InstCombineCompares.cpp revision 251662
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/Analysis/ConstantFolding.h" 16#include "llvm/Analysis/InstructionSimplify.h" 17#include "llvm/Analysis/MemoryBuiltins.h" 18#include "llvm/IR/DataLayout.h" 19#include "llvm/IR/IntrinsicInst.h" 20#include "llvm/Support/ConstantRange.h" 21#include "llvm/Support/GetElementPtrTypeIterator.h" 22#include "llvm/Support/PatternMatch.h" 23#include "llvm/Target/TargetLibraryInfo.h" 24using namespace llvm; 25using namespace PatternMatch; 26 27static ConstantInt *getOne(Constant *C) { 28 return ConstantInt::get(cast<IntegerType>(C->getType()), 1); 29} 30 31/// AddOne - Add one to a ConstantInt 32static Constant *AddOne(Constant *C) { 33 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 34} 35/// SubOne - Subtract one from a ConstantInt 36static Constant *SubOne(Constant *C) { 37 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1)); 38} 39 40static ConstantInt *ExtractElement(Constant *V, Constant *Idx) { 41 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx)); 42} 43 44static bool HasAddOverflow(ConstantInt *Result, 45 ConstantInt *In1, ConstantInt *In2, 46 bool IsSigned) { 47 if (!IsSigned) 48 return Result->getValue().ult(In1->getValue()); 49 50 if (In2->isNegative()) 51 return Result->getValue().sgt(In1->getValue()); 52 return Result->getValue().slt(In1->getValue()); 53} 54 55/// AddWithOverflow - Compute Result = In1+In2, returning true if the result 56/// overflowed for this type. 57static bool AddWithOverflow(Constant *&Result, Constant *In1, 58 Constant *In2, bool IsSigned = false) { 59 Result = ConstantExpr::getAdd(In1, In2); 60 61 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 62 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 63 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 64 if (HasAddOverflow(ExtractElement(Result, Idx), 65 ExtractElement(In1, Idx), 66 ExtractElement(In2, Idx), 67 IsSigned)) 68 return true; 69 } 70 return false; 71 } 72 73 return HasAddOverflow(cast<ConstantInt>(Result), 74 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 75 IsSigned); 76} 77 78static bool HasSubOverflow(ConstantInt *Result, 79 ConstantInt *In1, ConstantInt *In2, 80 bool IsSigned) { 81 if (!IsSigned) 82 return Result->getValue().ugt(In1->getValue()); 83 84 if (In2->isNegative()) 85 return Result->getValue().slt(In1->getValue()); 86 87 return Result->getValue().sgt(In1->getValue()); 88} 89 90/// SubWithOverflow - Compute Result = In1-In2, returning true if the result 91/// overflowed for this type. 92static bool SubWithOverflow(Constant *&Result, Constant *In1, 93 Constant *In2, bool IsSigned = false) { 94 Result = ConstantExpr::getSub(In1, In2); 95 96 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 97 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 98 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 99 if (HasSubOverflow(ExtractElement(Result, Idx), 100 ExtractElement(In1, Idx), 101 ExtractElement(In2, Idx), 102 IsSigned)) 103 return true; 104 } 105 return false; 106 } 107 108 return HasSubOverflow(cast<ConstantInt>(Result), 109 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 110 IsSigned); 111} 112 113/// isSignBitCheck - Given an exploded icmp instruction, return true if the 114/// comparison only checks the sign bit. If it only checks the sign bit, set 115/// TrueIfSigned if the result of the comparison is true when the input value is 116/// signed. 117static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS, 118 bool &TrueIfSigned) { 119 switch (pred) { 120 case ICmpInst::ICMP_SLT: // True if LHS s< 0 121 TrueIfSigned = true; 122 return RHS->isZero(); 123 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1 124 TrueIfSigned = true; 125 return RHS->isAllOnesValue(); 126 case ICmpInst::ICMP_SGT: // True if LHS s> -1 127 TrueIfSigned = false; 128 return RHS->isAllOnesValue(); 129 case ICmpInst::ICMP_UGT: 130 // True if LHS u> RHS and RHS == high-bit-mask - 1 131 TrueIfSigned = true; 132 return RHS->isMaxValue(true); 133 case ICmpInst::ICMP_UGE: 134 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc) 135 TrueIfSigned = true; 136 return RHS->getValue().isSignBit(); 137 default: 138 return false; 139 } 140} 141 142/// Returns true if the exploded icmp can be expressed as a signed comparison 143/// to zero and updates the predicate accordingly. 144/// The signedness of the comparison is preserved. 145static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) { 146 if (!ICmpInst::isSigned(pred)) 147 return false; 148 149 if (RHS->isZero()) 150 return ICmpInst::isRelational(pred); 151 152 if (RHS->isOne()) { 153 if (pred == ICmpInst::ICMP_SLT) { 154 pred = ICmpInst::ICMP_SLE; 155 return true; 156 } 157 } else if (RHS->isAllOnesValue()) { 158 if (pred == ICmpInst::ICMP_SGT) { 159 pred = ICmpInst::ICMP_SGE; 160 return true; 161 } 162 } 163 164 return false; 165} 166 167// isHighOnes - Return true if the constant is of the form 1+0+. 168// This is the same as lowones(~X). 169static bool isHighOnes(const ConstantInt *CI) { 170 return (~CI->getValue() + 1).isPowerOf2(); 171} 172 173/// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a 174/// set of known zero and one bits, compute the maximum and minimum values that 175/// could have the specified known zero and known one bits, returning them in 176/// min/max. 177static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero, 178 const APInt& KnownOne, 179 APInt& Min, APInt& Max) { 180 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 181 KnownZero.getBitWidth() == Min.getBitWidth() && 182 KnownZero.getBitWidth() == Max.getBitWidth() && 183 "KnownZero, KnownOne and Min, Max must have equal bitwidth."); 184 APInt UnknownBits = ~(KnownZero|KnownOne); 185 186 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign 187 // bit if it is unknown. 188 Min = KnownOne; 189 Max = KnownOne|UnknownBits; 190 191 if (UnknownBits.isNegative()) { // Sign bit is unknown 192 Min.setBit(Min.getBitWidth()-1); 193 Max.clearBit(Max.getBitWidth()-1); 194 } 195} 196 197// ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and 198// a set of known zero and one bits, compute the maximum and minimum values that 199// could have the specified known zero and known one bits, returning them in 200// min/max. 201static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero, 202 const APInt &KnownOne, 203 APInt &Min, APInt &Max) { 204 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 205 KnownZero.getBitWidth() == Min.getBitWidth() && 206 KnownZero.getBitWidth() == Max.getBitWidth() && 207 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."); 208 APInt UnknownBits = ~(KnownZero|KnownOne); 209 210 // The minimum value is when the unknown bits are all zeros. 211 Min = KnownOne; 212 // The maximum value is when the unknown bits are all ones. 213 Max = KnownOne|UnknownBits; 214} 215 216 217 218/// FoldCmpLoadFromIndexedGlobal - Called we see this pattern: 219/// cmp pred (load (gep GV, ...)), cmpcst 220/// where GV is a global variable with a constant initializer. Try to simplify 221/// this into some simple computation that does not need the load. For example 222/// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3". 223/// 224/// If AndCst is non-null, then the loaded value is masked with that constant 225/// before doing the comparison. This handles cases like "A[i]&4 == 0". 226Instruction *InstCombiner:: 227FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV, 228 CmpInst &ICI, ConstantInt *AndCst) { 229 // We need TD information to know the pointer size unless this is inbounds. 230 if (!GEP->isInBounds() && TD == 0) return 0; 231 232 Constant *Init = GV->getInitializer(); 233 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init)) 234 return 0; 235 236 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements(); 237 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays. 238 239 // There are many forms of this optimization we can handle, for now, just do 240 // the simple index into a single-dimensional array. 241 // 242 // Require: GEP GV, 0, i {{, constant indices}} 243 if (GEP->getNumOperands() < 3 || 244 !isa<ConstantInt>(GEP->getOperand(1)) || 245 !cast<ConstantInt>(GEP->getOperand(1))->isZero() || 246 isa<Constant>(GEP->getOperand(2))) 247 return 0; 248 249 // Check that indices after the variable are constants and in-range for the 250 // type they index. Collect the indices. This is typically for arrays of 251 // structs. 252 SmallVector<unsigned, 4> LaterIndices; 253 254 Type *EltTy = Init->getType()->getArrayElementType(); 255 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { 256 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i)); 257 if (Idx == 0) return 0; // Variable index. 258 259 uint64_t IdxVal = Idx->getZExtValue(); 260 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index. 261 262 if (StructType *STy = dyn_cast<StructType>(EltTy)) 263 EltTy = STy->getElementType(IdxVal); 264 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) { 265 if (IdxVal >= ATy->getNumElements()) return 0; 266 EltTy = ATy->getElementType(); 267 } else { 268 return 0; // Unknown type. 269 } 270 271 LaterIndices.push_back(IdxVal); 272 } 273 274 enum { Overdefined = -3, Undefined = -2 }; 275 276 // Variables for our state machines. 277 278 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form 279 // "i == 47 | i == 87", where 47 is the first index the condition is true for, 280 // and 87 is the second (and last) index. FirstTrueElement is -2 when 281 // undefined, otherwise set to the first true element. SecondTrueElement is 282 // -2 when undefined, -3 when overdefined and >= 0 when that index is true. 283 int FirstTrueElement = Undefined, SecondTrueElement = Undefined; 284 285 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the 286 // form "i != 47 & i != 87". Same state transitions as for true elements. 287 int FirstFalseElement = Undefined, SecondFalseElement = Undefined; 288 289 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these 290 /// define a state machine that triggers for ranges of values that the index 291 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. 292 /// This is -2 when undefined, -3 when overdefined, and otherwise the last 293 /// index in the range (inclusive). We use -2 for undefined here because we 294 /// use relative comparisons and don't want 0-1 to match -1. 295 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; 296 297 // MagicBitvector - This is a magic bitvector where we set a bit if the 298 // comparison is true for element 'i'. If there are 64 elements or less in 299 // the array, this will fully represent all the comparison results. 300 uint64_t MagicBitvector = 0; 301 302 303 // Scan the array and see if one of our patterns matches. 304 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1)); 305 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) { 306 Constant *Elt = Init->getAggregateElement(i); 307 if (Elt == 0) return 0; 308 309 // If this is indexing an array of structures, get the structure element. 310 if (!LaterIndices.empty()) 311 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices); 312 313 // If the element is masked, handle it. 314 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst); 315 316 // Find out if the comparison would be true or false for the i'th element. 317 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt, 318 CompareRHS, TD, TLI); 319 // If the result is undef for this element, ignore it. 320 if (isa<UndefValue>(C)) { 321 // Extend range state machines to cover this element in case there is an 322 // undef in the middle of the range. 323 if (TrueRangeEnd == (int)i-1) 324 TrueRangeEnd = i; 325 if (FalseRangeEnd == (int)i-1) 326 FalseRangeEnd = i; 327 continue; 328 } 329 330 // If we can't compute the result for any of the elements, we have to give 331 // up evaluating the entire conditional. 332 if (!isa<ConstantInt>(C)) return 0; 333 334 // Otherwise, we know if the comparison is true or false for this element, 335 // update our state machines. 336 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero(); 337 338 // State machine for single/double/range index comparison. 339 if (IsTrueForElt) { 340 // Update the TrueElement state machine. 341 if (FirstTrueElement == Undefined) 342 FirstTrueElement = TrueRangeEnd = i; // First true element. 343 else { 344 // Update double-compare state machine. 345 if (SecondTrueElement == Undefined) 346 SecondTrueElement = i; 347 else 348 SecondTrueElement = Overdefined; 349 350 // Update range state machine. 351 if (TrueRangeEnd == (int)i-1) 352 TrueRangeEnd = i; 353 else 354 TrueRangeEnd = Overdefined; 355 } 356 } else { 357 // Update the FalseElement state machine. 358 if (FirstFalseElement == Undefined) 359 FirstFalseElement = FalseRangeEnd = i; // First false element. 360 else { 361 // Update double-compare state machine. 362 if (SecondFalseElement == Undefined) 363 SecondFalseElement = i; 364 else 365 SecondFalseElement = Overdefined; 366 367 // Update range state machine. 368 if (FalseRangeEnd == (int)i-1) 369 FalseRangeEnd = i; 370 else 371 FalseRangeEnd = Overdefined; 372 } 373 } 374 375 376 // If this element is in range, update our magic bitvector. 377 if (i < 64 && IsTrueForElt) 378 MagicBitvector |= 1ULL << i; 379 380 // If all of our states become overdefined, bail out early. Since the 381 // predicate is expensive, only check it every 8 elements. This is only 382 // really useful for really huge arrays. 383 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && 384 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && 385 FalseRangeEnd == Overdefined) 386 return 0; 387 } 388 389 // Now that we've scanned the entire array, emit our new comparison(s). We 390 // order the state machines in complexity of the generated code. 391 Value *Idx = GEP->getOperand(2); 392 393 // If the index is larger than the pointer size of the target, truncate the 394 // index down like the GEP would do implicitly. We don't have to do this for 395 // an inbounds GEP because the index can't be out of range. 396 if (!GEP->isInBounds() && 397 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits()) 398 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext())); 399 400 // If the comparison is only true for one or two elements, emit direct 401 // comparisons. 402 if (SecondTrueElement != Overdefined) { 403 // None true -> false. 404 if (FirstTrueElement == Undefined) 405 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext())); 406 407 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement); 408 409 // True for one element -> 'i == 47'. 410 if (SecondTrueElement == Undefined) 411 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); 412 413 // True for two elements -> 'i == 47 | i == 72'. 414 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx); 415 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement); 416 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx); 417 return BinaryOperator::CreateOr(C1, C2); 418 } 419 420 // If the comparison is only false for one or two elements, emit direct 421 // comparisons. 422 if (SecondFalseElement != Overdefined) { 423 // None false -> true. 424 if (FirstFalseElement == Undefined) 425 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext())); 426 427 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement); 428 429 // False for one element -> 'i != 47'. 430 if (SecondFalseElement == Undefined) 431 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); 432 433 // False for two elements -> 'i != 47 & i != 72'. 434 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx); 435 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement); 436 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx); 437 return BinaryOperator::CreateAnd(C1, C2); 438 } 439 440 // If the comparison can be replaced with a range comparison for the elements 441 // where it is true, emit the range check. 442 if (TrueRangeEnd != Overdefined) { 443 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare"); 444 445 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1). 446 if (FirstTrueElement) { 447 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement); 448 Idx = Builder->CreateAdd(Idx, Offs); 449 } 450 451 Value *End = ConstantInt::get(Idx->getType(), 452 TrueRangeEnd-FirstTrueElement+1); 453 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); 454 } 455 456 // False range check. 457 if (FalseRangeEnd != Overdefined) { 458 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare"); 459 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). 460 if (FirstFalseElement) { 461 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement); 462 Idx = Builder->CreateAdd(Idx, Offs); 463 } 464 465 Value *End = ConstantInt::get(Idx->getType(), 466 FalseRangeEnd-FirstFalseElement); 467 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); 468 } 469 470 471 // If a magic bitvector captures the entire comparison state 472 // of this load, replace it with computation that does: 473 // ((magic_cst >> i) & 1) != 0 474 { 475 Type *Ty = 0; 476 477 // Look for an appropriate type: 478 // - The type of Idx if the magic fits 479 // - The smallest fitting legal type if we have a DataLayout 480 // - Default to i32 481 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth()) 482 Ty = Idx->getType(); 483 else if (TD) 484 Ty = TD->getSmallestLegalIntType(Init->getContext(), ArrayElementCount); 485 else if (ArrayElementCount <= 32) 486 Ty = Type::getInt32Ty(Init->getContext()); 487 488 if (Ty != 0) { 489 Value *V = Builder->CreateIntCast(Idx, Ty, false); 490 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V); 491 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V); 492 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0)); 493 } 494 } 495 496 return 0; 497} 498 499 500/// EvaluateGEPOffsetExpression - Return a value that can be used to compare 501/// the *offset* implied by a GEP to zero. For example, if we have &A[i], we 502/// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can 503/// be complex, and scales are involved. The above expression would also be 504/// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). 505/// This later form is less amenable to optimization though, and we are allowed 506/// to generate the first by knowing that pointer arithmetic doesn't overflow. 507/// 508/// If we can't emit an optimized form for this expression, this returns null. 509/// 510static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) { 511 DataLayout &TD = *IC.getDataLayout(); 512 gep_type_iterator GTI = gep_type_begin(GEP); 513 514 // Check to see if this gep only has a single variable index. If so, and if 515 // any constant indices are a multiple of its scale, then we can compute this 516 // in terms of the scale of the variable index. For example, if the GEP 517 // implies an offset of "12 + i*4", then we can codegen this as "3 + i", 518 // because the expression will cross zero at the same point. 519 unsigned i, e = GEP->getNumOperands(); 520 int64_t Offset = 0; 521 for (i = 1; i != e; ++i, ++GTI) { 522 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 523 // Compute the aggregate offset of constant indices. 524 if (CI->isZero()) continue; 525 526 // Handle a struct index, which adds its field offset to the pointer. 527 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 528 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 529 } else { 530 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 531 Offset += Size*CI->getSExtValue(); 532 } 533 } else { 534 // Found our variable index. 535 break; 536 } 537 } 538 539 // If there are no variable indices, we must have a constant offset, just 540 // evaluate it the general way. 541 if (i == e) return 0; 542 543 Value *VariableIdx = GEP->getOperand(i); 544 // Determine the scale factor of the variable element. For example, this is 545 // 4 if the variable index is into an array of i32. 546 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType()); 547 548 // Verify that there are no other variable indices. If so, emit the hard way. 549 for (++i, ++GTI; i != e; ++i, ++GTI) { 550 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i)); 551 if (!CI) return 0; 552 553 // Compute the aggregate offset of constant indices. 554 if (CI->isZero()) continue; 555 556 // Handle a struct index, which adds its field offset to the pointer. 557 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 558 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 559 } else { 560 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 561 Offset += Size*CI->getSExtValue(); 562 } 563 } 564 565 // Okay, we know we have a single variable index, which must be a 566 // pointer/array/vector index. If there is no offset, life is simple, return 567 // the index. 568 unsigned IntPtrWidth = TD.getPointerSizeInBits(); 569 if (Offset == 0) { 570 // Cast to intptrty in case a truncation occurs. If an extension is needed, 571 // we don't need to bother extending: the extension won't affect where the 572 // computation crosses zero. 573 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) { 574 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext()); 575 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy); 576 } 577 return VariableIdx; 578 } 579 580 // Otherwise, there is an index. The computation we will do will be modulo 581 // the pointer size, so get it. 582 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth); 583 584 Offset &= PtrSizeMask; 585 VariableScale &= PtrSizeMask; 586 587 // To do this transformation, any constant index must be a multiple of the 588 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i", 589 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a 590 // multiple of the variable scale. 591 int64_t NewOffs = Offset / (int64_t)VariableScale; 592 if (Offset != NewOffs*(int64_t)VariableScale) 593 return 0; 594 595 // Okay, we can do this evaluation. Start by converting the index to intptr. 596 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext()); 597 if (VariableIdx->getType() != IntPtrTy) 598 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy, 599 true /*Signed*/); 600 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs); 601 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset"); 602} 603 604/// FoldGEPICmp - Fold comparisons between a GEP instruction and something 605/// else. At this point we know that the GEP is on the LHS of the comparison. 606Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 607 ICmpInst::Predicate Cond, 608 Instruction &I) { 609 // Don't transform signed compares of GEPs into index compares. Even if the 610 // GEP is inbounds, the final add of the base pointer can have signed overflow 611 // and would change the result of the icmp. 612 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be 613 // the maximum signed value for the pointer type. 614 if (ICmpInst::isSigned(Cond)) 615 return 0; 616 617 // Look through bitcasts. 618 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS)) 619 RHS = BCI->getOperand(0); 620 621 Value *PtrBase = GEPLHS->getOperand(0); 622 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) { 623 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). 624 // This transformation (ignoring the base and scales) is valid because we 625 // know pointers can't overflow since the gep is inbounds. See if we can 626 // output an optimized form. 627 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this); 628 629 // If not, synthesize the offset the hard way. 630 if (Offset == 0) 631 Offset = EmitGEPOffset(GEPLHS); 632 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset, 633 Constant::getNullValue(Offset->getType())); 634 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) { 635 // If the base pointers are different, but the indices are the same, just 636 // compare the base pointer. 637 if (PtrBase != GEPRHS->getOperand(0)) { 638 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands(); 639 IndicesTheSame &= GEPLHS->getOperand(0)->getType() == 640 GEPRHS->getOperand(0)->getType(); 641 if (IndicesTheSame) 642 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 643 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 644 IndicesTheSame = false; 645 break; 646 } 647 648 // If all indices are the same, just compare the base pointers. 649 if (IndicesTheSame) 650 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), 651 GEPLHS->getOperand(0), GEPRHS->getOperand(0)); 652 653 // If we're comparing GEPs with two base pointers that only differ in type 654 // and both GEPs have only constant indices or just one use, then fold 655 // the compare with the adjusted indices. 656 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() && 657 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) && 658 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) && 659 PtrBase->stripPointerCasts() == 660 GEPRHS->getOperand(0)->stripPointerCasts()) { 661 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond), 662 EmitGEPOffset(GEPLHS), 663 EmitGEPOffset(GEPRHS)); 664 return ReplaceInstUsesWith(I, Cmp); 665 } 666 667 // Otherwise, the base pointers are different and the indices are 668 // different, bail out. 669 return 0; 670 } 671 672 // If one of the GEPs has all zero indices, recurse. 673 bool AllZeros = true; 674 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 675 if (!isa<Constant>(GEPLHS->getOperand(i)) || 676 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) { 677 AllZeros = false; 678 break; 679 } 680 if (AllZeros) 681 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0), 682 ICmpInst::getSwappedPredicate(Cond), I); 683 684 // If the other GEP has all zero indices, recurse. 685 AllZeros = true; 686 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 687 if (!isa<Constant>(GEPRHS->getOperand(i)) || 688 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) { 689 AllZeros = false; 690 break; 691 } 692 if (AllZeros) 693 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I); 694 695 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds(); 696 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) { 697 // If the GEPs only differ by one index, compare it. 698 unsigned NumDifferences = 0; // Keep track of # differences. 699 unsigned DiffOperand = 0; // The operand that differs. 700 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 701 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 702 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() != 703 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) { 704 // Irreconcilable differences. 705 NumDifferences = 2; 706 break; 707 } else { 708 if (NumDifferences++) break; 709 DiffOperand = i; 710 } 711 } 712 713 if (NumDifferences == 0) // SAME GEP? 714 return ReplaceInstUsesWith(I, // No comparison is needed here. 715 ConstantInt::get(Type::getInt1Ty(I.getContext()), 716 ICmpInst::isTrueWhenEqual(Cond))); 717 718 else if (NumDifferences == 1 && GEPsInBounds) { 719 Value *LHSV = GEPLHS->getOperand(DiffOperand); 720 Value *RHSV = GEPRHS->getOperand(DiffOperand); 721 // Make sure we do a signed comparison here. 722 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV); 723 } 724 } 725 726 // Only lower this if the icmp is the only user of the GEP or if we expect 727 // the result to fold to a constant! 728 if (TD && 729 GEPsInBounds && 730 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) && 731 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) { 732 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) 733 Value *L = EmitGEPOffset(GEPLHS); 734 Value *R = EmitGEPOffset(GEPRHS); 735 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R); 736 } 737 } 738 return 0; 739} 740 741/// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X". 742Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI, 743 Value *X, ConstantInt *CI, 744 ICmpInst::Predicate Pred, 745 Value *TheAdd) { 746 // If we have X+0, exit early (simplifying logic below) and let it get folded 747 // elsewhere. icmp X+0, X -> icmp X, X 748 if (CI->isZero()) { 749 bool isTrue = ICmpInst::isTrueWhenEqual(Pred); 750 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 751 } 752 753 // (X+4) == X -> false. 754 if (Pred == ICmpInst::ICMP_EQ) 755 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext())); 756 757 // (X+4) != X -> true. 758 if (Pred == ICmpInst::ICMP_NE) 759 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext())); 760 761 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, 762 // so the values can never be equal. Similarly for all other "or equals" 763 // operators. 764 765 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255 766 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253 767 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0 768 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 769 Value *R = 770 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI); 771 return new ICmpInst(ICmpInst::ICMP_UGT, X, R); 772 } 773 774 // (X+1) >u X --> X <u (0-1) --> X != 255 775 // (X+2) >u X --> X <u (0-2) --> X <u 254 776 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0 777 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) 778 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI)); 779 780 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits(); 781 ConstantInt *SMax = ConstantInt::get(X->getContext(), 782 APInt::getSignedMaxValue(BitWidth)); 783 784 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127 785 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125 786 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0 787 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1 788 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126 789 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127 790 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 791 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI)); 792 793 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127 794 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126 795 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1 796 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2 797 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126 798 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128 799 800 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); 801 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1); 802 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C)); 803} 804 805/// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS 806/// and CmpRHS are both known to be integer constants. 807Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI, 808 ConstantInt *DivRHS) { 809 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1)); 810 const APInt &CmpRHSV = CmpRHS->getValue(); 811 812 // FIXME: If the operand types don't match the type of the divide 813 // then don't attempt this transform. The code below doesn't have the 814 // logic to deal with a signed divide and an unsigned compare (and 815 // vice versa). This is because (x /s C1) <s C2 produces different 816 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even 817 // (x /u C1) <u C2. Simply casting the operands and result won't 818 // work. :( The if statement below tests that condition and bails 819 // if it finds it. 820 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv; 821 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned()) 822 return 0; 823 if (DivRHS->isZero()) 824 return 0; // The ProdOV computation fails on divide by zero. 825 if (DivIsSigned && DivRHS->isAllOnesValue()) 826 return 0; // The overflow computation also screws up here 827 if (DivRHS->isOne()) { 828 // This eliminates some funny cases with INT_MIN. 829 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X. 830 return &ICI; 831 } 832 833 // Compute Prod = CI * DivRHS. We are essentially solving an equation 834 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and 835 // C2 (CI). By solving for X we can turn this into a range check 836 // instead of computing a divide. 837 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS); 838 839 // Determine if the product overflows by seeing if the product is 840 // not equal to the divide. Make sure we do the same kind of divide 841 // as in the LHS instruction that we're folding. 842 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) : 843 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS; 844 845 // Get the ICmp opcode 846 ICmpInst::Predicate Pred = ICI.getPredicate(); 847 848 /// If the division is known to be exact, then there is no remainder from the 849 /// divide, so the covered range size is unit, otherwise it is the divisor. 850 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS; 851 852 // Figure out the interval that is being checked. For example, a comparison 853 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). 854 // Compute this interval based on the constants involved and the signedness of 855 // the compare/divide. This computes a half-open interval, keeping track of 856 // whether either value in the interval overflows. After analysis each 857 // overflow variable is set to 0 if it's corresponding bound variable is valid 858 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. 859 int LoOverflow = 0, HiOverflow = 0; 860 Constant *LoBound = 0, *HiBound = 0; 861 862 if (!DivIsSigned) { // udiv 863 // e.g. X/5 op 3 --> [15, 20) 864 LoBound = Prod; 865 HiOverflow = LoOverflow = ProdOV; 866 if (!HiOverflow) { 867 // If this is not an exact divide, then many values in the range collapse 868 // to the same result value. 869 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false); 870 } 871 872 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0. 873 if (CmpRHSV == 0) { // (X / pos) op 0 874 // Can't overflow. e.g. X/2 op 0 --> [-1, 2) 875 LoBound = ConstantExpr::getNeg(SubOne(RangeSize)); 876 HiBound = RangeSize; 877 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos 878 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) 879 HiOverflow = LoOverflow = ProdOV; 880 if (!HiOverflow) 881 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true); 882 } else { // (X / pos) op neg 883 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) 884 HiBound = AddOne(Prod); 885 LoOverflow = HiOverflow = ProdOV ? -1 : 0; 886 if (!LoOverflow) { 887 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 888 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0; 889 } 890 } 891 } else if (DivRHS->isNegative()) { // Divisor is < 0. 892 if (DivI->isExact()) 893 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 894 if (CmpRHSV == 0) { // (X / neg) op 0 895 // e.g. X/-5 op 0 --> [-4, 5) 896 LoBound = AddOne(RangeSize); 897 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 898 if (HiBound == DivRHS) { // -INTMIN = INTMIN 899 HiOverflow = 1; // [INTMIN+1, overflow) 900 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN 901 } 902 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos 903 // e.g. X/-5 op 3 --> [-19, -14) 904 HiBound = AddOne(Prod); 905 HiOverflow = LoOverflow = ProdOV ? -1 : 0; 906 if (!LoOverflow) 907 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0; 908 } else { // (X / neg) op neg 909 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) 910 LoOverflow = HiOverflow = ProdOV; 911 if (!HiOverflow) 912 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true); 913 } 914 915 // Dividing by a negative swaps the condition. LT <-> GT 916 Pred = ICmpInst::getSwappedPredicate(Pred); 917 } 918 919 Value *X = DivI->getOperand(0); 920 switch (Pred) { 921 default: llvm_unreachable("Unhandled icmp opcode!"); 922 case ICmpInst::ICMP_EQ: 923 if (LoOverflow && HiOverflow) 924 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 925 if (HiOverflow) 926 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 927 ICmpInst::ICMP_UGE, X, LoBound); 928 if (LoOverflow) 929 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 930 ICmpInst::ICMP_ULT, X, HiBound); 931 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 932 DivIsSigned, true)); 933 case ICmpInst::ICMP_NE: 934 if (LoOverflow && HiOverflow) 935 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 936 if (HiOverflow) 937 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 938 ICmpInst::ICMP_ULT, X, LoBound); 939 if (LoOverflow) 940 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 941 ICmpInst::ICMP_UGE, X, HiBound); 942 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 943 DivIsSigned, false)); 944 case ICmpInst::ICMP_ULT: 945 case ICmpInst::ICMP_SLT: 946 if (LoOverflow == +1) // Low bound is greater than input range. 947 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 948 if (LoOverflow == -1) // Low bound is less than input range. 949 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 950 return new ICmpInst(Pred, X, LoBound); 951 case ICmpInst::ICMP_UGT: 952 case ICmpInst::ICMP_SGT: 953 if (HiOverflow == +1) // High bound greater than input range. 954 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 955 if (HiOverflow == -1) // High bound less than input range. 956 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 957 if (Pred == ICmpInst::ICMP_UGT) 958 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound); 959 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound); 960 } 961} 962 963/// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)". 964Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr, 965 ConstantInt *ShAmt) { 966 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue(); 967 968 // Check that the shift amount is in range. If not, don't perform 969 // undefined shifts. When the shift is visited it will be 970 // simplified. 971 uint32_t TypeBits = CmpRHSV.getBitWidth(); 972 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 973 if (ShAmtVal >= TypeBits || ShAmtVal == 0) 974 return 0; 975 976 if (!ICI.isEquality()) { 977 // If we have an unsigned comparison and an ashr, we can't simplify this. 978 // Similarly for signed comparisons with lshr. 979 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr)) 980 return 0; 981 982 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv 983 // by a power of 2. Since we already have logic to simplify these, 984 // transform to div and then simplify the resultant comparison. 985 if (Shr->getOpcode() == Instruction::AShr && 986 (!Shr->isExact() || ShAmtVal == TypeBits - 1)) 987 return 0; 988 989 // Revisit the shift (to delete it). 990 Worklist.Add(Shr); 991 992 Constant *DivCst = 993 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal)); 994 995 Value *Tmp = 996 Shr->getOpcode() == Instruction::AShr ? 997 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) : 998 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()); 999 1000 ICI.setOperand(0, Tmp); 1001 1002 // If the builder folded the binop, just return it. 1003 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp); 1004 if (TheDiv == 0) 1005 return &ICI; 1006 1007 // Otherwise, fold this div/compare. 1008 assert(TheDiv->getOpcode() == Instruction::SDiv || 1009 TheDiv->getOpcode() == Instruction::UDiv); 1010 1011 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst)); 1012 assert(Res && "This div/cst should have folded!"); 1013 return Res; 1014 } 1015 1016 1017 // If we are comparing against bits always shifted out, the 1018 // comparison cannot succeed. 1019 APInt Comp = CmpRHSV << ShAmtVal; 1020 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp); 1021 if (Shr->getOpcode() == Instruction::LShr) 1022 Comp = Comp.lshr(ShAmtVal); 1023 else 1024 Comp = Comp.ashr(ShAmtVal); 1025 1026 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero. 1027 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1028 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1029 IsICMP_NE); 1030 return ReplaceInstUsesWith(ICI, Cst); 1031 } 1032 1033 // Otherwise, check to see if the bits shifted out are known to be zero. 1034 // If so, we can compare against the unshifted value: 1035 // (X & 4) >> 1 == 2 --> (X & 4) == 4. 1036 if (Shr->hasOneUse() && Shr->isExact()) 1037 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS); 1038 1039 if (Shr->hasOneUse()) { 1040 // Otherwise strength reduce the shift into an and. 1041 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal)); 1042 Constant *Mask = ConstantInt::get(ICI.getContext(), Val); 1043 1044 Value *And = Builder->CreateAnd(Shr->getOperand(0), 1045 Mask, Shr->getName()+".mask"); 1046 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS); 1047 } 1048 return 0; 1049} 1050 1051 1052/// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)". 1053/// 1054Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI, 1055 Instruction *LHSI, 1056 ConstantInt *RHS) { 1057 const APInt &RHSV = RHS->getValue(); 1058 1059 switch (LHSI->getOpcode()) { 1060 case Instruction::Trunc: 1061 if (ICI.isEquality() && LHSI->hasOneUse()) { 1062 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all 1063 // of the high bits truncated out of x are known. 1064 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(), 1065 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1066 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0); 1067 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne); 1068 1069 // If all the high bits are known, we can do this xform. 1070 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) { 1071 // Pull in the high bits from known-ones set. 1072 APInt NewRHS = RHS->getValue().zext(SrcBits); 1073 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits); 1074 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1075 ConstantInt::get(ICI.getContext(), NewRHS)); 1076 } 1077 } 1078 break; 1079 1080 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI) 1081 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1082 // If this is a comparison that tests the signbit (X < 0) or (x > -1), 1083 // fold the xor. 1084 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) || 1085 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) { 1086 Value *CompareVal = LHSI->getOperand(0); 1087 1088 // If the sign bit of the XorCST is not set, there is no change to 1089 // the operation, just stop using the Xor. 1090 if (!XorCST->isNegative()) { 1091 ICI.setOperand(0, CompareVal); 1092 Worklist.Add(LHSI); 1093 return &ICI; 1094 } 1095 1096 // Was the old condition true if the operand is positive? 1097 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT; 1098 1099 // If so, the new one isn't. 1100 isTrueIfPositive ^= true; 1101 1102 if (isTrueIfPositive) 1103 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, 1104 SubOne(RHS)); 1105 else 1106 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, 1107 AddOne(RHS)); 1108 } 1109 1110 if (LHSI->hasOneUse()) { 1111 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit)) 1112 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) { 1113 const APInt &SignBit = XorCST->getValue(); 1114 ICmpInst::Predicate Pred = ICI.isSigned() 1115 ? ICI.getUnsignedPredicate() 1116 : ICI.getSignedPredicate(); 1117 return new ICmpInst(Pred, LHSI->getOperand(0), 1118 ConstantInt::get(ICI.getContext(), 1119 RHSV ^ SignBit)); 1120 } 1121 1122 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A) 1123 if (!ICI.isEquality() && XorCST->isMaxValue(true)) { 1124 const APInt &NotSignBit = XorCST->getValue(); 1125 ICmpInst::Predicate Pred = ICI.isSigned() 1126 ? ICI.getUnsignedPredicate() 1127 : ICI.getSignedPredicate(); 1128 Pred = ICI.getSwappedPredicate(Pred); 1129 return new ICmpInst(Pred, LHSI->getOperand(0), 1130 ConstantInt::get(ICI.getContext(), 1131 RHSV ^ NotSignBit)); 1132 } 1133 } 1134 } 1135 break; 1136 case Instruction::And: // (icmp pred (and X, AndCST), RHS) 1137 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) && 1138 LHSI->getOperand(0)->hasOneUse()) { 1139 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1)); 1140 1141 // If the LHS is an AND of a truncating cast, we can widen the 1142 // and/compare to be the input width without changing the value 1143 // produced, eliminating a cast. 1144 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) { 1145 // We can do this transformation if either the AND constant does not 1146 // have its sign bit set or if it is an equality comparison. 1147 // Extending a relational comparison when we're checking the sign 1148 // bit would not work. 1149 if (ICI.isEquality() || 1150 (!AndCST->isNegative() && RHSV.isNonNegative())) { 1151 Value *NewAnd = 1152 Builder->CreateAnd(Cast->getOperand(0), 1153 ConstantExpr::getZExt(AndCST, Cast->getSrcTy())); 1154 NewAnd->takeName(LHSI); 1155 return new ICmpInst(ICI.getPredicate(), NewAnd, 1156 ConstantExpr::getZExt(RHS, Cast->getSrcTy())); 1157 } 1158 } 1159 1160 // If the LHS is an AND of a zext, and we have an equality compare, we can 1161 // shrink the and/compare to the smaller type, eliminating the cast. 1162 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) { 1163 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy()); 1164 // Make sure we don't compare the upper bits, SimplifyDemandedBits 1165 // should fold the icmp to true/false in that case. 1166 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) { 1167 Value *NewAnd = 1168 Builder->CreateAnd(Cast->getOperand(0), 1169 ConstantExpr::getTrunc(AndCST, Ty)); 1170 NewAnd->takeName(LHSI); 1171 return new ICmpInst(ICI.getPredicate(), NewAnd, 1172 ConstantExpr::getTrunc(RHS, Ty)); 1173 } 1174 } 1175 1176 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare 1177 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This 1178 // happens a LOT in code produced by the C front-end, for bitfield 1179 // access. 1180 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0)); 1181 if (Shift && !Shift->isShift()) 1182 Shift = 0; 1183 1184 ConstantInt *ShAmt; 1185 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0; 1186 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift. 1187 Type *AndTy = AndCST->getType(); // Type of the and. 1188 1189 // We can fold this as long as we can't shift unknown bits 1190 // into the mask. This can only happen with signed shift 1191 // rights, as they sign-extend. 1192 if (ShAmt) { 1193 bool CanFold = Shift->isLogicalShift(); 1194 if (!CanFold) { 1195 // To test for the bad case of the signed shr, see if any 1196 // of the bits shifted in could be tested after the mask. 1197 uint32_t TyBits = Ty->getPrimitiveSizeInBits(); 1198 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits); 1199 1200 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits(); 1201 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) & 1202 AndCST->getValue()) == 0) 1203 CanFold = true; 1204 } 1205 1206 if (CanFold) { 1207 Constant *NewCst; 1208 if (Shift->getOpcode() == Instruction::Shl) 1209 NewCst = ConstantExpr::getLShr(RHS, ShAmt); 1210 else 1211 NewCst = ConstantExpr::getShl(RHS, ShAmt); 1212 1213 // Check to see if we are shifting out any of the bits being 1214 // compared. 1215 if (ConstantExpr::get(Shift->getOpcode(), 1216 NewCst, ShAmt) != RHS) { 1217 // If we shifted bits out, the fold is not going to work out. 1218 // As a special case, check to see if this means that the 1219 // result is always true or false now. 1220 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1221 return ReplaceInstUsesWith(ICI, 1222 ConstantInt::getFalse(ICI.getContext())); 1223 if (ICI.getPredicate() == ICmpInst::ICMP_NE) 1224 return ReplaceInstUsesWith(ICI, 1225 ConstantInt::getTrue(ICI.getContext())); 1226 } else { 1227 ICI.setOperand(1, NewCst); 1228 Constant *NewAndCST; 1229 if (Shift->getOpcode() == Instruction::Shl) 1230 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt); 1231 else 1232 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt); 1233 LHSI->setOperand(1, NewAndCST); 1234 LHSI->setOperand(0, Shift->getOperand(0)); 1235 Worklist.Add(Shift); // Shift is dead. 1236 return &ICI; 1237 } 1238 } 1239 } 1240 1241 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is 1242 // preferable because it allows the C<<Y expression to be hoisted out 1243 // of a loop if Y is invariant and X is not. 1244 if (Shift && Shift->hasOneUse() && RHSV == 0 && 1245 ICI.isEquality() && !Shift->isArithmeticShift() && 1246 !isa<Constant>(Shift->getOperand(0))) { 1247 // Compute C << Y. 1248 Value *NS; 1249 if (Shift->getOpcode() == Instruction::LShr) { 1250 NS = Builder->CreateShl(AndCST, Shift->getOperand(1)); 1251 } else { 1252 // Insert a logical shift. 1253 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1)); 1254 } 1255 1256 // Compute X & (C << Y). 1257 Value *NewAnd = 1258 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName()); 1259 1260 ICI.setOperand(0, NewAnd); 1261 return &ICI; 1262 } 1263 1264 // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any 1265 // bit set in (X & AndCST) will produce a result greater than RHSV. 1266 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) { 1267 unsigned NTZ = AndCST->getValue().countTrailingZeros(); 1268 if ((NTZ < AndCST->getBitWidth()) && 1269 APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV)) 1270 return new ICmpInst(ICmpInst::ICMP_NE, LHSI, 1271 Constant::getNullValue(RHS->getType())); 1272 } 1273 } 1274 1275 // Try to optimize things like "A[i]&42 == 0" to index computations. 1276 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) { 1277 if (GetElementPtrInst *GEP = 1278 dyn_cast<GetElementPtrInst>(LI->getOperand(0))) 1279 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 1280 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 1281 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) { 1282 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1)); 1283 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C)) 1284 return Res; 1285 } 1286 } 1287 break; 1288 1289 case Instruction::Or: { 1290 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse()) 1291 break; 1292 Value *P, *Q; 1293 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) { 1294 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 1295 // -> and (icmp eq P, null), (icmp eq Q, null). 1296 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P, 1297 Constant::getNullValue(P->getType())); 1298 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q, 1299 Constant::getNullValue(Q->getType())); 1300 Instruction *Op; 1301 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1302 Op = BinaryOperator::CreateAnd(ICIP, ICIQ); 1303 else 1304 Op = BinaryOperator::CreateOr(ICIP, ICIQ); 1305 return Op; 1306 } 1307 break; 1308 } 1309 1310 case Instruction::Mul: { // (icmp pred (mul X, Val), CI) 1311 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1312 if (!Val) break; 1313 1314 // If this is a signed comparison to 0 and the mul is sign preserving, 1315 // use the mul LHS operand instead. 1316 ICmpInst::Predicate pred = ICI.getPredicate(); 1317 if (isSignTest(pred, RHS) && !Val->isZero() && 1318 cast<BinaryOperator>(LHSI)->hasNoSignedWrap()) 1319 return new ICmpInst(Val->isNegative() ? 1320 ICmpInst::getSwappedPredicate(pred) : pred, 1321 LHSI->getOperand(0), 1322 Constant::getNullValue(RHS->getType())); 1323 1324 break; 1325 } 1326 1327 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI) 1328 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1329 if (!ShAmt) break; 1330 1331 uint32_t TypeBits = RHSV.getBitWidth(); 1332 1333 // Check that the shift amount is in range. If not, don't perform 1334 // undefined shifts. When the shift is visited it will be 1335 // simplified. 1336 if (ShAmt->uge(TypeBits)) 1337 break; 1338 1339 if (ICI.isEquality()) { 1340 // If we are comparing against bits always shifted out, the 1341 // comparison cannot succeed. 1342 Constant *Comp = 1343 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), 1344 ShAmt); 1345 if (Comp != RHS) {// Comparing against a bit that we know is zero. 1346 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1347 Constant *Cst = 1348 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE); 1349 return ReplaceInstUsesWith(ICI, Cst); 1350 } 1351 1352 // If the shift is NUW, then it is just shifting out zeros, no need for an 1353 // AND. 1354 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap()) 1355 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1356 ConstantExpr::getLShr(RHS, ShAmt)); 1357 1358 // If the shift is NSW and we compare to 0, then it is just shifting out 1359 // sign bits, no need for an AND either. 1360 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0) 1361 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1362 ConstantExpr::getLShr(RHS, ShAmt)); 1363 1364 if (LHSI->hasOneUse()) { 1365 // Otherwise strength reduce the shift into an and. 1366 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 1367 Constant *Mask = 1368 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits, 1369 TypeBits-ShAmtVal)); 1370 1371 Value *And = 1372 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask"); 1373 return new ICmpInst(ICI.getPredicate(), And, 1374 ConstantExpr::getLShr(RHS, ShAmt)); 1375 } 1376 } 1377 1378 // If this is a signed comparison to 0 and the shift is sign preserving, 1379 // use the shift LHS operand instead. 1380 ICmpInst::Predicate pred = ICI.getPredicate(); 1381 if (isSignTest(pred, RHS) && 1382 cast<BinaryOperator>(LHSI)->hasNoSignedWrap()) 1383 return new ICmpInst(pred, 1384 LHSI->getOperand(0), 1385 Constant::getNullValue(RHS->getType())); 1386 1387 // Otherwise, if this is a comparison of the sign bit, simplify to and/test. 1388 bool TrueIfSigned = false; 1389 if (LHSI->hasOneUse() && 1390 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) { 1391 // (X << 31) <s 0 --> (X&1) != 0 1392 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(), 1393 APInt::getOneBitSet(TypeBits, 1394 TypeBits-ShAmt->getZExtValue()-1)); 1395 Value *And = 1396 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask"); 1397 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, 1398 And, Constant::getNullValue(And->getType())); 1399 } 1400 1401 // Transform (icmp pred iM (shl iM %v, N), CI) 1402 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N)) 1403 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N. 1404 // This enables to get rid of the shift in favor of a trunc which can be 1405 // free on the target. It has the additional benefit of comparing to a 1406 // smaller constant, which will be target friendly. 1407 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1); 1408 if (LHSI->hasOneUse() && 1409 Amt != 0 && RHSV.countTrailingZeros() >= Amt) { 1410 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt); 1411 Constant *NCI = ConstantExpr::getTrunc( 1412 ConstantExpr::getAShr(RHS, 1413 ConstantInt::get(RHS->getType(), Amt)), 1414 NTy); 1415 return new ICmpInst(ICI.getPredicate(), 1416 Builder->CreateTrunc(LHSI->getOperand(0), NTy), 1417 NCI); 1418 } 1419 1420 break; 1421 } 1422 1423 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI) 1424 case Instruction::AShr: { 1425 // Handle equality comparisons of shift-by-constant. 1426 BinaryOperator *BO = cast<BinaryOperator>(LHSI); 1427 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1428 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt)) 1429 return Res; 1430 } 1431 1432 // Handle exact shr's. 1433 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) { 1434 if (RHSV.isMinValue()) 1435 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS); 1436 } 1437 break; 1438 } 1439 1440 case Instruction::SDiv: 1441 case Instruction::UDiv: 1442 // Fold: icmp pred ([us]div X, C1), C2 -> range test 1443 // Fold this div into the comparison, producing a range check. 1444 // Determine, based on the divide type, what the range is being 1445 // checked. If there is an overflow on the low or high side, remember 1446 // it, otherwise compute the range [low, hi) bounding the new value. 1447 // See: InsertRangeTest above for the kinds of replacements possible. 1448 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) 1449 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI), 1450 DivRHS)) 1451 return R; 1452 break; 1453 1454 case Instruction::Add: 1455 // Fold: icmp pred (add X, C1), C2 1456 if (!ICI.isEquality()) { 1457 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1458 if (!LHSC) break; 1459 const APInt &LHSV = LHSC->getValue(); 1460 1461 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV) 1462 .subtract(LHSV); 1463 1464 if (ICI.isSigned()) { 1465 if (CR.getLower().isSignBit()) { 1466 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0), 1467 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1468 } else if (CR.getUpper().isSignBit()) { 1469 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0), 1470 ConstantInt::get(ICI.getContext(),CR.getLower())); 1471 } 1472 } else { 1473 if (CR.getLower().isMinValue()) { 1474 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), 1475 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1476 } else if (CR.getUpper().isMinValue()) { 1477 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), 1478 ConstantInt::get(ICI.getContext(),CR.getLower())); 1479 } 1480 } 1481 } 1482 break; 1483 } 1484 1485 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS. 1486 if (ICI.isEquality()) { 1487 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1488 1489 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and 1490 // the second operand is a constant, simplify a bit. 1491 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) { 1492 switch (BO->getOpcode()) { 1493 case Instruction::SRem: 1494 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. 1495 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){ 1496 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue(); 1497 if (V.sgt(1) && V.isPowerOf2()) { 1498 Value *NewRem = 1499 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1), 1500 BO->getName()); 1501 return new ICmpInst(ICI.getPredicate(), NewRem, 1502 Constant::getNullValue(BO->getType())); 1503 } 1504 } 1505 break; 1506 case Instruction::Add: 1507 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants. 1508 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1509 if (BO->hasOneUse()) 1510 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1511 ConstantExpr::getSub(RHS, BOp1C)); 1512 } else if (RHSV == 0) { 1513 // Replace ((add A, B) != 0) with (A != -B) if A or B is 1514 // efficiently invertible, or if the add has just this one use. 1515 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1); 1516 1517 if (Value *NegVal = dyn_castNegVal(BOp1)) 1518 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal); 1519 if (Value *NegVal = dyn_castNegVal(BOp0)) 1520 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1); 1521 if (BO->hasOneUse()) { 1522 Value *Neg = Builder->CreateNeg(BOp1); 1523 Neg->takeName(BO); 1524 return new ICmpInst(ICI.getPredicate(), BOp0, Neg); 1525 } 1526 } 1527 break; 1528 case Instruction::Xor: 1529 // For the xor case, we can xor two constants together, eliminating 1530 // the explicit xor. 1531 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) { 1532 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1533 ConstantExpr::getXor(RHS, BOC)); 1534 } else if (RHSV == 0) { 1535 // Replace ((xor A, B) != 0) with (A != B) 1536 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1537 BO->getOperand(1)); 1538 } 1539 break; 1540 case Instruction::Sub: 1541 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants. 1542 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) { 1543 if (BO->hasOneUse()) 1544 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1), 1545 ConstantExpr::getSub(BOp0C, RHS)); 1546 } else if (RHSV == 0) { 1547 // Replace ((sub A, B) != 0) with (A != B) 1548 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1549 BO->getOperand(1)); 1550 } 1551 break; 1552 case Instruction::Or: 1553 // If bits are being or'd in that are not present in the constant we 1554 // are comparing against, then the comparison could never succeed! 1555 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1556 Constant *NotCI = ConstantExpr::getNot(RHS); 1557 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue()) 1558 return ReplaceInstUsesWith(ICI, 1559 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1560 isICMP_NE)); 1561 } 1562 break; 1563 1564 case Instruction::And: 1565 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1566 // If bits are being compared against that are and'd out, then the 1567 // comparison can never succeed! 1568 if ((RHSV & ~BOC->getValue()) != 0) 1569 return ReplaceInstUsesWith(ICI, 1570 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1571 isICMP_NE)); 1572 1573 // If we have ((X & C) == C), turn it into ((X & C) != 0). 1574 if (RHS == BOC && RHSV.isPowerOf2()) 1575 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : 1576 ICmpInst::ICMP_NE, LHSI, 1577 Constant::getNullValue(RHS->getType())); 1578 1579 // Don't perform the following transforms if the AND has multiple uses 1580 if (!BO->hasOneUse()) 1581 break; 1582 1583 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0 1584 if (BOC->getValue().isSignBit()) { 1585 Value *X = BO->getOperand(0); 1586 Constant *Zero = Constant::getNullValue(X->getType()); 1587 ICmpInst::Predicate pred = isICMP_NE ? 1588 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; 1589 return new ICmpInst(pred, X, Zero); 1590 } 1591 1592 // ((X & ~7) == 0) --> X < 8 1593 if (RHSV == 0 && isHighOnes(BOC)) { 1594 Value *X = BO->getOperand(0); 1595 Constant *NegX = ConstantExpr::getNeg(BOC); 1596 ICmpInst::Predicate pred = isICMP_NE ? 1597 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; 1598 return new ICmpInst(pred, X, NegX); 1599 } 1600 } 1601 break; 1602 case Instruction::Mul: 1603 if (RHSV == 0 && BO->hasNoSignedWrap()) { 1604 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1605 // The trivial case (mul X, 0) is handled by InstSimplify 1606 // General case : (mul X, C) != 0 iff X != 0 1607 // (mul X, C) == 0 iff X == 0 1608 if (!BOC->isZero()) 1609 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1610 Constant::getNullValue(RHS->getType())); 1611 } 1612 } 1613 break; 1614 default: break; 1615 } 1616 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) { 1617 // Handle icmp {eq|ne} <intrinsic>, intcst. 1618 switch (II->getIntrinsicID()) { 1619 case Intrinsic::bswap: 1620 Worklist.Add(II); 1621 ICI.setOperand(0, II->getArgOperand(0)); 1622 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap())); 1623 return &ICI; 1624 case Intrinsic::ctlz: 1625 case Intrinsic::cttz: 1626 // ctz(A) == bitwidth(a) -> A == 0 and likewise for != 1627 if (RHSV == RHS->getType()->getBitWidth()) { 1628 Worklist.Add(II); 1629 ICI.setOperand(0, II->getArgOperand(0)); 1630 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0)); 1631 return &ICI; 1632 } 1633 break; 1634 case Intrinsic::ctpop: 1635 // popcount(A) == 0 -> A == 0 and likewise for != 1636 if (RHS->isZero()) { 1637 Worklist.Add(II); 1638 ICI.setOperand(0, II->getArgOperand(0)); 1639 ICI.setOperand(1, RHS); 1640 return &ICI; 1641 } 1642 break; 1643 default: 1644 break; 1645 } 1646 } 1647 } 1648 return 0; 1649} 1650 1651/// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst). 1652/// We only handle extending casts so far. 1653/// 1654Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) { 1655 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0)); 1656 Value *LHSCIOp = LHSCI->getOperand(0); 1657 Type *SrcTy = LHSCIOp->getType(); 1658 Type *DestTy = LHSCI->getType(); 1659 Value *RHSCIOp; 1660 1661 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the 1662 // integer type is the same size as the pointer type. 1663 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt && 1664 TD->getPointerSizeInBits() == 1665 cast<IntegerType>(DestTy)->getBitWidth()) { 1666 Value *RHSOp = 0; 1667 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) { 1668 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy); 1669 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) { 1670 RHSOp = RHSC->getOperand(0); 1671 // If the pointer types don't match, insert a bitcast. 1672 if (LHSCIOp->getType() != RHSOp->getType()) 1673 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType()); 1674 } 1675 1676 if (RHSOp) 1677 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp); 1678 } 1679 1680 // The code below only handles extension cast instructions, so far. 1681 // Enforce this. 1682 if (LHSCI->getOpcode() != Instruction::ZExt && 1683 LHSCI->getOpcode() != Instruction::SExt) 1684 return 0; 1685 1686 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt; 1687 bool isSignedCmp = ICI.isSigned(); 1688 1689 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) { 1690 // Not an extension from the same type? 1691 RHSCIOp = CI->getOperand(0); 1692 if (RHSCIOp->getType() != LHSCIOp->getType()) 1693 return 0; 1694 1695 // If the signedness of the two casts doesn't agree (i.e. one is a sext 1696 // and the other is a zext), then we can't handle this. 1697 if (CI->getOpcode() != LHSCI->getOpcode()) 1698 return 0; 1699 1700 // Deal with equality cases early. 1701 if (ICI.isEquality()) 1702 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1703 1704 // A signed comparison of sign extended values simplifies into a 1705 // signed comparison. 1706 if (isSignedCmp && isSignedExt) 1707 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1708 1709 // The other three cases all fold into an unsigned comparison. 1710 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp); 1711 } 1712 1713 // If we aren't dealing with a constant on the RHS, exit early 1714 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1)); 1715 if (!CI) 1716 return 0; 1717 1718 // Compute the constant that would happen if we truncated to SrcTy then 1719 // reextended to DestTy. 1720 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy); 1721 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), 1722 Res1, DestTy); 1723 1724 // If the re-extended constant didn't change... 1725 if (Res2 == CI) { 1726 // Deal with equality cases early. 1727 if (ICI.isEquality()) 1728 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1729 1730 // A signed comparison of sign extended values simplifies into a 1731 // signed comparison. 1732 if (isSignedExt && isSignedCmp) 1733 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1734 1735 // The other three cases all fold into an unsigned comparison. 1736 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1); 1737 } 1738 1739 // The re-extended constant changed so the constant cannot be represented 1740 // in the shorter type. Consequently, we cannot emit a simple comparison. 1741 // All the cases that fold to true or false will have already been handled 1742 // by SimplifyICmpInst, so only deal with the tricky case. 1743 1744 if (isSignedCmp || !isSignedExt) 1745 return 0; 1746 1747 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases 1748 // should have been folded away previously and not enter in here. 1749 1750 // We're performing an unsigned comp with a sign extended value. 1751 // This is true if the input is >= 0. [aka >s -1] 1752 Constant *NegOne = Constant::getAllOnesValue(SrcTy); 1753 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName()); 1754 1755 // Finally, return the value computed. 1756 if (ICI.getPredicate() == ICmpInst::ICMP_ULT) 1757 return ReplaceInstUsesWith(ICI, Result); 1758 1759 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"); 1760 return BinaryOperator::CreateNot(Result); 1761} 1762 1763/// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form: 1764/// I = icmp ugt (add (add A, B), CI2), CI1 1765/// If this is of the form: 1766/// sum = a + b 1767/// if (sum+128 >u 255) 1768/// Then replace it with llvm.sadd.with.overflow.i8. 1769/// 1770static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B, 1771 ConstantInt *CI2, ConstantInt *CI1, 1772 InstCombiner &IC) { 1773 // The transformation we're trying to do here is to transform this into an 1774 // llvm.sadd.with.overflow. To do this, we have to replace the original add 1775 // with a narrower add, and discard the add-with-constant that is part of the 1776 // range check (if we can't eliminate it, this isn't profitable). 1777 1778 // In order to eliminate the add-with-constant, the compare can be its only 1779 // use. 1780 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0)); 1781 if (!AddWithCst->hasOneUse()) return 0; 1782 1783 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow. 1784 if (!CI2->getValue().isPowerOf2()) return 0; 1785 unsigned NewWidth = CI2->getValue().countTrailingZeros(); 1786 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0; 1787 1788 // The width of the new add formed is 1 more than the bias. 1789 ++NewWidth; 1790 1791 // Check to see that CI1 is an all-ones value with NewWidth bits. 1792 if (CI1->getBitWidth() == NewWidth || 1793 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth)) 1794 return 0; 1795 1796 // This is only really a signed overflow check if the inputs have been 1797 // sign-extended; check for that condition. For example, if CI2 is 2^31 and 1798 // the operands of the add are 64 bits wide, we need at least 33 sign bits. 1799 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1; 1800 if (IC.ComputeNumSignBits(A) < NeededSignBits || 1801 IC.ComputeNumSignBits(B) < NeededSignBits) 1802 return 0; 1803 1804 // In order to replace the original add with a narrower 1805 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant 1806 // and truncates that discard the high bits of the add. Verify that this is 1807 // the case. 1808 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0)); 1809 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end(); 1810 UI != E; ++UI) { 1811 if (*UI == AddWithCst) continue; 1812 1813 // Only accept truncates for now. We would really like a nice recursive 1814 // predicate like SimplifyDemandedBits, but which goes downwards the use-def 1815 // chain to see which bits of a value are actually demanded. If the 1816 // original add had another add which was then immediately truncated, we 1817 // could still do the transformation. 1818 TruncInst *TI = dyn_cast<TruncInst>(*UI); 1819 if (TI == 0 || 1820 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0; 1821 } 1822 1823 // If the pattern matches, truncate the inputs to the narrower type and 1824 // use the sadd_with_overflow intrinsic to efficiently compute both the 1825 // result and the overflow bit. 1826 Module *M = I.getParent()->getParent()->getParent(); 1827 1828 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth); 1829 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow, 1830 NewType); 1831 1832 InstCombiner::BuilderTy *Builder = IC.Builder; 1833 1834 // Put the new code above the original add, in case there are any uses of the 1835 // add between the add and the compare. 1836 Builder->SetInsertPoint(OrigAdd); 1837 1838 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc"); 1839 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc"); 1840 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd"); 1841 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result"); 1842 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType()); 1843 1844 // The inner add was the result of the narrow add, zero extended to the 1845 // wider type. Replace it with the result computed by the intrinsic. 1846 IC.ReplaceInstUsesWith(*OrigAdd, ZExt); 1847 1848 // The original icmp gets replaced with the overflow value. 1849 return ExtractValueInst::Create(Call, 1, "sadd.overflow"); 1850} 1851 1852static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV, 1853 InstCombiner &IC) { 1854 // Don't bother doing this transformation for pointers, don't do it for 1855 // vectors. 1856 if (!isa<IntegerType>(OrigAddV->getType())) return 0; 1857 1858 // If the add is a constant expr, then we don't bother transforming it. 1859 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV); 1860 if (OrigAdd == 0) return 0; 1861 1862 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1); 1863 1864 // Put the new code above the original add, in case there are any uses of the 1865 // add between the add and the compare. 1866 InstCombiner::BuilderTy *Builder = IC.Builder; 1867 Builder->SetInsertPoint(OrigAdd); 1868 1869 Module *M = I.getParent()->getParent()->getParent(); 1870 Type *Ty = LHS->getType(); 1871 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty); 1872 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd"); 1873 Value *Add = Builder->CreateExtractValue(Call, 0); 1874 1875 IC.ReplaceInstUsesWith(*OrigAdd, Add); 1876 1877 // The original icmp gets replaced with the overflow value. 1878 return ExtractValueInst::Create(Call, 1, "uadd.overflow"); 1879} 1880 1881// DemandedBitsLHSMask - When performing a comparison against a constant, 1882// it is possible that not all the bits in the LHS are demanded. This helper 1883// method computes the mask that IS demanded. 1884static APInt DemandedBitsLHSMask(ICmpInst &I, 1885 unsigned BitWidth, bool isSignCheck) { 1886 if (isSignCheck) 1887 return APInt::getSignBit(BitWidth); 1888 1889 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1)); 1890 if (!CI) return APInt::getAllOnesValue(BitWidth); 1891 const APInt &RHS = CI->getValue(); 1892 1893 switch (I.getPredicate()) { 1894 // For a UGT comparison, we don't care about any bits that 1895 // correspond to the trailing ones of the comparand. The value of these 1896 // bits doesn't impact the outcome of the comparison, because any value 1897 // greater than the RHS must differ in a bit higher than these due to carry. 1898 case ICmpInst::ICMP_UGT: { 1899 unsigned trailingOnes = RHS.countTrailingOnes(); 1900 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes); 1901 return ~lowBitsSet; 1902 } 1903 1904 // Similarly, for a ULT comparison, we don't care about the trailing zeros. 1905 // Any value less than the RHS must differ in a higher bit because of carries. 1906 case ICmpInst::ICMP_ULT: { 1907 unsigned trailingZeros = RHS.countTrailingZeros(); 1908 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros); 1909 return ~lowBitsSet; 1910 } 1911 1912 default: 1913 return APInt::getAllOnesValue(BitWidth); 1914 } 1915 1916} 1917 1918Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { 1919 bool Changed = false; 1920 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1921 1922 /// Orders the operands of the compare so that they are listed from most 1923 /// complex to least complex. This puts constants before unary operators, 1924 /// before binary operators. 1925 if (getComplexity(Op0) < getComplexity(Op1)) { 1926 I.swapOperands(); 1927 std::swap(Op0, Op1); 1928 Changed = true; 1929 } 1930 1931 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD)) 1932 return ReplaceInstUsesWith(I, V); 1933 1934 // comparing -val or val with non-zero is the same as just comparing val 1935 // ie, abs(val) != 0 -> val != 0 1936 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) 1937 { 1938 Value *Cond, *SelectTrue, *SelectFalse; 1939 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue), 1940 m_Value(SelectFalse)))) { 1941 if (Value *V = dyn_castNegVal(SelectTrue)) { 1942 if (V == SelectFalse) 1943 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 1944 } 1945 else if (Value *V = dyn_castNegVal(SelectFalse)) { 1946 if (V == SelectTrue) 1947 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 1948 } 1949 } 1950 } 1951 1952 Type *Ty = Op0->getType(); 1953 1954 // icmp's with boolean values can always be turned into bitwise operations 1955 if (Ty->isIntegerTy(1)) { 1956 switch (I.getPredicate()) { 1957 default: llvm_unreachable("Invalid icmp instruction!"); 1958 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B) 1959 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp"); 1960 return BinaryOperator::CreateNot(Xor); 1961 } 1962 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B 1963 return BinaryOperator::CreateXor(Op0, Op1); 1964 1965 case ICmpInst::ICMP_UGT: 1966 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult 1967 // FALL THROUGH 1968 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B 1969 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1970 return BinaryOperator::CreateAnd(Not, Op1); 1971 } 1972 case ICmpInst::ICMP_SGT: 1973 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt 1974 // FALL THROUGH 1975 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B 1976 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1977 return BinaryOperator::CreateAnd(Not, Op0); 1978 } 1979 case ICmpInst::ICMP_UGE: 1980 std::swap(Op0, Op1); // Change icmp uge -> icmp ule 1981 // FALL THROUGH 1982 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B 1983 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1984 return BinaryOperator::CreateOr(Not, Op1); 1985 } 1986 case ICmpInst::ICMP_SGE: 1987 std::swap(Op0, Op1); // Change icmp sge -> icmp sle 1988 // FALL THROUGH 1989 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B 1990 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1991 return BinaryOperator::CreateOr(Not, Op0); 1992 } 1993 } 1994 } 1995 1996 unsigned BitWidth = 0; 1997 if (Ty->isIntOrIntVectorTy()) 1998 BitWidth = Ty->getScalarSizeInBits(); 1999 else if (TD) // Pointers require TD info to get their size. 2000 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType()); 2001 2002 bool isSignBit = false; 2003 2004 // See if we are doing a comparison with a constant. 2005 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2006 Value *A = 0, *B = 0; 2007 2008 // Match the following pattern, which is a common idiom when writing 2009 // overflow-safe integer arithmetic function. The source performs an 2010 // addition in wider type, and explicitly checks for overflow using 2011 // comparisons against INT_MIN and INT_MAX. Simplify this by using the 2012 // sadd_with_overflow intrinsic. 2013 // 2014 // TODO: This could probably be generalized to handle other overflow-safe 2015 // operations if we worked out the formulas to compute the appropriate 2016 // magic constants. 2017 // 2018 // sum = a + b 2019 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8 2020 { 2021 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI 2022 if (I.getPredicate() == ICmpInst::ICMP_UGT && 2023 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2)))) 2024 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this)) 2025 return Res; 2026 } 2027 2028 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B) 2029 if (I.isEquality() && CI->isZero() && 2030 match(Op0, m_Sub(m_Value(A), m_Value(B)))) { 2031 // (icmp cond A B) if cond is equality 2032 return new ICmpInst(I.getPredicate(), A, B); 2033 } 2034 2035 // If we have an icmp le or icmp ge instruction, turn it into the 2036 // appropriate icmp lt or icmp gt instruction. This allows us to rely on 2037 // them being folded in the code below. The SimplifyICmpInst code has 2038 // already handled the edge cases for us, so we just assert on them. 2039 switch (I.getPredicate()) { 2040 default: break; 2041 case ICmpInst::ICMP_ULE: 2042 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE 2043 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, 2044 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2045 case ICmpInst::ICMP_SLE: 2046 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE 2047 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 2048 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2049 case ICmpInst::ICMP_UGE: 2050 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE 2051 return new ICmpInst(ICmpInst::ICMP_UGT, Op0, 2052 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2053 case ICmpInst::ICMP_SGE: 2054 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE 2055 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 2056 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2057 } 2058 2059 // If this comparison is a normal comparison, it demands all 2060 // bits, if it is a sign bit comparison, it only demands the sign bit. 2061 bool UnusedBit; 2062 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit); 2063 } 2064 2065 // See if we can fold the comparison based on range information we can get 2066 // by checking whether bits are known to be zero or one in the input. 2067 if (BitWidth != 0) { 2068 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0); 2069 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0); 2070 2071 if (SimplifyDemandedBits(I.getOperandUse(0), 2072 DemandedBitsLHSMask(I, BitWidth, isSignBit), 2073 Op0KnownZero, Op0KnownOne, 0)) 2074 return &I; 2075 if (SimplifyDemandedBits(I.getOperandUse(1), 2076 APInt::getAllOnesValue(BitWidth), 2077 Op1KnownZero, Op1KnownOne, 0)) 2078 return &I; 2079 2080 // Given the known and unknown bits, compute a range that the LHS could be 2081 // in. Compute the Min, Max and RHS values based on the known bits. For the 2082 // EQ and NE we use unsigned values. 2083 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0); 2084 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0); 2085 if (I.isSigned()) { 2086 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 2087 Op0Min, Op0Max); 2088 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 2089 Op1Min, Op1Max); 2090 } else { 2091 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 2092 Op0Min, Op0Max); 2093 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 2094 Op1Min, Op1Max); 2095 } 2096 2097 // If Min and Max are known to be the same, then SimplifyDemandedBits 2098 // figured out that the LHS is a constant. Just constant fold this now so 2099 // that code below can assume that Min != Max. 2100 if (!isa<Constant>(Op0) && Op0Min == Op0Max) 2101 return new ICmpInst(I.getPredicate(), 2102 ConstantInt::get(Op0->getType(), Op0Min), Op1); 2103 if (!isa<Constant>(Op1) && Op1Min == Op1Max) 2104 return new ICmpInst(I.getPredicate(), Op0, 2105 ConstantInt::get(Op1->getType(), Op1Min)); 2106 2107 // Based on the range information we know about the LHS, see if we can 2108 // simplify this comparison. For example, (x&4) < 8 is always true. 2109 switch (I.getPredicate()) { 2110 default: llvm_unreachable("Unknown icmp opcode!"); 2111 case ICmpInst::ICMP_EQ: { 2112 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2113 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2114 2115 // If all bits are known zero except for one, then we know at most one 2116 // bit is set. If the comparison is against zero, then this is a check 2117 // to see if *that* bit is set. 2118 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2119 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2120 // If the LHS is an AND with the same constant, look through it. 2121 Value *LHS = 0; 2122 ConstantInt *LHSC = 0; 2123 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2124 LHSC->getValue() != Op0KnownZeroInverted) 2125 LHS = Op0; 2126 2127 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2128 // then turn "((1 << x)&8) == 0" into "x != 3". 2129 Value *X = 0; 2130 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2131 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2132 return new ICmpInst(ICmpInst::ICMP_NE, X, 2133 ConstantInt::get(X->getType(), CmpVal)); 2134 } 2135 2136 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2137 // then turn "((8 >>u x)&1) == 0" into "x != 3". 2138 const APInt *CI; 2139 if (Op0KnownZeroInverted == 1 && 2140 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2141 return new ICmpInst(ICmpInst::ICMP_NE, X, 2142 ConstantInt::get(X->getType(), 2143 CI->countTrailingZeros())); 2144 } 2145 2146 break; 2147 } 2148 case ICmpInst::ICMP_NE: { 2149 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2150 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2151 2152 // If all bits are known zero except for one, then we know at most one 2153 // bit is set. If the comparison is against zero, then this is a check 2154 // to see if *that* bit is set. 2155 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2156 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2157 // If the LHS is an AND with the same constant, look through it. 2158 Value *LHS = 0; 2159 ConstantInt *LHSC = 0; 2160 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2161 LHSC->getValue() != Op0KnownZeroInverted) 2162 LHS = Op0; 2163 2164 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2165 // then turn "((1 << x)&8) != 0" into "x == 3". 2166 Value *X = 0; 2167 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2168 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2169 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2170 ConstantInt::get(X->getType(), CmpVal)); 2171 } 2172 2173 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2174 // then turn "((8 >>u x)&1) != 0" into "x == 3". 2175 const APInt *CI; 2176 if (Op0KnownZeroInverted == 1 && 2177 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2178 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2179 ConstantInt::get(X->getType(), 2180 CI->countTrailingZeros())); 2181 } 2182 2183 break; 2184 } 2185 case ICmpInst::ICMP_ULT: 2186 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B) 2187 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2188 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B) 2189 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2190 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B) 2191 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2192 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2193 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C 2194 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2195 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2196 2197 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear 2198 if (CI->isMinValue(true)) 2199 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 2200 Constant::getAllOnesValue(Op0->getType())); 2201 } 2202 break; 2203 case ICmpInst::ICMP_UGT: 2204 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B) 2205 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2206 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B) 2207 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2208 2209 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B) 2210 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2211 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2212 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C 2213 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2214 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2215 2216 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set 2217 if (CI->isMaxValue(true)) 2218 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 2219 Constant::getNullValue(Op0->getType())); 2220 } 2221 break; 2222 case ICmpInst::ICMP_SLT: 2223 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C) 2224 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2225 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C) 2226 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2227 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B) 2228 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2229 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2230 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C 2231 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2232 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2233 } 2234 break; 2235 case ICmpInst::ICMP_SGT: 2236 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B) 2237 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2238 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B) 2239 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2240 2241 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B) 2242 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2243 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2244 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C 2245 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2246 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2247 } 2248 break; 2249 case ICmpInst::ICMP_SGE: 2250 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"); 2251 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B) 2252 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2253 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B) 2254 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2255 break; 2256 case ICmpInst::ICMP_SLE: 2257 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"); 2258 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B) 2259 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2260 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B) 2261 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2262 break; 2263 case ICmpInst::ICMP_UGE: 2264 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"); 2265 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B) 2266 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2267 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B) 2268 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2269 break; 2270 case ICmpInst::ICMP_ULE: 2271 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"); 2272 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B) 2273 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2274 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B) 2275 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2276 break; 2277 } 2278 2279 // Turn a signed comparison into an unsigned one if both operands 2280 // are known to have the same sign. 2281 if (I.isSigned() && 2282 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) || 2283 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative()))) 2284 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1); 2285 } 2286 2287 // Test if the ICmpInst instruction is used exclusively by a select as 2288 // part of a minimum or maximum operation. If so, refrain from doing 2289 // any other folding. This helps out other analyses which understand 2290 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 2291 // and CodeGen. And in this case, at least one of the comparison 2292 // operands has at least one user besides the compare (the select), 2293 // which would often largely negate the benefit of folding anyway. 2294 if (I.hasOneUse()) 2295 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin())) 2296 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) || 2297 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1)) 2298 return 0; 2299 2300 // See if we are doing a comparison between a constant and an instruction that 2301 // can be folded into the comparison. 2302 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2303 // Since the RHS is a ConstantInt (CI), if the left hand side is an 2304 // instruction, see if that instruction also has constants so that the 2305 // instruction can be folded into the icmp 2306 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2307 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI)) 2308 return Res; 2309 } 2310 2311 // Handle icmp with constant (but not simple integer constant) RHS 2312 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2313 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2314 switch (LHSI->getOpcode()) { 2315 case Instruction::GetElementPtr: 2316 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null 2317 if (RHSC->isNullValue() && 2318 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices()) 2319 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2320 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2321 break; 2322 case Instruction::PHI: 2323 // Only fold icmp into the PHI if the phi and icmp are in the same 2324 // block. If in the same block, we're encouraging jump threading. If 2325 // not, we are just pessimizing the code by making an i1 phi. 2326 if (LHSI->getParent() == I.getParent()) 2327 if (Instruction *NV = FoldOpIntoPhi(I)) 2328 return NV; 2329 break; 2330 case Instruction::Select: { 2331 // If either operand of the select is a constant, we can fold the 2332 // comparison into the select arms, which will cause one to be 2333 // constant folded and the select turned into a bitwise or. 2334 Value *Op1 = 0, *Op2 = 0; 2335 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) 2336 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2337 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) 2338 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2339 2340 // We only want to perform this transformation if it will not lead to 2341 // additional code. This is true if either both sides of the select 2342 // fold to a constant (in which case the icmp is replaced with a select 2343 // which will usually simplify) or this is the only user of the 2344 // select (in which case we are trading a select+icmp for a simpler 2345 // select+icmp). 2346 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) { 2347 if (!Op1) 2348 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1), 2349 RHSC, I.getName()); 2350 if (!Op2) 2351 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2), 2352 RHSC, I.getName()); 2353 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2354 } 2355 break; 2356 } 2357 case Instruction::IntToPtr: 2358 // icmp pred inttoptr(X), null -> icmp pred X, 0 2359 if (RHSC->isNullValue() && TD && 2360 TD->getIntPtrType(RHSC->getContext()) == 2361 LHSI->getOperand(0)->getType()) 2362 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2363 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2364 break; 2365 2366 case Instruction::Load: 2367 // Try to optimize things like "A[i] > 4" to index computations. 2368 if (GetElementPtrInst *GEP = 2369 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2370 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2371 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2372 !cast<LoadInst>(LHSI)->isVolatile()) 2373 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2374 return Res; 2375 } 2376 break; 2377 } 2378 } 2379 2380 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now. 2381 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0)) 2382 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I)) 2383 return NI; 2384 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) 2385 if (Instruction *NI = FoldGEPICmp(GEP, Op0, 2386 ICmpInst::getSwappedPredicate(I.getPredicate()), I)) 2387 return NI; 2388 2389 // Test to see if the operands of the icmp are casted versions of other 2390 // values. If the ptr->ptr cast can be stripped off both arguments, we do so 2391 // now. 2392 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) { 2393 if (Op0->getType()->isPointerTy() && 2394 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) { 2395 // We keep moving the cast from the left operand over to the right 2396 // operand, where it can often be eliminated completely. 2397 Op0 = CI->getOperand(0); 2398 2399 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast 2400 // so eliminate it as well. 2401 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1)) 2402 Op1 = CI2->getOperand(0); 2403 2404 // If Op1 is a constant, we can fold the cast into the constant. 2405 if (Op0->getType() != Op1->getType()) { 2406 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 2407 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType()); 2408 } else { 2409 // Otherwise, cast the RHS right before the icmp 2410 Op1 = Builder->CreateBitCast(Op1, Op0->getType()); 2411 } 2412 } 2413 return new ICmpInst(I.getPredicate(), Op0, Op1); 2414 } 2415 } 2416 2417 if (isa<CastInst>(Op0)) { 2418 // Handle the special case of: icmp (cast bool to X), <cst> 2419 // This comes up when you have code like 2420 // int X = A < B; 2421 // if (X) ... 2422 // For generality, we handle any zero-extension of any operand comparison 2423 // with a constant or another cast from the same type. 2424 if (isa<Constant>(Op1) || isa<CastInst>(Op1)) 2425 if (Instruction *R = visitICmpInstWithCastAndCast(I)) 2426 return R; 2427 } 2428 2429 // Special logic for binary operators. 2430 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0); 2431 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1); 2432 if (BO0 || BO1) { 2433 CmpInst::Predicate Pred = I.getPredicate(); 2434 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false; 2435 if (BO0 && isa<OverflowingBinaryOperator>(BO0)) 2436 NoOp0WrapProblem = ICmpInst::isEquality(Pred) || 2437 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) || 2438 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap()); 2439 if (BO1 && isa<OverflowingBinaryOperator>(BO1)) 2440 NoOp1WrapProblem = ICmpInst::isEquality(Pred) || 2441 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) || 2442 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap()); 2443 2444 // Analyze the case when either Op0 or Op1 is an add instruction. 2445 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null). 2446 Value *A = 0, *B = 0, *C = 0, *D = 0; 2447 if (BO0 && BO0->getOpcode() == Instruction::Add) 2448 A = BO0->getOperand(0), B = BO0->getOperand(1); 2449 if (BO1 && BO1->getOpcode() == Instruction::Add) 2450 C = BO1->getOperand(0), D = BO1->getOperand(1); 2451 2452 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2453 if ((A == Op1 || B == Op1) && NoOp0WrapProblem) 2454 return new ICmpInst(Pred, A == Op1 ? B : A, 2455 Constant::getNullValue(Op1->getType())); 2456 2457 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2458 if ((C == Op0 || D == Op0) && NoOp1WrapProblem) 2459 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()), 2460 C == Op0 ? D : C); 2461 2462 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow. 2463 if (A && C && (A == C || A == D || B == C || B == D) && 2464 NoOp0WrapProblem && NoOp1WrapProblem && 2465 // Try not to increase register pressure. 2466 BO0->hasOneUse() && BO1->hasOneUse()) { 2467 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2468 Value *Y, *Z; 2469 if (A == C) { 2470 // C + B == C + D -> B == D 2471 Y = B; 2472 Z = D; 2473 } else if (A == D) { 2474 // D + B == C + D -> B == C 2475 Y = B; 2476 Z = C; 2477 } else if (B == C) { 2478 // A + C == C + D -> A == D 2479 Y = A; 2480 Z = D; 2481 } else { 2482 assert(B == D); 2483 // A + D == C + D -> A == C 2484 Y = A; 2485 Z = C; 2486 } 2487 return new ICmpInst(Pred, Y, Z); 2488 } 2489 2490 // icmp slt (X + -1), Y -> icmp sle X, Y 2491 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT && 2492 match(B, m_AllOnes())) 2493 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1); 2494 2495 // icmp sge (X + -1), Y -> icmp sgt X, Y 2496 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE && 2497 match(B, m_AllOnes())) 2498 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1); 2499 2500 // icmp sle (X + 1), Y -> icmp slt X, Y 2501 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && 2502 match(B, m_One())) 2503 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1); 2504 2505 // icmp sgt (X + 1), Y -> icmp sge X, Y 2506 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && 2507 match(B, m_One())) 2508 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1); 2509 2510 // if C1 has greater magnitude than C2: 2511 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y 2512 // s.t. C3 = C1 - C2 2513 // 2514 // if C2 has greater magnitude than C1: 2515 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3) 2516 // s.t. C3 = C2 - C1 2517 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem && 2518 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned()) 2519 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B)) 2520 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) { 2521 const APInt &AP1 = C1->getValue(); 2522 const APInt &AP2 = C2->getValue(); 2523 if (AP1.isNegative() == AP2.isNegative()) { 2524 APInt AP1Abs = C1->getValue().abs(); 2525 APInt AP2Abs = C2->getValue().abs(); 2526 if (AP1Abs.uge(AP2Abs)) { 2527 ConstantInt *C3 = Builder->getInt(AP1 - AP2); 2528 Value *NewAdd = Builder->CreateNSWAdd(A, C3); 2529 return new ICmpInst(Pred, NewAdd, C); 2530 } else { 2531 ConstantInt *C3 = Builder->getInt(AP2 - AP1); 2532 Value *NewAdd = Builder->CreateNSWAdd(C, C3); 2533 return new ICmpInst(Pred, A, NewAdd); 2534 } 2535 } 2536 } 2537 2538 2539 // Analyze the case when either Op0 or Op1 is a sub instruction. 2540 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null). 2541 A = 0; B = 0; C = 0; D = 0; 2542 if (BO0 && BO0->getOpcode() == Instruction::Sub) 2543 A = BO0->getOperand(0), B = BO0->getOperand(1); 2544 if (BO1 && BO1->getOpcode() == Instruction::Sub) 2545 C = BO1->getOperand(0), D = BO1->getOperand(1); 2546 2547 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow. 2548 if (A == Op1 && NoOp0WrapProblem) 2549 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B); 2550 2551 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow. 2552 if (C == Op0 && NoOp1WrapProblem) 2553 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType())); 2554 2555 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow. 2556 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem && 2557 // Try not to increase register pressure. 2558 BO0->hasOneUse() && BO1->hasOneUse()) 2559 return new ICmpInst(Pred, A, C); 2560 2561 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow. 2562 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem && 2563 // Try not to increase register pressure. 2564 BO0->hasOneUse() && BO1->hasOneUse()) 2565 return new ICmpInst(Pred, D, B); 2566 2567 BinaryOperator *SRem = NULL; 2568 // icmp (srem X, Y), Y 2569 if (BO0 && BO0->getOpcode() == Instruction::SRem && 2570 Op1 == BO0->getOperand(1)) 2571 SRem = BO0; 2572 // icmp Y, (srem X, Y) 2573 else if (BO1 && BO1->getOpcode() == Instruction::SRem && 2574 Op0 == BO1->getOperand(1)) 2575 SRem = BO1; 2576 if (SRem) { 2577 // We don't check hasOneUse to avoid increasing register pressure because 2578 // the value we use is the same value this instruction was already using. 2579 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) { 2580 default: break; 2581 case ICmpInst::ICMP_EQ: 2582 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2583 case ICmpInst::ICMP_NE: 2584 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2585 case ICmpInst::ICMP_SGT: 2586 case ICmpInst::ICMP_SGE: 2587 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1), 2588 Constant::getAllOnesValue(SRem->getType())); 2589 case ICmpInst::ICMP_SLT: 2590 case ICmpInst::ICMP_SLE: 2591 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1), 2592 Constant::getNullValue(SRem->getType())); 2593 } 2594 } 2595 2596 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && 2597 BO0->hasOneUse() && BO1->hasOneUse() && 2598 BO0->getOperand(1) == BO1->getOperand(1)) { 2599 switch (BO0->getOpcode()) { 2600 default: break; 2601 case Instruction::Add: 2602 case Instruction::Sub: 2603 case Instruction::Xor: 2604 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b 2605 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2606 BO1->getOperand(0)); 2607 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b 2608 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2609 if (CI->getValue().isSignBit()) { 2610 ICmpInst::Predicate Pred = I.isSigned() 2611 ? I.getUnsignedPredicate() 2612 : I.getSignedPredicate(); 2613 return new ICmpInst(Pred, BO0->getOperand(0), 2614 BO1->getOperand(0)); 2615 } 2616 2617 if (CI->isMaxValue(true)) { 2618 ICmpInst::Predicate Pred = I.isSigned() 2619 ? I.getUnsignedPredicate() 2620 : I.getSignedPredicate(); 2621 Pred = I.getSwappedPredicate(Pred); 2622 return new ICmpInst(Pred, BO0->getOperand(0), 2623 BO1->getOperand(0)); 2624 } 2625 } 2626 break; 2627 case Instruction::Mul: 2628 if (!I.isEquality()) 2629 break; 2630 2631 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2632 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask 2633 // Mask = -1 >> count-trailing-zeros(Cst). 2634 if (!CI->isZero() && !CI->isOne()) { 2635 const APInt &AP = CI->getValue(); 2636 ConstantInt *Mask = ConstantInt::get(I.getContext(), 2637 APInt::getLowBitsSet(AP.getBitWidth(), 2638 AP.getBitWidth() - 2639 AP.countTrailingZeros())); 2640 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask); 2641 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask); 2642 return new ICmpInst(I.getPredicate(), And1, And2); 2643 } 2644 } 2645 break; 2646 case Instruction::UDiv: 2647 case Instruction::LShr: 2648 if (I.isSigned()) 2649 break; 2650 // fall-through 2651 case Instruction::SDiv: 2652 case Instruction::AShr: 2653 if (!BO0->isExact() || !BO1->isExact()) 2654 break; 2655 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2656 BO1->getOperand(0)); 2657 case Instruction::Shl: { 2658 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap(); 2659 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap(); 2660 if (!NUW && !NSW) 2661 break; 2662 if (!NSW && I.isSigned()) 2663 break; 2664 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2665 BO1->getOperand(0)); 2666 } 2667 } 2668 } 2669 } 2670 2671 { Value *A, *B; 2672 // Transform (A & ~B) == 0 --> (A & B) != 0 2673 // and (A & ~B) != 0 --> (A & B) == 0 2674 // if A is a power of 2. 2675 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) && 2676 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality()) 2677 return new ICmpInst(I.getInversePredicate(), 2678 Builder->CreateAnd(A, B), 2679 Op1); 2680 2681 // ~x < ~y --> y < x 2682 // ~x < cst --> ~cst < x 2683 if (match(Op0, m_Not(m_Value(A)))) { 2684 if (match(Op1, m_Not(m_Value(B)))) 2685 return new ICmpInst(I.getPredicate(), B, A); 2686 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1)) 2687 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A); 2688 } 2689 2690 // (a+b) <u a --> llvm.uadd.with.overflow. 2691 // (a+b) <u b --> llvm.uadd.with.overflow. 2692 if (I.getPredicate() == ICmpInst::ICMP_ULT && 2693 match(Op0, m_Add(m_Value(A), m_Value(B))) && 2694 (Op1 == A || Op1 == B)) 2695 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this)) 2696 return R; 2697 2698 // a >u (a+b) --> llvm.uadd.with.overflow. 2699 // b >u (a+b) --> llvm.uadd.with.overflow. 2700 if (I.getPredicate() == ICmpInst::ICMP_UGT && 2701 match(Op1, m_Add(m_Value(A), m_Value(B))) && 2702 (Op0 == A || Op0 == B)) 2703 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this)) 2704 return R; 2705 } 2706 2707 if (I.isEquality()) { 2708 Value *A, *B, *C, *D; 2709 2710 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 2711 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0 2712 Value *OtherVal = A == Op1 ? B : A; 2713 return new ICmpInst(I.getPredicate(), OtherVal, 2714 Constant::getNullValue(A->getType())); 2715 } 2716 2717 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) { 2718 // A^c1 == C^c2 --> A == C^(c1^c2) 2719 ConstantInt *C1, *C2; 2720 if (match(B, m_ConstantInt(C1)) && 2721 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) { 2722 Constant *NC = ConstantInt::get(I.getContext(), 2723 C1->getValue() ^ C2->getValue()); 2724 Value *Xor = Builder->CreateXor(C, NC); 2725 return new ICmpInst(I.getPredicate(), A, Xor); 2726 } 2727 2728 // A^B == A^D -> B == D 2729 if (A == C) return new ICmpInst(I.getPredicate(), B, D); 2730 if (A == D) return new ICmpInst(I.getPredicate(), B, C); 2731 if (B == C) return new ICmpInst(I.getPredicate(), A, D); 2732 if (B == D) return new ICmpInst(I.getPredicate(), A, C); 2733 } 2734 } 2735 2736 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && 2737 (A == Op0 || B == Op0)) { 2738 // A == (A^B) -> B == 0 2739 Value *OtherVal = A == Op0 ? B : A; 2740 return new ICmpInst(I.getPredicate(), OtherVal, 2741 Constant::getNullValue(A->getType())); 2742 } 2743 2744 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0 2745 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) && 2746 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) { 2747 Value *X = 0, *Y = 0, *Z = 0; 2748 2749 if (A == C) { 2750 X = B; Y = D; Z = A; 2751 } else if (A == D) { 2752 X = B; Y = C; Z = A; 2753 } else if (B == C) { 2754 X = A; Y = D; Z = B; 2755 } else if (B == D) { 2756 X = A; Y = C; Z = B; 2757 } 2758 2759 if (X) { // Build (X^Y) & Z 2760 Op1 = Builder->CreateXor(X, Y); 2761 Op1 = Builder->CreateAnd(Op1, Z); 2762 I.setOperand(0, Op1); 2763 I.setOperand(1, Constant::getNullValue(Op1->getType())); 2764 return &I; 2765 } 2766 } 2767 2768 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B) 2769 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B) 2770 ConstantInt *Cst1; 2771 if ((Op0->hasOneUse() && 2772 match(Op0, m_ZExt(m_Value(A))) && 2773 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) || 2774 (Op1->hasOneUse() && 2775 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) && 2776 match(Op1, m_ZExt(m_Value(A))))) { 2777 APInt Pow2 = Cst1->getValue() + 1; 2778 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) && 2779 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth()) 2780 return new ICmpInst(I.getPredicate(), A, 2781 Builder->CreateTrunc(B, A->getType())); 2782 } 2783 2784 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to 2785 // "icmp (and X, mask), cst" 2786 uint64_t ShAmt = 0; 2787 if (Op0->hasOneUse() && 2788 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), 2789 m_ConstantInt(ShAmt))))) && 2790 match(Op1, m_ConstantInt(Cst1)) && 2791 // Only do this when A has multiple uses. This is most important to do 2792 // when it exposes other optimizations. 2793 !A->hasOneUse()) { 2794 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits(); 2795 2796 if (ShAmt < ASize) { 2797 APInt MaskV = 2798 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits()); 2799 MaskV <<= ShAmt; 2800 2801 APInt CmpV = Cst1->getValue().zext(ASize); 2802 CmpV <<= ShAmt; 2803 2804 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV)); 2805 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV)); 2806 } 2807 } 2808 } 2809 2810 { 2811 Value *X; ConstantInt *Cst; 2812 // icmp X+Cst, X 2813 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X) 2814 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0); 2815 2816 // icmp X, X+Cst 2817 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X) 2818 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1); 2819 } 2820 return Changed ? &I : 0; 2821} 2822 2823 2824 2825 2826 2827 2828/// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible. 2829/// 2830Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I, 2831 Instruction *LHSI, 2832 Constant *RHSC) { 2833 if (!isa<ConstantFP>(RHSC)) return 0; 2834 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF(); 2835 2836 // Get the width of the mantissa. We don't want to hack on conversions that 2837 // might lose information from the integer, e.g. "i64 -> float" 2838 int MantissaWidth = LHSI->getType()->getFPMantissaWidth(); 2839 if (MantissaWidth == -1) return 0; // Unknown. 2840 2841 // Check to see that the input is converted from an integer type that is small 2842 // enough that preserves all bits. TODO: check here for "known" sign bits. 2843 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e. 2844 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits(); 2845 2846 // If this is a uitofp instruction, we need an extra bit to hold the sign. 2847 bool LHSUnsigned = isa<UIToFPInst>(LHSI); 2848 if (LHSUnsigned) 2849 ++InputSize; 2850 2851 // If the conversion would lose info, don't hack on this. 2852 if ((int)InputSize > MantissaWidth) 2853 return 0; 2854 2855 // Otherwise, we can potentially simplify the comparison. We know that it 2856 // will always come through as an integer value and we know the constant is 2857 // not a NAN (it would have been previously simplified). 2858 assert(!RHS.isNaN() && "NaN comparison not already folded!"); 2859 2860 ICmpInst::Predicate Pred; 2861 switch (I.getPredicate()) { 2862 default: llvm_unreachable("Unexpected predicate!"); 2863 case FCmpInst::FCMP_UEQ: 2864 case FCmpInst::FCMP_OEQ: 2865 Pred = ICmpInst::ICMP_EQ; 2866 break; 2867 case FCmpInst::FCMP_UGT: 2868 case FCmpInst::FCMP_OGT: 2869 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT; 2870 break; 2871 case FCmpInst::FCMP_UGE: 2872 case FCmpInst::FCMP_OGE: 2873 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE; 2874 break; 2875 case FCmpInst::FCMP_ULT: 2876 case FCmpInst::FCMP_OLT: 2877 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT; 2878 break; 2879 case FCmpInst::FCMP_ULE: 2880 case FCmpInst::FCMP_OLE: 2881 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE; 2882 break; 2883 case FCmpInst::FCMP_UNE: 2884 case FCmpInst::FCMP_ONE: 2885 Pred = ICmpInst::ICMP_NE; 2886 break; 2887 case FCmpInst::FCMP_ORD: 2888 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2889 case FCmpInst::FCMP_UNO: 2890 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2891 } 2892 2893 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType()); 2894 2895 // Now we know that the APFloat is a normal number, zero or inf. 2896 2897 // See if the FP constant is too large for the integer. For example, 2898 // comparing an i8 to 300.0. 2899 unsigned IntWidth = IntTy->getScalarSizeInBits(); 2900 2901 if (!LHSUnsigned) { 2902 // If the RHS value is > SignedMax, fold the comparison. This handles +INF 2903 // and large values. 2904 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false); 2905 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true, 2906 APFloat::rmNearestTiesToEven); 2907 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0 2908 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT || 2909 Pred == ICmpInst::ICMP_SLE) 2910 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2911 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2912 } 2913 } else { 2914 // If the RHS value is > UnsignedMax, fold the comparison. This handles 2915 // +INF and large values. 2916 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false); 2917 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false, 2918 APFloat::rmNearestTiesToEven); 2919 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0 2920 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT || 2921 Pred == ICmpInst::ICMP_ULE) 2922 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2923 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2924 } 2925 } 2926 2927 if (!LHSUnsigned) { 2928 // See if the RHS value is < SignedMin. 2929 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false); 2930 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true, 2931 APFloat::rmNearestTiesToEven); 2932 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0 2933 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT || 2934 Pred == ICmpInst::ICMP_SGE) 2935 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2936 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2937 } 2938 } else { 2939 // See if the RHS value is < UnsignedMin. 2940 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false); 2941 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true, 2942 APFloat::rmNearestTiesToEven); 2943 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0 2944 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT || 2945 Pred == ICmpInst::ICMP_UGE) 2946 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2947 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2948 } 2949 } 2950 2951 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or 2952 // [0, UMAX], but it may still be fractional. See if it is fractional by 2953 // casting the FP value to the integer value and back, checking for equality. 2954 // Don't do this for zero, because -0.0 is not fractional. 2955 Constant *RHSInt = LHSUnsigned 2956 ? ConstantExpr::getFPToUI(RHSC, IntTy) 2957 : ConstantExpr::getFPToSI(RHSC, IntTy); 2958 if (!RHS.isZero()) { 2959 bool Equal = LHSUnsigned 2960 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC 2961 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC; 2962 if (!Equal) { 2963 // If we had a comparison against a fractional value, we have to adjust 2964 // the compare predicate and sometimes the value. RHSC is rounded towards 2965 // zero at this point. 2966 switch (Pred) { 2967 default: llvm_unreachable("Unexpected integer comparison!"); 2968 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true 2969 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2970 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false 2971 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2972 case ICmpInst::ICMP_ULE: 2973 // (float)int <= 4.4 --> int <= 4 2974 // (float)int <= -4.4 --> false 2975 if (RHS.isNegative()) 2976 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2977 break; 2978 case ICmpInst::ICMP_SLE: 2979 // (float)int <= 4.4 --> int <= 4 2980 // (float)int <= -4.4 --> int < -4 2981 if (RHS.isNegative()) 2982 Pred = ICmpInst::ICMP_SLT; 2983 break; 2984 case ICmpInst::ICMP_ULT: 2985 // (float)int < -4.4 --> false 2986 // (float)int < 4.4 --> int <= 4 2987 if (RHS.isNegative()) 2988 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2989 Pred = ICmpInst::ICMP_ULE; 2990 break; 2991 case ICmpInst::ICMP_SLT: 2992 // (float)int < -4.4 --> int < -4 2993 // (float)int < 4.4 --> int <= 4 2994 if (!RHS.isNegative()) 2995 Pred = ICmpInst::ICMP_SLE; 2996 break; 2997 case ICmpInst::ICMP_UGT: 2998 // (float)int > 4.4 --> int > 4 2999 // (float)int > -4.4 --> true 3000 if (RHS.isNegative()) 3001 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 3002 break; 3003 case ICmpInst::ICMP_SGT: 3004 // (float)int > 4.4 --> int > 4 3005 // (float)int > -4.4 --> int >= -4 3006 if (RHS.isNegative()) 3007 Pred = ICmpInst::ICMP_SGE; 3008 break; 3009 case ICmpInst::ICMP_UGE: 3010 // (float)int >= -4.4 --> true 3011 // (float)int >= 4.4 --> int > 4 3012 if (RHS.isNegative()) 3013 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 3014 Pred = ICmpInst::ICMP_UGT; 3015 break; 3016 case ICmpInst::ICMP_SGE: 3017 // (float)int >= -4.4 --> int >= -4 3018 // (float)int >= 4.4 --> int > 4 3019 if (!RHS.isNegative()) 3020 Pred = ICmpInst::ICMP_SGT; 3021 break; 3022 } 3023 } 3024 } 3025 3026 // Lower this FP comparison into an appropriate integer version of the 3027 // comparison. 3028 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt); 3029} 3030 3031Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { 3032 bool Changed = false; 3033 3034 /// Orders the operands of the compare so that they are listed from most 3035 /// complex to least complex. This puts constants before unary operators, 3036 /// before binary operators. 3037 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) { 3038 I.swapOperands(); 3039 Changed = true; 3040 } 3041 3042 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 3043 3044 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD)) 3045 return ReplaceInstUsesWith(I, V); 3046 3047 // Simplify 'fcmp pred X, X' 3048 if (Op0 == Op1) { 3049 switch (I.getPredicate()) { 3050 default: llvm_unreachable("Unknown predicate!"); 3051 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y) 3052 case FCmpInst::FCMP_ULT: // True if unordered or less than 3053 case FCmpInst::FCMP_UGT: // True if unordered or greater than 3054 case FCmpInst::FCMP_UNE: // True if unordered or not equal 3055 // Canonicalize these to be 'fcmp uno %X, 0.0'. 3056 I.setPredicate(FCmpInst::FCMP_UNO); 3057 I.setOperand(1, Constant::getNullValue(Op0->getType())); 3058 return &I; 3059 3060 case FCmpInst::FCMP_ORD: // True if ordered (no nans) 3061 case FCmpInst::FCMP_OEQ: // True if ordered and equal 3062 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal 3063 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal 3064 // Canonicalize these to be 'fcmp ord %X, 0.0'. 3065 I.setPredicate(FCmpInst::FCMP_ORD); 3066 I.setOperand(1, Constant::getNullValue(Op0->getType())); 3067 return &I; 3068 } 3069 } 3070 3071 // Handle fcmp with constant RHS 3072 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 3073 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 3074 switch (LHSI->getOpcode()) { 3075 case Instruction::FPExt: { 3076 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless 3077 FPExtInst *LHSExt = cast<FPExtInst>(LHSI); 3078 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC); 3079 if (!RHSF) 3080 break; 3081 3082 const fltSemantics *Sem; 3083 // FIXME: This shouldn't be here. 3084 if (LHSExt->getSrcTy()->isHalfTy()) 3085 Sem = &APFloat::IEEEhalf; 3086 else if (LHSExt->getSrcTy()->isFloatTy()) 3087 Sem = &APFloat::IEEEsingle; 3088 else if (LHSExt->getSrcTy()->isDoubleTy()) 3089 Sem = &APFloat::IEEEdouble; 3090 else if (LHSExt->getSrcTy()->isFP128Ty()) 3091 Sem = &APFloat::IEEEquad; 3092 else if (LHSExt->getSrcTy()->isX86_FP80Ty()) 3093 Sem = &APFloat::x87DoubleExtended; 3094 else if (LHSExt->getSrcTy()->isPPC_FP128Ty()) 3095 Sem = &APFloat::PPCDoubleDouble; 3096 else 3097 break; 3098 3099 bool Lossy; 3100 APFloat F = RHSF->getValueAPF(); 3101 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy); 3102 3103 // Avoid lossy conversions and denormals. Zero is a special case 3104 // that's OK to convert. 3105 APFloat Fabs = F; 3106 Fabs.clearSign(); 3107 if (!Lossy && 3108 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) != 3109 APFloat::cmpLessThan) || Fabs.isZero())) 3110 3111 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 3112 ConstantFP::get(RHSC->getContext(), F)); 3113 break; 3114 } 3115 case Instruction::PHI: 3116 // Only fold fcmp into the PHI if the phi and fcmp are in the same 3117 // block. If in the same block, we're encouraging jump threading. If 3118 // not, we are just pessimizing the code by making an i1 phi. 3119 if (LHSI->getParent() == I.getParent()) 3120 if (Instruction *NV = FoldOpIntoPhi(I)) 3121 return NV; 3122 break; 3123 case Instruction::SIToFP: 3124 case Instruction::UIToFP: 3125 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC)) 3126 return NV; 3127 break; 3128 case Instruction::Select: { 3129 // If either operand of the select is a constant, we can fold the 3130 // comparison into the select arms, which will cause one to be 3131 // constant folded and the select turned into a bitwise or. 3132 Value *Op1 = 0, *Op2 = 0; 3133 if (LHSI->hasOneUse()) { 3134 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) { 3135 // Fold the known value into the constant operand. 3136 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 3137 // Insert a new FCmp of the other select operand. 3138 Op2 = Builder->CreateFCmp(I.getPredicate(), 3139 LHSI->getOperand(2), RHSC, I.getName()); 3140 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) { 3141 // Fold the known value into the constant operand. 3142 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 3143 // Insert a new FCmp of the other select operand. 3144 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1), 3145 RHSC, I.getName()); 3146 } 3147 } 3148 3149 if (Op1) 3150 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 3151 break; 3152 } 3153 case Instruction::FSub: { 3154 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C 3155 Value *Op; 3156 if (match(LHSI, m_FNeg(m_Value(Op)))) 3157 return new FCmpInst(I.getSwappedPredicate(), Op, 3158 ConstantExpr::getFNeg(RHSC)); 3159 break; 3160 } 3161 case Instruction::Load: 3162 if (GetElementPtrInst *GEP = 3163 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 3164 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 3165 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 3166 !cast<LoadInst>(LHSI)->isVolatile()) 3167 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 3168 return Res; 3169 } 3170 break; 3171 case Instruction::Call: { 3172 CallInst *CI = cast<CallInst>(LHSI); 3173 LibFunc::Func Func; 3174 // Various optimization for fabs compared with zero. 3175 if (RHSC->isNullValue() && CI->getCalledFunction() && 3176 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) && 3177 TLI->has(Func)) { 3178 if (Func == LibFunc::fabs || Func == LibFunc::fabsf || 3179 Func == LibFunc::fabsl) { 3180 switch (I.getPredicate()) { 3181 default: break; 3182 // fabs(x) < 0 --> false 3183 case FCmpInst::FCMP_OLT: 3184 return ReplaceInstUsesWith(I, Builder->getFalse()); 3185 // fabs(x) > 0 --> x != 0 3186 case FCmpInst::FCMP_OGT: 3187 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), 3188 RHSC); 3189 // fabs(x) <= 0 --> x == 0 3190 case FCmpInst::FCMP_OLE: 3191 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), 3192 RHSC); 3193 // fabs(x) >= 0 --> !isnan(x) 3194 case FCmpInst::FCMP_OGE: 3195 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), 3196 RHSC); 3197 // fabs(x) == 0 --> x == 0 3198 // fabs(x) != 0 --> x != 0 3199 case FCmpInst::FCMP_OEQ: 3200 case FCmpInst::FCMP_UEQ: 3201 case FCmpInst::FCMP_ONE: 3202 case FCmpInst::FCMP_UNE: 3203 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), 3204 RHSC); 3205 } 3206 } 3207 } 3208 } 3209 } 3210 } 3211 3212 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y 3213 Value *X, *Y; 3214 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) 3215 return new FCmpInst(I.getSwappedPredicate(), X, Y); 3216 3217 // fcmp (fpext x), (fpext y) -> fcmp x, y 3218 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0)) 3219 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1)) 3220 if (LHSExt->getSrcTy() == RHSExt->getSrcTy()) 3221 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 3222 RHSExt->getOperand(0)); 3223 3224 return Changed ? &I : 0; 3225} 3226