InstCombineCompares.cpp revision e5116f840eedbc3cffa31adc683b4e37d53f2c7a
1//===- InstCombineCompares.cpp --------------------------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the visitICmp and visitFCmp functions. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/IntrinsicInst.h" 16#include "llvm/Analysis/InstructionSimplify.h" 17#include "llvm/Analysis/MemoryBuiltins.h" 18#include "llvm/Target/TargetData.h" 19#include "llvm/Support/ConstantRange.h" 20#include "llvm/Support/GetElementPtrTypeIterator.h" 21#include "llvm/Support/PatternMatch.h" 22using namespace llvm; 23using namespace PatternMatch; 24 25static ConstantInt *getOne(Constant *C) { 26 return ConstantInt::get(cast<IntegerType>(C->getType()), 1); 27} 28 29/// AddOne - Add one to a ConstantInt 30static Constant *AddOne(Constant *C) { 31 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 32} 33/// SubOne - Subtract one from a ConstantInt 34static Constant *SubOne(Constant *C) { 35 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1)); 36} 37 38static ConstantInt *ExtractElement(Constant *V, Constant *Idx) { 39 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx)); 40} 41 42static bool HasAddOverflow(ConstantInt *Result, 43 ConstantInt *In1, ConstantInt *In2, 44 bool IsSigned) { 45 if (IsSigned) 46 if (In2->getValue().isNegative()) 47 return Result->getValue().sgt(In1->getValue()); 48 else 49 return Result->getValue().slt(In1->getValue()); 50 else 51 return Result->getValue().ult(In1->getValue()); 52} 53 54/// AddWithOverflow - Compute Result = In1+In2, returning true if the result 55/// overflowed for this type. 56static bool AddWithOverflow(Constant *&Result, Constant *In1, 57 Constant *In2, bool IsSigned = false) { 58 Result = ConstantExpr::getAdd(In1, In2); 59 60 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 61 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 62 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 63 if (HasAddOverflow(ExtractElement(Result, Idx), 64 ExtractElement(In1, Idx), 65 ExtractElement(In2, Idx), 66 IsSigned)) 67 return true; 68 } 69 return false; 70 } 71 72 return HasAddOverflow(cast<ConstantInt>(Result), 73 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 74 IsSigned); 75} 76 77static bool HasSubOverflow(ConstantInt *Result, 78 ConstantInt *In1, ConstantInt *In2, 79 bool IsSigned) { 80 if (IsSigned) 81 if (In2->getValue().isNegative()) 82 return Result->getValue().slt(In1->getValue()); 83 else 84 return Result->getValue().sgt(In1->getValue()); 85 else 86 return Result->getValue().ugt(In1->getValue()); 87} 88 89/// SubWithOverflow - Compute Result = In1-In2, returning true if the result 90/// overflowed for this type. 91static bool SubWithOverflow(Constant *&Result, Constant *In1, 92 Constant *In2, bool IsSigned = false) { 93 Result = ConstantExpr::getSub(In1, In2); 94 95 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 96 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 97 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 98 if (HasSubOverflow(ExtractElement(Result, Idx), 99 ExtractElement(In1, Idx), 100 ExtractElement(In2, Idx), 101 IsSigned)) 102 return true; 103 } 104 return false; 105 } 106 107 return HasSubOverflow(cast<ConstantInt>(Result), 108 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 109 IsSigned); 110} 111 112/// isSignBitCheck - Given an exploded icmp instruction, return true if the 113/// comparison only checks the sign bit. If it only checks the sign bit, set 114/// TrueIfSigned if the result of the comparison is true when the input value is 115/// signed. 116static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS, 117 bool &TrueIfSigned) { 118 switch (pred) { 119 case ICmpInst::ICMP_SLT: // True if LHS s< 0 120 TrueIfSigned = true; 121 return RHS->isZero(); 122 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1 123 TrueIfSigned = true; 124 return RHS->isAllOnesValue(); 125 case ICmpInst::ICMP_SGT: // True if LHS s> -1 126 TrueIfSigned = false; 127 return RHS->isAllOnesValue(); 128 case ICmpInst::ICMP_UGT: 129 // True if LHS u> RHS and RHS == high-bit-mask - 1 130 TrueIfSigned = true; 131 return RHS->getValue() == 132 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits()); 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// isHighOnes - Return true if the constant is of the form 1+0+. 143// This is the same as lowones(~X). 144static bool isHighOnes(const ConstantInt *CI) { 145 return (~CI->getValue() + 1).isPowerOf2(); 146} 147 148/// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a 149/// set of known zero and one bits, compute the maximum and minimum values that 150/// could have the specified known zero and known one bits, returning them in 151/// min/max. 152static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero, 153 const APInt& KnownOne, 154 APInt& Min, APInt& Max) { 155 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 156 KnownZero.getBitWidth() == Min.getBitWidth() && 157 KnownZero.getBitWidth() == Max.getBitWidth() && 158 "KnownZero, KnownOne and Min, Max must have equal bitwidth."); 159 APInt UnknownBits = ~(KnownZero|KnownOne); 160 161 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign 162 // bit if it is unknown. 163 Min = KnownOne; 164 Max = KnownOne|UnknownBits; 165 166 if (UnknownBits.isNegative()) { // Sign bit is unknown 167 Min.setBit(Min.getBitWidth()-1); 168 Max.clearBit(Max.getBitWidth()-1); 169 } 170} 171 172// ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and 173// a set of known zero and one bits, compute the maximum and minimum values that 174// could have the specified known zero and known one bits, returning them in 175// min/max. 176static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero, 177 const APInt &KnownOne, 178 APInt &Min, APInt &Max) { 179 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 180 KnownZero.getBitWidth() == Min.getBitWidth() && 181 KnownZero.getBitWidth() == Max.getBitWidth() && 182 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."); 183 APInt UnknownBits = ~(KnownZero|KnownOne); 184 185 // The minimum value is when the unknown bits are all zeros. 186 Min = KnownOne; 187 // The maximum value is when the unknown bits are all ones. 188 Max = KnownOne|UnknownBits; 189} 190 191 192 193/// FoldCmpLoadFromIndexedGlobal - Called we see this pattern: 194/// cmp pred (load (gep GV, ...)), cmpcst 195/// where GV is a global variable with a constant initializer. Try to simplify 196/// this into some simple computation that does not need the load. For example 197/// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3". 198/// 199/// If AndCst is non-null, then the loaded value is masked with that constant 200/// before doing the comparison. This handles cases like "A[i]&4 == 0". 201Instruction *InstCombiner:: 202FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV, 203 CmpInst &ICI, ConstantInt *AndCst) { 204 // We need TD information to know the pointer size unless this is inbounds. 205 if (!GEP->isInBounds() && TD == 0) return 0; 206 207 ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer()); 208 if (Init == 0 || Init->getNumOperands() > 1024) return 0; 209 210 // There are many forms of this optimization we can handle, for now, just do 211 // the simple index into a single-dimensional array. 212 // 213 // Require: GEP GV, 0, i {{, constant indices}} 214 if (GEP->getNumOperands() < 3 || 215 !isa<ConstantInt>(GEP->getOperand(1)) || 216 !cast<ConstantInt>(GEP->getOperand(1))->isZero() || 217 isa<Constant>(GEP->getOperand(2))) 218 return 0; 219 220 // Check that indices after the variable are constants and in-range for the 221 // type they index. Collect the indices. This is typically for arrays of 222 // structs. 223 SmallVector<unsigned, 4> LaterIndices; 224 225 const Type *EltTy = cast<ArrayType>(Init->getType())->getElementType(); 226 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { 227 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i)); 228 if (Idx == 0) return 0; // Variable index. 229 230 uint64_t IdxVal = Idx->getZExtValue(); 231 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index. 232 233 if (const StructType *STy = dyn_cast<StructType>(EltTy)) 234 EltTy = STy->getElementType(IdxVal); 235 else if (const ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) { 236 if (IdxVal >= ATy->getNumElements()) return 0; 237 EltTy = ATy->getElementType(); 238 } else { 239 return 0; // Unknown type. 240 } 241 242 LaterIndices.push_back(IdxVal); 243 } 244 245 enum { Overdefined = -3, Undefined = -2 }; 246 247 // Variables for our state machines. 248 249 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form 250 // "i == 47 | i == 87", where 47 is the first index the condition is true for, 251 // and 87 is the second (and last) index. FirstTrueElement is -2 when 252 // undefined, otherwise set to the first true element. SecondTrueElement is 253 // -2 when undefined, -3 when overdefined and >= 0 when that index is true. 254 int FirstTrueElement = Undefined, SecondTrueElement = Undefined; 255 256 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the 257 // form "i != 47 & i != 87". Same state transitions as for true elements. 258 int FirstFalseElement = Undefined, SecondFalseElement = Undefined; 259 260 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these 261 /// define a state machine that triggers for ranges of values that the index 262 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. 263 /// This is -2 when undefined, -3 when overdefined, and otherwise the last 264 /// index in the range (inclusive). We use -2 for undefined here because we 265 /// use relative comparisons and don't want 0-1 to match -1. 266 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; 267 268 // MagicBitvector - This is a magic bitvector where we set a bit if the 269 // comparison is true for element 'i'. If there are 64 elements or less in 270 // the array, this will fully represent all the comparison results. 271 uint64_t MagicBitvector = 0; 272 273 274 // Scan the array and see if one of our patterns matches. 275 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1)); 276 for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) { 277 Constant *Elt = Init->getOperand(i); 278 279 // If this is indexing an array of structures, get the structure element. 280 if (!LaterIndices.empty()) 281 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices.data(), 282 LaterIndices.size()); 283 284 // If the element is masked, handle it. 285 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst); 286 287 // Find out if the comparison would be true or false for the i'th element. 288 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt, 289 CompareRHS, TD); 290 // If the result is undef for this element, ignore it. 291 if (isa<UndefValue>(C)) { 292 // Extend range state machines to cover this element in case there is an 293 // undef in the middle of the range. 294 if (TrueRangeEnd == (int)i-1) 295 TrueRangeEnd = i; 296 if (FalseRangeEnd == (int)i-1) 297 FalseRangeEnd = i; 298 continue; 299 } 300 301 // If we can't compute the result for any of the elements, we have to give 302 // up evaluating the entire conditional. 303 if (!isa<ConstantInt>(C)) return 0; 304 305 // Otherwise, we know if the comparison is true or false for this element, 306 // update our state machines. 307 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero(); 308 309 // State machine for single/double/range index comparison. 310 if (IsTrueForElt) { 311 // Update the TrueElement state machine. 312 if (FirstTrueElement == Undefined) 313 FirstTrueElement = TrueRangeEnd = i; // First true element. 314 else { 315 // Update double-compare state machine. 316 if (SecondTrueElement == Undefined) 317 SecondTrueElement = i; 318 else 319 SecondTrueElement = Overdefined; 320 321 // Update range state machine. 322 if (TrueRangeEnd == (int)i-1) 323 TrueRangeEnd = i; 324 else 325 TrueRangeEnd = Overdefined; 326 } 327 } else { 328 // Update the FalseElement state machine. 329 if (FirstFalseElement == Undefined) 330 FirstFalseElement = FalseRangeEnd = i; // First false element. 331 else { 332 // Update double-compare state machine. 333 if (SecondFalseElement == Undefined) 334 SecondFalseElement = i; 335 else 336 SecondFalseElement = Overdefined; 337 338 // Update range state machine. 339 if (FalseRangeEnd == (int)i-1) 340 FalseRangeEnd = i; 341 else 342 FalseRangeEnd = Overdefined; 343 } 344 } 345 346 347 // If this element is in range, update our magic bitvector. 348 if (i < 64 && IsTrueForElt) 349 MagicBitvector |= 1ULL << i; 350 351 // If all of our states become overdefined, bail out early. Since the 352 // predicate is expensive, only check it every 8 elements. This is only 353 // really useful for really huge arrays. 354 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && 355 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && 356 FalseRangeEnd == Overdefined) 357 return 0; 358 } 359 360 // Now that we've scanned the entire array, emit our new comparison(s). We 361 // order the state machines in complexity of the generated code. 362 Value *Idx = GEP->getOperand(2); 363 364 // If the index is larger than the pointer size of the target, truncate the 365 // index down like the GEP would do implicitly. We don't have to do this for 366 // an inbounds GEP because the index can't be out of range. 367 if (!GEP->isInBounds() && 368 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits()) 369 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext())); 370 371 // If the comparison is only true for one or two elements, emit direct 372 // comparisons. 373 if (SecondTrueElement != Overdefined) { 374 // None true -> false. 375 if (FirstTrueElement == Undefined) 376 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext())); 377 378 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement); 379 380 // True for one element -> 'i == 47'. 381 if (SecondTrueElement == Undefined) 382 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); 383 384 // True for two elements -> 'i == 47 | i == 72'. 385 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx); 386 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement); 387 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx); 388 return BinaryOperator::CreateOr(C1, C2); 389 } 390 391 // If the comparison is only false for one or two elements, emit direct 392 // comparisons. 393 if (SecondFalseElement != Overdefined) { 394 // None false -> true. 395 if (FirstFalseElement == Undefined) 396 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext())); 397 398 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement); 399 400 // False for one element -> 'i != 47'. 401 if (SecondFalseElement == Undefined) 402 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); 403 404 // False for two elements -> 'i != 47 & i != 72'. 405 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx); 406 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement); 407 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx); 408 return BinaryOperator::CreateAnd(C1, C2); 409 } 410 411 // If the comparison can be replaced with a range comparison for the elements 412 // where it is true, emit the range check. 413 if (TrueRangeEnd != Overdefined) { 414 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare"); 415 416 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1). 417 if (FirstTrueElement) { 418 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement); 419 Idx = Builder->CreateAdd(Idx, Offs); 420 } 421 422 Value *End = ConstantInt::get(Idx->getType(), 423 TrueRangeEnd-FirstTrueElement+1); 424 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); 425 } 426 427 // False range check. 428 if (FalseRangeEnd != Overdefined) { 429 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare"); 430 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). 431 if (FirstFalseElement) { 432 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement); 433 Idx = Builder->CreateAdd(Idx, Offs); 434 } 435 436 Value *End = ConstantInt::get(Idx->getType(), 437 FalseRangeEnd-FirstFalseElement); 438 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); 439 } 440 441 442 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state 443 // of this load, replace it with computation that does: 444 // ((magic_cst >> i) & 1) != 0 445 if (Init->getNumOperands() <= 32 || 446 (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) { 447 const Type *Ty; 448 if (Init->getNumOperands() <= 32) 449 Ty = Type::getInt32Ty(Init->getContext()); 450 else 451 Ty = Type::getInt64Ty(Init->getContext()); 452 Value *V = Builder->CreateIntCast(Idx, Ty, false); 453 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V); 454 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V); 455 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0)); 456 } 457 458 return 0; 459} 460 461 462/// EvaluateGEPOffsetExpression - Return a value that can be used to compare 463/// the *offset* implied by a GEP to zero. For example, if we have &A[i], we 464/// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can 465/// be complex, and scales are involved. The above expression would also be 466/// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). 467/// This later form is less amenable to optimization though, and we are allowed 468/// to generate the first by knowing that pointer arithmetic doesn't overflow. 469/// 470/// If we can't emit an optimized form for this expression, this returns null. 471/// 472static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I, 473 InstCombiner &IC) { 474 TargetData &TD = *IC.getTargetData(); 475 gep_type_iterator GTI = gep_type_begin(GEP); 476 477 // Check to see if this gep only has a single variable index. If so, and if 478 // any constant indices are a multiple of its scale, then we can compute this 479 // in terms of the scale of the variable index. For example, if the GEP 480 // implies an offset of "12 + i*4", then we can codegen this as "3 + i", 481 // because the expression will cross zero at the same point. 482 unsigned i, e = GEP->getNumOperands(); 483 int64_t Offset = 0; 484 for (i = 1; i != e; ++i, ++GTI) { 485 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 486 // Compute the aggregate offset of constant indices. 487 if (CI->isZero()) continue; 488 489 // Handle a struct index, which adds its field offset to the pointer. 490 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 491 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 492 } else { 493 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 494 Offset += Size*CI->getSExtValue(); 495 } 496 } else { 497 // Found our variable index. 498 break; 499 } 500 } 501 502 // If there are no variable indices, we must have a constant offset, just 503 // evaluate it the general way. 504 if (i == e) return 0; 505 506 Value *VariableIdx = GEP->getOperand(i); 507 // Determine the scale factor of the variable element. For example, this is 508 // 4 if the variable index is into an array of i32. 509 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType()); 510 511 // Verify that there are no other variable indices. If so, emit the hard way. 512 for (++i, ++GTI; i != e; ++i, ++GTI) { 513 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i)); 514 if (!CI) return 0; 515 516 // Compute the aggregate offset of constant indices. 517 if (CI->isZero()) continue; 518 519 // Handle a struct index, which adds its field offset to the pointer. 520 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 521 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 522 } else { 523 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 524 Offset += Size*CI->getSExtValue(); 525 } 526 } 527 528 // Okay, we know we have a single variable index, which must be a 529 // pointer/array/vector index. If there is no offset, life is simple, return 530 // the index. 531 unsigned IntPtrWidth = TD.getPointerSizeInBits(); 532 if (Offset == 0) { 533 // Cast to intptrty in case a truncation occurs. If an extension is needed, 534 // we don't need to bother extending: the extension won't affect where the 535 // computation crosses zero. 536 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) 537 VariableIdx = new TruncInst(VariableIdx, 538 TD.getIntPtrType(VariableIdx->getContext()), 539 VariableIdx->getName(), &I); 540 return VariableIdx; 541 } 542 543 // Otherwise, there is an index. The computation we will do will be modulo 544 // the pointer size, so get it. 545 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth); 546 547 Offset &= PtrSizeMask; 548 VariableScale &= PtrSizeMask; 549 550 // To do this transformation, any constant index must be a multiple of the 551 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i", 552 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a 553 // multiple of the variable scale. 554 int64_t NewOffs = Offset / (int64_t)VariableScale; 555 if (Offset != NewOffs*(int64_t)VariableScale) 556 return 0; 557 558 // Okay, we can do this evaluation. Start by converting the index to intptr. 559 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext()); 560 if (VariableIdx->getType() != IntPtrTy) 561 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy, 562 true /*SExt*/, 563 VariableIdx->getName(), &I); 564 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs); 565 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I); 566} 567 568/// FoldGEPICmp - Fold comparisons between a GEP instruction and something 569/// else. At this point we know that the GEP is on the LHS of the comparison. 570Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 571 ICmpInst::Predicate Cond, 572 Instruction &I) { 573 // Look through bitcasts. 574 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS)) 575 RHS = BCI->getOperand(0); 576 577 Value *PtrBase = GEPLHS->getOperand(0); 578 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) { 579 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). 580 // This transformation (ignoring the base and scales) is valid because we 581 // know pointers can't overflow since the gep is inbounds. See if we can 582 // output an optimized form. 583 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this); 584 585 // If not, synthesize the offset the hard way. 586 if (Offset == 0) 587 Offset = EmitGEPOffset(GEPLHS); 588 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset, 589 Constant::getNullValue(Offset->getType())); 590 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) { 591 // If the base pointers are different, but the indices are the same, just 592 // compare the base pointer. 593 if (PtrBase != GEPRHS->getOperand(0)) { 594 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands(); 595 IndicesTheSame &= GEPLHS->getOperand(0)->getType() == 596 GEPRHS->getOperand(0)->getType(); 597 if (IndicesTheSame) 598 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 599 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 600 IndicesTheSame = false; 601 break; 602 } 603 604 // If all indices are the same, just compare the base pointers. 605 if (IndicesTheSame) 606 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), 607 GEPLHS->getOperand(0), GEPRHS->getOperand(0)); 608 609 // Otherwise, the base pointers are different and the indices are 610 // different, bail out. 611 return 0; 612 } 613 614 // If one of the GEPs has all zero indices, recurse. 615 bool AllZeros = true; 616 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 617 if (!isa<Constant>(GEPLHS->getOperand(i)) || 618 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) { 619 AllZeros = false; 620 break; 621 } 622 if (AllZeros) 623 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0), 624 ICmpInst::getSwappedPredicate(Cond), I); 625 626 // If the other GEP has all zero indices, recurse. 627 AllZeros = true; 628 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 629 if (!isa<Constant>(GEPRHS->getOperand(i)) || 630 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) { 631 AllZeros = false; 632 break; 633 } 634 if (AllZeros) 635 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I); 636 637 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) { 638 // If the GEPs only differ by one index, compare it. 639 unsigned NumDifferences = 0; // Keep track of # differences. 640 unsigned DiffOperand = 0; // The operand that differs. 641 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 642 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 643 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() != 644 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) { 645 // Irreconcilable differences. 646 NumDifferences = 2; 647 break; 648 } else { 649 if (NumDifferences++) break; 650 DiffOperand = i; 651 } 652 } 653 654 if (NumDifferences == 0) // SAME GEP? 655 return ReplaceInstUsesWith(I, // No comparison is needed here. 656 ConstantInt::get(Type::getInt1Ty(I.getContext()), 657 ICmpInst::isTrueWhenEqual(Cond))); 658 659 else if (NumDifferences == 1) { 660 Value *LHSV = GEPLHS->getOperand(DiffOperand); 661 Value *RHSV = GEPRHS->getOperand(DiffOperand); 662 // Make sure we do a signed comparison here. 663 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV); 664 } 665 } 666 667 // Only lower this if the icmp is the only user of the GEP or if we expect 668 // the result to fold to a constant! 669 if (TD && 670 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) && 671 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) { 672 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) 673 Value *L = EmitGEPOffset(GEPLHS); 674 Value *R = EmitGEPOffset(GEPRHS); 675 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R); 676 } 677 } 678 return 0; 679} 680 681/// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X". 682Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI, 683 Value *X, ConstantInt *CI, 684 ICmpInst::Predicate Pred, 685 Value *TheAdd) { 686 // If we have X+0, exit early (simplifying logic below) and let it get folded 687 // elsewhere. icmp X+0, X -> icmp X, X 688 if (CI->isZero()) { 689 bool isTrue = ICmpInst::isTrueWhenEqual(Pred); 690 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 691 } 692 693 // (X+4) == X -> false. 694 if (Pred == ICmpInst::ICMP_EQ) 695 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext())); 696 697 // (X+4) != X -> true. 698 if (Pred == ICmpInst::ICMP_NE) 699 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext())); 700 701 // If this is an instruction (as opposed to constantexpr) get NUW/NSW info. 702 bool isNUW = false, isNSW = false; 703 if (BinaryOperator *Add = dyn_cast<BinaryOperator>(TheAdd)) { 704 isNUW = Add->hasNoUnsignedWrap(); 705 isNSW = Add->hasNoSignedWrap(); 706 } 707 708 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, 709 // so the values can never be equal. Similiarly for all other "or equals" 710 // operators. 711 712 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255 713 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253 714 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0 715 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 716 // If this is an NUW add, then this is always false. 717 if (isNUW) 718 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext())); 719 720 Value *R = 721 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI); 722 return new ICmpInst(ICmpInst::ICMP_UGT, X, R); 723 } 724 725 // (X+1) >u X --> X <u (0-1) --> X != 255 726 // (X+2) >u X --> X <u (0-2) --> X <u 254 727 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0 728 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { 729 // If this is an NUW add, then this is always true. 730 if (isNUW) 731 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext())); 732 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI)); 733 } 734 735 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits(); 736 ConstantInt *SMax = ConstantInt::get(X->getContext(), 737 APInt::getSignedMaxValue(BitWidth)); 738 739 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127 740 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125 741 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0 742 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1 743 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126 744 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127 745 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) { 746 // If this is an NSW add, then we have two cases: if the constant is 747 // positive, then this is always false, if negative, this is always true. 748 if (isNSW) { 749 bool isTrue = CI->getValue().isNegative(); 750 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 751 } 752 753 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI)); 754 } 755 756 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127 757 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126 758 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1 759 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2 760 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126 761 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128 762 763 // If this is an NSW add, then we have two cases: if the constant is 764 // positive, then this is always true, if negative, this is always false. 765 if (isNSW) { 766 bool isTrue = !CI->getValue().isNegative(); 767 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 768 } 769 770 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); 771 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1); 772 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C)); 773} 774 775/// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS 776/// and CmpRHS are both known to be integer constants. 777Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI, 778 ConstantInt *DivRHS) { 779 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1)); 780 const APInt &CmpRHSV = CmpRHS->getValue(); 781 782 // FIXME: If the operand types don't match the type of the divide 783 // then don't attempt this transform. The code below doesn't have the 784 // logic to deal with a signed divide and an unsigned compare (and 785 // vice versa). This is because (x /s C1) <s C2 produces different 786 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even 787 // (x /u C1) <u C2. Simply casting the operands and result won't 788 // work. :( The if statement below tests that condition and bails 789 // if it finds it. 790 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv; 791 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned()) 792 return 0; 793 if (DivRHS->isZero()) 794 return 0; // The ProdOV computation fails on divide by zero. 795 if (DivIsSigned && DivRHS->isAllOnesValue()) 796 return 0; // The overflow computation also screws up here 797 if (DivRHS->isOne()) { 798 // This eliminates some funny cases with INT_MIN. 799 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X. 800 return &ICI; 801 } 802 803 // Compute Prod = CI * DivRHS. We are essentially solving an equation 804 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and 805 // C2 (CI). By solving for X we can turn this into a range check 806 // instead of computing a divide. 807 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS); 808 809 // Determine if the product overflows by seeing if the product is 810 // not equal to the divide. Make sure we do the same kind of divide 811 // as in the LHS instruction that we're folding. 812 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) : 813 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS; 814 815 // Get the ICmp opcode 816 ICmpInst::Predicate Pred = ICI.getPredicate(); 817 818 /// If the division is known to be exact, then there is no remainder from the 819 /// divide, so the covered range size is unit, otherwise it is the divisor. 820 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS; 821 822 // Figure out the interval that is being checked. For example, a comparison 823 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). 824 // Compute this interval based on the constants involved and the signedness of 825 // the compare/divide. This computes a half-open interval, keeping track of 826 // whether either value in the interval overflows. After analysis each 827 // overflow variable is set to 0 if it's corresponding bound variable is valid 828 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. 829 int LoOverflow = 0, HiOverflow = 0; 830 Constant *LoBound = 0, *HiBound = 0; 831 832 if (!DivIsSigned) { // udiv 833 // e.g. X/5 op 3 --> [15, 20) 834 LoBound = Prod; 835 HiOverflow = LoOverflow = ProdOV; 836 if (!HiOverflow) { 837 // If this is not an exact divide, then many values in the range collapse 838 // to the same result value. 839 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false); 840 } 841 842 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0. 843 if (CmpRHSV == 0) { // (X / pos) op 0 844 // Can't overflow. e.g. X/2 op 0 --> [-1, 2) 845 LoBound = ConstantExpr::getNeg(SubOne(RangeSize)); 846 HiBound = RangeSize; 847 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos 848 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) 849 HiOverflow = LoOverflow = ProdOV; 850 if (!HiOverflow) 851 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true); 852 } else { // (X / pos) op neg 853 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) 854 HiBound = AddOne(Prod); 855 LoOverflow = HiOverflow = ProdOV ? -1 : 0; 856 if (!LoOverflow) { 857 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 858 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0; 859 } 860 } 861 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0. 862 if (DivI->isExact()) 863 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 864 if (CmpRHSV == 0) { // (X / neg) op 0 865 // e.g. X/-5 op 0 --> [-4, 5) 866 LoBound = AddOne(RangeSize); 867 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 868 if (HiBound == DivRHS) { // -INTMIN = INTMIN 869 HiOverflow = 1; // [INTMIN+1, overflow) 870 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN 871 } 872 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos 873 // e.g. X/-5 op 3 --> [-19, -14) 874 HiBound = AddOne(Prod); 875 HiOverflow = LoOverflow = ProdOV ? -1 : 0; 876 if (!LoOverflow) 877 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0; 878 } else { // (X / neg) op neg 879 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) 880 LoOverflow = HiOverflow = ProdOV; 881 if (!HiOverflow) 882 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true); 883 } 884 885 // Dividing by a negative swaps the condition. LT <-> GT 886 Pred = ICmpInst::getSwappedPredicate(Pred); 887 } 888 889 Value *X = DivI->getOperand(0); 890 switch (Pred) { 891 default: llvm_unreachable("Unhandled icmp opcode!"); 892 case ICmpInst::ICMP_EQ: 893 if (LoOverflow && HiOverflow) 894 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 895 if (HiOverflow) 896 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 897 ICmpInst::ICMP_UGE, X, LoBound); 898 if (LoOverflow) 899 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 900 ICmpInst::ICMP_ULT, X, HiBound); 901 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 902 DivIsSigned, true)); 903 case ICmpInst::ICMP_NE: 904 if (LoOverflow && HiOverflow) 905 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 906 if (HiOverflow) 907 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 908 ICmpInst::ICMP_ULT, X, LoBound); 909 if (LoOverflow) 910 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 911 ICmpInst::ICMP_UGE, X, HiBound); 912 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 913 DivIsSigned, false)); 914 case ICmpInst::ICMP_ULT: 915 case ICmpInst::ICMP_SLT: 916 if (LoOverflow == +1) // Low bound is greater than input range. 917 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 918 if (LoOverflow == -1) // Low bound is less than input range. 919 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 920 return new ICmpInst(Pred, X, LoBound); 921 case ICmpInst::ICMP_UGT: 922 case ICmpInst::ICMP_SGT: 923 if (HiOverflow == +1) // High bound greater than input range. 924 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 925 if (HiOverflow == -1) // High bound less than input range. 926 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 927 if (Pred == ICmpInst::ICMP_UGT) 928 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound); 929 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound); 930 } 931} 932 933/// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)". 934Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr, 935 ConstantInt *ShAmt) { 936 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue(); 937 938 // Check that the shift amount is in range. If not, don't perform 939 // undefined shifts. When the shift is visited it will be 940 // simplified. 941 uint32_t TypeBits = CmpRHSV.getBitWidth(); 942 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 943 if (ShAmtVal >= TypeBits || ShAmtVal == 0) 944 return 0; 945 946 if (!ICI.isEquality()) { 947 // If we have an unsigned comparison and an ashr, we can't simplify this. 948 // Similarly for signed comparisons with lshr. 949 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr)) 950 return 0; 951 952 // Otherwise, all lshr and all exact ashr's are equivalent to a udiv/sdiv by 953 // a power of 2. Since we already have logic to simplify these, transform 954 // to div and then simplify the resultant comparison. 955 if (Shr->getOpcode() == Instruction::AShr && 956 !Shr->isExact()) 957 return 0; 958 959 // Revisit the shift (to delete it). 960 Worklist.Add(Shr); 961 962 Constant *DivCst = 963 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal)); 964 965 Value *Tmp = 966 Shr->getOpcode() == Instruction::AShr ? 967 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) : 968 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()); 969 970 ICI.setOperand(0, Tmp); 971 972 // If the builder folded the binop, just return it. 973 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp); 974 if (TheDiv == 0) 975 return &ICI; 976 977 // Otherwise, fold this div/compare. 978 assert(TheDiv->getOpcode() == Instruction::SDiv || 979 TheDiv->getOpcode() == Instruction::UDiv); 980 981 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst)); 982 assert(Res && "This div/cst should have folded!"); 983 return Res; 984 } 985 986 987 // If we are comparing against bits always shifted out, the 988 // comparison cannot succeed. 989 APInt Comp = CmpRHSV << ShAmtVal; 990 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp); 991 if (Shr->getOpcode() == Instruction::LShr) 992 Comp = Comp.lshr(ShAmtVal); 993 else 994 Comp = Comp.ashr(ShAmtVal); 995 996 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero. 997 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 998 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 999 IsICMP_NE); 1000 return ReplaceInstUsesWith(ICI, Cst); 1001 } 1002 1003 // Otherwise, check to see if the bits shifted out are known to be zero. 1004 // If so, we can compare against the unshifted value: 1005 // (X & 4) >> 1 == 2 --> (X & 4) == 4. 1006 if (Shr->hasOneUse() && Shr->isExact()) 1007 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS); 1008 1009 if (Shr->hasOneUse()) { 1010 // Otherwise strength reduce the shift into an and. 1011 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal)); 1012 Constant *Mask = ConstantInt::get(ICI.getContext(), Val); 1013 1014 Value *And = Builder->CreateAnd(Shr->getOperand(0), 1015 Mask, Shr->getName()+".mask"); 1016 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS); 1017 } 1018 return 0; 1019} 1020 1021 1022/// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)". 1023/// 1024Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI, 1025 Instruction *LHSI, 1026 ConstantInt *RHS) { 1027 const APInt &RHSV = RHS->getValue(); 1028 1029 switch (LHSI->getOpcode()) { 1030 case Instruction::Trunc: 1031 if (ICI.isEquality() && LHSI->hasOneUse()) { 1032 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all 1033 // of the high bits truncated out of x are known. 1034 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(), 1035 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1036 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits)); 1037 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0); 1038 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne); 1039 1040 // If all the high bits are known, we can do this xform. 1041 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) { 1042 // Pull in the high bits from known-ones set. 1043 APInt NewRHS = RHS->getValue().zext(SrcBits); 1044 NewRHS |= KnownOne; 1045 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1046 ConstantInt::get(ICI.getContext(), NewRHS)); 1047 } 1048 } 1049 break; 1050 1051 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI) 1052 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1053 // If this is a comparison that tests the signbit (X < 0) or (x > -1), 1054 // fold the xor. 1055 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) || 1056 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) { 1057 Value *CompareVal = LHSI->getOperand(0); 1058 1059 // If the sign bit of the XorCST is not set, there is no change to 1060 // the operation, just stop using the Xor. 1061 if (!XorCST->getValue().isNegative()) { 1062 ICI.setOperand(0, CompareVal); 1063 Worklist.Add(LHSI); 1064 return &ICI; 1065 } 1066 1067 // Was the old condition true if the operand is positive? 1068 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT; 1069 1070 // If so, the new one isn't. 1071 isTrueIfPositive ^= true; 1072 1073 if (isTrueIfPositive) 1074 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, 1075 SubOne(RHS)); 1076 else 1077 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, 1078 AddOne(RHS)); 1079 } 1080 1081 if (LHSI->hasOneUse()) { 1082 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit)) 1083 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) { 1084 const APInt &SignBit = XorCST->getValue(); 1085 ICmpInst::Predicate Pred = ICI.isSigned() 1086 ? ICI.getUnsignedPredicate() 1087 : ICI.getSignedPredicate(); 1088 return new ICmpInst(Pred, LHSI->getOperand(0), 1089 ConstantInt::get(ICI.getContext(), 1090 RHSV ^ SignBit)); 1091 } 1092 1093 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A) 1094 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) { 1095 const APInt &NotSignBit = XorCST->getValue(); 1096 ICmpInst::Predicate Pred = ICI.isSigned() 1097 ? ICI.getUnsignedPredicate() 1098 : ICI.getSignedPredicate(); 1099 Pred = ICI.getSwappedPredicate(Pred); 1100 return new ICmpInst(Pred, LHSI->getOperand(0), 1101 ConstantInt::get(ICI.getContext(), 1102 RHSV ^ NotSignBit)); 1103 } 1104 } 1105 } 1106 break; 1107 case Instruction::And: // (icmp pred (and X, AndCST), RHS) 1108 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) && 1109 LHSI->getOperand(0)->hasOneUse()) { 1110 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1)); 1111 1112 // If the LHS is an AND of a truncating cast, we can widen the 1113 // and/compare to be the input width without changing the value 1114 // produced, eliminating a cast. 1115 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) { 1116 // We can do this transformation if either the AND constant does not 1117 // have its sign bit set or if it is an equality comparison. 1118 // Extending a relational comparison when we're checking the sign 1119 // bit would not work. 1120 if (Cast->hasOneUse() && 1121 (ICI.isEquality() || 1122 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) { 1123 uint32_t BitWidth = 1124 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth(); 1125 APInt NewCST = AndCST->getValue().zext(BitWidth); 1126 APInt NewCI = RHSV.zext(BitWidth); 1127 Value *NewAnd = 1128 Builder->CreateAnd(Cast->getOperand(0), 1129 ConstantInt::get(ICI.getContext(), NewCST), 1130 LHSI->getName()); 1131 return new ICmpInst(ICI.getPredicate(), NewAnd, 1132 ConstantInt::get(ICI.getContext(), NewCI)); 1133 } 1134 } 1135 1136 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare 1137 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This 1138 // happens a LOT in code produced by the C front-end, for bitfield 1139 // access. 1140 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0)); 1141 if (Shift && !Shift->isShift()) 1142 Shift = 0; 1143 1144 ConstantInt *ShAmt; 1145 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0; 1146 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift. 1147 const Type *AndTy = AndCST->getType(); // Type of the and. 1148 1149 // We can fold this as long as we can't shift unknown bits 1150 // into the mask. This can only happen with signed shift 1151 // rights, as they sign-extend. 1152 if (ShAmt) { 1153 bool CanFold = Shift->isLogicalShift(); 1154 if (!CanFold) { 1155 // To test for the bad case of the signed shr, see if any 1156 // of the bits shifted in could be tested after the mask. 1157 uint32_t TyBits = Ty->getPrimitiveSizeInBits(); 1158 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits); 1159 1160 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits(); 1161 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) & 1162 AndCST->getValue()) == 0) 1163 CanFold = true; 1164 } 1165 1166 if (CanFold) { 1167 Constant *NewCst; 1168 if (Shift->getOpcode() == Instruction::Shl) 1169 NewCst = ConstantExpr::getLShr(RHS, ShAmt); 1170 else 1171 NewCst = ConstantExpr::getShl(RHS, ShAmt); 1172 1173 // Check to see if we are shifting out any of the bits being 1174 // compared. 1175 if (ConstantExpr::get(Shift->getOpcode(), 1176 NewCst, ShAmt) != RHS) { 1177 // If we shifted bits out, the fold is not going to work out. 1178 // As a special case, check to see if this means that the 1179 // result is always true or false now. 1180 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1181 return ReplaceInstUsesWith(ICI, 1182 ConstantInt::getFalse(ICI.getContext())); 1183 if (ICI.getPredicate() == ICmpInst::ICMP_NE) 1184 return ReplaceInstUsesWith(ICI, 1185 ConstantInt::getTrue(ICI.getContext())); 1186 } else { 1187 ICI.setOperand(1, NewCst); 1188 Constant *NewAndCST; 1189 if (Shift->getOpcode() == Instruction::Shl) 1190 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt); 1191 else 1192 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt); 1193 LHSI->setOperand(1, NewAndCST); 1194 LHSI->setOperand(0, Shift->getOperand(0)); 1195 Worklist.Add(Shift); // Shift is dead. 1196 return &ICI; 1197 } 1198 } 1199 } 1200 1201 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is 1202 // preferable because it allows the C<<Y expression to be hoisted out 1203 // of a loop if Y is invariant and X is not. 1204 if (Shift && Shift->hasOneUse() && RHSV == 0 && 1205 ICI.isEquality() && !Shift->isArithmeticShift() && 1206 !isa<Constant>(Shift->getOperand(0))) { 1207 // Compute C << Y. 1208 Value *NS; 1209 if (Shift->getOpcode() == Instruction::LShr) { 1210 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp"); 1211 } else { 1212 // Insert a logical shift. 1213 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp"); 1214 } 1215 1216 // Compute X & (C << Y). 1217 Value *NewAnd = 1218 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName()); 1219 1220 ICI.setOperand(0, NewAnd); 1221 return &ICI; 1222 } 1223 } 1224 1225 // Try to optimize things like "A[i]&42 == 0" to index computations. 1226 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) { 1227 if (GetElementPtrInst *GEP = 1228 dyn_cast<GetElementPtrInst>(LI->getOperand(0))) 1229 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 1230 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 1231 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) { 1232 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1)); 1233 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C)) 1234 return Res; 1235 } 1236 } 1237 break; 1238 1239 case Instruction::Or: { 1240 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse()) 1241 break; 1242 Value *P, *Q; 1243 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) { 1244 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 1245 // -> and (icmp eq P, null), (icmp eq Q, null). 1246 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P, 1247 Constant::getNullValue(P->getType())); 1248 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q, 1249 Constant::getNullValue(Q->getType())); 1250 Instruction *Op; 1251 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1252 Op = BinaryOperator::CreateAnd(ICIP, ICIQ); 1253 else 1254 Op = BinaryOperator::CreateOr(ICIP, ICIQ); 1255 return Op; 1256 } 1257 break; 1258 } 1259 1260 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI) 1261 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1262 if (!ShAmt) break; 1263 1264 uint32_t TypeBits = RHSV.getBitWidth(); 1265 1266 // Check that the shift amount is in range. If not, don't perform 1267 // undefined shifts. When the shift is visited it will be 1268 // simplified. 1269 if (ShAmt->uge(TypeBits)) 1270 break; 1271 1272 if (ICI.isEquality()) { 1273 // If we are comparing against bits always shifted out, the 1274 // comparison cannot succeed. 1275 Constant *Comp = 1276 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), 1277 ShAmt); 1278 if (Comp != RHS) {// Comparing against a bit that we know is zero. 1279 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1280 Constant *Cst = 1281 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE); 1282 return ReplaceInstUsesWith(ICI, Cst); 1283 } 1284 1285 // If the shift is NUW, then it is just shifting out zeros, no need for an 1286 // AND. 1287 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap()) 1288 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1289 ConstantExpr::getLShr(RHS, ShAmt)); 1290 1291 if (LHSI->hasOneUse()) { 1292 // Otherwise strength reduce the shift into an and. 1293 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 1294 Constant *Mask = 1295 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits, 1296 TypeBits-ShAmtVal)); 1297 1298 Value *And = 1299 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask"); 1300 return new ICmpInst(ICI.getPredicate(), And, 1301 ConstantExpr::getLShr(RHS, ShAmt)); 1302 } 1303 } 1304 1305 // Otherwise, if this is a comparison of the sign bit, simplify to and/test. 1306 bool TrueIfSigned = false; 1307 if (LHSI->hasOneUse() && 1308 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) { 1309 // (X << 31) <s 0 --> (X&1) != 0 1310 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(), 1311 APInt::getOneBitSet(TypeBits, 1312 TypeBits-ShAmt->getZExtValue()-1)); 1313 Value *And = 1314 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask"); 1315 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, 1316 And, Constant::getNullValue(And->getType())); 1317 } 1318 break; 1319 } 1320 1321 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI) 1322 case Instruction::AShr: 1323 // Only handle equality comparisons of shift-by-constant. 1324 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) 1325 if (Instruction *Res = FoldICmpShrCst(ICI, cast<BinaryOperator>(LHSI), 1326 ShAmt)) 1327 return Res; 1328 break; 1329 1330 case Instruction::SDiv: 1331 case Instruction::UDiv: 1332 // Fold: icmp pred ([us]div X, C1), C2 -> range test 1333 // Fold this div into the comparison, producing a range check. 1334 // Determine, based on the divide type, what the range is being 1335 // checked. If there is an overflow on the low or high side, remember 1336 // it, otherwise compute the range [low, hi) bounding the new value. 1337 // See: InsertRangeTest above for the kinds of replacements possible. 1338 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) 1339 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI), 1340 DivRHS)) 1341 return R; 1342 break; 1343 1344 case Instruction::Add: 1345 // Fold: icmp pred (add X, C1), C2 1346 if (!ICI.isEquality()) { 1347 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1348 if (!LHSC) break; 1349 const APInt &LHSV = LHSC->getValue(); 1350 1351 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV) 1352 .subtract(LHSV); 1353 1354 if (ICI.isSigned()) { 1355 if (CR.getLower().isSignBit()) { 1356 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0), 1357 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1358 } else if (CR.getUpper().isSignBit()) { 1359 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0), 1360 ConstantInt::get(ICI.getContext(),CR.getLower())); 1361 } 1362 } else { 1363 if (CR.getLower().isMinValue()) { 1364 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), 1365 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1366 } else if (CR.getUpper().isMinValue()) { 1367 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), 1368 ConstantInt::get(ICI.getContext(),CR.getLower())); 1369 } 1370 } 1371 } 1372 break; 1373 } 1374 1375 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS. 1376 if (ICI.isEquality()) { 1377 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1378 1379 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and 1380 // the second operand is a constant, simplify a bit. 1381 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) { 1382 switch (BO->getOpcode()) { 1383 case Instruction::SRem: 1384 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. 1385 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){ 1386 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue(); 1387 if (V.sgt(1) && V.isPowerOf2()) { 1388 Value *NewRem = 1389 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1), 1390 BO->getName()); 1391 return new ICmpInst(ICI.getPredicate(), NewRem, 1392 Constant::getNullValue(BO->getType())); 1393 } 1394 } 1395 break; 1396 case Instruction::Add: 1397 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants. 1398 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1399 if (BO->hasOneUse()) 1400 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1401 ConstantExpr::getSub(RHS, BOp1C)); 1402 } else if (RHSV == 0) { 1403 // Replace ((add A, B) != 0) with (A != -B) if A or B is 1404 // efficiently invertible, or if the add has just this one use. 1405 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1); 1406 1407 if (Value *NegVal = dyn_castNegVal(BOp1)) 1408 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal); 1409 else if (Value *NegVal = dyn_castNegVal(BOp0)) 1410 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1); 1411 else if (BO->hasOneUse()) { 1412 Value *Neg = Builder->CreateNeg(BOp1); 1413 Neg->takeName(BO); 1414 return new ICmpInst(ICI.getPredicate(), BOp0, Neg); 1415 } 1416 } 1417 break; 1418 case Instruction::Xor: 1419 // For the xor case, we can xor two constants together, eliminating 1420 // the explicit xor. 1421 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) 1422 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1423 ConstantExpr::getXor(RHS, BOC)); 1424 1425 // FALLTHROUGH 1426 case Instruction::Sub: 1427 // Replace (([sub|xor] A, B) != 0) with (A != B) 1428 if (RHSV == 0) 1429 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1430 BO->getOperand(1)); 1431 break; 1432 1433 case Instruction::Or: 1434 // If bits are being or'd in that are not present in the constant we 1435 // are comparing against, then the comparison could never succeed! 1436 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1437 Constant *NotCI = ConstantExpr::getNot(RHS); 1438 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue()) 1439 return ReplaceInstUsesWith(ICI, 1440 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1441 isICMP_NE)); 1442 } 1443 break; 1444 1445 case Instruction::And: 1446 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1447 // If bits are being compared against that are and'd out, then the 1448 // comparison can never succeed! 1449 if ((RHSV & ~BOC->getValue()) != 0) 1450 return ReplaceInstUsesWith(ICI, 1451 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1452 isICMP_NE)); 1453 1454 // If we have ((X & C) == C), turn it into ((X & C) != 0). 1455 if (RHS == BOC && RHSV.isPowerOf2()) 1456 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : 1457 ICmpInst::ICMP_NE, LHSI, 1458 Constant::getNullValue(RHS->getType())); 1459 1460 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0 1461 if (BOC->getValue().isSignBit()) { 1462 Value *X = BO->getOperand(0); 1463 Constant *Zero = Constant::getNullValue(X->getType()); 1464 ICmpInst::Predicate pred = isICMP_NE ? 1465 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; 1466 return new ICmpInst(pred, X, Zero); 1467 } 1468 1469 // ((X & ~7) == 0) --> X < 8 1470 if (RHSV == 0 && isHighOnes(BOC)) { 1471 Value *X = BO->getOperand(0); 1472 Constant *NegX = ConstantExpr::getNeg(BOC); 1473 ICmpInst::Predicate pred = isICMP_NE ? 1474 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; 1475 return new ICmpInst(pred, X, NegX); 1476 } 1477 } 1478 default: break; 1479 } 1480 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) { 1481 // Handle icmp {eq|ne} <intrinsic>, intcst. 1482 switch (II->getIntrinsicID()) { 1483 case Intrinsic::bswap: 1484 Worklist.Add(II); 1485 ICI.setOperand(0, II->getArgOperand(0)); 1486 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap())); 1487 return &ICI; 1488 case Intrinsic::ctlz: 1489 case Intrinsic::cttz: 1490 // ctz(A) == bitwidth(a) -> A == 0 and likewise for != 1491 if (RHSV == RHS->getType()->getBitWidth()) { 1492 Worklist.Add(II); 1493 ICI.setOperand(0, II->getArgOperand(0)); 1494 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0)); 1495 return &ICI; 1496 } 1497 break; 1498 case Intrinsic::ctpop: 1499 // popcount(A) == 0 -> A == 0 and likewise for != 1500 if (RHS->isZero()) { 1501 Worklist.Add(II); 1502 ICI.setOperand(0, II->getArgOperand(0)); 1503 ICI.setOperand(1, RHS); 1504 return &ICI; 1505 } 1506 break; 1507 default: 1508 break; 1509 } 1510 } 1511 } 1512 return 0; 1513} 1514 1515/// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst). 1516/// We only handle extending casts so far. 1517/// 1518Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) { 1519 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0)); 1520 Value *LHSCIOp = LHSCI->getOperand(0); 1521 const Type *SrcTy = LHSCIOp->getType(); 1522 const Type *DestTy = LHSCI->getType(); 1523 Value *RHSCIOp; 1524 1525 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the 1526 // integer type is the same size as the pointer type. 1527 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt && 1528 TD->getPointerSizeInBits() == 1529 cast<IntegerType>(DestTy)->getBitWidth()) { 1530 Value *RHSOp = 0; 1531 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) { 1532 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy); 1533 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) { 1534 RHSOp = RHSC->getOperand(0); 1535 // If the pointer types don't match, insert a bitcast. 1536 if (LHSCIOp->getType() != RHSOp->getType()) 1537 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType()); 1538 } 1539 1540 if (RHSOp) 1541 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp); 1542 } 1543 1544 // The code below only handles extension cast instructions, so far. 1545 // Enforce this. 1546 if (LHSCI->getOpcode() != Instruction::ZExt && 1547 LHSCI->getOpcode() != Instruction::SExt) 1548 return 0; 1549 1550 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt; 1551 bool isSignedCmp = ICI.isSigned(); 1552 1553 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) { 1554 // Not an extension from the same type? 1555 RHSCIOp = CI->getOperand(0); 1556 if (RHSCIOp->getType() != LHSCIOp->getType()) 1557 return 0; 1558 1559 // If the signedness of the two casts doesn't agree (i.e. one is a sext 1560 // and the other is a zext), then we can't handle this. 1561 if (CI->getOpcode() != LHSCI->getOpcode()) 1562 return 0; 1563 1564 // Deal with equality cases early. 1565 if (ICI.isEquality()) 1566 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1567 1568 // A signed comparison of sign extended values simplifies into a 1569 // signed comparison. 1570 if (isSignedCmp && isSignedExt) 1571 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1572 1573 // The other three cases all fold into an unsigned comparison. 1574 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp); 1575 } 1576 1577 // If we aren't dealing with a constant on the RHS, exit early 1578 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1)); 1579 if (!CI) 1580 return 0; 1581 1582 // Compute the constant that would happen if we truncated to SrcTy then 1583 // reextended to DestTy. 1584 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy); 1585 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), 1586 Res1, DestTy); 1587 1588 // If the re-extended constant didn't change... 1589 if (Res2 == CI) { 1590 // Deal with equality cases early. 1591 if (ICI.isEquality()) 1592 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1593 1594 // A signed comparison of sign extended values simplifies into a 1595 // signed comparison. 1596 if (isSignedExt && isSignedCmp) 1597 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1598 1599 // The other three cases all fold into an unsigned comparison. 1600 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1); 1601 } 1602 1603 // The re-extended constant changed so the constant cannot be represented 1604 // in the shorter type. Consequently, we cannot emit a simple comparison. 1605 // All the cases that fold to true or false will have already been handled 1606 // by SimplifyICmpInst, so only deal with the tricky case. 1607 1608 if (isSignedCmp || !isSignedExt) 1609 return 0; 1610 1611 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases 1612 // should have been folded away previously and not enter in here. 1613 1614 // We're performing an unsigned comp with a sign extended value. 1615 // This is true if the input is >= 0. [aka >s -1] 1616 Constant *NegOne = Constant::getAllOnesValue(SrcTy); 1617 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName()); 1618 1619 // Finally, return the value computed. 1620 if (ICI.getPredicate() == ICmpInst::ICMP_ULT) 1621 return ReplaceInstUsesWith(ICI, Result); 1622 1623 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"); 1624 return BinaryOperator::CreateNot(Result); 1625} 1626 1627/// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form: 1628/// I = icmp ugt (add (add A, B), CI2), CI1 1629/// If this is of the form: 1630/// sum = a + b 1631/// if (sum+128 >u 255) 1632/// Then replace it with llvm.sadd.with.overflow.i8. 1633/// 1634static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B, 1635 ConstantInt *CI2, ConstantInt *CI1, 1636 InstCombiner &IC) { 1637 // The transformation we're trying to do here is to transform this into an 1638 // llvm.sadd.with.overflow. To do this, we have to replace the original add 1639 // with a narrower add, and discard the add-with-constant that is part of the 1640 // range check (if we can't eliminate it, this isn't profitable). 1641 1642 // In order to eliminate the add-with-constant, the compare can be its only 1643 // use. 1644 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0)); 1645 if (!AddWithCst->hasOneUse()) return 0; 1646 1647 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow. 1648 if (!CI2->getValue().isPowerOf2()) return 0; 1649 unsigned NewWidth = CI2->getValue().countTrailingZeros(); 1650 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0; 1651 1652 // The width of the new add formed is 1 more than the bias. 1653 ++NewWidth; 1654 1655 // Check to see that CI1 is an all-ones value with NewWidth bits. 1656 if (CI1->getBitWidth() == NewWidth || 1657 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth)) 1658 return 0; 1659 1660 // In order to replace the original add with a narrower 1661 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant 1662 // and truncates that discard the high bits of the add. Verify that this is 1663 // the case. 1664 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0)); 1665 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end(); 1666 UI != E; ++UI) { 1667 if (*UI == AddWithCst) continue; 1668 1669 // Only accept truncates for now. We would really like a nice recursive 1670 // predicate like SimplifyDemandedBits, but which goes downwards the use-def 1671 // chain to see which bits of a value are actually demanded. If the 1672 // original add had another add which was then immediately truncated, we 1673 // could still do the transformation. 1674 TruncInst *TI = dyn_cast<TruncInst>(*UI); 1675 if (TI == 0 || 1676 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0; 1677 } 1678 1679 // If the pattern matches, truncate the inputs to the narrower type and 1680 // use the sadd_with_overflow intrinsic to efficiently compute both the 1681 // result and the overflow bit. 1682 Module *M = I.getParent()->getParent()->getParent(); 1683 1684 const Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth); 1685 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow, 1686 &NewType, 1); 1687 1688 InstCombiner::BuilderTy *Builder = IC.Builder; 1689 1690 // Put the new code above the original add, in case there are any uses of the 1691 // add between the add and the compare. 1692 Builder->SetInsertPoint(OrigAdd); 1693 1694 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc"); 1695 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc"); 1696 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd"); 1697 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result"); 1698 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType()); 1699 1700 // The inner add was the result of the narrow add, zero extended to the 1701 // wider type. Replace it with the result computed by the intrinsic. 1702 IC.ReplaceInstUsesWith(*OrigAdd, ZExt); 1703 1704 // The original icmp gets replaced with the overflow value. 1705 return ExtractValueInst::Create(Call, 1, "sadd.overflow"); 1706} 1707 1708static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV, 1709 InstCombiner &IC) { 1710 // Don't bother doing this transformation for pointers, don't do it for 1711 // vectors. 1712 if (!isa<IntegerType>(OrigAddV->getType())) return 0; 1713 1714 // If the add is a constant expr, then we don't bother transforming it. 1715 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV); 1716 if (OrigAdd == 0) return 0; 1717 1718 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1); 1719 1720 // Put the new code above the original add, in case there are any uses of the 1721 // add between the add and the compare. 1722 InstCombiner::BuilderTy *Builder = IC.Builder; 1723 Builder->SetInsertPoint(OrigAdd); 1724 1725 Module *M = I.getParent()->getParent()->getParent(); 1726 const Type *Ty = LHS->getType(); 1727 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, &Ty,1); 1728 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd"); 1729 Value *Add = Builder->CreateExtractValue(Call, 0); 1730 1731 IC.ReplaceInstUsesWith(*OrigAdd, Add); 1732 1733 // The original icmp gets replaced with the overflow value. 1734 return ExtractValueInst::Create(Call, 1, "uadd.overflow"); 1735} 1736 1737// DemandedBitsLHSMask - When performing a comparison against a constant, 1738// it is possible that not all the bits in the LHS are demanded. This helper 1739// method computes the mask that IS demanded. 1740static APInt DemandedBitsLHSMask(ICmpInst &I, 1741 unsigned BitWidth, bool isSignCheck) { 1742 if (isSignCheck) 1743 return APInt::getSignBit(BitWidth); 1744 1745 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1)); 1746 if (!CI) return APInt::getAllOnesValue(BitWidth); 1747 const APInt &RHS = CI->getValue(); 1748 1749 switch (I.getPredicate()) { 1750 // For a UGT comparison, we don't care about any bits that 1751 // correspond to the trailing ones of the comparand. The value of these 1752 // bits doesn't impact the outcome of the comparison, because any value 1753 // greater than the RHS must differ in a bit higher than these due to carry. 1754 case ICmpInst::ICMP_UGT: { 1755 unsigned trailingOnes = RHS.countTrailingOnes(); 1756 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes); 1757 return ~lowBitsSet; 1758 } 1759 1760 // Similarly, for a ULT comparison, we don't care about the trailing zeros. 1761 // Any value less than the RHS must differ in a higher bit because of carries. 1762 case ICmpInst::ICMP_ULT: { 1763 unsigned trailingZeros = RHS.countTrailingZeros(); 1764 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros); 1765 return ~lowBitsSet; 1766 } 1767 1768 default: 1769 return APInt::getAllOnesValue(BitWidth); 1770 } 1771 1772} 1773 1774Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { 1775 bool Changed = false; 1776 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1777 1778 /// Orders the operands of the compare so that they are listed from most 1779 /// complex to least complex. This puts constants before unary operators, 1780 /// before binary operators. 1781 if (getComplexity(Op0) < getComplexity(Op1)) { 1782 I.swapOperands(); 1783 std::swap(Op0, Op1); 1784 Changed = true; 1785 } 1786 1787 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD)) 1788 return ReplaceInstUsesWith(I, V); 1789 1790 const Type *Ty = Op0->getType(); 1791 1792 // icmp's with boolean values can always be turned into bitwise operations 1793 if (Ty->isIntegerTy(1)) { 1794 switch (I.getPredicate()) { 1795 default: llvm_unreachable("Invalid icmp instruction!"); 1796 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B) 1797 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp"); 1798 return BinaryOperator::CreateNot(Xor); 1799 } 1800 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B 1801 return BinaryOperator::CreateXor(Op0, Op1); 1802 1803 case ICmpInst::ICMP_UGT: 1804 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult 1805 // FALL THROUGH 1806 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B 1807 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1808 return BinaryOperator::CreateAnd(Not, Op1); 1809 } 1810 case ICmpInst::ICMP_SGT: 1811 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt 1812 // FALL THROUGH 1813 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B 1814 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1815 return BinaryOperator::CreateAnd(Not, Op0); 1816 } 1817 case ICmpInst::ICMP_UGE: 1818 std::swap(Op0, Op1); // Change icmp uge -> icmp ule 1819 // FALL THROUGH 1820 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B 1821 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1822 return BinaryOperator::CreateOr(Not, Op1); 1823 } 1824 case ICmpInst::ICMP_SGE: 1825 std::swap(Op0, Op1); // Change icmp sge -> icmp sle 1826 // FALL THROUGH 1827 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B 1828 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1829 return BinaryOperator::CreateOr(Not, Op0); 1830 } 1831 } 1832 } 1833 1834 unsigned BitWidth = 0; 1835 if (Ty->isIntOrIntVectorTy()) 1836 BitWidth = Ty->getScalarSizeInBits(); 1837 else if (TD) // Pointers require TD info to get their size. 1838 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType()); 1839 1840 bool isSignBit = false; 1841 1842 // See if we are doing a comparison with a constant. 1843 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 1844 Value *A = 0, *B = 0; 1845 1846 // Match the following pattern, which is a common idiom when writing 1847 // overflow-safe integer arithmetic function. The source performs an 1848 // addition in wider type, and explicitly checks for overflow using 1849 // comparisons against INT_MIN and INT_MAX. Simplify this by using the 1850 // sadd_with_overflow intrinsic. 1851 // 1852 // TODO: This could probably be generalized to handle other overflow-safe 1853 // operations if we worked out the formulas to compute the appropriate 1854 // magic constants. 1855 // 1856 // sum = a + b 1857 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8 1858 { 1859 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI 1860 if (I.getPredicate() == ICmpInst::ICMP_UGT && 1861 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2)))) 1862 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this)) 1863 return Res; 1864 } 1865 1866 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B) 1867 if (I.isEquality() && CI->isZero() && 1868 match(Op0, m_Sub(m_Value(A), m_Value(B)))) { 1869 // (icmp cond A B) if cond is equality 1870 return new ICmpInst(I.getPredicate(), A, B); 1871 } 1872 1873 // If we have an icmp le or icmp ge instruction, turn it into the 1874 // appropriate icmp lt or icmp gt instruction. This allows us to rely on 1875 // them being folded in the code below. The SimplifyICmpInst code has 1876 // already handled the edge cases for us, so we just assert on them. 1877 switch (I.getPredicate()) { 1878 default: break; 1879 case ICmpInst::ICMP_ULE: 1880 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE 1881 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, 1882 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1883 case ICmpInst::ICMP_SLE: 1884 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE 1885 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 1886 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1887 case ICmpInst::ICMP_UGE: 1888 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE 1889 return new ICmpInst(ICmpInst::ICMP_UGT, Op0, 1890 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1891 case ICmpInst::ICMP_SGE: 1892 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE 1893 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 1894 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1895 } 1896 1897 // If this comparison is a normal comparison, it demands all 1898 // bits, if it is a sign bit comparison, it only demands the sign bit. 1899 bool UnusedBit; 1900 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit); 1901 } 1902 1903 // See if we can fold the comparison based on range information we can get 1904 // by checking whether bits are known to be zero or one in the input. 1905 if (BitWidth != 0) { 1906 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0); 1907 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0); 1908 1909 if (SimplifyDemandedBits(I.getOperandUse(0), 1910 DemandedBitsLHSMask(I, BitWidth, isSignBit), 1911 Op0KnownZero, Op0KnownOne, 0)) 1912 return &I; 1913 if (SimplifyDemandedBits(I.getOperandUse(1), 1914 APInt::getAllOnesValue(BitWidth), 1915 Op1KnownZero, Op1KnownOne, 0)) 1916 return &I; 1917 1918 // Given the known and unknown bits, compute a range that the LHS could be 1919 // in. Compute the Min, Max and RHS values based on the known bits. For the 1920 // EQ and NE we use unsigned values. 1921 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0); 1922 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0); 1923 if (I.isSigned()) { 1924 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 1925 Op0Min, Op0Max); 1926 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 1927 Op1Min, Op1Max); 1928 } else { 1929 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 1930 Op0Min, Op0Max); 1931 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 1932 Op1Min, Op1Max); 1933 } 1934 1935 // If Min and Max are known to be the same, then SimplifyDemandedBits 1936 // figured out that the LHS is a constant. Just constant fold this now so 1937 // that code below can assume that Min != Max. 1938 if (!isa<Constant>(Op0) && Op0Min == Op0Max) 1939 return new ICmpInst(I.getPredicate(), 1940 ConstantInt::get(I.getContext(), Op0Min), Op1); 1941 if (!isa<Constant>(Op1) && Op1Min == Op1Max) 1942 return new ICmpInst(I.getPredicate(), Op0, 1943 ConstantInt::get(I.getContext(), Op1Min)); 1944 1945 // Based on the range information we know about the LHS, see if we can 1946 // simplify this comparison. For example, (x&4) < 8 is always true. 1947 switch (I.getPredicate()) { 1948 default: llvm_unreachable("Unknown icmp opcode!"); 1949 case ICmpInst::ICMP_EQ: { 1950 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 1951 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 1952 1953 // If all bits are known zero except for one, then we know at most one 1954 // bit is set. If the comparison is against zero, then this is a check 1955 // to see if *that* bit is set. 1956 APInt Op0KnownZeroInverted = ~Op0KnownZero; 1957 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 1958 // If the LHS is an AND with the same constant, look through it. 1959 Value *LHS = 0; 1960 ConstantInt *LHSC = 0; 1961 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 1962 LHSC->getValue() != Op0KnownZeroInverted) 1963 LHS = Op0; 1964 1965 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 1966 // then turn "((1 << x)&8) == 0" into "x != 3". 1967 Value *X = 0; 1968 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 1969 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 1970 return new ICmpInst(ICmpInst::ICMP_NE, X, 1971 ConstantInt::get(X->getType(), CmpVal)); 1972 } 1973 1974 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 1975 // then turn "((8 >>u x)&1) == 0" into "x != 3". 1976 const APInt *CI; 1977 if (Op0KnownZeroInverted == 1 && 1978 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 1979 return new ICmpInst(ICmpInst::ICMP_NE, X, 1980 ConstantInt::get(X->getType(), 1981 CI->countTrailingZeros())); 1982 } 1983 1984 break; 1985 } 1986 case ICmpInst::ICMP_NE: { 1987 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 1988 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 1989 1990 // If all bits are known zero except for one, then we know at most one 1991 // bit is set. If the comparison is against zero, then this is a check 1992 // to see if *that* bit is set. 1993 APInt Op0KnownZeroInverted = ~Op0KnownZero; 1994 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 1995 // If the LHS is an AND with the same constant, look through it. 1996 Value *LHS = 0; 1997 ConstantInt *LHSC = 0; 1998 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 1999 LHSC->getValue() != Op0KnownZeroInverted) 2000 LHS = Op0; 2001 2002 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2003 // then turn "((1 << x)&8) != 0" into "x == 3". 2004 Value *X = 0; 2005 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2006 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2007 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2008 ConstantInt::get(X->getType(), CmpVal)); 2009 } 2010 2011 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2012 // then turn "((8 >>u x)&1) != 0" into "x == 3". 2013 const APInt *CI; 2014 if (Op0KnownZeroInverted == 1 && 2015 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2016 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2017 ConstantInt::get(X->getType(), 2018 CI->countTrailingZeros())); 2019 } 2020 2021 break; 2022 } 2023 case ICmpInst::ICMP_ULT: 2024 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B) 2025 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2026 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B) 2027 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2028 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B) 2029 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2030 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2031 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C 2032 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2033 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2034 2035 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear 2036 if (CI->isMinValue(true)) 2037 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 2038 Constant::getAllOnesValue(Op0->getType())); 2039 } 2040 break; 2041 case ICmpInst::ICMP_UGT: 2042 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B) 2043 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2044 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B) 2045 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2046 2047 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B) 2048 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2049 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2050 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C 2051 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2052 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2053 2054 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set 2055 if (CI->isMaxValue(true)) 2056 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 2057 Constant::getNullValue(Op0->getType())); 2058 } 2059 break; 2060 case ICmpInst::ICMP_SLT: 2061 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C) 2062 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2063 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C) 2064 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2065 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B) 2066 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2067 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2068 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C 2069 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2070 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2071 } 2072 break; 2073 case ICmpInst::ICMP_SGT: 2074 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B) 2075 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2076 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B) 2077 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2078 2079 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B) 2080 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2081 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2082 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C 2083 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2084 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2085 } 2086 break; 2087 case ICmpInst::ICMP_SGE: 2088 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"); 2089 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B) 2090 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2091 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B) 2092 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2093 break; 2094 case ICmpInst::ICMP_SLE: 2095 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"); 2096 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B) 2097 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2098 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B) 2099 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2100 break; 2101 case ICmpInst::ICMP_UGE: 2102 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"); 2103 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B) 2104 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2105 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B) 2106 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2107 break; 2108 case ICmpInst::ICMP_ULE: 2109 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"); 2110 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B) 2111 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2112 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B) 2113 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2114 break; 2115 } 2116 2117 // Turn a signed comparison into an unsigned one if both operands 2118 // are known to have the same sign. 2119 if (I.isSigned() && 2120 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) || 2121 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative()))) 2122 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1); 2123 } 2124 2125 // Test if the ICmpInst instruction is used exclusively by a select as 2126 // part of a minimum or maximum operation. If so, refrain from doing 2127 // any other folding. This helps out other analyses which understand 2128 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 2129 // and CodeGen. And in this case, at least one of the comparison 2130 // operands has at least one user besides the compare (the select), 2131 // which would often largely negate the benefit of folding anyway. 2132 if (I.hasOneUse()) 2133 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin())) 2134 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) || 2135 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1)) 2136 return 0; 2137 2138 // See if we are doing a comparison between a constant and an instruction that 2139 // can be folded into the comparison. 2140 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2141 // Since the RHS is a ConstantInt (CI), if the left hand side is an 2142 // instruction, see if that instruction also has constants so that the 2143 // instruction can be folded into the icmp 2144 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2145 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI)) 2146 return Res; 2147 } 2148 2149 // Handle icmp with constant (but not simple integer constant) RHS 2150 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2151 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2152 switch (LHSI->getOpcode()) { 2153 case Instruction::GetElementPtr: 2154 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null 2155 if (RHSC->isNullValue() && 2156 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices()) 2157 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2158 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2159 break; 2160 case Instruction::PHI: 2161 // Only fold icmp into the PHI if the phi and icmp are in the same 2162 // block. If in the same block, we're encouraging jump threading. If 2163 // not, we are just pessimizing the code by making an i1 phi. 2164 if (LHSI->getParent() == I.getParent()) 2165 if (Instruction *NV = FoldOpIntoPhi(I)) 2166 return NV; 2167 break; 2168 case Instruction::Select: { 2169 // If either operand of the select is a constant, we can fold the 2170 // comparison into the select arms, which will cause one to be 2171 // constant folded and the select turned into a bitwise or. 2172 Value *Op1 = 0, *Op2 = 0; 2173 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) 2174 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2175 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) 2176 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2177 2178 // We only want to perform this transformation if it will not lead to 2179 // additional code. This is true if either both sides of the select 2180 // fold to a constant (in which case the icmp is replaced with a select 2181 // which will usually simplify) or this is the only user of the 2182 // select (in which case we are trading a select+icmp for a simpler 2183 // select+icmp). 2184 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) { 2185 if (!Op1) 2186 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1), 2187 RHSC, I.getName()); 2188 if (!Op2) 2189 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2), 2190 RHSC, I.getName()); 2191 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2192 } 2193 break; 2194 } 2195 case Instruction::IntToPtr: 2196 // icmp pred inttoptr(X), null -> icmp pred X, 0 2197 if (RHSC->isNullValue() && TD && 2198 TD->getIntPtrType(RHSC->getContext()) == 2199 LHSI->getOperand(0)->getType()) 2200 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2201 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2202 break; 2203 2204 case Instruction::Load: 2205 // Try to optimize things like "A[i] > 4" to index computations. 2206 if (GetElementPtrInst *GEP = 2207 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2208 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2209 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2210 !cast<LoadInst>(LHSI)->isVolatile()) 2211 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2212 return Res; 2213 } 2214 break; 2215 } 2216 } 2217 2218 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now. 2219 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0)) 2220 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I)) 2221 return NI; 2222 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) 2223 if (Instruction *NI = FoldGEPICmp(GEP, Op0, 2224 ICmpInst::getSwappedPredicate(I.getPredicate()), I)) 2225 return NI; 2226 2227 // Test to see if the operands of the icmp are casted versions of other 2228 // values. If the ptr->ptr cast can be stripped off both arguments, we do so 2229 // now. 2230 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) { 2231 if (Op0->getType()->isPointerTy() && 2232 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) { 2233 // We keep moving the cast from the left operand over to the right 2234 // operand, where it can often be eliminated completely. 2235 Op0 = CI->getOperand(0); 2236 2237 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast 2238 // so eliminate it as well. 2239 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1)) 2240 Op1 = CI2->getOperand(0); 2241 2242 // If Op1 is a constant, we can fold the cast into the constant. 2243 if (Op0->getType() != Op1->getType()) { 2244 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 2245 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType()); 2246 } else { 2247 // Otherwise, cast the RHS right before the icmp 2248 Op1 = Builder->CreateBitCast(Op1, Op0->getType()); 2249 } 2250 } 2251 return new ICmpInst(I.getPredicate(), Op0, Op1); 2252 } 2253 } 2254 2255 if (isa<CastInst>(Op0)) { 2256 // Handle the special case of: icmp (cast bool to X), <cst> 2257 // This comes up when you have code like 2258 // int X = A < B; 2259 // if (X) ... 2260 // For generality, we handle any zero-extension of any operand comparison 2261 // with a constant or another cast from the same type. 2262 if (isa<Constant>(Op1) || isa<CastInst>(Op1)) 2263 if (Instruction *R = visitICmpInstWithCastAndCast(I)) 2264 return R; 2265 } 2266 2267 // See if it's the same type of instruction on the left and right. 2268 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) { 2269 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) { 2270 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() && 2271 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) { 2272 switch (Op0I->getOpcode()) { 2273 default: break; 2274 case Instruction::Add: 2275 case Instruction::Sub: 2276 case Instruction::Xor: 2277 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b 2278 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0), 2279 Op1I->getOperand(0)); 2280 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b 2281 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) { 2282 if (CI->getValue().isSignBit()) { 2283 ICmpInst::Predicate Pred = I.isSigned() 2284 ? I.getUnsignedPredicate() 2285 : I.getSignedPredicate(); 2286 return new ICmpInst(Pred, Op0I->getOperand(0), 2287 Op1I->getOperand(0)); 2288 } 2289 2290 if (CI->getValue().isMaxSignedValue()) { 2291 ICmpInst::Predicate Pred = I.isSigned() 2292 ? I.getUnsignedPredicate() 2293 : I.getSignedPredicate(); 2294 Pred = I.getSwappedPredicate(Pred); 2295 return new ICmpInst(Pred, Op0I->getOperand(0), 2296 Op1I->getOperand(0)); 2297 } 2298 } 2299 break; 2300 case Instruction::Mul: 2301 if (!I.isEquality()) 2302 break; 2303 2304 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) { 2305 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask 2306 // Mask = -1 >> count-trailing-zeros(Cst). 2307 if (!CI->isZero() && !CI->isOne()) { 2308 const APInt &AP = CI->getValue(); 2309 ConstantInt *Mask = ConstantInt::get(I.getContext(), 2310 APInt::getLowBitsSet(AP.getBitWidth(), 2311 AP.getBitWidth() - 2312 AP.countTrailingZeros())); 2313 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask); 2314 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask); 2315 return new ICmpInst(I.getPredicate(), And1, And2); 2316 } 2317 } 2318 break; 2319 } 2320 } 2321 } 2322 } 2323 2324 { Value *A, *B; 2325 // ~x < ~y --> y < x 2326 // ~x < cst --> ~cst < x 2327 if (match(Op0, m_Not(m_Value(A)))) { 2328 if (match(Op1, m_Not(m_Value(B)))) 2329 return new ICmpInst(I.getPredicate(), B, A); 2330 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1)) 2331 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A); 2332 } 2333 2334 // (a+b) <u a --> llvm.uadd.with.overflow. 2335 // (a+b) <u b --> llvm.uadd.with.overflow. 2336 if (I.getPredicate() == ICmpInst::ICMP_ULT && 2337 match(Op0, m_Add(m_Value(A), m_Value(B))) && 2338 (Op1 == A || Op1 == B)) 2339 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this)) 2340 return R; 2341 2342 // a >u (a+b) --> llvm.uadd.with.overflow. 2343 // b >u (a+b) --> llvm.uadd.with.overflow. 2344 if (I.getPredicate() == ICmpInst::ICMP_UGT && 2345 match(Op1, m_Add(m_Value(A), m_Value(B))) && 2346 (Op0 == A || Op0 == B)) 2347 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this)) 2348 return R; 2349 } 2350 2351 if (I.isEquality()) { 2352 Value *A, *B, *C, *D; 2353 2354 // -x == -y --> x == y 2355 if (match(Op0, m_Neg(m_Value(A))) && 2356 match(Op1, m_Neg(m_Value(B)))) 2357 return new ICmpInst(I.getPredicate(), A, B); 2358 2359 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 2360 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0 2361 Value *OtherVal = A == Op1 ? B : A; 2362 return new ICmpInst(I.getPredicate(), OtherVal, 2363 Constant::getNullValue(A->getType())); 2364 } 2365 2366 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) { 2367 // A^c1 == C^c2 --> A == C^(c1^c2) 2368 ConstantInt *C1, *C2; 2369 if (match(B, m_ConstantInt(C1)) && 2370 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) { 2371 Constant *NC = ConstantInt::get(I.getContext(), 2372 C1->getValue() ^ C2->getValue()); 2373 Value *Xor = Builder->CreateXor(C, NC, "tmp"); 2374 return new ICmpInst(I.getPredicate(), A, Xor); 2375 } 2376 2377 // A^B == A^D -> B == D 2378 if (A == C) return new ICmpInst(I.getPredicate(), B, D); 2379 if (A == D) return new ICmpInst(I.getPredicate(), B, C); 2380 if (B == C) return new ICmpInst(I.getPredicate(), A, D); 2381 if (B == D) return new ICmpInst(I.getPredicate(), A, C); 2382 } 2383 } 2384 2385 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && 2386 (A == Op0 || B == Op0)) { 2387 // A == (A^B) -> B == 0 2388 Value *OtherVal = A == Op0 ? B : A; 2389 return new ICmpInst(I.getPredicate(), OtherVal, 2390 Constant::getNullValue(A->getType())); 2391 } 2392 2393 // (A-B) == A -> B == 0 2394 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B)))) 2395 return new ICmpInst(I.getPredicate(), B, 2396 Constant::getNullValue(B->getType())); 2397 2398 // A == (A-B) -> B == 0 2399 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B)))) 2400 return new ICmpInst(I.getPredicate(), B, 2401 Constant::getNullValue(B->getType())); 2402 2403 // (A+B) == A -> B == 0 2404 if (match(Op0, m_Add(m_Specific(Op1), m_Value(B))) || 2405 match(Op0, m_Add(m_Value(B), m_Specific(Op1)))) 2406 return new ICmpInst(I.getPredicate(), B, 2407 Constant::getNullValue(B->getType())); 2408 2409 // A == (A+B) -> B == 0 2410 if (match(Op1, m_Add(m_Specific(Op0), m_Value(B))) || 2411 match(Op1, m_Add(m_Value(B), m_Specific(Op0)))) 2412 return new ICmpInst(I.getPredicate(), B, 2413 Constant::getNullValue(B->getType())); 2414 2415 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0 2416 if (Op0->hasOneUse() && Op1->hasOneUse() && 2417 match(Op0, m_And(m_Value(A), m_Value(B))) && 2418 match(Op1, m_And(m_Value(C), m_Value(D)))) { 2419 Value *X = 0, *Y = 0, *Z = 0; 2420 2421 if (A == C) { 2422 X = B; Y = D; Z = A; 2423 } else if (A == D) { 2424 X = B; Y = C; Z = A; 2425 } else if (B == C) { 2426 X = A; Y = D; Z = B; 2427 } else if (B == D) { 2428 X = A; Y = C; Z = B; 2429 } 2430 2431 if (X) { // Build (X^Y) & Z 2432 Op1 = Builder->CreateXor(X, Y, "tmp"); 2433 Op1 = Builder->CreateAnd(Op1, Z, "tmp"); 2434 I.setOperand(0, Op1); 2435 I.setOperand(1, Constant::getNullValue(Op1->getType())); 2436 return &I; 2437 } 2438 } 2439 } 2440 2441 { 2442 Value *X; ConstantInt *Cst; 2443 // icmp X+Cst, X 2444 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X) 2445 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0); 2446 2447 // icmp X, X+Cst 2448 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X) 2449 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1); 2450 } 2451 return Changed ? &I : 0; 2452} 2453 2454 2455 2456 2457 2458 2459/// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible. 2460/// 2461Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I, 2462 Instruction *LHSI, 2463 Constant *RHSC) { 2464 if (!isa<ConstantFP>(RHSC)) return 0; 2465 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF(); 2466 2467 // Get the width of the mantissa. We don't want to hack on conversions that 2468 // might lose information from the integer, e.g. "i64 -> float" 2469 int MantissaWidth = LHSI->getType()->getFPMantissaWidth(); 2470 if (MantissaWidth == -1) return 0; // Unknown. 2471 2472 // Check to see that the input is converted from an integer type that is small 2473 // enough that preserves all bits. TODO: check here for "known" sign bits. 2474 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e. 2475 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits(); 2476 2477 // If this is a uitofp instruction, we need an extra bit to hold the sign. 2478 bool LHSUnsigned = isa<UIToFPInst>(LHSI); 2479 if (LHSUnsigned) 2480 ++InputSize; 2481 2482 // If the conversion would lose info, don't hack on this. 2483 if ((int)InputSize > MantissaWidth) 2484 return 0; 2485 2486 // Otherwise, we can potentially simplify the comparison. We know that it 2487 // will always come through as an integer value and we know the constant is 2488 // not a NAN (it would have been previously simplified). 2489 assert(!RHS.isNaN() && "NaN comparison not already folded!"); 2490 2491 ICmpInst::Predicate Pred; 2492 switch (I.getPredicate()) { 2493 default: llvm_unreachable("Unexpected predicate!"); 2494 case FCmpInst::FCMP_UEQ: 2495 case FCmpInst::FCMP_OEQ: 2496 Pred = ICmpInst::ICMP_EQ; 2497 break; 2498 case FCmpInst::FCMP_UGT: 2499 case FCmpInst::FCMP_OGT: 2500 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT; 2501 break; 2502 case FCmpInst::FCMP_UGE: 2503 case FCmpInst::FCMP_OGE: 2504 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE; 2505 break; 2506 case FCmpInst::FCMP_ULT: 2507 case FCmpInst::FCMP_OLT: 2508 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT; 2509 break; 2510 case FCmpInst::FCMP_ULE: 2511 case FCmpInst::FCMP_OLE: 2512 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE; 2513 break; 2514 case FCmpInst::FCMP_UNE: 2515 case FCmpInst::FCMP_ONE: 2516 Pred = ICmpInst::ICMP_NE; 2517 break; 2518 case FCmpInst::FCMP_ORD: 2519 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2520 case FCmpInst::FCMP_UNO: 2521 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2522 } 2523 2524 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType()); 2525 2526 // Now we know that the APFloat is a normal number, zero or inf. 2527 2528 // See if the FP constant is too large for the integer. For example, 2529 // comparing an i8 to 300.0. 2530 unsigned IntWidth = IntTy->getScalarSizeInBits(); 2531 2532 if (!LHSUnsigned) { 2533 // If the RHS value is > SignedMax, fold the comparison. This handles +INF 2534 // and large values. 2535 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false); 2536 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true, 2537 APFloat::rmNearestTiesToEven); 2538 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0 2539 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT || 2540 Pred == ICmpInst::ICMP_SLE) 2541 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2542 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2543 } 2544 } else { 2545 // If the RHS value is > UnsignedMax, fold the comparison. This handles 2546 // +INF and large values. 2547 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false); 2548 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false, 2549 APFloat::rmNearestTiesToEven); 2550 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0 2551 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT || 2552 Pred == ICmpInst::ICMP_ULE) 2553 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2554 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2555 } 2556 } 2557 2558 if (!LHSUnsigned) { 2559 // See if the RHS value is < SignedMin. 2560 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false); 2561 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true, 2562 APFloat::rmNearestTiesToEven); 2563 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0 2564 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT || 2565 Pred == ICmpInst::ICMP_SGE) 2566 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2567 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2568 } 2569 } 2570 2571 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or 2572 // [0, UMAX], but it may still be fractional. See if it is fractional by 2573 // casting the FP value to the integer value and back, checking for equality. 2574 // Don't do this for zero, because -0.0 is not fractional. 2575 Constant *RHSInt = LHSUnsigned 2576 ? ConstantExpr::getFPToUI(RHSC, IntTy) 2577 : ConstantExpr::getFPToSI(RHSC, IntTy); 2578 if (!RHS.isZero()) { 2579 bool Equal = LHSUnsigned 2580 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC 2581 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC; 2582 if (!Equal) { 2583 // If we had a comparison against a fractional value, we have to adjust 2584 // the compare predicate and sometimes the value. RHSC is rounded towards 2585 // zero at this point. 2586 switch (Pred) { 2587 default: llvm_unreachable("Unexpected integer comparison!"); 2588 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true 2589 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2590 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false 2591 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2592 case ICmpInst::ICMP_ULE: 2593 // (float)int <= 4.4 --> int <= 4 2594 // (float)int <= -4.4 --> false 2595 if (RHS.isNegative()) 2596 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2597 break; 2598 case ICmpInst::ICMP_SLE: 2599 // (float)int <= 4.4 --> int <= 4 2600 // (float)int <= -4.4 --> int < -4 2601 if (RHS.isNegative()) 2602 Pred = ICmpInst::ICMP_SLT; 2603 break; 2604 case ICmpInst::ICMP_ULT: 2605 // (float)int < -4.4 --> false 2606 // (float)int < 4.4 --> int <= 4 2607 if (RHS.isNegative()) 2608 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2609 Pred = ICmpInst::ICMP_ULE; 2610 break; 2611 case ICmpInst::ICMP_SLT: 2612 // (float)int < -4.4 --> int < -4 2613 // (float)int < 4.4 --> int <= 4 2614 if (!RHS.isNegative()) 2615 Pred = ICmpInst::ICMP_SLE; 2616 break; 2617 case ICmpInst::ICMP_UGT: 2618 // (float)int > 4.4 --> int > 4 2619 // (float)int > -4.4 --> true 2620 if (RHS.isNegative()) 2621 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2622 break; 2623 case ICmpInst::ICMP_SGT: 2624 // (float)int > 4.4 --> int > 4 2625 // (float)int > -4.4 --> int >= -4 2626 if (RHS.isNegative()) 2627 Pred = ICmpInst::ICMP_SGE; 2628 break; 2629 case ICmpInst::ICMP_UGE: 2630 // (float)int >= -4.4 --> true 2631 // (float)int >= 4.4 --> int > 4 2632 if (!RHS.isNegative()) 2633 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2634 Pred = ICmpInst::ICMP_UGT; 2635 break; 2636 case ICmpInst::ICMP_SGE: 2637 // (float)int >= -4.4 --> int >= -4 2638 // (float)int >= 4.4 --> int > 4 2639 if (!RHS.isNegative()) 2640 Pred = ICmpInst::ICMP_SGT; 2641 break; 2642 } 2643 } 2644 } 2645 2646 // Lower this FP comparison into an appropriate integer version of the 2647 // comparison. 2648 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt); 2649} 2650 2651Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { 2652 bool Changed = false; 2653 2654 /// Orders the operands of the compare so that they are listed from most 2655 /// complex to least complex. This puts constants before unary operators, 2656 /// before binary operators. 2657 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) { 2658 I.swapOperands(); 2659 Changed = true; 2660 } 2661 2662 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2663 2664 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD)) 2665 return ReplaceInstUsesWith(I, V); 2666 2667 // Simplify 'fcmp pred X, X' 2668 if (Op0 == Op1) { 2669 switch (I.getPredicate()) { 2670 default: llvm_unreachable("Unknown predicate!"); 2671 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y) 2672 case FCmpInst::FCMP_ULT: // True if unordered or less than 2673 case FCmpInst::FCMP_UGT: // True if unordered or greater than 2674 case FCmpInst::FCMP_UNE: // True if unordered or not equal 2675 // Canonicalize these to be 'fcmp uno %X, 0.0'. 2676 I.setPredicate(FCmpInst::FCMP_UNO); 2677 I.setOperand(1, Constant::getNullValue(Op0->getType())); 2678 return &I; 2679 2680 case FCmpInst::FCMP_ORD: // True if ordered (no nans) 2681 case FCmpInst::FCMP_OEQ: // True if ordered and equal 2682 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal 2683 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal 2684 // Canonicalize these to be 'fcmp ord %X, 0.0'. 2685 I.setPredicate(FCmpInst::FCMP_ORD); 2686 I.setOperand(1, Constant::getNullValue(Op0->getType())); 2687 return &I; 2688 } 2689 } 2690 2691 // Handle fcmp with constant RHS 2692 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2693 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2694 switch (LHSI->getOpcode()) { 2695 case Instruction::PHI: 2696 // Only fold fcmp into the PHI if the phi and fcmp are in the same 2697 // block. If in the same block, we're encouraging jump threading. If 2698 // not, we are just pessimizing the code by making an i1 phi. 2699 if (LHSI->getParent() == I.getParent()) 2700 if (Instruction *NV = FoldOpIntoPhi(I)) 2701 return NV; 2702 break; 2703 case Instruction::SIToFP: 2704 case Instruction::UIToFP: 2705 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC)) 2706 return NV; 2707 break; 2708 case Instruction::Select: { 2709 // If either operand of the select is a constant, we can fold the 2710 // comparison into the select arms, which will cause one to be 2711 // constant folded and the select turned into a bitwise or. 2712 Value *Op1 = 0, *Op2 = 0; 2713 if (LHSI->hasOneUse()) { 2714 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) { 2715 // Fold the known value into the constant operand. 2716 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 2717 // Insert a new FCmp of the other select operand. 2718 Op2 = Builder->CreateFCmp(I.getPredicate(), 2719 LHSI->getOperand(2), RHSC, I.getName()); 2720 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) { 2721 // Fold the known value into the constant operand. 2722 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 2723 // Insert a new FCmp of the other select operand. 2724 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1), 2725 RHSC, I.getName()); 2726 } 2727 } 2728 2729 if (Op1) 2730 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2731 break; 2732 } 2733 case Instruction::Load: 2734 if (GetElementPtrInst *GEP = 2735 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2736 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2737 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2738 !cast<LoadInst>(LHSI)->isVolatile()) 2739 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2740 return Res; 2741 } 2742 break; 2743 } 2744 } 2745 2746 return Changed ? &I : 0; 2747} 2748