1 //===- InstCombineCompares.cpp --------------------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the visitICmp and visitFCmp functions. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "InstCombine.h" 15 #include "llvm/Analysis/ConstantFolding.h" 16 #include "llvm/Analysis/InstructionSimplify.h" 17 #include "llvm/Analysis/MemoryBuiltins.h" 18 #include "llvm/IR/DataLayout.h" 19 #include "llvm/IR/IntrinsicInst.h" 20 #include "llvm/Support/ConstantRange.h" 21 #include "llvm/Support/GetElementPtrTypeIterator.h" 22 #include "llvm/Support/PatternMatch.h" 23 #include "llvm/Target/TargetLibraryInfo.h" 24 using namespace llvm; 25 using namespace PatternMatch; 26 27 static ConstantInt *getOne(Constant *C) { 28 return ConstantInt::get(cast<IntegerType>(C->getType()), 1); 29 } 30 31 /// AddOne - Add one to a ConstantInt 32 static Constant *AddOne(Constant *C) { 33 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 34 } 35 /// SubOne - Subtract one from a ConstantInt 36 static Constant *SubOne(Constant *C) { 37 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1)); 38 } 39 40 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) { 41 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx)); 42 } 43 44 static bool HasAddOverflow(ConstantInt *Result, 45 ConstantInt *In1, ConstantInt *In2, 46 bool IsSigned) { 47 if (!IsSigned) 48 return Result->getValue().ult(In1->getValue()); 49 50 if (In2->isNegative()) 51 return Result->getValue().sgt(In1->getValue()); 52 return Result->getValue().slt(In1->getValue()); 53 } 54 55 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result 56 /// overflowed for this type. 57 static bool AddWithOverflow(Constant *&Result, Constant *In1, 58 Constant *In2, bool IsSigned = false) { 59 Result = ConstantExpr::getAdd(In1, In2); 60 61 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 62 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 63 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 64 if (HasAddOverflow(ExtractElement(Result, Idx), 65 ExtractElement(In1, Idx), 66 ExtractElement(In2, Idx), 67 IsSigned)) 68 return true; 69 } 70 return false; 71 } 72 73 return HasAddOverflow(cast<ConstantInt>(Result), 74 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 75 IsSigned); 76 } 77 78 static bool HasSubOverflow(ConstantInt *Result, 79 ConstantInt *In1, ConstantInt *In2, 80 bool IsSigned) { 81 if (!IsSigned) 82 return Result->getValue().ugt(In1->getValue()); 83 84 if (In2->isNegative()) 85 return Result->getValue().slt(In1->getValue()); 86 87 return Result->getValue().sgt(In1->getValue()); 88 } 89 90 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result 91 /// overflowed for this type. 92 static bool SubWithOverflow(Constant *&Result, Constant *In1, 93 Constant *In2, bool IsSigned = false) { 94 Result = ConstantExpr::getSub(In1, In2); 95 96 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 97 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 98 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 99 if (HasSubOverflow(ExtractElement(Result, Idx), 100 ExtractElement(In1, Idx), 101 ExtractElement(In2, Idx), 102 IsSigned)) 103 return true; 104 } 105 return false; 106 } 107 108 return HasSubOverflow(cast<ConstantInt>(Result), 109 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 110 IsSigned); 111 } 112 113 /// isSignBitCheck - Given an exploded icmp instruction, return true if the 114 /// comparison only checks the sign bit. If it only checks the sign bit, set 115 /// TrueIfSigned if the result of the comparison is true when the input value is 116 /// signed. 117 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS, 118 bool &TrueIfSigned) { 119 switch (pred) { 120 case ICmpInst::ICMP_SLT: // True if LHS s< 0 121 TrueIfSigned = true; 122 return RHS->isZero(); 123 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1 124 TrueIfSigned = true; 125 return RHS->isAllOnesValue(); 126 case ICmpInst::ICMP_SGT: // True if LHS s> -1 127 TrueIfSigned = false; 128 return RHS->isAllOnesValue(); 129 case ICmpInst::ICMP_UGT: 130 // True if LHS u> RHS and RHS == high-bit-mask - 1 131 TrueIfSigned = true; 132 return RHS->isMaxValue(true); 133 case ICmpInst::ICMP_UGE: 134 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc) 135 TrueIfSigned = true; 136 return RHS->getValue().isSignBit(); 137 default: 138 return false; 139 } 140 } 141 142 // isHighOnes - Return true if the constant is of the form 1+0+. 143 // This is the same as lowones(~X). 144 static 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. 152 static 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. 176 static 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". 201 Instruction *InstCombiner:: 202 FoldCmpLoadFromIndexedGlobal(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 Constant *Init = GV->getInitializer(); 208 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init)) 209 return 0; 210 211 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements(); 212 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays. 213 214 // There are many forms of this optimization we can handle, for now, just do 215 // the simple index into a single-dimensional array. 216 // 217 // Require: GEP GV, 0, i {{, constant indices}} 218 if (GEP->getNumOperands() < 3 || 219 !isa<ConstantInt>(GEP->getOperand(1)) || 220 !cast<ConstantInt>(GEP->getOperand(1))->isZero() || 221 isa<Constant>(GEP->getOperand(2))) 222 return 0; 223 224 // Check that indices after the variable are constants and in-range for the 225 // type they index. Collect the indices. This is typically for arrays of 226 // structs. 227 SmallVector<unsigned, 4> LaterIndices; 228 229 Type *EltTy = Init->getType()->getArrayElementType(); 230 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { 231 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i)); 232 if (Idx == 0) return 0; // Variable index. 233 234 uint64_t IdxVal = Idx->getZExtValue(); 235 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index. 236 237 if (StructType *STy = dyn_cast<StructType>(EltTy)) 238 EltTy = STy->getElementType(IdxVal); 239 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) { 240 if (IdxVal >= ATy->getNumElements()) return 0; 241 EltTy = ATy->getElementType(); 242 } else { 243 return 0; // Unknown type. 244 } 245 246 LaterIndices.push_back(IdxVal); 247 } 248 249 enum { Overdefined = -3, Undefined = -2 }; 250 251 // Variables for our state machines. 252 253 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form 254 // "i == 47 | i == 87", where 47 is the first index the condition is true for, 255 // and 87 is the second (and last) index. FirstTrueElement is -2 when 256 // undefined, otherwise set to the first true element. SecondTrueElement is 257 // -2 when undefined, -3 when overdefined and >= 0 when that index is true. 258 int FirstTrueElement = Undefined, SecondTrueElement = Undefined; 259 260 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the 261 // form "i != 47 & i != 87". Same state transitions as for true elements. 262 int FirstFalseElement = Undefined, SecondFalseElement = Undefined; 263 264 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these 265 /// define a state machine that triggers for ranges of values that the index 266 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. 267 /// This is -2 when undefined, -3 when overdefined, and otherwise the last 268 /// index in the range (inclusive). We use -2 for undefined here because we 269 /// use relative comparisons and don't want 0-1 to match -1. 270 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; 271 272 // MagicBitvector - This is a magic bitvector where we set a bit if the 273 // comparison is true for element 'i'. If there are 64 elements or less in 274 // the array, this will fully represent all the comparison results. 275 uint64_t MagicBitvector = 0; 276 277 278 // Scan the array and see if one of our patterns matches. 279 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1)); 280 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) { 281 Constant *Elt = Init->getAggregateElement(i); 282 if (Elt == 0) return 0; 283 284 // If this is indexing an array of structures, get the structure element. 285 if (!LaterIndices.empty()) 286 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices); 287 288 // If the element is masked, handle it. 289 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst); 290 291 // Find out if the comparison would be true or false for the i'th element. 292 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt, 293 CompareRHS, TD, TLI); 294 // If the result is undef for this element, ignore it. 295 if (isa<UndefValue>(C)) { 296 // Extend range state machines to cover this element in case there is an 297 // undef in the middle of the range. 298 if (TrueRangeEnd == (int)i-1) 299 TrueRangeEnd = i; 300 if (FalseRangeEnd == (int)i-1) 301 FalseRangeEnd = i; 302 continue; 303 } 304 305 // If we can't compute the result for any of the elements, we have to give 306 // up evaluating the entire conditional. 307 if (!isa<ConstantInt>(C)) return 0; 308 309 // Otherwise, we know if the comparison is true or false for this element, 310 // update our state machines. 311 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero(); 312 313 // State machine for single/double/range index comparison. 314 if (IsTrueForElt) { 315 // Update the TrueElement state machine. 316 if (FirstTrueElement == Undefined) 317 FirstTrueElement = TrueRangeEnd = i; // First true element. 318 else { 319 // Update double-compare state machine. 320 if (SecondTrueElement == Undefined) 321 SecondTrueElement = i; 322 else 323 SecondTrueElement = Overdefined; 324 325 // Update range state machine. 326 if (TrueRangeEnd == (int)i-1) 327 TrueRangeEnd = i; 328 else 329 TrueRangeEnd = Overdefined; 330 } 331 } else { 332 // Update the FalseElement state machine. 333 if (FirstFalseElement == Undefined) 334 FirstFalseElement = FalseRangeEnd = i; // First false element. 335 else { 336 // Update double-compare state machine. 337 if (SecondFalseElement == Undefined) 338 SecondFalseElement = i; 339 else 340 SecondFalseElement = Overdefined; 341 342 // Update range state machine. 343 if (FalseRangeEnd == (int)i-1) 344 FalseRangeEnd = i; 345 else 346 FalseRangeEnd = Overdefined; 347 } 348 } 349 350 351 // If this element is in range, update our magic bitvector. 352 if (i < 64 && IsTrueForElt) 353 MagicBitvector |= 1ULL << i; 354 355 // If all of our states become overdefined, bail out early. Since the 356 // predicate is expensive, only check it every 8 elements. This is only 357 // really useful for really huge arrays. 358 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && 359 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && 360 FalseRangeEnd == Overdefined) 361 return 0; 362 } 363 364 // Now that we've scanned the entire array, emit our new comparison(s). We 365 // order the state machines in complexity of the generated code. 366 Value *Idx = GEP->getOperand(2); 367 368 // If the index is larger than the pointer size of the target, truncate the 369 // index down like the GEP would do implicitly. We don't have to do this for 370 // an inbounds GEP because the index can't be out of range. 371 if (!GEP->isInBounds() && 372 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits()) 373 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext())); 374 375 // If the comparison is only true for one or two elements, emit direct 376 // comparisons. 377 if (SecondTrueElement != Overdefined) { 378 // None true -> false. 379 if (FirstTrueElement == Undefined) 380 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext())); 381 382 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement); 383 384 // True for one element -> 'i == 47'. 385 if (SecondTrueElement == Undefined) 386 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); 387 388 // True for two elements -> 'i == 47 | i == 72'. 389 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx); 390 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement); 391 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx); 392 return BinaryOperator::CreateOr(C1, C2); 393 } 394 395 // If the comparison is only false for one or two elements, emit direct 396 // comparisons. 397 if (SecondFalseElement != Overdefined) { 398 // None false -> true. 399 if (FirstFalseElement == Undefined) 400 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext())); 401 402 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement); 403 404 // False for one element -> 'i != 47'. 405 if (SecondFalseElement == Undefined) 406 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); 407 408 // False for two elements -> 'i != 47 & i != 72'. 409 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx); 410 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement); 411 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx); 412 return BinaryOperator::CreateAnd(C1, C2); 413 } 414 415 // If the comparison can be replaced with a range comparison for the elements 416 // where it is true, emit the range check. 417 if (TrueRangeEnd != Overdefined) { 418 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare"); 419 420 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1). 421 if (FirstTrueElement) { 422 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement); 423 Idx = Builder->CreateAdd(Idx, Offs); 424 } 425 426 Value *End = ConstantInt::get(Idx->getType(), 427 TrueRangeEnd-FirstTrueElement+1); 428 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); 429 } 430 431 // False range check. 432 if (FalseRangeEnd != Overdefined) { 433 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare"); 434 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). 435 if (FirstFalseElement) { 436 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement); 437 Idx = Builder->CreateAdd(Idx, Offs); 438 } 439 440 Value *End = ConstantInt::get(Idx->getType(), 441 FalseRangeEnd-FirstFalseElement); 442 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); 443 } 444 445 446 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state 447 // of this load, replace it with computation that does: 448 // ((magic_cst >> i) & 1) != 0 449 if (ArrayElementCount <= 32 || 450 (TD && ArrayElementCount <= 64 && TD->isLegalInteger(64))) { 451 Type *Ty; 452 if (ArrayElementCount <= 32) 453 Ty = Type::getInt32Ty(Init->getContext()); 454 else 455 Ty = Type::getInt64Ty(Init->getContext()); 456 Value *V = Builder->CreateIntCast(Idx, Ty, false); 457 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V); 458 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V); 459 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0)); 460 } 461 462 return 0; 463 } 464 465 466 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare 467 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we 468 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can 469 /// be complex, and scales are involved. The above expression would also be 470 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). 471 /// This later form is less amenable to optimization though, and we are allowed 472 /// to generate the first by knowing that pointer arithmetic doesn't overflow. 473 /// 474 /// If we can't emit an optimized form for this expression, this returns null. 475 /// 476 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) { 477 DataLayout &TD = *IC.getDataLayout(); 478 gep_type_iterator GTI = gep_type_begin(GEP); 479 480 // Check to see if this gep only has a single variable index. If so, and if 481 // any constant indices are a multiple of its scale, then we can compute this 482 // in terms of the scale of the variable index. For example, if the GEP 483 // implies an offset of "12 + i*4", then we can codegen this as "3 + i", 484 // because the expression will cross zero at the same point. 485 unsigned i, e = GEP->getNumOperands(); 486 int64_t Offset = 0; 487 for (i = 1; i != e; ++i, ++GTI) { 488 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 489 // Compute the aggregate offset of constant indices. 490 if (CI->isZero()) continue; 491 492 // Handle a struct index, which adds its field offset to the pointer. 493 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 494 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 495 } else { 496 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 497 Offset += Size*CI->getSExtValue(); 498 } 499 } else { 500 // Found our variable index. 501 break; 502 } 503 } 504 505 // If there are no variable indices, we must have a constant offset, just 506 // evaluate it the general way. 507 if (i == e) return 0; 508 509 Value *VariableIdx = GEP->getOperand(i); 510 // Determine the scale factor of the variable element. For example, this is 511 // 4 if the variable index is into an array of i32. 512 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType()); 513 514 // Verify that there are no other variable indices. If so, emit the hard way. 515 for (++i, ++GTI; i != e; ++i, ++GTI) { 516 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i)); 517 if (!CI) return 0; 518 519 // Compute the aggregate offset of constant indices. 520 if (CI->isZero()) continue; 521 522 // Handle a struct index, which adds its field offset to the pointer. 523 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 524 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 525 } else { 526 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 527 Offset += Size*CI->getSExtValue(); 528 } 529 } 530 531 // Okay, we know we have a single variable index, which must be a 532 // pointer/array/vector index. If there is no offset, life is simple, return 533 // the index. 534 unsigned IntPtrWidth = TD.getPointerSizeInBits(); 535 if (Offset == 0) { 536 // Cast to intptrty in case a truncation occurs. If an extension is needed, 537 // we don't need to bother extending: the extension won't affect where the 538 // computation crosses zero. 539 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) { 540 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext()); 541 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy); 542 } 543 return VariableIdx; 544 } 545 546 // Otherwise, there is an index. The computation we will do will be modulo 547 // the pointer size, so get it. 548 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth); 549 550 Offset &= PtrSizeMask; 551 VariableScale &= PtrSizeMask; 552 553 // To do this transformation, any constant index must be a multiple of the 554 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i", 555 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a 556 // multiple of the variable scale. 557 int64_t NewOffs = Offset / (int64_t)VariableScale; 558 if (Offset != NewOffs*(int64_t)VariableScale) 559 return 0; 560 561 // Okay, we can do this evaluation. Start by converting the index to intptr. 562 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext()); 563 if (VariableIdx->getType() != IntPtrTy) 564 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy, 565 true /*Signed*/); 566 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs); 567 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset"); 568 } 569 570 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something 571 /// else. At this point we know that the GEP is on the LHS of the comparison. 572 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 573 ICmpInst::Predicate Cond, 574 Instruction &I) { 575 // Don't transform signed compares of GEPs into index compares. Even if the 576 // GEP is inbounds, the final add of the base pointer can have signed overflow 577 // and would change the result of the icmp. 578 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be 579 // the maximum signed value for the pointer type. 580 if (ICmpInst::isSigned(Cond)) 581 return 0; 582 583 // Look through bitcasts. 584 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS)) 585 RHS = BCI->getOperand(0); 586 587 Value *PtrBase = GEPLHS->getOperand(0); 588 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) { 589 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). 590 // This transformation (ignoring the base and scales) is valid because we 591 // know pointers can't overflow since the gep is inbounds. See if we can 592 // output an optimized form. 593 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this); 594 595 // If not, synthesize the offset the hard way. 596 if (Offset == 0) 597 Offset = EmitGEPOffset(GEPLHS); 598 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset, 599 Constant::getNullValue(Offset->getType())); 600 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) { 601 // If the base pointers are different, but the indices are the same, just 602 // compare the base pointer. 603 if (PtrBase != GEPRHS->getOperand(0)) { 604 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands(); 605 IndicesTheSame &= GEPLHS->getOperand(0)->getType() == 606 GEPRHS->getOperand(0)->getType(); 607 if (IndicesTheSame) 608 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 609 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 610 IndicesTheSame = false; 611 break; 612 } 613 614 // If all indices are the same, just compare the base pointers. 615 if (IndicesTheSame) 616 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), 617 GEPLHS->getOperand(0), GEPRHS->getOperand(0)); 618 619 // If we're comparing GEPs with two base pointers that only differ in type 620 // and both GEPs have only constant indices or just one use, then fold 621 // the compare with the adjusted indices. 622 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() && 623 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) && 624 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) && 625 PtrBase->stripPointerCasts() == 626 GEPRHS->getOperand(0)->stripPointerCasts()) { 627 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond), 628 EmitGEPOffset(GEPLHS), 629 EmitGEPOffset(GEPRHS)); 630 return ReplaceInstUsesWith(I, Cmp); 631 } 632 633 // Otherwise, the base pointers are different and the indices are 634 // different, bail out. 635 return 0; 636 } 637 638 // If one of the GEPs has all zero indices, recurse. 639 bool AllZeros = true; 640 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 641 if (!isa<Constant>(GEPLHS->getOperand(i)) || 642 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) { 643 AllZeros = false; 644 break; 645 } 646 if (AllZeros) 647 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0), 648 ICmpInst::getSwappedPredicate(Cond), I); 649 650 // If the other GEP has all zero indices, recurse. 651 AllZeros = true; 652 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 653 if (!isa<Constant>(GEPRHS->getOperand(i)) || 654 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) { 655 AllZeros = false; 656 break; 657 } 658 if (AllZeros) 659 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I); 660 661 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds(); 662 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) { 663 // If the GEPs only differ by one index, compare it. 664 unsigned NumDifferences = 0; // Keep track of # differences. 665 unsigned DiffOperand = 0; // The operand that differs. 666 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 667 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 668 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() != 669 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) { 670 // Irreconcilable differences. 671 NumDifferences = 2; 672 break; 673 } else { 674 if (NumDifferences++) break; 675 DiffOperand = i; 676 } 677 } 678 679 if (NumDifferences == 0) // SAME GEP? 680 return ReplaceInstUsesWith(I, // No comparison is needed here. 681 ConstantInt::get(Type::getInt1Ty(I.getContext()), 682 ICmpInst::isTrueWhenEqual(Cond))); 683 684 else if (NumDifferences == 1 && GEPsInBounds) { 685 Value *LHSV = GEPLHS->getOperand(DiffOperand); 686 Value *RHSV = GEPRHS->getOperand(DiffOperand); 687 // Make sure we do a signed comparison here. 688 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV); 689 } 690 } 691 692 // Only lower this if the icmp is the only user of the GEP or if we expect 693 // the result to fold to a constant! 694 if (TD && 695 GEPsInBounds && 696 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) && 697 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) { 698 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) 699 Value *L = EmitGEPOffset(GEPLHS); 700 Value *R = EmitGEPOffset(GEPRHS); 701 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R); 702 } 703 } 704 return 0; 705 } 706 707 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X". 708 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI, 709 Value *X, ConstantInt *CI, 710 ICmpInst::Predicate Pred, 711 Value *TheAdd) { 712 // If we have X+0, exit early (simplifying logic below) and let it get folded 713 // elsewhere. icmp X+0, X -> icmp X, X 714 if (CI->isZero()) { 715 bool isTrue = ICmpInst::isTrueWhenEqual(Pred); 716 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 717 } 718 719 // (X+4) == X -> false. 720 if (Pred == ICmpInst::ICMP_EQ) 721 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext())); 722 723 // (X+4) != X -> true. 724 if (Pred == ICmpInst::ICMP_NE) 725 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext())); 726 727 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, 728 // so the values can never be equal. Similarly for all other "or equals" 729 // operators. 730 731 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255 732 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253 733 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0 734 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 735 Value *R = 736 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI); 737 return new ICmpInst(ICmpInst::ICMP_UGT, X, R); 738 } 739 740 // (X+1) >u X --> X <u (0-1) --> X != 255 741 // (X+2) >u X --> X <u (0-2) --> X <u 254 742 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0 743 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) 744 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI)); 745 746 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits(); 747 ConstantInt *SMax = ConstantInt::get(X->getContext(), 748 APInt::getSignedMaxValue(BitWidth)); 749 750 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127 751 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125 752 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0 753 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1 754 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126 755 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127 756 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 757 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI)); 758 759 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127 760 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126 761 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1 762 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2 763 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126 764 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128 765 766 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); 767 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1); 768 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C)); 769 } 770 771 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS 772 /// and CmpRHS are both known to be integer constants. 773 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI, 774 ConstantInt *DivRHS) { 775 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1)); 776 const APInt &CmpRHSV = CmpRHS->getValue(); 777 778 // FIXME: If the operand types don't match the type of the divide 779 // then don't attempt this transform. The code below doesn't have the 780 // logic to deal with a signed divide and an unsigned compare (and 781 // vice versa). This is because (x /s C1) <s C2 produces different 782 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even 783 // (x /u C1) <u C2. Simply casting the operands and result won't 784 // work. :( The if statement below tests that condition and bails 785 // if it finds it. 786 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv; 787 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned()) 788 return 0; 789 if (DivRHS->isZero()) 790 return 0; // The ProdOV computation fails on divide by zero. 791 if (DivIsSigned && DivRHS->isAllOnesValue()) 792 return 0; // The overflow computation also screws up here 793 if (DivRHS->isOne()) { 794 // This eliminates some funny cases with INT_MIN. 795 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X. 796 return &ICI; 797 } 798 799 // Compute Prod = CI * DivRHS. We are essentially solving an equation 800 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and 801 // C2 (CI). By solving for X we can turn this into a range check 802 // instead of computing a divide. 803 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS); 804 805 // Determine if the product overflows by seeing if the product is 806 // not equal to the divide. Make sure we do the same kind of divide 807 // as in the LHS instruction that we're folding. 808 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) : 809 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS; 810 811 // Get the ICmp opcode 812 ICmpInst::Predicate Pred = ICI.getPredicate(); 813 814 /// If the division is known to be exact, then there is no remainder from the 815 /// divide, so the covered range size is unit, otherwise it is the divisor. 816 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS; 817 818 // Figure out the interval that is being checked. For example, a comparison 819 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). 820 // Compute this interval based on the constants involved and the signedness of 821 // the compare/divide. This computes a half-open interval, keeping track of 822 // whether either value in the interval overflows. After analysis each 823 // overflow variable is set to 0 if it's corresponding bound variable is valid 824 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. 825 int LoOverflow = 0, HiOverflow = 0; 826 Constant *LoBound = 0, *HiBound = 0; 827 828 if (!DivIsSigned) { // udiv 829 // e.g. X/5 op 3 --> [15, 20) 830 LoBound = Prod; 831 HiOverflow = LoOverflow = ProdOV; 832 if (!HiOverflow) { 833 // If this is not an exact divide, then many values in the range collapse 834 // to the same result value. 835 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false); 836 } 837 838 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0. 839 if (CmpRHSV == 0) { // (X / pos) op 0 840 // Can't overflow. e.g. X/2 op 0 --> [-1, 2) 841 LoBound = ConstantExpr::getNeg(SubOne(RangeSize)); 842 HiBound = RangeSize; 843 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos 844 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) 845 HiOverflow = LoOverflow = ProdOV; 846 if (!HiOverflow) 847 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true); 848 } else { // (X / pos) op neg 849 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) 850 HiBound = AddOne(Prod); 851 LoOverflow = HiOverflow = ProdOV ? -1 : 0; 852 if (!LoOverflow) { 853 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 854 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0; 855 } 856 } 857 } else if (DivRHS->isNegative()) { // Divisor is < 0. 858 if (DivI->isExact()) 859 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 860 if (CmpRHSV == 0) { // (X / neg) op 0 861 // e.g. X/-5 op 0 --> [-4, 5) 862 LoBound = AddOne(RangeSize); 863 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 864 if (HiBound == DivRHS) { // -INTMIN = INTMIN 865 HiOverflow = 1; // [INTMIN+1, overflow) 866 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN 867 } 868 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos 869 // e.g. X/-5 op 3 --> [-19, -14) 870 HiBound = AddOne(Prod); 871 HiOverflow = LoOverflow = ProdOV ? -1 : 0; 872 if (!LoOverflow) 873 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0; 874 } else { // (X / neg) op neg 875 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) 876 LoOverflow = HiOverflow = ProdOV; 877 if (!HiOverflow) 878 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true); 879 } 880 881 // Dividing by a negative swaps the condition. LT <-> GT 882 Pred = ICmpInst::getSwappedPredicate(Pred); 883 } 884 885 Value *X = DivI->getOperand(0); 886 switch (Pred) { 887 default: llvm_unreachable("Unhandled icmp opcode!"); 888 case ICmpInst::ICMP_EQ: 889 if (LoOverflow && HiOverflow) 890 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 891 if (HiOverflow) 892 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 893 ICmpInst::ICMP_UGE, X, LoBound); 894 if (LoOverflow) 895 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 896 ICmpInst::ICMP_ULT, X, HiBound); 897 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 898 DivIsSigned, true)); 899 case ICmpInst::ICMP_NE: 900 if (LoOverflow && HiOverflow) 901 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 902 if (HiOverflow) 903 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 904 ICmpInst::ICMP_ULT, X, LoBound); 905 if (LoOverflow) 906 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 907 ICmpInst::ICMP_UGE, X, HiBound); 908 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 909 DivIsSigned, false)); 910 case ICmpInst::ICMP_ULT: 911 case ICmpInst::ICMP_SLT: 912 if (LoOverflow == +1) // Low bound is greater than input range. 913 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 914 if (LoOverflow == -1) // Low bound is less than input range. 915 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 916 return new ICmpInst(Pred, X, LoBound); 917 case ICmpInst::ICMP_UGT: 918 case ICmpInst::ICMP_SGT: 919 if (HiOverflow == +1) // High bound greater than input range. 920 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 921 if (HiOverflow == -1) // High bound less than input range. 922 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 923 if (Pred == ICmpInst::ICMP_UGT) 924 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound); 925 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound); 926 } 927 } 928 929 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)". 930 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr, 931 ConstantInt *ShAmt) { 932 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue(); 933 934 // Check that the shift amount is in range. If not, don't perform 935 // undefined shifts. When the shift is visited it will be 936 // simplified. 937 uint32_t TypeBits = CmpRHSV.getBitWidth(); 938 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 939 if (ShAmtVal >= TypeBits || ShAmtVal == 0) 940 return 0; 941 942 if (!ICI.isEquality()) { 943 // If we have an unsigned comparison and an ashr, we can't simplify this. 944 // Similarly for signed comparisons with lshr. 945 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr)) 946 return 0; 947 948 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv 949 // by a power of 2. Since we already have logic to simplify these, 950 // transform to div and then simplify the resultant comparison. 951 if (Shr->getOpcode() == Instruction::AShr && 952 (!Shr->isExact() || ShAmtVal == TypeBits - 1)) 953 return 0; 954 955 // Revisit the shift (to delete it). 956 Worklist.Add(Shr); 957 958 Constant *DivCst = 959 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal)); 960 961 Value *Tmp = 962 Shr->getOpcode() == Instruction::AShr ? 963 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) : 964 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()); 965 966 ICI.setOperand(0, Tmp); 967 968 // If the builder folded the binop, just return it. 969 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp); 970 if (TheDiv == 0) 971 return &ICI; 972 973 // Otherwise, fold this div/compare. 974 assert(TheDiv->getOpcode() == Instruction::SDiv || 975 TheDiv->getOpcode() == Instruction::UDiv); 976 977 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst)); 978 assert(Res && "This div/cst should have folded!"); 979 return Res; 980 } 981 982 983 // If we are comparing against bits always shifted out, the 984 // comparison cannot succeed. 985 APInt Comp = CmpRHSV << ShAmtVal; 986 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp); 987 if (Shr->getOpcode() == Instruction::LShr) 988 Comp = Comp.lshr(ShAmtVal); 989 else 990 Comp = Comp.ashr(ShAmtVal); 991 992 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero. 993 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 994 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 995 IsICMP_NE); 996 return ReplaceInstUsesWith(ICI, Cst); 997 } 998 999 // Otherwise, check to see if the bits shifted out are known to be zero. 1000 // If so, we can compare against the unshifted value: 1001 // (X & 4) >> 1 == 2 --> (X & 4) == 4. 1002 if (Shr->hasOneUse() && Shr->isExact()) 1003 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS); 1004 1005 if (Shr->hasOneUse()) { 1006 // Otherwise strength reduce the shift into an and. 1007 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal)); 1008 Constant *Mask = ConstantInt::get(ICI.getContext(), Val); 1009 1010 Value *And = Builder->CreateAnd(Shr->getOperand(0), 1011 Mask, Shr->getName()+".mask"); 1012 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS); 1013 } 1014 return 0; 1015 } 1016 1017 1018 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)". 1019 /// 1020 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI, 1021 Instruction *LHSI, 1022 ConstantInt *RHS) { 1023 const APInt &RHSV = RHS->getValue(); 1024 1025 switch (LHSI->getOpcode()) { 1026 case Instruction::Trunc: 1027 if (ICI.isEquality() && LHSI->hasOneUse()) { 1028 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all 1029 // of the high bits truncated out of x are known. 1030 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(), 1031 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1032 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0); 1033 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne); 1034 1035 // If all the high bits are known, we can do this xform. 1036 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) { 1037 // Pull in the high bits from known-ones set. 1038 APInt NewRHS = RHS->getValue().zext(SrcBits); 1039 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits); 1040 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1041 ConstantInt::get(ICI.getContext(), NewRHS)); 1042 } 1043 } 1044 break; 1045 1046 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI) 1047 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1048 // If this is a comparison that tests the signbit (X < 0) or (x > -1), 1049 // fold the xor. 1050 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) || 1051 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) { 1052 Value *CompareVal = LHSI->getOperand(0); 1053 1054 // If the sign bit of the XorCST is not set, there is no change to 1055 // the operation, just stop using the Xor. 1056 if (!XorCST->isNegative()) { 1057 ICI.setOperand(0, CompareVal); 1058 Worklist.Add(LHSI); 1059 return &ICI; 1060 } 1061 1062 // Was the old condition true if the operand is positive? 1063 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT; 1064 1065 // If so, the new one isn't. 1066 isTrueIfPositive ^= true; 1067 1068 if (isTrueIfPositive) 1069 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, 1070 SubOne(RHS)); 1071 else 1072 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, 1073 AddOne(RHS)); 1074 } 1075 1076 if (LHSI->hasOneUse()) { 1077 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit)) 1078 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) { 1079 const APInt &SignBit = XorCST->getValue(); 1080 ICmpInst::Predicate Pred = ICI.isSigned() 1081 ? ICI.getUnsignedPredicate() 1082 : ICI.getSignedPredicate(); 1083 return new ICmpInst(Pred, LHSI->getOperand(0), 1084 ConstantInt::get(ICI.getContext(), 1085 RHSV ^ SignBit)); 1086 } 1087 1088 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A) 1089 if (!ICI.isEquality() && XorCST->isMaxValue(true)) { 1090 const APInt &NotSignBit = XorCST->getValue(); 1091 ICmpInst::Predicate Pred = ICI.isSigned() 1092 ? ICI.getUnsignedPredicate() 1093 : ICI.getSignedPredicate(); 1094 Pred = ICI.getSwappedPredicate(Pred); 1095 return new ICmpInst(Pred, LHSI->getOperand(0), 1096 ConstantInt::get(ICI.getContext(), 1097 RHSV ^ NotSignBit)); 1098 } 1099 } 1100 } 1101 break; 1102 case Instruction::And: // (icmp pred (and X, AndCST), RHS) 1103 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) && 1104 LHSI->getOperand(0)->hasOneUse()) { 1105 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1)); 1106 1107 // If the LHS is an AND of a truncating cast, we can widen the 1108 // and/compare to be the input width without changing the value 1109 // produced, eliminating a cast. 1110 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) { 1111 // We can do this transformation if either the AND constant does not 1112 // have its sign bit set or if it is an equality comparison. 1113 // Extending a relational comparison when we're checking the sign 1114 // bit would not work. 1115 if (ICI.isEquality() || 1116 (!AndCST->isNegative() && RHSV.isNonNegative())) { 1117 Value *NewAnd = 1118 Builder->CreateAnd(Cast->getOperand(0), 1119 ConstantExpr::getZExt(AndCST, Cast->getSrcTy())); 1120 NewAnd->takeName(LHSI); 1121 return new ICmpInst(ICI.getPredicate(), NewAnd, 1122 ConstantExpr::getZExt(RHS, Cast->getSrcTy())); 1123 } 1124 } 1125 1126 // If the LHS is an AND of a zext, and we have an equality compare, we can 1127 // shrink the and/compare to the smaller type, eliminating the cast. 1128 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) { 1129 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy()); 1130 // Make sure we don't compare the upper bits, SimplifyDemandedBits 1131 // should fold the icmp to true/false in that case. 1132 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) { 1133 Value *NewAnd = 1134 Builder->CreateAnd(Cast->getOperand(0), 1135 ConstantExpr::getTrunc(AndCST, Ty)); 1136 NewAnd->takeName(LHSI); 1137 return new ICmpInst(ICI.getPredicate(), NewAnd, 1138 ConstantExpr::getTrunc(RHS, Ty)); 1139 } 1140 } 1141 1142 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare 1143 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This 1144 // happens a LOT in code produced by the C front-end, for bitfield 1145 // access. 1146 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0)); 1147 if (Shift && !Shift->isShift()) 1148 Shift = 0; 1149 1150 ConstantInt *ShAmt; 1151 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0; 1152 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift. 1153 Type *AndTy = AndCST->getType(); // Type of the and. 1154 1155 // We can fold this as long as we can't shift unknown bits 1156 // into the mask. This can only happen with signed shift 1157 // rights, as they sign-extend. 1158 if (ShAmt) { 1159 bool CanFold = Shift->isLogicalShift(); 1160 if (!CanFold) { 1161 // To test for the bad case of the signed shr, see if any 1162 // of the bits shifted in could be tested after the mask. 1163 uint32_t TyBits = Ty->getPrimitiveSizeInBits(); 1164 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits); 1165 1166 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits(); 1167 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) & 1168 AndCST->getValue()) == 0) 1169 CanFold = true; 1170 } 1171 1172 if (CanFold) { 1173 Constant *NewCst; 1174 if (Shift->getOpcode() == Instruction::Shl) 1175 NewCst = ConstantExpr::getLShr(RHS, ShAmt); 1176 else 1177 NewCst = ConstantExpr::getShl(RHS, ShAmt); 1178 1179 // Check to see if we are shifting out any of the bits being 1180 // compared. 1181 if (ConstantExpr::get(Shift->getOpcode(), 1182 NewCst, ShAmt) != RHS) { 1183 // If we shifted bits out, the fold is not going to work out. 1184 // As a special case, check to see if this means that the 1185 // result is always true or false now. 1186 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1187 return ReplaceInstUsesWith(ICI, 1188 ConstantInt::getFalse(ICI.getContext())); 1189 if (ICI.getPredicate() == ICmpInst::ICMP_NE) 1190 return ReplaceInstUsesWith(ICI, 1191 ConstantInt::getTrue(ICI.getContext())); 1192 } else { 1193 ICI.setOperand(1, NewCst); 1194 Constant *NewAndCST; 1195 if (Shift->getOpcode() == Instruction::Shl) 1196 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt); 1197 else 1198 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt); 1199 LHSI->setOperand(1, NewAndCST); 1200 LHSI->setOperand(0, Shift->getOperand(0)); 1201 Worklist.Add(Shift); // Shift is dead. 1202 return &ICI; 1203 } 1204 } 1205 } 1206 1207 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is 1208 // preferable because it allows the C<<Y expression to be hoisted out 1209 // of a loop if Y is invariant and X is not. 1210 if (Shift && Shift->hasOneUse() && RHSV == 0 && 1211 ICI.isEquality() && !Shift->isArithmeticShift() && 1212 !isa<Constant>(Shift->getOperand(0))) { 1213 // Compute C << Y. 1214 Value *NS; 1215 if (Shift->getOpcode() == Instruction::LShr) { 1216 NS = Builder->CreateShl(AndCST, Shift->getOperand(1)); 1217 } else { 1218 // Insert a logical shift. 1219 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1)); 1220 } 1221 1222 // Compute X & (C << Y). 1223 Value *NewAnd = 1224 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName()); 1225 1226 ICI.setOperand(0, NewAnd); 1227 return &ICI; 1228 } 1229 1230 // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any 1231 // bit set in (X & AndCST) will produce a result greater than RHSV. 1232 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) { 1233 unsigned NTZ = AndCST->getValue().countTrailingZeros(); 1234 if ((NTZ < AndCST->getBitWidth()) && 1235 APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV)) 1236 return new ICmpInst(ICmpInst::ICMP_NE, LHSI, 1237 Constant::getNullValue(RHS->getType())); 1238 } 1239 } 1240 1241 // Try to optimize things like "A[i]&42 == 0" to index computations. 1242 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) { 1243 if (GetElementPtrInst *GEP = 1244 dyn_cast<GetElementPtrInst>(LI->getOperand(0))) 1245 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 1246 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 1247 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) { 1248 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1)); 1249 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C)) 1250 return Res; 1251 } 1252 } 1253 break; 1254 1255 case Instruction::Or: { 1256 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse()) 1257 break; 1258 Value *P, *Q; 1259 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) { 1260 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 1261 // -> and (icmp eq P, null), (icmp eq Q, null). 1262 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P, 1263 Constant::getNullValue(P->getType())); 1264 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q, 1265 Constant::getNullValue(Q->getType())); 1266 Instruction *Op; 1267 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1268 Op = BinaryOperator::CreateAnd(ICIP, ICIQ); 1269 else 1270 Op = BinaryOperator::CreateOr(ICIP, ICIQ); 1271 return Op; 1272 } 1273 break; 1274 } 1275 1276 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI) 1277 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1278 if (!ShAmt) break; 1279 1280 uint32_t TypeBits = RHSV.getBitWidth(); 1281 1282 // Check that the shift amount is in range. If not, don't perform 1283 // undefined shifts. When the shift is visited it will be 1284 // simplified. 1285 if (ShAmt->uge(TypeBits)) 1286 break; 1287 1288 if (ICI.isEquality()) { 1289 // If we are comparing against bits always shifted out, the 1290 // comparison cannot succeed. 1291 Constant *Comp = 1292 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), 1293 ShAmt); 1294 if (Comp != RHS) {// Comparing against a bit that we know is zero. 1295 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1296 Constant *Cst = 1297 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE); 1298 return ReplaceInstUsesWith(ICI, Cst); 1299 } 1300 1301 // If the shift is NUW, then it is just shifting out zeros, no need for an 1302 // AND. 1303 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap()) 1304 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1305 ConstantExpr::getLShr(RHS, ShAmt)); 1306 1307 if (LHSI->hasOneUse()) { 1308 // Otherwise strength reduce the shift into an and. 1309 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 1310 Constant *Mask = 1311 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits, 1312 TypeBits-ShAmtVal)); 1313 1314 Value *And = 1315 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask"); 1316 return new ICmpInst(ICI.getPredicate(), And, 1317 ConstantExpr::getLShr(RHS, ShAmt)); 1318 } 1319 } 1320 1321 // Otherwise, if this is a comparison of the sign bit, simplify to and/test. 1322 bool TrueIfSigned = false; 1323 if (LHSI->hasOneUse() && 1324 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) { 1325 // (X << 31) <s 0 --> (X&1) != 0 1326 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(), 1327 APInt::getOneBitSet(TypeBits, 1328 TypeBits-ShAmt->getZExtValue()-1)); 1329 Value *And = 1330 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask"); 1331 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, 1332 And, Constant::getNullValue(And->getType())); 1333 } 1334 1335 // Transform (icmp pred iM (shl iM %v, N), CI) 1336 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N)) 1337 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N. 1338 // This enables to get rid of the shift in favor of a trunc which can be 1339 // free on the target. It has the additional benefit of comparing to a 1340 // smaller constant, which will be target friendly. 1341 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1); 1342 if (LHSI->hasOneUse() && 1343 Amt != 0 && RHSV.countTrailingZeros() >= Amt) { 1344 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt); 1345 Constant *NCI = ConstantExpr::getTrunc( 1346 ConstantExpr::getAShr(RHS, 1347 ConstantInt::get(RHS->getType(), Amt)), 1348 NTy); 1349 return new ICmpInst(ICI.getPredicate(), 1350 Builder->CreateTrunc(LHSI->getOperand(0), NTy), 1351 NCI); 1352 } 1353 1354 break; 1355 } 1356 1357 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI) 1358 case Instruction::AShr: { 1359 // Handle equality comparisons of shift-by-constant. 1360 BinaryOperator *BO = cast<BinaryOperator>(LHSI); 1361 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1362 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt)) 1363 return Res; 1364 } 1365 1366 // Handle exact shr's. 1367 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) { 1368 if (RHSV.isMinValue()) 1369 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS); 1370 } 1371 break; 1372 } 1373 1374 case Instruction::SDiv: 1375 case Instruction::UDiv: 1376 // Fold: icmp pred ([us]div X, C1), C2 -> range test 1377 // Fold this div into the comparison, producing a range check. 1378 // Determine, based on the divide type, what the range is being 1379 // checked. If there is an overflow on the low or high side, remember 1380 // it, otherwise compute the range [low, hi) bounding the new value. 1381 // See: InsertRangeTest above for the kinds of replacements possible. 1382 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) 1383 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI), 1384 DivRHS)) 1385 return R; 1386 break; 1387 1388 case Instruction::Add: 1389 // Fold: icmp pred (add X, C1), C2 1390 if (!ICI.isEquality()) { 1391 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1392 if (!LHSC) break; 1393 const APInt &LHSV = LHSC->getValue(); 1394 1395 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV) 1396 .subtract(LHSV); 1397 1398 if (ICI.isSigned()) { 1399 if (CR.getLower().isSignBit()) { 1400 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0), 1401 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1402 } else if (CR.getUpper().isSignBit()) { 1403 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0), 1404 ConstantInt::get(ICI.getContext(),CR.getLower())); 1405 } 1406 } else { 1407 if (CR.getLower().isMinValue()) { 1408 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), 1409 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1410 } else if (CR.getUpper().isMinValue()) { 1411 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), 1412 ConstantInt::get(ICI.getContext(),CR.getLower())); 1413 } 1414 } 1415 } 1416 break; 1417 } 1418 1419 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS. 1420 if (ICI.isEquality()) { 1421 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1422 1423 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and 1424 // the second operand is a constant, simplify a bit. 1425 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) { 1426 switch (BO->getOpcode()) { 1427 case Instruction::SRem: 1428 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. 1429 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){ 1430 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue(); 1431 if (V.sgt(1) && V.isPowerOf2()) { 1432 Value *NewRem = 1433 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1), 1434 BO->getName()); 1435 return new ICmpInst(ICI.getPredicate(), NewRem, 1436 Constant::getNullValue(BO->getType())); 1437 } 1438 } 1439 break; 1440 case Instruction::Add: 1441 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants. 1442 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1443 if (BO->hasOneUse()) 1444 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1445 ConstantExpr::getSub(RHS, BOp1C)); 1446 } else if (RHSV == 0) { 1447 // Replace ((add A, B) != 0) with (A != -B) if A or B is 1448 // efficiently invertible, or if the add has just this one use. 1449 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1); 1450 1451 if (Value *NegVal = dyn_castNegVal(BOp1)) 1452 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal); 1453 if (Value *NegVal = dyn_castNegVal(BOp0)) 1454 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1); 1455 if (BO->hasOneUse()) { 1456 Value *Neg = Builder->CreateNeg(BOp1); 1457 Neg->takeName(BO); 1458 return new ICmpInst(ICI.getPredicate(), BOp0, Neg); 1459 } 1460 } 1461 break; 1462 case Instruction::Xor: 1463 // For the xor case, we can xor two constants together, eliminating 1464 // the explicit xor. 1465 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) { 1466 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1467 ConstantExpr::getXor(RHS, BOC)); 1468 } else if (RHSV == 0) { 1469 // Replace ((xor A, B) != 0) with (A != B) 1470 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1471 BO->getOperand(1)); 1472 } 1473 break; 1474 case Instruction::Sub: 1475 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants. 1476 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) { 1477 if (BO->hasOneUse()) 1478 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1), 1479 ConstantExpr::getSub(BOp0C, RHS)); 1480 } else if (RHSV == 0) { 1481 // Replace ((sub A, B) != 0) with (A != B) 1482 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1483 BO->getOperand(1)); 1484 } 1485 break; 1486 case Instruction::Or: 1487 // If bits are being or'd in that are not present in the constant we 1488 // are comparing against, then the comparison could never succeed! 1489 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1490 Constant *NotCI = ConstantExpr::getNot(RHS); 1491 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue()) 1492 return ReplaceInstUsesWith(ICI, 1493 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1494 isICMP_NE)); 1495 } 1496 break; 1497 1498 case Instruction::And: 1499 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1500 // If bits are being compared against that are and'd out, then the 1501 // comparison can never succeed! 1502 if ((RHSV & ~BOC->getValue()) != 0) 1503 return ReplaceInstUsesWith(ICI, 1504 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1505 isICMP_NE)); 1506 1507 // If we have ((X & C) == C), turn it into ((X & C) != 0). 1508 if (RHS == BOC && RHSV.isPowerOf2()) 1509 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : 1510 ICmpInst::ICMP_NE, LHSI, 1511 Constant::getNullValue(RHS->getType())); 1512 1513 // Don't perform the following transforms if the AND has multiple uses 1514 if (!BO->hasOneUse()) 1515 break; 1516 1517 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0 1518 if (BOC->getValue().isSignBit()) { 1519 Value *X = BO->getOperand(0); 1520 Constant *Zero = Constant::getNullValue(X->getType()); 1521 ICmpInst::Predicate pred = isICMP_NE ? 1522 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; 1523 return new ICmpInst(pred, X, Zero); 1524 } 1525 1526 // ((X & ~7) == 0) --> X < 8 1527 if (RHSV == 0 && isHighOnes(BOC)) { 1528 Value *X = BO->getOperand(0); 1529 Constant *NegX = ConstantExpr::getNeg(BOC); 1530 ICmpInst::Predicate pred = isICMP_NE ? 1531 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; 1532 return new ICmpInst(pred, X, NegX); 1533 } 1534 } 1535 default: break; 1536 } 1537 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) { 1538 // Handle icmp {eq|ne} <intrinsic>, intcst. 1539 switch (II->getIntrinsicID()) { 1540 case Intrinsic::bswap: 1541 Worklist.Add(II); 1542 ICI.setOperand(0, II->getArgOperand(0)); 1543 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap())); 1544 return &ICI; 1545 case Intrinsic::ctlz: 1546 case Intrinsic::cttz: 1547 // ctz(A) == bitwidth(a) -> A == 0 and likewise for != 1548 if (RHSV == RHS->getType()->getBitWidth()) { 1549 Worklist.Add(II); 1550 ICI.setOperand(0, II->getArgOperand(0)); 1551 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0)); 1552 return &ICI; 1553 } 1554 break; 1555 case Intrinsic::ctpop: 1556 // popcount(A) == 0 -> A == 0 and likewise for != 1557 if (RHS->isZero()) { 1558 Worklist.Add(II); 1559 ICI.setOperand(0, II->getArgOperand(0)); 1560 ICI.setOperand(1, RHS); 1561 return &ICI; 1562 } 1563 break; 1564 default: 1565 break; 1566 } 1567 } 1568 } 1569 return 0; 1570 } 1571 1572 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst). 1573 /// We only handle extending casts so far. 1574 /// 1575 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) { 1576 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0)); 1577 Value *LHSCIOp = LHSCI->getOperand(0); 1578 Type *SrcTy = LHSCIOp->getType(); 1579 Type *DestTy = LHSCI->getType(); 1580 Value *RHSCIOp; 1581 1582 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the 1583 // integer type is the same size as the pointer type. 1584 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt && 1585 TD->getPointerSizeInBits() == 1586 cast<IntegerType>(DestTy)->getBitWidth()) { 1587 Value *RHSOp = 0; 1588 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) { 1589 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy); 1590 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) { 1591 RHSOp = RHSC->getOperand(0); 1592 // If the pointer types don't match, insert a bitcast. 1593 if (LHSCIOp->getType() != RHSOp->getType()) 1594 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType()); 1595 } 1596 1597 if (RHSOp) 1598 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp); 1599 } 1600 1601 // The code below only handles extension cast instructions, so far. 1602 // Enforce this. 1603 if (LHSCI->getOpcode() != Instruction::ZExt && 1604 LHSCI->getOpcode() != Instruction::SExt) 1605 return 0; 1606 1607 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt; 1608 bool isSignedCmp = ICI.isSigned(); 1609 1610 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) { 1611 // Not an extension from the same type? 1612 RHSCIOp = CI->getOperand(0); 1613 if (RHSCIOp->getType() != LHSCIOp->getType()) 1614 return 0; 1615 1616 // If the signedness of the two casts doesn't agree (i.e. one is a sext 1617 // and the other is a zext), then we can't handle this. 1618 if (CI->getOpcode() != LHSCI->getOpcode()) 1619 return 0; 1620 1621 // Deal with equality cases early. 1622 if (ICI.isEquality()) 1623 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1624 1625 // A signed comparison of sign extended values simplifies into a 1626 // signed comparison. 1627 if (isSignedCmp && isSignedExt) 1628 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1629 1630 // The other three cases all fold into an unsigned comparison. 1631 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp); 1632 } 1633 1634 // If we aren't dealing with a constant on the RHS, exit early 1635 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1)); 1636 if (!CI) 1637 return 0; 1638 1639 // Compute the constant that would happen if we truncated to SrcTy then 1640 // reextended to DestTy. 1641 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy); 1642 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), 1643 Res1, DestTy); 1644 1645 // If the re-extended constant didn't change... 1646 if (Res2 == CI) { 1647 // Deal with equality cases early. 1648 if (ICI.isEquality()) 1649 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1650 1651 // A signed comparison of sign extended values simplifies into a 1652 // signed comparison. 1653 if (isSignedExt && isSignedCmp) 1654 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1655 1656 // The other three cases all fold into an unsigned comparison. 1657 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1); 1658 } 1659 1660 // The re-extended constant changed so the constant cannot be represented 1661 // in the shorter type. Consequently, we cannot emit a simple comparison. 1662 // All the cases that fold to true or false will have already been handled 1663 // by SimplifyICmpInst, so only deal with the tricky case. 1664 1665 if (isSignedCmp || !isSignedExt) 1666 return 0; 1667 1668 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases 1669 // should have been folded away previously and not enter in here. 1670 1671 // We're performing an unsigned comp with a sign extended value. 1672 // This is true if the input is >= 0. [aka >s -1] 1673 Constant *NegOne = Constant::getAllOnesValue(SrcTy); 1674 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName()); 1675 1676 // Finally, return the value computed. 1677 if (ICI.getPredicate() == ICmpInst::ICMP_ULT) 1678 return ReplaceInstUsesWith(ICI, Result); 1679 1680 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"); 1681 return BinaryOperator::CreateNot(Result); 1682 } 1683 1684 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form: 1685 /// I = icmp ugt (add (add A, B), CI2), CI1 1686 /// If this is of the form: 1687 /// sum = a + b 1688 /// if (sum+128 >u 255) 1689 /// Then replace it with llvm.sadd.with.overflow.i8. 1690 /// 1691 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B, 1692 ConstantInt *CI2, ConstantInt *CI1, 1693 InstCombiner &IC) { 1694 // The transformation we're trying to do here is to transform this into an 1695 // llvm.sadd.with.overflow. To do this, we have to replace the original add 1696 // with a narrower add, and discard the add-with-constant that is part of the 1697 // range check (if we can't eliminate it, this isn't profitable). 1698 1699 // In order to eliminate the add-with-constant, the compare can be its only 1700 // use. 1701 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0)); 1702 if (!AddWithCst->hasOneUse()) return 0; 1703 1704 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow. 1705 if (!CI2->getValue().isPowerOf2()) return 0; 1706 unsigned NewWidth = CI2->getValue().countTrailingZeros(); 1707 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0; 1708 1709 // The width of the new add formed is 1 more than the bias. 1710 ++NewWidth; 1711 1712 // Check to see that CI1 is an all-ones value with NewWidth bits. 1713 if (CI1->getBitWidth() == NewWidth || 1714 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth)) 1715 return 0; 1716 1717 // This is only really a signed overflow check if the inputs have been 1718 // sign-extended; check for that condition. For example, if CI2 is 2^31 and 1719 // the operands of the add are 64 bits wide, we need at least 33 sign bits. 1720 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1; 1721 if (IC.ComputeNumSignBits(A) < NeededSignBits || 1722 IC.ComputeNumSignBits(B) < NeededSignBits) 1723 return 0; 1724 1725 // In order to replace the original add with a narrower 1726 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant 1727 // and truncates that discard the high bits of the add. Verify that this is 1728 // the case. 1729 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0)); 1730 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end(); 1731 UI != E; ++UI) { 1732 if (*UI == AddWithCst) continue; 1733 1734 // Only accept truncates for now. We would really like a nice recursive 1735 // predicate like SimplifyDemandedBits, but which goes downwards the use-def 1736 // chain to see which bits of a value are actually demanded. If the 1737 // original add had another add which was then immediately truncated, we 1738 // could still do the transformation. 1739 TruncInst *TI = dyn_cast<TruncInst>(*UI); 1740 if (TI == 0 || 1741 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0; 1742 } 1743 1744 // If the pattern matches, truncate the inputs to the narrower type and 1745 // use the sadd_with_overflow intrinsic to efficiently compute both the 1746 // result and the overflow bit. 1747 Module *M = I.getParent()->getParent()->getParent(); 1748 1749 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth); 1750 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow, 1751 NewType); 1752 1753 InstCombiner::BuilderTy *Builder = IC.Builder; 1754 1755 // Put the new code above the original add, in case there are any uses of the 1756 // add between the add and the compare. 1757 Builder->SetInsertPoint(OrigAdd); 1758 1759 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc"); 1760 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc"); 1761 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd"); 1762 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result"); 1763 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType()); 1764 1765 // The inner add was the result of the narrow add, zero extended to the 1766 // wider type. Replace it with the result computed by the intrinsic. 1767 IC.ReplaceInstUsesWith(*OrigAdd, ZExt); 1768 1769 // The original icmp gets replaced with the overflow value. 1770 return ExtractValueInst::Create(Call, 1, "sadd.overflow"); 1771 } 1772 1773 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV, 1774 InstCombiner &IC) { 1775 // Don't bother doing this transformation for pointers, don't do it for 1776 // vectors. 1777 if (!isa<IntegerType>(OrigAddV->getType())) return 0; 1778 1779 // If the add is a constant expr, then we don't bother transforming it. 1780 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV); 1781 if (OrigAdd == 0) return 0; 1782 1783 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1); 1784 1785 // Put the new code above the original add, in case there are any uses of the 1786 // add between the add and the compare. 1787 InstCombiner::BuilderTy *Builder = IC.Builder; 1788 Builder->SetInsertPoint(OrigAdd); 1789 1790 Module *M = I.getParent()->getParent()->getParent(); 1791 Type *Ty = LHS->getType(); 1792 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty); 1793 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd"); 1794 Value *Add = Builder->CreateExtractValue(Call, 0); 1795 1796 IC.ReplaceInstUsesWith(*OrigAdd, Add); 1797 1798 // The original icmp gets replaced with the overflow value. 1799 return ExtractValueInst::Create(Call, 1, "uadd.overflow"); 1800 } 1801 1802 // DemandedBitsLHSMask - When performing a comparison against a constant, 1803 // it is possible that not all the bits in the LHS are demanded. This helper 1804 // method computes the mask that IS demanded. 1805 static APInt DemandedBitsLHSMask(ICmpInst &I, 1806 unsigned BitWidth, bool isSignCheck) { 1807 if (isSignCheck) 1808 return APInt::getSignBit(BitWidth); 1809 1810 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1)); 1811 if (!CI) return APInt::getAllOnesValue(BitWidth); 1812 const APInt &RHS = CI->getValue(); 1813 1814 switch (I.getPredicate()) { 1815 // For a UGT comparison, we don't care about any bits that 1816 // correspond to the trailing ones of the comparand. The value of these 1817 // bits doesn't impact the outcome of the comparison, because any value 1818 // greater than the RHS must differ in a bit higher than these due to carry. 1819 case ICmpInst::ICMP_UGT: { 1820 unsigned trailingOnes = RHS.countTrailingOnes(); 1821 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes); 1822 return ~lowBitsSet; 1823 } 1824 1825 // Similarly, for a ULT comparison, we don't care about the trailing zeros. 1826 // Any value less than the RHS must differ in a higher bit because of carries. 1827 case ICmpInst::ICMP_ULT: { 1828 unsigned trailingZeros = RHS.countTrailingZeros(); 1829 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros); 1830 return ~lowBitsSet; 1831 } 1832 1833 default: 1834 return APInt::getAllOnesValue(BitWidth); 1835 } 1836 1837 } 1838 1839 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { 1840 bool Changed = false; 1841 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1842 1843 /// Orders the operands of the compare so that they are listed from most 1844 /// complex to least complex. This puts constants before unary operators, 1845 /// before binary operators. 1846 if (getComplexity(Op0) < getComplexity(Op1)) { 1847 I.swapOperands(); 1848 std::swap(Op0, Op1); 1849 Changed = true; 1850 } 1851 1852 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD)) 1853 return ReplaceInstUsesWith(I, V); 1854 1855 // comparing -val or val with non-zero is the same as just comparing val 1856 // ie, abs(val) != 0 -> val != 0 1857 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) 1858 { 1859 Value *Cond, *SelectTrue, *SelectFalse; 1860 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue), 1861 m_Value(SelectFalse)))) { 1862 if (Value *V = dyn_castNegVal(SelectTrue)) { 1863 if (V == SelectFalse) 1864 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 1865 } 1866 else if (Value *V = dyn_castNegVal(SelectFalse)) { 1867 if (V == SelectTrue) 1868 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 1869 } 1870 } 1871 } 1872 1873 Type *Ty = Op0->getType(); 1874 1875 // icmp's with boolean values can always be turned into bitwise operations 1876 if (Ty->isIntegerTy(1)) { 1877 switch (I.getPredicate()) { 1878 default: llvm_unreachable("Invalid icmp instruction!"); 1879 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B) 1880 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp"); 1881 return BinaryOperator::CreateNot(Xor); 1882 } 1883 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B 1884 return BinaryOperator::CreateXor(Op0, Op1); 1885 1886 case ICmpInst::ICMP_UGT: 1887 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult 1888 // FALL THROUGH 1889 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B 1890 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1891 return BinaryOperator::CreateAnd(Not, Op1); 1892 } 1893 case ICmpInst::ICMP_SGT: 1894 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt 1895 // FALL THROUGH 1896 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B 1897 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1898 return BinaryOperator::CreateAnd(Not, Op0); 1899 } 1900 case ICmpInst::ICMP_UGE: 1901 std::swap(Op0, Op1); // Change icmp uge -> icmp ule 1902 // FALL THROUGH 1903 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B 1904 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1905 return BinaryOperator::CreateOr(Not, Op1); 1906 } 1907 case ICmpInst::ICMP_SGE: 1908 std::swap(Op0, Op1); // Change icmp sge -> icmp sle 1909 // FALL THROUGH 1910 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B 1911 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1912 return BinaryOperator::CreateOr(Not, Op0); 1913 } 1914 } 1915 } 1916 1917 unsigned BitWidth = 0; 1918 if (Ty->isIntOrIntVectorTy()) 1919 BitWidth = Ty->getScalarSizeInBits(); 1920 else if (TD) // Pointers require TD info to get their size. 1921 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType()); 1922 1923 bool isSignBit = false; 1924 1925 // See if we are doing a comparison with a constant. 1926 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 1927 Value *A = 0, *B = 0; 1928 1929 // Match the following pattern, which is a common idiom when writing 1930 // overflow-safe integer arithmetic function. The source performs an 1931 // addition in wider type, and explicitly checks for overflow using 1932 // comparisons against INT_MIN and INT_MAX. Simplify this by using the 1933 // sadd_with_overflow intrinsic. 1934 // 1935 // TODO: This could probably be generalized to handle other overflow-safe 1936 // operations if we worked out the formulas to compute the appropriate 1937 // magic constants. 1938 // 1939 // sum = a + b 1940 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8 1941 { 1942 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI 1943 if (I.getPredicate() == ICmpInst::ICMP_UGT && 1944 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2)))) 1945 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this)) 1946 return Res; 1947 } 1948 1949 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B) 1950 if (I.isEquality() && CI->isZero() && 1951 match(Op0, m_Sub(m_Value(A), m_Value(B)))) { 1952 // (icmp cond A B) if cond is equality 1953 return new ICmpInst(I.getPredicate(), A, B); 1954 } 1955 1956 // If we have an icmp le or icmp ge instruction, turn it into the 1957 // appropriate icmp lt or icmp gt instruction. This allows us to rely on 1958 // them being folded in the code below. The SimplifyICmpInst code has 1959 // already handled the edge cases for us, so we just assert on them. 1960 switch (I.getPredicate()) { 1961 default: break; 1962 case ICmpInst::ICMP_ULE: 1963 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE 1964 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, 1965 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1966 case ICmpInst::ICMP_SLE: 1967 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE 1968 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 1969 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1970 case ICmpInst::ICMP_UGE: 1971 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE 1972 return new ICmpInst(ICmpInst::ICMP_UGT, Op0, 1973 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1974 case ICmpInst::ICMP_SGE: 1975 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE 1976 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 1977 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1978 } 1979 1980 // If this comparison is a normal comparison, it demands all 1981 // bits, if it is a sign bit comparison, it only demands the sign bit. 1982 bool UnusedBit; 1983 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit); 1984 } 1985 1986 // See if we can fold the comparison based on range information we can get 1987 // by checking whether bits are known to be zero or one in the input. 1988 if (BitWidth != 0) { 1989 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0); 1990 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0); 1991 1992 if (SimplifyDemandedBits(I.getOperandUse(0), 1993 DemandedBitsLHSMask(I, BitWidth, isSignBit), 1994 Op0KnownZero, Op0KnownOne, 0)) 1995 return &I; 1996 if (SimplifyDemandedBits(I.getOperandUse(1), 1997 APInt::getAllOnesValue(BitWidth), 1998 Op1KnownZero, Op1KnownOne, 0)) 1999 return &I; 2000 2001 // Given the known and unknown bits, compute a range that the LHS could be 2002 // in. Compute the Min, Max and RHS values based on the known bits. For the 2003 // EQ and NE we use unsigned values. 2004 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0); 2005 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0); 2006 if (I.isSigned()) { 2007 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 2008 Op0Min, Op0Max); 2009 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 2010 Op1Min, Op1Max); 2011 } else { 2012 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 2013 Op0Min, Op0Max); 2014 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 2015 Op1Min, Op1Max); 2016 } 2017 2018 // If Min and Max are known to be the same, then SimplifyDemandedBits 2019 // figured out that the LHS is a constant. Just constant fold this now so 2020 // that code below can assume that Min != Max. 2021 if (!isa<Constant>(Op0) && Op0Min == Op0Max) 2022 return new ICmpInst(I.getPredicate(), 2023 ConstantInt::get(Op0->getType(), Op0Min), Op1); 2024 if (!isa<Constant>(Op1) && Op1Min == Op1Max) 2025 return new ICmpInst(I.getPredicate(), Op0, 2026 ConstantInt::get(Op1->getType(), Op1Min)); 2027 2028 // Based on the range information we know about the LHS, see if we can 2029 // simplify this comparison. For example, (x&4) < 8 is always true. 2030 switch (I.getPredicate()) { 2031 default: llvm_unreachable("Unknown icmp opcode!"); 2032 case ICmpInst::ICMP_EQ: { 2033 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2034 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2035 2036 // If all bits are known zero except for one, then we know at most one 2037 // bit is set. If the comparison is against zero, then this is a check 2038 // to see if *that* bit is set. 2039 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2040 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2041 // If the LHS is an AND with the same constant, look through it. 2042 Value *LHS = 0; 2043 ConstantInt *LHSC = 0; 2044 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2045 LHSC->getValue() != Op0KnownZeroInverted) 2046 LHS = Op0; 2047 2048 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2049 // then turn "((1 << x)&8) == 0" into "x != 3". 2050 Value *X = 0; 2051 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2052 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2053 return new ICmpInst(ICmpInst::ICMP_NE, X, 2054 ConstantInt::get(X->getType(), CmpVal)); 2055 } 2056 2057 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2058 // then turn "((8 >>u x)&1) == 0" into "x != 3". 2059 const APInt *CI; 2060 if (Op0KnownZeroInverted == 1 && 2061 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2062 return new ICmpInst(ICmpInst::ICMP_NE, X, 2063 ConstantInt::get(X->getType(), 2064 CI->countTrailingZeros())); 2065 } 2066 2067 break; 2068 } 2069 case ICmpInst::ICMP_NE: { 2070 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2071 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2072 2073 // If all bits are known zero except for one, then we know at most one 2074 // bit is set. If the comparison is against zero, then this is a check 2075 // to see if *that* bit is set. 2076 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2077 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2078 // If the LHS is an AND with the same constant, look through it. 2079 Value *LHS = 0; 2080 ConstantInt *LHSC = 0; 2081 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2082 LHSC->getValue() != Op0KnownZeroInverted) 2083 LHS = Op0; 2084 2085 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2086 // then turn "((1 << x)&8) != 0" into "x == 3". 2087 Value *X = 0; 2088 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2089 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2090 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2091 ConstantInt::get(X->getType(), CmpVal)); 2092 } 2093 2094 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2095 // then turn "((8 >>u x)&1) != 0" into "x == 3". 2096 const APInt *CI; 2097 if (Op0KnownZeroInverted == 1 && 2098 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2099 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2100 ConstantInt::get(X->getType(), 2101 CI->countTrailingZeros())); 2102 } 2103 2104 break; 2105 } 2106 case ICmpInst::ICMP_ULT: 2107 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B) 2108 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2109 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B) 2110 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2111 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B) 2112 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2113 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2114 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C 2115 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2116 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2117 2118 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear 2119 if (CI->isMinValue(true)) 2120 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 2121 Constant::getAllOnesValue(Op0->getType())); 2122 } 2123 break; 2124 case ICmpInst::ICMP_UGT: 2125 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B) 2126 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2127 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B) 2128 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2129 2130 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B) 2131 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2132 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2133 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C 2134 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2135 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2136 2137 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set 2138 if (CI->isMaxValue(true)) 2139 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 2140 Constant::getNullValue(Op0->getType())); 2141 } 2142 break; 2143 case ICmpInst::ICMP_SLT: 2144 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C) 2145 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2146 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C) 2147 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2148 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B) 2149 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2150 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2151 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C 2152 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2153 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2154 } 2155 break; 2156 case ICmpInst::ICMP_SGT: 2157 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B) 2158 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2159 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B) 2160 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2161 2162 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B) 2163 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2164 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2165 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C 2166 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2167 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2168 } 2169 break; 2170 case ICmpInst::ICMP_SGE: 2171 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"); 2172 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B) 2173 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2174 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B) 2175 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2176 break; 2177 case ICmpInst::ICMP_SLE: 2178 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"); 2179 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B) 2180 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2181 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B) 2182 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2183 break; 2184 case ICmpInst::ICMP_UGE: 2185 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"); 2186 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B) 2187 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2188 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B) 2189 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2190 break; 2191 case ICmpInst::ICMP_ULE: 2192 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"); 2193 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B) 2194 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2195 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B) 2196 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2197 break; 2198 } 2199 2200 // Turn a signed comparison into an unsigned one if both operands 2201 // are known to have the same sign. 2202 if (I.isSigned() && 2203 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) || 2204 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative()))) 2205 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1); 2206 } 2207 2208 // Test if the ICmpInst instruction is used exclusively by a select as 2209 // part of a minimum or maximum operation. If so, refrain from doing 2210 // any other folding. This helps out other analyses which understand 2211 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 2212 // and CodeGen. And in this case, at least one of the comparison 2213 // operands has at least one user besides the compare (the select), 2214 // which would often largely negate the benefit of folding anyway. 2215 if (I.hasOneUse()) 2216 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin())) 2217 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) || 2218 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1)) 2219 return 0; 2220 2221 // See if we are doing a comparison between a constant and an instruction that 2222 // can be folded into the comparison. 2223 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2224 // Since the RHS is a ConstantInt (CI), if the left hand side is an 2225 // instruction, see if that instruction also has constants so that the 2226 // instruction can be folded into the icmp 2227 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2228 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI)) 2229 return Res; 2230 } 2231 2232 // Handle icmp with constant (but not simple integer constant) RHS 2233 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2234 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2235 switch (LHSI->getOpcode()) { 2236 case Instruction::GetElementPtr: 2237 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null 2238 if (RHSC->isNullValue() && 2239 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices()) 2240 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2241 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2242 break; 2243 case Instruction::PHI: 2244 // Only fold icmp into the PHI if the phi and icmp are in the same 2245 // block. If in the same block, we're encouraging jump threading. If 2246 // not, we are just pessimizing the code by making an i1 phi. 2247 if (LHSI->getParent() == I.getParent()) 2248 if (Instruction *NV = FoldOpIntoPhi(I)) 2249 return NV; 2250 break; 2251 case Instruction::Select: { 2252 // If either operand of the select is a constant, we can fold the 2253 // comparison into the select arms, which will cause one to be 2254 // constant folded and the select turned into a bitwise or. 2255 Value *Op1 = 0, *Op2 = 0; 2256 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) 2257 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2258 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) 2259 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2260 2261 // We only want to perform this transformation if it will not lead to 2262 // additional code. This is true if either both sides of the select 2263 // fold to a constant (in which case the icmp is replaced with a select 2264 // which will usually simplify) or this is the only user of the 2265 // select (in which case we are trading a select+icmp for a simpler 2266 // select+icmp). 2267 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) { 2268 if (!Op1) 2269 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1), 2270 RHSC, I.getName()); 2271 if (!Op2) 2272 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2), 2273 RHSC, I.getName()); 2274 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2275 } 2276 break; 2277 } 2278 case Instruction::IntToPtr: 2279 // icmp pred inttoptr(X), null -> icmp pred X, 0 2280 if (RHSC->isNullValue() && TD && 2281 TD->getIntPtrType(RHSC->getContext()) == 2282 LHSI->getOperand(0)->getType()) 2283 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2284 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2285 break; 2286 2287 case Instruction::Load: 2288 // Try to optimize things like "A[i] > 4" to index computations. 2289 if (GetElementPtrInst *GEP = 2290 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2291 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2292 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2293 !cast<LoadInst>(LHSI)->isVolatile()) 2294 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2295 return Res; 2296 } 2297 break; 2298 } 2299 } 2300 2301 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now. 2302 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0)) 2303 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I)) 2304 return NI; 2305 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) 2306 if (Instruction *NI = FoldGEPICmp(GEP, Op0, 2307 ICmpInst::getSwappedPredicate(I.getPredicate()), I)) 2308 return NI; 2309 2310 // Test to see if the operands of the icmp are casted versions of other 2311 // values. If the ptr->ptr cast can be stripped off both arguments, we do so 2312 // now. 2313 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) { 2314 if (Op0->getType()->isPointerTy() && 2315 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) { 2316 // We keep moving the cast from the left operand over to the right 2317 // operand, where it can often be eliminated completely. 2318 Op0 = CI->getOperand(0); 2319 2320 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast 2321 // so eliminate it as well. 2322 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1)) 2323 Op1 = CI2->getOperand(0); 2324 2325 // If Op1 is a constant, we can fold the cast into the constant. 2326 if (Op0->getType() != Op1->getType()) { 2327 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 2328 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType()); 2329 } else { 2330 // Otherwise, cast the RHS right before the icmp 2331 Op1 = Builder->CreateBitCast(Op1, Op0->getType()); 2332 } 2333 } 2334 return new ICmpInst(I.getPredicate(), Op0, Op1); 2335 } 2336 } 2337 2338 if (isa<CastInst>(Op0)) { 2339 // Handle the special case of: icmp (cast bool to X), <cst> 2340 // This comes up when you have code like 2341 // int X = A < B; 2342 // if (X) ... 2343 // For generality, we handle any zero-extension of any operand comparison 2344 // with a constant or another cast from the same type. 2345 if (isa<Constant>(Op1) || isa<CastInst>(Op1)) 2346 if (Instruction *R = visitICmpInstWithCastAndCast(I)) 2347 return R; 2348 } 2349 2350 // Special logic for binary operators. 2351 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0); 2352 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1); 2353 if (BO0 || BO1) { 2354 CmpInst::Predicate Pred = I.getPredicate(); 2355 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false; 2356 if (BO0 && isa<OverflowingBinaryOperator>(BO0)) 2357 NoOp0WrapProblem = ICmpInst::isEquality(Pred) || 2358 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) || 2359 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap()); 2360 if (BO1 && isa<OverflowingBinaryOperator>(BO1)) 2361 NoOp1WrapProblem = ICmpInst::isEquality(Pred) || 2362 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) || 2363 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap()); 2364 2365 // Analyze the case when either Op0 or Op1 is an add instruction. 2366 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null). 2367 Value *A = 0, *B = 0, *C = 0, *D = 0; 2368 if (BO0 && BO0->getOpcode() == Instruction::Add) 2369 A = BO0->getOperand(0), B = BO0->getOperand(1); 2370 if (BO1 && BO1->getOpcode() == Instruction::Add) 2371 C = BO1->getOperand(0), D = BO1->getOperand(1); 2372 2373 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2374 if ((A == Op1 || B == Op1) && NoOp0WrapProblem) 2375 return new ICmpInst(Pred, A == Op1 ? B : A, 2376 Constant::getNullValue(Op1->getType())); 2377 2378 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2379 if ((C == Op0 || D == Op0) && NoOp1WrapProblem) 2380 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()), 2381 C == Op0 ? D : C); 2382 2383 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow. 2384 if (A && C && (A == C || A == D || B == C || B == D) && 2385 NoOp0WrapProblem && NoOp1WrapProblem && 2386 // Try not to increase register pressure. 2387 BO0->hasOneUse() && BO1->hasOneUse()) { 2388 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2389 Value *Y, *Z; 2390 if (A == C) { 2391 // C + B == C + D -> B == D 2392 Y = B; 2393 Z = D; 2394 } else if (A == D) { 2395 // D + B == C + D -> B == C 2396 Y = B; 2397 Z = C; 2398 } else if (B == C) { 2399 // A + C == C + D -> A == D 2400 Y = A; 2401 Z = D; 2402 } else { 2403 assert(B == D); 2404 // A + D == C + D -> A == C 2405 Y = A; 2406 Z = C; 2407 } 2408 return new ICmpInst(Pred, Y, Z); 2409 } 2410 2411 // Analyze the case when either Op0 or Op1 is a sub instruction. 2412 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null). 2413 A = 0; B = 0; C = 0; D = 0; 2414 if (BO0 && BO0->getOpcode() == Instruction::Sub) 2415 A = BO0->getOperand(0), B = BO0->getOperand(1); 2416 if (BO1 && BO1->getOpcode() == Instruction::Sub) 2417 C = BO1->getOperand(0), D = BO1->getOperand(1); 2418 2419 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow. 2420 if (A == Op1 && NoOp0WrapProblem) 2421 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B); 2422 2423 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow. 2424 if (C == Op0 && NoOp1WrapProblem) 2425 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType())); 2426 2427 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow. 2428 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem && 2429 // Try not to increase register pressure. 2430 BO0->hasOneUse() && BO1->hasOneUse()) 2431 return new ICmpInst(Pred, A, C); 2432 2433 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow. 2434 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem && 2435 // Try not to increase register pressure. 2436 BO0->hasOneUse() && BO1->hasOneUse()) 2437 return new ICmpInst(Pred, D, B); 2438 2439 BinaryOperator *SRem = NULL; 2440 // icmp (srem X, Y), Y 2441 if (BO0 && BO0->getOpcode() == Instruction::SRem && 2442 Op1 == BO0->getOperand(1)) 2443 SRem = BO0; 2444 // icmp Y, (srem X, Y) 2445 else if (BO1 && BO1->getOpcode() == Instruction::SRem && 2446 Op0 == BO1->getOperand(1)) 2447 SRem = BO1; 2448 if (SRem) { 2449 // We don't check hasOneUse to avoid increasing register pressure because 2450 // the value we use is the same value this instruction was already using. 2451 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) { 2452 default: break; 2453 case ICmpInst::ICMP_EQ: 2454 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2455 case ICmpInst::ICMP_NE: 2456 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2457 case ICmpInst::ICMP_SGT: 2458 case ICmpInst::ICMP_SGE: 2459 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1), 2460 Constant::getAllOnesValue(SRem->getType())); 2461 case ICmpInst::ICMP_SLT: 2462 case ICmpInst::ICMP_SLE: 2463 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1), 2464 Constant::getNullValue(SRem->getType())); 2465 } 2466 } 2467 2468 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && 2469 BO0->hasOneUse() && BO1->hasOneUse() && 2470 BO0->getOperand(1) == BO1->getOperand(1)) { 2471 switch (BO0->getOpcode()) { 2472 default: break; 2473 case Instruction::Add: 2474 case Instruction::Sub: 2475 case Instruction::Xor: 2476 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b 2477 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2478 BO1->getOperand(0)); 2479 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b 2480 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2481 if (CI->getValue().isSignBit()) { 2482 ICmpInst::Predicate Pred = I.isSigned() 2483 ? I.getUnsignedPredicate() 2484 : I.getSignedPredicate(); 2485 return new ICmpInst(Pred, BO0->getOperand(0), 2486 BO1->getOperand(0)); 2487 } 2488 2489 if (CI->isMaxValue(true)) { 2490 ICmpInst::Predicate Pred = I.isSigned() 2491 ? I.getUnsignedPredicate() 2492 : I.getSignedPredicate(); 2493 Pred = I.getSwappedPredicate(Pred); 2494 return new ICmpInst(Pred, BO0->getOperand(0), 2495 BO1->getOperand(0)); 2496 } 2497 } 2498 break; 2499 case Instruction::Mul: 2500 if (!I.isEquality()) 2501 break; 2502 2503 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2504 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask 2505 // Mask = -1 >> count-trailing-zeros(Cst). 2506 if (!CI->isZero() && !CI->isOne()) { 2507 const APInt &AP = CI->getValue(); 2508 ConstantInt *Mask = ConstantInt::get(I.getContext(), 2509 APInt::getLowBitsSet(AP.getBitWidth(), 2510 AP.getBitWidth() - 2511 AP.countTrailingZeros())); 2512 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask); 2513 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask); 2514 return new ICmpInst(I.getPredicate(), And1, And2); 2515 } 2516 } 2517 break; 2518 case Instruction::UDiv: 2519 case Instruction::LShr: 2520 if (I.isSigned()) 2521 break; 2522 // fall-through 2523 case Instruction::SDiv: 2524 case Instruction::AShr: 2525 if (!BO0->isExact() || !BO1->isExact()) 2526 break; 2527 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2528 BO1->getOperand(0)); 2529 case Instruction::Shl: { 2530 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap(); 2531 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap(); 2532 if (!NUW && !NSW) 2533 break; 2534 if (!NSW && I.isSigned()) 2535 break; 2536 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2537 BO1->getOperand(0)); 2538 } 2539 } 2540 } 2541 } 2542 2543 { Value *A, *B; 2544 // ~x < ~y --> y < x 2545 // ~x < cst --> ~cst < x 2546 if (match(Op0, m_Not(m_Value(A)))) { 2547 if (match(Op1, m_Not(m_Value(B)))) 2548 return new ICmpInst(I.getPredicate(), B, A); 2549 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1)) 2550 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A); 2551 } 2552 2553 // (a+b) <u a --> llvm.uadd.with.overflow. 2554 // (a+b) <u b --> llvm.uadd.with.overflow. 2555 if (I.getPredicate() == ICmpInst::ICMP_ULT && 2556 match(Op0, m_Add(m_Value(A), m_Value(B))) && 2557 (Op1 == A || Op1 == B)) 2558 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this)) 2559 return R; 2560 2561 // a >u (a+b) --> llvm.uadd.with.overflow. 2562 // b >u (a+b) --> llvm.uadd.with.overflow. 2563 if (I.getPredicate() == ICmpInst::ICMP_UGT && 2564 match(Op1, m_Add(m_Value(A), m_Value(B))) && 2565 (Op0 == A || Op0 == B)) 2566 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this)) 2567 return R; 2568 } 2569 2570 if (I.isEquality()) { 2571 Value *A, *B, *C, *D; 2572 2573 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 2574 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0 2575 Value *OtherVal = A == Op1 ? B : A; 2576 return new ICmpInst(I.getPredicate(), OtherVal, 2577 Constant::getNullValue(A->getType())); 2578 } 2579 2580 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) { 2581 // A^c1 == C^c2 --> A == C^(c1^c2) 2582 ConstantInt *C1, *C2; 2583 if (match(B, m_ConstantInt(C1)) && 2584 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) { 2585 Constant *NC = ConstantInt::get(I.getContext(), 2586 C1->getValue() ^ C2->getValue()); 2587 Value *Xor = Builder->CreateXor(C, NC); 2588 return new ICmpInst(I.getPredicate(), A, Xor); 2589 } 2590 2591 // A^B == A^D -> B == D 2592 if (A == C) return new ICmpInst(I.getPredicate(), B, D); 2593 if (A == D) return new ICmpInst(I.getPredicate(), B, C); 2594 if (B == C) return new ICmpInst(I.getPredicate(), A, D); 2595 if (B == D) return new ICmpInst(I.getPredicate(), A, C); 2596 } 2597 } 2598 2599 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && 2600 (A == Op0 || B == Op0)) { 2601 // A == (A^B) -> B == 0 2602 Value *OtherVal = A == Op0 ? B : A; 2603 return new ICmpInst(I.getPredicate(), OtherVal, 2604 Constant::getNullValue(A->getType())); 2605 } 2606 2607 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0 2608 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) && 2609 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) { 2610 Value *X = 0, *Y = 0, *Z = 0; 2611 2612 if (A == C) { 2613 X = B; Y = D; Z = A; 2614 } else if (A == D) { 2615 X = B; Y = C; Z = A; 2616 } else if (B == C) { 2617 X = A; Y = D; Z = B; 2618 } else if (B == D) { 2619 X = A; Y = C; Z = B; 2620 } 2621 2622 if (X) { // Build (X^Y) & Z 2623 Op1 = Builder->CreateXor(X, Y); 2624 Op1 = Builder->CreateAnd(Op1, Z); 2625 I.setOperand(0, Op1); 2626 I.setOperand(1, Constant::getNullValue(Op1->getType())); 2627 return &I; 2628 } 2629 } 2630 2631 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B) 2632 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B) 2633 ConstantInt *Cst1; 2634 if ((Op0->hasOneUse() && 2635 match(Op0, m_ZExt(m_Value(A))) && 2636 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) || 2637 (Op1->hasOneUse() && 2638 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) && 2639 match(Op1, m_ZExt(m_Value(A))))) { 2640 APInt Pow2 = Cst1->getValue() + 1; 2641 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) && 2642 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth()) 2643 return new ICmpInst(I.getPredicate(), A, 2644 Builder->CreateTrunc(B, A->getType())); 2645 } 2646 2647 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to 2648 // "icmp (and X, mask), cst" 2649 uint64_t ShAmt = 0; 2650 if (Op0->hasOneUse() && 2651 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), 2652 m_ConstantInt(ShAmt))))) && 2653 match(Op1, m_ConstantInt(Cst1)) && 2654 // Only do this when A has multiple uses. This is most important to do 2655 // when it exposes other optimizations. 2656 !A->hasOneUse()) { 2657 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits(); 2658 2659 if (ShAmt < ASize) { 2660 APInt MaskV = 2661 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits()); 2662 MaskV <<= ShAmt; 2663 2664 APInt CmpV = Cst1->getValue().zext(ASize); 2665 CmpV <<= ShAmt; 2666 2667 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV)); 2668 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV)); 2669 } 2670 } 2671 } 2672 2673 { 2674 Value *X; ConstantInt *Cst; 2675 // icmp X+Cst, X 2676 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X) 2677 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0); 2678 2679 // icmp X, X+Cst 2680 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X) 2681 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1); 2682 } 2683 return Changed ? &I : 0; 2684 } 2685 2686 2687 2688 2689 2690 2691 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible. 2692 /// 2693 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I, 2694 Instruction *LHSI, 2695 Constant *RHSC) { 2696 if (!isa<ConstantFP>(RHSC)) return 0; 2697 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF(); 2698 2699 // Get the width of the mantissa. We don't want to hack on conversions that 2700 // might lose information from the integer, e.g. "i64 -> float" 2701 int MantissaWidth = LHSI->getType()->getFPMantissaWidth(); 2702 if (MantissaWidth == -1) return 0; // Unknown. 2703 2704 // Check to see that the input is converted from an integer type that is small 2705 // enough that preserves all bits. TODO: check here for "known" sign bits. 2706 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e. 2707 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits(); 2708 2709 // If this is a uitofp instruction, we need an extra bit to hold the sign. 2710 bool LHSUnsigned = isa<UIToFPInst>(LHSI); 2711 if (LHSUnsigned) 2712 ++InputSize; 2713 2714 // If the conversion would lose info, don't hack on this. 2715 if ((int)InputSize > MantissaWidth) 2716 return 0; 2717 2718 // Otherwise, we can potentially simplify the comparison. We know that it 2719 // will always come through as an integer value and we know the constant is 2720 // not a NAN (it would have been previously simplified). 2721 assert(!RHS.isNaN() && "NaN comparison not already folded!"); 2722 2723 ICmpInst::Predicate Pred; 2724 switch (I.getPredicate()) { 2725 default: llvm_unreachable("Unexpected predicate!"); 2726 case FCmpInst::FCMP_UEQ: 2727 case FCmpInst::FCMP_OEQ: 2728 Pred = ICmpInst::ICMP_EQ; 2729 break; 2730 case FCmpInst::FCMP_UGT: 2731 case FCmpInst::FCMP_OGT: 2732 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT; 2733 break; 2734 case FCmpInst::FCMP_UGE: 2735 case FCmpInst::FCMP_OGE: 2736 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE; 2737 break; 2738 case FCmpInst::FCMP_ULT: 2739 case FCmpInst::FCMP_OLT: 2740 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT; 2741 break; 2742 case FCmpInst::FCMP_ULE: 2743 case FCmpInst::FCMP_OLE: 2744 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE; 2745 break; 2746 case FCmpInst::FCMP_UNE: 2747 case FCmpInst::FCMP_ONE: 2748 Pred = ICmpInst::ICMP_NE; 2749 break; 2750 case FCmpInst::FCMP_ORD: 2751 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2752 case FCmpInst::FCMP_UNO: 2753 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2754 } 2755 2756 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType()); 2757 2758 // Now we know that the APFloat is a normal number, zero or inf. 2759 2760 // See if the FP constant is too large for the integer. For example, 2761 // comparing an i8 to 300.0. 2762 unsigned IntWidth = IntTy->getScalarSizeInBits(); 2763 2764 if (!LHSUnsigned) { 2765 // If the RHS value is > SignedMax, fold the comparison. This handles +INF 2766 // and large values. 2767 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false); 2768 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true, 2769 APFloat::rmNearestTiesToEven); 2770 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0 2771 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT || 2772 Pred == ICmpInst::ICMP_SLE) 2773 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2774 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2775 } 2776 } else { 2777 // If the RHS value is > UnsignedMax, fold the comparison. This handles 2778 // +INF and large values. 2779 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false); 2780 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false, 2781 APFloat::rmNearestTiesToEven); 2782 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0 2783 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT || 2784 Pred == ICmpInst::ICMP_ULE) 2785 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2786 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2787 } 2788 } 2789 2790 if (!LHSUnsigned) { 2791 // See if the RHS value is < SignedMin. 2792 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false); 2793 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true, 2794 APFloat::rmNearestTiesToEven); 2795 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0 2796 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT || 2797 Pred == ICmpInst::ICMP_SGE) 2798 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2799 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2800 } 2801 } else { 2802 // See if the RHS value is < UnsignedMin. 2803 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false); 2804 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true, 2805 APFloat::rmNearestTiesToEven); 2806 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0 2807 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT || 2808 Pred == ICmpInst::ICMP_UGE) 2809 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2810 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2811 } 2812 } 2813 2814 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or 2815 // [0, UMAX], but it may still be fractional. See if it is fractional by 2816 // casting the FP value to the integer value and back, checking for equality. 2817 // Don't do this for zero, because -0.0 is not fractional. 2818 Constant *RHSInt = LHSUnsigned 2819 ? ConstantExpr::getFPToUI(RHSC, IntTy) 2820 : ConstantExpr::getFPToSI(RHSC, IntTy); 2821 if (!RHS.isZero()) { 2822 bool Equal = LHSUnsigned 2823 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC 2824 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC; 2825 if (!Equal) { 2826 // If we had a comparison against a fractional value, we have to adjust 2827 // the compare predicate and sometimes the value. RHSC is rounded towards 2828 // zero at this point. 2829 switch (Pred) { 2830 default: llvm_unreachable("Unexpected integer comparison!"); 2831 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true 2832 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2833 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false 2834 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2835 case ICmpInst::ICMP_ULE: 2836 // (float)int <= 4.4 --> int <= 4 2837 // (float)int <= -4.4 --> false 2838 if (RHS.isNegative()) 2839 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2840 break; 2841 case ICmpInst::ICMP_SLE: 2842 // (float)int <= 4.4 --> int <= 4 2843 // (float)int <= -4.4 --> int < -4 2844 if (RHS.isNegative()) 2845 Pred = ICmpInst::ICMP_SLT; 2846 break; 2847 case ICmpInst::ICMP_ULT: 2848 // (float)int < -4.4 --> false 2849 // (float)int < 4.4 --> int <= 4 2850 if (RHS.isNegative()) 2851 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2852 Pred = ICmpInst::ICMP_ULE; 2853 break; 2854 case ICmpInst::ICMP_SLT: 2855 // (float)int < -4.4 --> int < -4 2856 // (float)int < 4.4 --> int <= 4 2857 if (!RHS.isNegative()) 2858 Pred = ICmpInst::ICMP_SLE; 2859 break; 2860 case ICmpInst::ICMP_UGT: 2861 // (float)int > 4.4 --> int > 4 2862 // (float)int > -4.4 --> true 2863 if (RHS.isNegative()) 2864 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2865 break; 2866 case ICmpInst::ICMP_SGT: 2867 // (float)int > 4.4 --> int > 4 2868 // (float)int > -4.4 --> int >= -4 2869 if (RHS.isNegative()) 2870 Pred = ICmpInst::ICMP_SGE; 2871 break; 2872 case ICmpInst::ICMP_UGE: 2873 // (float)int >= -4.4 --> true 2874 // (float)int >= 4.4 --> int > 4 2875 if (RHS.isNegative()) 2876 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2877 Pred = ICmpInst::ICMP_UGT; 2878 break; 2879 case ICmpInst::ICMP_SGE: 2880 // (float)int >= -4.4 --> int >= -4 2881 // (float)int >= 4.4 --> int > 4 2882 if (!RHS.isNegative()) 2883 Pred = ICmpInst::ICMP_SGT; 2884 break; 2885 } 2886 } 2887 } 2888 2889 // Lower this FP comparison into an appropriate integer version of the 2890 // comparison. 2891 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt); 2892 } 2893 2894 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { 2895 bool Changed = false; 2896 2897 /// Orders the operands of the compare so that they are listed from most 2898 /// complex to least complex. This puts constants before unary operators, 2899 /// before binary operators. 2900 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) { 2901 I.swapOperands(); 2902 Changed = true; 2903 } 2904 2905 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2906 2907 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD)) 2908 return ReplaceInstUsesWith(I, V); 2909 2910 // Simplify 'fcmp pred X, X' 2911 if (Op0 == Op1) { 2912 switch (I.getPredicate()) { 2913 default: llvm_unreachable("Unknown predicate!"); 2914 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y) 2915 case FCmpInst::FCMP_ULT: // True if unordered or less than 2916 case FCmpInst::FCMP_UGT: // True if unordered or greater than 2917 case FCmpInst::FCMP_UNE: // True if unordered or not equal 2918 // Canonicalize these to be 'fcmp uno %X, 0.0'. 2919 I.setPredicate(FCmpInst::FCMP_UNO); 2920 I.setOperand(1, Constant::getNullValue(Op0->getType())); 2921 return &I; 2922 2923 case FCmpInst::FCMP_ORD: // True if ordered (no nans) 2924 case FCmpInst::FCMP_OEQ: // True if ordered and equal 2925 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal 2926 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal 2927 // Canonicalize these to be 'fcmp ord %X, 0.0'. 2928 I.setPredicate(FCmpInst::FCMP_ORD); 2929 I.setOperand(1, Constant::getNullValue(Op0->getType())); 2930 return &I; 2931 } 2932 } 2933 2934 // Handle fcmp with constant RHS 2935 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2936 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2937 switch (LHSI->getOpcode()) { 2938 case Instruction::FPExt: { 2939 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless 2940 FPExtInst *LHSExt = cast<FPExtInst>(LHSI); 2941 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC); 2942 if (!RHSF) 2943 break; 2944 2945 const fltSemantics *Sem; 2946 // FIXME: This shouldn't be here. 2947 if (LHSExt->getSrcTy()->isHalfTy()) 2948 Sem = &APFloat::IEEEhalf; 2949 else if (LHSExt->getSrcTy()->isFloatTy()) 2950 Sem = &APFloat::IEEEsingle; 2951 else if (LHSExt->getSrcTy()->isDoubleTy()) 2952 Sem = &APFloat::IEEEdouble; 2953 else if (LHSExt->getSrcTy()->isFP128Ty()) 2954 Sem = &APFloat::IEEEquad; 2955 else if (LHSExt->getSrcTy()->isX86_FP80Ty()) 2956 Sem = &APFloat::x87DoubleExtended; 2957 else if (LHSExt->getSrcTy()->isPPC_FP128Ty()) 2958 Sem = &APFloat::PPCDoubleDouble; 2959 else 2960 break; 2961 2962 bool Lossy; 2963 APFloat F = RHSF->getValueAPF(); 2964 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy); 2965 2966 // Avoid lossy conversions and denormals. Zero is a special case 2967 // that's OK to convert. 2968 APFloat Fabs = F; 2969 Fabs.clearSign(); 2970 if (!Lossy && 2971 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) != 2972 APFloat::cmpLessThan) || Fabs.isZero())) 2973 2974 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 2975 ConstantFP::get(RHSC->getContext(), F)); 2976 break; 2977 } 2978 case Instruction::PHI: 2979 // Only fold fcmp into the PHI if the phi and fcmp are in the same 2980 // block. If in the same block, we're encouraging jump threading. If 2981 // not, we are just pessimizing the code by making an i1 phi. 2982 if (LHSI->getParent() == I.getParent()) 2983 if (Instruction *NV = FoldOpIntoPhi(I)) 2984 return NV; 2985 break; 2986 case Instruction::SIToFP: 2987 case Instruction::UIToFP: 2988 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC)) 2989 return NV; 2990 break; 2991 case Instruction::Select: { 2992 // If either operand of the select is a constant, we can fold the 2993 // comparison into the select arms, which will cause one to be 2994 // constant folded and the select turned into a bitwise or. 2995 Value *Op1 = 0, *Op2 = 0; 2996 if (LHSI->hasOneUse()) { 2997 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) { 2998 // Fold the known value into the constant operand. 2999 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 3000 // Insert a new FCmp of the other select operand. 3001 Op2 = Builder->CreateFCmp(I.getPredicate(), 3002 LHSI->getOperand(2), RHSC, I.getName()); 3003 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) { 3004 // Fold the known value into the constant operand. 3005 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 3006 // Insert a new FCmp of the other select operand. 3007 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1), 3008 RHSC, I.getName()); 3009 } 3010 } 3011 3012 if (Op1) 3013 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 3014 break; 3015 } 3016 case Instruction::FSub: { 3017 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C 3018 Value *Op; 3019 if (match(LHSI, m_FNeg(m_Value(Op)))) 3020 return new FCmpInst(I.getSwappedPredicate(), Op, 3021 ConstantExpr::getFNeg(RHSC)); 3022 break; 3023 } 3024 case Instruction::Load: 3025 if (GetElementPtrInst *GEP = 3026 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 3027 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 3028 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 3029 !cast<LoadInst>(LHSI)->isVolatile()) 3030 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 3031 return Res; 3032 } 3033 break; 3034 case Instruction::Call: { 3035 CallInst *CI = cast<CallInst>(LHSI); 3036 LibFunc::Func Func; 3037 // Various optimization for fabs compared with zero. 3038 if (RHSC->isNullValue() && CI->getCalledFunction() && 3039 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) && 3040 TLI->has(Func)) { 3041 if (Func == LibFunc::fabs || Func == LibFunc::fabsf || 3042 Func == LibFunc::fabsl) { 3043 switch (I.getPredicate()) { 3044 default: break; 3045 // fabs(x) < 0 --> false 3046 case FCmpInst::FCMP_OLT: 3047 return ReplaceInstUsesWith(I, Builder->getFalse()); 3048 // fabs(x) > 0 --> x != 0 3049 case FCmpInst::FCMP_OGT: 3050 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), 3051 RHSC); 3052 // fabs(x) <= 0 --> x == 0 3053 case FCmpInst::FCMP_OLE: 3054 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), 3055 RHSC); 3056 // fabs(x) >= 0 --> !isnan(x) 3057 case FCmpInst::FCMP_OGE: 3058 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), 3059 RHSC); 3060 // fabs(x) == 0 --> x == 0 3061 // fabs(x) != 0 --> x != 0 3062 case FCmpInst::FCMP_OEQ: 3063 case FCmpInst::FCMP_UEQ: 3064 case FCmpInst::FCMP_ONE: 3065 case FCmpInst::FCMP_UNE: 3066 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), 3067 RHSC); 3068 } 3069 } 3070 } 3071 } 3072 } 3073 } 3074 3075 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y 3076 Value *X, *Y; 3077 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) 3078 return new FCmpInst(I.getSwappedPredicate(), X, Y); 3079 3080 // fcmp (fpext x), (fpext y) -> fcmp x, y 3081 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0)) 3082 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1)) 3083 if (LHSExt->getSrcTy() == RHSExt->getSrcTy()) 3084 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 3085 RHSExt->getOperand(0)); 3086 3087 return Changed ? &I : 0; 3088 } 3089