1 //===- InstCombineCompares.cpp --------------------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the visitICmp and visitFCmp functions. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "InstCombine.h" 15 #include "llvm/IntrinsicInst.h" 16 #include "llvm/Analysis/ConstantFolding.h" 17 #include "llvm/Analysis/InstructionSimplify.h" 18 #include "llvm/Analysis/MemoryBuiltins.h" 19 #include "llvm/Target/TargetData.h" 20 #include "llvm/Support/ConstantRange.h" 21 #include "llvm/Support/GetElementPtrTypeIterator.h" 22 #include "llvm/Support/PatternMatch.h" 23 using namespace llvm; 24 using namespace PatternMatch; 25 26 static ConstantInt *getOne(Constant *C) { 27 return ConstantInt::get(cast<IntegerType>(C->getType()), 1); 28 } 29 30 /// AddOne - Add one to a ConstantInt 31 static Constant *AddOne(Constant *C) { 32 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 33 } 34 /// SubOne - Subtract one from a ConstantInt 35 static Constant *SubOne(Constant *C) { 36 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1)); 37 } 38 39 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) { 40 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx)); 41 } 42 43 static bool HasAddOverflow(ConstantInt *Result, 44 ConstantInt *In1, ConstantInt *In2, 45 bool IsSigned) { 46 if (!IsSigned) 47 return Result->getValue().ult(In1->getValue()); 48 49 if (In2->isNegative()) 50 return Result->getValue().sgt(In1->getValue()); 51 return Result->getValue().slt(In1->getValue()); 52 } 53 54 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result 55 /// overflowed for this type. 56 static bool AddWithOverflow(Constant *&Result, Constant *In1, 57 Constant *In2, bool IsSigned = false) { 58 Result = ConstantExpr::getAdd(In1, In2); 59 60 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 61 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 62 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 63 if (HasAddOverflow(ExtractElement(Result, Idx), 64 ExtractElement(In1, Idx), 65 ExtractElement(In2, Idx), 66 IsSigned)) 67 return true; 68 } 69 return false; 70 } 71 72 return HasAddOverflow(cast<ConstantInt>(Result), 73 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 74 IsSigned); 75 } 76 77 static bool HasSubOverflow(ConstantInt *Result, 78 ConstantInt *In1, ConstantInt *In2, 79 bool IsSigned) { 80 if (!IsSigned) 81 return Result->getValue().ugt(In1->getValue()); 82 83 if (In2->isNegative()) 84 return Result->getValue().slt(In1->getValue()); 85 86 return Result->getValue().sgt(In1->getValue()); 87 } 88 89 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result 90 /// overflowed for this type. 91 static bool SubWithOverflow(Constant *&Result, Constant *In1, 92 Constant *In2, bool IsSigned = false) { 93 Result = ConstantExpr::getSub(In1, In2); 94 95 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 96 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 97 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 98 if (HasSubOverflow(ExtractElement(Result, Idx), 99 ExtractElement(In1, Idx), 100 ExtractElement(In2, Idx), 101 IsSigned)) 102 return true; 103 } 104 return false; 105 } 106 107 return HasSubOverflow(cast<ConstantInt>(Result), 108 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 109 IsSigned); 110 } 111 112 /// isSignBitCheck - Given an exploded icmp instruction, return true if the 113 /// comparison only checks the sign bit. If it only checks the sign bit, set 114 /// TrueIfSigned if the result of the comparison is true when the input value is 115 /// signed. 116 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS, 117 bool &TrueIfSigned) { 118 switch (pred) { 119 case ICmpInst::ICMP_SLT: // True if LHS s< 0 120 TrueIfSigned = true; 121 return RHS->isZero(); 122 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1 123 TrueIfSigned = true; 124 return RHS->isAllOnesValue(); 125 case ICmpInst::ICMP_SGT: // True if LHS s> -1 126 TrueIfSigned = false; 127 return RHS->isAllOnesValue(); 128 case ICmpInst::ICMP_UGT: 129 // True if LHS u> RHS and RHS == high-bit-mask - 1 130 TrueIfSigned = true; 131 return RHS->isMaxValue(true); 132 case ICmpInst::ICMP_UGE: 133 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc) 134 TrueIfSigned = true; 135 return RHS->getValue().isSignBit(); 136 default: 137 return false; 138 } 139 } 140 141 // isHighOnes - Return true if the constant is of the form 1+0+. 142 // This is the same as lowones(~X). 143 static bool isHighOnes(const ConstantInt *CI) { 144 return (~CI->getValue() + 1).isPowerOf2(); 145 } 146 147 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a 148 /// set of known zero and one bits, compute the maximum and minimum values that 149 /// could have the specified known zero and known one bits, returning them in 150 /// min/max. 151 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero, 152 const APInt& KnownOne, 153 APInt& Min, APInt& Max) { 154 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 155 KnownZero.getBitWidth() == Min.getBitWidth() && 156 KnownZero.getBitWidth() == Max.getBitWidth() && 157 "KnownZero, KnownOne and Min, Max must have equal bitwidth."); 158 APInt UnknownBits = ~(KnownZero|KnownOne); 159 160 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign 161 // bit if it is unknown. 162 Min = KnownOne; 163 Max = KnownOne|UnknownBits; 164 165 if (UnknownBits.isNegative()) { // Sign bit is unknown 166 Min.setBit(Min.getBitWidth()-1); 167 Max.clearBit(Max.getBitWidth()-1); 168 } 169 } 170 171 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and 172 // a set of known zero and one bits, compute the maximum and minimum values that 173 // could have the specified known zero and known one bits, returning them in 174 // min/max. 175 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero, 176 const APInt &KnownOne, 177 APInt &Min, APInt &Max) { 178 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 179 KnownZero.getBitWidth() == Min.getBitWidth() && 180 KnownZero.getBitWidth() == Max.getBitWidth() && 181 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."); 182 APInt UnknownBits = ~(KnownZero|KnownOne); 183 184 // The minimum value is when the unknown bits are all zeros. 185 Min = KnownOne; 186 // The maximum value is when the unknown bits are all ones. 187 Max = KnownOne|UnknownBits; 188 } 189 190 191 192 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern: 193 /// cmp pred (load (gep GV, ...)), cmpcst 194 /// where GV is a global variable with a constant initializer. Try to simplify 195 /// this into some simple computation that does not need the load. For example 196 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3". 197 /// 198 /// If AndCst is non-null, then the loaded value is masked with that constant 199 /// before doing the comparison. This handles cases like "A[i]&4 == 0". 200 Instruction *InstCombiner:: 201 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV, 202 CmpInst &ICI, ConstantInt *AndCst) { 203 // We need TD information to know the pointer size unless this is inbounds. 204 if (!GEP->isInBounds() && TD == 0) return 0; 205 206 Constant *Init = GV->getInitializer(); 207 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init)) 208 return 0; 209 210 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements(); 211 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays. 212 213 // There are many forms of this optimization we can handle, for now, just do 214 // the simple index into a single-dimensional array. 215 // 216 // Require: GEP GV, 0, i {{, constant indices}} 217 if (GEP->getNumOperands() < 3 || 218 !isa<ConstantInt>(GEP->getOperand(1)) || 219 !cast<ConstantInt>(GEP->getOperand(1))->isZero() || 220 isa<Constant>(GEP->getOperand(2))) 221 return 0; 222 223 // Check that indices after the variable are constants and in-range for the 224 // type they index. Collect the indices. This is typically for arrays of 225 // structs. 226 SmallVector<unsigned, 4> LaterIndices; 227 228 Type *EltTy = Init->getType()->getArrayElementType(); 229 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { 230 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i)); 231 if (Idx == 0) return 0; // Variable index. 232 233 uint64_t IdxVal = Idx->getZExtValue(); 234 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index. 235 236 if (StructType *STy = dyn_cast<StructType>(EltTy)) 237 EltTy = STy->getElementType(IdxVal); 238 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) { 239 if (IdxVal >= ATy->getNumElements()) return 0; 240 EltTy = ATy->getElementType(); 241 } else { 242 return 0; // Unknown type. 243 } 244 245 LaterIndices.push_back(IdxVal); 246 } 247 248 enum { Overdefined = -3, Undefined = -2 }; 249 250 // Variables for our state machines. 251 252 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form 253 // "i == 47 | i == 87", where 47 is the first index the condition is true for, 254 // and 87 is the second (and last) index. FirstTrueElement is -2 when 255 // undefined, otherwise set to the first true element. SecondTrueElement is 256 // -2 when undefined, -3 when overdefined and >= 0 when that index is true. 257 int FirstTrueElement = Undefined, SecondTrueElement = Undefined; 258 259 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the 260 // form "i != 47 & i != 87". Same state transitions as for true elements. 261 int FirstFalseElement = Undefined, SecondFalseElement = Undefined; 262 263 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these 264 /// define a state machine that triggers for ranges of values that the index 265 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. 266 /// This is -2 when undefined, -3 when overdefined, and otherwise the last 267 /// index in the range (inclusive). We use -2 for undefined here because we 268 /// use relative comparisons and don't want 0-1 to match -1. 269 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; 270 271 // MagicBitvector - This is a magic bitvector where we set a bit if the 272 // comparison is true for element 'i'. If there are 64 elements or less in 273 // the array, this will fully represent all the comparison results. 274 uint64_t MagicBitvector = 0; 275 276 277 // Scan the array and see if one of our patterns matches. 278 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1)); 279 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) { 280 Constant *Elt = Init->getAggregateElement(i); 281 if (Elt == 0) return 0; 282 283 // If this is indexing an array of structures, get the structure element. 284 if (!LaterIndices.empty()) 285 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices); 286 287 // If the element is masked, handle it. 288 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst); 289 290 // Find out if the comparison would be true or false for the i'th element. 291 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt, 292 CompareRHS, TD, TLI); 293 // If the result is undef for this element, ignore it. 294 if (isa<UndefValue>(C)) { 295 // Extend range state machines to cover this element in case there is an 296 // undef in the middle of the range. 297 if (TrueRangeEnd == (int)i-1) 298 TrueRangeEnd = i; 299 if (FalseRangeEnd == (int)i-1) 300 FalseRangeEnd = i; 301 continue; 302 } 303 304 // If we can't compute the result for any of the elements, we have to give 305 // up evaluating the entire conditional. 306 if (!isa<ConstantInt>(C)) return 0; 307 308 // Otherwise, we know if the comparison is true or false for this element, 309 // update our state machines. 310 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero(); 311 312 // State machine for single/double/range index comparison. 313 if (IsTrueForElt) { 314 // Update the TrueElement state machine. 315 if (FirstTrueElement == Undefined) 316 FirstTrueElement = TrueRangeEnd = i; // First true element. 317 else { 318 // Update double-compare state machine. 319 if (SecondTrueElement == Undefined) 320 SecondTrueElement = i; 321 else 322 SecondTrueElement = Overdefined; 323 324 // Update range state machine. 325 if (TrueRangeEnd == (int)i-1) 326 TrueRangeEnd = i; 327 else 328 TrueRangeEnd = Overdefined; 329 } 330 } else { 331 // Update the FalseElement state machine. 332 if (FirstFalseElement == Undefined) 333 FirstFalseElement = FalseRangeEnd = i; // First false element. 334 else { 335 // Update double-compare state machine. 336 if (SecondFalseElement == Undefined) 337 SecondFalseElement = i; 338 else 339 SecondFalseElement = Overdefined; 340 341 // Update range state machine. 342 if (FalseRangeEnd == (int)i-1) 343 FalseRangeEnd = i; 344 else 345 FalseRangeEnd = Overdefined; 346 } 347 } 348 349 350 // If this element is in range, update our magic bitvector. 351 if (i < 64 && IsTrueForElt) 352 MagicBitvector |= 1ULL << i; 353 354 // If all of our states become overdefined, bail out early. Since the 355 // predicate is expensive, only check it every 8 elements. This is only 356 // really useful for really huge arrays. 357 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && 358 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && 359 FalseRangeEnd == Overdefined) 360 return 0; 361 } 362 363 // Now that we've scanned the entire array, emit our new comparison(s). We 364 // order the state machines in complexity of the generated code. 365 Value *Idx = GEP->getOperand(2); 366 367 // If the index is larger than the pointer size of the target, truncate the 368 // index down like the GEP would do implicitly. We don't have to do this for 369 // an inbounds GEP because the index can't be out of range. 370 if (!GEP->isInBounds() && 371 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits()) 372 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext())); 373 374 // If the comparison is only true for one or two elements, emit direct 375 // comparisons. 376 if (SecondTrueElement != Overdefined) { 377 // None true -> false. 378 if (FirstTrueElement == Undefined) 379 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext())); 380 381 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement); 382 383 // True for one element -> 'i == 47'. 384 if (SecondTrueElement == Undefined) 385 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); 386 387 // True for two elements -> 'i == 47 | i == 72'. 388 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx); 389 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement); 390 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx); 391 return BinaryOperator::CreateOr(C1, C2); 392 } 393 394 // If the comparison is only false for one or two elements, emit direct 395 // comparisons. 396 if (SecondFalseElement != Overdefined) { 397 // None false -> true. 398 if (FirstFalseElement == Undefined) 399 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext())); 400 401 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement); 402 403 // False for one element -> 'i != 47'. 404 if (SecondFalseElement == Undefined) 405 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); 406 407 // False for two elements -> 'i != 47 & i != 72'. 408 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx); 409 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement); 410 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx); 411 return BinaryOperator::CreateAnd(C1, C2); 412 } 413 414 // If the comparison can be replaced with a range comparison for the elements 415 // where it is true, emit the range check. 416 if (TrueRangeEnd != Overdefined) { 417 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare"); 418 419 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1). 420 if (FirstTrueElement) { 421 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement); 422 Idx = Builder->CreateAdd(Idx, Offs); 423 } 424 425 Value *End = ConstantInt::get(Idx->getType(), 426 TrueRangeEnd-FirstTrueElement+1); 427 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); 428 } 429 430 // False range check. 431 if (FalseRangeEnd != Overdefined) { 432 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare"); 433 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). 434 if (FirstFalseElement) { 435 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement); 436 Idx = Builder->CreateAdd(Idx, Offs); 437 } 438 439 Value *End = ConstantInt::get(Idx->getType(), 440 FalseRangeEnd-FirstFalseElement); 441 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); 442 } 443 444 445 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state 446 // of this load, replace it with computation that does: 447 // ((magic_cst >> i) & 1) != 0 448 if (ArrayElementCount <= 32 || 449 (TD && ArrayElementCount <= 64 && TD->isLegalInteger(64))) { 450 Type *Ty; 451 if (ArrayElementCount <= 32) 452 Ty = Type::getInt32Ty(Init->getContext()); 453 else 454 Ty = Type::getInt64Ty(Init->getContext()); 455 Value *V = Builder->CreateIntCast(Idx, Ty, false); 456 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V); 457 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V); 458 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0)); 459 } 460 461 return 0; 462 } 463 464 465 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare 466 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we 467 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can 468 /// be complex, and scales are involved. The above expression would also be 469 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). 470 /// This later form is less amenable to optimization though, and we are allowed 471 /// to generate the first by knowing that pointer arithmetic doesn't overflow. 472 /// 473 /// If we can't emit an optimized form for this expression, this returns null. 474 /// 475 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) { 476 TargetData &TD = *IC.getTargetData(); 477 gep_type_iterator GTI = gep_type_begin(GEP); 478 479 // Check to see if this gep only has a single variable index. If so, and if 480 // any constant indices are a multiple of its scale, then we can compute this 481 // in terms of the scale of the variable index. For example, if the GEP 482 // implies an offset of "12 + i*4", then we can codegen this as "3 + i", 483 // because the expression will cross zero at the same point. 484 unsigned i, e = GEP->getNumOperands(); 485 int64_t Offset = 0; 486 for (i = 1; i != e; ++i, ++GTI) { 487 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 488 // Compute the aggregate offset of constant indices. 489 if (CI->isZero()) continue; 490 491 // Handle a struct index, which adds its field offset to the pointer. 492 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 493 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 494 } else { 495 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 496 Offset += Size*CI->getSExtValue(); 497 } 498 } else { 499 // Found our variable index. 500 break; 501 } 502 } 503 504 // If there are no variable indices, we must have a constant offset, just 505 // evaluate it the general way. 506 if (i == e) return 0; 507 508 Value *VariableIdx = GEP->getOperand(i); 509 // Determine the scale factor of the variable element. For example, this is 510 // 4 if the variable index is into an array of i32. 511 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType()); 512 513 // Verify that there are no other variable indices. If so, emit the hard way. 514 for (++i, ++GTI; i != e; ++i, ++GTI) { 515 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i)); 516 if (!CI) return 0; 517 518 // Compute the aggregate offset of constant indices. 519 if (CI->isZero()) continue; 520 521 // Handle a struct index, which adds its field offset to the pointer. 522 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 523 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 524 } else { 525 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 526 Offset += Size*CI->getSExtValue(); 527 } 528 } 529 530 // Okay, we know we have a single variable index, which must be a 531 // pointer/array/vector index. If there is no offset, life is simple, return 532 // the index. 533 unsigned IntPtrWidth = TD.getPointerSizeInBits(); 534 if (Offset == 0) { 535 // Cast to intptrty in case a truncation occurs. If an extension is needed, 536 // we don't need to bother extending: the extension won't affect where the 537 // computation crosses zero. 538 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) { 539 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext()); 540 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy); 541 } 542 return VariableIdx; 543 } 544 545 // Otherwise, there is an index. The computation we will do will be modulo 546 // the pointer size, so get it. 547 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth); 548 549 Offset &= PtrSizeMask; 550 VariableScale &= PtrSizeMask; 551 552 // To do this transformation, any constant index must be a multiple of the 553 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i", 554 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a 555 // multiple of the variable scale. 556 int64_t NewOffs = Offset / (int64_t)VariableScale; 557 if (Offset != NewOffs*(int64_t)VariableScale) 558 return 0; 559 560 // Okay, we can do this evaluation. Start by converting the index to intptr. 561 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext()); 562 if (VariableIdx->getType() != IntPtrTy) 563 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy, 564 true /*Signed*/); 565 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs); 566 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset"); 567 } 568 569 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something 570 /// else. At this point we know that the GEP is on the LHS of the comparison. 571 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 572 ICmpInst::Predicate Cond, 573 Instruction &I) { 574 // Don't transform signed compares of GEPs into index compares. Even if the 575 // GEP is inbounds, the final add of the base pointer can have signed overflow 576 // and would change the result of the icmp. 577 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be 578 // the maximum signed value for the pointer type. 579 if (ICmpInst::isSigned(Cond)) 580 return 0; 581 582 // Look through bitcasts. 583 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS)) 584 RHS = BCI->getOperand(0); 585 586 Value *PtrBase = GEPLHS->getOperand(0); 587 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) { 588 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). 589 // This transformation (ignoring the base and scales) is valid because we 590 // know pointers can't overflow since the gep is inbounds. See if we can 591 // output an optimized form. 592 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this); 593 594 // If not, synthesize the offset the hard way. 595 if (Offset == 0) 596 Offset = EmitGEPOffset(GEPLHS); 597 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset, 598 Constant::getNullValue(Offset->getType())); 599 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) { 600 // If the base pointers are different, but the indices are the same, just 601 // compare the base pointer. 602 if (PtrBase != GEPRHS->getOperand(0)) { 603 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands(); 604 IndicesTheSame &= GEPLHS->getOperand(0)->getType() == 605 GEPRHS->getOperand(0)->getType(); 606 if (IndicesTheSame) 607 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 608 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 609 IndicesTheSame = false; 610 break; 611 } 612 613 // If all indices are the same, just compare the base pointers. 614 if (IndicesTheSame) 615 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), 616 GEPLHS->getOperand(0), GEPRHS->getOperand(0)); 617 618 // If we're comparing GEPs with two base pointers that only differ in type 619 // and both GEPs have only constant indices or just one use, then fold 620 // the compare with the adjusted indices. 621 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() && 622 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) && 623 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) && 624 PtrBase->stripPointerCasts() == 625 GEPRHS->getOperand(0)->stripPointerCasts()) { 626 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond), 627 EmitGEPOffset(GEPLHS), 628 EmitGEPOffset(GEPRHS)); 629 return ReplaceInstUsesWith(I, Cmp); 630 } 631 632 // Otherwise, the base pointers are different and the indices are 633 // different, bail out. 634 return 0; 635 } 636 637 // If one of the GEPs has all zero indices, recurse. 638 bool AllZeros = true; 639 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 640 if (!isa<Constant>(GEPLHS->getOperand(i)) || 641 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) { 642 AllZeros = false; 643 break; 644 } 645 if (AllZeros) 646 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0), 647 ICmpInst::getSwappedPredicate(Cond), I); 648 649 // If the other GEP has all zero indices, recurse. 650 AllZeros = true; 651 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 652 if (!isa<Constant>(GEPRHS->getOperand(i)) || 653 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) { 654 AllZeros = false; 655 break; 656 } 657 if (AllZeros) 658 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I); 659 660 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds(); 661 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) { 662 // If the GEPs only differ by one index, compare it. 663 unsigned NumDifferences = 0; // Keep track of # differences. 664 unsigned DiffOperand = 0; // The operand that differs. 665 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 666 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 667 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() != 668 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) { 669 // Irreconcilable differences. 670 NumDifferences = 2; 671 break; 672 } else { 673 if (NumDifferences++) break; 674 DiffOperand = i; 675 } 676 } 677 678 if (NumDifferences == 0) // SAME GEP? 679 return ReplaceInstUsesWith(I, // No comparison is needed here. 680 ConstantInt::get(Type::getInt1Ty(I.getContext()), 681 ICmpInst::isTrueWhenEqual(Cond))); 682 683 else if (NumDifferences == 1 && GEPsInBounds) { 684 Value *LHSV = GEPLHS->getOperand(DiffOperand); 685 Value *RHSV = GEPRHS->getOperand(DiffOperand); 686 // Make sure we do a signed comparison here. 687 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV); 688 } 689 } 690 691 // Only lower this if the icmp is the only user of the GEP or if we expect 692 // the result to fold to a constant! 693 if (TD && 694 GEPsInBounds && 695 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) && 696 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) { 697 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) 698 Value *L = EmitGEPOffset(GEPLHS); 699 Value *R = EmitGEPOffset(GEPRHS); 700 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R); 701 } 702 } 703 return 0; 704 } 705 706 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X". 707 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI, 708 Value *X, ConstantInt *CI, 709 ICmpInst::Predicate Pred, 710 Value *TheAdd) { 711 // If we have X+0, exit early (simplifying logic below) and let it get folded 712 // elsewhere. icmp X+0, X -> icmp X, X 713 if (CI->isZero()) { 714 bool isTrue = ICmpInst::isTrueWhenEqual(Pred); 715 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 716 } 717 718 // (X+4) == X -> false. 719 if (Pred == ICmpInst::ICMP_EQ) 720 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext())); 721 722 // (X+4) != X -> true. 723 if (Pred == ICmpInst::ICMP_NE) 724 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext())); 725 726 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, 727 // so the values can never be equal. Similarly for all other "or equals" 728 // operators. 729 730 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255 731 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253 732 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0 733 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 734 Value *R = 735 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI); 736 return new ICmpInst(ICmpInst::ICMP_UGT, X, R); 737 } 738 739 // (X+1) >u X --> X <u (0-1) --> X != 255 740 // (X+2) >u X --> X <u (0-2) --> X <u 254 741 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0 742 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) 743 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI)); 744 745 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits(); 746 ConstantInt *SMax = ConstantInt::get(X->getContext(), 747 APInt::getSignedMaxValue(BitWidth)); 748 749 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127 750 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125 751 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0 752 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1 753 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126 754 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127 755 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 756 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI)); 757 758 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127 759 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126 760 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1 761 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2 762 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126 763 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128 764 765 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); 766 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1); 767 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C)); 768 } 769 770 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS 771 /// and CmpRHS are both known to be integer constants. 772 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI, 773 ConstantInt *DivRHS) { 774 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1)); 775 const APInt &CmpRHSV = CmpRHS->getValue(); 776 777 // FIXME: If the operand types don't match the type of the divide 778 // then don't attempt this transform. The code below doesn't have the 779 // logic to deal with a signed divide and an unsigned compare (and 780 // vice versa). This is because (x /s C1) <s C2 produces different 781 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even 782 // (x /u C1) <u C2. Simply casting the operands and result won't 783 // work. :( The if statement below tests that condition and bails 784 // if it finds it. 785 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv; 786 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned()) 787 return 0; 788 if (DivRHS->isZero()) 789 return 0; // The ProdOV computation fails on divide by zero. 790 if (DivIsSigned && DivRHS->isAllOnesValue()) 791 return 0; // The overflow computation also screws up here 792 if (DivRHS->isOne()) { 793 // This eliminates some funny cases with INT_MIN. 794 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X. 795 return &ICI; 796 } 797 798 // Compute Prod = CI * DivRHS. We are essentially solving an equation 799 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and 800 // C2 (CI). By solving for X we can turn this into a range check 801 // instead of computing a divide. 802 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS); 803 804 // Determine if the product overflows by seeing if the product is 805 // not equal to the divide. Make sure we do the same kind of divide 806 // as in the LHS instruction that we're folding. 807 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) : 808 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS; 809 810 // Get the ICmp opcode 811 ICmpInst::Predicate Pred = ICI.getPredicate(); 812 813 /// If the division is known to be exact, then there is no remainder from the 814 /// divide, so the covered range size is unit, otherwise it is the divisor. 815 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS; 816 817 // Figure out the interval that is being checked. For example, a comparison 818 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). 819 // Compute this interval based on the constants involved and the signedness of 820 // the compare/divide. This computes a half-open interval, keeping track of 821 // whether either value in the interval overflows. After analysis each 822 // overflow variable is set to 0 if it's corresponding bound variable is valid 823 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. 824 int LoOverflow = 0, HiOverflow = 0; 825 Constant *LoBound = 0, *HiBound = 0; 826 827 if (!DivIsSigned) { // udiv 828 // e.g. X/5 op 3 --> [15, 20) 829 LoBound = Prod; 830 HiOverflow = LoOverflow = ProdOV; 831 if (!HiOverflow) { 832 // If this is not an exact divide, then many values in the range collapse 833 // to the same result value. 834 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false); 835 } 836 837 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0. 838 if (CmpRHSV == 0) { // (X / pos) op 0 839 // Can't overflow. e.g. X/2 op 0 --> [-1, 2) 840 LoBound = ConstantExpr::getNeg(SubOne(RangeSize)); 841 HiBound = RangeSize; 842 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos 843 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) 844 HiOverflow = LoOverflow = ProdOV; 845 if (!HiOverflow) 846 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true); 847 } else { // (X / pos) op neg 848 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) 849 HiBound = AddOne(Prod); 850 LoOverflow = HiOverflow = ProdOV ? -1 : 0; 851 if (!LoOverflow) { 852 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 853 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0; 854 } 855 } 856 } else if (DivRHS->isNegative()) { // Divisor is < 0. 857 if (DivI->isExact()) 858 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 859 if (CmpRHSV == 0) { // (X / neg) op 0 860 // e.g. X/-5 op 0 --> [-4, 5) 861 LoBound = AddOne(RangeSize); 862 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 863 if (HiBound == DivRHS) { // -INTMIN = INTMIN 864 HiOverflow = 1; // [INTMIN+1, overflow) 865 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN 866 } 867 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos 868 // e.g. X/-5 op 3 --> [-19, -14) 869 HiBound = AddOne(Prod); 870 HiOverflow = LoOverflow = ProdOV ? -1 : 0; 871 if (!LoOverflow) 872 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0; 873 } else { // (X / neg) op neg 874 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) 875 LoOverflow = HiOverflow = ProdOV; 876 if (!HiOverflow) 877 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true); 878 } 879 880 // Dividing by a negative swaps the condition. LT <-> GT 881 Pred = ICmpInst::getSwappedPredicate(Pred); 882 } 883 884 Value *X = DivI->getOperand(0); 885 switch (Pred) { 886 default: llvm_unreachable("Unhandled icmp opcode!"); 887 case ICmpInst::ICMP_EQ: 888 if (LoOverflow && HiOverflow) 889 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 890 if (HiOverflow) 891 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 892 ICmpInst::ICMP_UGE, X, LoBound); 893 if (LoOverflow) 894 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 895 ICmpInst::ICMP_ULT, X, HiBound); 896 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 897 DivIsSigned, true)); 898 case ICmpInst::ICMP_NE: 899 if (LoOverflow && HiOverflow) 900 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 901 if (HiOverflow) 902 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 903 ICmpInst::ICMP_ULT, X, LoBound); 904 if (LoOverflow) 905 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 906 ICmpInst::ICMP_UGE, X, HiBound); 907 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 908 DivIsSigned, false)); 909 case ICmpInst::ICMP_ULT: 910 case ICmpInst::ICMP_SLT: 911 if (LoOverflow == +1) // Low bound is greater than input range. 912 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 913 if (LoOverflow == -1) // Low bound is less than input range. 914 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 915 return new ICmpInst(Pred, X, LoBound); 916 case ICmpInst::ICMP_UGT: 917 case ICmpInst::ICMP_SGT: 918 if (HiOverflow == +1) // High bound greater than input range. 919 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext())); 920 if (HiOverflow == -1) // High bound less than input range. 921 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext())); 922 if (Pred == ICmpInst::ICMP_UGT) 923 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound); 924 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound); 925 } 926 } 927 928 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)". 929 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr, 930 ConstantInt *ShAmt) { 931 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue(); 932 933 // Check that the shift amount is in range. If not, don't perform 934 // undefined shifts. When the shift is visited it will be 935 // simplified. 936 uint32_t TypeBits = CmpRHSV.getBitWidth(); 937 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 938 if (ShAmtVal >= TypeBits || ShAmtVal == 0) 939 return 0; 940 941 if (!ICI.isEquality()) { 942 // If we have an unsigned comparison and an ashr, we can't simplify this. 943 // Similarly for signed comparisons with lshr. 944 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr)) 945 return 0; 946 947 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv 948 // by a power of 2. Since we already have logic to simplify these, 949 // transform to div and then simplify the resultant comparison. 950 if (Shr->getOpcode() == Instruction::AShr && 951 (!Shr->isExact() || ShAmtVal == TypeBits - 1)) 952 return 0; 953 954 // Revisit the shift (to delete it). 955 Worklist.Add(Shr); 956 957 Constant *DivCst = 958 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal)); 959 960 Value *Tmp = 961 Shr->getOpcode() == Instruction::AShr ? 962 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) : 963 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()); 964 965 ICI.setOperand(0, Tmp); 966 967 // If the builder folded the binop, just return it. 968 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp); 969 if (TheDiv == 0) 970 return &ICI; 971 972 // Otherwise, fold this div/compare. 973 assert(TheDiv->getOpcode() == Instruction::SDiv || 974 TheDiv->getOpcode() == Instruction::UDiv); 975 976 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst)); 977 assert(Res && "This div/cst should have folded!"); 978 return Res; 979 } 980 981 982 // If we are comparing against bits always shifted out, the 983 // comparison cannot succeed. 984 APInt Comp = CmpRHSV << ShAmtVal; 985 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp); 986 if (Shr->getOpcode() == Instruction::LShr) 987 Comp = Comp.lshr(ShAmtVal); 988 else 989 Comp = Comp.ashr(ShAmtVal); 990 991 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero. 992 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 993 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 994 IsICMP_NE); 995 return ReplaceInstUsesWith(ICI, Cst); 996 } 997 998 // Otherwise, check to see if the bits shifted out are known to be zero. 999 // If so, we can compare against the unshifted value: 1000 // (X & 4) >> 1 == 2 --> (X & 4) == 4. 1001 if (Shr->hasOneUse() && Shr->isExact()) 1002 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS); 1003 1004 if (Shr->hasOneUse()) { 1005 // Otherwise strength reduce the shift into an and. 1006 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal)); 1007 Constant *Mask = ConstantInt::get(ICI.getContext(), Val); 1008 1009 Value *And = Builder->CreateAnd(Shr->getOperand(0), 1010 Mask, Shr->getName()+".mask"); 1011 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS); 1012 } 1013 return 0; 1014 } 1015 1016 1017 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)". 1018 /// 1019 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI, 1020 Instruction *LHSI, 1021 ConstantInt *RHS) { 1022 const APInt &RHSV = RHS->getValue(); 1023 1024 switch (LHSI->getOpcode()) { 1025 case Instruction::Trunc: 1026 if (ICI.isEquality() && LHSI->hasOneUse()) { 1027 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all 1028 // of the high bits truncated out of x are known. 1029 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(), 1030 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1031 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0); 1032 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne); 1033 1034 // If all the high bits are known, we can do this xform. 1035 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) { 1036 // Pull in the high bits from known-ones set. 1037 APInt NewRHS = RHS->getValue().zext(SrcBits); 1038 NewRHS |= KnownOne; 1039 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1040 ConstantInt::get(ICI.getContext(), NewRHS)); 1041 } 1042 } 1043 break; 1044 1045 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI) 1046 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1047 // If this is a comparison that tests the signbit (X < 0) or (x > -1), 1048 // fold the xor. 1049 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) || 1050 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) { 1051 Value *CompareVal = LHSI->getOperand(0); 1052 1053 // If the sign bit of the XorCST is not set, there is no change to 1054 // the operation, just stop using the Xor. 1055 if (!XorCST->isNegative()) { 1056 ICI.setOperand(0, CompareVal); 1057 Worklist.Add(LHSI); 1058 return &ICI; 1059 } 1060 1061 // Was the old condition true if the operand is positive? 1062 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT; 1063 1064 // If so, the new one isn't. 1065 isTrueIfPositive ^= true; 1066 1067 if (isTrueIfPositive) 1068 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, 1069 SubOne(RHS)); 1070 else 1071 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, 1072 AddOne(RHS)); 1073 } 1074 1075 if (LHSI->hasOneUse()) { 1076 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit)) 1077 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) { 1078 const APInt &SignBit = XorCST->getValue(); 1079 ICmpInst::Predicate Pred = ICI.isSigned() 1080 ? ICI.getUnsignedPredicate() 1081 : ICI.getSignedPredicate(); 1082 return new ICmpInst(Pred, LHSI->getOperand(0), 1083 ConstantInt::get(ICI.getContext(), 1084 RHSV ^ SignBit)); 1085 } 1086 1087 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A) 1088 if (!ICI.isEquality() && XorCST->isMaxValue(true)) { 1089 const APInt &NotSignBit = XorCST->getValue(); 1090 ICmpInst::Predicate Pred = ICI.isSigned() 1091 ? ICI.getUnsignedPredicate() 1092 : ICI.getSignedPredicate(); 1093 Pred = ICI.getSwappedPredicate(Pred); 1094 return new ICmpInst(Pred, LHSI->getOperand(0), 1095 ConstantInt::get(ICI.getContext(), 1096 RHSV ^ NotSignBit)); 1097 } 1098 } 1099 } 1100 break; 1101 case Instruction::And: // (icmp pred (and X, AndCST), RHS) 1102 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) && 1103 LHSI->getOperand(0)->hasOneUse()) { 1104 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1)); 1105 1106 // If the LHS is an AND of a truncating cast, we can widen the 1107 // and/compare to be the input width without changing the value 1108 // produced, eliminating a cast. 1109 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) { 1110 // We can do this transformation if either the AND constant does not 1111 // have its sign bit set or if it is an equality comparison. 1112 // Extending a relational comparison when we're checking the sign 1113 // bit would not work. 1114 if (ICI.isEquality() || 1115 (!AndCST->isNegative() && RHSV.isNonNegative())) { 1116 Value *NewAnd = 1117 Builder->CreateAnd(Cast->getOperand(0), 1118 ConstantExpr::getZExt(AndCST, Cast->getSrcTy())); 1119 NewAnd->takeName(LHSI); 1120 return new ICmpInst(ICI.getPredicate(), NewAnd, 1121 ConstantExpr::getZExt(RHS, Cast->getSrcTy())); 1122 } 1123 } 1124 1125 // If the LHS is an AND of a zext, and we have an equality compare, we can 1126 // shrink the and/compare to the smaller type, eliminating the cast. 1127 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) { 1128 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy()); 1129 // Make sure we don't compare the upper bits, SimplifyDemandedBits 1130 // should fold the icmp to true/false in that case. 1131 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) { 1132 Value *NewAnd = 1133 Builder->CreateAnd(Cast->getOperand(0), 1134 ConstantExpr::getTrunc(AndCST, Ty)); 1135 NewAnd->takeName(LHSI); 1136 return new ICmpInst(ICI.getPredicate(), NewAnd, 1137 ConstantExpr::getTrunc(RHS, Ty)); 1138 } 1139 } 1140 1141 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare 1142 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This 1143 // happens a LOT in code produced by the C front-end, for bitfield 1144 // access. 1145 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0)); 1146 if (Shift && !Shift->isShift()) 1147 Shift = 0; 1148 1149 ConstantInt *ShAmt; 1150 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0; 1151 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift. 1152 Type *AndTy = AndCST->getType(); // Type of the and. 1153 1154 // We can fold this as long as we can't shift unknown bits 1155 // into the mask. This can only happen with signed shift 1156 // rights, as they sign-extend. 1157 if (ShAmt) { 1158 bool CanFold = Shift->isLogicalShift(); 1159 if (!CanFold) { 1160 // To test for the bad case of the signed shr, see if any 1161 // of the bits shifted in could be tested after the mask. 1162 uint32_t TyBits = Ty->getPrimitiveSizeInBits(); 1163 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits); 1164 1165 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits(); 1166 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) & 1167 AndCST->getValue()) == 0) 1168 CanFold = true; 1169 } 1170 1171 if (CanFold) { 1172 Constant *NewCst; 1173 if (Shift->getOpcode() == Instruction::Shl) 1174 NewCst = ConstantExpr::getLShr(RHS, ShAmt); 1175 else 1176 NewCst = ConstantExpr::getShl(RHS, ShAmt); 1177 1178 // Check to see if we are shifting out any of the bits being 1179 // compared. 1180 if (ConstantExpr::get(Shift->getOpcode(), 1181 NewCst, ShAmt) != RHS) { 1182 // If we shifted bits out, the fold is not going to work out. 1183 // As a special case, check to see if this means that the 1184 // result is always true or false now. 1185 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1186 return ReplaceInstUsesWith(ICI, 1187 ConstantInt::getFalse(ICI.getContext())); 1188 if (ICI.getPredicate() == ICmpInst::ICMP_NE) 1189 return ReplaceInstUsesWith(ICI, 1190 ConstantInt::getTrue(ICI.getContext())); 1191 } else { 1192 ICI.setOperand(1, NewCst); 1193 Constant *NewAndCST; 1194 if (Shift->getOpcode() == Instruction::Shl) 1195 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt); 1196 else 1197 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt); 1198 LHSI->setOperand(1, NewAndCST); 1199 LHSI->setOperand(0, Shift->getOperand(0)); 1200 Worklist.Add(Shift); // Shift is dead. 1201 return &ICI; 1202 } 1203 } 1204 } 1205 1206 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is 1207 // preferable because it allows the C<<Y expression to be hoisted out 1208 // of a loop if Y is invariant and X is not. 1209 if (Shift && Shift->hasOneUse() && RHSV == 0 && 1210 ICI.isEquality() && !Shift->isArithmeticShift() && 1211 !isa<Constant>(Shift->getOperand(0))) { 1212 // Compute C << Y. 1213 Value *NS; 1214 if (Shift->getOpcode() == Instruction::LShr) { 1215 NS = Builder->CreateShl(AndCST, Shift->getOperand(1)); 1216 } else { 1217 // Insert a logical shift. 1218 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1)); 1219 } 1220 1221 // Compute X & (C << Y). 1222 Value *NewAnd = 1223 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName()); 1224 1225 ICI.setOperand(0, NewAnd); 1226 return &ICI; 1227 } 1228 } 1229 1230 // Try to optimize things like "A[i]&42 == 0" to index computations. 1231 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) { 1232 if (GetElementPtrInst *GEP = 1233 dyn_cast<GetElementPtrInst>(LI->getOperand(0))) 1234 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 1235 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 1236 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) { 1237 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1)); 1238 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C)) 1239 return Res; 1240 } 1241 } 1242 break; 1243 1244 case Instruction::Or: { 1245 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse()) 1246 break; 1247 Value *P, *Q; 1248 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) { 1249 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 1250 // -> and (icmp eq P, null), (icmp eq Q, null). 1251 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P, 1252 Constant::getNullValue(P->getType())); 1253 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q, 1254 Constant::getNullValue(Q->getType())); 1255 Instruction *Op; 1256 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1257 Op = BinaryOperator::CreateAnd(ICIP, ICIQ); 1258 else 1259 Op = BinaryOperator::CreateOr(ICIP, ICIQ); 1260 return Op; 1261 } 1262 break; 1263 } 1264 1265 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI) 1266 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1267 if (!ShAmt) break; 1268 1269 uint32_t TypeBits = RHSV.getBitWidth(); 1270 1271 // Check that the shift amount is in range. If not, don't perform 1272 // undefined shifts. When the shift is visited it will be 1273 // simplified. 1274 if (ShAmt->uge(TypeBits)) 1275 break; 1276 1277 if (ICI.isEquality()) { 1278 // If we are comparing against bits always shifted out, the 1279 // comparison cannot succeed. 1280 Constant *Comp = 1281 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), 1282 ShAmt); 1283 if (Comp != RHS) {// Comparing against a bit that we know is zero. 1284 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1285 Constant *Cst = 1286 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE); 1287 return ReplaceInstUsesWith(ICI, Cst); 1288 } 1289 1290 // If the shift is NUW, then it is just shifting out zeros, no need for an 1291 // AND. 1292 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap()) 1293 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1294 ConstantExpr::getLShr(RHS, ShAmt)); 1295 1296 if (LHSI->hasOneUse()) { 1297 // Otherwise strength reduce the shift into an and. 1298 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 1299 Constant *Mask = 1300 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits, 1301 TypeBits-ShAmtVal)); 1302 1303 Value *And = 1304 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask"); 1305 return new ICmpInst(ICI.getPredicate(), And, 1306 ConstantExpr::getLShr(RHS, ShAmt)); 1307 } 1308 } 1309 1310 // Otherwise, if this is a comparison of the sign bit, simplify to and/test. 1311 bool TrueIfSigned = false; 1312 if (LHSI->hasOneUse() && 1313 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) { 1314 // (X << 31) <s 0 --> (X&1) != 0 1315 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(), 1316 APInt::getOneBitSet(TypeBits, 1317 TypeBits-ShAmt->getZExtValue()-1)); 1318 Value *And = 1319 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask"); 1320 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, 1321 And, Constant::getNullValue(And->getType())); 1322 } 1323 break; 1324 } 1325 1326 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI) 1327 case Instruction::AShr: { 1328 // Handle equality comparisons of shift-by-constant. 1329 BinaryOperator *BO = cast<BinaryOperator>(LHSI); 1330 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1331 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt)) 1332 return Res; 1333 } 1334 1335 // Handle exact shr's. 1336 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) { 1337 if (RHSV.isMinValue()) 1338 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS); 1339 } 1340 break; 1341 } 1342 1343 case Instruction::SDiv: 1344 case Instruction::UDiv: 1345 // Fold: icmp pred ([us]div X, C1), C2 -> range test 1346 // Fold this div into the comparison, producing a range check. 1347 // Determine, based on the divide type, what the range is being 1348 // checked. If there is an overflow on the low or high side, remember 1349 // it, otherwise compute the range [low, hi) bounding the new value. 1350 // See: InsertRangeTest above for the kinds of replacements possible. 1351 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) 1352 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI), 1353 DivRHS)) 1354 return R; 1355 break; 1356 1357 case Instruction::Add: 1358 // Fold: icmp pred (add X, C1), C2 1359 if (!ICI.isEquality()) { 1360 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1361 if (!LHSC) break; 1362 const APInt &LHSV = LHSC->getValue(); 1363 1364 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV) 1365 .subtract(LHSV); 1366 1367 if (ICI.isSigned()) { 1368 if (CR.getLower().isSignBit()) { 1369 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0), 1370 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1371 } else if (CR.getUpper().isSignBit()) { 1372 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0), 1373 ConstantInt::get(ICI.getContext(),CR.getLower())); 1374 } 1375 } else { 1376 if (CR.getLower().isMinValue()) { 1377 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), 1378 ConstantInt::get(ICI.getContext(),CR.getUpper())); 1379 } else if (CR.getUpper().isMinValue()) { 1380 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), 1381 ConstantInt::get(ICI.getContext(),CR.getLower())); 1382 } 1383 } 1384 } 1385 break; 1386 } 1387 1388 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS. 1389 if (ICI.isEquality()) { 1390 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1391 1392 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and 1393 // the second operand is a constant, simplify a bit. 1394 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) { 1395 switch (BO->getOpcode()) { 1396 case Instruction::SRem: 1397 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. 1398 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){ 1399 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue(); 1400 if (V.sgt(1) && V.isPowerOf2()) { 1401 Value *NewRem = 1402 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1), 1403 BO->getName()); 1404 return new ICmpInst(ICI.getPredicate(), NewRem, 1405 Constant::getNullValue(BO->getType())); 1406 } 1407 } 1408 break; 1409 case Instruction::Add: 1410 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants. 1411 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1412 if (BO->hasOneUse()) 1413 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1414 ConstantExpr::getSub(RHS, BOp1C)); 1415 } else if (RHSV == 0) { 1416 // Replace ((add A, B) != 0) with (A != -B) if A or B is 1417 // efficiently invertible, or if the add has just this one use. 1418 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1); 1419 1420 if (Value *NegVal = dyn_castNegVal(BOp1)) 1421 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal); 1422 if (Value *NegVal = dyn_castNegVal(BOp0)) 1423 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1); 1424 if (BO->hasOneUse()) { 1425 Value *Neg = Builder->CreateNeg(BOp1); 1426 Neg->takeName(BO); 1427 return new ICmpInst(ICI.getPredicate(), BOp0, Neg); 1428 } 1429 } 1430 break; 1431 case Instruction::Xor: 1432 // For the xor case, we can xor two constants together, eliminating 1433 // the explicit xor. 1434 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) { 1435 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1436 ConstantExpr::getXor(RHS, BOC)); 1437 } else if (RHSV == 0) { 1438 // Replace ((xor A, B) != 0) with (A != B) 1439 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1440 BO->getOperand(1)); 1441 } 1442 break; 1443 case Instruction::Sub: 1444 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants. 1445 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) { 1446 if (BO->hasOneUse()) 1447 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1), 1448 ConstantExpr::getSub(BOp0C, RHS)); 1449 } else if (RHSV == 0) { 1450 // Replace ((sub A, B) != 0) with (A != B) 1451 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1452 BO->getOperand(1)); 1453 } 1454 break; 1455 case Instruction::Or: 1456 // If bits are being or'd in that are not present in the constant we 1457 // are comparing against, then the comparison could never succeed! 1458 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1459 Constant *NotCI = ConstantExpr::getNot(RHS); 1460 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue()) 1461 return ReplaceInstUsesWith(ICI, 1462 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1463 isICMP_NE)); 1464 } 1465 break; 1466 1467 case Instruction::And: 1468 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1469 // If bits are being compared against that are and'd out, then the 1470 // comparison can never succeed! 1471 if ((RHSV & ~BOC->getValue()) != 0) 1472 return ReplaceInstUsesWith(ICI, 1473 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), 1474 isICMP_NE)); 1475 1476 // If we have ((X & C) == C), turn it into ((X & C) != 0). 1477 if (RHS == BOC && RHSV.isPowerOf2()) 1478 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : 1479 ICmpInst::ICMP_NE, LHSI, 1480 Constant::getNullValue(RHS->getType())); 1481 1482 // Don't perform the following transforms if the AND has multiple uses 1483 if (!BO->hasOneUse()) 1484 break; 1485 1486 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0 1487 if (BOC->getValue().isSignBit()) { 1488 Value *X = BO->getOperand(0); 1489 Constant *Zero = Constant::getNullValue(X->getType()); 1490 ICmpInst::Predicate pred = isICMP_NE ? 1491 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; 1492 return new ICmpInst(pred, X, Zero); 1493 } 1494 1495 // ((X & ~7) == 0) --> X < 8 1496 if (RHSV == 0 && isHighOnes(BOC)) { 1497 Value *X = BO->getOperand(0); 1498 Constant *NegX = ConstantExpr::getNeg(BOC); 1499 ICmpInst::Predicate pred = isICMP_NE ? 1500 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; 1501 return new ICmpInst(pred, X, NegX); 1502 } 1503 } 1504 default: break; 1505 } 1506 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) { 1507 // Handle icmp {eq|ne} <intrinsic>, intcst. 1508 switch (II->getIntrinsicID()) { 1509 case Intrinsic::bswap: 1510 Worklist.Add(II); 1511 ICI.setOperand(0, II->getArgOperand(0)); 1512 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap())); 1513 return &ICI; 1514 case Intrinsic::ctlz: 1515 case Intrinsic::cttz: 1516 // ctz(A) == bitwidth(a) -> A == 0 and likewise for != 1517 if (RHSV == RHS->getType()->getBitWidth()) { 1518 Worklist.Add(II); 1519 ICI.setOperand(0, II->getArgOperand(0)); 1520 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0)); 1521 return &ICI; 1522 } 1523 break; 1524 case Intrinsic::ctpop: 1525 // popcount(A) == 0 -> A == 0 and likewise for != 1526 if (RHS->isZero()) { 1527 Worklist.Add(II); 1528 ICI.setOperand(0, II->getArgOperand(0)); 1529 ICI.setOperand(1, RHS); 1530 return &ICI; 1531 } 1532 break; 1533 default: 1534 break; 1535 } 1536 } 1537 } 1538 return 0; 1539 } 1540 1541 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst). 1542 /// We only handle extending casts so far. 1543 /// 1544 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) { 1545 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0)); 1546 Value *LHSCIOp = LHSCI->getOperand(0); 1547 Type *SrcTy = LHSCIOp->getType(); 1548 Type *DestTy = LHSCI->getType(); 1549 Value *RHSCIOp; 1550 1551 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the 1552 // integer type is the same size as the pointer type. 1553 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt && 1554 TD->getPointerSizeInBits() == 1555 cast<IntegerType>(DestTy)->getBitWidth()) { 1556 Value *RHSOp = 0; 1557 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) { 1558 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy); 1559 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) { 1560 RHSOp = RHSC->getOperand(0); 1561 // If the pointer types don't match, insert a bitcast. 1562 if (LHSCIOp->getType() != RHSOp->getType()) 1563 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType()); 1564 } 1565 1566 if (RHSOp) 1567 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp); 1568 } 1569 1570 // The code below only handles extension cast instructions, so far. 1571 // Enforce this. 1572 if (LHSCI->getOpcode() != Instruction::ZExt && 1573 LHSCI->getOpcode() != Instruction::SExt) 1574 return 0; 1575 1576 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt; 1577 bool isSignedCmp = ICI.isSigned(); 1578 1579 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) { 1580 // Not an extension from the same type? 1581 RHSCIOp = CI->getOperand(0); 1582 if (RHSCIOp->getType() != LHSCIOp->getType()) 1583 return 0; 1584 1585 // If the signedness of the two casts doesn't agree (i.e. one is a sext 1586 // and the other is a zext), then we can't handle this. 1587 if (CI->getOpcode() != LHSCI->getOpcode()) 1588 return 0; 1589 1590 // Deal with equality cases early. 1591 if (ICI.isEquality()) 1592 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1593 1594 // A signed comparison of sign extended values simplifies into a 1595 // signed comparison. 1596 if (isSignedCmp && isSignedExt) 1597 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1598 1599 // The other three cases all fold into an unsigned comparison. 1600 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp); 1601 } 1602 1603 // If we aren't dealing with a constant on the RHS, exit early 1604 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1)); 1605 if (!CI) 1606 return 0; 1607 1608 // Compute the constant that would happen if we truncated to SrcTy then 1609 // reextended to DestTy. 1610 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy); 1611 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), 1612 Res1, DestTy); 1613 1614 // If the re-extended constant didn't change... 1615 if (Res2 == CI) { 1616 // Deal with equality cases early. 1617 if (ICI.isEquality()) 1618 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1619 1620 // A signed comparison of sign extended values simplifies into a 1621 // signed comparison. 1622 if (isSignedExt && isSignedCmp) 1623 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1624 1625 // The other three cases all fold into an unsigned comparison. 1626 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1); 1627 } 1628 1629 // The re-extended constant changed so the constant cannot be represented 1630 // in the shorter type. Consequently, we cannot emit a simple comparison. 1631 // All the cases that fold to true or false will have already been handled 1632 // by SimplifyICmpInst, so only deal with the tricky case. 1633 1634 if (isSignedCmp || !isSignedExt) 1635 return 0; 1636 1637 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases 1638 // should have been folded away previously and not enter in here. 1639 1640 // We're performing an unsigned comp with a sign extended value. 1641 // This is true if the input is >= 0. [aka >s -1] 1642 Constant *NegOne = Constant::getAllOnesValue(SrcTy); 1643 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName()); 1644 1645 // Finally, return the value computed. 1646 if (ICI.getPredicate() == ICmpInst::ICMP_ULT) 1647 return ReplaceInstUsesWith(ICI, Result); 1648 1649 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"); 1650 return BinaryOperator::CreateNot(Result); 1651 } 1652 1653 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form: 1654 /// I = icmp ugt (add (add A, B), CI2), CI1 1655 /// If this is of the form: 1656 /// sum = a + b 1657 /// if (sum+128 >u 255) 1658 /// Then replace it with llvm.sadd.with.overflow.i8. 1659 /// 1660 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B, 1661 ConstantInt *CI2, ConstantInt *CI1, 1662 InstCombiner &IC) { 1663 // The transformation we're trying to do here is to transform this into an 1664 // llvm.sadd.with.overflow. To do this, we have to replace the original add 1665 // with a narrower add, and discard the add-with-constant that is part of the 1666 // range check (if we can't eliminate it, this isn't profitable). 1667 1668 // In order to eliminate the add-with-constant, the compare can be its only 1669 // use. 1670 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0)); 1671 if (!AddWithCst->hasOneUse()) return 0; 1672 1673 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow. 1674 if (!CI2->getValue().isPowerOf2()) return 0; 1675 unsigned NewWidth = CI2->getValue().countTrailingZeros(); 1676 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0; 1677 1678 // The width of the new add formed is 1 more than the bias. 1679 ++NewWidth; 1680 1681 // Check to see that CI1 is an all-ones value with NewWidth bits. 1682 if (CI1->getBitWidth() == NewWidth || 1683 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth)) 1684 return 0; 1685 1686 // This is only really a signed overflow check if the inputs have been 1687 // sign-extended; check for that condition. For example, if CI2 is 2^31 and 1688 // the operands of the add are 64 bits wide, we need at least 33 sign bits. 1689 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1; 1690 if (IC.ComputeNumSignBits(A) < NeededSignBits || 1691 IC.ComputeNumSignBits(B) < NeededSignBits) 1692 return 0; 1693 1694 // In order to replace the original add with a narrower 1695 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant 1696 // and truncates that discard the high bits of the add. Verify that this is 1697 // the case. 1698 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0)); 1699 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end(); 1700 UI != E; ++UI) { 1701 if (*UI == AddWithCst) continue; 1702 1703 // Only accept truncates for now. We would really like a nice recursive 1704 // predicate like SimplifyDemandedBits, but which goes downwards the use-def 1705 // chain to see which bits of a value are actually demanded. If the 1706 // original add had another add which was then immediately truncated, we 1707 // could still do the transformation. 1708 TruncInst *TI = dyn_cast<TruncInst>(*UI); 1709 if (TI == 0 || 1710 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0; 1711 } 1712 1713 // If the pattern matches, truncate the inputs to the narrower type and 1714 // use the sadd_with_overflow intrinsic to efficiently compute both the 1715 // result and the overflow bit. 1716 Module *M = I.getParent()->getParent()->getParent(); 1717 1718 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth); 1719 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow, 1720 NewType); 1721 1722 InstCombiner::BuilderTy *Builder = IC.Builder; 1723 1724 // Put the new code above the original add, in case there are any uses of the 1725 // add between the add and the compare. 1726 Builder->SetInsertPoint(OrigAdd); 1727 1728 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc"); 1729 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc"); 1730 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd"); 1731 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result"); 1732 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType()); 1733 1734 // The inner add was the result of the narrow add, zero extended to the 1735 // wider type. Replace it with the result computed by the intrinsic. 1736 IC.ReplaceInstUsesWith(*OrigAdd, ZExt); 1737 1738 // The original icmp gets replaced with the overflow value. 1739 return ExtractValueInst::Create(Call, 1, "sadd.overflow"); 1740 } 1741 1742 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV, 1743 InstCombiner &IC) { 1744 // Don't bother doing this transformation for pointers, don't do it for 1745 // vectors. 1746 if (!isa<IntegerType>(OrigAddV->getType())) return 0; 1747 1748 // If the add is a constant expr, then we don't bother transforming it. 1749 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV); 1750 if (OrigAdd == 0) return 0; 1751 1752 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1); 1753 1754 // Put the new code above the original add, in case there are any uses of the 1755 // add between the add and the compare. 1756 InstCombiner::BuilderTy *Builder = IC.Builder; 1757 Builder->SetInsertPoint(OrigAdd); 1758 1759 Module *M = I.getParent()->getParent()->getParent(); 1760 Type *Ty = LHS->getType(); 1761 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty); 1762 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd"); 1763 Value *Add = Builder->CreateExtractValue(Call, 0); 1764 1765 IC.ReplaceInstUsesWith(*OrigAdd, Add); 1766 1767 // The original icmp gets replaced with the overflow value. 1768 return ExtractValueInst::Create(Call, 1, "uadd.overflow"); 1769 } 1770 1771 // DemandedBitsLHSMask - When performing a comparison against a constant, 1772 // it is possible that not all the bits in the LHS are demanded. This helper 1773 // method computes the mask that IS demanded. 1774 static APInt DemandedBitsLHSMask(ICmpInst &I, 1775 unsigned BitWidth, bool isSignCheck) { 1776 if (isSignCheck) 1777 return APInt::getSignBit(BitWidth); 1778 1779 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1)); 1780 if (!CI) return APInt::getAllOnesValue(BitWidth); 1781 const APInt &RHS = CI->getValue(); 1782 1783 switch (I.getPredicate()) { 1784 // For a UGT comparison, we don't care about any bits that 1785 // correspond to the trailing ones of the comparand. The value of these 1786 // bits doesn't impact the outcome of the comparison, because any value 1787 // greater than the RHS must differ in a bit higher than these due to carry. 1788 case ICmpInst::ICMP_UGT: { 1789 unsigned trailingOnes = RHS.countTrailingOnes(); 1790 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes); 1791 return ~lowBitsSet; 1792 } 1793 1794 // Similarly, for a ULT comparison, we don't care about the trailing zeros. 1795 // Any value less than the RHS must differ in a higher bit because of carries. 1796 case ICmpInst::ICMP_ULT: { 1797 unsigned trailingZeros = RHS.countTrailingZeros(); 1798 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros); 1799 return ~lowBitsSet; 1800 } 1801 1802 default: 1803 return APInt::getAllOnesValue(BitWidth); 1804 } 1805 1806 } 1807 1808 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { 1809 bool Changed = false; 1810 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1811 1812 /// Orders the operands of the compare so that they are listed from most 1813 /// complex to least complex. This puts constants before unary operators, 1814 /// before binary operators. 1815 if (getComplexity(Op0) < getComplexity(Op1)) { 1816 I.swapOperands(); 1817 std::swap(Op0, Op1); 1818 Changed = true; 1819 } 1820 1821 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD)) 1822 return ReplaceInstUsesWith(I, V); 1823 1824 // comparing -val or val with non-zero is the same as just comparing val 1825 // ie, abs(val) != 0 -> val != 0 1826 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) 1827 { 1828 Value *Cond, *SelectTrue, *SelectFalse; 1829 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue), 1830 m_Value(SelectFalse)))) { 1831 if (Value *V = dyn_castNegVal(SelectTrue)) { 1832 if (V == SelectFalse) 1833 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 1834 } 1835 else if (Value *V = dyn_castNegVal(SelectFalse)) { 1836 if (V == SelectTrue) 1837 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 1838 } 1839 } 1840 } 1841 1842 Type *Ty = Op0->getType(); 1843 1844 // icmp's with boolean values can always be turned into bitwise operations 1845 if (Ty->isIntegerTy(1)) { 1846 switch (I.getPredicate()) { 1847 default: llvm_unreachable("Invalid icmp instruction!"); 1848 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B) 1849 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp"); 1850 return BinaryOperator::CreateNot(Xor); 1851 } 1852 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B 1853 return BinaryOperator::CreateXor(Op0, Op1); 1854 1855 case ICmpInst::ICMP_UGT: 1856 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult 1857 // FALL THROUGH 1858 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B 1859 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1860 return BinaryOperator::CreateAnd(Not, Op1); 1861 } 1862 case ICmpInst::ICMP_SGT: 1863 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt 1864 // FALL THROUGH 1865 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B 1866 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1867 return BinaryOperator::CreateAnd(Not, Op0); 1868 } 1869 case ICmpInst::ICMP_UGE: 1870 std::swap(Op0, Op1); // Change icmp uge -> icmp ule 1871 // FALL THROUGH 1872 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B 1873 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 1874 return BinaryOperator::CreateOr(Not, Op1); 1875 } 1876 case ICmpInst::ICMP_SGE: 1877 std::swap(Op0, Op1); // Change icmp sge -> icmp sle 1878 // FALL THROUGH 1879 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B 1880 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 1881 return BinaryOperator::CreateOr(Not, Op0); 1882 } 1883 } 1884 } 1885 1886 unsigned BitWidth = 0; 1887 if (Ty->isIntOrIntVectorTy()) 1888 BitWidth = Ty->getScalarSizeInBits(); 1889 else if (TD) // Pointers require TD info to get their size. 1890 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType()); 1891 1892 bool isSignBit = false; 1893 1894 // See if we are doing a comparison with a constant. 1895 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 1896 Value *A = 0, *B = 0; 1897 1898 // Match the following pattern, which is a common idiom when writing 1899 // overflow-safe integer arithmetic function. The source performs an 1900 // addition in wider type, and explicitly checks for overflow using 1901 // comparisons against INT_MIN and INT_MAX. Simplify this by using the 1902 // sadd_with_overflow intrinsic. 1903 // 1904 // TODO: This could probably be generalized to handle other overflow-safe 1905 // operations if we worked out the formulas to compute the appropriate 1906 // magic constants. 1907 // 1908 // sum = a + b 1909 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8 1910 { 1911 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI 1912 if (I.getPredicate() == ICmpInst::ICMP_UGT && 1913 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2)))) 1914 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this)) 1915 return Res; 1916 } 1917 1918 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B) 1919 if (I.isEquality() && CI->isZero() && 1920 match(Op0, m_Sub(m_Value(A), m_Value(B)))) { 1921 // (icmp cond A B) if cond is equality 1922 return new ICmpInst(I.getPredicate(), A, B); 1923 } 1924 1925 // If we have an icmp le or icmp ge instruction, turn it into the 1926 // appropriate icmp lt or icmp gt instruction. This allows us to rely on 1927 // them being folded in the code below. The SimplifyICmpInst code has 1928 // already handled the edge cases for us, so we just assert on them. 1929 switch (I.getPredicate()) { 1930 default: break; 1931 case ICmpInst::ICMP_ULE: 1932 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE 1933 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, 1934 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1935 case ICmpInst::ICMP_SLE: 1936 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE 1937 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 1938 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 1939 case ICmpInst::ICMP_UGE: 1940 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE 1941 return new ICmpInst(ICmpInst::ICMP_UGT, Op0, 1942 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1943 case ICmpInst::ICMP_SGE: 1944 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE 1945 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 1946 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 1947 } 1948 1949 // If this comparison is a normal comparison, it demands all 1950 // bits, if it is a sign bit comparison, it only demands the sign bit. 1951 bool UnusedBit; 1952 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit); 1953 } 1954 1955 // See if we can fold the comparison based on range information we can get 1956 // by checking whether bits are known to be zero or one in the input. 1957 if (BitWidth != 0) { 1958 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0); 1959 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0); 1960 1961 if (SimplifyDemandedBits(I.getOperandUse(0), 1962 DemandedBitsLHSMask(I, BitWidth, isSignBit), 1963 Op0KnownZero, Op0KnownOne, 0)) 1964 return &I; 1965 if (SimplifyDemandedBits(I.getOperandUse(1), 1966 APInt::getAllOnesValue(BitWidth), 1967 Op1KnownZero, Op1KnownOne, 0)) 1968 return &I; 1969 1970 // Given the known and unknown bits, compute a range that the LHS could be 1971 // in. Compute the Min, Max and RHS values based on the known bits. For the 1972 // EQ and NE we use unsigned values. 1973 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0); 1974 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0); 1975 if (I.isSigned()) { 1976 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 1977 Op0Min, Op0Max); 1978 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 1979 Op1Min, Op1Max); 1980 } else { 1981 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 1982 Op0Min, Op0Max); 1983 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 1984 Op1Min, Op1Max); 1985 } 1986 1987 // If Min and Max are known to be the same, then SimplifyDemandedBits 1988 // figured out that the LHS is a constant. Just constant fold this now so 1989 // that code below can assume that Min != Max. 1990 if (!isa<Constant>(Op0) && Op0Min == Op0Max) 1991 return new ICmpInst(I.getPredicate(), 1992 ConstantInt::get(Op0->getType(), Op0Min), Op1); 1993 if (!isa<Constant>(Op1) && Op1Min == Op1Max) 1994 return new ICmpInst(I.getPredicate(), Op0, 1995 ConstantInt::get(Op1->getType(), Op1Min)); 1996 1997 // Based on the range information we know about the LHS, see if we can 1998 // simplify this comparison. For example, (x&4) < 8 is always true. 1999 switch (I.getPredicate()) { 2000 default: llvm_unreachable("Unknown icmp opcode!"); 2001 case ICmpInst::ICMP_EQ: { 2002 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2003 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2004 2005 // If all bits are known zero except for one, then we know at most one 2006 // bit is set. If the comparison is against zero, then this is a check 2007 // to see if *that* bit is set. 2008 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2009 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2010 // If the LHS is an AND with the same constant, look through it. 2011 Value *LHS = 0; 2012 ConstantInt *LHSC = 0; 2013 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2014 LHSC->getValue() != Op0KnownZeroInverted) 2015 LHS = Op0; 2016 2017 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2018 // then turn "((1 << x)&8) == 0" into "x != 3". 2019 Value *X = 0; 2020 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2021 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2022 return new ICmpInst(ICmpInst::ICMP_NE, X, 2023 ConstantInt::get(X->getType(), CmpVal)); 2024 } 2025 2026 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2027 // then turn "((8 >>u x)&1) == 0" into "x != 3". 2028 const APInt *CI; 2029 if (Op0KnownZeroInverted == 1 && 2030 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2031 return new ICmpInst(ICmpInst::ICMP_NE, X, 2032 ConstantInt::get(X->getType(), 2033 CI->countTrailingZeros())); 2034 } 2035 2036 break; 2037 } 2038 case ICmpInst::ICMP_NE: { 2039 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2040 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2041 2042 // If all bits are known zero except for one, then we know at most one 2043 // bit is set. If the comparison is against zero, then this is a check 2044 // to see if *that* bit is set. 2045 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2046 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2047 // If the LHS is an AND with the same constant, look through it. 2048 Value *LHS = 0; 2049 ConstantInt *LHSC = 0; 2050 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2051 LHSC->getValue() != Op0KnownZeroInverted) 2052 LHS = Op0; 2053 2054 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2055 // then turn "((1 << x)&8) != 0" into "x == 3". 2056 Value *X = 0; 2057 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2058 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2059 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2060 ConstantInt::get(X->getType(), CmpVal)); 2061 } 2062 2063 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2064 // then turn "((8 >>u x)&1) != 0" into "x == 3". 2065 const APInt *CI; 2066 if (Op0KnownZeroInverted == 1 && 2067 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2068 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2069 ConstantInt::get(X->getType(), 2070 CI->countTrailingZeros())); 2071 } 2072 2073 break; 2074 } 2075 case ICmpInst::ICMP_ULT: 2076 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B) 2077 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2078 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B) 2079 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2080 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B) 2081 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2082 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2083 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C 2084 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2085 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2086 2087 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear 2088 if (CI->isMinValue(true)) 2089 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 2090 Constant::getAllOnesValue(Op0->getType())); 2091 } 2092 break; 2093 case ICmpInst::ICMP_UGT: 2094 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B) 2095 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2096 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B) 2097 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2098 2099 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B) 2100 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2101 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2102 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C 2103 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2104 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2105 2106 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set 2107 if (CI->isMaxValue(true)) 2108 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 2109 Constant::getNullValue(Op0->getType())); 2110 } 2111 break; 2112 case ICmpInst::ICMP_SLT: 2113 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C) 2114 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2115 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C) 2116 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2117 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B) 2118 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2119 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2120 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C 2121 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2122 ConstantInt::get(CI->getContext(), CI->getValue()-1)); 2123 } 2124 break; 2125 case ICmpInst::ICMP_SGT: 2126 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B) 2127 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2128 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B) 2129 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2130 2131 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B) 2132 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2133 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2134 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C 2135 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2136 ConstantInt::get(CI->getContext(), CI->getValue()+1)); 2137 } 2138 break; 2139 case ICmpInst::ICMP_SGE: 2140 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"); 2141 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B) 2142 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2143 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B) 2144 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2145 break; 2146 case ICmpInst::ICMP_SLE: 2147 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"); 2148 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B) 2149 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2150 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B) 2151 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2152 break; 2153 case ICmpInst::ICMP_UGE: 2154 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"); 2155 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B) 2156 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2157 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B) 2158 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2159 break; 2160 case ICmpInst::ICMP_ULE: 2161 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"); 2162 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B) 2163 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2164 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B) 2165 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2166 break; 2167 } 2168 2169 // Turn a signed comparison into an unsigned one if both operands 2170 // are known to have the same sign. 2171 if (I.isSigned() && 2172 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) || 2173 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative()))) 2174 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1); 2175 } 2176 2177 // Test if the ICmpInst instruction is used exclusively by a select as 2178 // part of a minimum or maximum operation. If so, refrain from doing 2179 // any other folding. This helps out other analyses which understand 2180 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 2181 // and CodeGen. And in this case, at least one of the comparison 2182 // operands has at least one user besides the compare (the select), 2183 // which would often largely negate the benefit of folding anyway. 2184 if (I.hasOneUse()) 2185 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin())) 2186 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) || 2187 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1)) 2188 return 0; 2189 2190 // See if we are doing a comparison between a constant and an instruction that 2191 // can be folded into the comparison. 2192 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2193 // Since the RHS is a ConstantInt (CI), if the left hand side is an 2194 // instruction, see if that instruction also has constants so that the 2195 // instruction can be folded into the icmp 2196 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2197 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI)) 2198 return Res; 2199 } 2200 2201 // Handle icmp with constant (but not simple integer constant) RHS 2202 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2203 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2204 switch (LHSI->getOpcode()) { 2205 case Instruction::GetElementPtr: 2206 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null 2207 if (RHSC->isNullValue() && 2208 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices()) 2209 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2210 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2211 break; 2212 case Instruction::PHI: 2213 // Only fold icmp into the PHI if the phi and icmp are in the same 2214 // block. If in the same block, we're encouraging jump threading. If 2215 // not, we are just pessimizing the code by making an i1 phi. 2216 if (LHSI->getParent() == I.getParent()) 2217 if (Instruction *NV = FoldOpIntoPhi(I)) 2218 return NV; 2219 break; 2220 case Instruction::Select: { 2221 // If either operand of the select is a constant, we can fold the 2222 // comparison into the select arms, which will cause one to be 2223 // constant folded and the select turned into a bitwise or. 2224 Value *Op1 = 0, *Op2 = 0; 2225 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) 2226 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2227 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) 2228 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2229 2230 // We only want to perform this transformation if it will not lead to 2231 // additional code. This is true if either both sides of the select 2232 // fold to a constant (in which case the icmp is replaced with a select 2233 // which will usually simplify) or this is the only user of the 2234 // select (in which case we are trading a select+icmp for a simpler 2235 // select+icmp). 2236 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) { 2237 if (!Op1) 2238 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1), 2239 RHSC, I.getName()); 2240 if (!Op2) 2241 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2), 2242 RHSC, I.getName()); 2243 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2244 } 2245 break; 2246 } 2247 case Instruction::IntToPtr: 2248 // icmp pred inttoptr(X), null -> icmp pred X, 0 2249 if (RHSC->isNullValue() && TD && 2250 TD->getIntPtrType(RHSC->getContext()) == 2251 LHSI->getOperand(0)->getType()) 2252 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2253 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2254 break; 2255 2256 case Instruction::Load: 2257 // Try to optimize things like "A[i] > 4" to index computations. 2258 if (GetElementPtrInst *GEP = 2259 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2260 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2261 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2262 !cast<LoadInst>(LHSI)->isVolatile()) 2263 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2264 return Res; 2265 } 2266 break; 2267 } 2268 } 2269 2270 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now. 2271 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0)) 2272 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I)) 2273 return NI; 2274 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) 2275 if (Instruction *NI = FoldGEPICmp(GEP, Op0, 2276 ICmpInst::getSwappedPredicate(I.getPredicate()), I)) 2277 return NI; 2278 2279 // Test to see if the operands of the icmp are casted versions of other 2280 // values. If the ptr->ptr cast can be stripped off both arguments, we do so 2281 // now. 2282 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) { 2283 if (Op0->getType()->isPointerTy() && 2284 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) { 2285 // We keep moving the cast from the left operand over to the right 2286 // operand, where it can often be eliminated completely. 2287 Op0 = CI->getOperand(0); 2288 2289 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast 2290 // so eliminate it as well. 2291 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1)) 2292 Op1 = CI2->getOperand(0); 2293 2294 // If Op1 is a constant, we can fold the cast into the constant. 2295 if (Op0->getType() != Op1->getType()) { 2296 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 2297 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType()); 2298 } else { 2299 // Otherwise, cast the RHS right before the icmp 2300 Op1 = Builder->CreateBitCast(Op1, Op0->getType()); 2301 } 2302 } 2303 return new ICmpInst(I.getPredicate(), Op0, Op1); 2304 } 2305 } 2306 2307 if (isa<CastInst>(Op0)) { 2308 // Handle the special case of: icmp (cast bool to X), <cst> 2309 // This comes up when you have code like 2310 // int X = A < B; 2311 // if (X) ... 2312 // For generality, we handle any zero-extension of any operand comparison 2313 // with a constant or another cast from the same type. 2314 if (isa<Constant>(Op1) || isa<CastInst>(Op1)) 2315 if (Instruction *R = visitICmpInstWithCastAndCast(I)) 2316 return R; 2317 } 2318 2319 // Special logic for binary operators. 2320 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0); 2321 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1); 2322 if (BO0 || BO1) { 2323 CmpInst::Predicate Pred = I.getPredicate(); 2324 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false; 2325 if (BO0 && isa<OverflowingBinaryOperator>(BO0)) 2326 NoOp0WrapProblem = ICmpInst::isEquality(Pred) || 2327 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) || 2328 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap()); 2329 if (BO1 && isa<OverflowingBinaryOperator>(BO1)) 2330 NoOp1WrapProblem = ICmpInst::isEquality(Pred) || 2331 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) || 2332 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap()); 2333 2334 // Analyze the case when either Op0 or Op1 is an add instruction. 2335 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null). 2336 Value *A = 0, *B = 0, *C = 0, *D = 0; 2337 if (BO0 && BO0->getOpcode() == Instruction::Add) 2338 A = BO0->getOperand(0), B = BO0->getOperand(1); 2339 if (BO1 && BO1->getOpcode() == Instruction::Add) 2340 C = BO1->getOperand(0), D = BO1->getOperand(1); 2341 2342 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2343 if ((A == Op1 || B == Op1) && NoOp0WrapProblem) 2344 return new ICmpInst(Pred, A == Op1 ? B : A, 2345 Constant::getNullValue(Op1->getType())); 2346 2347 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2348 if ((C == Op0 || D == Op0) && NoOp1WrapProblem) 2349 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()), 2350 C == Op0 ? D : C); 2351 2352 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow. 2353 if (A && C && (A == C || A == D || B == C || B == D) && 2354 NoOp0WrapProblem && NoOp1WrapProblem && 2355 // Try not to increase register pressure. 2356 BO0->hasOneUse() && BO1->hasOneUse()) { 2357 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2358 Value *Y = (A == C || A == D) ? B : A; 2359 Value *Z = (C == A || C == B) ? D : C; 2360 return new ICmpInst(Pred, Y, Z); 2361 } 2362 2363 // Analyze the case when either Op0 or Op1 is a sub instruction. 2364 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null). 2365 A = 0; B = 0; C = 0; D = 0; 2366 if (BO0 && BO0->getOpcode() == Instruction::Sub) 2367 A = BO0->getOperand(0), B = BO0->getOperand(1); 2368 if (BO1 && BO1->getOpcode() == Instruction::Sub) 2369 C = BO1->getOperand(0), D = BO1->getOperand(1); 2370 2371 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow. 2372 if (A == Op1 && NoOp0WrapProblem) 2373 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B); 2374 2375 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow. 2376 if (C == Op0 && NoOp1WrapProblem) 2377 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType())); 2378 2379 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow. 2380 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem && 2381 // Try not to increase register pressure. 2382 BO0->hasOneUse() && BO1->hasOneUse()) 2383 return new ICmpInst(Pred, A, C); 2384 2385 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow. 2386 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem && 2387 // Try not to increase register pressure. 2388 BO0->hasOneUse() && BO1->hasOneUse()) 2389 return new ICmpInst(Pred, D, B); 2390 2391 BinaryOperator *SRem = NULL; 2392 // icmp (srem X, Y), Y 2393 if (BO0 && BO0->getOpcode() == Instruction::SRem && 2394 Op1 == BO0->getOperand(1)) 2395 SRem = BO0; 2396 // icmp Y, (srem X, Y) 2397 else if (BO1 && BO1->getOpcode() == Instruction::SRem && 2398 Op0 == BO1->getOperand(1)) 2399 SRem = BO1; 2400 if (SRem) { 2401 // We don't check hasOneUse to avoid increasing register pressure because 2402 // the value we use is the same value this instruction was already using. 2403 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) { 2404 default: break; 2405 case ICmpInst::ICMP_EQ: 2406 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2407 case ICmpInst::ICMP_NE: 2408 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2409 case ICmpInst::ICMP_SGT: 2410 case ICmpInst::ICMP_SGE: 2411 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1), 2412 Constant::getAllOnesValue(SRem->getType())); 2413 case ICmpInst::ICMP_SLT: 2414 case ICmpInst::ICMP_SLE: 2415 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1), 2416 Constant::getNullValue(SRem->getType())); 2417 } 2418 } 2419 2420 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && 2421 BO0->hasOneUse() && BO1->hasOneUse() && 2422 BO0->getOperand(1) == BO1->getOperand(1)) { 2423 switch (BO0->getOpcode()) { 2424 default: break; 2425 case Instruction::Add: 2426 case Instruction::Sub: 2427 case Instruction::Xor: 2428 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b 2429 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2430 BO1->getOperand(0)); 2431 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b 2432 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2433 if (CI->getValue().isSignBit()) { 2434 ICmpInst::Predicate Pred = I.isSigned() 2435 ? I.getUnsignedPredicate() 2436 : I.getSignedPredicate(); 2437 return new ICmpInst(Pred, BO0->getOperand(0), 2438 BO1->getOperand(0)); 2439 } 2440 2441 if (CI->isMaxValue(true)) { 2442 ICmpInst::Predicate Pred = I.isSigned() 2443 ? I.getUnsignedPredicate() 2444 : I.getSignedPredicate(); 2445 Pred = I.getSwappedPredicate(Pred); 2446 return new ICmpInst(Pred, BO0->getOperand(0), 2447 BO1->getOperand(0)); 2448 } 2449 } 2450 break; 2451 case Instruction::Mul: 2452 if (!I.isEquality()) 2453 break; 2454 2455 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2456 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask 2457 // Mask = -1 >> count-trailing-zeros(Cst). 2458 if (!CI->isZero() && !CI->isOne()) { 2459 const APInt &AP = CI->getValue(); 2460 ConstantInt *Mask = ConstantInt::get(I.getContext(), 2461 APInt::getLowBitsSet(AP.getBitWidth(), 2462 AP.getBitWidth() - 2463 AP.countTrailingZeros())); 2464 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask); 2465 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask); 2466 return new ICmpInst(I.getPredicate(), And1, And2); 2467 } 2468 } 2469 break; 2470 case Instruction::UDiv: 2471 case Instruction::LShr: 2472 if (I.isSigned()) 2473 break; 2474 // fall-through 2475 case Instruction::SDiv: 2476 case Instruction::AShr: 2477 if (!BO0->isExact() || !BO1->isExact()) 2478 break; 2479 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2480 BO1->getOperand(0)); 2481 case Instruction::Shl: { 2482 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap(); 2483 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap(); 2484 if (!NUW && !NSW) 2485 break; 2486 if (!NSW && I.isSigned()) 2487 break; 2488 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2489 BO1->getOperand(0)); 2490 } 2491 } 2492 } 2493 } 2494 2495 { Value *A, *B; 2496 // ~x < ~y --> y < x 2497 // ~x < cst --> ~cst < x 2498 if (match(Op0, m_Not(m_Value(A)))) { 2499 if (match(Op1, m_Not(m_Value(B)))) 2500 return new ICmpInst(I.getPredicate(), B, A); 2501 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1)) 2502 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A); 2503 } 2504 2505 // (a+b) <u a --> llvm.uadd.with.overflow. 2506 // (a+b) <u b --> llvm.uadd.with.overflow. 2507 if (I.getPredicate() == ICmpInst::ICMP_ULT && 2508 match(Op0, m_Add(m_Value(A), m_Value(B))) && 2509 (Op1 == A || Op1 == B)) 2510 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this)) 2511 return R; 2512 2513 // a >u (a+b) --> llvm.uadd.with.overflow. 2514 // b >u (a+b) --> llvm.uadd.with.overflow. 2515 if (I.getPredicate() == ICmpInst::ICMP_UGT && 2516 match(Op1, m_Add(m_Value(A), m_Value(B))) && 2517 (Op0 == A || Op0 == B)) 2518 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this)) 2519 return R; 2520 } 2521 2522 if (I.isEquality()) { 2523 Value *A, *B, *C, *D; 2524 2525 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 2526 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0 2527 Value *OtherVal = A == Op1 ? B : A; 2528 return new ICmpInst(I.getPredicate(), OtherVal, 2529 Constant::getNullValue(A->getType())); 2530 } 2531 2532 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) { 2533 // A^c1 == C^c2 --> A == C^(c1^c2) 2534 ConstantInt *C1, *C2; 2535 if (match(B, m_ConstantInt(C1)) && 2536 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) { 2537 Constant *NC = ConstantInt::get(I.getContext(), 2538 C1->getValue() ^ C2->getValue()); 2539 Value *Xor = Builder->CreateXor(C, NC); 2540 return new ICmpInst(I.getPredicate(), A, Xor); 2541 } 2542 2543 // A^B == A^D -> B == D 2544 if (A == C) return new ICmpInst(I.getPredicate(), B, D); 2545 if (A == D) return new ICmpInst(I.getPredicate(), B, C); 2546 if (B == C) return new ICmpInst(I.getPredicate(), A, D); 2547 if (B == D) return new ICmpInst(I.getPredicate(), A, C); 2548 } 2549 } 2550 2551 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && 2552 (A == Op0 || B == Op0)) { 2553 // A == (A^B) -> B == 0 2554 Value *OtherVal = A == Op0 ? B : A; 2555 return new ICmpInst(I.getPredicate(), OtherVal, 2556 Constant::getNullValue(A->getType())); 2557 } 2558 2559 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0 2560 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) && 2561 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) { 2562 Value *X = 0, *Y = 0, *Z = 0; 2563 2564 if (A == C) { 2565 X = B; Y = D; Z = A; 2566 } else if (A == D) { 2567 X = B; Y = C; Z = A; 2568 } else if (B == C) { 2569 X = A; Y = D; Z = B; 2570 } else if (B == D) { 2571 X = A; Y = C; Z = B; 2572 } 2573 2574 if (X) { // Build (X^Y) & Z 2575 Op1 = Builder->CreateXor(X, Y); 2576 Op1 = Builder->CreateAnd(Op1, Z); 2577 I.setOperand(0, Op1); 2578 I.setOperand(1, Constant::getNullValue(Op1->getType())); 2579 return &I; 2580 } 2581 } 2582 2583 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to 2584 // "icmp (and X, mask), cst" 2585 uint64_t ShAmt = 0; 2586 ConstantInt *Cst1; 2587 if (Op0->hasOneUse() && 2588 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), 2589 m_ConstantInt(ShAmt))))) && 2590 match(Op1, m_ConstantInt(Cst1)) && 2591 // Only do this when A has multiple uses. This is most important to do 2592 // when it exposes other optimizations. 2593 !A->hasOneUse()) { 2594 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits(); 2595 2596 if (ShAmt < ASize) { 2597 APInt MaskV = 2598 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits()); 2599 MaskV <<= ShAmt; 2600 2601 APInt CmpV = Cst1->getValue().zext(ASize); 2602 CmpV <<= ShAmt; 2603 2604 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV)); 2605 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV)); 2606 } 2607 } 2608 } 2609 2610 { 2611 Value *X; ConstantInt *Cst; 2612 // icmp X+Cst, X 2613 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X) 2614 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0); 2615 2616 // icmp X, X+Cst 2617 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X) 2618 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1); 2619 } 2620 return Changed ? &I : 0; 2621 } 2622 2623 2624 2625 2626 2627 2628 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible. 2629 /// 2630 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I, 2631 Instruction *LHSI, 2632 Constant *RHSC) { 2633 if (!isa<ConstantFP>(RHSC)) return 0; 2634 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF(); 2635 2636 // Get the width of the mantissa. We don't want to hack on conversions that 2637 // might lose information from the integer, e.g. "i64 -> float" 2638 int MantissaWidth = LHSI->getType()->getFPMantissaWidth(); 2639 if (MantissaWidth == -1) return 0; // Unknown. 2640 2641 // Check to see that the input is converted from an integer type that is small 2642 // enough that preserves all bits. TODO: check here for "known" sign bits. 2643 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e. 2644 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits(); 2645 2646 // If this is a uitofp instruction, we need an extra bit to hold the sign. 2647 bool LHSUnsigned = isa<UIToFPInst>(LHSI); 2648 if (LHSUnsigned) 2649 ++InputSize; 2650 2651 // If the conversion would lose info, don't hack on this. 2652 if ((int)InputSize > MantissaWidth) 2653 return 0; 2654 2655 // Otherwise, we can potentially simplify the comparison. We know that it 2656 // will always come through as an integer value and we know the constant is 2657 // not a NAN (it would have been previously simplified). 2658 assert(!RHS.isNaN() && "NaN comparison not already folded!"); 2659 2660 ICmpInst::Predicate Pred; 2661 switch (I.getPredicate()) { 2662 default: llvm_unreachable("Unexpected predicate!"); 2663 case FCmpInst::FCMP_UEQ: 2664 case FCmpInst::FCMP_OEQ: 2665 Pred = ICmpInst::ICMP_EQ; 2666 break; 2667 case FCmpInst::FCMP_UGT: 2668 case FCmpInst::FCMP_OGT: 2669 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT; 2670 break; 2671 case FCmpInst::FCMP_UGE: 2672 case FCmpInst::FCMP_OGE: 2673 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE; 2674 break; 2675 case FCmpInst::FCMP_ULT: 2676 case FCmpInst::FCMP_OLT: 2677 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT; 2678 break; 2679 case FCmpInst::FCMP_ULE: 2680 case FCmpInst::FCMP_OLE: 2681 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE; 2682 break; 2683 case FCmpInst::FCMP_UNE: 2684 case FCmpInst::FCMP_ONE: 2685 Pred = ICmpInst::ICMP_NE; 2686 break; 2687 case FCmpInst::FCMP_ORD: 2688 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2689 case FCmpInst::FCMP_UNO: 2690 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2691 } 2692 2693 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType()); 2694 2695 // Now we know that the APFloat is a normal number, zero or inf. 2696 2697 // See if the FP constant is too large for the integer. For example, 2698 // comparing an i8 to 300.0. 2699 unsigned IntWidth = IntTy->getScalarSizeInBits(); 2700 2701 if (!LHSUnsigned) { 2702 // If the RHS value is > SignedMax, fold the comparison. This handles +INF 2703 // and large values. 2704 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false); 2705 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true, 2706 APFloat::rmNearestTiesToEven); 2707 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0 2708 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT || 2709 Pred == ICmpInst::ICMP_SLE) 2710 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2711 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2712 } 2713 } else { 2714 // If the RHS value is > UnsignedMax, fold the comparison. This handles 2715 // +INF and large values. 2716 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false); 2717 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false, 2718 APFloat::rmNearestTiesToEven); 2719 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0 2720 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT || 2721 Pred == ICmpInst::ICMP_ULE) 2722 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2723 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2724 } 2725 } 2726 2727 if (!LHSUnsigned) { 2728 // See if the RHS value is < SignedMin. 2729 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false); 2730 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true, 2731 APFloat::rmNearestTiesToEven); 2732 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0 2733 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT || 2734 Pred == ICmpInst::ICMP_SGE) 2735 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2736 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2737 } 2738 } else { 2739 // See if the RHS value is < UnsignedMin. 2740 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false); 2741 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true, 2742 APFloat::rmNearestTiesToEven); 2743 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0 2744 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT || 2745 Pred == ICmpInst::ICMP_UGE) 2746 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2747 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2748 } 2749 } 2750 2751 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or 2752 // [0, UMAX], but it may still be fractional. See if it is fractional by 2753 // casting the FP value to the integer value and back, checking for equality. 2754 // Don't do this for zero, because -0.0 is not fractional. 2755 Constant *RHSInt = LHSUnsigned 2756 ? ConstantExpr::getFPToUI(RHSC, IntTy) 2757 : ConstantExpr::getFPToSI(RHSC, IntTy); 2758 if (!RHS.isZero()) { 2759 bool Equal = LHSUnsigned 2760 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC 2761 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC; 2762 if (!Equal) { 2763 // If we had a comparison against a fractional value, we have to adjust 2764 // the compare predicate and sometimes the value. RHSC is rounded towards 2765 // zero at this point. 2766 switch (Pred) { 2767 default: llvm_unreachable("Unexpected integer comparison!"); 2768 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true 2769 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2770 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false 2771 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2772 case ICmpInst::ICMP_ULE: 2773 // (float)int <= 4.4 --> int <= 4 2774 // (float)int <= -4.4 --> false 2775 if (RHS.isNegative()) 2776 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2777 break; 2778 case ICmpInst::ICMP_SLE: 2779 // (float)int <= 4.4 --> int <= 4 2780 // (float)int <= -4.4 --> int < -4 2781 if (RHS.isNegative()) 2782 Pred = ICmpInst::ICMP_SLT; 2783 break; 2784 case ICmpInst::ICMP_ULT: 2785 // (float)int < -4.4 --> false 2786 // (float)int < 4.4 --> int <= 4 2787 if (RHS.isNegative()) 2788 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext())); 2789 Pred = ICmpInst::ICMP_ULE; 2790 break; 2791 case ICmpInst::ICMP_SLT: 2792 // (float)int < -4.4 --> int < -4 2793 // (float)int < 4.4 --> int <= 4 2794 if (!RHS.isNegative()) 2795 Pred = ICmpInst::ICMP_SLE; 2796 break; 2797 case ICmpInst::ICMP_UGT: 2798 // (float)int > 4.4 --> int > 4 2799 // (float)int > -4.4 --> true 2800 if (RHS.isNegative()) 2801 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2802 break; 2803 case ICmpInst::ICMP_SGT: 2804 // (float)int > 4.4 --> int > 4 2805 // (float)int > -4.4 --> int >= -4 2806 if (RHS.isNegative()) 2807 Pred = ICmpInst::ICMP_SGE; 2808 break; 2809 case ICmpInst::ICMP_UGE: 2810 // (float)int >= -4.4 --> true 2811 // (float)int >= 4.4 --> int > 4 2812 if (!RHS.isNegative()) 2813 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext())); 2814 Pred = ICmpInst::ICMP_UGT; 2815 break; 2816 case ICmpInst::ICMP_SGE: 2817 // (float)int >= -4.4 --> int >= -4 2818 // (float)int >= 4.4 --> int > 4 2819 if (!RHS.isNegative()) 2820 Pred = ICmpInst::ICMP_SGT; 2821 break; 2822 } 2823 } 2824 } 2825 2826 // Lower this FP comparison into an appropriate integer version of the 2827 // comparison. 2828 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt); 2829 } 2830 2831 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { 2832 bool Changed = false; 2833 2834 /// Orders the operands of the compare so that they are listed from most 2835 /// complex to least complex. This puts constants before unary operators, 2836 /// before binary operators. 2837 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) { 2838 I.swapOperands(); 2839 Changed = true; 2840 } 2841 2842 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2843 2844 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD)) 2845 return ReplaceInstUsesWith(I, V); 2846 2847 // Simplify 'fcmp pred X, X' 2848 if (Op0 == Op1) { 2849 switch (I.getPredicate()) { 2850 default: llvm_unreachable("Unknown predicate!"); 2851 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y) 2852 case FCmpInst::FCMP_ULT: // True if unordered or less than 2853 case FCmpInst::FCMP_UGT: // True if unordered or greater than 2854 case FCmpInst::FCMP_UNE: // True if unordered or not equal 2855 // Canonicalize these to be 'fcmp uno %X, 0.0'. 2856 I.setPredicate(FCmpInst::FCMP_UNO); 2857 I.setOperand(1, Constant::getNullValue(Op0->getType())); 2858 return &I; 2859 2860 case FCmpInst::FCMP_ORD: // True if ordered (no nans) 2861 case FCmpInst::FCMP_OEQ: // True if ordered and equal 2862 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal 2863 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal 2864 // Canonicalize these to be 'fcmp ord %X, 0.0'. 2865 I.setPredicate(FCmpInst::FCMP_ORD); 2866 I.setOperand(1, Constant::getNullValue(Op0->getType())); 2867 return &I; 2868 } 2869 } 2870 2871 // Handle fcmp with constant RHS 2872 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2873 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2874 switch (LHSI->getOpcode()) { 2875 case Instruction::FPExt: { 2876 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless 2877 FPExtInst *LHSExt = cast<FPExtInst>(LHSI); 2878 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC); 2879 if (!RHSF) 2880 break; 2881 2882 // We can't convert a PPC double double. 2883 if (RHSF->getType()->isPPC_FP128Ty()) 2884 break; 2885 2886 const fltSemantics *Sem; 2887 // FIXME: This shouldn't be here. 2888 if (LHSExt->getSrcTy()->isHalfTy()) 2889 Sem = &APFloat::IEEEhalf; 2890 else if (LHSExt->getSrcTy()->isFloatTy()) 2891 Sem = &APFloat::IEEEsingle; 2892 else if (LHSExt->getSrcTy()->isDoubleTy()) 2893 Sem = &APFloat::IEEEdouble; 2894 else if (LHSExt->getSrcTy()->isFP128Ty()) 2895 Sem = &APFloat::IEEEquad; 2896 else if (LHSExt->getSrcTy()->isX86_FP80Ty()) 2897 Sem = &APFloat::x87DoubleExtended; 2898 else 2899 break; 2900 2901 bool Lossy; 2902 APFloat F = RHSF->getValueAPF(); 2903 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy); 2904 2905 // Avoid lossy conversions and denormals. Zero is a special case 2906 // that's OK to convert. 2907 APFloat Fabs = F; 2908 Fabs.clearSign(); 2909 if (!Lossy && 2910 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) != 2911 APFloat::cmpLessThan) || Fabs.isZero())) 2912 2913 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 2914 ConstantFP::get(RHSC->getContext(), F)); 2915 break; 2916 } 2917 case Instruction::PHI: 2918 // Only fold fcmp into the PHI if the phi and fcmp are in the same 2919 // block. If in the same block, we're encouraging jump threading. If 2920 // not, we are just pessimizing the code by making an i1 phi. 2921 if (LHSI->getParent() == I.getParent()) 2922 if (Instruction *NV = FoldOpIntoPhi(I)) 2923 return NV; 2924 break; 2925 case Instruction::SIToFP: 2926 case Instruction::UIToFP: 2927 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC)) 2928 return NV; 2929 break; 2930 case Instruction::Select: { 2931 // If either operand of the select is a constant, we can fold the 2932 // comparison into the select arms, which will cause one to be 2933 // constant folded and the select turned into a bitwise or. 2934 Value *Op1 = 0, *Op2 = 0; 2935 if (LHSI->hasOneUse()) { 2936 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) { 2937 // Fold the known value into the constant operand. 2938 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 2939 // Insert a new FCmp of the other select operand. 2940 Op2 = Builder->CreateFCmp(I.getPredicate(), 2941 LHSI->getOperand(2), RHSC, I.getName()); 2942 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) { 2943 // Fold the known value into the constant operand. 2944 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 2945 // Insert a new FCmp of the other select operand. 2946 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1), 2947 RHSC, I.getName()); 2948 } 2949 } 2950 2951 if (Op1) 2952 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2953 break; 2954 } 2955 case Instruction::FSub: { 2956 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C 2957 Value *Op; 2958 if (match(LHSI, m_FNeg(m_Value(Op)))) 2959 return new FCmpInst(I.getSwappedPredicate(), Op, 2960 ConstantExpr::getFNeg(RHSC)); 2961 break; 2962 } 2963 case Instruction::Load: 2964 if (GetElementPtrInst *GEP = 2965 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2966 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2967 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2968 !cast<LoadInst>(LHSI)->isVolatile()) 2969 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2970 return Res; 2971 } 2972 break; 2973 } 2974 } 2975 2976 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y 2977 Value *X, *Y; 2978 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) 2979 return new FCmpInst(I.getSwappedPredicate(), X, Y); 2980 2981 // fcmp (fpext x), (fpext y) -> fcmp x, y 2982 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0)) 2983 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1)) 2984 if (LHSExt->getSrcTy() == RHSExt->getSrcTy()) 2985 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 2986 RHSExt->getOperand(0)); 2987 2988 return Changed ? &I : 0; 2989 } 2990