InstCombineCompares.cpp revision 204642
1//===- InstCombineCompares.cpp --------------------------------------------===// 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 the visitICmp and visitFCmp functions. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/IntrinsicInst.h" 16#include "llvm/Analysis/InstructionSimplify.h" 17#include "llvm/Analysis/MemoryBuiltins.h" 18#include "llvm/Target/TargetData.h" 19#include "llvm/Support/ConstantRange.h" 20#include "llvm/Support/GetElementPtrTypeIterator.h" 21#include "llvm/Support/PatternMatch.h" 22using namespace llvm; 23using namespace PatternMatch; 24 25/// AddOne - Add one to a ConstantInt 26static Constant *AddOne(Constant *C) { 27 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 28} 29/// SubOne - Subtract one from a ConstantInt 30static Constant *SubOne(ConstantInt *C) { 31 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1)); 32} 33 34static ConstantInt *ExtractElement(Constant *V, Constant *Idx) { 35 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx)); 36} 37 38static bool HasAddOverflow(ConstantInt *Result, 39 ConstantInt *In1, ConstantInt *In2, 40 bool IsSigned) { 41 if (IsSigned) 42 if (In2->getValue().isNegative()) 43 return Result->getValue().sgt(In1->getValue()); 44 else 45 return Result->getValue().slt(In1->getValue()); 46 else 47 return Result->getValue().ult(In1->getValue()); 48} 49 50/// AddWithOverflow - Compute Result = In1+In2, returning true if the result 51/// overflowed for this type. 52static bool AddWithOverflow(Constant *&Result, Constant *In1, 53 Constant *In2, bool IsSigned = false) { 54 Result = ConstantExpr::getAdd(In1, In2); 55 56 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 57 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 58 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 59 if (HasAddOverflow(ExtractElement(Result, Idx), 60 ExtractElement(In1, Idx), 61 ExtractElement(In2, Idx), 62 IsSigned)) 63 return true; 64 } 65 return false; 66 } 67 68 return HasAddOverflow(cast<ConstantInt>(Result), 69 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 70 IsSigned); 71} 72 73static bool HasSubOverflow(ConstantInt *Result, 74 ConstantInt *In1, ConstantInt *In2, 75 bool IsSigned) { 76 if (IsSigned) 77 if (In2->getValue().isNegative()) 78 return Result->getValue().slt(In1->getValue()); 79 else 80 return Result->getValue().sgt(In1->getValue()); 81 else 82 return Result->getValue().ugt(In1->getValue()); 83} 84 85/// SubWithOverflow - Compute Result = In1-In2, returning true if the result 86/// overflowed for this type. 87static bool SubWithOverflow(Constant *&Result, Constant *In1, 88 Constant *In2, bool IsSigned = false) { 89 Result = ConstantExpr::getSub(In1, In2); 90 91 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 92 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 93 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 94 if (HasSubOverflow(ExtractElement(Result, Idx), 95 ExtractElement(In1, Idx), 96 ExtractElement(In2, Idx), 97 IsSigned)) 98 return true; 99 } 100 return false; 101 } 102 103 return HasSubOverflow(cast<ConstantInt>(Result), 104 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 105 IsSigned); 106} 107 108/// isSignBitCheck - Given an exploded icmp instruction, return true if the 109/// comparison only checks the sign bit. If it only checks the sign bit, set 110/// TrueIfSigned if the result of the comparison is true when the input value is 111/// signed. 112static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS, 113 bool &TrueIfSigned) { 114 switch (pred) { 115 case ICmpInst::ICMP_SLT: // True if LHS s< 0 116 TrueIfSigned = true; 117 return RHS->isZero(); 118 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1 119 TrueIfSigned = true; 120 return RHS->isAllOnesValue(); 121 case ICmpInst::ICMP_SGT: // True if LHS s> -1 122 TrueIfSigned = false; 123 return RHS->isAllOnesValue(); 124 case ICmpInst::ICMP_UGT: 125 // True if LHS u> RHS and RHS == high-bit-mask - 1 126 TrueIfSigned = true; 127 return RHS->getValue() == 128 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits()); 129 case ICmpInst::ICMP_UGE: 130 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc) 131 TrueIfSigned = true; 132 return RHS->getValue().isSignBit(); 133 default: 134 return false; 135 } 136} 137 138// isHighOnes - Return true if the constant is of the form 1+0+. 139// This is the same as lowones(~X). 140static bool isHighOnes(const ConstantInt *CI) { 141 return (~CI->getValue() + 1).isPowerOf2(); 142} 143 144/// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a 145/// set of known zero and one bits, compute the maximum and minimum values that 146/// could have the specified known zero and known one bits, returning them in 147/// min/max. 148static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero, 149 const APInt& KnownOne, 150 APInt& Min, APInt& Max) { 151 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 152 KnownZero.getBitWidth() == Min.getBitWidth() && 153 KnownZero.getBitWidth() == Max.getBitWidth() && 154 "KnownZero, KnownOne and Min, Max must have equal bitwidth."); 155 APInt UnknownBits = ~(KnownZero|KnownOne); 156 157 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign 158 // bit if it is unknown. 159 Min = KnownOne; 160 Max = KnownOne|UnknownBits; 161 162 if (UnknownBits.isNegative()) { // Sign bit is unknown 163 Min.set(Min.getBitWidth()-1); 164 Max.clear(Max.getBitWidth()-1); 165 } 166} 167 168// ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and 169// a set of known zero and one bits, compute the maximum and minimum values that 170// could have the specified known zero and known one bits, returning them in 171// min/max. 172static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero, 173 const APInt &KnownOne, 174 APInt &Min, APInt &Max) { 175 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 176 KnownZero.getBitWidth() == Min.getBitWidth() && 177 KnownZero.getBitWidth() == Max.getBitWidth() && 178 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."); 179 APInt UnknownBits = ~(KnownZero|KnownOne); 180 181 // The minimum value is when the unknown bits are all zeros. 182 Min = KnownOne; 183 // The maximum value is when the unknown bits are all ones. 184 Max = KnownOne|UnknownBits; 185} 186 187 188 189/// FoldCmpLoadFromIndexedGlobal - Called we see this pattern: 190/// cmp pred (load (gep GV, ...)), cmpcst 191/// where GV is a global variable with a constant initializer. Try to simplify 192/// this into some simple computation that does not need the load. For example 193/// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3". 194/// 195/// If AndCst is non-null, then the loaded value is masked with that constant 196/// before doing the comparison. This handles cases like "A[i]&4 == 0". 197Instruction *InstCombiner:: 198FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV, 199 CmpInst &ICI, ConstantInt *AndCst) { 200 // We need TD information to know the pointer size unless this is inbounds. 201 if (!GEP->isInBounds() && TD == 0) return 0; 202 203 ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer()); 204 if (Init == 0 || Init->getNumOperands() > 1024) return 0; 205 206 // There are many forms of this optimization we can handle, for now, just do 207 // the simple index into a single-dimensional array. 208 // 209 // Require: GEP GV, 0, i {{, constant indices}} 210 if (GEP->getNumOperands() < 3 || 211 !isa<ConstantInt>(GEP->getOperand(1)) || 212 !cast<ConstantInt>(GEP->getOperand(1))->isZero() || 213 isa<Constant>(GEP->getOperand(2))) 214 return 0; 215 216 // Check that indices after the variable are constants and in-range for the 217 // type they index. Collect the indices. This is typically for arrays of 218 // structs. 219 SmallVector<unsigned, 4> LaterIndices; 220 221 const Type *EltTy = cast<ArrayType>(Init->getType())->getElementType(); 222 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { 223 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i)); 224 if (Idx == 0) return 0; // Variable index. 225 226 uint64_t IdxVal = Idx->getZExtValue(); 227 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index. 228 229 if (const StructType *STy = dyn_cast<StructType>(EltTy)) 230 EltTy = STy->getElementType(IdxVal); 231 else if (const ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) { 232 if (IdxVal >= ATy->getNumElements()) return 0; 233 EltTy = ATy->getElementType(); 234 } else { 235 return 0; // Unknown type. 236 } 237 238 LaterIndices.push_back(IdxVal); 239 } 240 241 enum { Overdefined = -3, Undefined = -2 }; 242 243 // Variables for our state machines. 244 245 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form 246 // "i == 47 | i == 87", where 47 is the first index the condition is true for, 247 // and 87 is the second (and last) index. FirstTrueElement is -2 when 248 // undefined, otherwise set to the first true element. SecondTrueElement is 249 // -2 when undefined, -3 when overdefined and >= 0 when that index is true. 250 int FirstTrueElement = Undefined, SecondTrueElement = Undefined; 251 252 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the 253 // form "i != 47 & i != 87". Same state transitions as for true elements. 254 int FirstFalseElement = Undefined, SecondFalseElement = Undefined; 255 256 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these 257 /// define a state machine that triggers for ranges of values that the index 258 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. 259 /// This is -2 when undefined, -3 when overdefined, and otherwise the last 260 /// index in the range (inclusive). We use -2 for undefined here because we 261 /// use relative comparisons and don't want 0-1 to match -1. 262 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; 263 264 // MagicBitvector - This is a magic bitvector where we set a bit if the 265 // comparison is true for element 'i'. If there are 64 elements or less in 266 // the array, this will fully represent all the comparison results. 267 uint64_t MagicBitvector = 0; 268 269 270 // Scan the array and see if one of our patterns matches. 271 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1)); 272 for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) { 273 Constant *Elt = Init->getOperand(i); 274 275 // If this is indexing an array of structures, get the structure element. 276 if (!LaterIndices.empty()) 277 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices.data(), 278 LaterIndices.size()); 279 280 // If the element is masked, handle it. 281 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst); 282 283 // Find out if the comparison would be true or false for the i'th element. 284 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt, 285 CompareRHS, TD); 286 // If the result is undef for this element, ignore it. 287 if (isa<UndefValue>(C)) { 288 // Extend range state machines to cover this element in case there is an 289 // undef in the middle of the range. 290 if (TrueRangeEnd == (int)i-1) 291 TrueRangeEnd = i; 292 if (FalseRangeEnd == (int)i-1) 293 FalseRangeEnd = i; 294 continue; 295 } 296 297 // If we can't compute the result for any of the elements, we have to give 298 // up evaluating the entire conditional. 299 if (!isa<ConstantInt>(C)) return 0; 300 301 // Otherwise, we know if the comparison is true or false for this element, 302 // update our state machines. 303 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero(); 304 305 // State machine for single/double/range index comparison. 306 if (IsTrueForElt) { 307 // Update the TrueElement state machine. 308 if (FirstTrueElement == Undefined) 309 FirstTrueElement = TrueRangeEnd = i; // First true element. 310 else { 311 // Update double-compare state machine. 312 if (SecondTrueElement == Undefined) 313 SecondTrueElement = i; 314 else 315 SecondTrueElement = Overdefined; 316 317 // Update range state machine. 318 if (TrueRangeEnd == (int)i-1) 319 TrueRangeEnd = i; 320 else 321 TrueRangeEnd = Overdefined; 322 } 323 } else { 324 // Update the FalseElement state machine. 325 if (FirstFalseElement == Undefined) 326 FirstFalseElement = FalseRangeEnd = i; // First false element. 327 else { 328 // Update double-compare state machine. 329 if (SecondFalseElement == Undefined) 330 SecondFalseElement = i; 331 else 332 SecondFalseElement = Overdefined; 333 334 // Update range state machine. 335 if (FalseRangeEnd == (int)i-1) 336 FalseRangeEnd = i; 337 else 338 FalseRangeEnd = Overdefined; 339 } 340 } 341 342 343 // If this element is in range, update our magic bitvector. 344 if (i < 64 && IsTrueForElt) 345 MagicBitvector |= 1ULL << i; 346 347 // If all of our states become overdefined, bail out early. Since the 348 // predicate is expensive, only check it every 8 elements. This is only 349 // really useful for really huge arrays. 350 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && 351 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && 352 FalseRangeEnd == Overdefined) 353 return 0; 354 } 355 356 // Now that we've scanned the entire array, emit our new comparison(s). We 357 // order the state machines in complexity of the generated code. 358 Value *Idx = GEP->getOperand(2); 359 360 // If the index is larger than the pointer size of the target, truncate the 361 // index down like the GEP would do implicitly. We don't have to do this for 362 // an inbounds GEP because the index can't be out of range. 363 if (!GEP->isInBounds() && 364 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits()) 365 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext())); 366 367 // If the comparison is only true for one or two elements, emit direct 368 // comparisons. 369 if (SecondTrueElement != Overdefined) { 370 // None true -> false. 371 if (FirstTrueElement == Undefined) 372 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext())); 373 374 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement); 375 376 // True for one element -> 'i == 47'. 377 if (SecondTrueElement == Undefined) 378 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); 379 380 // True for two elements -> 'i == 47 | i == 72'. 381 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx); 382 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement); 383 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx); 384 return BinaryOperator::CreateOr(C1, C2); 385 } 386 387 // If the comparison is only false for one or two elements, emit direct 388 // comparisons. 389 if (SecondFalseElement != Overdefined) { 390 // None false -> true. 391 if (FirstFalseElement == Undefined) 392 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext())); 393 394 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement); 395 396 // False for one element -> 'i != 47'. 397 if (SecondFalseElement == Undefined) 398 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); 399 400 // False for two elements -> 'i != 47 & i != 72'. 401 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx); 402 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement); 403 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx); 404 return BinaryOperator::CreateAnd(C1, C2); 405 } 406 407 // If the comparison can be replaced with a range comparison for the elements 408 // where it is true, emit the range check. 409 if (TrueRangeEnd != Overdefined) { 410 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare"); 411 412 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1). 413 if (FirstTrueElement) { 414 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement); 415 Idx = Builder->CreateAdd(Idx, Offs); 416 } 417 418 Value *End = ConstantInt::get(Idx->getType(), 419 TrueRangeEnd-FirstTrueElement+1); 420 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); 421 } 422 423 // False range check. 424 if (FalseRangeEnd != Overdefined) { 425 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare"); 426 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). 427 if (FirstFalseElement) { 428 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement); 429 Idx = Builder->CreateAdd(Idx, Offs); 430 } 431 432 Value *End = ConstantInt::get(Idx->getType(), 433 FalseRangeEnd-FirstFalseElement); 434 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); 435 } 436 437 438 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state 439 // of this load, replace it with computation that does: 440 // ((magic_cst >> i) & 1) != 0 441 if (Init->getNumOperands() <= 32 || 442 (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) { 443 const Type *Ty; 444 if (Init->getNumOperands() <= 32) 445 Ty = Type::getInt32Ty(Init->getContext()); 446 else 447 Ty = Type::getInt64Ty(Init->getContext()); 448 Value *V = Builder->CreateIntCast(Idx, Ty, false); 449 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V); 450 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V); 451 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0)); 452 } 453 454 return 0; 455} 456 457 458/// EvaluateGEPOffsetExpression - Return a value that can be used to compare 459/// the *offset* implied by a GEP to zero. For example, if we have &A[i], we 460/// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can 461/// be complex, and scales are involved. The above expression would also be 462/// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). 463/// This later form is less amenable to optimization though, and we are allowed 464/// to generate the first by knowing that pointer arithmetic doesn't overflow. 465/// 466/// If we can't emit an optimized form for this expression, this returns null. 467/// 468static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I, 469 InstCombiner &IC) { 470 TargetData &TD = *IC.getTargetData(); 471 gep_type_iterator GTI = gep_type_begin(GEP); 472 473 // Check to see if this gep only has a single variable index. If so, and if 474 // any constant indices are a multiple of its scale, then we can compute this 475 // in terms of the scale of the variable index. For example, if the GEP 476 // implies an offset of "12 + i*4", then we can codegen this as "3 + i", 477 // because the expression will cross zero at the same point. 478 unsigned i, e = GEP->getNumOperands(); 479 int64_t Offset = 0; 480 for (i = 1; i != e; ++i, ++GTI) { 481 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 482 // Compute the aggregate offset of constant indices. 483 if (CI->isZero()) continue; 484 485 // Handle a struct index, which adds its field offset to the pointer. 486 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 487 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 488 } else { 489 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 490 Offset += Size*CI->getSExtValue(); 491 } 492 } else { 493 // Found our variable index. 494 break; 495 } 496 } 497 498 // If there are no variable indices, we must have a constant offset, just 499 // evaluate it the general way. 500 if (i == e) return 0; 501 502 Value *VariableIdx = GEP->getOperand(i); 503 // Determine the scale factor of the variable element. For example, this is 504 // 4 if the variable index is into an array of i32. 505 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType()); 506 507 // Verify that there are no other variable indices. If so, emit the hard way. 508 for (++i, ++GTI; i != e; ++i, ++GTI) { 509 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i)); 510 if (!CI) return 0; 511 512 // Compute the aggregate offset of constant indices. 513 if (CI->isZero()) continue; 514 515 // Handle a struct index, which adds its field offset to the pointer. 516 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 517 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 518 } else { 519 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 520 Offset += Size*CI->getSExtValue(); 521 } 522 } 523 524 // Okay, we know we have a single variable index, which must be a 525 // pointer/array/vector index. If there is no offset, life is simple, return 526 // the index. 527 unsigned IntPtrWidth = TD.getPointerSizeInBits(); 528 if (Offset == 0) { 529 // Cast to intptrty in case a truncation occurs. If an extension is needed, 530 // we don't need to bother extending: the extension won't affect where the 531 // computation crosses zero. 532 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) 533 VariableIdx = new TruncInst(VariableIdx, 534 TD.getIntPtrType(VariableIdx->getContext()), 535 VariableIdx->getName(), &I); 536 return VariableIdx; 537 } 538 539 // Otherwise, there is an index. The computation we will do will be modulo 540 // the pointer size, so get it. 541 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth); 542 543 Offset &= PtrSizeMask; 544 VariableScale &= PtrSizeMask; 545 546 // To do this transformation, any constant index must be a multiple of the 547 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i", 548 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a 549 // multiple of the variable scale. 550 int64_t NewOffs = Offset / (int64_t)VariableScale; 551 if (Offset != NewOffs*(int64_t)VariableScale) 552 return 0; 553 554 // Okay, we can do this evaluation. Start by converting the index to intptr. 555 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext()); 556 if (VariableIdx->getType() != IntPtrTy) 557 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy, 558 true /*SExt*/, 559 VariableIdx->getName(), &I); 560 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs); 561 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I); 562} 563 564/// FoldGEPICmp - Fold comparisons between a GEP instruction and something 565/// else. At this point we know that the GEP is on the LHS of the comparison. 566Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 567 ICmpInst::Predicate Cond, 568 Instruction &I) { 569 // Look through bitcasts. 570 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS)) 571 RHS = BCI->getOperand(0); 572 573 Value *PtrBase = GEPLHS->getOperand(0); 574 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) { 575 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). 576 // This transformation (ignoring the base and scales) is valid because we 577 // know pointers can't overflow since the gep is inbounds. See if we can 578 // output an optimized form. 579 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this); 580 581 // If not, synthesize the offset the hard way. 582 if (Offset == 0) 583 Offset = EmitGEPOffset(GEPLHS); 584 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset, 585 Constant::getNullValue(Offset->getType())); 586 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) { 587 // If the base pointers are different, but the indices are the same, just 588 // compare the base pointer. 589 if (PtrBase != GEPRHS->getOperand(0)) { 590 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands(); 591 IndicesTheSame &= GEPLHS->getOperand(0)->getType() == 592 GEPRHS->getOperand(0)->getType(); 593 if (IndicesTheSame) 594 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 595 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 596 IndicesTheSame = false; 597 break; 598 } 599 600 // If all indices are the same, just compare the base pointers. 601 if (IndicesTheSame) 602 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), 603 GEPLHS->getOperand(0), GEPRHS->getOperand(0)); 604 605 // Otherwise, the base pointers are different and the indices are 606 // different, bail out. 607 return 0; 608 } 609 610 // If one of the GEPs has all zero indices, recurse. 611 bool AllZeros = true; 612 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 613 if (!isa<Constant>(GEPLHS->getOperand(i)) || 614 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) { 615 AllZeros = false; 616 break; 617 } 618 if (AllZeros) 619 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0), 620 ICmpInst::getSwappedPredicate(Cond), I); 621 622 // If the other GEP has all zero indices, recurse. 623 AllZeros = true; 624 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 625 if (!isa<Constant>(GEPRHS->getOperand(i)) || 626 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) { 627 AllZeros = false; 628 break; 629 } 630 if (AllZeros) 631 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I); 632 633 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) { 634 // If the GEPs only differ by one index, compare it. 635 unsigned NumDifferences = 0; // Keep track of # differences. 636 unsigned DiffOperand = 0; // The operand that differs. 637 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 638 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 639 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() != 640 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) { 641 // Irreconcilable differences. 642 NumDifferences = 2; 643 break; 644 } else { 645 if (NumDifferences++) break; 646 DiffOperand = i; 647 } 648 } 649 650 if (NumDifferences == 0) // SAME GEP? 651 return ReplaceInstUsesWith(I, // No comparison is needed here. 652 ConstantInt::get(Type::getInt1Ty(I.getContext()), 653 ICmpInst::isTrueWhenEqual(Cond))); 654 655 else if (NumDifferences == 1) { 656 Value *LHSV = GEPLHS->getOperand(DiffOperand); 657 Value *RHSV = GEPRHS->getOperand(DiffOperand); 658 // Make sure we do a signed comparison here. 659 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV); 660 } 661 } 662 663 // Only lower this if the icmp is the only user of the GEP or if we expect 664 // the result to fold to a constant! 665 if (TD && 666 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) && 667 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) { 668 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) 669 Value *L = EmitGEPOffset(GEPLHS); 670 Value *R = EmitGEPOffset(GEPRHS); 671 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R); 672 } 673 } 674 return 0; 675} 676 677/// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X". 678Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI, 679 Value *X, ConstantInt *CI, 680 ICmpInst::Predicate Pred, 681 Value *TheAdd) { 682 // If we have X+0, exit early (simplifying logic below) and let it get folded 683 // elsewhere. icmp X+0, X -> icmp X, X 684 if (CI->isZero()) { 685 bool isTrue = ICmpInst::isTrueWhenEqual(Pred); 686 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 687 } 688 689 // (X+4) == X -> false. 690 if (Pred == ICmpInst::ICMP_EQ) 691 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext())); 692 693 // (X+4) != X -> true. 694 if (Pred == ICmpInst::ICMP_NE) 695 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext())); 696 697 // If this is an instruction (as opposed to constantexpr) get NUW/NSW info. 698 bool isNUW = false, isNSW = false; 699 if (BinaryOperator *Add = dyn_cast<BinaryOperator>(TheAdd)) { 700 isNUW = Add->hasNoUnsignedWrap(); 701 isNSW = Add->hasNoSignedWrap(); 702 } 703 704 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, 705 // so the values can never be equal. Similiarly for all other "or equals" 706 // operators. 707 708 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255 709 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253 710 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0 711 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 712 // If this is an NUW add, then this is always false. 713 if (isNUW) 714 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext())); 715 716 Value *R = 717 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI); 718 return new ICmpInst(ICmpInst::ICMP_UGT, X, R); 719 } 720 721 // (X+1) >u X --> X <u (0-1) --> X != 255 722 // (X+2) >u X --> X <u (0-2) --> X <u 254 723 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0 724 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { 725 // If this is an NUW add, then this is always true. 726 if (isNUW) 727 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext())); 728 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI)); 729 } 730 731 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits(); 732 ConstantInt *SMax = ConstantInt::get(X->getContext(), 733 APInt::getSignedMaxValue(BitWidth)); 734 735 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127 736 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125 737 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0 738 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1 739 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126 740 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127 741 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) { 742 // If this is an NSW add, then we have two cases: if the constant is 743 // positive, then this is always false, if negative, this is always true. 744 if (isNSW) { 745 bool isTrue = CI->getValue().isNegative(); 746 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 747 } 748 749 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI)); 750 } 751 752 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127 753 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126 754 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1 755 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2 756 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126 757 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128 758 759 // If this is an NSW add, then we have two cases: if the constant is 760 // positive, then this is always true, if negative, this is always false. 761 if (isNSW) { 762 bool isTrue = !CI->getValue().isNegative(); 763 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 764 } 765 766 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); 767 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1); 768 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C)); 769} 770 771/// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS 772/// and CmpRHS are both known to be integer constants. 773Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI, 774 ConstantInt *DivRHS) { 775 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1)); 776 const APInt &CmpRHSV = CmpRHS->getValue(); 777 778 // FIXME: If the operand types don't match the type of the divide 779 // then don't attempt this transform. The code below doesn't have the 780 // logic to deal with a signed divide and an unsigned compare (and 781 // vice versa). This is because (x /s C1) <s C2 produces different 782 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even 783 // (x /u C1) <u C2. Simply casting the operands and result won't 784 // work. :( The if statement below tests that condition and bails 785 // if it finds it. 786 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv; 787 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned()) 788 return 0; 789 if (DivRHS->isZero()) 790 return 0; // The ProdOV computation fails on divide by zero. 791 if (DivIsSigned && DivRHS->isAllOnesValue()) 792 return 0; // The overflow computation also screws up here 793 if (DivRHS->isOne()) 794 return 0; // Not worth bothering, and eliminates some funny cases 795 // with INT_MIN. 796 797 // Compute Prod = CI * DivRHS. We are essentially solving an equation 798 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and 799 // C2 (CI). By solving for X we can turn this into a range check 800 // instead of computing a divide. 801 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS); 802 803 // Determine if the product overflows by seeing if the product is 804 // not equal to the divide. Make sure we do the same kind of divide 805 // as in the LHS instruction that we're folding. 806 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) : 807 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS; 808 809 // Get the ICmp opcode 810 ICmpInst::Predicate Pred = ICI.getPredicate(); 811 812 // Figure out the interval that is being checked. For example, a comparison 813 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). 814 // Compute this interval based on the constants involved and the signedness of 815 // the compare/divide. This computes a half-open interval, keeping track of 816 // whether either value in the interval overflows. After analysis each 817 // overflow variable is set to 0 if it's corresponding bound variable is valid 818 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. 819 int LoOverflow = 0, HiOverflow = 0; 820 Constant *LoBound = 0, *HiBound = 0; 821 822 if (!DivIsSigned) { // udiv 823 // e.g. X/5 op 3 --> [15, 20) 824 LoBound = Prod; 825 HiOverflow = LoOverflow = ProdOV; 826 if (!HiOverflow) 827 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false); 828 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0. 829 if (CmpRHSV == 0) { // (X / pos) op 0 830 // Can't overflow. e.g. X/2 op 0 --> [-1, 2) 831 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS))); 832 HiBound = DivRHS; 833 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos 834 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) 835 HiOverflow = LoOverflow = ProdOV; 836 if (!HiOverflow) 837 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true); 838 } else { // (X / pos) op neg 839 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) 840 HiBound = AddOne(Prod); 841 LoOverflow = HiOverflow = ProdOV ? -1 : 0; 842 if (!LoOverflow) { 843 ConstantInt* DivNeg = 844 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS)); 845 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0; 846 } 847 } 848 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0. 849 if (CmpRHSV == 0) { // (X / neg) op 0 850 // e.g. X/-5 op 0 --> [-4, 5) 851 LoBound = AddOne(DivRHS); 852 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS)); 853 if (HiBound == DivRHS) { // -INTMIN = INTMIN 854 HiOverflow = 1; // [INTMIN+1, overflow) 855 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN 856 } 857 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos 858 // e.g. X/-5 op 3 --> [-19, -14) 859 HiBound = AddOne(Prod); 860 HiOverflow = LoOverflow = ProdOV ? -1 : 0; 861 if (!LoOverflow) 862 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0; 863 } else { // (X / neg) op neg 864 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) 865 LoOverflow = HiOverflow = ProdOV; 866 if (!HiOverflow) 867 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true); 868 } 869 870 // Dividing by a negative swaps the condition. LT <-> GT 871 Pred = ICmpInst::getSwappedPredicate(Pred); 872 } 873 874 Value *X = DivI->getOperand(0); 875 switch (Pred) { 876 default: llvm_unreachable("Unhandled icmp opcode!"); 877 case ICmpInst::ICMP_EQ: 878 if (LoOverflow && HiOverflow) 879 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 880 else if (HiOverflow) 881 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 882 ICmpInst::ICMP_UGE, X, LoBound); 883 else if (LoOverflow) 884 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 885 ICmpInst::ICMP_ULT, X, HiBound); 886 else 887 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI); 888 case ICmpInst::ICMP_NE: 889 if (LoOverflow && HiOverflow) 890 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 891 else if (HiOverflow) 892 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 893 ICmpInst::ICMP_ULT, X, LoBound); 894 else if (LoOverflow) 895 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 896 ICmpInst::ICMP_UGE, X, HiBound); 897 else 898 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI); 899 case ICmpInst::ICMP_ULT: 900 case ICmpInst::ICMP_SLT: 901 if (LoOverflow == +1) // Low bound is greater than input range. 902 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 903 if (LoOverflow == -1) // Low bound is less than input range. 904 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 905 return new ICmpInst(Pred, X, LoBound); 906 case ICmpInst::ICMP_UGT: 907 case ICmpInst::ICMP_SGT: 908 if (HiOverflow == +1) // High bound greater than input range. 909 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 910 else if (HiOverflow == -1) // High bound less than input range. 911 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 912 if (Pred == ICmpInst::ICMP_UGT) 913 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound); 914 else 915 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound); 916 } 917} 918 919 920/// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)". 921/// 922Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI, 923 Instruction *LHSI, 924 ConstantInt *RHS) { 925 const APInt &RHSV = RHS->getValue(); 926 927 switch (LHSI->getOpcode()) { 928 case Instruction::Trunc: 929 if (ICI.isEquality() && LHSI->hasOneUse()) { 930 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all 931 // of the high bits truncated out of x are known. 932 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(), 933 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits(); 934 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits)); 935 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0); 936 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne); 937 938 // If all the high bits are known, we can do this xform. 939 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) { 940 // Pull in the high bits from known-ones set. 941 APInt NewRHS(RHS->getValue()); 942 NewRHS.zext(SrcBits); 943 NewRHS |= KnownOne; 944 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 945 ConstantInt::get(ICI.getContext(), NewRHS)); 946 } 947 } 948 break; 949 950 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI) 951 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 952 // If this is a comparison that tests the signbit (X < 0) or (x > -1), 953 // fold the xor. 954 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) || 955 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) { 956 Value *CompareVal = LHSI->getOperand(0); 957 958 // If the sign bit of the XorCST is not set, there is no change to 959 // the operation, just stop using the Xor. 960 if (!XorCST->getValue().isNegative()) { 961 ICI.setOperand(0, CompareVal); 962 Worklist.Add(LHSI); 963 return &ICI; 964 } 965 966 // Was the old condition true if the operand is positive? 967 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT; 968 969 // If so, the new one isn't. 970 isTrueIfPositive ^= true; 971 972 if (isTrueIfPositive) 973 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, 974 SubOne(RHS)); 975 else 976 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, 977 AddOne(RHS)); 978 } 979 980 if (LHSI->hasOneUse()) { 981 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit)) 982 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) { 983 const APInt &SignBit = XorCST->getValue(); 984 ICmpInst::Predicate Pred = ICI.isSigned() 985 ? ICI.getUnsignedPredicate() 986 : ICI.getSignedPredicate(); 987 return new ICmpInst(Pred, LHSI->getOperand(0), 988 ConstantInt::get(ICI.getContext(), 989 RHSV ^ SignBit)); 990 } 991 992 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A) 993 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) { 994 const APInt &NotSignBit = XorCST->getValue(); 995 ICmpInst::Predicate Pred = ICI.isSigned() 996 ? ICI.getUnsignedPredicate() 997 : ICI.getSignedPredicate(); 998 Pred = ICI.getSwappedPredicate(Pred); 999 return new ICmpInst(Pred, LHSI->getOperand(0), 1000 ConstantInt::get(ICI.getContext(), 1001 RHSV ^ NotSignBit)); 1002 } 1003 } 1004 } 1005 break; 1006 case Instruction::And: // (icmp pred (and X, AndCST), RHS) 1007 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) && 1008 LHSI->getOperand(0)->hasOneUse()) { 1009 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1)); 1010 1011 // If the LHS is an AND of a truncating cast, we can widen the 1012 // and/compare to be the input width without changing the value 1013 // produced, eliminating a cast. 1014 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) { 1015 // We can do this transformation if either the AND constant does not 1016 // have its sign bit set or if it is an equality comparison. 1017 // Extending a relational comparison when we're checking the sign 1018 // bit would not work. 1019 if (Cast->hasOneUse() && 1020 (ICI.isEquality() || 1021 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) { 1022 uint32_t BitWidth = 1023 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth(); 1024 APInt NewCST = AndCST->getValue(); 1025 NewCST.zext(BitWidth); 1026 APInt NewCI = RHSV; 1027 NewCI.zext(BitWidth); 1028 Value *NewAnd = 1029 Builder->CreateAnd(Cast->getOperand(0), 1030 ConstantInt::get(ICI.getContext(), NewCST), 1031 LHSI->getName()); 1032 return new ICmpInst(ICI.getPredicate(), NewAnd, 1033 ConstantInt::get(ICI.getContext(), NewCI)); 1034 } 1035 } 1036 1037 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare 1038 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This 1039 // happens a LOT in code produced by the C front-end, for bitfield 1040 // access. 1041 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0)); 1042 if (Shift && !Shift->isShift()) 1043 Shift = 0; 1044 1045 ConstantInt *ShAmt; 1046 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0; 1047 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift. 1048 const Type *AndTy = AndCST->getType(); // Type of the and. 1049 1050 // We can fold this as long as we can't shift unknown bits 1051 // into the mask. This can only happen with signed shift 1052 // rights, as they sign-extend. 1053 if (ShAmt) { 1054 bool CanFold = Shift->isLogicalShift(); 1055 if (!CanFold) { 1056 // To test for the bad case of the signed shr, see if any 1057 // of the bits shifted in could be tested after the mask. 1058 uint32_t TyBits = Ty->getPrimitiveSizeInBits(); 1059 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits); 1060 1061 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits(); 1062 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) & 1063 AndCST->getValue()) == 0) 1064 CanFold = true; 1065 } 1066 1067 if (CanFold) { 1068 Constant *NewCst; 1069 if (Shift->getOpcode() == Instruction::Shl) 1070 NewCst = ConstantExpr::getLShr(RHS, ShAmt); 1071 else 1072 NewCst = ConstantExpr::getShl(RHS, ShAmt); 1073 1074 // Check to see if we are shifting out any of the bits being 1075 // compared. 1076 if (ConstantExpr::get(Shift->getOpcode(), 1077 NewCst, ShAmt) != RHS) { 1078 // If we shifted bits out, the fold is not going to work out. 1079 // As a special case, check to see if this means that the 1080 // result is always true or false now. 1081 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1082 return ReplaceInstUsesWith(ICI, 1083 ConstantInt::getFalse(ICI.getContext())); 1084 if (ICI.getPredicate() == ICmpInst::ICMP_NE) 1085 return ReplaceInstUsesWith(ICI, 1086 ConstantInt::getTrue(ICI.getContext())); 1087 } else { 1088 ICI.setOperand(1, NewCst); 1089 Constant *NewAndCST; 1090 if (Shift->getOpcode() == Instruction::Shl) 1091 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt); 1092 else 1093 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt); 1094 LHSI->setOperand(1, NewAndCST); 1095 LHSI->setOperand(0, Shift->getOperand(0)); 1096 Worklist.Add(Shift); // Shift is dead. 1097 return &ICI; 1098 } 1099 } 1100 } 1101 1102 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is 1103 // preferable because it allows the C<<Y expression to be hoisted out 1104 // of a loop if Y is invariant and X is not. 1105 if (Shift && Shift->hasOneUse() && RHSV == 0 && 1106 ICI.isEquality() && !Shift->isArithmeticShift() && 1107 !isa<Constant>(Shift->getOperand(0))) { 1108 // Compute C << Y. 1109 Value *NS; 1110 if (Shift->getOpcode() == Instruction::LShr) { 1111 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp"); 1112 } else { 1113 // Insert a logical shift. 1114 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp"); 1115 } 1116 1117 // Compute X & (C << Y). 1118 Value *NewAnd = 1119 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName()); 1120 1121 ICI.setOperand(0, NewAnd); 1122 return &ICI; 1123 } 1124 } 1125 1126 // Try to optimize things like "A[i]&42 == 0" to index computations. 1127 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) { 1128 if (GetElementPtrInst *GEP = 1129 dyn_cast<GetElementPtrInst>(LI->getOperand(0))) 1130 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 1131 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 1132 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) { 1133 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1)); 1134 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C)) 1135 return Res; 1136 } 1137 } 1138 break; 1139 1140 case Instruction::Or: { 1141 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse()) 1142 break; 1143 Value *P, *Q; 1144 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) { 1145 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 1146 // -> and (icmp eq P, null), (icmp eq Q, null). 1147 1148 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P, 1149 Constant::getNullValue(P->getType())); 1150 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q, 1151 Constant::getNullValue(Q->getType())); 1152 Instruction *Op; 1153 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1154 Op = BinaryOperator::CreateAnd(ICIP, ICIQ); 1155 else 1156 Op = BinaryOperator::CreateOr(ICIP, ICIQ); 1157 return Op; 1158 } 1159 break; 1160 } 1161 1162 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI) 1163 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1164 if (!ShAmt) break; 1165 1166 uint32_t TypeBits = RHSV.getBitWidth(); 1167 1168 // Check that the shift amount is in range. If not, don't perform 1169 // undefined shifts. When the shift is visited it will be 1170 // simplified. 1171 if (ShAmt->uge(TypeBits)) 1172 break; 1173 1174 if (ICI.isEquality()) { 1175 // If we are comparing against bits always shifted out, the 1176 // comparison cannot succeed. 1177 Constant *Comp = 1178 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), 1179 ShAmt); 1180 if (Comp != RHS) {// Comparing against a bit that we know is zero. 1181 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1182 Constant *Cst = 1183 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE); 1184 return ReplaceInstUsesWith(ICI, Cst); 1185 } 1186 1187 if (LHSI->hasOneUse()) { 1188 // Otherwise strength reduce the shift into an and. 1189 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 1190 Constant *Mask = 1191 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits, 1192 TypeBits-ShAmtVal)); 1193 1194 Value *And = 1195 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask"); 1196 return new ICmpInst(ICI.getPredicate(), And, 1197 ConstantInt::get(ICI.getContext(), 1198 RHSV.lshr(ShAmtVal))); 1199 } 1200 } 1201 1202 // Otherwise, if this is a comparison of the sign bit, simplify to and/test. 1203 bool TrueIfSigned = false; 1204 if (LHSI->hasOneUse() && 1205 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) { 1206 // (X << 31) <s 0 --> (X&1) != 0 1207 Constant *Mask = ConstantInt::get(ICI.getContext(), APInt(TypeBits, 1) << 1208 (TypeBits-ShAmt->getZExtValue()-1)); 1209 Value *And = 1210 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask"); 1211 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, 1212 And, Constant::getNullValue(And->getType())); 1213 } 1214 break; 1215 } 1216 1217 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI) 1218 case Instruction::AShr: { 1219 // Only handle equality comparisons of shift-by-constant. 1220 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1221 if (!ShAmt || !ICI.isEquality()) break; 1222 1223 // Check that the shift amount is in range. If not, don't perform 1224 // undefined shifts. When the shift is visited it will be 1225 // simplified. 1226 uint32_t TypeBits = RHSV.getBitWidth(); 1227 if (ShAmt->uge(TypeBits)) 1228 break; 1229 1230 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 1231 1232 // If we are comparing against bits always shifted out, the 1233 // comparison cannot succeed. 1234 APInt Comp = RHSV << ShAmtVal; 1235 if (LHSI->getOpcode() == Instruction::LShr) 1236 Comp = Comp.lshr(ShAmtVal); 1237 else 1238 Comp = Comp.ashr(ShAmtVal); 1239 1240 if (Comp != RHSV) { // Comparing against a bit that we know is zero. 1241 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1242 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1243 IsICMP_NE); 1244 return ReplaceInstUsesWith(ICI, Cst); 1245 } 1246 1247 // Otherwise, check to see if the bits shifted out are known to be zero. 1248 // If so, we can compare against the unshifted value: 1249 // (X & 4) >> 1 == 2 --> (X & 4) == 4. 1250 if (LHSI->hasOneUse() && 1251 MaskedValueIsZero(LHSI->getOperand(0), 1252 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) { 1253 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1254 ConstantExpr::getShl(RHS, ShAmt)); 1255 } 1256 1257 if (LHSI->hasOneUse()) { 1258 // Otherwise strength reduce the shift into an and. 1259 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal)); 1260 Constant *Mask = ConstantInt::get(ICI.getContext(), Val); 1261 1262 Value *And = Builder->CreateAnd(LHSI->getOperand(0), 1263 Mask, LHSI->getName()+".mask"); 1264 return new ICmpInst(ICI.getPredicate(), And, 1265 ConstantExpr::getShl(RHS, ShAmt)); 1266 } 1267 break; 1268 } 1269 1270 case Instruction::SDiv: 1271 case Instruction::UDiv: 1272 // Fold: icmp pred ([us]div X, C1), C2 -> range test 1273 // Fold this div into the comparison, producing a range check. 1274 // Determine, based on the divide type, what the range is being 1275 // checked. If there is an overflow on the low or high side, remember 1276 // it, otherwise compute the range [low, hi) bounding the new value. 1277 // See: InsertRangeTest above for the kinds of replacements possible. 1278 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) 1279 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI), 1280 DivRHS)) 1281 return R; 1282 break; 1283 1284 case Instruction::Add: 1285 // Fold: icmp pred (add X, C1), C2 1286 if (!ICI.isEquality()) { 1287 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1288 if (!LHSC) break; 1289 const APInt &LHSV = LHSC->getValue(); 1290 1291 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV) 1292 .subtract(LHSV); 1293 1294 if (ICI.isSigned()) { 1295 if (CR.getLower().isSignBit()) { 1296 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0), 1297 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1298 } else if (CR.getUpper().isSignBit()) { 1299 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0), 1300 ConstantInt::get(ICI.getContext(),CR.getLower())); 1301 } 1302 } else { 1303 if (CR.getLower().isMinValue()) { 1304 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), 1305 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1306 } else if (CR.getUpper().isMinValue()) { 1307 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), 1308 ConstantInt::get(ICI.getContext(),CR.getLower())); 1309 } 1310 } 1311 } 1312 break; 1313 } 1314 1315 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS. 1316 if (ICI.isEquality()) { 1317 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1318 1319 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and 1320 // the second operand is a constant, simplify a bit. 1321 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) { 1322 switch (BO->getOpcode()) { 1323 case Instruction::SRem: 1324 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. 1325 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){ 1326 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue(); 1327 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) { 1328 Value *NewRem = 1329 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1), 1330 BO->getName()); 1331 return new ICmpInst(ICI.getPredicate(), NewRem, 1332 Constant::getNullValue(BO->getType())); 1333 } 1334 } 1335 break; 1336 case Instruction::Add: 1337 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants. 1338 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1339 if (BO->hasOneUse()) 1340 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1341 ConstantExpr::getSub(RHS, BOp1C)); 1342 } else if (RHSV == 0) { 1343 // Replace ((add A, B) != 0) with (A != -B) if A or B is 1344 // efficiently invertible, or if the add has just this one use. 1345 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1); 1346 1347 if (Value *NegVal = dyn_castNegVal(BOp1)) 1348 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal); 1349 else if (Value *NegVal = dyn_castNegVal(BOp0)) 1350 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1); 1351 else if (BO->hasOneUse()) { 1352 Value *Neg = Builder->CreateNeg(BOp1); 1353 Neg->takeName(BO); 1354 return new ICmpInst(ICI.getPredicate(), BOp0, Neg); 1355 } 1356 } 1357 break; 1358 case Instruction::Xor: 1359 // For the xor case, we can xor two constants together, eliminating 1360 // the explicit xor. 1361 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) 1362 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1363 ConstantExpr::getXor(RHS, BOC)); 1364 1365 // FALLTHROUGH 1366 case Instruction::Sub: 1367 // Replace (([sub|xor] A, B) != 0) with (A != B) 1368 if (RHSV == 0) 1369 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1370 BO->getOperand(1)); 1371 break; 1372 1373 case Instruction::Or: 1374 // If bits are being or'd in that are not present in the constant we 1375 // are comparing against, then the comparison could never succeed! 1376 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) { 1377 Constant *NotCI = ConstantExpr::getNot(RHS); 1378 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue()) 1379 return ReplaceInstUsesWith(ICI, 1380 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1381 isICMP_NE)); 1382 } 1383 break; 1384 1385 case Instruction::And: 1386 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1387 // If bits are being compared against that are and'd out, then the 1388 // comparison can never succeed! 1389 if ((RHSV & ~BOC->getValue()) != 0) 1390 return ReplaceInstUsesWith(ICI, 1391 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1392 isICMP_NE)); 1393 1394 // If we have ((X & C) == C), turn it into ((X & C) != 0). 1395 if (RHS == BOC && RHSV.isPowerOf2()) 1396 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : 1397 ICmpInst::ICMP_NE, LHSI, 1398 Constant::getNullValue(RHS->getType())); 1399 1400 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0 1401 if (BOC->getValue().isSignBit()) { 1402 Value *X = BO->getOperand(0); 1403 Constant *Zero = Constant::getNullValue(X->getType()); 1404 ICmpInst::Predicate pred = isICMP_NE ? 1405 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; 1406 return new ICmpInst(pred, X, Zero); 1407 } 1408 1409 // ((X & ~7) == 0) --> X < 8 1410 if (RHSV == 0 && isHighOnes(BOC)) { 1411 Value *X = BO->getOperand(0); 1412 Constant *NegX = ConstantExpr::getNeg(BOC); 1413 ICmpInst::Predicate pred = isICMP_NE ? 1414 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; 1415 return new ICmpInst(pred, X, NegX); 1416 } 1417 } 1418 default: break; 1419 } 1420 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) { 1421 // Handle icmp {eq|ne} <intrinsic>, intcst. 1422 switch (II->getIntrinsicID()) { 1423 case Intrinsic::bswap: 1424 Worklist.Add(II); 1425 ICI.setOperand(0, II->getOperand(1)); 1426 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap())); 1427 return &ICI; 1428 case Intrinsic::ctlz: 1429 case Intrinsic::cttz: 1430 // ctz(A) == bitwidth(a) -> A == 0 and likewise for != 1431 if (RHSV == RHS->getType()->getBitWidth()) { 1432 Worklist.Add(II); 1433 ICI.setOperand(0, II->getOperand(1)); 1434 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0)); 1435 return &ICI; 1436 } 1437 break; 1438 case Intrinsic::ctpop: 1439 // popcount(A) == 0 -> A == 0 and likewise for != 1440 if (RHS->isZero()) { 1441 Worklist.Add(II); 1442 ICI.setOperand(0, II->getOperand(1)); 1443 ICI.setOperand(1, RHS); 1444 return &ICI; 1445 } 1446 break; 1447 default: 1448 break; 1449 } 1450 } 1451 } 1452 return 0; 1453} 1454 1455/// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst). 1456/// We only handle extending casts so far. 1457/// 1458Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) { 1459 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0)); 1460 Value *LHSCIOp = LHSCI->getOperand(0); 1461 const Type *SrcTy = LHSCIOp->getType(); 1462 const Type *DestTy = LHSCI->getType(); 1463 Value *RHSCIOp; 1464 1465 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the 1466 // integer type is the same size as the pointer type. 1467 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt && 1468 TD->getPointerSizeInBits() == 1469 cast<IntegerType>(DestTy)->getBitWidth()) { 1470 Value *RHSOp = 0; 1471 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) { 1472 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy); 1473 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) { 1474 RHSOp = RHSC->getOperand(0); 1475 // If the pointer types don't match, insert a bitcast. 1476 if (LHSCIOp->getType() != RHSOp->getType()) 1477 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType()); 1478 } 1479 1480 if (RHSOp) 1481 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp); 1482 } 1483 1484 // The code below only handles extension cast instructions, so far. 1485 // Enforce this. 1486 if (LHSCI->getOpcode() != Instruction::ZExt && 1487 LHSCI->getOpcode() != Instruction::SExt) 1488 return 0; 1489 1490 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt; 1491 bool isSignedCmp = ICI.isSigned(); 1492 1493 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) { 1494 // Not an extension from the same type? 1495 RHSCIOp = CI->getOperand(0); 1496 if (RHSCIOp->getType() != LHSCIOp->getType()) 1497 return 0; 1498 1499 // If the signedness of the two casts doesn't agree (i.e. one is a sext 1500 // and the other is a zext), then we can't handle this. 1501 if (CI->getOpcode() != LHSCI->getOpcode()) 1502 return 0; 1503 1504 // Deal with equality cases early. 1505 if (ICI.isEquality()) 1506 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1507 1508 // A signed comparison of sign extended values simplifies into a 1509 // signed comparison. 1510 if (isSignedCmp && isSignedExt) 1511 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1512 1513 // The other three cases all fold into an unsigned comparison. 1514 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp); 1515 } 1516 1517 // If we aren't dealing with a constant on the RHS, exit early 1518 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1)); 1519 if (!CI) 1520 return 0; 1521 1522 // Compute the constant that would happen if we truncated to SrcTy then 1523 // reextended to DestTy. 1524 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy); 1525 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), 1526 Res1, DestTy); 1527 1528 // If the re-extended constant didn't change... 1529 if (Res2 == CI) { 1530 // Deal with equality cases early. 1531 if (ICI.isEquality()) 1532 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1533 1534 // A signed comparison of sign extended values simplifies into a 1535 // signed comparison. 1536 if (isSignedExt && isSignedCmp) 1537 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1538 1539 // The other three cases all fold into an unsigned comparison. 1540 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1); 1541 } 1542 1543 // The re-extended constant changed so the constant cannot be represented 1544 // in the shorter type. Consequently, we cannot emit a simple comparison. 1545 1546 // First, handle some easy cases. We know the result cannot be equal at this 1547 // point so handle the ICI.isEquality() cases 1548 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1549 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 1550 if (ICI.getPredicate() == ICmpInst::ICMP_NE) 1551 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 1552 1553 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases 1554 // should have been folded away previously and not enter in here. 1555 Value *Result; 1556 if (isSignedCmp) { 1557 // We're performing a signed comparison. 1558 if (cast<ConstantInt>(CI)->getValue().isNegative()) 1559 Result = ConstantInt::getFalse(ICI.getContext()); // X < (small) --> false 1560 else 1561 Result = ConstantInt::getTrue(ICI.getContext()); // X < (large) --> true 1562 } else { 1563 // We're performing an unsigned comparison. 1564 if (isSignedExt) { 1565 // We're performing an unsigned comp with a sign extended value. 1566 // This is true if the input is >= 0. [aka >s -1] 1567 Constant *NegOne = Constant::getAllOnesValue(SrcTy); 1568 Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName()); 1569 } else { 1570 // Unsigned extend & unsigned compare -> always true. 1571 Result = ConstantInt::getTrue(ICI.getContext()); 1572 } 1573 } 1574 1575 // Finally, return the value computed. 1576 if (ICI.getPredicate() == ICmpInst::ICMP_ULT || 1577 ICI.getPredicate() == ICmpInst::ICMP_SLT) 1578 return ReplaceInstUsesWith(ICI, Result); 1579 1580 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT || 1581 ICI.getPredicate()==ICmpInst::ICMP_SGT) && 1582 "ICmp should be folded!"); 1583 if (Constant *CI = dyn_cast<Constant>(Result)) 1584 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI)); 1585 return BinaryOperator::CreateNot(Result); 1586} 1587 1588 1589 1590Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { 1591 bool Changed = false; 1592 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1593 1594 /// Orders the operands of the compare so that they are listed from most 1595 /// complex to least complex. This puts constants before unary operators, 1596 /// before binary operators. 1597 if (getComplexity(Op0) < getComplexity(Op1)) { 1598 I.swapOperands(); 1599 std::swap(Op0, Op1); 1600 Changed = true; 1601 } 1602 1603 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD)) 1604 return ReplaceInstUsesWith(I, V); 1605 1606 const Type *Ty = Op0->getType(); 1607 1608 // icmp's with boolean values can always be turned into bitwise operations 1609 if (Ty->isIntegerTy(1)) { 1610 switch (I.getPredicate()) { 1611 default: llvm_unreachable("Invalid icmp instruction!"); 1612 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B) 1613 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp"); 1614 return BinaryOperator::CreateNot(Xor); 1615 } 1616 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B 1617 return BinaryOperator::CreateXor(Op0, Op1); 1618 1619 case ICmpInst::ICMP_UGT: 1620 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult 1621 // FALL THROUGH 1622 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B 1623 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1624 return BinaryOperator::CreateAnd(Not, Op1); 1625 } 1626 case ICmpInst::ICMP_SGT: 1627 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt 1628 // FALL THROUGH 1629 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B 1630 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1631 return BinaryOperator::CreateAnd(Not, Op0); 1632 } 1633 case ICmpInst::ICMP_UGE: 1634 std::swap(Op0, Op1); // Change icmp uge -> icmp ule 1635 // FALL THROUGH 1636 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B 1637 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1638 return BinaryOperator::CreateOr(Not, Op1); 1639 } 1640 case ICmpInst::ICMP_SGE: 1641 std::swap(Op0, Op1); // Change icmp sge -> icmp sle 1642 // FALL THROUGH 1643 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B 1644 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1645 return BinaryOperator::CreateOr(Not, Op0); 1646 } 1647 } 1648 } 1649 1650 unsigned BitWidth = 0; 1651 if (TD) 1652 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType()); 1653 else if (Ty->isIntOrIntVectorTy()) 1654 BitWidth = Ty->getScalarSizeInBits(); 1655 1656 bool isSignBit = false; 1657 1658 // See if we are doing a comparison with a constant. 1659 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 1660 Value *A = 0, *B = 0; 1661 1662 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B) 1663 if (I.isEquality() && CI->isZero() && 1664 match(Op0, m_Sub(m_Value(A), m_Value(B)))) { 1665 // (icmp cond A B) if cond is equality 1666 return new ICmpInst(I.getPredicate(), A, B); 1667 } 1668 1669 // If we have an icmp le or icmp ge instruction, turn it into the 1670 // appropriate icmp lt or icmp gt instruction. This allows us to rely on 1671 // them being folded in the code below. The SimplifyICmpInst code has 1672 // already handled the edge cases for us, so we just assert on them. 1673 switch (I.getPredicate()) { 1674 default: break; 1675 case ICmpInst::ICMP_ULE: 1676 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE 1677 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, 1678 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1679 case ICmpInst::ICMP_SLE: 1680 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE 1681 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 1682 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1683 case ICmpInst::ICMP_UGE: 1684 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE 1685 return new ICmpInst(ICmpInst::ICMP_UGT, Op0, 1686 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1687 case ICmpInst::ICMP_SGE: 1688 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE 1689 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 1690 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1691 } 1692 1693 // If this comparison is a normal comparison, it demands all 1694 // bits, if it is a sign bit comparison, it only demands the sign bit. 1695 bool UnusedBit; 1696 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit); 1697 } 1698 1699 // See if we can fold the comparison based on range information we can get 1700 // by checking whether bits are known to be zero or one in the input. 1701 if (BitWidth != 0) { 1702 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0); 1703 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0); 1704 1705 if (SimplifyDemandedBits(I.getOperandUse(0), 1706 isSignBit ? APInt::getSignBit(BitWidth) 1707 : APInt::getAllOnesValue(BitWidth), 1708 Op0KnownZero, Op0KnownOne, 0)) 1709 return &I; 1710 if (SimplifyDemandedBits(I.getOperandUse(1), 1711 APInt::getAllOnesValue(BitWidth), 1712 Op1KnownZero, Op1KnownOne, 0)) 1713 return &I; 1714 1715 // Given the known and unknown bits, compute a range that the LHS could be 1716 // in. Compute the Min, Max and RHS values based on the known bits. For the 1717 // EQ and NE we use unsigned values. 1718 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0); 1719 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0); 1720 if (I.isSigned()) { 1721 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 1722 Op0Min, Op0Max); 1723 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 1724 Op1Min, Op1Max); 1725 } else { 1726 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 1727 Op0Min, Op0Max); 1728 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 1729 Op1Min, Op1Max); 1730 } 1731 1732 // If Min and Max are known to be the same, then SimplifyDemandedBits 1733 // figured out that the LHS is a constant. Just constant fold this now so 1734 // that code below can assume that Min != Max. 1735 if (!isa<Constant>(Op0) && Op0Min == Op0Max) 1736 return new ICmpInst(I.getPredicate(), 1737 ConstantInt::get(I.getContext(), Op0Min), Op1); 1738 if (!isa<Constant>(Op1) && Op1Min == Op1Max) 1739 return new ICmpInst(I.getPredicate(), Op0, 1740 ConstantInt::get(I.getContext(), Op1Min)); 1741 1742 // Based on the range information we know about the LHS, see if we can 1743 // simplify this comparison. For example, (x&4) < 8 is always true. 1744 switch (I.getPredicate()) { 1745 default: llvm_unreachable("Unknown icmp opcode!"); 1746 case ICmpInst::ICMP_EQ: 1747 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 1748 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1749 break; 1750 case ICmpInst::ICMP_NE: 1751 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 1752 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1753 break; 1754 case ICmpInst::ICMP_ULT: 1755 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B) 1756 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1757 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B) 1758 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1759 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B) 1760 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 1761 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 1762 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C 1763 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 1764 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1765 1766 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear 1767 if (CI->isMinValue(true)) 1768 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 1769 Constant::getAllOnesValue(Op0->getType())); 1770 } 1771 break; 1772 case ICmpInst::ICMP_UGT: 1773 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B) 1774 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1775 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B) 1776 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1777 1778 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B) 1779 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 1780 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 1781 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C 1782 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 1783 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1784 1785 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set 1786 if (CI->isMaxValue(true)) 1787 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 1788 Constant::getNullValue(Op0->getType())); 1789 } 1790 break; 1791 case ICmpInst::ICMP_SLT: 1792 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C) 1793 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1794 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C) 1795 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1796 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B) 1797 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 1798 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 1799 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C 1800 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 1801 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1802 } 1803 break; 1804 case ICmpInst::ICMP_SGT: 1805 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B) 1806 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1807 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B) 1808 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1809 1810 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B) 1811 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 1812 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 1813 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C 1814 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 1815 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1816 } 1817 break; 1818 case ICmpInst::ICMP_SGE: 1819 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"); 1820 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B) 1821 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1822 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B) 1823 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1824 break; 1825 case ICmpInst::ICMP_SLE: 1826 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"); 1827 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B) 1828 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1829 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B) 1830 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1831 break; 1832 case ICmpInst::ICMP_UGE: 1833 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"); 1834 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B) 1835 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1836 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B) 1837 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1838 break; 1839 case ICmpInst::ICMP_ULE: 1840 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"); 1841 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B) 1842 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1843 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B) 1844 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1845 break; 1846 } 1847 1848 // Turn a signed comparison into an unsigned one if both operands 1849 // are known to have the same sign. 1850 if (I.isSigned() && 1851 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) || 1852 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative()))) 1853 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1); 1854 } 1855 1856 // Test if the ICmpInst instruction is used exclusively by a select as 1857 // part of a minimum or maximum operation. If so, refrain from doing 1858 // any other folding. This helps out other analyses which understand 1859 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 1860 // and CodeGen. And in this case, at least one of the comparison 1861 // operands has at least one user besides the compare (the select), 1862 // which would often largely negate the benefit of folding anyway. 1863 if (I.hasOneUse()) 1864 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin())) 1865 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) || 1866 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1)) 1867 return 0; 1868 1869 // See if we are doing a comparison between a constant and an instruction that 1870 // can be folded into the comparison. 1871 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 1872 // Since the RHS is a ConstantInt (CI), if the left hand side is an 1873 // instruction, see if that instruction also has constants so that the 1874 // instruction can be folded into the icmp 1875 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 1876 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI)) 1877 return Res; 1878 } 1879 1880 // Handle icmp with constant (but not simple integer constant) RHS 1881 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 1882 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 1883 switch (LHSI->getOpcode()) { 1884 case Instruction::GetElementPtr: 1885 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null 1886 if (RHSC->isNullValue() && 1887 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices()) 1888 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 1889 Constant::getNullValue(LHSI->getOperand(0)->getType())); 1890 break; 1891 case Instruction::PHI: 1892 // Only fold icmp into the PHI if the phi and icmp are in the same 1893 // block. If in the same block, we're encouraging jump threading. If 1894 // not, we are just pessimizing the code by making an i1 phi. 1895 if (LHSI->getParent() == I.getParent()) 1896 if (Instruction *NV = FoldOpIntoPhi(I, true)) 1897 return NV; 1898 break; 1899 case Instruction::Select: { 1900 // If either operand of the select is a constant, we can fold the 1901 // comparison into the select arms, which will cause one to be 1902 // constant folded and the select turned into a bitwise or. 1903 Value *Op1 = 0, *Op2 = 0; 1904 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) 1905 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 1906 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) 1907 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 1908 1909 // We only want to perform this transformation if it will not lead to 1910 // additional code. This is true if either both sides of the select 1911 // fold to a constant (in which case the icmp is replaced with a select 1912 // which will usually simplify) or this is the only user of the 1913 // select (in which case we are trading a select+icmp for a simpler 1914 // select+icmp). 1915 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) { 1916 if (!Op1) 1917 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1), 1918 RHSC, I.getName()); 1919 if (!Op2) 1920 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2), 1921 RHSC, I.getName()); 1922 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 1923 } 1924 break; 1925 } 1926 case Instruction::Call: 1927 // If we have (malloc != null), and if the malloc has a single use, we 1928 // can assume it is successful and remove the malloc. 1929 if (isMalloc(LHSI) && LHSI->hasOneUse() && 1930 isa<ConstantPointerNull>(RHSC)) { 1931 // Need to explicitly erase malloc call here, instead of adding it to 1932 // Worklist, because it won't get DCE'd from the Worklist since 1933 // isInstructionTriviallyDead() returns false for function calls. 1934 // It is OK to replace LHSI/MallocCall with Undef because the 1935 // instruction that uses it will be erased via Worklist. 1936 if (extractMallocCall(LHSI)) { 1937 LHSI->replaceAllUsesWith(UndefValue::get(LHSI->getType())); 1938 EraseInstFromFunction(*LHSI); 1939 return ReplaceInstUsesWith(I, 1940 ConstantInt::get(Type::getInt1Ty(I.getContext()), 1941 !I.isTrueWhenEqual())); 1942 } 1943 if (CallInst* MallocCall = extractMallocCallFromBitCast(LHSI)) 1944 if (MallocCall->hasOneUse()) { 1945 MallocCall->replaceAllUsesWith( 1946 UndefValue::get(MallocCall->getType())); 1947 EraseInstFromFunction(*MallocCall); 1948 Worklist.Add(LHSI); // The malloc's bitcast use. 1949 return ReplaceInstUsesWith(I, 1950 ConstantInt::get(Type::getInt1Ty(I.getContext()), 1951 !I.isTrueWhenEqual())); 1952 } 1953 } 1954 break; 1955 case Instruction::IntToPtr: 1956 // icmp pred inttoptr(X), null -> icmp pred X, 0 1957 if (RHSC->isNullValue() && TD && 1958 TD->getIntPtrType(RHSC->getContext()) == 1959 LHSI->getOperand(0)->getType()) 1960 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 1961 Constant::getNullValue(LHSI->getOperand(0)->getType())); 1962 break; 1963 1964 case Instruction::Load: 1965 // Try to optimize things like "A[i] > 4" to index computations. 1966 if (GetElementPtrInst *GEP = 1967 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 1968 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 1969 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 1970 !cast<LoadInst>(LHSI)->isVolatile()) 1971 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 1972 return Res; 1973 } 1974 break; 1975 } 1976 } 1977 1978 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now. 1979 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0)) 1980 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I)) 1981 return NI; 1982 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) 1983 if (Instruction *NI = FoldGEPICmp(GEP, Op0, 1984 ICmpInst::getSwappedPredicate(I.getPredicate()), I)) 1985 return NI; 1986 1987 // Test to see if the operands of the icmp are casted versions of other 1988 // values. If the ptr->ptr cast can be stripped off both arguments, we do so 1989 // now. 1990 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) { 1991 if (Op0->getType()->isPointerTy() && 1992 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) { 1993 // We keep moving the cast from the left operand over to the right 1994 // operand, where it can often be eliminated completely. 1995 Op0 = CI->getOperand(0); 1996 1997 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast 1998 // so eliminate it as well. 1999 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1)) 2000 Op1 = CI2->getOperand(0); 2001 2002 // If Op1 is a constant, we can fold the cast into the constant. 2003 if (Op0->getType() != Op1->getType()) { 2004 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 2005 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType()); 2006 } else { 2007 // Otherwise, cast the RHS right before the icmp 2008 Op1 = Builder->CreateBitCast(Op1, Op0->getType()); 2009 } 2010 } 2011 return new ICmpInst(I.getPredicate(), Op0, Op1); 2012 } 2013 } 2014 2015 if (isa<CastInst>(Op0)) { 2016 // Handle the special case of: icmp (cast bool to X), <cst> 2017 // This comes up when you have code like 2018 // int X = A < B; 2019 // if (X) ... 2020 // For generality, we handle any zero-extension of any operand comparison 2021 // with a constant or another cast from the same type. 2022 if (isa<Constant>(Op1) || isa<CastInst>(Op1)) 2023 if (Instruction *R = visitICmpInstWithCastAndCast(I)) 2024 return R; 2025 } 2026 2027 // See if it's the same type of instruction on the left and right. 2028 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) { 2029 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) { 2030 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() && 2031 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) { 2032 switch (Op0I->getOpcode()) { 2033 default: break; 2034 case Instruction::Add: 2035 case Instruction::Sub: 2036 case Instruction::Xor: 2037 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b 2038 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0), 2039 Op1I->getOperand(0)); 2040 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b 2041 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) { 2042 if (CI->getValue().isSignBit()) { 2043 ICmpInst::Predicate Pred = I.isSigned() 2044 ? I.getUnsignedPredicate() 2045 : I.getSignedPredicate(); 2046 return new ICmpInst(Pred, Op0I->getOperand(0), 2047 Op1I->getOperand(0)); 2048 } 2049 2050 if (CI->getValue().isMaxSignedValue()) { 2051 ICmpInst::Predicate Pred = I.isSigned() 2052 ? I.getUnsignedPredicate() 2053 : I.getSignedPredicate(); 2054 Pred = I.getSwappedPredicate(Pred); 2055 return new ICmpInst(Pred, Op0I->getOperand(0), 2056 Op1I->getOperand(0)); 2057 } 2058 } 2059 break; 2060 case Instruction::Mul: 2061 if (!I.isEquality()) 2062 break; 2063 2064 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) { 2065 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask 2066 // Mask = -1 >> count-trailing-zeros(Cst). 2067 if (!CI->isZero() && !CI->isOne()) { 2068 const APInt &AP = CI->getValue(); 2069 ConstantInt *Mask = ConstantInt::get(I.getContext(), 2070 APInt::getLowBitsSet(AP.getBitWidth(), 2071 AP.getBitWidth() - 2072 AP.countTrailingZeros())); 2073 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask); 2074 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask); 2075 return new ICmpInst(I.getPredicate(), And1, And2); 2076 } 2077 } 2078 break; 2079 } 2080 } 2081 } 2082 } 2083 2084 // ~x < ~y --> y < x 2085 { Value *A, *B; 2086 if (match(Op0, m_Not(m_Value(A))) && 2087 match(Op1, m_Not(m_Value(B)))) 2088 return new ICmpInst(I.getPredicate(), B, A); 2089 } 2090 2091 if (I.isEquality()) { 2092 Value *A, *B, *C, *D; 2093 2094 // -x == -y --> x == y 2095 if (match(Op0, m_Neg(m_Value(A))) && 2096 match(Op1, m_Neg(m_Value(B)))) 2097 return new ICmpInst(I.getPredicate(), A, B); 2098 2099 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 2100 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0 2101 Value *OtherVal = A == Op1 ? B : A; 2102 return new ICmpInst(I.getPredicate(), OtherVal, 2103 Constant::getNullValue(A->getType())); 2104 } 2105 2106 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) { 2107 // A^c1 == C^c2 --> A == C^(c1^c2) 2108 ConstantInt *C1, *C2; 2109 if (match(B, m_ConstantInt(C1)) && 2110 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) { 2111 Constant *NC = ConstantInt::get(I.getContext(), 2112 C1->getValue() ^ C2->getValue()); 2113 Value *Xor = Builder->CreateXor(C, NC, "tmp"); 2114 return new ICmpInst(I.getPredicate(), A, Xor); 2115 } 2116 2117 // A^B == A^D -> B == D 2118 if (A == C) return new ICmpInst(I.getPredicate(), B, D); 2119 if (A == D) return new ICmpInst(I.getPredicate(), B, C); 2120 if (B == C) return new ICmpInst(I.getPredicate(), A, D); 2121 if (B == D) return new ICmpInst(I.getPredicate(), A, C); 2122 } 2123 } 2124 2125 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && 2126 (A == Op0 || B == Op0)) { 2127 // A == (A^B) -> B == 0 2128 Value *OtherVal = A == Op0 ? B : A; 2129 return new ICmpInst(I.getPredicate(), OtherVal, 2130 Constant::getNullValue(A->getType())); 2131 } 2132 2133 // (A-B) == A -> B == 0 2134 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B)))) 2135 return new ICmpInst(I.getPredicate(), B, 2136 Constant::getNullValue(B->getType())); 2137 2138 // A == (A-B) -> B == 0 2139 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B)))) 2140 return new ICmpInst(I.getPredicate(), B, 2141 Constant::getNullValue(B->getType())); 2142 2143 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0 2144 if (Op0->hasOneUse() && Op1->hasOneUse() && 2145 match(Op0, m_And(m_Value(A), m_Value(B))) && 2146 match(Op1, m_And(m_Value(C), m_Value(D)))) { 2147 Value *X = 0, *Y = 0, *Z = 0; 2148 2149 if (A == C) { 2150 X = B; Y = D; Z = A; 2151 } else if (A == D) { 2152 X = B; Y = C; Z = A; 2153 } else if (B == C) { 2154 X = A; Y = D; Z = B; 2155 } else if (B == D) { 2156 X = A; Y = C; Z = B; 2157 } 2158 2159 if (X) { // Build (X^Y) & Z 2160 Op1 = Builder->CreateXor(X, Y, "tmp"); 2161 Op1 = Builder->CreateAnd(Op1, Z, "tmp"); 2162 I.setOperand(0, Op1); 2163 I.setOperand(1, Constant::getNullValue(Op1->getType())); 2164 return &I; 2165 } 2166 } 2167 } 2168 2169 { 2170 Value *X; ConstantInt *Cst; 2171 // icmp X+Cst, X 2172 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X) 2173 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0); 2174 2175 // icmp X, X+Cst 2176 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X) 2177 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1); 2178 } 2179 return Changed ? &I : 0; 2180} 2181 2182 2183 2184 2185 2186 2187/// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible. 2188/// 2189Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I, 2190 Instruction *LHSI, 2191 Constant *RHSC) { 2192 if (!isa<ConstantFP>(RHSC)) return 0; 2193 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF(); 2194 2195 // Get the width of the mantissa. We don't want to hack on conversions that 2196 // might lose information from the integer, e.g. "i64 -> float" 2197 int MantissaWidth = LHSI->getType()->getFPMantissaWidth(); 2198 if (MantissaWidth == -1) return 0; // Unknown. 2199 2200 // Check to see that the input is converted from an integer type that is small 2201 // enough that preserves all bits. TODO: check here for "known" sign bits. 2202 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e. 2203 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits(); 2204 2205 // If this is a uitofp instruction, we need an extra bit to hold the sign. 2206 bool LHSUnsigned = isa<UIToFPInst>(LHSI); 2207 if (LHSUnsigned) 2208 ++InputSize; 2209 2210 // If the conversion would lose info, don't hack on this. 2211 if ((int)InputSize > MantissaWidth) 2212 return 0; 2213 2214 // Otherwise, we can potentially simplify the comparison. We know that it 2215 // will always come through as an integer value and we know the constant is 2216 // not a NAN (it would have been previously simplified). 2217 assert(!RHS.isNaN() && "NaN comparison not already folded!"); 2218 2219 ICmpInst::Predicate Pred; 2220 switch (I.getPredicate()) { 2221 default: llvm_unreachable("Unexpected predicate!"); 2222 case FCmpInst::FCMP_UEQ: 2223 case FCmpInst::FCMP_OEQ: 2224 Pred = ICmpInst::ICMP_EQ; 2225 break; 2226 case FCmpInst::FCMP_UGT: 2227 case FCmpInst::FCMP_OGT: 2228 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT; 2229 break; 2230 case FCmpInst::FCMP_UGE: 2231 case FCmpInst::FCMP_OGE: 2232 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE; 2233 break; 2234 case FCmpInst::FCMP_ULT: 2235 case FCmpInst::FCMP_OLT: 2236 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT; 2237 break; 2238 case FCmpInst::FCMP_ULE: 2239 case FCmpInst::FCMP_OLE: 2240 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE; 2241 break; 2242 case FCmpInst::FCMP_UNE: 2243 case FCmpInst::FCMP_ONE: 2244 Pred = ICmpInst::ICMP_NE; 2245 break; 2246 case FCmpInst::FCMP_ORD: 2247 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2248 case FCmpInst::FCMP_UNO: 2249 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2250 } 2251 2252 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType()); 2253 2254 // Now we know that the APFloat is a normal number, zero or inf. 2255 2256 // See if the FP constant is too large for the integer. For example, 2257 // comparing an i8 to 300.0. 2258 unsigned IntWidth = IntTy->getScalarSizeInBits(); 2259 2260 if (!LHSUnsigned) { 2261 // If the RHS value is > SignedMax, fold the comparison. This handles +INF 2262 // and large values. 2263 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false); 2264 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true, 2265 APFloat::rmNearestTiesToEven); 2266 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0 2267 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT || 2268 Pred == ICmpInst::ICMP_SLE) 2269 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2270 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2271 } 2272 } else { 2273 // If the RHS value is > UnsignedMax, fold the comparison. This handles 2274 // +INF and large values. 2275 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false); 2276 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false, 2277 APFloat::rmNearestTiesToEven); 2278 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0 2279 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT || 2280 Pred == ICmpInst::ICMP_ULE) 2281 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2282 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2283 } 2284 } 2285 2286 if (!LHSUnsigned) { 2287 // See if the RHS value is < SignedMin. 2288 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false); 2289 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true, 2290 APFloat::rmNearestTiesToEven); 2291 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0 2292 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT || 2293 Pred == ICmpInst::ICMP_SGE) 2294 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2295 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2296 } 2297 } 2298 2299 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or 2300 // [0, UMAX], but it may still be fractional. See if it is fractional by 2301 // casting the FP value to the integer value and back, checking for equality. 2302 // Don't do this for zero, because -0.0 is not fractional. 2303 Constant *RHSInt = LHSUnsigned 2304 ? ConstantExpr::getFPToUI(RHSC, IntTy) 2305 : ConstantExpr::getFPToSI(RHSC, IntTy); 2306 if (!RHS.isZero()) { 2307 bool Equal = LHSUnsigned 2308 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC 2309 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC; 2310 if (!Equal) { 2311 // If we had a comparison against a fractional value, we have to adjust 2312 // the compare predicate and sometimes the value. RHSC is rounded towards 2313 // zero at this point. 2314 switch (Pred) { 2315 default: llvm_unreachable("Unexpected integer comparison!"); 2316 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true 2317 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2318 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false 2319 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2320 case ICmpInst::ICMP_ULE: 2321 // (float)int <= 4.4 --> int <= 4 2322 // (float)int <= -4.4 --> false 2323 if (RHS.isNegative()) 2324 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2325 break; 2326 case ICmpInst::ICMP_SLE: 2327 // (float)int <= 4.4 --> int <= 4 2328 // (float)int <= -4.4 --> int < -4 2329 if (RHS.isNegative()) 2330 Pred = ICmpInst::ICMP_SLT; 2331 break; 2332 case ICmpInst::ICMP_ULT: 2333 // (float)int < -4.4 --> false 2334 // (float)int < 4.4 --> int <= 4 2335 if (RHS.isNegative()) 2336 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2337 Pred = ICmpInst::ICMP_ULE; 2338 break; 2339 case ICmpInst::ICMP_SLT: 2340 // (float)int < -4.4 --> int < -4 2341 // (float)int < 4.4 --> int <= 4 2342 if (!RHS.isNegative()) 2343 Pred = ICmpInst::ICMP_SLE; 2344 break; 2345 case ICmpInst::ICMP_UGT: 2346 // (float)int > 4.4 --> int > 4 2347 // (float)int > -4.4 --> true 2348 if (RHS.isNegative()) 2349 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2350 break; 2351 case ICmpInst::ICMP_SGT: 2352 // (float)int > 4.4 --> int > 4 2353 // (float)int > -4.4 --> int >= -4 2354 if (RHS.isNegative()) 2355 Pred = ICmpInst::ICMP_SGE; 2356 break; 2357 case ICmpInst::ICMP_UGE: 2358 // (float)int >= -4.4 --> true 2359 // (float)int >= 4.4 --> int > 4 2360 if (!RHS.isNegative()) 2361 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2362 Pred = ICmpInst::ICMP_UGT; 2363 break; 2364 case ICmpInst::ICMP_SGE: 2365 // (float)int >= -4.4 --> int >= -4 2366 // (float)int >= 4.4 --> int > 4 2367 if (!RHS.isNegative()) 2368 Pred = ICmpInst::ICMP_SGT; 2369 break; 2370 } 2371 } 2372 } 2373 2374 // Lower this FP comparison into an appropriate integer version of the 2375 // comparison. 2376 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt); 2377} 2378 2379Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { 2380 bool Changed = false; 2381 2382 /// Orders the operands of the compare so that they are listed from most 2383 /// complex to least complex. This puts constants before unary operators, 2384 /// before binary operators. 2385 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) { 2386 I.swapOperands(); 2387 Changed = true; 2388 } 2389 2390 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2391 2392 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD)) 2393 return ReplaceInstUsesWith(I, V); 2394 2395 // Simplify 'fcmp pred X, X' 2396 if (Op0 == Op1) { 2397 switch (I.getPredicate()) { 2398 default: llvm_unreachable("Unknown predicate!"); 2399 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y) 2400 case FCmpInst::FCMP_ULT: // True if unordered or less than 2401 case FCmpInst::FCMP_UGT: // True if unordered or greater than 2402 case FCmpInst::FCMP_UNE: // True if unordered or not equal 2403 // Canonicalize these to be 'fcmp uno %X, 0.0'. 2404 I.setPredicate(FCmpInst::FCMP_UNO); 2405 I.setOperand(1, Constant::getNullValue(Op0->getType())); 2406 return &I; 2407 2408 case FCmpInst::FCMP_ORD: // True if ordered (no nans) 2409 case FCmpInst::FCMP_OEQ: // True if ordered and equal 2410 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal 2411 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal 2412 // Canonicalize these to be 'fcmp ord %X, 0.0'. 2413 I.setPredicate(FCmpInst::FCMP_ORD); 2414 I.setOperand(1, Constant::getNullValue(Op0->getType())); 2415 return &I; 2416 } 2417 } 2418 2419 // Handle fcmp with constant RHS 2420 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2421 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2422 switch (LHSI->getOpcode()) { 2423 case Instruction::PHI: 2424 // Only fold fcmp into the PHI if the phi and fcmp are in the same 2425 // block. If in the same block, we're encouraging jump threading. If 2426 // not, we are just pessimizing the code by making an i1 phi. 2427 if (LHSI->getParent() == I.getParent()) 2428 if (Instruction *NV = FoldOpIntoPhi(I, true)) 2429 return NV; 2430 break; 2431 case Instruction::SIToFP: 2432 case Instruction::UIToFP: 2433 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC)) 2434 return NV; 2435 break; 2436 case Instruction::Select: { 2437 // If either operand of the select is a constant, we can fold the 2438 // comparison into the select arms, which will cause one to be 2439 // constant folded and the select turned into a bitwise or. 2440 Value *Op1 = 0, *Op2 = 0; 2441 if (LHSI->hasOneUse()) { 2442 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) { 2443 // Fold the known value into the constant operand. 2444 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 2445 // Insert a new FCmp of the other select operand. 2446 Op2 = Builder->CreateFCmp(I.getPredicate(), 2447 LHSI->getOperand(2), RHSC, I.getName()); 2448 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) { 2449 // Fold the known value into the constant operand. 2450 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 2451 // Insert a new FCmp of the other select operand. 2452 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1), 2453 RHSC, I.getName()); 2454 } 2455 } 2456 2457 if (Op1) 2458 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2459 break; 2460 } 2461 case Instruction::Load: 2462 if (GetElementPtrInst *GEP = 2463 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2464 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2465 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2466 !cast<LoadInst>(LHSI)->isVolatile()) 2467 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2468 return Res; 2469 } 2470 break; 2471 } 2472 } 2473 2474 return Changed ? &I : 0; 2475} 2476