ConstantFold.cpp revision 3892baac42daadf760d140168720be1d6e585aa8
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 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 (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 (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, 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 (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 (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 (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 (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 (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 (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 (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 (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, 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 Constant *Cond, 324 Constant *V1, Constant *V2) { 325 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond)) 326 return CB->getZExtValue() ? V1 : V2; 327 328 if (isa<UndefValue>(V1)) return V2; 329 if (isa<UndefValue>(V2)) return V1; 330 if (isa<UndefValue>(Cond)) return V1; 331 if (V1 == V2) return V1; 332 return 0; 333} 334 335Constant *llvm::ConstantFoldExtractElementInstruction(LLVMContext &Context, 336 Constant *Val, 337 Constant *Idx) { 338 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 339 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType()); 340 if (Val->isNullValue()) // ee(zero, x) -> zero 341 return Constant::getNullValue( 342 cast<VectorType>(Val->getType())->getElementType()); 343 344 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 345 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 346 return CVal->getOperand(CIdx->getZExtValue()); 347 } else if (isa<UndefValue>(Idx)) { 348 // ee({w,x,y,z}, undef) -> w (an arbitrary value). 349 return CVal->getOperand(0); 350 } 351 } 352 return 0; 353} 354 355Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, 356 Constant *Val, 357 Constant *Elt, 358 Constant *Idx) { 359 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 360 if (!CIdx) return 0; 361 APInt idxVal = CIdx->getValue(); 362 if (isa<UndefValue>(Val)) { 363 // Insertion of scalar constant into vector undef 364 // Optimize away insertion of undef 365 if (isa<UndefValue>(Elt)) 366 return Val; 367 // Otherwise break the aggregate undef into multiple undefs and do 368 // the insertion 369 unsigned numOps = 370 cast<VectorType>(Val->getType())->getNumElements(); 371 std::vector<Constant*> Ops; 372 Ops.reserve(numOps); 373 for (unsigned i = 0; i < numOps; ++i) { 374 Constant *Op = 375 (idxVal == i) ? Elt : UndefValue::get(Elt->getType()); 376 Ops.push_back(Op); 377 } 378 return ConstantVector::get(Ops); 379 } 380 if (isa<ConstantAggregateZero>(Val)) { 381 // Insertion of scalar constant into vector aggregate zero 382 // Optimize away insertion of zero 383 if (Elt->isNullValue()) 384 return Val; 385 // Otherwise break the aggregate zero into multiple zeros and do 386 // the insertion 387 unsigned numOps = 388 cast<VectorType>(Val->getType())->getNumElements(); 389 std::vector<Constant*> Ops; 390 Ops.reserve(numOps); 391 for (unsigned i = 0; i < numOps; ++i) { 392 Constant *Op = 393 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType()); 394 Ops.push_back(Op); 395 } 396 return ConstantVector::get(Ops); 397 } 398 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 399 // Insertion of scalar constant into vector constant 400 std::vector<Constant*> Ops; 401 Ops.reserve(CVal->getNumOperands()); 402 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) { 403 Constant *Op = 404 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i)); 405 Ops.push_back(Op); 406 } 407 return ConstantVector::get(Ops); 408 } 409 410 return 0; 411} 412 413/// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef 414/// return the specified element value. Otherwise return null. 415static Constant *GetVectorElement(LLVMContext &Context, Constant *C, 416 unsigned EltNo) { 417 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) 418 return CV->getOperand(EltNo); 419 420 const Type *EltTy = cast<VectorType>(C->getType())->getElementType(); 421 if (isa<ConstantAggregateZero>(C)) 422 return Constant::getNullValue(EltTy); 423 if (isa<UndefValue>(C)) 424 return UndefValue::get(EltTy); 425 return 0; 426} 427 428Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, 429 Constant *V1, 430 Constant *V2, 431 Constant *Mask) { 432 // Undefined shuffle mask -> undefined value. 433 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType()); 434 435 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements(); 436 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements(); 437 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType(); 438 439 // Loop over the shuffle mask, evaluating each element. 440 SmallVector<Constant*, 32> Result; 441 for (unsigned i = 0; i != MaskNumElts; ++i) { 442 Constant *InElt = GetVectorElement(Context, Mask, i); 443 if (InElt == 0) return 0; 444 445 if (isa<UndefValue>(InElt)) 446 InElt = UndefValue::get(EltTy); 447 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) { 448 unsigned Elt = CI->getZExtValue(); 449 if (Elt >= SrcNumElts*2) 450 InElt = UndefValue::get(EltTy); 451 else if (Elt >= SrcNumElts) 452 InElt = GetVectorElement(Context, V2, Elt - SrcNumElts); 453 else 454 InElt = GetVectorElement(Context, V1, Elt); 455 if (InElt == 0) return 0; 456 } else { 457 // Unknown value. 458 return 0; 459 } 460 Result.push_back(InElt); 461 } 462 463 return ConstantVector::get(&Result[0], Result.size()); 464} 465 466Constant *llvm::ConstantFoldExtractValueInstruction(LLVMContext &Context, 467 Constant *Agg, 468 const unsigned *Idxs, 469 unsigned NumIdx) { 470 // Base case: no indices, so return the entire value. 471 if (NumIdx == 0) 472 return Agg; 473 474 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef 475 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(), 476 Idxs, 477 Idxs + NumIdx)); 478 479 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0 480 return 481 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), 482 Idxs, 483 Idxs + NumIdx)); 484 485 // Otherwise recurse. 486 return ConstantFoldExtractValueInstruction(Context, Agg->getOperand(*Idxs), 487 Idxs+1, NumIdx-1); 488} 489 490Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, 491 Constant *Agg, 492 Constant *Val, 493 const unsigned *Idxs, 494 unsigned NumIdx) { 495 // Base case: no indices, so replace the entire value. 496 if (NumIdx == 0) 497 return Val; 498 499 if (isa<UndefValue>(Agg)) { 500 // Insertion of constant into aggregate undef 501 // Optimize away insertion of undef. 502 if (isa<UndefValue>(Val)) 503 return Agg; 504 505 // 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(); 513 514 std::vector<Constant*> Ops(numOps); 515 for (unsigned i = 0; i < numOps; ++i) { 516 const Type *MemberTy = AggTy->getTypeAtIndex(i); 517 Constant *Op = 518 (*Idxs == i) ? 519 ConstantFoldInsertValueInstruction(Context, UndefValue::get(MemberTy), 520 Val, Idxs+1, NumIdx-1) : 521 UndefValue::get(MemberTy); 522 Ops[i] = Op; 523 } 524 525 if (const StructType* ST = dyn_cast<StructType>(AggTy)) 526 return ConstantStruct::get(Context, Ops, ST->isPacked()); 527 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 528 } 529 530 if (isa<ConstantAggregateZero>(Agg)) { 531 // Insertion of constant into aggregate zero 532 // Optimize away insertion of zero. 533 if (Val->isNullValue()) 534 return Agg; 535 536 // Otherwise break the aggregate zero into multiple zeros and do 537 // the insertion. 538 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 539 unsigned numOps; 540 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 541 numOps = AR->getNumElements(); 542 else 543 numOps = cast<StructType>(AggTy)->getNumElements(); 544 545 std::vector<Constant*> Ops(numOps); 546 for (unsigned i = 0; i < numOps; ++i) { 547 const Type *MemberTy = AggTy->getTypeAtIndex(i); 548 Constant *Op = 549 (*Idxs == i) ? 550 ConstantFoldInsertValueInstruction(Context, 551 Constant::getNullValue(MemberTy), 552 Val, Idxs+1, NumIdx-1) : 553 Constant::getNullValue(MemberTy); 554 Ops[i] = Op; 555 } 556 557 if (const StructType* ST = dyn_cast<StructType>(AggTy)) 558 return ConstantStruct::get(Context, Ops, ST->isPacked()); 559 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 560 } 561 562 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) { 563 // Insertion of constant into aggregate constant. 564 std::vector<Constant*> Ops(Agg->getNumOperands()); 565 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { 566 Constant *Op = 567 (*Idxs == i) ? 568 ConstantFoldInsertValueInstruction(Context, Agg->getOperand(i), 569 Val, Idxs+1, NumIdx-1) : 570 Agg->getOperand(i); 571 Ops[i] = Op; 572 } 573 574 if (const StructType* ST = dyn_cast<StructType>(Agg->getType())) 575 return ConstantStruct::get(Context, Ops, ST->isPacked()); 576 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops); 577 } 578 579 return 0; 580} 581 582 583Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, 584 unsigned Opcode, 585 Constant *C1, Constant *C2) { 586 // No compile-time operations on this type yet. 587 if (C1->getType() == Type::getPPC_FP128Ty(Context)) 588 return 0; 589 590 // Handle UndefValue up front. 591 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 592 switch (Opcode) { 593 case Instruction::Xor: 594 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 595 // Handle undef ^ undef -> 0 special case. This is a common 596 // idiom (misuse). 597 return Constant::getNullValue(C1->getType()); 598 // Fallthrough 599 case Instruction::Add: 600 case Instruction::Sub: 601 return UndefValue::get(C1->getType()); 602 case Instruction::Mul: 603 case Instruction::And: 604 return Constant::getNullValue(C1->getType()); 605 case Instruction::UDiv: 606 case Instruction::SDiv: 607 case Instruction::URem: 608 case Instruction::SRem: 609 if (!isa<UndefValue>(C2)) // undef / X -> 0 610 return Constant::getNullValue(C1->getType()); 611 return C2; // X / undef -> undef 612 case Instruction::Or: // X | undef -> -1 613 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType())) 614 return Constant::getAllOnesValue(PTy); 615 return Constant::getAllOnesValue(C1->getType()); 616 case Instruction::LShr: 617 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 618 return C1; // undef lshr undef -> undef 619 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 620 // undef lshr X -> 0 621 case Instruction::AShr: 622 if (!isa<UndefValue>(C2)) 623 return C1; // undef ashr X --> undef 624 else if (isa<UndefValue>(C1)) 625 return C1; // undef ashr undef -> undef 626 else 627 return C1; // X ashr undef --> X 628 case Instruction::Shl: 629 // undef << X -> 0 or X << undef -> 0 630 return Constant::getNullValue(C1->getType()); 631 } 632 } 633 634 // Handle simplifications when the RHS is a constant int. 635 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 636 switch (Opcode) { 637 case Instruction::Add: 638 if (CI2->equalsInt(0)) return C1; // X + 0 == X 639 break; 640 case Instruction::Sub: 641 if (CI2->equalsInt(0)) return C1; // X - 0 == X 642 break; 643 case Instruction::Mul: 644 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 645 if (CI2->equalsInt(1)) 646 return C1; // X * 1 == X 647 break; 648 case Instruction::UDiv: 649 case Instruction::SDiv: 650 if (CI2->equalsInt(1)) 651 return C1; // X / 1 == X 652 if (CI2->equalsInt(0)) 653 return UndefValue::get(CI2->getType()); // X / 0 == undef 654 break; 655 case Instruction::URem: 656 case Instruction::SRem: 657 if (CI2->equalsInt(1)) 658 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 659 if (CI2->equalsInt(0)) 660 return UndefValue::get(CI2->getType()); // X % 0 == undef 661 break; 662 case Instruction::And: 663 if (CI2->isZero()) return C2; // X & 0 == 0 664 if (CI2->isAllOnesValue()) 665 return C1; // X & -1 == X 666 667 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 668 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 669 if (CE1->getOpcode() == Instruction::ZExt) { 670 unsigned DstWidth = CI2->getType()->getBitWidth(); 671 unsigned SrcWidth = 672 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 673 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 674 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 675 return C1; 676 } 677 678 // If and'ing the address of a global with a constant, fold it. 679 if (CE1->getOpcode() == Instruction::PtrToInt && 680 isa<GlobalValue>(CE1->getOperand(0))) { 681 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 682 683 // Functions are at least 4-byte aligned. 684 unsigned GVAlign = GV->getAlignment(); 685 if (isa<Function>(GV)) 686 GVAlign = std::max(GVAlign, 4U); 687 688 if (GVAlign > 1) { 689 unsigned DstWidth = CI2->getType()->getBitWidth(); 690 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 691 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 692 693 // If checking bits we know are clear, return zero. 694 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 695 return Constant::getNullValue(CI2->getType()); 696 } 697 } 698 } 699 break; 700 case Instruction::Or: 701 if (CI2->equalsInt(0)) return C1; // X | 0 == X 702 if (CI2->isAllOnesValue()) 703 return C2; // X | -1 == -1 704 break; 705 case Instruction::Xor: 706 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 707 break; 708 case Instruction::AShr: 709 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 710 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 711 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 712 return ConstantExpr::getLShr(C1, C2); 713 break; 714 } 715 } 716 717 // At this point we know neither constant is an UndefValue. 718 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 719 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 720 using namespace APIntOps; 721 const APInt &C1V = CI1->getValue(); 722 const APInt &C2V = CI2->getValue(); 723 switch (Opcode) { 724 default: 725 break; 726 case Instruction::Add: 727 return ConstantInt::get(Context, C1V + C2V); 728 case Instruction::Sub: 729 return ConstantInt::get(Context, C1V - C2V); 730 case Instruction::Mul: 731 return ConstantInt::get(Context, C1V * C2V); 732 case Instruction::UDiv: 733 assert(!CI2->isNullValue() && "Div by zero handled above"); 734 return ConstantInt::get(Context, C1V.udiv(C2V)); 735 case Instruction::SDiv: 736 assert(!CI2->isNullValue() && "Div by zero handled above"); 737 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 738 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 739 return ConstantInt::get(Context, C1V.sdiv(C2V)); 740 case Instruction::URem: 741 assert(!CI2->isNullValue() && "Div by zero handled above"); 742 return ConstantInt::get(Context, C1V.urem(C2V)); 743 case Instruction::SRem: 744 assert(!CI2->isNullValue() && "Div by zero handled above"); 745 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 746 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 747 return ConstantInt::get(Context, C1V.srem(C2V)); 748 case Instruction::And: 749 return ConstantInt::get(Context, C1V & C2V); 750 case Instruction::Or: 751 return ConstantInt::get(Context, C1V | C2V); 752 case Instruction::Xor: 753 return ConstantInt::get(Context, C1V ^ C2V); 754 case Instruction::Shl: { 755 uint32_t shiftAmt = C2V.getZExtValue(); 756 if (shiftAmt < C1V.getBitWidth()) 757 return ConstantInt::get(Context, C1V.shl(shiftAmt)); 758 else 759 return UndefValue::get(C1->getType()); // too big shift is undef 760 } 761 case Instruction::LShr: { 762 uint32_t shiftAmt = C2V.getZExtValue(); 763 if (shiftAmt < C1V.getBitWidth()) 764 return ConstantInt::get(Context, C1V.lshr(shiftAmt)); 765 else 766 return UndefValue::get(C1->getType()); // too big shift is undef 767 } 768 case Instruction::AShr: { 769 uint32_t shiftAmt = C2V.getZExtValue(); 770 if (shiftAmt < C1V.getBitWidth()) 771 return ConstantInt::get(Context, C1V.ashr(shiftAmt)); 772 else 773 return UndefValue::get(C1->getType()); // too big shift is undef 774 } 775 } 776 } 777 778 switch (Opcode) { 779 case Instruction::SDiv: 780 case Instruction::UDiv: 781 case Instruction::URem: 782 case Instruction::SRem: 783 case Instruction::LShr: 784 case Instruction::AShr: 785 case Instruction::Shl: 786 if (CI1->equalsInt(0)) return C1; 787 break; 788 default: 789 break; 790 } 791 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 792 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 793 APFloat C1V = CFP1->getValueAPF(); 794 APFloat C2V = CFP2->getValueAPF(); 795 APFloat C3V = C1V; // copy for modification 796 switch (Opcode) { 797 default: 798 break; 799 case Instruction::FAdd: 800 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 801 return ConstantFP::get(Context, C3V); 802 case Instruction::FSub: 803 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 804 return ConstantFP::get(Context, C3V); 805 case Instruction::FMul: 806 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 807 return ConstantFP::get(Context, C3V); 808 case Instruction::FDiv: 809 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 810 return ConstantFP::get(Context, C3V); 811 case Instruction::FRem: 812 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 813 return ConstantFP::get(Context, C3V); 814 } 815 } 816 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 817 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1); 818 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2); 819 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) && 820 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) { 821 std::vector<Constant*> Res; 822 const Type* EltTy = VTy->getElementType(); 823 Constant *C1 = 0; 824 Constant *C2 = 0; 825 switch (Opcode) { 826 default: 827 break; 828 case Instruction::Add: 829 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 830 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 831 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 832 Res.push_back(ConstantExpr::getAdd(C1, 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(C1, C2)); 840 } 841 return ConstantVector::get(Res); 842 case Instruction::Sub: 843 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 844 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 845 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 846 Res.push_back(ConstantExpr::getSub(C1, C2)); 847 } 848 return ConstantVector::get(Res); 849 case Instruction::FSub: 850 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 851 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 852 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 853 Res.push_back(ConstantExpr::getFSub(C1, C2)); 854 } 855 return ConstantVector::get(Res); 856 case Instruction::Mul: 857 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 858 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 859 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 860 Res.push_back(ConstantExpr::getMul(C1, C2)); 861 } 862 return ConstantVector::get(Res); 863 case Instruction::FMul: 864 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 865 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 866 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 867 Res.push_back(ConstantExpr::getFMul(C1, C2)); 868 } 869 return ConstantVector::get(Res); 870 case Instruction::UDiv: 871 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 872 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 873 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 874 Res.push_back(ConstantExpr::getUDiv(C1, C2)); 875 } 876 return ConstantVector::get(Res); 877 case Instruction::SDiv: 878 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 879 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 880 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 881 Res.push_back(ConstantExpr::getSDiv(C1, C2)); 882 } 883 return ConstantVector::get(Res); 884 case Instruction::FDiv: 885 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 886 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 887 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 888 Res.push_back(ConstantExpr::getFDiv(C1, C2)); 889 } 890 return ConstantVector::get(Res); 891 case Instruction::URem: 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::getURem(C1, C2)); 896 } 897 return ConstantVector::get(Res); 898 case Instruction::SRem: 899 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 900 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 901 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 902 Res.push_back(ConstantExpr::getSRem(C1, C2)); 903 } 904 return ConstantVector::get(Res); 905 case Instruction::FRem: 906 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 907 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 908 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 909 Res.push_back(ConstantExpr::getFRem(C1, C2)); 910 } 911 return ConstantVector::get(Res); 912 case Instruction::And: 913 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 914 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 915 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 916 Res.push_back(ConstantExpr::getAnd(C1, C2)); 917 } 918 return ConstantVector::get(Res); 919 case Instruction::Or: 920 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 921 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 922 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 923 Res.push_back(ConstantExpr::getOr(C1, C2)); 924 } 925 return ConstantVector::get(Res); 926 case Instruction::Xor: 927 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 928 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 929 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 930 Res.push_back(ConstantExpr::getXor(C1, C2)); 931 } 932 return ConstantVector::get(Res); 933 case Instruction::LShr: 934 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 935 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 936 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 937 Res.push_back(ConstantExpr::getLShr(C1, C2)); 938 } 939 return ConstantVector::get(Res); 940 case Instruction::AShr: 941 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 942 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 943 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 944 Res.push_back(ConstantExpr::getAShr(C1, C2)); 945 } 946 return ConstantVector::get(Res); 947 case Instruction::Shl: 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::getShl(C1, C2)); 952 } 953 return ConstantVector::get(Res); 954 } 955 } 956 } 957 958 if (isa<ConstantExpr>(C1)) { 959 // There are many possible foldings we could do here. We should probably 960 // at least fold add of a pointer with an integer into the appropriate 961 // getelementptr. This will improve alias analysis a bit. 962 } else if (isa<ConstantExpr>(C2)) { 963 // If C2 is a constant expr and C1 isn't, flop them around and fold the 964 // other way if possible. 965 switch (Opcode) { 966 case Instruction::Add: 967 case Instruction::FAdd: 968 case Instruction::Mul: 969 case Instruction::FMul: 970 case Instruction::And: 971 case Instruction::Or: 972 case Instruction::Xor: 973 // No change of opcode required. 974 return ConstantFoldBinaryInstruction(Context, Opcode, C2, C1); 975 976 case Instruction::Shl: 977 case Instruction::LShr: 978 case Instruction::AShr: 979 case Instruction::Sub: 980 case Instruction::FSub: 981 case Instruction::SDiv: 982 case Instruction::UDiv: 983 case Instruction::FDiv: 984 case Instruction::URem: 985 case Instruction::SRem: 986 case Instruction::FRem: 987 default: // These instructions cannot be flopped around. 988 break; 989 } 990 } 991 992 // i1 can be simplified in many cases. 993 if (C1->getType() == Type::getInt1Ty(Context)) { 994 switch (Opcode) { 995 case Instruction::Add: 996 case Instruction::Sub: 997 return ConstantExpr::getXor(C1, C2); 998 case Instruction::Mul: 999 return ConstantExpr::getAnd(C1, C2); 1000 case Instruction::Shl: 1001 case Instruction::LShr: 1002 case Instruction::AShr: 1003 // We can assume that C2 == 0. If it were one the result would be 1004 // undefined because the shift value is as large as the bitwidth. 1005 return C1; 1006 case Instruction::SDiv: 1007 case Instruction::UDiv: 1008 // We can assume that C2 == 1. If it were zero the result would be 1009 // undefined through division by zero. 1010 return C1; 1011 case Instruction::URem: 1012 case Instruction::SRem: 1013 // We can assume that C2 == 1. If it were zero the result would be 1014 // undefined through division by zero. 1015 return ConstantInt::getFalse(Context); 1016 default: 1017 break; 1018 } 1019 } 1020 1021 // We don't know how to fold this. 1022 return 0; 1023} 1024 1025/// isZeroSizedType - This type is zero sized if its an array or structure of 1026/// zero sized types. The only leaf zero sized type is an empty structure. 1027static bool isMaybeZeroSizedType(const Type *Ty) { 1028 if (isa<OpaqueType>(Ty)) return true; // Can't say. 1029 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1030 1031 // If all of elements have zero size, this does too. 1032 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1033 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1034 return true; 1035 1036 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1037 return isMaybeZeroSizedType(ATy->getElementType()); 1038 } 1039 return false; 1040} 1041 1042/// IdxCompare - Compare the two constants as though they were getelementptr 1043/// indices. This allows coersion of the types to be the same thing. 1044/// 1045/// If the two constants are the "same" (after coersion), return 0. If the 1046/// first is less than the second, return -1, if the second is less than the 1047/// first, return 1. If the constants are not integral, return -2. 1048/// 1049static int IdxCompare(LLVMContext &Context, Constant *C1, Constant *C2, 1050 const Type *ElTy) { 1051 if (C1 == C2) return 0; 1052 1053 // Ok, we found a different index. If they are not ConstantInt, we can't do 1054 // anything with them. 1055 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1056 return -2; // don't know! 1057 1058 // Ok, we have two differing integer indices. Sign extend them to be the same 1059 // type. Long is always big enough, so we use it. 1060 if (C1->getType() != Type::getInt64Ty(Context)) 1061 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(Context)); 1062 1063 if (C2->getType() != Type::getInt64Ty(Context)) 1064 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(Context)); 1065 1066 if (C1 == C2) return 0; // They are equal 1067 1068 // If the type being indexed over is really just a zero sized type, there is 1069 // no pointer difference being made here. 1070 if (isMaybeZeroSizedType(ElTy)) 1071 return -2; // dunno. 1072 1073 // If they are really different, now that they are the same type, then we 1074 // found a difference! 1075 if (cast<ConstantInt>(C1)->getSExtValue() < 1076 cast<ConstantInt>(C2)->getSExtValue()) 1077 return -1; 1078 else 1079 return 1; 1080} 1081 1082/// evaluateFCmpRelation - This function determines if there is anything we can 1083/// decide about the two constants provided. This doesn't need to handle simple 1084/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1085/// If we can determine that the two constants have a particular relation to 1086/// each other, we should return the corresponding FCmpInst predicate, 1087/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1088/// ConstantFoldCompareInstruction. 1089/// 1090/// To simplify this code we canonicalize the relation so that the first 1091/// operand is always the most "complex" of the two. We consider ConstantFP 1092/// to be the simplest, and ConstantExprs to be the most complex. 1093static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, 1094 Constant *V1, Constant *V2) { 1095 assert(V1->getType() == V2->getType() && 1096 "Cannot compare values of different types!"); 1097 1098 // No compile-time operations on this type yet. 1099 if (V1->getType() == Type::getPPC_FP128Ty(Context)) 1100 return FCmpInst::BAD_FCMP_PREDICATE; 1101 1102 // Handle degenerate case quickly 1103 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1104 1105 if (!isa<ConstantExpr>(V1)) { 1106 if (!isa<ConstantExpr>(V2)) { 1107 // We distilled thisUse the standard constant folder for a few cases 1108 ConstantInt *R = 0; 1109 R = dyn_cast<ConstantInt>( 1110 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1111 if (R && !R->isZero()) 1112 return FCmpInst::FCMP_OEQ; 1113 R = dyn_cast<ConstantInt>( 1114 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1115 if (R && !R->isZero()) 1116 return FCmpInst::FCMP_OLT; 1117 R = dyn_cast<ConstantInt>( 1118 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1119 if (R && !R->isZero()) 1120 return FCmpInst::FCMP_OGT; 1121 1122 // Nothing more we can do 1123 return FCmpInst::BAD_FCMP_PREDICATE; 1124 } 1125 1126 // If the first operand is simple and second is ConstantExpr, swap operands. 1127 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(Context, V2, V1); 1128 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1129 return FCmpInst::getSwappedPredicate(SwappedRelation); 1130 } else { 1131 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1132 // constantexpr or a simple constant. 1133 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1134 switch (CE1->getOpcode()) { 1135 case Instruction::FPTrunc: 1136 case Instruction::FPExt: 1137 case Instruction::UIToFP: 1138 case Instruction::SIToFP: 1139 // We might be able to do something with these but we don't right now. 1140 break; 1141 default: 1142 break; 1143 } 1144 } 1145 // There are MANY other foldings that we could perform here. They will 1146 // probably be added on demand, as they seem needed. 1147 return FCmpInst::BAD_FCMP_PREDICATE; 1148} 1149 1150/// evaluateICmpRelation - This function determines if there is anything we can 1151/// decide about the two constants provided. This doesn't need to handle simple 1152/// things like integer comparisons, but should instead handle ConstantExprs 1153/// and GlobalValues. If we can determine that the two constants have a 1154/// particular relation to each other, we should return the corresponding ICmp 1155/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1156/// 1157/// To simplify this code we canonicalize the relation so that the first 1158/// operand is always the most "complex" of the two. We consider simple 1159/// constants (like ConstantInt) to be the simplest, followed by 1160/// GlobalValues, followed by ConstantExpr's (the most complex). 1161/// 1162static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, 1163 Constant *V1, 1164 Constant *V2, 1165 bool isSigned) { 1166 assert(V1->getType() == V2->getType() && 1167 "Cannot compare different types of values!"); 1168 if (V1 == V2) return ICmpInst::ICMP_EQ; 1169 1170 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1)) { 1171 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2)) { 1172 // We distilled this down to a simple case, use the standard constant 1173 // folder. 1174 ConstantInt *R = 0; 1175 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1176 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1177 if (R && !R->isZero()) 1178 return pred; 1179 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1180 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1181 if (R && !R->isZero()) 1182 return pred; 1183 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1184 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1185 if (R && !R->isZero()) 1186 return pred; 1187 1188 // If we couldn't figure it out, bail. 1189 return ICmpInst::BAD_ICMP_PREDICATE; 1190 } 1191 1192 // If the first operand is simple, swap operands. 1193 ICmpInst::Predicate SwappedRelation = 1194 evaluateICmpRelation(Context, V2, V1, isSigned); 1195 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1196 return ICmpInst::getSwappedPredicate(SwappedRelation); 1197 1198 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(V1)) { 1199 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1200 ICmpInst::Predicate SwappedRelation = 1201 evaluateICmpRelation(Context, V2, V1, isSigned); 1202 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1203 return ICmpInst::getSwappedPredicate(SwappedRelation); 1204 else 1205 return ICmpInst::BAD_ICMP_PREDICATE; 1206 } 1207 1208 // Now we know that the RHS is a GlobalValue or simple constant, 1209 // which (since the types must match) means that it's a ConstantPointerNull. 1210 if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) { 1211 // Don't try to decide equality of aliases. 1212 if (!isa<GlobalAlias>(CPR1) && !isa<GlobalAlias>(CPR2)) 1213 if (!CPR1->hasExternalWeakLinkage() || !CPR2->hasExternalWeakLinkage()) 1214 return ICmpInst::ICMP_NE; 1215 } else { 1216 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1217 // GlobalVals can never be null. Don't try to evaluate aliases. 1218 if (!CPR1->hasExternalWeakLinkage() && !isa<GlobalAlias>(CPR1)) 1219 return ICmpInst::ICMP_NE; 1220 } 1221 } else { 1222 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1223 // constantexpr, a CPR, or a simple constant. 1224 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1225 Constant *CE1Op0 = CE1->getOperand(0); 1226 1227 switch (CE1->getOpcode()) { 1228 case Instruction::Trunc: 1229 case Instruction::FPTrunc: 1230 case Instruction::FPExt: 1231 case Instruction::FPToUI: 1232 case Instruction::FPToSI: 1233 break; // We can't evaluate floating point casts or truncations. 1234 1235 case Instruction::UIToFP: 1236 case Instruction::SIToFP: 1237 case Instruction::BitCast: 1238 case Instruction::ZExt: 1239 case Instruction::SExt: 1240 // If the cast is not actually changing bits, and the second operand is a 1241 // null pointer, do the comparison with the pre-casted value. 1242 if (V2->isNullValue() && 1243 (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) { 1244 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1245 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1246 return evaluateICmpRelation(Context, CE1Op0, 1247 Constant::getNullValue(CE1Op0->getType()), 1248 isSigned); 1249 } 1250 break; 1251 1252 case Instruction::GetElementPtr: 1253 // Ok, since this is a getelementptr, we know that the constant has a 1254 // pointer type. Check the various cases. 1255 if (isa<ConstantPointerNull>(V2)) { 1256 // If we are comparing a GEP to a null pointer, check to see if the base 1257 // of the GEP equals the null pointer. 1258 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1259 if (GV->hasExternalWeakLinkage()) 1260 // Weak linkage GVals could be zero or not. We're comparing that 1261 // to null pointer so its greater-or-equal 1262 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1263 else 1264 // If its not weak linkage, the GVal must have a non-zero address 1265 // so the result is greater-than 1266 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1267 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1268 // If we are indexing from a null pointer, check to see if we have any 1269 // non-zero indices. 1270 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1271 if (!CE1->getOperand(i)->isNullValue()) 1272 // Offsetting from null, must not be equal. 1273 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1274 // Only zero indexes from null, must still be zero. 1275 return ICmpInst::ICMP_EQ; 1276 } 1277 // Otherwise, we can't really say if the first operand is null or not. 1278 } else if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) { 1279 if (isa<ConstantPointerNull>(CE1Op0)) { 1280 if (CPR2->hasExternalWeakLinkage()) 1281 // Weak linkage GVals could be zero or not. We're comparing it to 1282 // a null pointer, so its less-or-equal 1283 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1284 else 1285 // If its not weak linkage, the GVal must have a non-zero address 1286 // so the result is less-than 1287 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1288 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(CE1Op0)) { 1289 if (CPR1 == CPR2) { 1290 // If this is a getelementptr of the same global, then it must be 1291 // different. Because the types must match, the getelementptr could 1292 // only have at most one index, and because we fold getelementptr's 1293 // with a single zero index, it must be nonzero. 1294 assert(CE1->getNumOperands() == 2 && 1295 !CE1->getOperand(1)->isNullValue() && 1296 "Suprising getelementptr!"); 1297 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1298 } else { 1299 // If they are different globals, we don't know what the value is, 1300 // but they can't be equal. 1301 return ICmpInst::ICMP_NE; 1302 } 1303 } 1304 } else { 1305 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1306 Constant *CE2Op0 = CE2->getOperand(0); 1307 1308 // There are MANY other foldings that we could perform here. They will 1309 // probably be added on demand, as they seem needed. 1310 switch (CE2->getOpcode()) { 1311 default: break; 1312 case Instruction::GetElementPtr: 1313 // By far the most common case to handle is when the base pointers are 1314 // obviously to the same or different globals. 1315 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1316 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal 1317 return ICmpInst::ICMP_NE; 1318 // Ok, we know that both getelementptr instructions are based on the 1319 // same global. From this, we can precisely determine the relative 1320 // ordering of the resultant pointers. 1321 unsigned i = 1; 1322 1323 // The logic below assumes that the result of the comparison 1324 // can be determined by finding the first index that differs. 1325 // This doesn't work if there is over-indexing in any 1326 // subsequent indices, so check for that case first. 1327 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1328 !CE2->isGEPWithNoNotionalOverIndexing()) 1329 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1330 1331 // Compare all of the operands the GEP's have in common. 1332 gep_type_iterator GTI = gep_type_begin(CE1); 1333 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1334 ++i, ++GTI) 1335 switch (IdxCompare(Context, CE1->getOperand(i), 1336 CE2->getOperand(i), GTI.getIndexedType())) { 1337 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1338 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1339 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1340 } 1341 1342 // Ok, we ran out of things they have in common. If any leftovers 1343 // are non-zero then we have a difference, otherwise we are equal. 1344 for (; i < CE1->getNumOperands(); ++i) 1345 if (!CE1->getOperand(i)->isNullValue()) { 1346 if (isa<ConstantInt>(CE1->getOperand(i))) 1347 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1348 else 1349 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1350 } 1351 1352 for (; i < CE2->getNumOperands(); ++i) 1353 if (!CE2->getOperand(i)->isNullValue()) { 1354 if (isa<ConstantInt>(CE2->getOperand(i))) 1355 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1356 else 1357 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1358 } 1359 return ICmpInst::ICMP_EQ; 1360 } 1361 } 1362 } 1363 default: 1364 break; 1365 } 1366 } 1367 1368 return ICmpInst::BAD_ICMP_PREDICATE; 1369} 1370 1371Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, 1372 unsigned short pred, 1373 Constant *C1, Constant *C2) { 1374 const Type *ResultTy; 1375 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1376 ResultTy = VectorType::get(Type::getInt1Ty(Context), VT->getNumElements()); 1377 else 1378 ResultTy = Type::getInt1Ty(Context); 1379 1380 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1381 if (pred == FCmpInst::FCMP_FALSE) 1382 return Constant::getNullValue(ResultTy); 1383 1384 if (pred == FCmpInst::FCMP_TRUE) 1385 return Constant::getAllOnesValue(ResultTy); 1386 1387 // Handle some degenerate cases first 1388 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) 1389 return UndefValue::get(ResultTy); 1390 1391 // No compile-time operations on this type yet. 1392 if (C1->getType() == Type::getPPC_FP128Ty(Context)) 1393 return 0; 1394 1395 // icmp eq/ne(null,GV) -> false/true 1396 if (C1->isNullValue()) { 1397 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1398 // Don't try to evaluate aliases. External weak GV can be null. 1399 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1400 if (pred == ICmpInst::ICMP_EQ) 1401 return ConstantInt::getFalse(Context); 1402 else if (pred == ICmpInst::ICMP_NE) 1403 return ConstantInt::getTrue(Context); 1404 } 1405 // icmp eq/ne(GV,null) -> false/true 1406 } else if (C2->isNullValue()) { 1407 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 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 } 1416 1417 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1418 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1419 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1420 switch (pred) { 1421 default: llvm_unreachable("Invalid ICmp Predicate"); return 0; 1422 case ICmpInst::ICMP_EQ: 1423 return ConstantInt::get(Type::getInt1Ty(Context), V1 == V2); 1424 case ICmpInst::ICMP_NE: 1425 return ConstantInt::get(Type::getInt1Ty(Context), V1 != V2); 1426 case ICmpInst::ICMP_SLT: 1427 return ConstantInt::get(Type::getInt1Ty(Context), V1.slt(V2)); 1428 case ICmpInst::ICMP_SGT: 1429 return ConstantInt::get(Type::getInt1Ty(Context), V1.sgt(V2)); 1430 case ICmpInst::ICMP_SLE: 1431 return ConstantInt::get(Type::getInt1Ty(Context), V1.sle(V2)); 1432 case ICmpInst::ICMP_SGE: 1433 return ConstantInt::get(Type::getInt1Ty(Context), V1.sge(V2)); 1434 case ICmpInst::ICMP_ULT: 1435 return ConstantInt::get(Type::getInt1Ty(Context), V1.ult(V2)); 1436 case ICmpInst::ICMP_UGT: 1437 return ConstantInt::get(Type::getInt1Ty(Context), V1.ugt(V2)); 1438 case ICmpInst::ICMP_ULE: 1439 return ConstantInt::get(Type::getInt1Ty(Context), V1.ule(V2)); 1440 case ICmpInst::ICMP_UGE: 1441 return ConstantInt::get(Type::getInt1Ty(Context), V1.uge(V2)); 1442 } 1443 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1444 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1445 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1446 APFloat::cmpResult R = C1V.compare(C2V); 1447 switch (pred) { 1448 default: llvm_unreachable("Invalid FCmp Predicate"); return 0; 1449 case FCmpInst::FCMP_FALSE: return ConstantInt::getFalse(Context); 1450 case FCmpInst::FCMP_TRUE: return ConstantInt::getTrue(Context); 1451 case FCmpInst::FCMP_UNO: 1452 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered); 1453 case FCmpInst::FCMP_ORD: 1454 return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpUnordered); 1455 case FCmpInst::FCMP_UEQ: 1456 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || 1457 R==APFloat::cmpEqual); 1458 case FCmpInst::FCMP_OEQ: 1459 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpEqual); 1460 case FCmpInst::FCMP_UNE: 1461 return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpEqual); 1462 case FCmpInst::FCMP_ONE: 1463 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan || 1464 R==APFloat::cmpGreaterThan); 1465 case FCmpInst::FCMP_ULT: 1466 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || 1467 R==APFloat::cmpLessThan); 1468 case FCmpInst::FCMP_OLT: 1469 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan); 1470 case FCmpInst::FCMP_UGT: 1471 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpUnordered || 1472 R==APFloat::cmpGreaterThan); 1473 case FCmpInst::FCMP_OGT: 1474 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpGreaterThan); 1475 case FCmpInst::FCMP_ULE: 1476 return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpGreaterThan); 1477 case FCmpInst::FCMP_OLE: 1478 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpLessThan || 1479 R==APFloat::cmpEqual); 1480 case FCmpInst::FCMP_UGE: 1481 return ConstantInt::get(Type::getInt1Ty(Context), R!=APFloat::cmpLessThan); 1482 case FCmpInst::FCMP_OGE: 1483 return ConstantInt::get(Type::getInt1Ty(Context), R==APFloat::cmpGreaterThan || 1484 R==APFloat::cmpEqual); 1485 } 1486 } else if (isa<VectorType>(C1->getType())) { 1487 SmallVector<Constant*, 16> C1Elts, C2Elts; 1488 C1->getVectorElements(Context, C1Elts); 1489 C2->getVectorElements(Context, C2Elts); 1490 1491 // If we can constant fold the comparison of each element, constant fold 1492 // the whole vector comparison. 1493 SmallVector<Constant*, 4> ResElts; 1494 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) { 1495 // Compare the elements, producing an i1 result or constant expr. 1496 ResElts.push_back( 1497 ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); 1498 } 1499 return ConstantVector::get(&ResElts[0], ResElts.size()); 1500 } 1501 1502 if (C1->getType()->isFloatingPoint()) { 1503 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1504 switch (evaluateFCmpRelation(Context, C1, C2)) { 1505 default: llvm_unreachable("Unknown relation!"); 1506 case FCmpInst::FCMP_UNO: 1507 case FCmpInst::FCMP_ORD: 1508 case FCmpInst::FCMP_UEQ: 1509 case FCmpInst::FCMP_UNE: 1510 case FCmpInst::FCMP_ULT: 1511 case FCmpInst::FCMP_UGT: 1512 case FCmpInst::FCMP_ULE: 1513 case FCmpInst::FCMP_UGE: 1514 case FCmpInst::FCMP_TRUE: 1515 case FCmpInst::FCMP_FALSE: 1516 case FCmpInst::BAD_FCMP_PREDICATE: 1517 break; // Couldn't determine anything about these constants. 1518 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1519 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1520 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1521 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1522 break; 1523 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1524 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1525 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1526 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1527 break; 1528 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1529 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1530 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1531 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1532 break; 1533 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1534 // We can only partially decide this relation. 1535 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1536 Result = 0; 1537 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1538 Result = 1; 1539 break; 1540 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1541 // We can only partially decide this relation. 1542 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1543 Result = 0; 1544 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1545 Result = 1; 1546 break; 1547 case ICmpInst::ICMP_NE: // We know that C1 != C2 1548 // We can only partially decide this relation. 1549 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1550 Result = 0; 1551 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1552 Result = 1; 1553 break; 1554 } 1555 1556 // If we evaluated the result, return it now. 1557 if (Result != -1) 1558 return ConstantInt::get(Type::getInt1Ty(Context), Result); 1559 1560 } else { 1561 // Evaluate the relation between the two constants, per the predicate. 1562 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1563 switch (evaluateICmpRelation(Context, C1, C2, CmpInst::isSigned(pred))) { 1564 default: llvm_unreachable("Unknown relational!"); 1565 case ICmpInst::BAD_ICMP_PREDICATE: 1566 break; // Couldn't determine anything about these constants. 1567 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1568 // If we know the constants are equal, we can decide the result of this 1569 // computation precisely. 1570 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1571 break; 1572 case ICmpInst::ICMP_ULT: 1573 switch (pred) { 1574 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1575 Result = 1; break; 1576 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1577 Result = 0; break; 1578 } 1579 break; 1580 case ICmpInst::ICMP_SLT: 1581 switch (pred) { 1582 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1583 Result = 1; break; 1584 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1585 Result = 0; break; 1586 } 1587 break; 1588 case ICmpInst::ICMP_UGT: 1589 switch (pred) { 1590 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1591 Result = 1; break; 1592 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1593 Result = 0; break; 1594 } 1595 break; 1596 case ICmpInst::ICMP_SGT: 1597 switch (pred) { 1598 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1599 Result = 1; break; 1600 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1601 Result = 0; break; 1602 } 1603 break; 1604 case ICmpInst::ICMP_ULE: 1605 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1606 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1607 break; 1608 case ICmpInst::ICMP_SLE: 1609 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1610 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1611 break; 1612 case ICmpInst::ICMP_UGE: 1613 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1614 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1615 break; 1616 case ICmpInst::ICMP_SGE: 1617 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1618 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1619 break; 1620 case ICmpInst::ICMP_NE: 1621 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1622 if (pred == ICmpInst::ICMP_NE) Result = 1; 1623 break; 1624 } 1625 1626 // If we evaluated the result, return it now. 1627 if (Result != -1) 1628 return ConstantInt::get(Type::getInt1Ty(Context), Result); 1629 1630 if (!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) { 1631 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1632 // other way if possible. 1633 switch (pred) { 1634 case ICmpInst::ICMP_EQ: 1635 case ICmpInst::ICMP_NE: 1636 // No change of predicate required. 1637 return ConstantFoldCompareInstruction(Context, pred, C2, C1); 1638 1639 case ICmpInst::ICMP_ULT: 1640 case ICmpInst::ICMP_SLT: 1641 case ICmpInst::ICMP_UGT: 1642 case ICmpInst::ICMP_SGT: 1643 case ICmpInst::ICMP_ULE: 1644 case ICmpInst::ICMP_SLE: 1645 case ICmpInst::ICMP_UGE: 1646 case ICmpInst::ICMP_SGE: 1647 // Change the predicate as necessary to swap the operands. 1648 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1649 return ConstantFoldCompareInstruction(Context, pred, C2, C1); 1650 1651 default: // These predicates cannot be flopped around. 1652 break; 1653 } 1654 } 1655 } 1656 return 0; 1657} 1658 1659/// isInBoundsIndices - Test whether the given sequence of *normalized* indices 1660/// is "inbounds". 1661static bool isInBoundsIndices(Constant *const *Idxs, size_t NumIdx) { 1662 // No indices means nothing that could be out of bounds. 1663 if (NumIdx == 0) return true; 1664 1665 // If the first index is zero, it's in bounds. 1666 if (Idxs[0]->isNullValue()) return true; 1667 1668 // If the first index is one and all the rest are zero, it's in bounds, 1669 // by the one-past-the-end rule. 1670 if (!cast<ConstantInt>(Idxs[0])->isOne()) 1671 return false; 1672 for (unsigned i = 1, e = NumIdx; i != e; ++i) 1673 if (!Idxs[i]->isNullValue()) 1674 return false; 1675 return true; 1676} 1677 1678Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, 1679 Constant *C, 1680 bool inBounds, 1681 Constant* const *Idxs, 1682 unsigned NumIdx) { 1683 if (NumIdx == 0 || 1684 (NumIdx == 1 && Idxs[0]->isNullValue())) 1685 return C; 1686 1687 if (isa<UndefValue>(C)) { 1688 const PointerType *Ptr = cast<PointerType>(C->getType()); 1689 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, 1690 (Value **)Idxs, 1691 (Value **)Idxs+NumIdx); 1692 assert(Ty != 0 && "Invalid indices for GEP!"); 1693 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 1694 } 1695 1696 Constant *Idx0 = Idxs[0]; 1697 if (C->isNullValue()) { 1698 bool isNull = true; 1699 for (unsigned i = 0, e = NumIdx; i != e; ++i) 1700 if (!Idxs[i]->isNullValue()) { 1701 isNull = false; 1702 break; 1703 } 1704 if (isNull) { 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 ConstantPointerNull::get( 1711 PointerType::get(Ty,Ptr->getAddressSpace())); 1712 } 1713 } 1714 1715 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 1716 // Combine Indices - If the source pointer to this getelementptr instruction 1717 // is a getelementptr instruction, combine the indices of the two 1718 // getelementptr instructions into a single instruction. 1719 // 1720 if (CE->getOpcode() == Instruction::GetElementPtr) { 1721 const Type *LastTy = 0; 1722 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 1723 I != E; ++I) 1724 LastTy = *I; 1725 1726 if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) { 1727 SmallVector<Value*, 16> NewIndices; 1728 NewIndices.reserve(NumIdx + CE->getNumOperands()); 1729 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 1730 NewIndices.push_back(CE->getOperand(i)); 1731 1732 // Add the last index of the source with the first index of the new GEP. 1733 // Make sure to handle the case when they are actually different types. 1734 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 1735 // Otherwise it must be an array. 1736 if (!Idx0->isNullValue()) { 1737 const Type *IdxTy = Combined->getType(); 1738 if (IdxTy != Idx0->getType()) { 1739 Constant *C1 = 1740 ConstantExpr::getSExtOrBitCast(Idx0, Type::getInt64Ty(Context)); 1741 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, 1742 Type::getInt64Ty(Context)); 1743 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 1744 } else { 1745 Combined = 1746 ConstantExpr::get(Instruction::Add, Idx0, Combined); 1747 } 1748 } 1749 1750 NewIndices.push_back(Combined); 1751 NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx); 1752 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ? 1753 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0), 1754 &NewIndices[0], 1755 NewIndices.size()) : 1756 ConstantExpr::getGetElementPtr(CE->getOperand(0), 1757 &NewIndices[0], 1758 NewIndices.size()); 1759 } 1760 } 1761 1762 // Implement folding of: 1763 // int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*), 1764 // long 0, long 0) 1765 // To: int* getelementptr ([3 x int]* %X, long 0, long 0) 1766 // 1767 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) { 1768 if (const PointerType *SPT = 1769 dyn_cast<PointerType>(CE->getOperand(0)->getType())) 1770 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType())) 1771 if (const ArrayType *CAT = 1772 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType())) 1773 if (CAT->getElementType() == SAT->getElementType()) 1774 return inBounds ? 1775 ConstantExpr::getInBoundsGetElementPtr( 1776 (Constant*)CE->getOperand(0), Idxs, NumIdx) : 1777 ConstantExpr::getGetElementPtr( 1778 (Constant*)CE->getOperand(0), Idxs, NumIdx); 1779 } 1780 1781 // Fold: getelementptr (i8* inttoptr (i64 1 to i8*), i32 -1) 1782 // Into: inttoptr (i64 0 to i8*) 1783 // This happens with pointers to member functions in C++. 1784 if (CE->getOpcode() == Instruction::IntToPtr && NumIdx == 1 && 1785 isa<ConstantInt>(CE->getOperand(0)) && isa<ConstantInt>(Idxs[0]) && 1786 cast<PointerType>(CE->getType())->getElementType() == Type::getInt8Ty(Context)) { 1787 Constant *Base = CE->getOperand(0); 1788 Constant *Offset = Idxs[0]; 1789 1790 // Convert the smaller integer to the larger type. 1791 if (Offset->getType()->getPrimitiveSizeInBits() < 1792 Base->getType()->getPrimitiveSizeInBits()) 1793 Offset = ConstantExpr::getSExt(Offset, Base->getType()); 1794 else if (Base->getType()->getPrimitiveSizeInBits() < 1795 Offset->getType()->getPrimitiveSizeInBits()) 1796 Base = ConstantExpr::getZExt(Base, Offset->getType()); 1797 1798 Base = ConstantExpr::getAdd(Base, Offset); 1799 return ConstantExpr::getIntToPtr(Base, CE->getType()); 1800 } 1801 } 1802 1803 // Check to see if any array indices are not within the corresponding 1804 // notional array bounds. If so, try to determine if they can be factored 1805 // out into preceding dimensions. 1806 bool Unknown = false; 1807 SmallVector<Constant *, 8> NewIdxs; 1808 const Type *Ty = C->getType(); 1809 const Type *Prev = 0; 1810 for (unsigned i = 0; i != NumIdx; 1811 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 1812 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 1813 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 1814 if (ATy->getNumElements() <= INT64_MAX && 1815 ATy->getNumElements() != 0 && 1816 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) { 1817 if (isa<SequentialType>(Prev)) { 1818 // It's out of range, but we can factor it into the prior 1819 // dimension. 1820 NewIdxs.resize(NumIdx); 1821 ConstantInt *Factor = ConstantInt::get(CI->getType(), 1822 ATy->getNumElements()); 1823 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 1824 1825 Constant *PrevIdx = Idxs[i-1]; 1826 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 1827 1828 // Before adding, extend both operands to i64 to avoid 1829 // overflow trouble. 1830 if (PrevIdx->getType() != Type::getInt64Ty(Context)) 1831 PrevIdx = ConstantExpr::getSExt(PrevIdx, 1832 Type::getInt64Ty(Context)); 1833 if (Div->getType() != Type::getInt64Ty(Context)) 1834 Div = ConstantExpr::getSExt(Div, 1835 Type::getInt64Ty(Context)); 1836 1837 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 1838 } else { 1839 // It's out of range, but the prior dimension is a struct 1840 // so we can't do anything about it. 1841 Unknown = true; 1842 } 1843 } 1844 } else { 1845 // We don't know if it's in range or not. 1846 Unknown = true; 1847 } 1848 } 1849 1850 // If we did any factoring, start over with the adjusted indices. 1851 if (!NewIdxs.empty()) { 1852 for (unsigned i = 0; i != NumIdx; ++i) 1853 if (!NewIdxs[i]) NewIdxs[i] = Idxs[i]; 1854 return inBounds ? 1855 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(), 1856 NewIdxs.size()) : 1857 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size()); 1858 } 1859 1860 // If all indices are known integers and normalized, we can do a simple 1861 // check for the "inbounds" property. 1862 if (!Unknown && !inBounds && 1863 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx)) 1864 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx); 1865 1866 return 0; 1867} 1868