InstCombineAndOrXor.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
1//===- InstCombineAndOrXor.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 visitAnd, visitOr, and visitXor functions. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/Analysis/InstructionSimplify.h" 16#include "llvm/IR/ConstantRange.h" 17#include "llvm/IR/Intrinsics.h" 18#include "llvm/IR/PatternMatch.h" 19#include "llvm/Transforms/Utils/CmpInstAnalysis.h" 20using namespace llvm; 21using namespace PatternMatch; 22 23/// isFreeToInvert - Return true if the specified value is free to invert (apply 24/// ~ to). This happens in cases where the ~ can be eliminated. 25static inline bool isFreeToInvert(Value *V) { 26 // ~(~(X)) -> X. 27 if (BinaryOperator::isNot(V)) 28 return true; 29 30 // Constants can be considered to be not'ed values. 31 if (isa<ConstantInt>(V)) 32 return true; 33 34 // Compares can be inverted if they have a single use. 35 if (CmpInst *CI = dyn_cast<CmpInst>(V)) 36 return CI->hasOneUse(); 37 38 return false; 39} 40 41static inline Value *dyn_castNotVal(Value *V) { 42 // If this is not(not(x)) don't return that this is a not: we want the two 43 // not's to be folded first. 44 if (BinaryOperator::isNot(V)) { 45 Value *Operand = BinaryOperator::getNotArgument(V); 46 if (!isFreeToInvert(Operand)) 47 return Operand; 48 } 49 50 // Constants can be considered to be not'ed values... 51 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 52 return ConstantInt::get(C->getType(), ~C->getValue()); 53 return 0; 54} 55 56/// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp 57/// predicate into a three bit mask. It also returns whether it is an ordered 58/// predicate by reference. 59static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) { 60 isOrdered = false; 61 switch (CC) { 62 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000 63 case FCmpInst::FCMP_UNO: return 0; // 000 64 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001 65 case FCmpInst::FCMP_UGT: return 1; // 001 66 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010 67 case FCmpInst::FCMP_UEQ: return 2; // 010 68 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011 69 case FCmpInst::FCMP_UGE: return 3; // 011 70 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100 71 case FCmpInst::FCMP_ULT: return 4; // 100 72 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101 73 case FCmpInst::FCMP_UNE: return 5; // 101 74 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110 75 case FCmpInst::FCMP_ULE: return 6; // 110 76 // True -> 7 77 default: 78 // Not expecting FCMP_FALSE and FCMP_TRUE; 79 llvm_unreachable("Unexpected FCmp predicate!"); 80 } 81} 82 83/// getNewICmpValue - This is the complement of getICmpCode, which turns an 84/// opcode and two operands into either a constant true or false, or a brand 85/// new ICmp instruction. The sign is passed in to determine which kind 86/// of predicate to use in the new icmp instruction. 87static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS, 88 InstCombiner::BuilderTy *Builder) { 89 ICmpInst::Predicate NewPred; 90 if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred)) 91 return NewConstant; 92 return Builder->CreateICmp(NewPred, LHS, RHS); 93} 94 95/// getFCmpValue - This is the complement of getFCmpCode, which turns an 96/// opcode and two operands into either a FCmp instruction. isordered is passed 97/// in to determine which kind of predicate to use in the new fcmp instruction. 98static Value *getFCmpValue(bool isordered, unsigned code, 99 Value *LHS, Value *RHS, 100 InstCombiner::BuilderTy *Builder) { 101 CmpInst::Predicate Pred; 102 switch (code) { 103 default: llvm_unreachable("Illegal FCmp code!"); 104 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break; 105 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break; 106 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break; 107 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break; 108 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break; 109 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break; 110 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break; 111 case 7: 112 if (!isordered) return ConstantInt::getTrue(LHS->getContext()); 113 Pred = FCmpInst::FCMP_ORD; break; 114 } 115 return Builder->CreateFCmp(Pred, LHS, RHS); 116} 117 118// OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where 119// the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is 120// guaranteed to be a binary operator. 121Instruction *InstCombiner::OptAndOp(Instruction *Op, 122 ConstantInt *OpRHS, 123 ConstantInt *AndRHS, 124 BinaryOperator &TheAnd) { 125 Value *X = Op->getOperand(0); 126 Constant *Together = 0; 127 if (!Op->isShift()) 128 Together = ConstantExpr::getAnd(AndRHS, OpRHS); 129 130 switch (Op->getOpcode()) { 131 case Instruction::Xor: 132 if (Op->hasOneUse()) { 133 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2) 134 Value *And = Builder->CreateAnd(X, AndRHS); 135 And->takeName(Op); 136 return BinaryOperator::CreateXor(And, Together); 137 } 138 break; 139 case Instruction::Or: 140 if (Op->hasOneUse()){ 141 if (Together != OpRHS) { 142 // (X | C1) & C2 --> (X | (C1&C2)) & C2 143 Value *Or = Builder->CreateOr(X, Together); 144 Or->takeName(Op); 145 return BinaryOperator::CreateAnd(Or, AndRHS); 146 } 147 148 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together); 149 if (TogetherCI && !TogetherCI->isZero()){ 150 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1 151 // NOTE: This reduces the number of bits set in the & mask, which 152 // can expose opportunities for store narrowing. 153 Together = ConstantExpr::getXor(AndRHS, Together); 154 Value *And = Builder->CreateAnd(X, Together); 155 And->takeName(Op); 156 return BinaryOperator::CreateOr(And, OpRHS); 157 } 158 } 159 160 break; 161 case Instruction::Add: 162 if (Op->hasOneUse()) { 163 // Adding a one to a single bit bit-field should be turned into an XOR 164 // of the bit. First thing to check is to see if this AND is with a 165 // single bit constant. 166 const APInt &AndRHSV = AndRHS->getValue(); 167 168 // If there is only one bit set. 169 if (AndRHSV.isPowerOf2()) { 170 // Ok, at this point, we know that we are masking the result of the 171 // ADD down to exactly one bit. If the constant we are adding has 172 // no bits set below this bit, then we can eliminate the ADD. 173 const APInt& AddRHS = OpRHS->getValue(); 174 175 // Check to see if any bits below the one bit set in AndRHSV are set. 176 if ((AddRHS & (AndRHSV-1)) == 0) { 177 // If not, the only thing that can effect the output of the AND is 178 // the bit specified by AndRHSV. If that bit is set, the effect of 179 // the XOR is to toggle the bit. If it is clear, then the ADD has 180 // no effect. 181 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop 182 TheAnd.setOperand(0, X); 183 return &TheAnd; 184 } else { 185 // Pull the XOR out of the AND. 186 Value *NewAnd = Builder->CreateAnd(X, AndRHS); 187 NewAnd->takeName(Op); 188 return BinaryOperator::CreateXor(NewAnd, AndRHS); 189 } 190 } 191 } 192 } 193 break; 194 195 case Instruction::Shl: { 196 // We know that the AND will not produce any of the bits shifted in, so if 197 // the anded constant includes them, clear them now! 198 // 199 uint32_t BitWidth = AndRHS->getType()->getBitWidth(); 200 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth); 201 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal)); 202 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask); 203 204 if (CI->getValue() == ShlMask) 205 // Masking out bits that the shift already masks. 206 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and. 207 208 if (CI != AndRHS) { // Reducing bits set in and. 209 TheAnd.setOperand(1, CI); 210 return &TheAnd; 211 } 212 break; 213 } 214 case Instruction::LShr: { 215 // We know that the AND will not produce any of the bits shifted in, so if 216 // the anded constant includes them, clear them now! This only applies to 217 // unsigned shifts, because a signed shr may bring in set bits! 218 // 219 uint32_t BitWidth = AndRHS->getType()->getBitWidth(); 220 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth); 221 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal)); 222 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask); 223 224 if (CI->getValue() == ShrMask) 225 // Masking out bits that the shift already masks. 226 return ReplaceInstUsesWith(TheAnd, Op); 227 228 if (CI != AndRHS) { 229 TheAnd.setOperand(1, CI); // Reduce bits set in and cst. 230 return &TheAnd; 231 } 232 break; 233 } 234 case Instruction::AShr: 235 // Signed shr. 236 // See if this is shifting in some sign extension, then masking it out 237 // with an and. 238 if (Op->hasOneUse()) { 239 uint32_t BitWidth = AndRHS->getType()->getBitWidth(); 240 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth); 241 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal)); 242 Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask); 243 if (C == AndRHS) { // Masking out bits shifted in. 244 // (Val ashr C1) & C2 -> (Val lshr C1) & C2 245 // Make the argument unsigned. 246 Value *ShVal = Op->getOperand(0); 247 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName()); 248 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName()); 249 } 250 } 251 break; 252 } 253 return 0; 254} 255 256/// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise 257/// (V < Lo || V >= Hi). In practice, we emit the more efficient 258/// (V-Lo) \<u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates 259/// whether to treat the V, Lo and HI as signed or not. IB is the location to 260/// insert new instructions. 261Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi, 262 bool isSigned, bool Inside) { 263 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ? 264 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() && 265 "Lo is not <= Hi in range emission code!"); 266 267 if (Inside) { 268 if (Lo == Hi) // Trivially false. 269 return Builder->getFalse(); 270 271 // V >= Min && V < Hi --> V < Hi 272 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) { 273 ICmpInst::Predicate pred = (isSigned ? 274 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT); 275 return Builder->CreateICmp(pred, V, Hi); 276 } 277 278 // Emit V-Lo <u Hi-Lo 279 Constant *NegLo = ConstantExpr::getNeg(Lo); 280 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off"); 281 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi); 282 return Builder->CreateICmpULT(Add, UpperBound); 283 } 284 285 if (Lo == Hi) // Trivially true. 286 return Builder->getTrue(); 287 288 // V < Min || V >= Hi -> V > Hi-1 289 Hi = SubOne(cast<ConstantInt>(Hi)); 290 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) { 291 ICmpInst::Predicate pred = (isSigned ? 292 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT); 293 return Builder->CreateICmp(pred, V, Hi); 294 } 295 296 // Emit V-Lo >u Hi-1-Lo 297 // Note that Hi has already had one subtracted from it, above. 298 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo)); 299 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off"); 300 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi); 301 return Builder->CreateICmpUGT(Add, LowerBound); 302} 303 304// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with 305// any number of 0s on either side. The 1s are allowed to wrap from LSB to 306// MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is 307// not, since all 1s are not contiguous. 308static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) { 309 const APInt& V = Val->getValue(); 310 uint32_t BitWidth = Val->getType()->getBitWidth(); 311 if (!APIntOps::isShiftedMask(BitWidth, V)) return false; 312 313 // look for the first zero bit after the run of ones 314 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros(); 315 // look for the first non-zero bit 316 ME = V.getActiveBits(); 317 return true; 318} 319 320/// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask, 321/// where isSub determines whether the operator is a sub. If we can fold one of 322/// the following xforms: 323/// 324/// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask 325/// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0 326/// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0 327/// 328/// return (A +/- B). 329/// 330Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS, 331 ConstantInt *Mask, bool isSub, 332 Instruction &I) { 333 Instruction *LHSI = dyn_cast<Instruction>(LHS); 334 if (!LHSI || LHSI->getNumOperands() != 2 || 335 !isa<ConstantInt>(LHSI->getOperand(1))) return 0; 336 337 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1)); 338 339 switch (LHSI->getOpcode()) { 340 default: return 0; 341 case Instruction::And: 342 if (ConstantExpr::getAnd(N, Mask) == Mask) { 343 // If the AndRHS is a power of two minus one (0+1+), this is simple. 344 if ((Mask->getValue().countLeadingZeros() + 345 Mask->getValue().countPopulation()) == 346 Mask->getValue().getBitWidth()) 347 break; 348 349 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+ 350 // part, we don't need any explicit masks to take them out of A. If that 351 // is all N is, ignore it. 352 uint32_t MB = 0, ME = 0; 353 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive 354 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth(); 355 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1)); 356 if (MaskedValueIsZero(RHS, Mask)) 357 break; 358 } 359 } 360 return 0; 361 case Instruction::Or: 362 case Instruction::Xor: 363 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0 364 if ((Mask->getValue().countLeadingZeros() + 365 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth() 366 && ConstantExpr::getAnd(N, Mask)->isNullValue()) 367 break; 368 return 0; 369 } 370 371 if (isSub) 372 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold"); 373 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold"); 374} 375 376/// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C) 377/// One of A and B is considered the mask, the other the value. This is 378/// described as the "AMask" or "BMask" part of the enum. If the enum 379/// contains only "Mask", then both A and B can be considered masks. 380/// If A is the mask, then it was proven, that (A & C) == C. This 381/// is trivial if C == A, or C == 0. If both A and C are constants, this 382/// proof is also easy. 383/// For the following explanations we assume that A is the mask. 384/// The part "AllOnes" declares, that the comparison is true only 385/// if (A & B) == A, or all bits of A are set in B. 386/// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes 387/// The part "AllZeroes" declares, that the comparison is true only 388/// if (A & B) == 0, or all bits of A are cleared in B. 389/// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes 390/// The part "Mixed" declares, that (A & B) == C and C might or might not 391/// contain any number of one bits and zero bits. 392/// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed 393/// The Part "Not" means, that in above descriptions "==" should be replaced 394/// by "!=". 395/// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes 396/// If the mask A contains a single bit, then the following is equivalent: 397/// (icmp eq (A & B), A) equals (icmp ne (A & B), 0) 398/// (icmp ne (A & B), A) equals (icmp eq (A & B), 0) 399enum MaskedICmpType { 400 FoldMskICmp_AMask_AllOnes = 1, 401 FoldMskICmp_AMask_NotAllOnes = 2, 402 FoldMskICmp_BMask_AllOnes = 4, 403 FoldMskICmp_BMask_NotAllOnes = 8, 404 FoldMskICmp_Mask_AllZeroes = 16, 405 FoldMskICmp_Mask_NotAllZeroes = 32, 406 FoldMskICmp_AMask_Mixed = 64, 407 FoldMskICmp_AMask_NotMixed = 128, 408 FoldMskICmp_BMask_Mixed = 256, 409 FoldMskICmp_BMask_NotMixed = 512 410}; 411 412/// return the set of pattern classes (from MaskedICmpType) 413/// that (icmp SCC (A & B), C) satisfies 414static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C, 415 ICmpInst::Predicate SCC) 416{ 417 ConstantInt *ACst = dyn_cast<ConstantInt>(A); 418 ConstantInt *BCst = dyn_cast<ConstantInt>(B); 419 ConstantInt *CCst = dyn_cast<ConstantInt>(C); 420 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ); 421 bool icmp_abit = (ACst != 0 && !ACst->isZero() && 422 ACst->getValue().isPowerOf2()); 423 bool icmp_bbit = (BCst != 0 && !BCst->isZero() && 424 BCst->getValue().isPowerOf2()); 425 unsigned result = 0; 426 if (CCst != 0 && CCst->isZero()) { 427 // if C is zero, then both A and B qualify as mask 428 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes | 429 FoldMskICmp_Mask_AllZeroes | 430 FoldMskICmp_AMask_Mixed | 431 FoldMskICmp_BMask_Mixed) 432 : (FoldMskICmp_Mask_NotAllZeroes | 433 FoldMskICmp_Mask_NotAllZeroes | 434 FoldMskICmp_AMask_NotMixed | 435 FoldMskICmp_BMask_NotMixed)); 436 if (icmp_abit) 437 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes | 438 FoldMskICmp_AMask_NotMixed) 439 : (FoldMskICmp_AMask_AllOnes | 440 FoldMskICmp_AMask_Mixed)); 441 if (icmp_bbit) 442 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes | 443 FoldMskICmp_BMask_NotMixed) 444 : (FoldMskICmp_BMask_AllOnes | 445 FoldMskICmp_BMask_Mixed)); 446 return result; 447 } 448 if (A == C) { 449 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes | 450 FoldMskICmp_AMask_Mixed) 451 : (FoldMskICmp_AMask_NotAllOnes | 452 FoldMskICmp_AMask_NotMixed)); 453 if (icmp_abit) 454 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes | 455 FoldMskICmp_AMask_NotMixed) 456 : (FoldMskICmp_Mask_AllZeroes | 457 FoldMskICmp_AMask_Mixed)); 458 } else if (ACst != 0 && CCst != 0 && 459 ConstantExpr::getAnd(ACst, CCst) == CCst) { 460 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed 461 : FoldMskICmp_AMask_NotMixed); 462 } 463 if (B == C) { 464 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes | 465 FoldMskICmp_BMask_Mixed) 466 : (FoldMskICmp_BMask_NotAllOnes | 467 FoldMskICmp_BMask_NotMixed)); 468 if (icmp_bbit) 469 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes | 470 FoldMskICmp_BMask_NotMixed) 471 : (FoldMskICmp_Mask_AllZeroes | 472 FoldMskICmp_BMask_Mixed)); 473 } else if (BCst != 0 && CCst != 0 && 474 ConstantExpr::getAnd(BCst, CCst) == CCst) { 475 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed 476 : FoldMskICmp_BMask_NotMixed); 477 } 478 return result; 479} 480 481/// Convert an analysis of a masked ICmp into its equivalent if all boolean 482/// operations had the opposite sense. Since each "NotXXX" flag (recording !=) 483/// is adjacent to the corresponding normal flag (recording ==), this just 484/// involves swapping those bits over. 485static unsigned conjugateICmpMask(unsigned Mask) { 486 unsigned NewMask; 487 NewMask = (Mask & (FoldMskICmp_AMask_AllOnes | FoldMskICmp_BMask_AllOnes | 488 FoldMskICmp_Mask_AllZeroes | FoldMskICmp_AMask_Mixed | 489 FoldMskICmp_BMask_Mixed)) 490 << 1; 491 492 NewMask |= 493 (Mask & (FoldMskICmp_AMask_NotAllOnes | FoldMskICmp_BMask_NotAllOnes | 494 FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_AMask_NotMixed | 495 FoldMskICmp_BMask_NotMixed)) 496 >> 1; 497 498 return NewMask; 499} 500 501/// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z) 502/// if possible. The returned predicate is either == or !=. Returns false if 503/// decomposition fails. 504static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred, 505 Value *&X, Value *&Y, Value *&Z) { 506 ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)); 507 if (!C) 508 return false; 509 510 switch (I->getPredicate()) { 511 default: 512 return false; 513 case ICmpInst::ICMP_SLT: 514 // X < 0 is equivalent to (X & SignBit) != 0. 515 if (!C->isZero()) 516 return false; 517 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth())); 518 Pred = ICmpInst::ICMP_NE; 519 break; 520 case ICmpInst::ICMP_SGT: 521 // X > -1 is equivalent to (X & SignBit) == 0. 522 if (!C->isAllOnesValue()) 523 return false; 524 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth())); 525 Pred = ICmpInst::ICMP_EQ; 526 break; 527 case ICmpInst::ICMP_ULT: 528 // X <u 2^n is equivalent to (X & ~(2^n-1)) == 0. 529 if (!C->getValue().isPowerOf2()) 530 return false; 531 Y = ConstantInt::get(I->getContext(), -C->getValue()); 532 Pred = ICmpInst::ICMP_EQ; 533 break; 534 case ICmpInst::ICMP_UGT: 535 // X >u 2^n-1 is equivalent to (X & ~(2^n-1)) != 0. 536 if (!(C->getValue() + 1).isPowerOf2()) 537 return false; 538 Y = ConstantInt::get(I->getContext(), ~C->getValue()); 539 Pred = ICmpInst::ICMP_NE; 540 break; 541 } 542 543 X = I->getOperand(0); 544 Z = ConstantInt::getNullValue(C->getType()); 545 return true; 546} 547 548/// foldLogOpOfMaskedICmpsHelper: 549/// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) 550/// return the set of pattern classes (from MaskedICmpType) 551/// that both LHS and RHS satisfy 552static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A, 553 Value*& B, Value*& C, 554 Value*& D, Value*& E, 555 ICmpInst *LHS, ICmpInst *RHS, 556 ICmpInst::Predicate &LHSCC, 557 ICmpInst::Predicate &RHSCC) { 558 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0; 559 // vectors are not (yet?) supported 560 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0; 561 562 // Here comes the tricky part: 563 // LHS might be of the form L11 & L12 == X, X == L21 & L22, 564 // and L11 & L12 == L21 & L22. The same goes for RHS. 565 // Now we must find those components L** and R**, that are equal, so 566 // that we can extract the parameters A, B, C, D, and E for the canonical 567 // above. 568 Value *L1 = LHS->getOperand(0); 569 Value *L2 = LHS->getOperand(1); 570 Value *L11,*L12,*L21,*L22; 571 // Check whether the icmp can be decomposed into a bit test. 572 if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) { 573 L21 = L22 = L1 = 0; 574 } else { 575 // Look for ANDs in the LHS icmp. 576 if (!L1->getType()->isIntegerTy()) { 577 // You can icmp pointers, for example. They really aren't masks. 578 L11 = L12 = 0; 579 } else if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) { 580 // Any icmp can be viewed as being trivially masked; if it allows us to 581 // remove one, it's worth it. 582 L11 = L1; 583 L12 = Constant::getAllOnesValue(L1->getType()); 584 } 585 586 if (!L2->getType()->isIntegerTy()) { 587 // You can icmp pointers, for example. They really aren't masks. 588 L21 = L22 = 0; 589 } else if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) { 590 L21 = L2; 591 L22 = Constant::getAllOnesValue(L2->getType()); 592 } 593 } 594 595 // Bail if LHS was a icmp that can't be decomposed into an equality. 596 if (!ICmpInst::isEquality(LHSCC)) 597 return 0; 598 599 Value *R1 = RHS->getOperand(0); 600 Value *R2 = RHS->getOperand(1); 601 Value *R11,*R12; 602 bool ok = false; 603 if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) { 604 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 605 A = R11; D = R12; 606 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 607 A = R12; D = R11; 608 } else { 609 return 0; 610 } 611 E = R2; R1 = 0; ok = true; 612 } else if (R1->getType()->isIntegerTy()) { 613 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) { 614 // As before, model no mask as a trivial mask if it'll let us do an 615 // optimisation. 616 R11 = R1; 617 R12 = Constant::getAllOnesValue(R1->getType()); 618 } 619 620 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 621 A = R11; D = R12; E = R2; ok = true; 622 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 623 A = R12; D = R11; E = R2; ok = true; 624 } 625 } 626 627 // Bail if RHS was a icmp that can't be decomposed into an equality. 628 if (!ICmpInst::isEquality(RHSCC)) 629 return 0; 630 631 // Look for ANDs in on the right side of the RHS icmp. 632 if (!ok && R2->getType()->isIntegerTy()) { 633 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) { 634 R11 = R2; 635 R12 = Constant::getAllOnesValue(R2->getType()); 636 } 637 638 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 639 A = R11; D = R12; E = R1; ok = true; 640 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 641 A = R12; D = R11; E = R1; ok = true; 642 } else { 643 return 0; 644 } 645 } 646 if (!ok) 647 return 0; 648 649 if (L11 == A) { 650 B = L12; C = L2; 651 } else if (L12 == A) { 652 B = L11; C = L2; 653 } else if (L21 == A) { 654 B = L22; C = L1; 655 } else if (L22 == A) { 656 B = L21; C = L1; 657 } 658 659 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC); 660 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC); 661 return left_type & right_type; 662} 663/// foldLogOpOfMaskedICmps: 664/// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) 665/// into a single (icmp(A & X) ==/!= Y) 666static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, 667 llvm::InstCombiner::BuilderTy* Builder) { 668 Value *A = 0, *B = 0, *C = 0, *D = 0, *E = 0; 669 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate(); 670 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS, 671 LHSCC, RHSCC); 672 if (mask == 0) return 0; 673 assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) && 674 "foldLogOpOfMaskedICmpsHelper must return an equality predicate."); 675 676 // In full generality: 677 // (icmp (A & B) Op C) | (icmp (A & D) Op E) 678 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ] 679 // 680 // If the latter can be converted into (icmp (A & X) Op Y) then the former is 681 // equivalent to (icmp (A & X) !Op Y). 682 // 683 // Therefore, we can pretend for the rest of this function that we're dealing 684 // with the conjunction, provided we flip the sense of any comparisons (both 685 // input and output). 686 687 // In most cases we're going to produce an EQ for the "&&" case. 688 ICmpInst::Predicate NEWCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE; 689 if (!IsAnd) { 690 // Convert the masking analysis into its equivalent with negated 691 // comparisons. 692 mask = conjugateICmpMask(mask); 693 } 694 695 if (mask & FoldMskICmp_Mask_AllZeroes) { 696 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0) 697 // -> (icmp eq (A & (B|D)), 0) 698 Value* newOr = Builder->CreateOr(B, D); 699 Value* newAnd = Builder->CreateAnd(A, newOr); 700 // we can't use C as zero, because we might actually handle 701 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 702 // with B and D, having a single bit set 703 Value* zero = Constant::getNullValue(A->getType()); 704 return Builder->CreateICmp(NEWCC, newAnd, zero); 705 } 706 if (mask & FoldMskICmp_BMask_AllOnes) { 707 // (icmp eq (A & B), B) & (icmp eq (A & D), D) 708 // -> (icmp eq (A & (B|D)), (B|D)) 709 Value* newOr = Builder->CreateOr(B, D); 710 Value* newAnd = Builder->CreateAnd(A, newOr); 711 return Builder->CreateICmp(NEWCC, newAnd, newOr); 712 } 713 if (mask & FoldMskICmp_AMask_AllOnes) { 714 // (icmp eq (A & B), A) & (icmp eq (A & D), A) 715 // -> (icmp eq (A & (B&D)), A) 716 Value* newAnd1 = Builder->CreateAnd(B, D); 717 Value* newAnd = Builder->CreateAnd(A, newAnd1); 718 return Builder->CreateICmp(NEWCC, newAnd, A); 719 } 720 721 // Remaining cases assume at least that B and D are constant, and depend on 722 // their actual values. This isn't strictly, necessary, just a "handle the 723 // easy cases for now" decision. 724 ConstantInt *BCst = dyn_cast<ConstantInt>(B); 725 if (BCst == 0) return 0; 726 ConstantInt *DCst = dyn_cast<ConstantInt>(D); 727 if (DCst == 0) return 0; 728 729 if (mask & (FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_BMask_NotAllOnes)) { 730 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and 731 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 732 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0) 733 // Only valid if one of the masks is a superset of the other (check "B&D" is 734 // the same as either B or D). 735 APInt NewMask = BCst->getValue() & DCst->getValue(); 736 737 if (NewMask == BCst->getValue()) 738 return LHS; 739 else if (NewMask == DCst->getValue()) 740 return RHS; 741 } 742 if (mask & FoldMskICmp_AMask_NotAllOnes) { 743 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 744 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A) 745 // Only valid if one of the masks is a superset of the other (check "B|D" is 746 // the same as either B or D). 747 APInt NewMask = BCst->getValue() | DCst->getValue(); 748 749 if (NewMask == BCst->getValue()) 750 return LHS; 751 else if (NewMask == DCst->getValue()) 752 return RHS; 753 } 754 if (mask & FoldMskICmp_BMask_Mixed) { 755 // (icmp eq (A & B), C) & (icmp eq (A & D), E) 756 // We already know that B & C == C && D & E == E. 757 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of 758 // C and E, which are shared by both the mask B and the mask D, don't 759 // contradict, then we can transform to 760 // -> (icmp eq (A & (B|D)), (C|E)) 761 // Currently, we only handle the case of B, C, D, and E being constant. 762 // we can't simply use C and E, because we might actually handle 763 // (icmp ne (A & B), B) & (icmp eq (A & D), D) 764 // with B and D, having a single bit set 765 ConstantInt *CCst = dyn_cast<ConstantInt>(C); 766 if (CCst == 0) return 0; 767 if (LHSCC != NEWCC) 768 CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) ); 769 ConstantInt *ECst = dyn_cast<ConstantInt>(E); 770 if (ECst == 0) return 0; 771 if (RHSCC != NEWCC) 772 ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) ); 773 ConstantInt* MCst = dyn_cast<ConstantInt>( 774 ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst), 775 ConstantExpr::getXor(CCst, ECst)) ); 776 // if there is a conflict we should actually return a false for the 777 // whole construct 778 if (!MCst->isZero()) 779 return 0; 780 Value *newOr1 = Builder->CreateOr(B, D); 781 Value *newOr2 = ConstantExpr::getOr(CCst, ECst); 782 Value *newAnd = Builder->CreateAnd(A, newOr1); 783 return Builder->CreateICmp(NEWCC, newAnd, newOr2); 784 } 785 return 0; 786} 787 788/// FoldAndOfICmps - Fold (icmp)&(icmp) if possible. 789Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) { 790 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate(); 791 792 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B) 793 if (PredicatesFoldable(LHSCC, RHSCC)) { 794 if (LHS->getOperand(0) == RHS->getOperand(1) && 795 LHS->getOperand(1) == RHS->getOperand(0)) 796 LHS->swapOperands(); 797 if (LHS->getOperand(0) == RHS->getOperand(0) && 798 LHS->getOperand(1) == RHS->getOperand(1)) { 799 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 800 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS); 801 bool isSigned = LHS->isSigned() || RHS->isSigned(); 802 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder); 803 } 804 } 805 806 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E) 807 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder)) 808 return V; 809 810 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2). 811 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0); 812 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1)); 813 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1)); 814 if (LHSCst == 0 || RHSCst == 0) return 0; 815 816 if (LHSCst == RHSCst && LHSCC == RHSCC) { 817 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C) 818 // where C is a power of 2 819 if (LHSCC == ICmpInst::ICMP_ULT && 820 LHSCst->getValue().isPowerOf2()) { 821 Value *NewOr = Builder->CreateOr(Val, Val2); 822 return Builder->CreateICmp(LHSCC, NewOr, LHSCst); 823 } 824 825 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0) 826 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) { 827 Value *NewOr = Builder->CreateOr(Val, Val2); 828 return Builder->CreateICmp(LHSCC, NewOr, LHSCst); 829 } 830 } 831 832 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2 833 // where CMAX is the all ones value for the truncated type, 834 // iff the lower bits of C2 and CA are zero. 835 if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC && 836 LHS->hasOneUse() && RHS->hasOneUse()) { 837 Value *V; 838 ConstantInt *AndCst, *SmallCst = 0, *BigCst = 0; 839 840 // (trunc x) == C1 & (and x, CA) == C2 841 // (and x, CA) == C2 & (trunc x) == C1 842 if (match(Val2, m_Trunc(m_Value(V))) && 843 match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) { 844 SmallCst = RHSCst; 845 BigCst = LHSCst; 846 } else if (match(Val, m_Trunc(m_Value(V))) && 847 match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) { 848 SmallCst = LHSCst; 849 BigCst = RHSCst; 850 } 851 852 if (SmallCst && BigCst) { 853 unsigned BigBitSize = BigCst->getType()->getBitWidth(); 854 unsigned SmallBitSize = SmallCst->getType()->getBitWidth(); 855 856 // Check that the low bits are zero. 857 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize); 858 if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) { 859 Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue()); 860 APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue(); 861 Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N); 862 return Builder->CreateICmp(LHSCC, NewAnd, NewVal); 863 } 864 } 865 } 866 867 // From here on, we only handle: 868 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler. 869 if (Val != Val2) return 0; 870 871 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere. 872 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE || 873 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE || 874 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE || 875 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE) 876 return 0; 877 878 // Make a constant range that's the intersection of the two icmp ranges. 879 // If the intersection is empty, we know that the result is false. 880 ConstantRange LHSRange = 881 ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue()); 882 ConstantRange RHSRange = 883 ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue()); 884 885 if (LHSRange.intersectWith(RHSRange).isEmptySet()) 886 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 887 888 // We can't fold (ugt x, C) & (sgt x, C2). 889 if (!PredicatesFoldable(LHSCC, RHSCC)) 890 return 0; 891 892 // Ensure that the larger constant is on the RHS. 893 bool ShouldSwap; 894 if (CmpInst::isSigned(LHSCC) || 895 (ICmpInst::isEquality(LHSCC) && 896 CmpInst::isSigned(RHSCC))) 897 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue()); 898 else 899 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue()); 900 901 if (ShouldSwap) { 902 std::swap(LHS, RHS); 903 std::swap(LHSCst, RHSCst); 904 std::swap(LHSCC, RHSCC); 905 } 906 907 // At this point, we know we have two icmp instructions 908 // comparing a value against two constants and and'ing the result 909 // together. Because of the above check, we know that we only have 910 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know 911 // (from the icmp folding check above), that the two constants 912 // are not equal and that the larger constant is on the RHS 913 assert(LHSCst != RHSCst && "Compares not folded above?"); 914 915 switch (LHSCC) { 916 default: llvm_unreachable("Unknown integer condition code!"); 917 case ICmpInst::ICMP_EQ: 918 switch (RHSCC) { 919 default: llvm_unreachable("Unknown integer condition code!"); 920 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13 921 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13 922 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13 923 return LHS; 924 } 925 case ICmpInst::ICMP_NE: 926 switch (RHSCC) { 927 default: llvm_unreachable("Unknown integer condition code!"); 928 case ICmpInst::ICMP_ULT: 929 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13 930 return Builder->CreateICmpULT(Val, LHSCst); 931 break; // (X != 13 & X u< 15) -> no change 932 case ICmpInst::ICMP_SLT: 933 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13 934 return Builder->CreateICmpSLT(Val, LHSCst); 935 break; // (X != 13 & X s< 15) -> no change 936 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15 937 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15 938 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15 939 return RHS; 940 case ICmpInst::ICMP_NE: 941 // Special case to get the ordering right when the values wrap around 942 // zero. 943 if (LHSCst->getValue() == 0 && RHSCst->getValue().isAllOnesValue()) 944 std::swap(LHSCst, RHSCst); 945 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1 946 Constant *AddCST = ConstantExpr::getNeg(LHSCst); 947 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off"); 948 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1), 949 Val->getName()+".cmp"); 950 } 951 break; // (X != 13 & X != 15) -> no change 952 } 953 break; 954 case ICmpInst::ICMP_ULT: 955 switch (RHSCC) { 956 default: llvm_unreachable("Unknown integer condition code!"); 957 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false 958 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false 959 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 960 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change 961 break; 962 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13 963 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13 964 return LHS; 965 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change 966 break; 967 } 968 break; 969 case ICmpInst::ICMP_SLT: 970 switch (RHSCC) { 971 default: llvm_unreachable("Unknown integer condition code!"); 972 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change 973 break; 974 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13 975 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13 976 return LHS; 977 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change 978 break; 979 } 980 break; 981 case ICmpInst::ICMP_UGT: 982 switch (RHSCC) { 983 default: llvm_unreachable("Unknown integer condition code!"); 984 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15 985 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15 986 return RHS; 987 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change 988 break; 989 case ICmpInst::ICMP_NE: 990 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14 991 return Builder->CreateICmp(LHSCC, Val, RHSCst); 992 break; // (X u> 13 & X != 15) -> no change 993 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1 994 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true); 995 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change 996 break; 997 } 998 break; 999 case ICmpInst::ICMP_SGT: 1000 switch (RHSCC) { 1001 default: llvm_unreachable("Unknown integer condition code!"); 1002 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15 1003 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15 1004 return RHS; 1005 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change 1006 break; 1007 case ICmpInst::ICMP_NE: 1008 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14 1009 return Builder->CreateICmp(LHSCC, Val, RHSCst); 1010 break; // (X s> 13 & X != 15) -> no change 1011 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1 1012 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true); 1013 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change 1014 break; 1015 } 1016 break; 1017 } 1018 1019 return 0; 1020} 1021 1022/// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of 1023/// instcombine, this returns a Value which should already be inserted into the 1024/// function. 1025Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) { 1026 if (LHS->getPredicate() == FCmpInst::FCMP_ORD && 1027 RHS->getPredicate() == FCmpInst::FCMP_ORD) { 1028 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) 1029 return 0; 1030 1031 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y) 1032 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1))) 1033 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) { 1034 // If either of the constants are nans, then the whole thing returns 1035 // false. 1036 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN()) 1037 return Builder->getFalse(); 1038 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0)); 1039 } 1040 1041 // Handle vector zeros. This occurs because the canonical form of 1042 // "fcmp ord x,x" is "fcmp ord x, 0". 1043 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) && 1044 isa<ConstantAggregateZero>(RHS->getOperand(1))) 1045 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0)); 1046 return 0; 1047 } 1048 1049 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1); 1050 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1); 1051 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate(); 1052 1053 1054 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) { 1055 // Swap RHS operands to match LHS. 1056 Op1CC = FCmpInst::getSwappedPredicate(Op1CC); 1057 std::swap(Op1LHS, Op1RHS); 1058 } 1059 1060 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) { 1061 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y). 1062 if (Op0CC == Op1CC) 1063 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS); 1064 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE) 1065 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 1066 if (Op0CC == FCmpInst::FCMP_TRUE) 1067 return RHS; 1068 if (Op1CC == FCmpInst::FCMP_TRUE) 1069 return LHS; 1070 1071 bool Op0Ordered; 1072 bool Op1Ordered; 1073 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered); 1074 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered); 1075 // uno && ord -> false 1076 if (Op0Pred == 0 && Op1Pred == 0 && Op0Ordered != Op1Ordered) 1077 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 1078 if (Op1Pred == 0) { 1079 std::swap(LHS, RHS); 1080 std::swap(Op0Pred, Op1Pred); 1081 std::swap(Op0Ordered, Op1Ordered); 1082 } 1083 if (Op0Pred == 0) { 1084 // uno && ueq -> uno && (uno || eq) -> uno 1085 // ord && olt -> ord && (ord && lt) -> olt 1086 if (!Op0Ordered && (Op0Ordered == Op1Ordered)) 1087 return LHS; 1088 if (Op0Ordered && (Op0Ordered == Op1Ordered)) 1089 return RHS; 1090 1091 // uno && oeq -> uno && (ord && eq) -> false 1092 if (!Op0Ordered) 1093 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 1094 // ord && ueq -> ord && (uno || eq) -> oeq 1095 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder); 1096 } 1097 } 1098 1099 return 0; 1100} 1101 1102 1103Instruction *InstCombiner::visitAnd(BinaryOperator &I) { 1104 bool Changed = SimplifyAssociativeOrCommutative(I); 1105 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1106 1107 if (Value *V = SimplifyAndInst(Op0, Op1, DL)) 1108 return ReplaceInstUsesWith(I, V); 1109 1110 // (A|B)&(A|C) -> A|(B&C) etc 1111 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1112 return ReplaceInstUsesWith(I, V); 1113 1114 // See if we can simplify any instructions used by the instruction whose sole 1115 // purpose is to compute bits we don't care about. 1116 if (SimplifyDemandedInstructionBits(I)) 1117 return &I; 1118 1119 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) { 1120 const APInt &AndRHSMask = AndRHS->getValue(); 1121 1122 // Optimize a variety of ((val OP C1) & C2) combinations... 1123 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) { 1124 Value *Op0LHS = Op0I->getOperand(0); 1125 Value *Op0RHS = Op0I->getOperand(1); 1126 switch (Op0I->getOpcode()) { 1127 default: break; 1128 case Instruction::Xor: 1129 case Instruction::Or: { 1130 // If the mask is only needed on one incoming arm, push it up. 1131 if (!Op0I->hasOneUse()) break; 1132 1133 APInt NotAndRHS(~AndRHSMask); 1134 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) { 1135 // Not masking anything out for the LHS, move to RHS. 1136 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS, 1137 Op0RHS->getName()+".masked"); 1138 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS); 1139 } 1140 if (!isa<Constant>(Op0RHS) && 1141 MaskedValueIsZero(Op0RHS, NotAndRHS)) { 1142 // Not masking anything out for the RHS, move to LHS. 1143 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS, 1144 Op0LHS->getName()+".masked"); 1145 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS); 1146 } 1147 1148 break; 1149 } 1150 case Instruction::Add: 1151 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS. 1152 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0 1153 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0 1154 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I)) 1155 return BinaryOperator::CreateAnd(V, AndRHS); 1156 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I)) 1157 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes 1158 break; 1159 1160 case Instruction::Sub: 1161 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS. 1162 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0 1163 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0 1164 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I)) 1165 return BinaryOperator::CreateAnd(V, AndRHS); 1166 1167 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS 1168 // has 1's for all bits that the subtraction with A might affect. 1169 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) { 1170 uint32_t BitWidth = AndRHSMask.getBitWidth(); 1171 uint32_t Zeros = AndRHSMask.countLeadingZeros(); 1172 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros); 1173 1174 if (MaskedValueIsZero(Op0LHS, Mask)) { 1175 Value *NewNeg = Builder->CreateNeg(Op0RHS); 1176 return BinaryOperator::CreateAnd(NewNeg, AndRHS); 1177 } 1178 } 1179 break; 1180 1181 case Instruction::Shl: 1182 case Instruction::LShr: 1183 // (1 << x) & 1 --> zext(x == 0) 1184 // (1 >> x) & 1 --> zext(x == 0) 1185 if (AndRHSMask == 1 && Op0LHS == AndRHS) { 1186 Value *NewICmp = 1187 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType())); 1188 return new ZExtInst(NewICmp, I.getType()); 1189 } 1190 break; 1191 } 1192 1193 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) 1194 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I)) 1195 return Res; 1196 } 1197 1198 // If this is an integer truncation, and if the source is an 'and' with 1199 // immediate, transform it. This frequently occurs for bitfield accesses. 1200 { 1201 Value *X = 0; ConstantInt *YC = 0; 1202 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) { 1203 // Change: and (trunc (and X, YC) to T), C2 1204 // into : and (trunc X to T), trunc(YC) & C2 1205 // This will fold the two constants together, which may allow 1206 // other simplifications. 1207 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk"); 1208 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType()); 1209 C3 = ConstantExpr::getAnd(C3, AndRHS); 1210 return BinaryOperator::CreateAnd(NewCast, C3); 1211 } 1212 } 1213 1214 // Try to fold constant and into select arguments. 1215 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 1216 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1217 return R; 1218 if (isa<PHINode>(Op0)) 1219 if (Instruction *NV = FoldOpIntoPhi(I)) 1220 return NV; 1221 } 1222 1223 1224 // (~A & ~B) == (~(A | B)) - De Morgan's Law 1225 if (Value *Op0NotVal = dyn_castNotVal(Op0)) 1226 if (Value *Op1NotVal = dyn_castNotVal(Op1)) 1227 if (Op0->hasOneUse() && Op1->hasOneUse()) { 1228 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal, 1229 I.getName()+".demorgan"); 1230 return BinaryOperator::CreateNot(Or); 1231 } 1232 1233 { 1234 Value *A = 0, *B = 0, *C = 0, *D = 0; 1235 // (A|B) & ~(A&B) -> A^B 1236 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1237 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) && 1238 ((A == C && B == D) || (A == D && B == C))) 1239 return BinaryOperator::CreateXor(A, B); 1240 1241 // ~(A&B) & (A|B) -> A^B 1242 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1243 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) && 1244 ((A == C && B == D) || (A == D && B == C))) 1245 return BinaryOperator::CreateXor(A, B); 1246 1247 // A&(A^B) => A & ~B 1248 { 1249 Value *tmpOp0 = Op0; 1250 Value *tmpOp1 = Op1; 1251 if (Op0->hasOneUse() && 1252 match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 1253 if (A == Op1 || B == Op1 ) { 1254 tmpOp1 = Op0; 1255 tmpOp0 = Op1; 1256 // Simplify below 1257 } 1258 } 1259 1260 if (tmpOp1->hasOneUse() && 1261 match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) { 1262 if (B == tmpOp0) { 1263 std::swap(A, B); 1264 } 1265 // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if 1266 // A is originally -1 (or a vector of -1 and undefs), then we enter 1267 // an endless loop. By checking that A is non-constant we ensure that 1268 // we will never get to the loop. 1269 if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B 1270 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B)); 1271 } 1272 } 1273 1274 // (A&((~A)|B)) -> A&B 1275 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) || 1276 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1))))) 1277 return BinaryOperator::CreateAnd(A, Op1); 1278 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) || 1279 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0))))) 1280 return BinaryOperator::CreateAnd(A, Op0); 1281 } 1282 1283 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) 1284 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0)) 1285 if (Value *Res = FoldAndOfICmps(LHS, RHS)) 1286 return ReplaceInstUsesWith(I, Res); 1287 1288 // If and'ing two fcmp, try combine them into one. 1289 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 1290 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 1291 if (Value *Res = FoldAndOfFCmps(LHS, RHS)) 1292 return ReplaceInstUsesWith(I, Res); 1293 1294 1295 // fold (and (cast A), (cast B)) -> (cast (and A, B)) 1296 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) 1297 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) { 1298 Type *SrcTy = Op0C->getOperand(0)->getType(); 1299 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ? 1300 SrcTy == Op1C->getOperand(0)->getType() && 1301 SrcTy->isIntOrIntVectorTy()) { 1302 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0); 1303 1304 // Only do this if the casts both really cause code to be generated. 1305 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) && 1306 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) { 1307 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName()); 1308 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); 1309 } 1310 1311 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the 1312 // cast is otherwise not optimizable. This happens for vector sexts. 1313 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp)) 1314 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp)) 1315 if (Value *Res = FoldAndOfICmps(LHS, RHS)) 1316 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 1317 1318 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the 1319 // cast is otherwise not optimizable. This happens for vector sexts. 1320 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp)) 1321 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp)) 1322 if (Value *Res = FoldAndOfFCmps(LHS, RHS)) 1323 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 1324 } 1325 } 1326 1327 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts. 1328 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) { 1329 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0)) 1330 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() && 1331 SI0->getOperand(1) == SI1->getOperand(1) && 1332 (SI0->hasOneUse() || SI1->hasOneUse())) { 1333 Value *NewOp = 1334 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0), 1335 SI0->getName()); 1336 return BinaryOperator::Create(SI1->getOpcode(), NewOp, 1337 SI1->getOperand(1)); 1338 } 1339 } 1340 1341 { 1342 Value *X = 0; 1343 bool OpsSwapped = false; 1344 // Canonicalize SExt or Not to the LHS 1345 if (match(Op1, m_SExt(m_Value())) || 1346 match(Op1, m_Not(m_Value()))) { 1347 std::swap(Op0, Op1); 1348 OpsSwapped = true; 1349 } 1350 1351 // Fold (and (sext bool to A), B) --> (select bool, B, 0) 1352 if (match(Op0, m_SExt(m_Value(X))) && 1353 X->getType()->getScalarType()->isIntegerTy(1)) { 1354 Value *Zero = Constant::getNullValue(Op1->getType()); 1355 return SelectInst::Create(X, Op1, Zero); 1356 } 1357 1358 // Fold (and ~(sext bool to A), B) --> (select bool, 0, B) 1359 if (match(Op0, m_Not(m_SExt(m_Value(X)))) && 1360 X->getType()->getScalarType()->isIntegerTy(1)) { 1361 Value *Zero = Constant::getNullValue(Op0->getType()); 1362 return SelectInst::Create(X, Zero, Op1); 1363 } 1364 1365 if (OpsSwapped) 1366 std::swap(Op0, Op1); 1367 } 1368 1369 return Changed ? &I : 0; 1370} 1371 1372/// CollectBSwapParts - Analyze the specified subexpression and see if it is 1373/// capable of providing pieces of a bswap. The subexpression provides pieces 1374/// of a bswap if it is proven that each of the non-zero bytes in the output of 1375/// the expression came from the corresponding "byte swapped" byte in some other 1376/// value. For example, if the current subexpression is "(shl i32 %X, 24)" then 1377/// we know that the expression deposits the low byte of %X into the high byte 1378/// of the bswap result and that all other bytes are zero. This expression is 1379/// accepted, the high byte of ByteValues is set to X to indicate a correct 1380/// match. 1381/// 1382/// This function returns true if the match was unsuccessful and false if so. 1383/// On entry to the function the "OverallLeftShift" is a signed integer value 1384/// indicating the number of bytes that the subexpression is later shifted. For 1385/// example, if the expression is later right shifted by 16 bits, the 1386/// OverallLeftShift value would be -2 on entry. This is used to specify which 1387/// byte of ByteValues is actually being set. 1388/// 1389/// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding 1390/// byte is masked to zero by a user. For example, in (X & 255), X will be 1391/// processed with a bytemask of 1. Because bytemask is 32-bits, this limits 1392/// this function to working on up to 32-byte (256 bit) values. ByteMask is 1393/// always in the local (OverallLeftShift) coordinate space. 1394/// 1395static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask, 1396 SmallVectorImpl<Value *> &ByteValues) { 1397 if (Instruction *I = dyn_cast<Instruction>(V)) { 1398 // If this is an or instruction, it may be an inner node of the bswap. 1399 if (I->getOpcode() == Instruction::Or) { 1400 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask, 1401 ByteValues) || 1402 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask, 1403 ByteValues); 1404 } 1405 1406 // If this is a logical shift by a constant multiple of 8, recurse with 1407 // OverallLeftShift and ByteMask adjusted. 1408 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) { 1409 unsigned ShAmt = 1410 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U); 1411 // Ensure the shift amount is defined and of a byte value. 1412 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size())) 1413 return true; 1414 1415 unsigned ByteShift = ShAmt >> 3; 1416 if (I->getOpcode() == Instruction::Shl) { 1417 // X << 2 -> collect(X, +2) 1418 OverallLeftShift += ByteShift; 1419 ByteMask >>= ByteShift; 1420 } else { 1421 // X >>u 2 -> collect(X, -2) 1422 OverallLeftShift -= ByteShift; 1423 ByteMask <<= ByteShift; 1424 ByteMask &= (~0U >> (32-ByteValues.size())); 1425 } 1426 1427 if (OverallLeftShift >= (int)ByteValues.size()) return true; 1428 if (OverallLeftShift <= -(int)ByteValues.size()) return true; 1429 1430 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask, 1431 ByteValues); 1432 } 1433 1434 // If this is a logical 'and' with a mask that clears bytes, clear the 1435 // corresponding bytes in ByteMask. 1436 if (I->getOpcode() == Instruction::And && 1437 isa<ConstantInt>(I->getOperand(1))) { 1438 // Scan every byte of the and mask, seeing if the byte is either 0 or 255. 1439 unsigned NumBytes = ByteValues.size(); 1440 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255); 1441 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue(); 1442 1443 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) { 1444 // If this byte is masked out by a later operation, we don't care what 1445 // the and mask is. 1446 if ((ByteMask & (1 << i)) == 0) 1447 continue; 1448 1449 // If the AndMask is all zeros for this byte, clear the bit. 1450 APInt MaskB = AndMask & Byte; 1451 if (MaskB == 0) { 1452 ByteMask &= ~(1U << i); 1453 continue; 1454 } 1455 1456 // If the AndMask is not all ones for this byte, it's not a bytezap. 1457 if (MaskB != Byte) 1458 return true; 1459 1460 // Otherwise, this byte is kept. 1461 } 1462 1463 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask, 1464 ByteValues); 1465 } 1466 } 1467 1468 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be 1469 // the input value to the bswap. Some observations: 1) if more than one byte 1470 // is demanded from this input, then it could not be successfully assembled 1471 // into a byteswap. At least one of the two bytes would not be aligned with 1472 // their ultimate destination. 1473 if (!isPowerOf2_32(ByteMask)) return true; 1474 unsigned InputByteNo = countTrailingZeros(ByteMask); 1475 1476 // 2) The input and ultimate destinations must line up: if byte 3 of an i32 1477 // is demanded, it needs to go into byte 0 of the result. This means that the 1478 // byte needs to be shifted until it lands in the right byte bucket. The 1479 // shift amount depends on the position: if the byte is coming from the high 1480 // part of the value (e.g. byte 3) then it must be shifted right. If from the 1481 // low part, it must be shifted left. 1482 unsigned DestByteNo = InputByteNo + OverallLeftShift; 1483 if (ByteValues.size()-1-DestByteNo != InputByteNo) 1484 return true; 1485 1486 // If the destination byte value is already defined, the values are or'd 1487 // together, which isn't a bswap (unless it's an or of the same bits). 1488 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V) 1489 return true; 1490 ByteValues[DestByteNo] = V; 1491 return false; 1492} 1493 1494/// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom. 1495/// If so, insert the new bswap intrinsic and return it. 1496Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) { 1497 IntegerType *ITy = dyn_cast<IntegerType>(I.getType()); 1498 if (!ITy || ITy->getBitWidth() % 16 || 1499 // ByteMask only allows up to 32-byte values. 1500 ITy->getBitWidth() > 32*8) 1501 return 0; // Can only bswap pairs of bytes. Can't do vectors. 1502 1503 /// ByteValues - For each byte of the result, we keep track of which value 1504 /// defines each byte. 1505 SmallVector<Value*, 8> ByteValues; 1506 ByteValues.resize(ITy->getBitWidth()/8); 1507 1508 // Try to find all the pieces corresponding to the bswap. 1509 uint32_t ByteMask = ~0U >> (32-ByteValues.size()); 1510 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues)) 1511 return 0; 1512 1513 // Check to see if all of the bytes come from the same value. 1514 Value *V = ByteValues[0]; 1515 if (V == 0) return 0; // Didn't find a byte? Must be zero. 1516 1517 // Check to make sure that all of the bytes come from the same value. 1518 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i) 1519 if (ByteValues[i] != V) 1520 return 0; 1521 Module *M = I.getParent()->getParent()->getParent(); 1522 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy); 1523 return CallInst::Create(F, V); 1524} 1525 1526/// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check 1527/// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then 1528/// we can simplify this expression to "cond ? C : D or B". 1529static Instruction *MatchSelectFromAndOr(Value *A, Value *B, 1530 Value *C, Value *D) { 1531 // If A is not a select of -1/0, this cannot match. 1532 Value *Cond = 0; 1533 if (!match(A, m_SExt(m_Value(Cond))) || 1534 !Cond->getType()->isIntegerTy(1)) 1535 return 0; 1536 1537 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B. 1538 if (match(D, m_Not(m_SExt(m_Specific(Cond))))) 1539 return SelectInst::Create(Cond, C, B); 1540 if (match(D, m_SExt(m_Not(m_Specific(Cond))))) 1541 return SelectInst::Create(Cond, C, B); 1542 1543 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D. 1544 if (match(B, m_Not(m_SExt(m_Specific(Cond))))) 1545 return SelectInst::Create(Cond, C, D); 1546 if (match(B, m_SExt(m_Not(m_Specific(Cond))))) 1547 return SelectInst::Create(Cond, C, D); 1548 return 0; 1549} 1550 1551/// FoldOrOfICmps - Fold (icmp)|(icmp) if possible. 1552Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) { 1553 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate(); 1554 1555 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2) 1556 // if K1 and K2 are a one-bit mask. 1557 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1)); 1558 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1)); 1559 1560 if (LHS->getPredicate() == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero() && 1561 RHS->getPredicate() == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) { 1562 1563 BinaryOperator *LAnd = dyn_cast<BinaryOperator>(LHS->getOperand(0)); 1564 BinaryOperator *RAnd = dyn_cast<BinaryOperator>(RHS->getOperand(0)); 1565 if (LAnd && RAnd && LAnd->hasOneUse() && RHS->hasOneUse() && 1566 LAnd->getOpcode() == Instruction::And && 1567 RAnd->getOpcode() == Instruction::And) { 1568 1569 Value *Mask = 0; 1570 Value *Masked = 0; 1571 if (LAnd->getOperand(0) == RAnd->getOperand(0) && 1572 isKnownToBeAPowerOfTwo(LAnd->getOperand(1)) && 1573 isKnownToBeAPowerOfTwo(RAnd->getOperand(1))) { 1574 Mask = Builder->CreateOr(LAnd->getOperand(1), RAnd->getOperand(1)); 1575 Masked = Builder->CreateAnd(LAnd->getOperand(0), Mask); 1576 } else if (LAnd->getOperand(1) == RAnd->getOperand(1) && 1577 isKnownToBeAPowerOfTwo(LAnd->getOperand(0)) && 1578 isKnownToBeAPowerOfTwo(RAnd->getOperand(0))) { 1579 Mask = Builder->CreateOr(LAnd->getOperand(0), RAnd->getOperand(0)); 1580 Masked = Builder->CreateAnd(LAnd->getOperand(1), Mask); 1581 } 1582 1583 if (Masked) 1584 return Builder->CreateICmp(ICmpInst::ICMP_NE, Masked, Mask); 1585 } 1586 } 1587 1588 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B) 1589 if (PredicatesFoldable(LHSCC, RHSCC)) { 1590 if (LHS->getOperand(0) == RHS->getOperand(1) && 1591 LHS->getOperand(1) == RHS->getOperand(0)) 1592 LHS->swapOperands(); 1593 if (LHS->getOperand(0) == RHS->getOperand(0) && 1594 LHS->getOperand(1) == RHS->getOperand(1)) { 1595 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 1596 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS); 1597 bool isSigned = LHS->isSigned() || RHS->isSigned(); 1598 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder); 1599 } 1600 } 1601 1602 // handle (roughly): 1603 // (icmp ne (A & B), C) | (icmp ne (A & D), E) 1604 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder)) 1605 return V; 1606 1607 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0); 1608 if (LHS->hasOneUse() || RHS->hasOneUse()) { 1609 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1) 1610 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1) 1611 Value *A = 0, *B = 0; 1612 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) { 1613 B = Val; 1614 if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1)) 1615 A = Val2; 1616 else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2) 1617 A = RHS->getOperand(1); 1618 } 1619 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1) 1620 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1) 1621 else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) { 1622 B = Val2; 1623 if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1)) 1624 A = Val; 1625 else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val) 1626 A = LHS->getOperand(1); 1627 } 1628 if (A && B) 1629 return Builder->CreateICmp( 1630 ICmpInst::ICMP_UGE, 1631 Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A); 1632 } 1633 1634 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2). 1635 if (LHSCst == 0 || RHSCst == 0) return 0; 1636 1637 if (LHSCst == RHSCst && LHSCC == RHSCC) { 1638 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0) 1639 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) { 1640 Value *NewOr = Builder->CreateOr(Val, Val2); 1641 return Builder->CreateICmp(LHSCC, NewOr, LHSCst); 1642 } 1643 } 1644 1645 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1) 1646 // iff C2 + CA == C1. 1647 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) { 1648 ConstantInt *AddCst; 1649 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst)))) 1650 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue()) 1651 return Builder->CreateICmpULE(Val, LHSCst); 1652 } 1653 1654 // From here on, we only handle: 1655 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler. 1656 if (Val != Val2) return 0; 1657 1658 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere. 1659 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE || 1660 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE || 1661 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE || 1662 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE) 1663 return 0; 1664 1665 // We can't fold (ugt x, C) | (sgt x, C2). 1666 if (!PredicatesFoldable(LHSCC, RHSCC)) 1667 return 0; 1668 1669 // Ensure that the larger constant is on the RHS. 1670 bool ShouldSwap; 1671 if (CmpInst::isSigned(LHSCC) || 1672 (ICmpInst::isEquality(LHSCC) && 1673 CmpInst::isSigned(RHSCC))) 1674 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue()); 1675 else 1676 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue()); 1677 1678 if (ShouldSwap) { 1679 std::swap(LHS, RHS); 1680 std::swap(LHSCst, RHSCst); 1681 std::swap(LHSCC, RHSCC); 1682 } 1683 1684 // At this point, we know we have two icmp instructions 1685 // comparing a value against two constants and or'ing the result 1686 // together. Because of the above check, we know that we only have 1687 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the 1688 // icmp folding check above), that the two constants are not 1689 // equal. 1690 assert(LHSCst != RHSCst && "Compares not folded above?"); 1691 1692 switch (LHSCC) { 1693 default: llvm_unreachable("Unknown integer condition code!"); 1694 case ICmpInst::ICMP_EQ: 1695 switch (RHSCC) { 1696 default: llvm_unreachable("Unknown integer condition code!"); 1697 case ICmpInst::ICMP_EQ: 1698 if (LHS->getOperand(0) == RHS->getOperand(0)) { 1699 // if LHSCst and RHSCst differ only by one bit: 1700 // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1 1701 assert(LHSCst->getValue().ule(LHSCst->getValue())); 1702 1703 APInt Xor = LHSCst->getValue() ^ RHSCst->getValue(); 1704 if (Xor.isPowerOf2()) { 1705 Value *NegCst = Builder->getInt(~Xor); 1706 Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst); 1707 return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst); 1708 } 1709 } 1710 1711 if (LHSCst == SubOne(RHSCst)) { 1712 // (X == 13 | X == 14) -> X-13 <u 2 1713 Constant *AddCST = ConstantExpr::getNeg(LHSCst); 1714 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off"); 1715 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst); 1716 return Builder->CreateICmpULT(Add, AddCST); 1717 } 1718 1719 break; // (X == 13 | X == 15) -> no change 1720 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change 1721 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change 1722 break; 1723 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15 1724 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15 1725 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15 1726 return RHS; 1727 } 1728 break; 1729 case ICmpInst::ICMP_NE: 1730 switch (RHSCC) { 1731 default: llvm_unreachable("Unknown integer condition code!"); 1732 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13 1733 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13 1734 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13 1735 return LHS; 1736 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true 1737 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true 1738 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true 1739 return Builder->getTrue(); 1740 } 1741 case ICmpInst::ICMP_ULT: 1742 switch (RHSCC) { 1743 default: llvm_unreachable("Unknown integer condition code!"); 1744 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change 1745 break; 1746 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2 1747 // If RHSCst is [us]MAXINT, it is always false. Not handling 1748 // this can cause overflow. 1749 if (RHSCst->isMaxValue(false)) 1750 return LHS; 1751 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false); 1752 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change 1753 break; 1754 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15 1755 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15 1756 return RHS; 1757 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change 1758 break; 1759 } 1760 break; 1761 case ICmpInst::ICMP_SLT: 1762 switch (RHSCC) { 1763 default: llvm_unreachable("Unknown integer condition code!"); 1764 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change 1765 break; 1766 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2 1767 // If RHSCst is [us]MAXINT, it is always false. Not handling 1768 // this can cause overflow. 1769 if (RHSCst->isMaxValue(true)) 1770 return LHS; 1771 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false); 1772 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change 1773 break; 1774 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15 1775 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15 1776 return RHS; 1777 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change 1778 break; 1779 } 1780 break; 1781 case ICmpInst::ICMP_UGT: 1782 switch (RHSCC) { 1783 default: llvm_unreachable("Unknown integer condition code!"); 1784 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13 1785 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13 1786 return LHS; 1787 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change 1788 break; 1789 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true 1790 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true 1791 return Builder->getTrue(); 1792 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change 1793 break; 1794 } 1795 break; 1796 case ICmpInst::ICMP_SGT: 1797 switch (RHSCC) { 1798 default: llvm_unreachable("Unknown integer condition code!"); 1799 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13 1800 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13 1801 return LHS; 1802 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change 1803 break; 1804 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true 1805 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true 1806 return Builder->getTrue(); 1807 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change 1808 break; 1809 } 1810 break; 1811 } 1812 return 0; 1813} 1814 1815/// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of 1816/// instcombine, this returns a Value which should already be inserted into the 1817/// function. 1818Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) { 1819 if (LHS->getPredicate() == FCmpInst::FCMP_UNO && 1820 RHS->getPredicate() == FCmpInst::FCMP_UNO && 1821 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) { 1822 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1))) 1823 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) { 1824 // If either of the constants are nans, then the whole thing returns 1825 // true. 1826 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN()) 1827 return Builder->getTrue(); 1828 1829 // Otherwise, no need to compare the two constants, compare the 1830 // rest. 1831 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0)); 1832 } 1833 1834 // Handle vector zeros. This occurs because the canonical form of 1835 // "fcmp uno x,x" is "fcmp uno x, 0". 1836 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) && 1837 isa<ConstantAggregateZero>(RHS->getOperand(1))) 1838 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0)); 1839 1840 return 0; 1841 } 1842 1843 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1); 1844 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1); 1845 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate(); 1846 1847 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) { 1848 // Swap RHS operands to match LHS. 1849 Op1CC = FCmpInst::getSwappedPredicate(Op1CC); 1850 std::swap(Op1LHS, Op1RHS); 1851 } 1852 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) { 1853 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y). 1854 if (Op0CC == Op1CC) 1855 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS); 1856 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE) 1857 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1); 1858 if (Op0CC == FCmpInst::FCMP_FALSE) 1859 return RHS; 1860 if (Op1CC == FCmpInst::FCMP_FALSE) 1861 return LHS; 1862 bool Op0Ordered; 1863 bool Op1Ordered; 1864 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered); 1865 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered); 1866 if (Op0Ordered == Op1Ordered) { 1867 // If both are ordered or unordered, return a new fcmp with 1868 // or'ed predicates. 1869 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder); 1870 } 1871 } 1872 return 0; 1873} 1874 1875/// FoldOrWithConstants - This helper function folds: 1876/// 1877/// ((A | B) & C1) | (B & C2) 1878/// 1879/// into: 1880/// 1881/// (A & C1) | B 1882/// 1883/// when the XOR of the two constants is "all ones" (-1). 1884Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op, 1885 Value *A, Value *B, Value *C) { 1886 ConstantInt *CI1 = dyn_cast<ConstantInt>(C); 1887 if (!CI1) return 0; 1888 1889 Value *V1 = 0; 1890 ConstantInt *CI2 = 0; 1891 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0; 1892 1893 APInt Xor = CI1->getValue() ^ CI2->getValue(); 1894 if (!Xor.isAllOnesValue()) return 0; 1895 1896 if (V1 == A || V1 == B) { 1897 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1); 1898 return BinaryOperator::CreateOr(NewOp, V1); 1899 } 1900 1901 return 0; 1902} 1903 1904Instruction *InstCombiner::visitOr(BinaryOperator &I) { 1905 bool Changed = SimplifyAssociativeOrCommutative(I); 1906 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1907 1908 if (Value *V = SimplifyOrInst(Op0, Op1, DL)) 1909 return ReplaceInstUsesWith(I, V); 1910 1911 // (A&B)|(A&C) -> A&(B|C) etc 1912 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1913 return ReplaceInstUsesWith(I, V); 1914 1915 // See if we can simplify any instructions used by the instruction whose sole 1916 // purpose is to compute bits we don't care about. 1917 if (SimplifyDemandedInstructionBits(I)) 1918 return &I; 1919 1920 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) { 1921 ConstantInt *C1 = 0; Value *X = 0; 1922 // (X & C1) | C2 --> (X | C2) & (C1|C2) 1923 // iff (C1 & C2) == 0. 1924 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && 1925 (RHS->getValue() & C1->getValue()) != 0 && 1926 Op0->hasOneUse()) { 1927 Value *Or = Builder->CreateOr(X, RHS); 1928 Or->takeName(Op0); 1929 return BinaryOperator::CreateAnd(Or, 1930 Builder->getInt(RHS->getValue() | C1->getValue())); 1931 } 1932 1933 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2) 1934 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && 1935 Op0->hasOneUse()) { 1936 Value *Or = Builder->CreateOr(X, RHS); 1937 Or->takeName(Op0); 1938 return BinaryOperator::CreateXor(Or, 1939 Builder->getInt(C1->getValue() & ~RHS->getValue())); 1940 } 1941 1942 // Try to fold constant and into select arguments. 1943 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 1944 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1945 return R; 1946 1947 if (isa<PHINode>(Op0)) 1948 if (Instruction *NV = FoldOpIntoPhi(I)) 1949 return NV; 1950 } 1951 1952 Value *A = 0, *B = 0; 1953 ConstantInt *C1 = 0, *C2 = 0; 1954 1955 // (A | B) | C and A | (B | C) -> bswap if possible. 1956 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible. 1957 if (match(Op0, m_Or(m_Value(), m_Value())) || 1958 match(Op1, m_Or(m_Value(), m_Value())) || 1959 (match(Op0, m_LogicalShift(m_Value(), m_Value())) && 1960 match(Op1, m_LogicalShift(m_Value(), m_Value())))) { 1961 if (Instruction *BSwap = MatchBSwap(I)) 1962 return BSwap; 1963 } 1964 1965 // (X^C)|Y -> (X|Y)^C iff Y&C == 0 1966 if (Op0->hasOneUse() && 1967 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) && 1968 MaskedValueIsZero(Op1, C1->getValue())) { 1969 Value *NOr = Builder->CreateOr(A, Op1); 1970 NOr->takeName(Op0); 1971 return BinaryOperator::CreateXor(NOr, C1); 1972 } 1973 1974 // Y|(X^C) -> (X|Y)^C iff Y&C == 0 1975 if (Op1->hasOneUse() && 1976 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) && 1977 MaskedValueIsZero(Op0, C1->getValue())) { 1978 Value *NOr = Builder->CreateOr(A, Op0); 1979 NOr->takeName(Op0); 1980 return BinaryOperator::CreateXor(NOr, C1); 1981 } 1982 1983 // (A & C)|(B & D) 1984 Value *C = 0, *D = 0; 1985 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 1986 match(Op1, m_And(m_Value(B), m_Value(D)))) { 1987 Value *V1 = 0, *V2 = 0; 1988 C1 = dyn_cast<ConstantInt>(C); 1989 C2 = dyn_cast<ConstantInt>(D); 1990 if (C1 && C2) { // (A & C1)|(B & C2) 1991 // If we have: ((V + N) & C1) | (V & C2) 1992 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0 1993 // replace with V+N. 1994 if (C1->getValue() == ~C2->getValue()) { 1995 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+ 1996 match(A, m_Add(m_Value(V1), m_Value(V2)))) { 1997 // Add commutes, try both ways. 1998 if (V1 == B && MaskedValueIsZero(V2, C2->getValue())) 1999 return ReplaceInstUsesWith(I, A); 2000 if (V2 == B && MaskedValueIsZero(V1, C2->getValue())) 2001 return ReplaceInstUsesWith(I, A); 2002 } 2003 // Or commutes, try both ways. 2004 if ((C1->getValue() & (C1->getValue()+1)) == 0 && 2005 match(B, m_Add(m_Value(V1), m_Value(V2)))) { 2006 // Add commutes, try both ways. 2007 if (V1 == A && MaskedValueIsZero(V2, C1->getValue())) 2008 return ReplaceInstUsesWith(I, B); 2009 if (V2 == A && MaskedValueIsZero(V1, C1->getValue())) 2010 return ReplaceInstUsesWith(I, B); 2011 } 2012 } 2013 2014 if ((C1->getValue() & C2->getValue()) == 0) { 2015 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2) 2016 // iff (C1&C2) == 0 and (N&~C1) == 0 2017 if (match(A, m_Or(m_Value(V1), m_Value(V2))) && 2018 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N) 2019 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V) 2020 return BinaryOperator::CreateAnd(A, 2021 Builder->getInt(C1->getValue()|C2->getValue())); 2022 // Or commutes, try both ways. 2023 if (match(B, m_Or(m_Value(V1), m_Value(V2))) && 2024 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N) 2025 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V) 2026 return BinaryOperator::CreateAnd(B, 2027 Builder->getInt(C1->getValue()|C2->getValue())); 2028 2029 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2) 2030 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0. 2031 ConstantInt *C3 = 0, *C4 = 0; 2032 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) && 2033 (C3->getValue() & ~C1->getValue()) == 0 && 2034 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) && 2035 (C4->getValue() & ~C2->getValue()) == 0) { 2036 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield"); 2037 return BinaryOperator::CreateAnd(V2, 2038 Builder->getInt(C1->getValue()|C2->getValue())); 2039 } 2040 } 2041 } 2042 2043 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants. 2044 // Don't do this for vector select idioms, the code generator doesn't handle 2045 // them well yet. 2046 if (!I.getType()->isVectorTy()) { 2047 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D)) 2048 return Match; 2049 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C)) 2050 return Match; 2051 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D)) 2052 return Match; 2053 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C)) 2054 return Match; 2055 } 2056 2057 // ((A&~B)|(~A&B)) -> A^B 2058 if ((match(C, m_Not(m_Specific(D))) && 2059 match(B, m_Not(m_Specific(A))))) 2060 return BinaryOperator::CreateXor(A, D); 2061 // ((~B&A)|(~A&B)) -> A^B 2062 if ((match(A, m_Not(m_Specific(D))) && 2063 match(B, m_Not(m_Specific(C))))) 2064 return BinaryOperator::CreateXor(C, D); 2065 // ((A&~B)|(B&~A)) -> A^B 2066 if ((match(C, m_Not(m_Specific(B))) && 2067 match(D, m_Not(m_Specific(A))))) 2068 return BinaryOperator::CreateXor(A, B); 2069 // ((~B&A)|(B&~A)) -> A^B 2070 if ((match(A, m_Not(m_Specific(B))) && 2071 match(D, m_Not(m_Specific(C))))) 2072 return BinaryOperator::CreateXor(C, B); 2073 2074 // ((A|B)&1)|(B&-2) -> (A&1) | B 2075 if (match(A, m_Or(m_Value(V1), m_Specific(B))) || 2076 match(A, m_Or(m_Specific(B), m_Value(V1)))) { 2077 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C); 2078 if (Ret) return Ret; 2079 } 2080 // (B&-2)|((A|B)&1) -> (A&1) | B 2081 if (match(B, m_Or(m_Specific(A), m_Value(V1))) || 2082 match(B, m_Or(m_Value(V1), m_Specific(A)))) { 2083 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D); 2084 if (Ret) return Ret; 2085 } 2086 } 2087 2088 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts. 2089 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) { 2090 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0)) 2091 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() && 2092 SI0->getOperand(1) == SI1->getOperand(1) && 2093 (SI0->hasOneUse() || SI1->hasOneUse())) { 2094 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0), 2095 SI0->getName()); 2096 return BinaryOperator::Create(SI1->getOpcode(), NewOp, 2097 SI1->getOperand(1)); 2098 } 2099 } 2100 2101 // (~A | ~B) == (~(A & B)) - De Morgan's Law 2102 if (Value *Op0NotVal = dyn_castNotVal(Op0)) 2103 if (Value *Op1NotVal = dyn_castNotVal(Op1)) 2104 if (Op0->hasOneUse() && Op1->hasOneUse()) { 2105 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal, 2106 I.getName()+".demorgan"); 2107 return BinaryOperator::CreateNot(And); 2108 } 2109 2110 // Canonicalize xor to the RHS. 2111 bool SwappedForXor = false; 2112 if (match(Op0, m_Xor(m_Value(), m_Value()))) { 2113 std::swap(Op0, Op1); 2114 SwappedForXor = true; 2115 } 2116 2117 // A | ( A ^ B) -> A | B 2118 // A | (~A ^ B) -> A | ~B 2119 // (A & B) | (A ^ B) 2120 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) { 2121 if (Op0 == A || Op0 == B) 2122 return BinaryOperator::CreateOr(A, B); 2123 2124 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) || 2125 match(Op0, m_And(m_Specific(B), m_Specific(A)))) 2126 return BinaryOperator::CreateOr(A, B); 2127 2128 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) { 2129 Value *Not = Builder->CreateNot(B, B->getName()+".not"); 2130 return BinaryOperator::CreateOr(Not, Op0); 2131 } 2132 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) { 2133 Value *Not = Builder->CreateNot(A, A->getName()+".not"); 2134 return BinaryOperator::CreateOr(Not, Op0); 2135 } 2136 } 2137 2138 // A | ~(A | B) -> A | ~B 2139 // A | ~(A ^ B) -> A | ~B 2140 if (match(Op1, m_Not(m_Value(A)))) 2141 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A)) 2142 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) && 2143 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or || 2144 B->getOpcode() == Instruction::Xor)) { 2145 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) : 2146 B->getOperand(0); 2147 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not"); 2148 return BinaryOperator::CreateOr(Not, Op0); 2149 } 2150 2151 if (SwappedForXor) 2152 std::swap(Op0, Op1); 2153 2154 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) 2155 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0))) 2156 if (Value *Res = FoldOrOfICmps(LHS, RHS)) 2157 return ReplaceInstUsesWith(I, Res); 2158 2159 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y) 2160 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 2161 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 2162 if (Value *Res = FoldOrOfFCmps(LHS, RHS)) 2163 return ReplaceInstUsesWith(I, Res); 2164 2165 // fold (or (cast A), (cast B)) -> (cast (or A, B)) 2166 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { 2167 CastInst *Op1C = dyn_cast<CastInst>(Op1); 2168 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ? 2169 Type *SrcTy = Op0C->getOperand(0)->getType(); 2170 if (SrcTy == Op1C->getOperand(0)->getType() && 2171 SrcTy->isIntOrIntVectorTy()) { 2172 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0); 2173 2174 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) && 2175 // Only do this if the casts both really cause code to be 2176 // generated. 2177 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) && 2178 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) { 2179 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName()); 2180 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); 2181 } 2182 2183 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the 2184 // cast is otherwise not optimizable. This happens for vector sexts. 2185 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp)) 2186 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp)) 2187 if (Value *Res = FoldOrOfICmps(LHS, RHS)) 2188 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 2189 2190 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the 2191 // cast is otherwise not optimizable. This happens for vector sexts. 2192 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp)) 2193 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp)) 2194 if (Value *Res = FoldOrOfFCmps(LHS, RHS)) 2195 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 2196 } 2197 } 2198 } 2199 2200 // or(sext(A), B) -> A ? -1 : B where A is an i1 2201 // or(A, sext(B)) -> B ? -1 : A where B is an i1 2202 if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1)) 2203 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1); 2204 if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1)) 2205 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0); 2206 2207 // Note: If we've gotten to the point of visiting the outer OR, then the 2208 // inner one couldn't be simplified. If it was a constant, then it won't 2209 // be simplified by a later pass either, so we try swapping the inner/outer 2210 // ORs in the hopes that we'll be able to simplify it this way. 2211 // (X|C) | V --> (X|V) | C 2212 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) && 2213 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) { 2214 Value *Inner = Builder->CreateOr(A, Op1); 2215 Inner->takeName(Op0); 2216 return BinaryOperator::CreateOr(Inner, C1); 2217 } 2218 2219 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D)) 2220 // Since this OR statement hasn't been optimized further yet, we hope 2221 // that this transformation will allow the new ORs to be optimized. 2222 { 2223 Value *X = 0, *Y = 0; 2224 if (Op0->hasOneUse() && Op1->hasOneUse() && 2225 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) && 2226 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) { 2227 Value *orTrue = Builder->CreateOr(A, C); 2228 Value *orFalse = Builder->CreateOr(B, D); 2229 return SelectInst::Create(X, orTrue, orFalse); 2230 } 2231 } 2232 2233 return Changed ? &I : 0; 2234} 2235 2236Instruction *InstCombiner::visitXor(BinaryOperator &I) { 2237 bool Changed = SimplifyAssociativeOrCommutative(I); 2238 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2239 2240 if (Value *V = SimplifyXorInst(Op0, Op1, DL)) 2241 return ReplaceInstUsesWith(I, V); 2242 2243 // (A&B)^(A&C) -> A&(B^C) etc 2244 if (Value *V = SimplifyUsingDistributiveLaws(I)) 2245 return ReplaceInstUsesWith(I, V); 2246 2247 // See if we can simplify any instructions used by the instruction whose sole 2248 // purpose is to compute bits we don't care about. 2249 if (SimplifyDemandedInstructionBits(I)) 2250 return &I; 2251 2252 // Is this a ~ operation? 2253 if (Value *NotOp = dyn_castNotVal(&I)) { 2254 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) { 2255 if (Op0I->getOpcode() == Instruction::And || 2256 Op0I->getOpcode() == Instruction::Or) { 2257 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law 2258 // ~(~X | Y) === (X & ~Y) - De Morgan's Law 2259 if (dyn_castNotVal(Op0I->getOperand(1))) 2260 Op0I->swapOperands(); 2261 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) { 2262 Value *NotY = 2263 Builder->CreateNot(Op0I->getOperand(1), 2264 Op0I->getOperand(1)->getName()+".not"); 2265 if (Op0I->getOpcode() == Instruction::And) 2266 return BinaryOperator::CreateOr(Op0NotVal, NotY); 2267 return BinaryOperator::CreateAnd(Op0NotVal, NotY); 2268 } 2269 2270 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law 2271 // ~(X | Y) === (~X & ~Y) - De Morgan's Law 2272 if (isFreeToInvert(Op0I->getOperand(0)) && 2273 isFreeToInvert(Op0I->getOperand(1))) { 2274 Value *NotX = 2275 Builder->CreateNot(Op0I->getOperand(0), "notlhs"); 2276 Value *NotY = 2277 Builder->CreateNot(Op0I->getOperand(1), "notrhs"); 2278 if (Op0I->getOpcode() == Instruction::And) 2279 return BinaryOperator::CreateOr(NotX, NotY); 2280 return BinaryOperator::CreateAnd(NotX, NotY); 2281 } 2282 2283 } else if (Op0I->getOpcode() == Instruction::AShr) { 2284 // ~(~X >>s Y) --> (X >>s Y) 2285 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) 2286 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1)); 2287 } 2288 } 2289 } 2290 2291 2292 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) { 2293 if (RHS->isOne() && Op0->hasOneUse()) 2294 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B 2295 if (CmpInst *CI = dyn_cast<CmpInst>(Op0)) 2296 return CmpInst::Create(CI->getOpcode(), 2297 CI->getInversePredicate(), 2298 CI->getOperand(0), CI->getOperand(1)); 2299 2300 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp). 2301 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { 2302 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) { 2303 if (CI->hasOneUse() && Op0C->hasOneUse()) { 2304 Instruction::CastOps Opcode = Op0C->getOpcode(); 2305 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) && 2306 (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(), 2307 Op0C->getDestTy()))) { 2308 CI->setPredicate(CI->getInversePredicate()); 2309 return CastInst::Create(Opcode, CI, Op0C->getType()); 2310 } 2311 } 2312 } 2313 } 2314 2315 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) { 2316 // ~(c-X) == X-c-1 == X+(-c-1) 2317 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue()) 2318 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) { 2319 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C); 2320 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C, 2321 ConstantInt::get(I.getType(), 1)); 2322 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS); 2323 } 2324 2325 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) { 2326 if (Op0I->getOpcode() == Instruction::Add) { 2327 // ~(X-c) --> (-c-1)-X 2328 if (RHS->isAllOnesValue()) { 2329 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI); 2330 return BinaryOperator::CreateSub( 2331 ConstantExpr::getSub(NegOp0CI, 2332 ConstantInt::get(I.getType(), 1)), 2333 Op0I->getOperand(0)); 2334 } else if (RHS->getValue().isSignBit()) { 2335 // (X + C) ^ signbit -> (X + C + signbit) 2336 Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue()); 2337 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C); 2338 2339 } 2340 } else if (Op0I->getOpcode() == Instruction::Or) { 2341 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0 2342 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) { 2343 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS); 2344 // Anything in both C1 and C2 is known to be zero, remove it from 2345 // NewRHS. 2346 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS); 2347 NewRHS = ConstantExpr::getAnd(NewRHS, 2348 ConstantExpr::getNot(CommonBits)); 2349 Worklist.Add(Op0I); 2350 I.setOperand(0, Op0I->getOperand(0)); 2351 I.setOperand(1, NewRHS); 2352 return &I; 2353 } 2354 } else if (Op0I->getOpcode() == Instruction::LShr) { 2355 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3) 2356 // E1 = "X ^ C1" 2357 BinaryOperator *E1; 2358 ConstantInt *C1; 2359 if (Op0I->hasOneUse() && 2360 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) && 2361 E1->getOpcode() == Instruction::Xor && 2362 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) { 2363 // fold (C1 >> C2) ^ C3 2364 ConstantInt *C2 = Op0CI, *C3 = RHS; 2365 APInt FoldConst = C1->getValue().lshr(C2->getValue()); 2366 FoldConst ^= C3->getValue(); 2367 // Prepare the two operands. 2368 Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2); 2369 Opnd0->takeName(Op0I); 2370 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc()); 2371 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst); 2372 2373 return BinaryOperator::CreateXor(Opnd0, FoldVal); 2374 } 2375 } 2376 } 2377 } 2378 2379 // Try to fold constant and into select arguments. 2380 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 2381 if (Instruction *R = FoldOpIntoSelect(I, SI)) 2382 return R; 2383 if (isa<PHINode>(Op0)) 2384 if (Instruction *NV = FoldOpIntoPhi(I)) 2385 return NV; 2386 } 2387 2388 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1); 2389 if (Op1I) { 2390 Value *A, *B; 2391 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) { 2392 if (A == Op0) { // B^(B|A) == (A|B)^B 2393 Op1I->swapOperands(); 2394 I.swapOperands(); 2395 std::swap(Op0, Op1); 2396 } else if (B == Op0) { // B^(A|B) == (A|B)^B 2397 I.swapOperands(); // Simplified below. 2398 std::swap(Op0, Op1); 2399 } 2400 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && 2401 Op1I->hasOneUse()){ 2402 if (A == Op0) { // A^(A&B) -> A^(B&A) 2403 Op1I->swapOperands(); 2404 std::swap(A, B); 2405 } 2406 if (B == Op0) { // A^(B&A) -> (B&A)^A 2407 I.swapOperands(); // Simplified below. 2408 std::swap(Op0, Op1); 2409 } 2410 } 2411 } 2412 2413 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0); 2414 if (Op0I) { 2415 Value *A, *B; 2416 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && 2417 Op0I->hasOneUse()) { 2418 if (A == Op1) // (B|A)^B == (A|B)^B 2419 std::swap(A, B); 2420 if (B == Op1) // (A|B)^B == A & ~B 2421 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1)); 2422 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && 2423 Op0I->hasOneUse()){ 2424 if (A == Op1) // (A&B)^A -> (B&A)^A 2425 std::swap(A, B); 2426 if (B == Op1 && // (B&A)^A == ~B & A 2427 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C 2428 return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1); 2429 } 2430 } 2431 } 2432 2433 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts. 2434 if (Op0I && Op1I && Op0I->isShift() && 2435 Op0I->getOpcode() == Op1I->getOpcode() && 2436 Op0I->getOperand(1) == Op1I->getOperand(1) && 2437 (Op0I->hasOneUse() || Op1I->hasOneUse())) { 2438 Value *NewOp = 2439 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0), 2440 Op0I->getName()); 2441 return BinaryOperator::Create(Op1I->getOpcode(), NewOp, 2442 Op1I->getOperand(1)); 2443 } 2444 2445 if (Op0I && Op1I) { 2446 Value *A, *B, *C, *D; 2447 // (A & B)^(A | B) -> A ^ B 2448 if (match(Op0I, m_And(m_Value(A), m_Value(B))) && 2449 match(Op1I, m_Or(m_Value(C), m_Value(D)))) { 2450 if ((A == C && B == D) || (A == D && B == C)) 2451 return BinaryOperator::CreateXor(A, B); 2452 } 2453 // (A | B)^(A & B) -> A ^ B 2454 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && 2455 match(Op1I, m_And(m_Value(C), m_Value(D)))) { 2456 if ((A == C && B == D) || (A == D && B == C)) 2457 return BinaryOperator::CreateXor(A, B); 2458 } 2459 } 2460 2461 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B) 2462 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) 2463 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0))) 2464 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) { 2465 if (LHS->getOperand(0) == RHS->getOperand(1) && 2466 LHS->getOperand(1) == RHS->getOperand(0)) 2467 LHS->swapOperands(); 2468 if (LHS->getOperand(0) == RHS->getOperand(0) && 2469 LHS->getOperand(1) == RHS->getOperand(1)) { 2470 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 2471 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS); 2472 bool isSigned = LHS->isSigned() || RHS->isSigned(); 2473 return ReplaceInstUsesWith(I, 2474 getNewICmpValue(isSigned, Code, Op0, Op1, 2475 Builder)); 2476 } 2477 } 2478 2479 // fold (xor (cast A), (cast B)) -> (cast (xor A, B)) 2480 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { 2481 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) 2482 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind? 2483 Type *SrcTy = Op0C->getOperand(0)->getType(); 2484 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() && 2485 // Only do this if the casts both really cause code to be generated. 2486 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0), 2487 I.getType()) && 2488 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0), 2489 I.getType())) { 2490 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0), 2491 Op1C->getOperand(0), I.getName()); 2492 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); 2493 } 2494 } 2495 } 2496 2497 return Changed ? &I : 0; 2498} 2499