), 548 Val, Idxs+1, NumIdx-1) : 549 Constant::getNullValue(MemberTy); 550 Ops[i] = const_cast<Constant*>(Op); 551 } 552 if (isa<StructType>(AggTy)) 553 return ConstantStruct::get(Context, Ops); 554 else 555 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 556 } 557 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) { 558 // Insertion of constant into aggregate constant 559 std::vector<Constant*> Ops(Agg->getNumOperands()); 560 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { 561 const Constant *Op = 562 (*Idxs == i) ? 563 ConstantFoldInsertValueInstruction(Context, Agg->getOperand(i), 564 Val, Idxs+1, NumIdx-1) : 565 Agg->getOperand(i); 566 Ops[i] = const_cast<Constant*>(Op); 567 } 568 Constant *C; 569 if (isa<StructType>(Agg->getType())) 570 C = ConstantStruct::get(Context, Ops); 571 else 572 C = ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops); 573 return C; 574 } 575 576 return 0; 577} 578 579 580Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, 581 unsigned Opcode, 582 const Constant *C1, 583 const Constant *C2) { 584 // No compile-time operations on this type yet. 585 if (C1->getType() == Type::getPPC_FP128Ty(Context)) 586 return 0; 587 588 // Handle UndefValue up front 589 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 590 switch (Opcode) { 591 case Instruction::Xor: 592 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 593 // Handle undef ^ undef -> 0 special case. This is a common 594 // idiom (misuse). 595 return Constant::getNullValue(C1->getType()); 596 // Fallthrough 597 case Instruction::Add: 598 case Instruction::Sub: 599 return UndefValue::get(C1->getType()); 600 case Instruction::Mul: 601 case Instruction::And: 602 return Constant::getNullValue(C1->getType()); 603 case Instruction::UDiv: 604 case Instruction::SDiv: 605 case Instruction::URem: 606 case Instruction::SRem: 607 if (!isa<UndefValue>(C2)) // undef / X -> 0 608 return Constant::getNullValue(C1->getType()); 609 return const_cast<Constant*>(C2); // X / undef -> undef 610 case Instruction::Or: // X | undef -> -1 611 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType())) 612 return Constant::getAllOnesValue(PTy); 613 return Constant::getAllOnesValue(C1->getType()); 614 case Instruction::LShr: 615 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 616 return const_cast<Constant*>(C1); // undef lshr undef -> undef 617 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 618 // undef lshr X -> 0 619 case Instruction::AShr: 620 if (!isa<UndefValue>(C2)) 621 return const_cast<Constant*>(C1); // undef ashr X --> undef 622 else if (isa<UndefValue>(C1)) 623 return const_cast<Constant*>(C1); // undef ashr undef -> undef 624 else 625 return const_cast<Constant*>(C1); // X ashr undef --> X 626 case Instruction::Shl: 627 // undef << X -> 0 or X << undef -> 0 628 return Constant::getNullValue(C1->getType()); 629 } 630 } 631 632 // Handle simplifications when the RHS is a constant int. 633 if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 634 switch (Opcode) { 635 case Instruction::Add: 636 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X + 0 == X 637 break; 638 case Instruction::Sub: 639 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X - 0 == X 640 break; 641 case Instruction::Mul: 642 if (CI2->equalsInt(0)) return const_cast<Constant*>(C2); // X * 0 == 0 643 if (CI2->equalsInt(1)) 644 return const_cast<Constant*>(C1); // X * 1 == X 645 break; 646 case Instruction::UDiv: 647 case Instruction::SDiv: 648 if (CI2->equalsInt(1)) 649 return const_cast<Constant*>(C1); // X / 1 == X 650 if (CI2->equalsInt(0)) 651 return UndefValue::get(CI2->getType()); // X / 0 == undef 652 break; 653 case Instruction::URem: 654 case Instruction::SRem: 655 if (CI2->equalsInt(1)) 656 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 657 if (CI2->equalsInt(0)) 658 return UndefValue::get(CI2->getType()); // X % 0 == undef 659 break; 660 case Instruction::And: 661 if (CI2->isZero()) return const_cast<Constant*>(C2); // X & 0 == 0 662 if (CI2->isAllOnesValue()) 663 return const_cast<Constant*>(C1); // X & -1 == X 664 665 if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 666 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 667 if (CE1->getOpcode() == Instruction::ZExt) { 668 unsigned DstWidth = CI2->getType()->getBitWidth(); 669 unsigned SrcWidth = 670 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 671 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 672 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 673 return const_cast<Constant*>(C1); 674 } 675 676 // If and'ing the address of a global with a constant, fold it. 677 if (CE1->getOpcode() == Instruction::PtrToInt && 678 isa<GlobalValue>(CE1->getOperand(0))) { 679 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 680 681 // Functions are at least 4-byte aligned. 682 unsigned GVAlign = GV->getAlignment(); 683 if (isa<Function>(GV)) 684 GVAlign = std::max(GVAlign, 4U); 685 686 if (GVAlign > 1) { 687 unsigned DstWidth = CI2->getType()->getBitWidth(); 688 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 689 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 690 691 // If checking bits we know are clear, return zero. 692 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 693 return Constant::getNullValue(CI2->getType()); 694 } 695 } 696 } 697 break; 698 case Instruction::Or: 699 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X | 0 == X 700 if (CI2->isAllOnesValue()) 701 return const_cast<Constant*>(C2); // X | -1 == -1 702 break; 703 case Instruction::Xor: 704 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X ^ 0 == X 705 break; 706 case Instruction::AShr: 707 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 708 if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 709 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 710 return ConstantExpr::getLShr(const_cast<Constant*>(C1), 711 const_cast<Constant*>(C2)); 712 break; 713 } 714 } 715 716 // At this point we know neither constant is an UndefValue. 717 if (const ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 718 if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 719 using namespace APIntOps; 720 const APInt &C1V = CI1->getValue(); 721 const APInt &C2V = CI2->getValue(); 722 switch (Opcode) { 723 default: 724 break; 725 case Instruction::Add: 726 return ConstantInt::get(Context, C1V + C2V); 727 case Instruction::Sub: 728 return ConstantInt::get(Context, C1V - C2V); 729 case Instruction::Mul: 730 return ConstantInt::get(Context, C1V * C2V); 731 case Instruction::UDiv: 732 assert(!CI2->isNullValue() && "Div by zero handled above"); 733 return ConstantInt::get(Context, C1V.udiv(C2V)); 734 case Instruction::SDiv: 735 assert(!CI2->isNullValue() && "Div by zero handled above"); 736 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 737 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 738 return ConstantInt::get(Context, C1V.sdiv(C2V)); 739 case Instruction::URem: 740 assert(!CI2->isNullValue() && "Div by zero handled above"); 741 return ConstantInt::get(Context, C1V.urem(C2V)); 742 case Instruction::SRem: 743 assert(!CI2->isNullValue() && "Div by zero handled above"); 744 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 745 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 746 return ConstantInt::get(Context, C1V.srem(C2V)); 747 case Instruction::And: 748 return ConstantInt::get(Context, C1V & C2V); 749 case Instruction::Or: 750 return ConstantInt::get(Context, C1V | C2V); 751 case Instruction::Xor: 752 return ConstantInt::get(Context, C1V ^ C2V); 753 case Instruction::Shl: { 754 uint32_t shiftAmt = C2V.getZExtValue(); 755 if (shiftAmt < C1V.getBitWidth()) 756 return ConstantInt::get(Context, C1V.shl(shiftAmt)); 757 else 758 return UndefValue::get(C1->getType()); // too big shift is undef 759 } 760 case Instruction::LShr: { 761 uint32_t shiftAmt = C2V.getZExtValue(); 762 if (shiftAmt < C1V.getBitWidth()) 763 return ConstantInt::get(Context, C1V.lshr(shiftAmt)); 764 else 765 return UndefValue::get(C1->getType()); // too big shift is undef 766 } 767 case Instruction::AShr: { 768 uint32_t shiftAmt = C2V.getZExtValue(); 769 if (shiftAmt < C1V.getBitWidth()) 770 return ConstantInt::get(Context, C1V.ashr(shiftAmt)); 771 else 772 return UndefValue::get(C1->getType()); // too big shift is undef 773 } 774 } 775 } 776 777 switch (Opcode) { 778 case Instruction::SDiv: 779 case Instruction::UDiv: 780 case Instruction::URem: 781 case Instruction::SRem: 782 case Instruction::LShr: 783 case Instruction::AShr: 784 case Instruction::Shl: 785 if (CI1->equalsInt(0)) return const_cast<Constant*>(C1); 786 break; 787 default: 788 break; 789 } 790 } else if (const ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 791 if (const ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 792 APFloat C1V = CFP1->getValueAPF(); 793 APFloat C2V = CFP2->getValueAPF(); 794 APFloat C3V = C1V; // copy for modification 795 switch (Opcode) { 796 default: 797 break; 798 case Instruction::FAdd: 799 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 800 return ConstantFP::get(Context, C3V); 801 case Instruction::FSub: 802 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 803 return ConstantFP::get(Context, C3V); 804 case Instruction::FMul: 805 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 806 return ConstantFP::get(Context, C3V); 807 case Instruction::FDiv: 808 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 809 return ConstantFP::get(Context, C3V); 810 case Instruction::FRem: 811 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 812 return ConstantFP::get(Context, C3V); 813 } 814 } 815 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 816 const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1); 817 const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2); 818 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) && 819 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) { 820 std::vector<Constant*> Res; 821 const Type* EltTy = VTy->getElementType(); 822 const Constant *C1 = 0; 823 const Constant *C2 = 0; 824 switch (Opcode) { 825 default: 826 break; 827 case Instruction::Add: 828 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 829 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 830 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 831 Res.push_back(ConstantExpr::getAdd(const_cast<Constant*>(C1), 832 const_cast<Constant*>(C2))); 833 } 834 return ConstantVector::get(Res); 835 case Instruction::FAdd: 836 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 837 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 838 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 839 Res.push_back(ConstantExpr::getFAdd(const_cast<Constant*>(C1), 840 const_cast<Constant*>(C2))); 841 } 842 return ConstantVector::get(Res); 843 case Instruction::Sub: 844 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 845 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 846 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 847 Res.push_back(ConstantExpr::getSub(const_cast<Constant*>(C1), 848 const_cast<Constant*>(C2))); 849 } 850 return ConstantVector::get(Res); 851 case Instruction::FSub: 852 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 853 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 854 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 855 Res.push_back(ConstantExpr::getFSub(const_cast<Constant*>(C1), 856 const_cast<Constant*>(C2))); 857 } 858 return ConstantVector::get(Res); 859 case Instruction::Mul: 860 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 861 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 862 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 863 Res.push_back(ConstantExpr::getMul(const_cast<Constant*>(C1), 864 const_cast<Constant*>(C2))); 865 } 866 return ConstantVector::get(Res); 867 case Instruction::FMul: 868 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 869 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 870 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 871 Res.push_back(ConstantExpr::getFMul(const_cast<Constant*>(C1), 872 const_cast<Constant*>(C2))); 873 } 874 return ConstantVector::get(Res); 875 case Instruction::UDiv: 876 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 877 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 878 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 879 Res.push_back(ConstantExpr::getUDiv(const_cast<Constant*>(C1), 880 const_cast<Constant*>(C2))); 881 } 882 return ConstantVector::get(Res); 883 case Instruction::SDiv: 884 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 885 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 886 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 887 Res.push_back(ConstantExpr::getSDiv(const_cast<Constant*>(C1), 888 const_cast<Constant*>(C2))); 889 } 890 return ConstantVector::get(Res); 891 case Instruction::FDiv: 892 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 893 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 894 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 895 Res.push_back(ConstantExpr::getFDiv(const_cast<Constant*>(C1), 896 const_cast<Constant*>(C2))); 897 } 898 return ConstantVector::get(Res); 899 case Instruction::URem: 900 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 901 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 902 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 903 Res.push_back(ConstantExpr::getURem(const_cast<Constant*>(C1), 904 const_cast<Constant*>(C2))); 905 } 906 return ConstantVector::get(Res); 907 case Instruction::SRem: 908 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 909 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 910 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 911 Res.push_back(ConstantExpr::getSRem(const_cast<Constant*>(C1), 912 const_cast<Constant*>(C2))); 913 } 914 return ConstantVector::get(Res); 915 case Instruction::FRem: 916 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 917 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 918 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 919 Res.push_back(ConstantExpr::getFRem(const_cast<Constant*>(C1), 920 const_cast<Constant*>(C2))); 921 } 922 return ConstantVector::get(Res); 923 case Instruction::And: 924 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 925 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 926 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 927 Res.push_back(ConstantExpr::getAnd(const_cast<Constant*>(C1), 928 const_cast<Constant*>(C2))); 929 } 930 return ConstantVector::get(Res); 931 case Instruction::Or: 932 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 933 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 934 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 935 Res.push_back(ConstantExpr::getOr(const_cast<Constant*>(C1), 936 const_cast<Constant*>(C2))); 937 } 938 return ConstantVector::get(Res); 939 case Instruction::Xor: 940 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 941 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 942 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 943 Res.push_back(ConstantExpr::getXor(const_cast<Constant*>(C1), 944 const_cast<Constant*>(C2))); 945 } 946 return ConstantVector::get(Res); 947 case Instruction::LShr: 948 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 949 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 950 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 951 Res.push_back(ConstantExpr::getLShr(const_cast<Constant*>(C1), 952 const_cast<Constant*>(C2))); 953 } 954 return ConstantVector::get(Res); 955 case Instruction::AShr: 956 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 957 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 958 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 959 Res.push_back(ConstantExpr::getAShr(const_cast<Constant*>(C1), 960 const_cast<Constant*>(C2))); 961 } 962 return ConstantVector::get(Res); 963 case Instruction::Shl: 964 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 965 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 966 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 967 Res.push_back(ConstantExpr::getShl(const_cast<Constant*>(C1), 968 const_cast<Constant*>(C2))); 969 } 970 return ConstantVector::get(Res); 971 } 972 } 973 } 974 975 if (isa<ConstantExpr>(C1)) { 976 // There are many possible foldings we could do here. We should probably 977 // at least fold add of a pointer with an integer into the appropriate 978 // getelementptr. This will improve alias analysis a bit. 979 } else if (isa<ConstantExpr>(C2)) { 980 // If C2 is a constant expr and C1 isn't, flop them around and fold the 981 // other way if possible. 982 switch (Opcode) { 983 case Instruction::Add: 984 case Instruction::FAdd: 985 case Instruction::Mul: 986 case Instruction::FMul: 987 case Instruction::And: 988 case Instruction::Or: 989 case Instruction::Xor: 990 // No change of opcode required. 991 return ConstantFoldBinaryInstruction(Context, Opcode, C2, C1); 992 993 case Instruction::Shl: 994 case Instruction::LShr: 995 case Instruction::AShr: 996 case Instruction::Sub: 997 case Instruction::FSub: 998 case Instruction::SDiv: 999 case Instruction::UDiv: 1000 case Instruction::FDiv: 1001 case Instruction::URem: 1002 case Instruction::SRem: 1003 case Instruction::FRem: 1004 default: // These instructions cannot be flopped around. 1005 break; 1006 } 1007 } 1008 1009 // We don't know how to fold this. 1010 return 0; 1011} 1012 1013/// isZeroSizedType - This type is zero sized if its an array or structure of 1014/// zero sized types. The only leaf zero sized type is an empty structure. 1015static bool isMaybeZeroSizedType(const Type *Ty) { 1016 if (isa<OpaqueType>(Ty)) return true; // Can't say. 1017 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1018 1019 // If all of elements have zero size, this does too. 1020 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1021 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1022 return true; 1023 1024 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1025 return isMaybeZeroSizedType(ATy->getElementType()); 1026 } 1027 return false; 1028} 1029 1030/// IdxCompare - Compare the two constants as though they were getelementptr 1031/// indices. This allows coersion of the types to be the same thing. 1032/// 1033/// If the two constants are the "same" (after coersion), return 0. If the 1034/// first is less than the second, return -1, if the second is less than the 1035/// first, return 1. If the constants are not integral, return -2. 1036/// 1037static int IdxCompare(LLVMContext &Context, Constant *C1, Constant *C2, 1038 const Type *ElTy) { 1039 if (C1 == C2) return 0; 1040 1041 // Ok, we found a different index. If they are not ConstantInt, we can't do 1042 // anything with them. 1043 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1044 return -2; // don't know! 1045 1046 // Ok, we have two differing integer indices. Sign extend them to be the same 1047 // type. Long is always big enough, so we use it. 1048 if (C1->getType() != Type::getInt64Ty(Context)) 1049 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(Context)); 1050 1051 if (C2->getType() != Type::getInt64Ty(Context)) 1052 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(Context)); 1053 1054 if (C1 == C2) return 0; // They are equal 1055 1056 // If the type being indexed over is really just a zero sized type, there is 1057 // no pointer difference being made here. 1058 if (isMaybeZeroSizedType(ElTy)) 1059 return -2; // dunno. 1060 1061 // If they are really different, now that they are the same type, then we 1062 // found a difference! 1063 if (cast<ConstantInt>(C1)->getSExtValue() < 1064 cast<ConstantInt>(C2)->getSExtValue()) 1065 return -1; 1066 else 1067 return 1; 1068} 1069 1070/// evaluateFCmpRelation - This function determines if there is anything we can 1071/// decide about the two constants provided. This doesn't need to handle simple 1072/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1073/// If we can determine that the two constants have a particular relation to 1074/// each other, we should return the corresponding FCmpInst predicate, 1075/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1076/// ConstantFoldCompareInstruction. 1077/// 1078/// To simplify this code we canonicalize the relation so that the first 1079/// operand is always the most "complex" of the two. We consider ConstantFP 1080/// to be the simplest, and ConstantExprs to be the most complex. 1081static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, 1082 const Constant *V1, 1083 const Constant *V2) { 1084 assert(V1->getType() == V2->getType() && 1085 "Cannot compare values of different types!"); 1086 1087 // No compile-time operations on this type yet. 1088 if (V1->getType() == Type::getPPC_FP128Ty(Context)) 1089 return FCmpInst::BAD_FCMP_PREDICATE; 1090 1091 // Handle degenerate case quickly 1092 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1093 1094 if (!isa<ConstantExpr>(V1)) { 1095 if (!isa<ConstantExpr>(V2)) { 1096 // We distilled thisUse the standard constant folder for a few cases 1097 ConstantInt *R = 0; 1098 Constant *C1 = const_cast<Constant*>(V1); 1099 Constant *C2 = const_cast<Constant*>(V2); 1100 R = dyn_cast<ConstantInt>( 1101 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, C1, C2)); 1102 if (R && !R->isZero()) 1103 return FCmpInst::FCMP_OEQ; 1104 R = dyn_cast<ConstantInt>( 1105 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, C1, C2)); 1106 if (R && !R->isZero()) 1107 return FCmpInst::FCMP_OLT; 1108 R = dyn_cast<ConstantInt>( 1109 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, C1, C2)); 1110 if (R && !R->isZero()) 1111 return FCmpInst::FCMP_OGT; 1112 1113 // Nothing more we can do 1114 return FCmpInst::BAD_FCMP_PREDICATE; 1115 } 1116 1117 // If the first operand is simple and second is ConstantExpr, swap operands. 1118 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(Context, V2, V1); 1119 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1120 return FCmpInst::getSwappedPredicate(SwappedRelation); 1121 } else { 1122 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1123 // constantexpr or a simple constant. 1124 const ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1125 switch (CE1->getOpcode()) { 1126 case Instruction::FPTrunc: 1127 case Instruction::FPExt: 1128 case Instruction::UIToFP: 1129 case Instruction::SIToFP: 1130 // We might be able to do something with these but we don't right now. 1131 break; 1132 default: 1133 break; 1134 } 1135 } 1136 // There are MANY other foldings that we could perform here. They will 1137 // probably be added on demand, as they seem needed. 1138 return FCmpInst::BAD_FCMP_PREDICATE; 1139} 1140 1141/// evaluateICmpRelation - This function determines if there is anything we can 1142/// decide about the two constants provided. This doesn't need to handle simple 1143/// things like integer comparisons, but should instead handle ConstantExprs 1144/// and GlobalValues. If we can determine that the two constants have a 1145/// particular relation to each other, we should return the corresponding ICmp 1146/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1147/// 1148/// To simplify this code we canonicalize the relation so that the first 1149/// operand is always the most "complex" of the two. We consider simple 1150/// constants (like ConstantInt) to be the simplest, followed by 1151/// GlobalValues, followed by ConstantExpr's (the most complex). 1152/// 1153static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, 1154 const Constant *V1, 1155 const Constant *V2, 1156 bool isSigned) { 1157 assert(V1->getType() == V2->getType() && 1158 "Cannot compare different types of values!"); 1159 if (V1 == V2) return ICmpInst::ICMP_EQ; 1160 1161 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1)) { 1162 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2)) { 1163 // We distilled this down to a simple case, use the standard constant 1164 // folder. 1165 ConstantInt *R = 0; 1166 Constant *C1 = const_cast<Constant*>(V1); 1167 Constant *C2 = const_cast<Constant*>(V2); 1168 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1169 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2)); 1170 if (R && !R->isZero()) 1171 return pred; 1172 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1173 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2)); 1174 if (R && !R->isZero()) 1175 return pred; 1176 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1177 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2)); 1178 if (R && !R->isZero()) 1179 return pred; 1180 1181 // If we couldn't figure it out, bail. 1182 return ICmpInst::BAD_ICMP_PREDICATE; 1183 } 1184 1185 // If the first operand is simple, swap operands. 1186 ICmpInst::Predicate SwappedRelation = 1187 evaluateICmpRelation(Context, V2, V1, isSigned); 1188 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1189 return ICmpInst::getSwappedPredicate(SwappedRelation); 1190 1191 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(V1)) { 1192 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1193 ICmpInst::Predicate SwappedRelation = 1194 evaluateICmpRelation(Context, V2, V1, isSigned); 1195 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1196 return ICmpInst::getSwappedPredicate(SwappedRelation); 1197 else 1198 return ICmpInst::BAD_ICMP_PREDICATE; 1199 } 1200 1201 // Now we know that the RHS is a GlobalValue or simple constant, 1202 // which (since the types must match) means that it's a ConstantPointerNull. 1203 if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) { 1204 // Don't try to decide equality of aliases. 1205 if (!isa<GlobalAlias>(CPR1) && !isa<GlobalAlias>(CPR2)) 1206 if (!CPR1->hasExternalWeakLinkage() || !CPR2->hasExternalWeakLinkage()) 1207 return ICmpInst::ICMP_NE; 1208 } else { 1209 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1210 // GlobalVals can never be null. Don't try to evaluate aliases. 1211 if (!CPR1->hasExternalWeakLinkage() && !isa<GlobalAlias>(CPR1)) 1212 return ICmpInst::ICMP_NE; 1213 } 1214 } else { 1215 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1216 // constantexpr, a CPR, or a simple constant. 1217 const ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1218 const Constant *CE1Op0 = CE1->getOperand(0); 1219 1220 switch (CE1->getOpcode()) { 1221 case Instruction::Trunc: 1222 case Instruction::FPTrunc: 1223 case Instruction::FPExt: 1224 case Instruction::FPToUI: 1225 case Instruction::FPToSI: 1226 break; // We can't evaluate floating point casts or truncations. 1227 1228 case Instruction::UIToFP: 1229 case Instruction::SIToFP: 1230 case Instruction::BitCast: 1231 case Instruction::ZExt: 1232 case Instruction::SExt: 1233 // If the cast is not actually changing bits, and the second operand is a 1234 // null pointer, do the comparison with the pre-casted value. 1235 if (V2->isNullValue() && 1236 (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) { 1237 bool sgnd = isSigned; 1238 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1239 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1240 return evaluateICmpRelation(Context, CE1Op0, 1241 Constant::getNullValue(CE1Op0->getType()), 1242 sgnd); 1243 } 1244 1245 // If the dest type is a pointer type, and the RHS is a constantexpr cast 1246 // from the same type as the src of the LHS, evaluate the inputs. This is 1247 // important for things like "icmp eq (cast 4 to int*), (cast 5 to int*)", 1248 // which happens a lot in compilers with tagged integers. 1249 if (const ConstantExpr *CE2 = dyn_cast<ConstantExpr>(V2)) 1250 if (CE2->isCast() && isa<PointerType>(CE1->getType()) && 1251 CE1->getOperand(0)->getType() == CE2->getOperand(0)->getType() && 1252 CE1->getOperand(0)->getType()->isInteger()) { 1253 bool sgnd = isSigned; 1254 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1255 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1256 return evaluateICmpRelation(Context, CE1->getOperand(0), 1257 CE2->getOperand(0), sgnd); 1258 } 1259 break; 1260 1261 case Instruction::GetElementPtr: 1262 // Ok, since this is a getelementptr, we know that the constant has a 1263 // pointer type. Check the various cases. 1264 if (isa<ConstantPointerNull>(V2)) { 1265 // If we are comparing a GEP to a null pointer, check to see if the base 1266 // of the GEP equals the null pointer. 1267 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1268 if (GV->hasExternalWeakLinkage()) 1269 // Weak linkage GVals could be zero or not. We're comparing that 1270 // to null pointer so its greater-or-equal 1271 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1272 else 1273 // If its not weak linkage, the GVal must have a non-zero address 1274 // so the result is greater-than 1275 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1276 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1277 // If we are indexing from a null pointer, check to see if we have any 1278 // non-zero indices. 1279 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1280 if (!CE1->getOperand(i)->isNullValue()) 1281 // Offsetting from null, must not be equal. 1282 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1283 // Only zero indexes from null, must still be zero. 1284 return ICmpInst::ICMP_EQ; 1285 } 1286 // Otherwise, we can't really say if the first operand is null or not. 1287 } else if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) { 1288 if (isa<ConstantPointerNull>(CE1Op0)) { 1289 if (CPR2->hasExternalWeakLinkage()) 1290 // Weak linkage GVals could be zero or not. We're comparing it to 1291 // a null pointer, so its less-or-equal 1292 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1293 else 1294 // If its not weak linkage, the GVal must have a non-zero address 1295 // so the result is less-than 1296 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1297 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(CE1Op0)) { 1298 if (CPR1 == CPR2) { 1299 // If this is a getelementptr of the same global, then it must be 1300 // different. Because the types must match, the getelementptr could 1301 // only have at most one index, and because we fold getelementptr's 1302 // with a single zero index, it must be nonzero. 1303 assert(CE1->getNumOperands() == 2 && 1304 !CE1->getOperand(1)->isNullValue() && 1305 "Suprising getelementptr!"); 1306 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1307 } else { 1308 // If they are different globals, we don't know what the value is, 1309 // but they can't be equal. 1310 return ICmpInst::ICMP_NE; 1311 } 1312 } 1313 } else { 1314 const ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1315 const Constant *CE2Op0 = CE2->getOperand(0); 1316 1317 // There are MANY other foldings that we could perform here. They will 1318 // probably be added on demand, as they seem needed. 1319 switch (CE2->getOpcode()) { 1320 default: break; 1321 case Instruction::GetElementPtr: 1322 // By far the most common case to handle is when the base pointers are 1323 // obviously to the same or different globals. 1324 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1325 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal 1326 return ICmpInst::ICMP_NE; 1327 // Ok, we know that both getelementptr instructions are based on the 1328 // same global. From this, we can precisely determine the relative 1329 // ordering of the resultant pointers. 1330 unsigned i = 1; 1331 1332 // The logic below assumes that the result of the comparison 1333 // can be determined by finding the first index that differs. 1334 // This doesn't work if there is over-indexing in any 1335 // subsequent indices, so check for that case first. 1336 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1337 !CE2->isGEPWithNoNotionalOverIndexing()) 1338 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1339 1340 // Compare all of the operands the GEP's have in common. 1341 gep_type_iterator GTI = gep_type_begin(CE1); 1342 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1343 ++i, ++GTI) 1344 switch (IdxCompare(Context, CE1->getOperand(i), 1345 CE2->getOperand(i), GTI.getIndexedType())) { 1346 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1347 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1348 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1349 } 1350 1351 // Ok, we ran out of things they have in common. If any leftovers 1352 // are non-zero then we have a difference, otherwise we are equal. 1353 for (; i < CE1->getNumOperands(); ++i) 1354 if (!CE1->getOperand(i)->isNullValue()) { 1355 if (isa<ConstantInt>(CE1->getOperand(i))) 1356 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1357 else 1358 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1359 } 1360 1361 for (; i < CE2->getNumOperands(); ++i) 1362 if (!CE2->getOperand(i)->isNullValue()) { 1363 if (isa<ConstantInt>(CE2->getOperand(i))) 1364 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1365 else 1366 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1367 } 1368 return ICmpInst::ICMP_EQ; 1369 } 1370 } 1371 } 1372 default: 1373 break; 1374 } 1375 } 1376 1377 return ICmpInst::BAD_ICMP_PREDICATE; 1378} 1379 1380Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, 1381 unsigned short pred, 1382 const Constant *C1, 1383 const Constant *C2) { 1384 const Type *ResultTy; 1385 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1386 ResultTy = VectorType::get(Type::getInt1Ty(Context), VT->getNumElements()); 1387 else 1388 ResultTy = Type::getInt1Ty(Context); 1389 1390 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1391 if (pred == FCmpInst::FCMP_FALSE) 1392 return Constant::getNullValue(ResultTy); 1393 1394 if (pred == FCmpInst::FCMP_TRUE) 1395 return Constant::getAllOnesValue(ResultTy); 1396 1397 // Handle some degenerate cases first 1398 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) 1399 return UndefValue::get(ResultTy); 1400 1401 // No compile-time operations on this type yet. 1402 if (C1->getType() == Type::getPPC_FP128Ty(Context)) 1403 return 0; 1404 1405 // icmp eq/ne(null,GV) -> false/true 1406 if (C1->isNullValue()) { 1407 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1408 // Don't try to evaluate aliases. External weak GV can be null. 1409 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1410 if (pred == ICmpInst::ICMP_EQ) 1411 return ConstantInt::getFalse(Context); 1412 else if (pred == ICmpInst::ICMP_NE) 1413 return ConstantInt::getTrue(Context); 1414 } 1415 // icmp eq/ne(GV,null) -> false/true 1416 } else if (C2->isNullValue()) { 1417 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1418 // Don't try to evaluate aliases. External weak GV can be null. 1419 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1420 if (pred == ICmpInst::ICMP_EQ) 1421 return ConstantInt::getFalse(Context); 1422 else if (pred == ICmpInst::ICMP_NE) 1423 return ConstantInt::getTrue(Context); 1424 } 1425 } 1426 1427 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1428 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1429 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1430 switch (pred) { 1431 default: llvm_unreachable("Invalid ICmp Predicate"); return 0; 1432 case ICmpInst::ICMP_EQ: 1433 return ConstantInt::get(Type::getInt1Ty(Context), V1 == V2); 1434 case ICmpInst::ICMP_NE: 1435 return ConstantInt::get(Type::getInt1Ty(Context), V1 != V2); 1436 case ICmpInst::ICMP_SLT: 1437 return ConstantInt::get(Type::getInt1Ty(Context), V1.slt(V2)); 1438 case ICmpInst::ICMP_SGT: 1439 return ConstantInt::get(Type::getInt1Ty(Context), V1.sgt(V2)); 1440 case ICmpInst::ICMP_SLE: 1441 return ConstantInt::get(Type::getInt1Ty(Context), V1.sle(V2)); 1442 case ICmpInst::ICMP_SGE: 1443 return ConstantInt::get(Type::getInt1Ty(Context), V1.sge(V2)); 1444 case ICmpInst::ICMP_ULT: 1445 return ConstantInt::get(Type::getInt1Ty(Context), V1.ult(V2)); 1446 case ICmpInst::ICMP_UGT: 1447 return ConstantInt::get(Type::getInt1Ty(Context), V1.ugt(V2)); 1448 case ICmpInst::ICMP_ULE: 1449 return ConstantInt::get(Type::getInt1Ty(Context), V1.ule(V2)); 1450 case ICmpInst::ICMP_UGE: 1451 return ConstantInt::get(Type::getInt1Ty(Context), V1.uge(V2)); 1452 } 1453 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1454 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1455 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1456 APFloat::cmpResult R = C1V.compare(C2V); 1457 switch (pred) { 1458 default: llvm_unreachable("Invalid FCmp Predicate"); return 0; 1459 case FCmpInst::FCMP_FALSE: return ConstantInt::getFalse(Context); 1460 case FCmpInst::FCMP_TRUE: return ConstantInt::getTrue(Context); 1461 case FCmpInst::FCMP_UNO: 1462 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered); 1463 case FCmpInst::FCMP_ORD: 1464 return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpUnordered); 1465 case FCmpInst::FCMP_UEQ: 1466 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || 1467 R==APFloat::cmpEqual); 1468 case FCmpInst::FCMP_OEQ: 1469 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpEqual); 1470 case FCmpInst::FCMP_UNE: 1471 return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpEqual); 1472 case FCmpInst::FCMP_ONE: 1473 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan || 1474 R==APFloat::cmpGreaterThan); 1475 case FCmpInst::FCMP_ULT: 1476 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || 1477 R==APFloat::cmpLessThan); 1478 case FCmpInst::FCMP_OLT: 1479 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan); 1480 case FCmpInst::FCMP_UGT: 1481 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || 1482 R==APFloat::cmpGreaterThan); 1483 case FCmpInst::FCMP_OGT: 1484 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpGreaterThan); 1485 case FCmpInst::FCMP_ULE: 1486 return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpGreaterThan); 1487 case FCmpInst::FCMP_OLE: 1488 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan || 1489 R==APFloat::cmpEqual); 1490 case FCmpInst::FCMP_UGE: 1491 return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpLessThan); 1492 case FCmpInst::FCMP_OGE: 1493 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpGreaterThan || 1494 R==APFloat::cmpEqual); 1495 } 1496 } else if (isa<VectorType>(C1->getType())) { 1497 SmallVector<Constant*, 16> C1Elts, C2Elts; 1498 C1->getVectorElements(Context, C1Elts); 1499 C2->getVectorElements(Context, C2Elts); 1500 1501 // If we can constant fold the comparison of each element, constant fold 1502 // the whole vector comparison. 1503 SmallVector<Constant*, 4> ResElts; 1504 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) { 1505 // Compare the elements, producing an i1 result or constant expr. 1506 ResElts.push_back( 1507 ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); 1508 } 1509 return ConstantVector::get(&ResElts[0], ResElts.size()); 1510 } 1511 1512 if (C1->getType()->isFloatingPoint()) { 1513 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1514 switch (evaluateFCmpRelation(Context, C1, C2)) { 1515 default: llvm_unreachable("Unknown relation!"); 1516 case FCmpInst::FCMP_UNO: 1517 case FCmpInst::FCMP_ORD: 1518 case FCmpInst::FCMP_UEQ: 1519 case FCmpInst::FCMP_UNE: 1520 case FCmpInst::FCMP_ULT: 1521 case FCmpInst::FCMP_UGT: 1522 case FCmpInst::FCMP_ULE: 1523 case FCmpInst::FCMP_UGE: 1524 case FCmpInst::FCMP_TRUE: 1525 case FCmpInst::FCMP_FALSE: 1526 case FCmpInst::BAD_FCMP_PREDICATE: 1527 break; // Couldn't determine anything about these constants. 1528 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1529 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1530 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1531 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1532 break; 1533 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1534 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1535 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1536 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1537 break; 1538 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1539 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1540 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1541 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1542 break; 1543 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1544 // We can only partially decide this relation. 1545 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1546 Result = 0; 1547 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1548 Result = 1; 1549 break; 1550 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1551 // We can only partially decide this relation. 1552 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1553 Result = 0; 1554 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1555 Result = 1; 1556 break; 1557 case ICmpInst::ICMP_NE: // We know that C1 != C2 1558 // We can only partially decide this relation. 1559 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1560 Result = 0; 1561 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1562 Result = 1; 1563 break; 1564 } 1565 1566 // If we evaluated the result, return it now. 1567 if (Result != -1) 1568 return ConstantInt::get(Type::getInt1Ty(Context), Result); 1569 1570 } else { 1571 // Evaluate the relation between the two constants, per the predicate. 1572 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1573 switch (evaluateICmpRelation(Context, C1, C2, CmpInst::isSigned(pred))) { 1574 default: llvm_unreachable("Unknown relational!"); 1575 case ICmpInst::BAD_ICMP_PREDICATE: 1576 break; // Couldn't determine anything about these constants. 1577 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1578 // If we know the constants are equal, we can decide the result of this 1579 // computation precisely. 1580 Result = (pred == ICmpInst::ICMP_EQ || 1581 pred == ICmpInst::ICMP_ULE || 1582 pred == ICmpInst::ICMP_SLE || 1583 pred == ICmpInst::ICMP_UGE || 1584 pred == ICmpInst::ICMP_SGE); 1585 break; 1586 case ICmpInst::ICMP_ULT: 1587 // If we know that C1 < C2, we can decide the result of this computation 1588 // precisely. 1589 Result = (pred == ICmpInst::ICMP_ULT || 1590 pred == ICmpInst::ICMP_NE || 1591 pred == ICmpInst::ICMP_ULE); 1592 break; 1593 case ICmpInst::ICMP_SLT: 1594 // If we know that C1 < C2, we can decide the result of this computation 1595 // precisely. 1596 Result = (pred == ICmpInst::ICMP_SLT || 1597 pred == ICmpInst::ICMP_NE || 1598 pred == ICmpInst::ICMP_SLE); 1599 break; 1600 case ICmpInst::ICMP_UGT: 1601 // If we know that C1 > C2, we can decide the result of this computation 1602 // precisely. 1603 Result = (pred == ICmpInst::ICMP_UGT || 1604 pred == ICmpInst::ICMP_NE || 1605 pred == ICmpInst::ICMP_UGE); 1606 break; 1607 case ICmpInst::ICMP_SGT: 1608 // If we know that C1 > C2, we can decide the result of this computation 1609 // precisely. 1610 Result = (pred == ICmpInst::ICMP_SGT || 1611 pred == ICmpInst::ICMP_NE || 1612 pred == ICmpInst::ICMP_SGE); 1613 break; 1614 case ICmpInst::ICMP_ULE: 1615 // If we know that C1 <= C2, we can only partially decide this relation. 1616 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1617 if (pred == ICmpInst::ICMP_ULT) Result = 1; 1618 break; 1619 case ICmpInst::ICMP_SLE: 1620 // If we know that C1 <= C2, we can only partially decide this relation. 1621 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1622 if (pred == ICmpInst::ICMP_SLT) Result = 1; 1623 break; 1624 1625 case ICmpInst::ICMP_UGE: 1626 // If we know that C1 >= C2, we can only partially decide this relation. 1627 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1628 if (pred == ICmpInst::ICMP_UGT) Result = 1; 1629 break; 1630 case ICmpInst::ICMP_SGE: 1631 // If we know that C1 >= C2, we can only partially decide this relation. 1632 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1633 if (pred == ICmpInst::ICMP_SGT) Result = 1; 1634 break; 1635 1636 case ICmpInst::ICMP_NE: 1637 // If we know that C1 != C2, we can only partially decide this relation. 1638 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1639 if (pred == ICmpInst::ICMP_NE) Result = 1; 1640 break; 1641 } 1642 1643 // If we evaluated the result, return it now. 1644 if (Result != -1) 1645 return ConstantInt::get(Type::getInt1Ty(Context), Result); 1646 1647 if (!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) { 1648 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1649 // other way if possible. 1650 switch (pred) { 1651 case ICmpInst::ICMP_EQ: 1652 case ICmpInst::ICMP_NE: 1653 // No change of predicate required. 1654 return ConstantFoldCompareInstruction(Context, pred, C2, C1); 1655 1656 case ICmpInst::ICMP_ULT: 1657 case ICmpInst::ICMP_SLT: 1658 case ICmpInst::ICMP_UGT: 1659 case ICmpInst::ICMP_SGT: 1660 case ICmpInst::ICMP_ULE: 1661 case ICmpInst::ICMP_SLE: 1662 case ICmpInst::ICMP_UGE: 1663 case ICmpInst::ICMP_SGE: 1664 // Change the predicate as necessary to swap the operands. 1665 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1666 return ConstantFoldCompareInstruction(Context, pred, C2, C1); 1667 1668 default: // These predicates cannot be flopped around. 1669 break; 1670 } 1671 } 1672 } 1673 return 0; 1674} 1675 1676/// isInBoundsIndices - Test whether the given sequence of *normalized* indices 1677/// is "inbounds". 1678static bool isInBoundsIndices(Constant *const *Idxs, size_t NumIdx) { 1679 // No indices means nothing that could be out of bounds. 1680 if (NumIdx == 0) return true; 1681 1682 // If the first index is zero, it's in bounds. 1683 if (Idxs[0]->isNullValue()) return true; 1684 1685 // If the first index is one and all the rest are zero, it's in bounds, 1686 // by the one-past-the-end rule. 1687 if (!cast<ConstantInt>(Idxs[0])->isOne()) 1688 return false; 1689 for (unsigned i = 1, e = NumIdx; i != e; ++i) 1690 if (!Idxs[i]->isNullValue()) 1691 return false; 1692 return true; 1693} 1694 1695Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, 1696 const Constant *C, 1697 bool inBounds, 1698 Constant* const *Idxs, 1699 unsigned NumIdx) { 1700 if (NumIdx == 0 || 1701 (NumIdx == 1 && Idxs[0]->isNullValue())) 1702 return const_cast<Constant*>(C); 1703 1704 if (isa<UndefValue>(C)) { 1705 const PointerType *Ptr = cast<PointerType>(C->getType()); 1706 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, 1707 (Value **)Idxs, 1708 (Value **)Idxs+NumIdx); 1709 assert(Ty != 0 && "Invalid indices for GEP!"); 1710 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 1711 } 1712 1713 Constant *Idx0 = Idxs[0]; 1714 if (C->isNullValue()) { 1715 bool isNull = true; 1716 for (unsigned i = 0, e = NumIdx; i != e; ++i) 1717 if (!Idxs[i]->isNullValue()) { 1718 isNull = false; 1719 break; 1720 } 1721 if (isNull) { 1722 const PointerType *Ptr = cast<PointerType>(C->getType()); 1723 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, 1724 (Value**)Idxs, 1725 (Value**)Idxs+NumIdx); 1726 assert(Ty != 0 && "Invalid indices for GEP!"); 1727 return ConstantPointerNull::get( 1728 PointerType::get(Ty,Ptr->getAddressSpace())); 1729 } 1730 } 1731 1732 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(const_cast<Constant*>(C))) { 1733 // Combine Indices - If the source pointer to this getelementptr instruction 1734 // is a getelementptr instruction, combine the indices of the two 1735 // getelementptr instructions into a single instruction. 1736 // 1737 if (CE->getOpcode() == Instruction::GetElementPtr) { 1738 const Type *LastTy = 0; 1739 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 1740 I != E; ++I) 1741 LastTy = *I; 1742 1743 if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) { 1744 SmallVector<Value*, 16> NewIndices; 1745 NewIndices.reserve(NumIdx + CE->getNumOperands()); 1746 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 1747 NewIndices.push_back(CE->getOperand(i)); 1748 1749 // Add the last index of the source with the first index of the new GEP. 1750 // Make sure to handle the case when they are actually different types. 1751 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 1752 // Otherwise it must be an array. 1753 if (!Idx0->isNullValue()) { 1754 const Type *IdxTy = Combined->getType(); 1755 if (IdxTy != Idx0->getType()) { 1756 Constant *C1 = 1757 ConstantExpr::getSExtOrBitCast(Idx0, Type::getInt64Ty(Context)); 1758 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, 1759 Type::getInt64Ty(Context)); 1760 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 1761 } else { 1762 Combined = 1763 ConstantExpr::get(Instruction::Add, Idx0, Combined); 1764 } 1765 } 1766 1767 NewIndices.push_back(Combined); 1768 NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx); 1769 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ? 1770 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0), 1771 &NewIndices[0], 1772 NewIndices.size()) : 1773 ConstantExpr::getGetElementPtr(CE->getOperand(0), 1774 &NewIndices[0], 1775 NewIndices.size()); 1776 } 1777 } 1778 1779 // Implement folding of: 1780 // int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*), 1781 // long 0, long 0) 1782 // To: int* getelementptr ([3 x int]* %X, long 0, long 0) 1783 // 1784 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) { 1785 if (const PointerType *SPT = 1786 dyn_cast<PointerType>(CE->getOperand(0)->getType())) 1787 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType())) 1788 if (const ArrayType *CAT = 1789 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType())) 1790 if (CAT->getElementType() == SAT->getElementType()) 1791 return inBounds ? 1792 ConstantExpr::getInBoundsGetElementPtr( 1793 (Constant*)CE->getOperand(0), Idxs, NumIdx) : 1794 ConstantExpr::getGetElementPtr( 1795 (Constant*)CE->getOperand(0), Idxs, NumIdx); 1796 } 1797 1798 // Fold: getelementptr (i8* inttoptr (i64 1 to i8*), i32 -1) 1799 // Into: inttoptr (i64 0 to i8*) 1800 // This happens with pointers to member functions in C++. 1801 if (CE->getOpcode() == Instruction::IntToPtr && NumIdx == 1 && 1802 isa<ConstantInt>(CE->getOperand(0)) && isa<ConstantInt>(Idxs[0]) && 1803 cast<PointerType>(CE->getType())->getElementType() == Type::getInt8Ty(Context)) { 1804 Constant *Base = CE->getOperand(0); 1805 Constant *Offset = Idxs[0]; 1806 1807 // Convert the smaller integer to the larger type. 1808 if (Offset->getType()->getPrimitiveSizeInBits() < 1809 Base->getType()->getPrimitiveSizeInBits()) 1810 Offset = ConstantExpr::getSExt(Offset, Base->getType()); 1811 else if (Base->getType()->getPrimitiveSizeInBits() < 1812 Offset->getType()->getPrimitiveSizeInBits()) 1813 Base = ConstantExpr::getZExt(Base, Offset->getType()); 1814 1815 Base = ConstantExpr::getAdd(Base, Offset); 1816 return ConstantExpr::getIntToPtr(Base, CE->getType()); 1817 } 1818 } 1819 1820 // Check to see if any array indices are not within the corresponding 1821 // notional array bounds. If so, try to determine if they can be factored 1822 // out into preceding dimensions. 1823 bool Unknown = false; 1824 SmallVector<Constant *, 8> NewIdxs; 1825 const Type *Ty = C->getType(); 1826 const Type *Prev = 0; 1827 for (unsigned i = 0; i != NumIdx; 1828 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 1829 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 1830 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 1831 if (ATy->getNumElements() <= INT64_MAX && 1832 ATy->getNumElements() != 0 && 1833 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) { 1834 if (isa<SequentialType>(Prev)) { 1835 // It's out of range, but we can factor it into the prior 1836 // dimension. 1837 NewIdxs.resize(NumIdx); 1838 ConstantInt *Factor = ConstantInt::get(CI->getType(), 1839 ATy->getNumElements()); 1840 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 1841 1842 Constant *PrevIdx = Idxs[i-1]; 1843 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 1844 1845 // Before adding, extend both operands to i64 to avoid 1846 // overflow trouble. 1847 if (PrevIdx->getType() != Type::getInt64Ty(Context)) 1848 PrevIdx = ConstantExpr::getSExt(PrevIdx, 1849 Type::getInt64Ty(Context)); 1850 if (Div->getType() != Type::getInt64Ty(Context)) 1851 Div = ConstantExpr::getSExt(Div, 1852 Type::getInt64Ty(Context)); 1853 1854 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 1855 } else { 1856 // It's out of range, but the prior dimension is a struct 1857 // so we can't do anything about it. 1858 Unknown = true; 1859 } 1860 } 1861 } else { 1862 // We don't know if it's in range or not. 1863 Unknown = true; 1864 } 1865 } 1866 1867 // If we did any factoring, start over with the adjusted indices. 1868 if (!NewIdxs.empty()) { 1869 for (unsigned i = 0; i != NumIdx; ++i) 1870 if (!NewIdxs[i]) NewIdxs[i] = Idxs[i]; 1871 return inBounds ? 1872 ConstantExpr::getGetElementPtr(const_cast<Constant*>(C), 1873 NewIdxs.data(), NewIdxs.size()) : 1874 ConstantExpr::getInBoundsGetElementPtr(const_cast<Constant*>(C), 1875 NewIdxs.data(), NewIdxs.size()); 1876 } 1877 1878 // If all indices are known integers and normalized, we can do a simple 1879 // check for the "inbounds" property. 1880 if (!Unknown && !inBounds && 1881 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx)) 1882 return ConstantExpr::getInBoundsGetElementPtr(const_cast<Constant*>(C), 1883 Idxs, NumIdx); 1884 1885 return 0; 1886} 1887