InstCombineAndOrXor.cpp revision 298c45e845e1743f86b060cd280e8729cd4ba468
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/Intrinsics.h" 16#include "llvm/Analysis/InstructionSimplify.h" 17#include "llvm/Support/PatternMatch.h" 18using namespace llvm; 19using namespace PatternMatch; 20 21 22/// AddOne - Add one to a ConstantInt. 23static Constant *AddOne(Constant *C) { 24 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 25} 26/// SubOne - Subtract one from a ConstantInt. 27static Constant *SubOne(ConstantInt *C) { 28 return ConstantInt::get(C->getContext(), C->getValue()-1); 29} 30 31/// isFreeToInvert - Return true if the specified value is free to invert (apply 32/// ~ to). This happens in cases where the ~ can be eliminated. 33static inline bool isFreeToInvert(Value *V) { 34 // ~(~(X)) -> X. 35 if (BinaryOperator::isNot(V)) 36 return true; 37 38 // Constants can be considered to be not'ed values. 39 if (isa<ConstantInt>(V)) 40 return true; 41 42 // Compares can be inverted if they have a single use. 43 if (CmpInst *CI = dyn_cast<CmpInst>(V)) 44 return CI->hasOneUse(); 45 46 return false; 47} 48 49static inline Value *dyn_castNotVal(Value *V) { 50 // If this is not(not(x)) don't return that this is a not: we want the two 51 // not's to be folded first. 52 if (BinaryOperator::isNot(V)) { 53 Value *Operand = BinaryOperator::getNotArgument(V); 54 if (!isFreeToInvert(Operand)) 55 return Operand; 56 } 57 58 // Constants can be considered to be not'ed values... 59 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 60 return ConstantInt::get(C->getType(), ~C->getValue()); 61 return 0; 62} 63 64 65/// getICmpCode - Encode a icmp predicate into a three bit mask. These bits 66/// are carefully arranged to allow folding of expressions such as: 67/// 68/// (A < B) | (A > B) --> (A != B) 69/// 70/// Note that this is only valid if the first and second predicates have the 71/// same sign. Is illegal to do: (A u< B) | (A s> B) 72/// 73/// Three bits are used to represent the condition, as follows: 74/// 0 A > B 75/// 1 A == B 76/// 2 A < B 77/// 78/// <=> Value Definition 79/// 000 0 Always false 80/// 001 1 A > B 81/// 010 2 A == B 82/// 011 3 A >= B 83/// 100 4 A < B 84/// 101 5 A != B 85/// 110 6 A <= B 86/// 111 7 Always true 87/// 88static unsigned getICmpCode(const ICmpInst *ICI) { 89 switch (ICI->getPredicate()) { 90 // False -> 0 91 case ICmpInst::ICMP_UGT: return 1; // 001 92 case ICmpInst::ICMP_SGT: return 1; // 001 93 case ICmpInst::ICMP_EQ: return 2; // 010 94 case ICmpInst::ICMP_UGE: return 3; // 011 95 case ICmpInst::ICMP_SGE: return 3; // 011 96 case ICmpInst::ICMP_ULT: return 4; // 100 97 case ICmpInst::ICMP_SLT: return 4; // 100 98 case ICmpInst::ICMP_NE: return 5; // 101 99 case ICmpInst::ICMP_ULE: return 6; // 110 100 case ICmpInst::ICMP_SLE: return 6; // 110 101 // True -> 7 102 default: 103 llvm_unreachable("Invalid ICmp predicate!"); 104 return 0; 105 } 106} 107 108/// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp 109/// predicate into a three bit mask. It also returns whether it is an ordered 110/// predicate by reference. 111static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) { 112 isOrdered = false; 113 switch (CC) { 114 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000 115 case FCmpInst::FCMP_UNO: return 0; // 000 116 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001 117 case FCmpInst::FCMP_UGT: return 1; // 001 118 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010 119 case FCmpInst::FCMP_UEQ: return 2; // 010 120 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011 121 case FCmpInst::FCMP_UGE: return 3; // 011 122 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100 123 case FCmpInst::FCMP_ULT: return 4; // 100 124 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101 125 case FCmpInst::FCMP_UNE: return 5; // 101 126 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110 127 case FCmpInst::FCMP_ULE: return 6; // 110 128 // True -> 7 129 default: 130 // Not expecting FCMP_FALSE and FCMP_TRUE; 131 llvm_unreachable("Unexpected FCmp predicate!"); 132 return 0; 133 } 134} 135 136/// getICmpValue - This is the complement of getICmpCode, which turns an 137/// opcode and two operands into either a constant true or false, or a brand 138/// new ICmp instruction. The sign is passed in to determine which kind 139/// of predicate to use in the new icmp instruction. 140static Value *getICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS, 141 InstCombiner::BuilderTy *Builder) { 142 CmpInst::Predicate Pred; 143 switch (Code) { 144 default: assert(0 && "Illegal ICmp code!"); 145 case 0: // False. 146 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 147 case 1: Pred = Sign ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break; 148 case 2: Pred = ICmpInst::ICMP_EQ; break; 149 case 3: Pred = Sign ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break; 150 case 4: Pred = Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break; 151 case 5: Pred = ICmpInst::ICMP_NE; break; 152 case 6: Pred = Sign ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break; 153 case 7: // True. 154 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1); 155 } 156 return Builder->CreateICmp(Pred, LHS, RHS); 157} 158 159/// getFCmpValue - This is the complement of getFCmpCode, which turns an 160/// opcode and two operands into either a FCmp instruction. isordered is passed 161/// in to determine which kind of predicate to use in the new fcmp instruction. 162static Value *getFCmpValue(bool isordered, unsigned code, 163 Value *LHS, Value *RHS, 164 InstCombiner::BuilderTy *Builder) { 165 CmpInst::Predicate Pred; 166 switch (code) { 167 default: assert(0 && "Illegal FCmp code!"); 168 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break; 169 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break; 170 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break; 171 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break; 172 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break; 173 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break; 174 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break; 175 case 7: return ConstantInt::getTrue(LHS->getContext()); 176 } 177 return Builder->CreateFCmp(Pred, LHS, RHS); 178} 179 180/// PredicatesFoldable - Return true if both predicates match sign or if at 181/// least one of them is an equality comparison (which is signless). 182static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) { 183 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) || 184 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) || 185 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1)); 186} 187 188// OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where 189// the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is 190// guaranteed to be a binary operator. 191Instruction *InstCombiner::OptAndOp(Instruction *Op, 192 ConstantInt *OpRHS, 193 ConstantInt *AndRHS, 194 BinaryOperator &TheAnd) { 195 Value *X = Op->getOperand(0); 196 Constant *Together = 0; 197 if (!Op->isShift()) 198 Together = ConstantExpr::getAnd(AndRHS, OpRHS); 199 200 switch (Op->getOpcode()) { 201 case Instruction::Xor: 202 if (Op->hasOneUse()) { 203 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2) 204 Value *And = Builder->CreateAnd(X, AndRHS); 205 And->takeName(Op); 206 return BinaryOperator::CreateXor(And, Together); 207 } 208 break; 209 case Instruction::Or: 210 if (Op->hasOneUse() && Together != OpRHS) { 211 // (X | C1) & C2 --> (X | (C1&C2)) & C2 212 Value *Or = Builder->CreateOr(X, Together); 213 Or->takeName(Op); 214 return BinaryOperator::CreateAnd(Or, AndRHS); 215 } 216 break; 217 case Instruction::Add: 218 if (Op->hasOneUse()) { 219 // Adding a one to a single bit bit-field should be turned into an XOR 220 // of the bit. First thing to check is to see if this AND is with a 221 // single bit constant. 222 const APInt &AndRHSV = cast<ConstantInt>(AndRHS)->getValue(); 223 224 // If there is only one bit set. 225 if (AndRHSV.isPowerOf2()) { 226 // Ok, at this point, we know that we are masking the result of the 227 // ADD down to exactly one bit. If the constant we are adding has 228 // no bits set below this bit, then we can eliminate the ADD. 229 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue(); 230 231 // Check to see if any bits below the one bit set in AndRHSV are set. 232 if ((AddRHS & (AndRHSV-1)) == 0) { 233 // If not, the only thing that can effect the output of the AND is 234 // the bit specified by AndRHSV. If that bit is set, the effect of 235 // the XOR is to toggle the bit. If it is clear, then the ADD has 236 // no effect. 237 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop 238 TheAnd.setOperand(0, X); 239 return &TheAnd; 240 } else { 241 // Pull the XOR out of the AND. 242 Value *NewAnd = Builder->CreateAnd(X, AndRHS); 243 NewAnd->takeName(Op); 244 return BinaryOperator::CreateXor(NewAnd, AndRHS); 245 } 246 } 247 } 248 } 249 break; 250 251 case Instruction::Shl: { 252 // We know that the AND will not produce any of the bits shifted in, so if 253 // the anded constant includes them, clear them now! 254 // 255 uint32_t BitWidth = AndRHS->getType()->getBitWidth(); 256 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth); 257 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal)); 258 ConstantInt *CI = ConstantInt::get(AndRHS->getContext(), 259 AndRHS->getValue() & ShlMask); 260 261 if (CI->getValue() == ShlMask) { 262 // Masking out bits that the shift already masks 263 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and. 264 } else if (CI != AndRHS) { // Reducing bits set in and. 265 TheAnd.setOperand(1, CI); 266 return &TheAnd; 267 } 268 break; 269 } 270 case Instruction::LShr: { 271 // We know that the AND will not produce any of the bits shifted in, so if 272 // the anded constant includes them, clear them now! This only applies to 273 // unsigned shifts, because a signed shr may bring in set bits! 274 // 275 uint32_t BitWidth = AndRHS->getType()->getBitWidth(); 276 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth); 277 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal)); 278 ConstantInt *CI = ConstantInt::get(Op->getContext(), 279 AndRHS->getValue() & ShrMask); 280 281 if (CI->getValue() == ShrMask) { 282 // Masking out bits that the shift already masks. 283 return ReplaceInstUsesWith(TheAnd, Op); 284 } else if (CI != AndRHS) { 285 TheAnd.setOperand(1, CI); // Reduce bits set in and cst. 286 return &TheAnd; 287 } 288 break; 289 } 290 case Instruction::AShr: 291 // Signed shr. 292 // See if this is shifting in some sign extension, then masking it out 293 // with an and. 294 if (Op->hasOneUse()) { 295 uint32_t BitWidth = AndRHS->getType()->getBitWidth(); 296 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth); 297 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal)); 298 Constant *C = ConstantInt::get(Op->getContext(), 299 AndRHS->getValue() & ShrMask); 300 if (C == AndRHS) { // Masking out bits shifted in. 301 // (Val ashr C1) & C2 -> (Val lshr C1) & C2 302 // Make the argument unsigned. 303 Value *ShVal = Op->getOperand(0); 304 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName()); 305 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName()); 306 } 307 } 308 break; 309 } 310 return 0; 311} 312 313 314/// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is 315/// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient 316/// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates 317/// whether to treat the V, Lo and HI as signed or not. IB is the location to 318/// insert new instructions. 319Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi, 320 bool isSigned, bool Inside) { 321 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ? 322 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() && 323 "Lo is not <= Hi in range emission code!"); 324 325 if (Inside) { 326 if (Lo == Hi) // Trivially false. 327 return ConstantInt::getFalse(V->getContext()); 328 329 // V >= Min && V < Hi --> V < Hi 330 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) { 331 ICmpInst::Predicate pred = (isSigned ? 332 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT); 333 return Builder->CreateICmp(pred, V, Hi); 334 } 335 336 // Emit V-Lo <u Hi-Lo 337 Constant *NegLo = ConstantExpr::getNeg(Lo); 338 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off"); 339 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi); 340 return Builder->CreateICmpULT(Add, UpperBound); 341 } 342 343 if (Lo == Hi) // Trivially true. 344 return ConstantInt::getTrue(V->getContext()); 345 346 // V < Min || V >= Hi -> V > Hi-1 347 Hi = SubOne(cast<ConstantInt>(Hi)); 348 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) { 349 ICmpInst::Predicate pred = (isSigned ? 350 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT); 351 return Builder->CreateICmp(pred, V, Hi); 352 } 353 354 // Emit V-Lo >u Hi-1-Lo 355 // Note that Hi has already had one subtracted from it, above. 356 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo)); 357 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off"); 358 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi); 359 return Builder->CreateICmpUGT(Add, LowerBound); 360} 361 362// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with 363// any number of 0s on either side. The 1s are allowed to wrap from LSB to 364// MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is 365// not, since all 1s are not contiguous. 366static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) { 367 const APInt& V = Val->getValue(); 368 uint32_t BitWidth = Val->getType()->getBitWidth(); 369 if (!APIntOps::isShiftedMask(BitWidth, V)) return false; 370 371 // look for the first zero bit after the run of ones 372 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros(); 373 // look for the first non-zero bit 374 ME = V.getActiveBits(); 375 return true; 376} 377 378/// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask, 379/// where isSub determines whether the operator is a sub. If we can fold one of 380/// the following xforms: 381/// 382/// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask 383/// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0 384/// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0 385/// 386/// return (A +/- B). 387/// 388Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS, 389 ConstantInt *Mask, bool isSub, 390 Instruction &I) { 391 Instruction *LHSI = dyn_cast<Instruction>(LHS); 392 if (!LHSI || LHSI->getNumOperands() != 2 || 393 !isa<ConstantInt>(LHSI->getOperand(1))) return 0; 394 395 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1)); 396 397 switch (LHSI->getOpcode()) { 398 default: return 0; 399 case Instruction::And: 400 if (ConstantExpr::getAnd(N, Mask) == Mask) { 401 // If the AndRHS is a power of two minus one (0+1+), this is simple. 402 if ((Mask->getValue().countLeadingZeros() + 403 Mask->getValue().countPopulation()) == 404 Mask->getValue().getBitWidth()) 405 break; 406 407 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+ 408 // part, we don't need any explicit masks to take them out of A. If that 409 // is all N is, ignore it. 410 uint32_t MB = 0, ME = 0; 411 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive 412 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth(); 413 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1)); 414 if (MaskedValueIsZero(RHS, Mask)) 415 break; 416 } 417 } 418 return 0; 419 case Instruction::Or: 420 case Instruction::Xor: 421 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0 422 if ((Mask->getValue().countLeadingZeros() + 423 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth() 424 && ConstantExpr::getAnd(N, Mask)->isNullValue()) 425 break; 426 return 0; 427 } 428 429 if (isSub) 430 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold"); 431 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold"); 432} 433 434/// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C) 435/// One of A and B is considered the mask, the other the value. This is 436/// described as the "AMask" or "BMask" part of the enum. If the enum 437/// contains only "Mask", then both A and B can be considered masks. 438/// If A is the mask, then it was proven, that (A & C) == C. This 439/// is trivial if C == A, or C == 0. If both A and C are constants, this 440/// proof is also easy. 441/// For the following explanations we assume that A is the mask. 442/// The part "AllOnes" declares, that the comparison is true only 443/// if (A & B) == A, or all bits of A are set in B. 444/// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes 445/// The part "AllZeroes" declares, that the comparison is true only 446/// if (A & B) == 0, or all bits of A are cleared in B. 447/// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes 448/// The part "Mixed" declares, that (A & B) == C and C might or might not 449/// contain any number of one bits and zero bits. 450/// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed 451/// The Part "Not" means, that in above descriptions "==" should be replaced 452/// by "!=". 453/// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes 454/// If the mask A contains a single bit, then the following is equivalent: 455/// (icmp eq (A & B), A) equals (icmp ne (A & B), 0) 456/// (icmp ne (A & B), A) equals (icmp eq (A & B), 0) 457enum MaskedICmpType { 458 FoldMskICmp_AMask_AllOnes = 1, 459 FoldMskICmp_AMask_NotAllOnes = 2, 460 FoldMskICmp_BMask_AllOnes = 4, 461 FoldMskICmp_BMask_NotAllOnes = 8, 462 FoldMskICmp_Mask_AllZeroes = 16, 463 FoldMskICmp_Mask_NotAllZeroes = 32, 464 FoldMskICmp_AMask_Mixed = 64, 465 FoldMskICmp_AMask_NotMixed = 128, 466 FoldMskICmp_BMask_Mixed = 256, 467 FoldMskICmp_BMask_NotMixed = 512 468}; 469 470/// return the set of pattern classes (from MaskedICmpType) 471/// that (icmp SCC (A & B), C) satisfies 472static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C, 473 ICmpInst::Predicate SCC) 474{ 475 ConstantInt *ACst = dyn_cast<ConstantInt>(A); 476 ConstantInt *BCst = dyn_cast<ConstantInt>(B); 477 ConstantInt *CCst = dyn_cast<ConstantInt>(C); 478 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ); 479 bool icmp_abit = (ACst != 0 && !ACst->isZero() && 480 ACst->getValue().isPowerOf2()); 481 bool icmp_bbit = (BCst != 0 && !BCst->isZero() && 482 BCst->getValue().isPowerOf2()); 483 unsigned result = 0; 484 if (CCst != 0 && CCst->isZero()) { 485 // if C is zero, then both A and B qualify as mask 486 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes | 487 FoldMskICmp_Mask_AllZeroes | 488 FoldMskICmp_AMask_Mixed | 489 FoldMskICmp_BMask_Mixed) 490 : (FoldMskICmp_Mask_NotAllZeroes | 491 FoldMskICmp_Mask_NotAllZeroes | 492 FoldMskICmp_AMask_NotMixed | 493 FoldMskICmp_BMask_NotMixed)); 494 if (icmp_abit) 495 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes | 496 FoldMskICmp_AMask_NotMixed) 497 : (FoldMskICmp_AMask_AllOnes | 498 FoldMskICmp_AMask_Mixed)); 499 if (icmp_bbit) 500 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes | 501 FoldMskICmp_BMask_NotMixed) 502 : (FoldMskICmp_BMask_AllOnes | 503 FoldMskICmp_BMask_Mixed)); 504 return result; 505 } 506 if (A == C) { 507 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes | 508 FoldMskICmp_AMask_Mixed) 509 : (FoldMskICmp_AMask_NotAllOnes | 510 FoldMskICmp_AMask_NotMixed)); 511 if (icmp_abit) 512 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes | 513 FoldMskICmp_AMask_NotMixed) 514 : (FoldMskICmp_Mask_AllZeroes | 515 FoldMskICmp_AMask_Mixed)); 516 } 517 else if (ACst != 0 && CCst != 0 && 518 ConstantExpr::getAnd(ACst, CCst) == CCst) { 519 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed 520 : FoldMskICmp_AMask_NotMixed); 521 } 522 if (B == C) 523 { 524 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes | 525 FoldMskICmp_BMask_Mixed) 526 : (FoldMskICmp_BMask_NotAllOnes | 527 FoldMskICmp_BMask_NotMixed)); 528 if (icmp_bbit) 529 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes | 530 FoldMskICmp_BMask_NotMixed) 531 : (FoldMskICmp_Mask_AllZeroes | 532 FoldMskICmp_BMask_Mixed)); 533 } 534 else if (BCst != 0 && CCst != 0 && 535 ConstantExpr::getAnd(BCst, CCst) == CCst) { 536 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed 537 : FoldMskICmp_BMask_NotMixed); 538 } 539 return result; 540} 541 542/// foldLogOpOfMaskedICmpsHelper: 543/// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) 544/// return the set of pattern classes (from MaskedICmpType) 545/// that both LHS and RHS satisfy 546static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A, 547 Value*& B, Value*& C, 548 Value*& D, Value*& E, 549 ICmpInst *LHS, ICmpInst *RHS) { 550 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate(); 551 if (LHSCC != ICmpInst::ICMP_EQ && LHSCC != ICmpInst::ICMP_NE) return 0; 552 if (RHSCC != ICmpInst::ICMP_EQ && RHSCC != ICmpInst::ICMP_NE) return 0; 553 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0; 554 // vectors are not (yet?) supported 555 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0; 556 557 // Here comes the tricky part: 558 // LHS might be of the form L11 & L12 == X, X == L21 & L22, 559 // and L11 & L12 == L21 & L22. The same goes for RHS. 560 // Now we must find those components L** and R**, that are equal, so 561 // that we can extract the parameters A, B, C, D, and E for the canonical 562 // above. 563 Value *L1 = LHS->getOperand(0); 564 Value *L2 = LHS->getOperand(1); 565 Value *L11,*L12,*L21,*L22; 566 if (match(L1, m_And(m_Value(L11), m_Value(L12)))) { 567 if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) 568 L21 = L22 = 0; 569 } 570 else { 571 if (!match(L2, m_And(m_Value(L11), m_Value(L12)))) 572 return 0; 573 std::swap(L1, L2); 574 L21 = L22 = 0; 575 } 576 577 Value *R1 = RHS->getOperand(0); 578 Value *R2 = RHS->getOperand(1); 579 Value *R11,*R12; 580 bool ok = false; 581 if (match(R1, m_And(m_Value(R11), m_Value(R12)))) { 582 if (R11 != 0 && (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22)) { 583 A = R11; D = R12; E = R2; ok = true; 584 } 585 else 586 if (R12 != 0 && (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22)) { 587 A = R12; D = R11; E = R2; ok = true; 588 } 589 } 590 if (!ok && match(R2, m_And(m_Value(R11), m_Value(R12)))) { 591 if (R11 != 0 && (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22)) { 592 A = R11; D = R12; E = R1; ok = true; 593 } 594 else 595 if (R12 != 0 && (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22)) { 596 A = R12; D = R11; E = R1; ok = true; 597 } 598 else 599 return 0; 600 } 601 if (!ok) 602 return 0; 603 604 if (L11 == A) { 605 B = L12; C = L2; 606 } 607 else if (L12 == A) { 608 B = L11; C = L2; 609 } 610 else if (L21 == A) { 611 B = L22; C = L1; 612 } 613 else if (L22 == A) { 614 B = L21; C = L1; 615 } 616 617 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC); 618 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC); 619 return left_type & right_type; 620} 621/// foldLogOpOfMaskedICmps: 622/// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) 623/// into a single (icmp(A & X) ==/!= Y) 624static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, 625 ICmpInst::Predicate NEWCC, 626 llvm::InstCombiner::BuilderTy* Builder) { 627 Value *A = 0, *B = 0, *C = 0, *D = 0, *E = 0; 628 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS); 629 if (mask == 0) return 0; 630 631 if (NEWCC == ICmpInst::ICMP_NE) 632 mask >>= 1; // treat "Not"-states as normal states 633 634 if (mask & FoldMskICmp_Mask_AllZeroes) { 635 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0) 636 // -> (icmp eq (A & (B|D)), 0) 637 Value* newOr = Builder->CreateOr(B, D); 638 Value* newAnd = Builder->CreateAnd(A, newOr); 639 // we can't use C as zero, because we might actually handle 640 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 641 // with B and D, having a single bit set 642 Value* zero = Constant::getNullValue(A->getType()); 643 return Builder->CreateICmp(NEWCC, newAnd, zero); 644 } 645 else if (mask & FoldMskICmp_BMask_AllOnes) { 646 // (icmp eq (A & B), B) & (icmp eq (A & D), D) 647 // -> (icmp eq (A & (B|D)), (B|D)) 648 Value* newOr = Builder->CreateOr(B, D); 649 Value* newAnd = Builder->CreateAnd(A, newOr); 650 return Builder->CreateICmp(NEWCC, newAnd, newOr); 651 } 652 else if (mask & FoldMskICmp_AMask_AllOnes) { 653 // (icmp eq (A & B), A) & (icmp eq (A & D), A) 654 // -> (icmp eq (A & (B&D)), A) 655 Value* newAnd1 = Builder->CreateAnd(B, D); 656 Value* newAnd = Builder->CreateAnd(A, newAnd1); 657 return Builder->CreateICmp(NEWCC, newAnd, A); 658 } 659 else if (mask & FoldMskICmp_BMask_Mixed) { 660 // (icmp eq (A & B), C) & (icmp eq (A & D), E) 661 // We already know that B & C == C && D & E == E. 662 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of 663 // C and E, which are shared by both the mask B and the mask D, don't 664 // contradict, then we can transform to 665 // -> (icmp eq (A & (B|D)), (C|E)) 666 // Currently, we only handle the case of B, C, D, and E being constant. 667 ConstantInt *BCst = dyn_cast<ConstantInt>(B); 668 if (BCst == 0) return 0; 669 ConstantInt *DCst = dyn_cast<ConstantInt>(D); 670 if (DCst == 0) return 0; 671 // we can't simply use C and E, because we might actually handle 672 // (icmp ne (A & B), B) & (icmp eq (A & D), D) 673 // with B and D, having a single bit set 674 675 ConstantInt *CCst = dyn_cast<ConstantInt>(C); 676 if (CCst == 0) return 0; 677 if (LHS->getPredicate() != NEWCC) 678 CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) ); 679 ConstantInt *ECst = dyn_cast<ConstantInt>(E); 680 if (ECst == 0) return 0; 681 if (RHS->getPredicate() != NEWCC) 682 ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) ); 683 ConstantInt* MCst = dyn_cast<ConstantInt>( 684 ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst), 685 ConstantExpr::getXor(CCst, ECst)) ); 686 // if there is a conflict we should actually return a false for the 687 // whole construct 688 if (!MCst->isZero()) 689 return 0; 690 Value* newOr1 = Builder->CreateOr(B, D); 691 Value* newOr2 = ConstantExpr::getOr(CCst, ECst); 692 Value* newAnd = Builder->CreateAnd(A, newOr1); 693 return Builder->CreateICmp(NEWCC, newAnd, newOr2); 694 } 695 return 0; 696} 697 698/// FoldAndOfICmps - Fold (icmp)&(icmp) if possible. 699Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) { 700 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate(); 701 702 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B) 703 if (PredicatesFoldable(LHSCC, RHSCC)) { 704 if (LHS->getOperand(0) == RHS->getOperand(1) && 705 LHS->getOperand(1) == RHS->getOperand(0)) 706 LHS->swapOperands(); 707 if (LHS->getOperand(0) == RHS->getOperand(0) && 708 LHS->getOperand(1) == RHS->getOperand(1)) { 709 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 710 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS); 711 bool isSigned = LHS->isSigned() || RHS->isSigned(); 712 return getICmpValue(isSigned, Code, Op0, Op1, Builder); 713 } 714 } 715 716 { 717 // handle (roughly): 718 // (icmp eq (A & B), C) & (icmp eq (A & D), E) 719 Value* fold = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_EQ, Builder); 720 if (fold) return fold; 721 } 722 723 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2). 724 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0); 725 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1)); 726 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1)); 727 if (LHSCst == 0 || RHSCst == 0) return 0; 728 729 if (LHSCst == RHSCst && LHSCC == RHSCC) { 730 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C) 731 // where C is a power of 2 732 if (LHSCC == ICmpInst::ICMP_ULT && 733 LHSCst->getValue().isPowerOf2()) { 734 Value *NewOr = Builder->CreateOr(Val, Val2); 735 return Builder->CreateICmp(LHSCC, NewOr, LHSCst); 736 } 737 738 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0) 739 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) { 740 Value *NewOr = Builder->CreateOr(Val, Val2); 741 return Builder->CreateICmp(LHSCC, NewOr, LHSCst); 742 } 743 } 744 745 // From here on, we only handle: 746 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler. 747 if (Val != Val2) return 0; 748 749 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere. 750 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE || 751 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE || 752 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE || 753 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE) 754 return 0; 755 756 // We can't fold (ugt x, C) & (sgt x, C2). 757 if (!PredicatesFoldable(LHSCC, RHSCC)) 758 return 0; 759 760 // Ensure that the larger constant is on the RHS. 761 bool ShouldSwap; 762 if (CmpInst::isSigned(LHSCC) || 763 (ICmpInst::isEquality(LHSCC) && 764 CmpInst::isSigned(RHSCC))) 765 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue()); 766 else 767 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue()); 768 769 if (ShouldSwap) { 770 std::swap(LHS, RHS); 771 std::swap(LHSCst, RHSCst); 772 std::swap(LHSCC, RHSCC); 773 } 774 775 // At this point, we know we have two icmp instructions 776 // comparing a value against two constants and and'ing the result 777 // together. Because of the above check, we know that we only have 778 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know 779 // (from the icmp folding check above), that the two constants 780 // are not equal and that the larger constant is on the RHS 781 assert(LHSCst != RHSCst && "Compares not folded above?"); 782 783 switch (LHSCC) { 784 default: llvm_unreachable("Unknown integer condition code!"); 785 case ICmpInst::ICMP_EQ: 786 switch (RHSCC) { 787 default: llvm_unreachable("Unknown integer condition code!"); 788 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false 789 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false 790 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false 791 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 792 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13 793 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13 794 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13 795 return LHS; 796 } 797 case ICmpInst::ICMP_NE: 798 switch (RHSCC) { 799 default: llvm_unreachable("Unknown integer condition code!"); 800 case ICmpInst::ICMP_ULT: 801 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13 802 return Builder->CreateICmpULT(Val, LHSCst); 803 break; // (X != 13 & X u< 15) -> no change 804 case ICmpInst::ICMP_SLT: 805 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13 806 return Builder->CreateICmpSLT(Val, LHSCst); 807 break; // (X != 13 & X s< 15) -> no change 808 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15 809 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15 810 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15 811 return RHS; 812 case ICmpInst::ICMP_NE: 813 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1 814 Constant *AddCST = ConstantExpr::getNeg(LHSCst); 815 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off"); 816 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1)); 817 } 818 break; // (X != 13 & X != 15) -> no change 819 } 820 break; 821 case ICmpInst::ICMP_ULT: 822 switch (RHSCC) { 823 default: llvm_unreachable("Unknown integer condition code!"); 824 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false 825 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false 826 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 827 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change 828 break; 829 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13 830 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13 831 return LHS; 832 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change 833 break; 834 } 835 break; 836 case ICmpInst::ICMP_SLT: 837 switch (RHSCC) { 838 default: llvm_unreachable("Unknown integer condition code!"); 839 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false 840 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false 841 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 842 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change 843 break; 844 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13 845 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13 846 return LHS; 847 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change 848 break; 849 } 850 break; 851 case ICmpInst::ICMP_UGT: 852 switch (RHSCC) { 853 default: llvm_unreachable("Unknown integer condition code!"); 854 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15 855 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15 856 return RHS; 857 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change 858 break; 859 case ICmpInst::ICMP_NE: 860 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14 861 return Builder->CreateICmp(LHSCC, Val, RHSCst); 862 break; // (X u> 13 & X != 15) -> no change 863 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1 864 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true); 865 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change 866 break; 867 } 868 break; 869 case ICmpInst::ICMP_SGT: 870 switch (RHSCC) { 871 default: llvm_unreachable("Unknown integer condition code!"); 872 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15 873 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15 874 return RHS; 875 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change 876 break; 877 case ICmpInst::ICMP_NE: 878 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14 879 return Builder->CreateICmp(LHSCC, Val, RHSCst); 880 break; // (X s> 13 & X != 15) -> no change 881 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1 882 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true); 883 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change 884 break; 885 } 886 break; 887 } 888 889 return 0; 890} 891 892/// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of 893/// instcombine, this returns a Value which should already be inserted into the 894/// function. 895Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) { 896 if (LHS->getPredicate() == FCmpInst::FCMP_ORD && 897 RHS->getPredicate() == FCmpInst::FCMP_ORD) { 898 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y) 899 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1))) 900 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) { 901 // If either of the constants are nans, then the whole thing returns 902 // false. 903 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN()) 904 return ConstantInt::getFalse(LHS->getContext()); 905 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0)); 906 } 907 908 // Handle vector zeros. This occurs because the canonical form of 909 // "fcmp ord x,x" is "fcmp ord x, 0". 910 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) && 911 isa<ConstantAggregateZero>(RHS->getOperand(1))) 912 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0)); 913 return 0; 914 } 915 916 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1); 917 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1); 918 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate(); 919 920 921 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) { 922 // Swap RHS operands to match LHS. 923 Op1CC = FCmpInst::getSwappedPredicate(Op1CC); 924 std::swap(Op1LHS, Op1RHS); 925 } 926 927 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) { 928 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y). 929 if (Op0CC == Op1CC) 930 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS); 931 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE) 932 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 933 if (Op0CC == FCmpInst::FCMP_TRUE) 934 return RHS; 935 if (Op1CC == FCmpInst::FCMP_TRUE) 936 return LHS; 937 938 bool Op0Ordered; 939 bool Op1Ordered; 940 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered); 941 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered); 942 if (Op1Pred == 0) { 943 std::swap(LHS, RHS); 944 std::swap(Op0Pred, Op1Pred); 945 std::swap(Op0Ordered, Op1Ordered); 946 } 947 if (Op0Pred == 0) { 948 // uno && ueq -> uno && (uno || eq) -> ueq 949 // ord && olt -> ord && (ord && lt) -> olt 950 if (Op0Ordered == Op1Ordered) 951 return RHS; 952 953 // uno && oeq -> uno && (ord && eq) -> false 954 // uno && ord -> false 955 if (!Op0Ordered) 956 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 957 // ord && ueq -> ord && (uno || eq) -> oeq 958 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder); 959 } 960 } 961 962 return 0; 963} 964 965 966Instruction *InstCombiner::visitAnd(BinaryOperator &I) { 967 bool Changed = SimplifyCommutative(I); 968 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 969 970 if (Value *V = SimplifyAndInst(Op0, Op1, TD)) 971 return ReplaceInstUsesWith(I, V); 972 973 // See if we can simplify any instructions used by the instruction whose sole 974 // purpose is to compute bits we don't care about. 975 if (SimplifyDemandedInstructionBits(I)) 976 return &I; 977 978 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) { 979 const APInt &AndRHSMask = AndRHS->getValue(); 980 APInt NotAndRHS(~AndRHSMask); 981 982 // Optimize a variety of ((val OP C1) & C2) combinations... 983 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) { 984 Value *Op0LHS = Op0I->getOperand(0); 985 Value *Op0RHS = Op0I->getOperand(1); 986 switch (Op0I->getOpcode()) { 987 default: break; 988 case Instruction::Xor: 989 case Instruction::Or: 990 // If the mask is only needed on one incoming arm, push it up. 991 if (!Op0I->hasOneUse()) break; 992 993 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) { 994 // Not masking anything out for the LHS, move to RHS. 995 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS, 996 Op0RHS->getName()+".masked"); 997 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS); 998 } 999 if (!isa<Constant>(Op0RHS) && 1000 MaskedValueIsZero(Op0RHS, NotAndRHS)) { 1001 // Not masking anything out for the RHS, move to LHS. 1002 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS, 1003 Op0LHS->getName()+".masked"); 1004 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS); 1005 } 1006 1007 break; 1008 case Instruction::Add: 1009 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS. 1010 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0 1011 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0 1012 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I)) 1013 return BinaryOperator::CreateAnd(V, AndRHS); 1014 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I)) 1015 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes 1016 break; 1017 1018 case Instruction::Sub: 1019 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS. 1020 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0 1021 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0 1022 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I)) 1023 return BinaryOperator::CreateAnd(V, AndRHS); 1024 1025 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS 1026 // has 1's for all bits that the subtraction with A might affect. 1027 if (Op0I->hasOneUse()) { 1028 uint32_t BitWidth = AndRHSMask.getBitWidth(); 1029 uint32_t Zeros = AndRHSMask.countLeadingZeros(); 1030 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros); 1031 1032 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS); 1033 if (!(A && A->isZero()) && // avoid infinite recursion. 1034 MaskedValueIsZero(Op0LHS, Mask)) { 1035 Value *NewNeg = Builder->CreateNeg(Op0RHS); 1036 return BinaryOperator::CreateAnd(NewNeg, AndRHS); 1037 } 1038 } 1039 break; 1040 1041 case Instruction::Shl: 1042 case Instruction::LShr: 1043 // (1 << x) & 1 --> zext(x == 0) 1044 // (1 >> x) & 1 --> zext(x == 0) 1045 if (AndRHSMask == 1 && Op0LHS == AndRHS) { 1046 Value *NewICmp = 1047 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType())); 1048 return new ZExtInst(NewICmp, I.getType()); 1049 } 1050 break; 1051 } 1052 1053 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) 1054 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I)) 1055 return Res; 1056 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) { 1057 // If this is an integer truncation or change from signed-to-unsigned, and 1058 // if the source is an and/or with immediate, transform it. This 1059 // frequently occurs for bitfield accesses. 1060 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) { 1061 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) && 1062 CastOp->getNumOperands() == 2) 1063 if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1))){ 1064 if (CastOp->getOpcode() == Instruction::And) { 1065 // Change: and (cast (and X, C1) to T), C2 1066 // into : and (cast X to T), trunc_or_bitcast(C1)&C2 1067 // This will fold the two constants together, which may allow 1068 // other simplifications. 1069 Value *NewCast = Builder->CreateTruncOrBitCast( 1070 CastOp->getOperand(0), I.getType(), 1071 CastOp->getName()+".shrunk"); 1072 // trunc_or_bitcast(C1)&C2 1073 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType()); 1074 C3 = ConstantExpr::getAnd(C3, AndRHS); 1075 return BinaryOperator::CreateAnd(NewCast, C3); 1076 } else if (CastOp->getOpcode() == Instruction::Or) { 1077 // Change: and (cast (or X, C1) to T), C2 1078 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2 1079 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType()); 1080 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) 1081 // trunc(C1)&C2 1082 return ReplaceInstUsesWith(I, AndRHS); 1083 } 1084 } 1085 } 1086 } 1087 1088 // Try to fold constant and into select arguments. 1089 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 1090 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1091 return R; 1092 if (isa<PHINode>(Op0)) 1093 if (Instruction *NV = FoldOpIntoPhi(I)) 1094 return NV; 1095 } 1096 1097 1098 // (~A & ~B) == (~(A | B)) - De Morgan's Law 1099 if (Value *Op0NotVal = dyn_castNotVal(Op0)) 1100 if (Value *Op1NotVal = dyn_castNotVal(Op1)) 1101 if (Op0->hasOneUse() && Op1->hasOneUse()) { 1102 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal, 1103 I.getName()+".demorgan"); 1104 return BinaryOperator::CreateNot(Or); 1105 } 1106 1107 { 1108 Value *A = 0, *B = 0, *C = 0, *D = 0; 1109 // (A|B) & ~(A&B) -> A^B 1110 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1111 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) && 1112 ((A == C && B == D) || (A == D && B == C))) 1113 return BinaryOperator::CreateXor(A, B); 1114 1115 // ~(A&B) & (A|B) -> A^B 1116 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1117 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) && 1118 ((A == C && B == D) || (A == D && B == C))) 1119 return BinaryOperator::CreateXor(A, B); 1120 1121 if (Op0->hasOneUse() && 1122 match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 1123 if (A == Op1) { // (A^B)&A -> A&(A^B) 1124 I.swapOperands(); // Simplify below 1125 std::swap(Op0, Op1); 1126 } else if (B == Op1) { // (A^B)&B -> B&(B^A) 1127 cast<BinaryOperator>(Op0)->swapOperands(); 1128 I.swapOperands(); // Simplify below 1129 std::swap(Op0, Op1); 1130 } 1131 } 1132 1133 if (Op1->hasOneUse() && 1134 match(Op1, m_Xor(m_Value(A), m_Value(B)))) { 1135 if (B == Op0) { // B&(A^B) -> B&(B^A) 1136 cast<BinaryOperator>(Op1)->swapOperands(); 1137 std::swap(A, B); 1138 } 1139 if (A == Op0) // A&(A^B) -> A & ~B 1140 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp")); 1141 } 1142 1143 // (A&((~A)|B)) -> A&B 1144 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) || 1145 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1))))) 1146 return BinaryOperator::CreateAnd(A, Op1); 1147 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) || 1148 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0))))) 1149 return BinaryOperator::CreateAnd(A, Op0); 1150 } 1151 1152 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) 1153 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0)) 1154 if (Value *Res = FoldAndOfICmps(LHS, RHS)) 1155 return ReplaceInstUsesWith(I, Res); 1156 1157 // If and'ing two fcmp, try combine them into one. 1158 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 1159 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 1160 if (Value *Res = FoldAndOfFCmps(LHS, RHS)) 1161 return ReplaceInstUsesWith(I, Res); 1162 1163 1164 // fold (and (cast A), (cast B)) -> (cast (and A, B)) 1165 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) 1166 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) { 1167 const Type *SrcTy = Op0C->getOperand(0)->getType(); 1168 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ? 1169 SrcTy == Op1C->getOperand(0)->getType() && 1170 SrcTy->isIntOrIntVectorTy()) { 1171 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0); 1172 1173 // Only do this if the casts both really cause code to be generated. 1174 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) && 1175 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) { 1176 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName()); 1177 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); 1178 } 1179 1180 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the 1181 // cast is otherwise not optimizable. This happens for vector sexts. 1182 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp)) 1183 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp)) 1184 if (Value *Res = FoldAndOfICmps(LHS, RHS)) 1185 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 1186 1187 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the 1188 // cast is otherwise not optimizable. This happens for vector sexts. 1189 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp)) 1190 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp)) 1191 if (Value *Res = FoldAndOfFCmps(LHS, RHS)) 1192 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 1193 } 1194 } 1195 1196 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts. 1197 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) { 1198 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0)) 1199 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() && 1200 SI0->getOperand(1) == SI1->getOperand(1) && 1201 (SI0->hasOneUse() || SI1->hasOneUse())) { 1202 Value *NewOp = 1203 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0), 1204 SI0->getName()); 1205 return BinaryOperator::Create(SI1->getOpcode(), NewOp, 1206 SI1->getOperand(1)); 1207 } 1208 } 1209 1210 return Changed ? &I : 0; 1211} 1212 1213/// CollectBSwapParts - Analyze the specified subexpression and see if it is 1214/// capable of providing pieces of a bswap. The subexpression provides pieces 1215/// of a bswap if it is proven that each of the non-zero bytes in the output of 1216/// the expression came from the corresponding "byte swapped" byte in some other 1217/// value. For example, if the current subexpression is "(shl i32 %X, 24)" then 1218/// we know that the expression deposits the low byte of %X into the high byte 1219/// of the bswap result and that all other bytes are zero. This expression is 1220/// accepted, the high byte of ByteValues is set to X to indicate a correct 1221/// match. 1222/// 1223/// This function returns true if the match was unsuccessful and false if so. 1224/// On entry to the function the "OverallLeftShift" is a signed integer value 1225/// indicating the number of bytes that the subexpression is later shifted. For 1226/// example, if the expression is later right shifted by 16 bits, the 1227/// OverallLeftShift value would be -2 on entry. This is used to specify which 1228/// byte of ByteValues is actually being set. 1229/// 1230/// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding 1231/// byte is masked to zero by a user. For example, in (X & 255), X will be 1232/// processed with a bytemask of 1. Because bytemask is 32-bits, this limits 1233/// this function to working on up to 32-byte (256 bit) values. ByteMask is 1234/// always in the local (OverallLeftShift) coordinate space. 1235/// 1236static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask, 1237 SmallVector<Value*, 8> &ByteValues) { 1238 if (Instruction *I = dyn_cast<Instruction>(V)) { 1239 // If this is an or instruction, it may be an inner node of the bswap. 1240 if (I->getOpcode() == Instruction::Or) { 1241 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask, 1242 ByteValues) || 1243 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask, 1244 ByteValues); 1245 } 1246 1247 // If this is a logical shift by a constant multiple of 8, recurse with 1248 // OverallLeftShift and ByteMask adjusted. 1249 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) { 1250 unsigned ShAmt = 1251 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U); 1252 // Ensure the shift amount is defined and of a byte value. 1253 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size())) 1254 return true; 1255 1256 unsigned ByteShift = ShAmt >> 3; 1257 if (I->getOpcode() == Instruction::Shl) { 1258 // X << 2 -> collect(X, +2) 1259 OverallLeftShift += ByteShift; 1260 ByteMask >>= ByteShift; 1261 } else { 1262 // X >>u 2 -> collect(X, -2) 1263 OverallLeftShift -= ByteShift; 1264 ByteMask <<= ByteShift; 1265 ByteMask &= (~0U >> (32-ByteValues.size())); 1266 } 1267 1268 if (OverallLeftShift >= (int)ByteValues.size()) return true; 1269 if (OverallLeftShift <= -(int)ByteValues.size()) return true; 1270 1271 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask, 1272 ByteValues); 1273 } 1274 1275 // If this is a logical 'and' with a mask that clears bytes, clear the 1276 // corresponding bytes in ByteMask. 1277 if (I->getOpcode() == Instruction::And && 1278 isa<ConstantInt>(I->getOperand(1))) { 1279 // Scan every byte of the and mask, seeing if the byte is either 0 or 255. 1280 unsigned NumBytes = ByteValues.size(); 1281 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255); 1282 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue(); 1283 1284 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) { 1285 // If this byte is masked out by a later operation, we don't care what 1286 // the and mask is. 1287 if ((ByteMask & (1 << i)) == 0) 1288 continue; 1289 1290 // If the AndMask is all zeros for this byte, clear the bit. 1291 APInt MaskB = AndMask & Byte; 1292 if (MaskB == 0) { 1293 ByteMask &= ~(1U << i); 1294 continue; 1295 } 1296 1297 // If the AndMask is not all ones for this byte, it's not a bytezap. 1298 if (MaskB != Byte) 1299 return true; 1300 1301 // Otherwise, this byte is kept. 1302 } 1303 1304 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask, 1305 ByteValues); 1306 } 1307 } 1308 1309 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be 1310 // the input value to the bswap. Some observations: 1) if more than one byte 1311 // is demanded from this input, then it could not be successfully assembled 1312 // into a byteswap. At least one of the two bytes would not be aligned with 1313 // their ultimate destination. 1314 if (!isPowerOf2_32(ByteMask)) return true; 1315 unsigned InputByteNo = CountTrailingZeros_32(ByteMask); 1316 1317 // 2) The input and ultimate destinations must line up: if byte 3 of an i32 1318 // is demanded, it needs to go into byte 0 of the result. This means that the 1319 // byte needs to be shifted until it lands in the right byte bucket. The 1320 // shift amount depends on the position: if the byte is coming from the high 1321 // part of the value (e.g. byte 3) then it must be shifted right. If from the 1322 // low part, it must be shifted left. 1323 unsigned DestByteNo = InputByteNo + OverallLeftShift; 1324 if (InputByteNo < ByteValues.size()/2) { 1325 if (ByteValues.size()-1-DestByteNo != InputByteNo) 1326 return true; 1327 } else { 1328 if (ByteValues.size()-1-DestByteNo != InputByteNo) 1329 return true; 1330 } 1331 1332 // If the destination byte value is already defined, the values are or'd 1333 // together, which isn't a bswap (unless it's an or of the same bits). 1334 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V) 1335 return true; 1336 ByteValues[DestByteNo] = V; 1337 return false; 1338} 1339 1340/// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom. 1341/// If so, insert the new bswap intrinsic and return it. 1342Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) { 1343 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType()); 1344 if (!ITy || ITy->getBitWidth() % 16 || 1345 // ByteMask only allows up to 32-byte values. 1346 ITy->getBitWidth() > 32*8) 1347 return 0; // Can only bswap pairs of bytes. Can't do vectors. 1348 1349 /// ByteValues - For each byte of the result, we keep track of which value 1350 /// defines each byte. 1351 SmallVector<Value*, 8> ByteValues; 1352 ByteValues.resize(ITy->getBitWidth()/8); 1353 1354 // Try to find all the pieces corresponding to the bswap. 1355 uint32_t ByteMask = ~0U >> (32-ByteValues.size()); 1356 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues)) 1357 return 0; 1358 1359 // Check to see if all of the bytes come from the same value. 1360 Value *V = ByteValues[0]; 1361 if (V == 0) return 0; // Didn't find a byte? Must be zero. 1362 1363 // Check to make sure that all of the bytes come from the same value. 1364 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i) 1365 if (ByteValues[i] != V) 1366 return 0; 1367 const Type *Tys[] = { ITy }; 1368 Module *M = I.getParent()->getParent()->getParent(); 1369 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1); 1370 return CallInst::Create(F, V); 1371} 1372 1373/// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check 1374/// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then 1375/// we can simplify this expression to "cond ? C : D or B". 1376static Instruction *MatchSelectFromAndOr(Value *A, Value *B, 1377 Value *C, Value *D) { 1378 // If A is not a select of -1/0, this cannot match. 1379 Value *Cond = 0; 1380 if (!match(A, m_SExt(m_Value(Cond))) || 1381 !Cond->getType()->isIntegerTy(1)) 1382 return 0; 1383 1384 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B. 1385 if (match(D, m_Not(m_SExt(m_Specific(Cond))))) 1386 return SelectInst::Create(Cond, C, B); 1387 if (match(D, m_SExt(m_Not(m_Specific(Cond))))) 1388 return SelectInst::Create(Cond, C, B); 1389 1390 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D. 1391 if (match(B, m_Not(m_SExt(m_Specific(Cond))))) 1392 return SelectInst::Create(Cond, C, D); 1393 if (match(B, m_SExt(m_Not(m_Specific(Cond))))) 1394 return SelectInst::Create(Cond, C, D); 1395 return 0; 1396} 1397 1398/// FoldOrOfICmps - Fold (icmp)|(icmp) if possible. 1399Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) { 1400 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate(); 1401 1402 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B) 1403 if (PredicatesFoldable(LHSCC, RHSCC)) { 1404 if (LHS->getOperand(0) == RHS->getOperand(1) && 1405 LHS->getOperand(1) == RHS->getOperand(0)) 1406 LHS->swapOperands(); 1407 if (LHS->getOperand(0) == RHS->getOperand(0) && 1408 LHS->getOperand(1) == RHS->getOperand(1)) { 1409 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 1410 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS); 1411 bool isSigned = LHS->isSigned() || RHS->isSigned(); 1412 return getICmpValue(isSigned, Code, Op0, Op1, Builder); 1413 } 1414 } 1415 1416 { 1417 // handle (roughly): 1418 // (icmp ne (A & B), C) | (icmp ne (A & D), E) 1419 Value* fold = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_NE, Builder); 1420 if (fold) return fold; 1421 } 1422 1423 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2). 1424 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0); 1425 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1)); 1426 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1)); 1427 if (LHSCst == 0 || RHSCst == 0) return 0; 1428 1429 if (LHSCst == RHSCst && LHSCC == RHSCC) { 1430 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0) 1431 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) { 1432 Value *NewOr = Builder->CreateOr(Val, Val2); 1433 return Builder->CreateICmp(LHSCC, NewOr, LHSCst); 1434 } 1435 } 1436 1437 // From here on, we only handle: 1438 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler. 1439 if (Val != Val2) return 0; 1440 1441 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere. 1442 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE || 1443 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE || 1444 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE || 1445 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE) 1446 return 0; 1447 1448 // We can't fold (ugt x, C) | (sgt x, C2). 1449 if (!PredicatesFoldable(LHSCC, RHSCC)) 1450 return 0; 1451 1452 // Ensure that the larger constant is on the RHS. 1453 bool ShouldSwap; 1454 if (CmpInst::isSigned(LHSCC) || 1455 (ICmpInst::isEquality(LHSCC) && 1456 CmpInst::isSigned(RHSCC))) 1457 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue()); 1458 else 1459 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue()); 1460 1461 if (ShouldSwap) { 1462 std::swap(LHS, RHS); 1463 std::swap(LHSCst, RHSCst); 1464 std::swap(LHSCC, RHSCC); 1465 } 1466 1467 // At this point, we know we have two icmp instructions 1468 // comparing a value against two constants and or'ing the result 1469 // together. Because of the above check, we know that we only have 1470 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the 1471 // icmp folding check above), that the two constants are not 1472 // equal. 1473 assert(LHSCst != RHSCst && "Compares not folded above?"); 1474 1475 switch (LHSCC) { 1476 default: llvm_unreachable("Unknown integer condition code!"); 1477 case ICmpInst::ICMP_EQ: 1478 switch (RHSCC) { 1479 default: llvm_unreachable("Unknown integer condition code!"); 1480 case ICmpInst::ICMP_EQ: 1481 if (LHSCst == SubOne(RHSCst)) { 1482 // (X == 13 | X == 14) -> X-13 <u 2 1483 Constant *AddCST = ConstantExpr::getNeg(LHSCst); 1484 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off"); 1485 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst); 1486 return Builder->CreateICmpULT(Add, AddCST); 1487 } 1488 break; // (X == 13 | X == 15) -> no change 1489 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change 1490 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change 1491 break; 1492 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15 1493 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15 1494 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15 1495 return RHS; 1496 } 1497 break; 1498 case ICmpInst::ICMP_NE: 1499 switch (RHSCC) { 1500 default: llvm_unreachable("Unknown integer condition code!"); 1501 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13 1502 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13 1503 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13 1504 return LHS; 1505 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true 1506 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true 1507 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true 1508 return ConstantInt::getTrue(LHS->getContext()); 1509 } 1510 break; 1511 case ICmpInst::ICMP_ULT: 1512 switch (RHSCC) { 1513 default: llvm_unreachable("Unknown integer condition code!"); 1514 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change 1515 break; 1516 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2 1517 // If RHSCst is [us]MAXINT, it is always false. Not handling 1518 // this can cause overflow. 1519 if (RHSCst->isMaxValue(false)) 1520 return LHS; 1521 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false); 1522 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change 1523 break; 1524 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15 1525 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15 1526 return RHS; 1527 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change 1528 break; 1529 } 1530 break; 1531 case ICmpInst::ICMP_SLT: 1532 switch (RHSCC) { 1533 default: llvm_unreachable("Unknown integer condition code!"); 1534 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change 1535 break; 1536 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2 1537 // If RHSCst is [us]MAXINT, it is always false. Not handling 1538 // this can cause overflow. 1539 if (RHSCst->isMaxValue(true)) 1540 return LHS; 1541 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false); 1542 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change 1543 break; 1544 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15 1545 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15 1546 return RHS; 1547 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change 1548 break; 1549 } 1550 break; 1551 case ICmpInst::ICMP_UGT: 1552 switch (RHSCC) { 1553 default: llvm_unreachable("Unknown integer condition code!"); 1554 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13 1555 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13 1556 return LHS; 1557 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change 1558 break; 1559 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true 1560 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true 1561 return ConstantInt::getTrue(LHS->getContext()); 1562 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change 1563 break; 1564 } 1565 break; 1566 case ICmpInst::ICMP_SGT: 1567 switch (RHSCC) { 1568 default: llvm_unreachable("Unknown integer condition code!"); 1569 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13 1570 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13 1571 return LHS; 1572 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change 1573 break; 1574 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true 1575 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true 1576 return ConstantInt::getTrue(LHS->getContext()); 1577 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change 1578 break; 1579 } 1580 break; 1581 } 1582 return 0; 1583} 1584 1585/// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of 1586/// instcombine, this returns a Value which should already be inserted into the 1587/// function. 1588Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) { 1589 if (LHS->getPredicate() == FCmpInst::FCMP_UNO && 1590 RHS->getPredicate() == FCmpInst::FCMP_UNO && 1591 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) { 1592 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1))) 1593 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) { 1594 // If either of the constants are nans, then the whole thing returns 1595 // true. 1596 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN()) 1597 return ConstantInt::getTrue(LHS->getContext()); 1598 1599 // Otherwise, no need to compare the two constants, compare the 1600 // rest. 1601 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0)); 1602 } 1603 1604 // Handle vector zeros. This occurs because the canonical form of 1605 // "fcmp uno x,x" is "fcmp uno x, 0". 1606 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) && 1607 isa<ConstantAggregateZero>(RHS->getOperand(1))) 1608 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0)); 1609 1610 return 0; 1611 } 1612 1613 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1); 1614 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1); 1615 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate(); 1616 1617 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) { 1618 // Swap RHS operands to match LHS. 1619 Op1CC = FCmpInst::getSwappedPredicate(Op1CC); 1620 std::swap(Op1LHS, Op1RHS); 1621 } 1622 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) { 1623 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y). 1624 if (Op0CC == Op1CC) 1625 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS); 1626 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE) 1627 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1); 1628 if (Op0CC == FCmpInst::FCMP_FALSE) 1629 return RHS; 1630 if (Op1CC == FCmpInst::FCMP_FALSE) 1631 return LHS; 1632 bool Op0Ordered; 1633 bool Op1Ordered; 1634 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered); 1635 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered); 1636 if (Op0Ordered == Op1Ordered) { 1637 // If both are ordered or unordered, return a new fcmp with 1638 // or'ed predicates. 1639 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder); 1640 } 1641 } 1642 return 0; 1643} 1644 1645/// FoldOrWithConstants - This helper function folds: 1646/// 1647/// ((A | B) & C1) | (B & C2) 1648/// 1649/// into: 1650/// 1651/// (A & C1) | B 1652/// 1653/// when the XOR of the two constants is "all ones" (-1). 1654Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op, 1655 Value *A, Value *B, Value *C) { 1656 ConstantInt *CI1 = dyn_cast<ConstantInt>(C); 1657 if (!CI1) return 0; 1658 1659 Value *V1 = 0; 1660 ConstantInt *CI2 = 0; 1661 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0; 1662 1663 APInt Xor = CI1->getValue() ^ CI2->getValue(); 1664 if (!Xor.isAllOnesValue()) return 0; 1665 1666 if (V1 == A || V1 == B) { 1667 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1); 1668 return BinaryOperator::CreateOr(NewOp, V1); 1669 } 1670 1671 return 0; 1672} 1673 1674Instruction *InstCombiner::visitOr(BinaryOperator &I) { 1675 bool Changed = SimplifyCommutative(I); 1676 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1677 1678 if (Value *V = SimplifyOrInst(Op0, Op1, TD)) 1679 return ReplaceInstUsesWith(I, V); 1680 1681 // See if we can simplify any instructions used by the instruction whose sole 1682 // purpose is to compute bits we don't care about. 1683 if (SimplifyDemandedInstructionBits(I)) 1684 return &I; 1685 1686 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) { 1687 ConstantInt *C1 = 0; Value *X = 0; 1688 // (X & C1) | C2 --> (X | C2) & (C1|C2) 1689 // iff (C1 & C2) == 0. 1690 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && 1691 (RHS->getValue() & C1->getValue()) != 0 && 1692 Op0->hasOneUse()) { 1693 Value *Or = Builder->CreateOr(X, RHS); 1694 Or->takeName(Op0); 1695 return BinaryOperator::CreateAnd(Or, 1696 ConstantInt::get(I.getContext(), 1697 RHS->getValue() | C1->getValue())); 1698 } 1699 1700 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2) 1701 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && 1702 Op0->hasOneUse()) { 1703 Value *Or = Builder->CreateOr(X, RHS); 1704 Or->takeName(Op0); 1705 return BinaryOperator::CreateXor(Or, 1706 ConstantInt::get(I.getContext(), 1707 C1->getValue() & ~RHS->getValue())); 1708 } 1709 1710 // Try to fold constant and into select arguments. 1711 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 1712 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1713 return R; 1714 1715 if (isa<PHINode>(Op0)) 1716 if (Instruction *NV = FoldOpIntoPhi(I)) 1717 return NV; 1718 } 1719 1720 Value *A = 0, *B = 0; 1721 ConstantInt *C1 = 0, *C2 = 0; 1722 1723 // (A | B) | C and A | (B | C) -> bswap if possible. 1724 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible. 1725 if (match(Op0, m_Or(m_Value(), m_Value())) || 1726 match(Op1, m_Or(m_Value(), m_Value())) || 1727 (match(Op0, m_Shift(m_Value(), m_Value())) && 1728 match(Op1, m_Shift(m_Value(), m_Value())))) { 1729 if (Instruction *BSwap = MatchBSwap(I)) 1730 return BSwap; 1731 } 1732 1733 // (X^C)|Y -> (X|Y)^C iff Y&C == 0 1734 if (Op0->hasOneUse() && 1735 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) && 1736 MaskedValueIsZero(Op1, C1->getValue())) { 1737 Value *NOr = Builder->CreateOr(A, Op1); 1738 NOr->takeName(Op0); 1739 return BinaryOperator::CreateXor(NOr, C1); 1740 } 1741 1742 // Y|(X^C) -> (X|Y)^C iff Y&C == 0 1743 if (Op1->hasOneUse() && 1744 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) && 1745 MaskedValueIsZero(Op0, C1->getValue())) { 1746 Value *NOr = Builder->CreateOr(A, Op0); 1747 NOr->takeName(Op0); 1748 return BinaryOperator::CreateXor(NOr, C1); 1749 } 1750 1751 // (A & C)|(B & D) 1752 Value *C = 0, *D = 0; 1753 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 1754 match(Op1, m_And(m_Value(B), m_Value(D)))) { 1755 Value *V1 = 0, *V2 = 0, *V3 = 0; 1756 C1 = dyn_cast<ConstantInt>(C); 1757 C2 = dyn_cast<ConstantInt>(D); 1758 if (C1 && C2) { // (A & C1)|(B & C2) 1759 // If we have: ((V + N) & C1) | (V & C2) 1760 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0 1761 // replace with V+N. 1762 if (C1->getValue() == ~C2->getValue()) { 1763 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+ 1764 match(A, m_Add(m_Value(V1), m_Value(V2)))) { 1765 // Add commutes, try both ways. 1766 if (V1 == B && MaskedValueIsZero(V2, C2->getValue())) 1767 return ReplaceInstUsesWith(I, A); 1768 if (V2 == B && MaskedValueIsZero(V1, C2->getValue())) 1769 return ReplaceInstUsesWith(I, A); 1770 } 1771 // Or commutes, try both ways. 1772 if ((C1->getValue() & (C1->getValue()+1)) == 0 && 1773 match(B, m_Add(m_Value(V1), m_Value(V2)))) { 1774 // Add commutes, try both ways. 1775 if (V1 == A && MaskedValueIsZero(V2, C1->getValue())) 1776 return ReplaceInstUsesWith(I, B); 1777 if (V2 == A && MaskedValueIsZero(V1, C1->getValue())) 1778 return ReplaceInstUsesWith(I, B); 1779 } 1780 } 1781 1782 if ((C1->getValue() & C2->getValue()) == 0) { 1783 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2) 1784 // iff (C1&C2) == 0 and (N&~C1) == 0 1785 if (match(A, m_Or(m_Value(V1), m_Value(V2))) && 1786 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N) 1787 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V) 1788 return BinaryOperator::CreateAnd(A, 1789 ConstantInt::get(A->getContext(), 1790 C1->getValue()|C2->getValue())); 1791 // Or commutes, try both ways. 1792 if (match(B, m_Or(m_Value(V1), m_Value(V2))) && 1793 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N) 1794 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V) 1795 return BinaryOperator::CreateAnd(B, 1796 ConstantInt::get(B->getContext(), 1797 C1->getValue()|C2->getValue())); 1798 1799 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2) 1800 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0. 1801 ConstantInt *C3 = 0, *C4 = 0; 1802 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) && 1803 (C3->getValue() & ~C1->getValue()) == 0 && 1804 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) && 1805 (C4->getValue() & ~C2->getValue()) == 0) { 1806 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield"); 1807 return BinaryOperator::CreateAnd(V2, 1808 ConstantInt::get(B->getContext(), 1809 C1->getValue()|C2->getValue())); 1810 } 1811 } 1812 } 1813 1814 // Check to see if we have any common things being and'ed. If so, find the 1815 // terms for V1 & (V2|V3). 1816 if (Op0->hasOneUse() || Op1->hasOneUse()) { 1817 V1 = 0; 1818 if (A == B) // (A & C)|(A & D) == A & (C|D) 1819 V1 = A, V2 = C, V3 = D; 1820 else if (A == D) // (A & C)|(B & A) == A & (B|C) 1821 V1 = A, V2 = B, V3 = C; 1822 else if (C == B) // (A & C)|(C & D) == C & (A|D) 1823 V1 = C, V2 = A, V3 = D; 1824 else if (C == D) // (A & C)|(B & C) == C & (A|B) 1825 V1 = C, V2 = A, V3 = B; 1826 1827 if (V1) { 1828 Value *Or = Builder->CreateOr(V2, V3, "tmp"); 1829 return BinaryOperator::CreateAnd(V1, Or); 1830 } 1831 } 1832 1833 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants. 1834 // Don't do this for vector select idioms, the code generator doesn't handle 1835 // them well yet. 1836 if (!I.getType()->isVectorTy()) { 1837 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D)) 1838 return Match; 1839 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C)) 1840 return Match; 1841 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D)) 1842 return Match; 1843 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C)) 1844 return Match; 1845 } 1846 1847 // ((A&~B)|(~A&B)) -> A^B 1848 if ((match(C, m_Not(m_Specific(D))) && 1849 match(B, m_Not(m_Specific(A))))) 1850 return BinaryOperator::CreateXor(A, D); 1851 // ((~B&A)|(~A&B)) -> A^B 1852 if ((match(A, m_Not(m_Specific(D))) && 1853 match(B, m_Not(m_Specific(C))))) 1854 return BinaryOperator::CreateXor(C, D); 1855 // ((A&~B)|(B&~A)) -> A^B 1856 if ((match(C, m_Not(m_Specific(B))) && 1857 match(D, m_Not(m_Specific(A))))) 1858 return BinaryOperator::CreateXor(A, B); 1859 // ((~B&A)|(B&~A)) -> A^B 1860 if ((match(A, m_Not(m_Specific(B))) && 1861 match(D, m_Not(m_Specific(C))))) 1862 return BinaryOperator::CreateXor(C, B); 1863 1864 // ((A|B)&1)|(B&-2) -> (A&1) | B 1865 if (match(A, m_Or(m_Value(V1), m_Specific(B))) || 1866 match(A, m_Or(m_Specific(B), m_Value(V1)))) { 1867 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C); 1868 if (Ret) return Ret; 1869 } 1870 // (B&-2)|((A|B)&1) -> (A&1) | B 1871 if (match(B, m_Or(m_Specific(A), m_Value(V1))) || 1872 match(B, m_Or(m_Value(V1), m_Specific(A)))) { 1873 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D); 1874 if (Ret) return Ret; 1875 } 1876 } 1877 1878 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts. 1879 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) { 1880 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0)) 1881 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() && 1882 SI0->getOperand(1) == SI1->getOperand(1) && 1883 (SI0->hasOneUse() || SI1->hasOneUse())) { 1884 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0), 1885 SI0->getName()); 1886 return BinaryOperator::Create(SI1->getOpcode(), NewOp, 1887 SI1->getOperand(1)); 1888 } 1889 } 1890 1891 // (~A | ~B) == (~(A & B)) - De Morgan's Law 1892 if (Value *Op0NotVal = dyn_castNotVal(Op0)) 1893 if (Value *Op1NotVal = dyn_castNotVal(Op1)) 1894 if (Op0->hasOneUse() && Op1->hasOneUse()) { 1895 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal, 1896 I.getName()+".demorgan"); 1897 return BinaryOperator::CreateNot(And); 1898 } 1899 1900 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) 1901 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0))) 1902 if (Value *Res = FoldOrOfICmps(LHS, RHS)) 1903 return ReplaceInstUsesWith(I, Res); 1904 1905 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y) 1906 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 1907 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 1908 if (Value *Res = FoldOrOfFCmps(LHS, RHS)) 1909 return ReplaceInstUsesWith(I, Res); 1910 1911 // fold (or (cast A), (cast B)) -> (cast (or A, B)) 1912 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { 1913 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) 1914 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ? 1915 const Type *SrcTy = Op0C->getOperand(0)->getType(); 1916 if (SrcTy == Op1C->getOperand(0)->getType() && 1917 SrcTy->isIntOrIntVectorTy()) { 1918 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0); 1919 1920 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) && 1921 // Only do this if the casts both really cause code to be 1922 // generated. 1923 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) && 1924 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) { 1925 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName()); 1926 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); 1927 } 1928 1929 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the 1930 // cast is otherwise not optimizable. This happens for vector sexts. 1931 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp)) 1932 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp)) 1933 if (Value *Res = FoldOrOfICmps(LHS, RHS)) 1934 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 1935 1936 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the 1937 // cast is otherwise not optimizable. This happens for vector sexts. 1938 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp)) 1939 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp)) 1940 if (Value *Res = FoldOrOfFCmps(LHS, RHS)) 1941 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 1942 } 1943 } 1944 } 1945 1946 return Changed ? &I : 0; 1947} 1948 1949Instruction *InstCombiner::visitXor(BinaryOperator &I) { 1950 bool Changed = SimplifyCommutative(I); 1951 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1952 1953 if (isa<UndefValue>(Op1)) { 1954 if (isa<UndefValue>(Op0)) 1955 // Handle undef ^ undef -> 0 special case. This is a common 1956 // idiom (misuse). 1957 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType())); 1958 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef 1959 } 1960 1961 // xor X, X = 0 1962 if (Op0 == Op1) 1963 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType())); 1964 1965 // See if we can simplify any instructions used by the instruction whose sole 1966 // purpose is to compute bits we don't care about. 1967 if (SimplifyDemandedInstructionBits(I)) 1968 return &I; 1969 if (I.getType()->isVectorTy()) 1970 if (isa<ConstantAggregateZero>(Op1)) 1971 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X 1972 1973 // Is this a ~ operation? 1974 if (Value *NotOp = dyn_castNotVal(&I)) { 1975 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) { 1976 if (Op0I->getOpcode() == Instruction::And || 1977 Op0I->getOpcode() == Instruction::Or) { 1978 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law 1979 // ~(~X | Y) === (X & ~Y) - De Morgan's Law 1980 if (dyn_castNotVal(Op0I->getOperand(1))) 1981 Op0I->swapOperands(); 1982 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) { 1983 Value *NotY = 1984 Builder->CreateNot(Op0I->getOperand(1), 1985 Op0I->getOperand(1)->getName()+".not"); 1986 if (Op0I->getOpcode() == Instruction::And) 1987 return BinaryOperator::CreateOr(Op0NotVal, NotY); 1988 return BinaryOperator::CreateAnd(Op0NotVal, NotY); 1989 } 1990 1991 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law 1992 // ~(X | Y) === (~X & ~Y) - De Morgan's Law 1993 if (isFreeToInvert(Op0I->getOperand(0)) && 1994 isFreeToInvert(Op0I->getOperand(1))) { 1995 Value *NotX = 1996 Builder->CreateNot(Op0I->getOperand(0), "notlhs"); 1997 Value *NotY = 1998 Builder->CreateNot(Op0I->getOperand(1), "notrhs"); 1999 if (Op0I->getOpcode() == Instruction::And) 2000 return BinaryOperator::CreateOr(NotX, NotY); 2001 return BinaryOperator::CreateAnd(NotX, NotY); 2002 } 2003 2004 } else if (Op0I->getOpcode() == Instruction::AShr) { 2005 // ~(~X >>s Y) --> (X >>s Y) 2006 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) 2007 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1)); 2008 } 2009 } 2010 } 2011 2012 2013 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) { 2014 if (RHS->isOne() && Op0->hasOneUse()) 2015 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B 2016 if (CmpInst *CI = dyn_cast<CmpInst>(Op0)) 2017 return CmpInst::Create(CI->getOpcode(), 2018 CI->getInversePredicate(), 2019 CI->getOperand(0), CI->getOperand(1)); 2020 2021 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp). 2022 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { 2023 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) { 2024 if (CI->hasOneUse() && Op0C->hasOneUse()) { 2025 Instruction::CastOps Opcode = Op0C->getOpcode(); 2026 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) && 2027 (RHS == ConstantExpr::getCast(Opcode, 2028 ConstantInt::getTrue(I.getContext()), 2029 Op0C->getDestTy()))) { 2030 CI->setPredicate(CI->getInversePredicate()); 2031 return CastInst::Create(Opcode, CI, Op0C->getType()); 2032 } 2033 } 2034 } 2035 } 2036 2037 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) { 2038 // ~(c-X) == X-c-1 == X+(-c-1) 2039 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue()) 2040 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) { 2041 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C); 2042 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C, 2043 ConstantInt::get(I.getType(), 1)); 2044 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS); 2045 } 2046 2047 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) { 2048 if (Op0I->getOpcode() == Instruction::Add) { 2049 // ~(X-c) --> (-c-1)-X 2050 if (RHS->isAllOnesValue()) { 2051 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI); 2052 return BinaryOperator::CreateSub( 2053 ConstantExpr::getSub(NegOp0CI, 2054 ConstantInt::get(I.getType(), 1)), 2055 Op0I->getOperand(0)); 2056 } else if (RHS->getValue().isSignBit()) { 2057 // (X + C) ^ signbit -> (X + C + signbit) 2058 Constant *C = ConstantInt::get(I.getContext(), 2059 RHS->getValue() + Op0CI->getValue()); 2060 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C); 2061 2062 } 2063 } else if (Op0I->getOpcode() == Instruction::Or) { 2064 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0 2065 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) { 2066 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS); 2067 // Anything in both C1 and C2 is known to be zero, remove it from 2068 // NewRHS. 2069 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS); 2070 NewRHS = ConstantExpr::getAnd(NewRHS, 2071 ConstantExpr::getNot(CommonBits)); 2072 Worklist.Add(Op0I); 2073 I.setOperand(0, Op0I->getOperand(0)); 2074 I.setOperand(1, NewRHS); 2075 return &I; 2076 } 2077 } 2078 } 2079 } 2080 2081 // Try to fold constant and into select arguments. 2082 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 2083 if (Instruction *R = FoldOpIntoSelect(I, SI)) 2084 return R; 2085 if (isa<PHINode>(Op0)) 2086 if (Instruction *NV = FoldOpIntoPhi(I)) 2087 return NV; 2088 } 2089 2090 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1 2091 if (X == Op1) 2092 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType())); 2093 2094 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1 2095 if (X == Op0) 2096 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType())); 2097 2098 2099 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1); 2100 if (Op1I) { 2101 Value *A, *B; 2102 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) { 2103 if (A == Op0) { // B^(B|A) == (A|B)^B 2104 Op1I->swapOperands(); 2105 I.swapOperands(); 2106 std::swap(Op0, Op1); 2107 } else if (B == Op0) { // B^(A|B) == (A|B)^B 2108 I.swapOperands(); // Simplified below. 2109 std::swap(Op0, Op1); 2110 } 2111 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) { 2112 return ReplaceInstUsesWith(I, B); // A^(A^B) == B 2113 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) { 2114 return ReplaceInstUsesWith(I, A); // A^(B^A) == B 2115 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && 2116 Op1I->hasOneUse()){ 2117 if (A == Op0) { // A^(A&B) -> A^(B&A) 2118 Op1I->swapOperands(); 2119 std::swap(A, B); 2120 } 2121 if (B == Op0) { // A^(B&A) -> (B&A)^A 2122 I.swapOperands(); // Simplified below. 2123 std::swap(Op0, Op1); 2124 } 2125 } 2126 } 2127 2128 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0); 2129 if (Op0I) { 2130 Value *A, *B; 2131 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && 2132 Op0I->hasOneUse()) { 2133 if (A == Op1) // (B|A)^B == (A|B)^B 2134 std::swap(A, B); 2135 if (B == Op1) // (A|B)^B == A & ~B 2136 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp")); 2137 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) { 2138 return ReplaceInstUsesWith(I, B); // (A^B)^A == B 2139 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) { 2140 return ReplaceInstUsesWith(I, A); // (B^A)^A == B 2141 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && 2142 Op0I->hasOneUse()){ 2143 if (A == Op1) // (A&B)^A -> (B&A)^A 2144 std::swap(A, B); 2145 if (B == Op1 && // (B&A)^A == ~B & A 2146 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C 2147 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1); 2148 } 2149 } 2150 } 2151 2152 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts. 2153 if (Op0I && Op1I && Op0I->isShift() && 2154 Op0I->getOpcode() == Op1I->getOpcode() && 2155 Op0I->getOperand(1) == Op1I->getOperand(1) && 2156 (Op1I->hasOneUse() || Op1I->hasOneUse())) { 2157 Value *NewOp = 2158 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0), 2159 Op0I->getName()); 2160 return BinaryOperator::Create(Op1I->getOpcode(), NewOp, 2161 Op1I->getOperand(1)); 2162 } 2163 2164 if (Op0I && Op1I) { 2165 Value *A, *B, *C, *D; 2166 // (A & B)^(A | B) -> A ^ B 2167 if (match(Op0I, m_And(m_Value(A), m_Value(B))) && 2168 match(Op1I, m_Or(m_Value(C), m_Value(D)))) { 2169 if ((A == C && B == D) || (A == D && B == C)) 2170 return BinaryOperator::CreateXor(A, B); 2171 } 2172 // (A | B)^(A & B) -> A ^ B 2173 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && 2174 match(Op1I, m_And(m_Value(C), m_Value(D)))) { 2175 if ((A == C && B == D) || (A == D && B == C)) 2176 return BinaryOperator::CreateXor(A, B); 2177 } 2178 2179 // (A & B)^(C & D) 2180 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) && 2181 match(Op0I, m_And(m_Value(A), m_Value(B))) && 2182 match(Op1I, m_And(m_Value(C), m_Value(D)))) { 2183 // (X & Y)^(X & Y) -> (Y^Z) & X 2184 Value *X = 0, *Y = 0, *Z = 0; 2185 if (A == C) 2186 X = A, Y = B, Z = D; 2187 else if (A == D) 2188 X = A, Y = B, Z = C; 2189 else if (B == C) 2190 X = B, Y = A, Z = D; 2191 else if (B == D) 2192 X = B, Y = A, Z = C; 2193 2194 if (X) { 2195 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName()); 2196 return BinaryOperator::CreateAnd(NewOp, X); 2197 } 2198 } 2199 } 2200 2201 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B) 2202 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) 2203 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0))) 2204 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) { 2205 if (LHS->getOperand(0) == RHS->getOperand(1) && 2206 LHS->getOperand(1) == RHS->getOperand(0)) 2207 LHS->swapOperands(); 2208 if (LHS->getOperand(0) == RHS->getOperand(0) && 2209 LHS->getOperand(1) == RHS->getOperand(1)) { 2210 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 2211 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS); 2212 bool isSigned = LHS->isSigned() || RHS->isSigned(); 2213 return ReplaceInstUsesWith(I, 2214 getICmpValue(isSigned, Code, Op0, Op1, Builder)); 2215 } 2216 } 2217 2218 // fold (xor (cast A), (cast B)) -> (cast (xor A, B)) 2219 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { 2220 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) 2221 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind? 2222 const Type *SrcTy = Op0C->getOperand(0)->getType(); 2223 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() && 2224 // Only do this if the casts both really cause code to be generated. 2225 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0), 2226 I.getType()) && 2227 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0), 2228 I.getType())) { 2229 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0), 2230 Op1C->getOperand(0), I.getName()); 2231 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); 2232 } 2233 } 2234 } 2235 2236 return Changed ? &I : 0; 2237} 2238