1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===// 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 routines for folding instructions into simpler forms 11 // that do not require creating new instructions. This does constant folding 12 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either 13 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value 14 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been 15 // simplified: This is usually true and assuming it simplifies the logic (if 16 // they have not been simplified then results are correct but maybe suboptimal). 17 // 18 //===----------------------------------------------------------------------===// 19 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/ADT/SetVector.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/Analysis/AliasAnalysis.h" 24 #include "llvm/Analysis/CaptureTracking.h" 25 #include "llvm/Analysis/ConstantFolding.h" 26 #include "llvm/Analysis/MemoryBuiltins.h" 27 #include "llvm/Analysis/ValueTracking.h" 28 #include "llvm/Analysis/VectorUtils.h" 29 #include "llvm/IR/ConstantRange.h" 30 #include "llvm/IR/DataLayout.h" 31 #include "llvm/IR/Dominators.h" 32 #include "llvm/IR/GetElementPtrTypeIterator.h" 33 #include "llvm/IR/GlobalAlias.h" 34 #include "llvm/IR/Operator.h" 35 #include "llvm/IR/PatternMatch.h" 36 #include "llvm/IR/ValueHandle.h" 37 #include <algorithm> 38 using namespace llvm; 39 using namespace llvm::PatternMatch; 40 41 #define DEBUG_TYPE "instsimplify" 42 43 enum { RecursionLimit = 3 }; 44 45 STATISTIC(NumExpand, "Number of expansions"); 46 STATISTIC(NumReassoc, "Number of reassociations"); 47 48 namespace { 49 struct Query { 50 const DataLayout &DL; 51 const TargetLibraryInfo *TLI; 52 const DominatorTree *DT; 53 AssumptionCache *AC; 54 const Instruction *CxtI; 55 56 Query(const DataLayout &DL, const TargetLibraryInfo *tli, 57 const DominatorTree *dt, AssumptionCache *ac = nullptr, 58 const Instruction *cxti = nullptr) 59 : DL(DL), TLI(tli), DT(dt), AC(ac), CxtI(cxti) {} 60 }; 61 } // end anonymous namespace 62 63 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned); 64 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &, 65 unsigned); 66 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &, 67 const Query &, unsigned); 68 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &, 69 unsigned); 70 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 71 const Query &Q, unsigned MaxRecurse); 72 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned); 73 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned); 74 static Value *SimplifyCastInst(unsigned, Value *, Type *, 75 const Query &, unsigned); 76 77 /// For a boolean type, or a vector of boolean type, return false, or 78 /// a vector with every element false, as appropriate for the type. 79 static Constant *getFalse(Type *Ty) { 80 assert(Ty->getScalarType()->isIntegerTy(1) && 81 "Expected i1 type or a vector of i1!"); 82 return Constant::getNullValue(Ty); 83 } 84 85 /// For a boolean type, or a vector of boolean type, return true, or 86 /// a vector with every element true, as appropriate for the type. 87 static Constant *getTrue(Type *Ty) { 88 assert(Ty->getScalarType()->isIntegerTy(1) && 89 "Expected i1 type or a vector of i1!"); 90 return Constant::getAllOnesValue(Ty); 91 } 92 93 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"? 94 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS, 95 Value *RHS) { 96 CmpInst *Cmp = dyn_cast<CmpInst>(V); 97 if (!Cmp) 98 return false; 99 CmpInst::Predicate CPred = Cmp->getPredicate(); 100 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1); 101 if (CPred == Pred && CLHS == LHS && CRHS == RHS) 102 return true; 103 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS && 104 CRHS == LHS; 105 } 106 107 /// Does the given value dominate the specified phi node? 108 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) { 109 Instruction *I = dyn_cast<Instruction>(V); 110 if (!I) 111 // Arguments and constants dominate all instructions. 112 return true; 113 114 // If we are processing instructions (and/or basic blocks) that have not been 115 // fully added to a function, the parent nodes may still be null. Simply 116 // return the conservative answer in these cases. 117 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent()) 118 return false; 119 120 // If we have a DominatorTree then do a precise test. 121 if (DT) { 122 if (!DT->isReachableFromEntry(P->getParent())) 123 return true; 124 if (!DT->isReachableFromEntry(I->getParent())) 125 return false; 126 return DT->dominates(I, P); 127 } 128 129 // Otherwise, if the instruction is in the entry block and is not an invoke, 130 // then it obviously dominates all phi nodes. 131 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() && 132 !isa<InvokeInst>(I)) 133 return true; 134 135 return false; 136 } 137 138 /// Simplify "A op (B op' C)" by distributing op over op', turning it into 139 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is 140 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS. 141 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)". 142 /// Returns the simplified value, or null if no simplification was performed. 143 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS, 144 unsigned OpcToExpand, const Query &Q, 145 unsigned MaxRecurse) { 146 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand; 147 // Recursion is always used, so bail out at once if we already hit the limit. 148 if (!MaxRecurse--) 149 return nullptr; 150 151 // Check whether the expression has the form "(A op' B) op C". 152 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS)) 153 if (Op0->getOpcode() == OpcodeToExpand) { 154 // It does! Try turning it into "(A op C) op' (B op C)". 155 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 156 // Do "A op C" and "B op C" both simplify? 157 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) 158 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 159 // They do! Return "L op' R" if it simplifies or is already available. 160 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 161 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand) 162 && L == B && R == A)) { 163 ++NumExpand; 164 return LHS; 165 } 166 // Otherwise return "L op' R" if it simplifies. 167 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 168 ++NumExpand; 169 return V; 170 } 171 } 172 } 173 174 // Check whether the expression has the form "A op (B op' C)". 175 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS)) 176 if (Op1->getOpcode() == OpcodeToExpand) { 177 // It does! Try turning it into "(A op B) op' (A op C)". 178 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 179 // Do "A op B" and "A op C" both simplify? 180 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) 181 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) { 182 // They do! Return "L op' R" if it simplifies or is already available. 183 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 184 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand) 185 && L == C && R == B)) { 186 ++NumExpand; 187 return RHS; 188 } 189 // Otherwise return "L op' R" if it simplifies. 190 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 191 ++NumExpand; 192 return V; 193 } 194 } 195 } 196 197 return nullptr; 198 } 199 200 /// Generic simplifications for associative binary operations. 201 /// Returns the simpler value, or null if none was found. 202 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS, 203 const Query &Q, unsigned MaxRecurse) { 204 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc; 205 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!"); 206 207 // Recursion is always used, so bail out at once if we already hit the limit. 208 if (!MaxRecurse--) 209 return nullptr; 210 211 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 212 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 213 214 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely. 215 if (Op0 && Op0->getOpcode() == Opcode) { 216 Value *A = Op0->getOperand(0); 217 Value *B = Op0->getOperand(1); 218 Value *C = RHS; 219 220 // Does "B op C" simplify? 221 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 222 // It does! Return "A op V" if it simplifies or is already available. 223 // If V equals B then "A op V" is just the LHS. 224 if (V == B) return LHS; 225 // Otherwise return "A op V" if it simplifies. 226 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) { 227 ++NumReassoc; 228 return W; 229 } 230 } 231 } 232 233 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely. 234 if (Op1 && Op1->getOpcode() == Opcode) { 235 Value *A = LHS; 236 Value *B = Op1->getOperand(0); 237 Value *C = Op1->getOperand(1); 238 239 // Does "A op B" simplify? 240 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) { 241 // It does! Return "V op C" if it simplifies or is already available. 242 // If V equals B then "V op C" is just the RHS. 243 if (V == B) return RHS; 244 // Otherwise return "V op C" if it simplifies. 245 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) { 246 ++NumReassoc; 247 return W; 248 } 249 } 250 } 251 252 // The remaining transforms require commutativity as well as associativity. 253 if (!Instruction::isCommutative(Opcode)) 254 return nullptr; 255 256 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely. 257 if (Op0 && Op0->getOpcode() == Opcode) { 258 Value *A = Op0->getOperand(0); 259 Value *B = Op0->getOperand(1); 260 Value *C = RHS; 261 262 // Does "C op A" simplify? 263 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 264 // It does! Return "V op B" if it simplifies or is already available. 265 // If V equals A then "V op B" is just the LHS. 266 if (V == A) return LHS; 267 // Otherwise return "V op B" if it simplifies. 268 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) { 269 ++NumReassoc; 270 return W; 271 } 272 } 273 } 274 275 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely. 276 if (Op1 && Op1->getOpcode() == Opcode) { 277 Value *A = LHS; 278 Value *B = Op1->getOperand(0); 279 Value *C = Op1->getOperand(1); 280 281 // Does "C op A" simplify? 282 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 283 // It does! Return "B op V" if it simplifies or is already available. 284 // If V equals C then "B op V" is just the RHS. 285 if (V == C) return RHS; 286 // Otherwise return "B op V" if it simplifies. 287 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) { 288 ++NumReassoc; 289 return W; 290 } 291 } 292 } 293 294 return nullptr; 295 } 296 297 /// In the case of a binary operation with a select instruction as an operand, 298 /// try to simplify the binop by seeing whether evaluating it on both branches 299 /// of the select results in the same value. Returns the common value if so, 300 /// otherwise returns null. 301 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS, 302 const Query &Q, unsigned MaxRecurse) { 303 // Recursion is always used, so bail out at once if we already hit the limit. 304 if (!MaxRecurse--) 305 return nullptr; 306 307 SelectInst *SI; 308 if (isa<SelectInst>(LHS)) { 309 SI = cast<SelectInst>(LHS); 310 } else { 311 assert(isa<SelectInst>(RHS) && "No select instruction operand!"); 312 SI = cast<SelectInst>(RHS); 313 } 314 315 // Evaluate the BinOp on the true and false branches of the select. 316 Value *TV; 317 Value *FV; 318 if (SI == LHS) { 319 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse); 320 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse); 321 } else { 322 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse); 323 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse); 324 } 325 326 // If they simplified to the same value, then return the common value. 327 // If they both failed to simplify then return null. 328 if (TV == FV) 329 return TV; 330 331 // If one branch simplified to undef, return the other one. 332 if (TV && isa<UndefValue>(TV)) 333 return FV; 334 if (FV && isa<UndefValue>(FV)) 335 return TV; 336 337 // If applying the operation did not change the true and false select values, 338 // then the result of the binop is the select itself. 339 if (TV == SI->getTrueValue() && FV == SI->getFalseValue()) 340 return SI; 341 342 // If one branch simplified and the other did not, and the simplified 343 // value is equal to the unsimplified one, return the simplified value. 344 // For example, select (cond, X, X & Z) & Z -> X & Z. 345 if ((FV && !TV) || (TV && !FV)) { 346 // Check that the simplified value has the form "X op Y" where "op" is the 347 // same as the original operation. 348 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV); 349 if (Simplified && Simplified->getOpcode() == Opcode) { 350 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS". 351 // We already know that "op" is the same as for the simplified value. See 352 // if the operands match too. If so, return the simplified value. 353 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue(); 354 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS; 355 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch; 356 if (Simplified->getOperand(0) == UnsimplifiedLHS && 357 Simplified->getOperand(1) == UnsimplifiedRHS) 358 return Simplified; 359 if (Simplified->isCommutative() && 360 Simplified->getOperand(1) == UnsimplifiedLHS && 361 Simplified->getOperand(0) == UnsimplifiedRHS) 362 return Simplified; 363 } 364 } 365 366 return nullptr; 367 } 368 369 /// In the case of a comparison with a select instruction, try to simplify the 370 /// comparison by seeing whether both branches of the select result in the same 371 /// value. Returns the common value if so, otherwise returns null. 372 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS, 373 Value *RHS, const Query &Q, 374 unsigned MaxRecurse) { 375 // Recursion is always used, so bail out at once if we already hit the limit. 376 if (!MaxRecurse--) 377 return nullptr; 378 379 // Make sure the select is on the LHS. 380 if (!isa<SelectInst>(LHS)) { 381 std::swap(LHS, RHS); 382 Pred = CmpInst::getSwappedPredicate(Pred); 383 } 384 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!"); 385 SelectInst *SI = cast<SelectInst>(LHS); 386 Value *Cond = SI->getCondition(); 387 Value *TV = SI->getTrueValue(); 388 Value *FV = SI->getFalseValue(); 389 390 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it. 391 // Does "cmp TV, RHS" simplify? 392 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse); 393 if (TCmp == Cond) { 394 // It not only simplified, it simplified to the select condition. Replace 395 // it with 'true'. 396 TCmp = getTrue(Cond->getType()); 397 } else if (!TCmp) { 398 // It didn't simplify. However if "cmp TV, RHS" is equal to the select 399 // condition then we can replace it with 'true'. Otherwise give up. 400 if (!isSameCompare(Cond, Pred, TV, RHS)) 401 return nullptr; 402 TCmp = getTrue(Cond->getType()); 403 } 404 405 // Does "cmp FV, RHS" simplify? 406 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse); 407 if (FCmp == Cond) { 408 // It not only simplified, it simplified to the select condition. Replace 409 // it with 'false'. 410 FCmp = getFalse(Cond->getType()); 411 } else if (!FCmp) { 412 // It didn't simplify. However if "cmp FV, RHS" is equal to the select 413 // condition then we can replace it with 'false'. Otherwise give up. 414 if (!isSameCompare(Cond, Pred, FV, RHS)) 415 return nullptr; 416 FCmp = getFalse(Cond->getType()); 417 } 418 419 // If both sides simplified to the same value, then use it as the result of 420 // the original comparison. 421 if (TCmp == FCmp) 422 return TCmp; 423 424 // The remaining cases only make sense if the select condition has the same 425 // type as the result of the comparison, so bail out if this is not so. 426 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy()) 427 return nullptr; 428 // If the false value simplified to false, then the result of the compare 429 // is equal to "Cond && TCmp". This also catches the case when the false 430 // value simplified to false and the true value to true, returning "Cond". 431 if (match(FCmp, m_Zero())) 432 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse)) 433 return V; 434 // If the true value simplified to true, then the result of the compare 435 // is equal to "Cond || FCmp". 436 if (match(TCmp, m_One())) 437 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse)) 438 return V; 439 // Finally, if the false value simplified to true and the true value to 440 // false, then the result of the compare is equal to "!Cond". 441 if (match(FCmp, m_One()) && match(TCmp, m_Zero())) 442 if (Value *V = 443 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()), 444 Q, MaxRecurse)) 445 return V; 446 447 return nullptr; 448 } 449 450 /// In the case of a binary operation with an operand that is a PHI instruction, 451 /// try to simplify the binop by seeing whether evaluating it on the incoming 452 /// phi values yields the same result for every value. If so returns the common 453 /// value, otherwise returns null. 454 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS, 455 const Query &Q, unsigned MaxRecurse) { 456 // Recursion is always used, so bail out at once if we already hit the limit. 457 if (!MaxRecurse--) 458 return nullptr; 459 460 PHINode *PI; 461 if (isa<PHINode>(LHS)) { 462 PI = cast<PHINode>(LHS); 463 // Bail out if RHS and the phi may be mutually interdependent due to a loop. 464 if (!ValueDominatesPHI(RHS, PI, Q.DT)) 465 return nullptr; 466 } else { 467 assert(isa<PHINode>(RHS) && "No PHI instruction operand!"); 468 PI = cast<PHINode>(RHS); 469 // Bail out if LHS and the phi may be mutually interdependent due to a loop. 470 if (!ValueDominatesPHI(LHS, PI, Q.DT)) 471 return nullptr; 472 } 473 474 // Evaluate the BinOp on the incoming phi values. 475 Value *CommonValue = nullptr; 476 for (Value *Incoming : PI->incoming_values()) { 477 // If the incoming value is the phi node itself, it can safely be skipped. 478 if (Incoming == PI) continue; 479 Value *V = PI == LHS ? 480 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) : 481 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse); 482 // If the operation failed to simplify, or simplified to a different value 483 // to previously, then give up. 484 if (!V || (CommonValue && V != CommonValue)) 485 return nullptr; 486 CommonValue = V; 487 } 488 489 return CommonValue; 490 } 491 492 /// In the case of a comparison with a PHI instruction, try to simplify the 493 /// comparison by seeing whether comparing with all of the incoming phi values 494 /// yields the same result every time. If so returns the common result, 495 /// otherwise returns null. 496 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS, 497 const Query &Q, unsigned MaxRecurse) { 498 // Recursion is always used, so bail out at once if we already hit the limit. 499 if (!MaxRecurse--) 500 return nullptr; 501 502 // Make sure the phi is on the LHS. 503 if (!isa<PHINode>(LHS)) { 504 std::swap(LHS, RHS); 505 Pred = CmpInst::getSwappedPredicate(Pred); 506 } 507 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!"); 508 PHINode *PI = cast<PHINode>(LHS); 509 510 // Bail out if RHS and the phi may be mutually interdependent due to a loop. 511 if (!ValueDominatesPHI(RHS, PI, Q.DT)) 512 return nullptr; 513 514 // Evaluate the BinOp on the incoming phi values. 515 Value *CommonValue = nullptr; 516 for (Value *Incoming : PI->incoming_values()) { 517 // If the incoming value is the phi node itself, it can safely be skipped. 518 if (Incoming == PI) continue; 519 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse); 520 // If the operation failed to simplify, or simplified to a different value 521 // to previously, then give up. 522 if (!V || (CommonValue && V != CommonValue)) 523 return nullptr; 524 CommonValue = V; 525 } 526 527 return CommonValue; 528 } 529 530 /// Given operands for an Add, see if we can fold the result. 531 /// If not, this returns null. 532 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 533 const Query &Q, unsigned MaxRecurse) { 534 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 535 if (Constant *CRHS = dyn_cast<Constant>(Op1)) 536 return ConstantFoldBinaryOpOperands(Instruction::Add, CLHS, CRHS, Q.DL); 537 538 // Canonicalize the constant to the RHS. 539 std::swap(Op0, Op1); 540 } 541 542 // X + undef -> undef 543 if (match(Op1, m_Undef())) 544 return Op1; 545 546 // X + 0 -> X 547 if (match(Op1, m_Zero())) 548 return Op0; 549 550 // X + (Y - X) -> Y 551 // (Y - X) + X -> Y 552 // Eg: X + -X -> 0 553 Value *Y = nullptr; 554 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) || 555 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1)))) 556 return Y; 557 558 // X + ~X -> -1 since ~X = -X-1 559 if (match(Op0, m_Not(m_Specific(Op1))) || 560 match(Op1, m_Not(m_Specific(Op0)))) 561 return Constant::getAllOnesValue(Op0->getType()); 562 563 /// i1 add -> xor. 564 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 565 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 566 return V; 567 568 // Try some generic simplifications for associative operations. 569 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q, 570 MaxRecurse)) 571 return V; 572 573 // Threading Add over selects and phi nodes is pointless, so don't bother. 574 // Threading over the select in "A + select(cond, B, C)" means evaluating 575 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and 576 // only if B and C are equal. If B and C are equal then (since we assume 577 // that operands have already been simplified) "select(cond, B, C)" should 578 // have been simplified to the common value of B and C already. Analysing 579 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly 580 // for threading over phi nodes. 581 582 return nullptr; 583 } 584 585 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 586 const DataLayout &DL, const TargetLibraryInfo *TLI, 587 const DominatorTree *DT, AssumptionCache *AC, 588 const Instruction *CxtI) { 589 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 590 RecursionLimit); 591 } 592 593 /// \brief Compute the base pointer and cumulative constant offsets for V. 594 /// 595 /// This strips all constant offsets off of V, leaving it the base pointer, and 596 /// accumulates the total constant offset applied in the returned constant. It 597 /// returns 0 if V is not a pointer, and returns the constant '0' if there are 598 /// no constant offsets applied. 599 /// 600 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't 601 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc. 602 /// folding. 603 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V, 604 bool AllowNonInbounds = false) { 605 assert(V->getType()->getScalarType()->isPointerTy()); 606 607 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType(); 608 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth()); 609 610 // Even though we don't look through PHI nodes, we could be called on an 611 // instruction in an unreachable block, which may be on a cycle. 612 SmallPtrSet<Value *, 4> Visited; 613 Visited.insert(V); 614 do { 615 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 616 if ((!AllowNonInbounds && !GEP->isInBounds()) || 617 !GEP->accumulateConstantOffset(DL, Offset)) 618 break; 619 V = GEP->getPointerOperand(); 620 } else if (Operator::getOpcode(V) == Instruction::BitCast) { 621 V = cast<Operator>(V)->getOperand(0); 622 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 623 if (GA->isInterposable()) 624 break; 625 V = GA->getAliasee(); 626 } else { 627 if (auto CS = CallSite(V)) 628 if (Value *RV = CS.getReturnedArgOperand()) { 629 V = RV; 630 continue; 631 } 632 break; 633 } 634 assert(V->getType()->getScalarType()->isPointerTy() && 635 "Unexpected operand type!"); 636 } while (Visited.insert(V).second); 637 638 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset); 639 if (V->getType()->isVectorTy()) 640 return ConstantVector::getSplat(V->getType()->getVectorNumElements(), 641 OffsetIntPtr); 642 return OffsetIntPtr; 643 } 644 645 /// \brief Compute the constant difference between two pointer values. 646 /// If the difference is not a constant, returns zero. 647 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS, 648 Value *RHS) { 649 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 650 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 651 652 // If LHS and RHS are not related via constant offsets to the same base 653 // value, there is nothing we can do here. 654 if (LHS != RHS) 655 return nullptr; 656 657 // Otherwise, the difference of LHS - RHS can be computed as: 658 // LHS - RHS 659 // = (LHSOffset + Base) - (RHSOffset + Base) 660 // = LHSOffset - RHSOffset 661 return ConstantExpr::getSub(LHSOffset, RHSOffset); 662 } 663 664 /// Given operands for a Sub, see if we can fold the result. 665 /// If not, this returns null. 666 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 667 const Query &Q, unsigned MaxRecurse) { 668 if (Constant *CLHS = dyn_cast<Constant>(Op0)) 669 if (Constant *CRHS = dyn_cast<Constant>(Op1)) 670 return ConstantFoldBinaryOpOperands(Instruction::Sub, CLHS, CRHS, Q.DL); 671 672 // X - undef -> undef 673 // undef - X -> undef 674 if (match(Op0, m_Undef()) || match(Op1, m_Undef())) 675 return UndefValue::get(Op0->getType()); 676 677 // X - 0 -> X 678 if (match(Op1, m_Zero())) 679 return Op0; 680 681 // X - X -> 0 682 if (Op0 == Op1) 683 return Constant::getNullValue(Op0->getType()); 684 685 // Is this a negation? 686 if (match(Op0, m_Zero())) { 687 // 0 - X -> 0 if the sub is NUW. 688 if (isNUW) 689 return Op0; 690 691 unsigned BitWidth = Op1->getType()->getScalarSizeInBits(); 692 APInt KnownZero(BitWidth, 0); 693 APInt KnownOne(BitWidth, 0); 694 computeKnownBits(Op1, KnownZero, KnownOne, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 695 if (KnownZero == ~APInt::getSignBit(BitWidth)) { 696 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then 697 // Op1 must be 0 because negating the minimum signed value is undefined. 698 if (isNSW) 699 return Op0; 700 701 // 0 - X -> X if X is 0 or the minimum signed value. 702 return Op1; 703 } 704 } 705 706 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies. 707 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X 708 Value *X = nullptr, *Y = nullptr, *Z = Op1; 709 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z 710 // See if "V === Y - Z" simplifies. 711 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1)) 712 // It does! Now see if "X + V" simplifies. 713 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) { 714 // It does, we successfully reassociated! 715 ++NumReassoc; 716 return W; 717 } 718 // See if "V === X - Z" simplifies. 719 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 720 // It does! Now see if "Y + V" simplifies. 721 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) { 722 // It does, we successfully reassociated! 723 ++NumReassoc; 724 return W; 725 } 726 } 727 728 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies. 729 // For example, X - (X + 1) -> -1 730 X = Op0; 731 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z) 732 // See if "V === X - Y" simplifies. 733 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 734 // It does! Now see if "V - Z" simplifies. 735 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) { 736 // It does, we successfully reassociated! 737 ++NumReassoc; 738 return W; 739 } 740 // See if "V === X - Z" simplifies. 741 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 742 // It does! Now see if "V - Y" simplifies. 743 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) { 744 // It does, we successfully reassociated! 745 ++NumReassoc; 746 return W; 747 } 748 } 749 750 // Z - (X - Y) -> (Z - X) + Y if everything simplifies. 751 // For example, X - (X - Y) -> Y. 752 Z = Op0; 753 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y) 754 // See if "V === Z - X" simplifies. 755 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1)) 756 // It does! Now see if "V + Y" simplifies. 757 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) { 758 // It does, we successfully reassociated! 759 ++NumReassoc; 760 return W; 761 } 762 763 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies. 764 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) && 765 match(Op1, m_Trunc(m_Value(Y)))) 766 if (X->getType() == Y->getType()) 767 // See if "V === X - Y" simplifies. 768 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 769 // It does! Now see if "trunc V" simplifies. 770 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(), 771 Q, MaxRecurse - 1)) 772 // It does, return the simplified "trunc V". 773 return W; 774 775 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...). 776 if (match(Op0, m_PtrToInt(m_Value(X))) && 777 match(Op1, m_PtrToInt(m_Value(Y)))) 778 if (Constant *Result = computePointerDifference(Q.DL, X, Y)) 779 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true); 780 781 // i1 sub -> xor. 782 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 783 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 784 return V; 785 786 // Threading Sub over selects and phi nodes is pointless, so don't bother. 787 // Threading over the select in "A - select(cond, B, C)" means evaluating 788 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and 789 // only if B and C are equal. If B and C are equal then (since we assume 790 // that operands have already been simplified) "select(cond, B, C)" should 791 // have been simplified to the common value of B and C already. Analysing 792 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly 793 // for threading over phi nodes. 794 795 return nullptr; 796 } 797 798 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 799 const DataLayout &DL, const TargetLibraryInfo *TLI, 800 const DominatorTree *DT, AssumptionCache *AC, 801 const Instruction *CxtI) { 802 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 803 RecursionLimit); 804 } 805 806 /// Given operands for an FAdd, see if we can fold the result. If not, this 807 /// returns null. 808 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 809 const Query &Q, unsigned MaxRecurse) { 810 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 811 if (Constant *CRHS = dyn_cast<Constant>(Op1)) 812 return ConstantFoldBinaryOpOperands(Instruction::FAdd, CLHS, CRHS, Q.DL); 813 814 // Canonicalize the constant to the RHS. 815 std::swap(Op0, Op1); 816 } 817 818 // fadd X, -0 ==> X 819 if (match(Op1, m_NegZero())) 820 return Op0; 821 822 // fadd X, 0 ==> X, when we know X is not -0 823 if (match(Op1, m_Zero()) && 824 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI))) 825 return Op0; 826 827 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0 828 // where nnan and ninf have to occur at least once somewhere in this 829 // expression 830 Value *SubOp = nullptr; 831 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0)))) 832 SubOp = Op1; 833 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1)))) 834 SubOp = Op0; 835 if (SubOp) { 836 Instruction *FSub = cast<Instruction>(SubOp); 837 if ((FMF.noNaNs() || FSub->hasNoNaNs()) && 838 (FMF.noInfs() || FSub->hasNoInfs())) 839 return Constant::getNullValue(Op0->getType()); 840 } 841 842 return nullptr; 843 } 844 845 /// Given operands for an FSub, see if we can fold the result. If not, this 846 /// returns null. 847 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 848 const Query &Q, unsigned MaxRecurse) { 849 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 850 if (Constant *CRHS = dyn_cast<Constant>(Op1)) 851 return ConstantFoldBinaryOpOperands(Instruction::FSub, CLHS, CRHS, Q.DL); 852 } 853 854 // fsub X, 0 ==> X 855 if (match(Op1, m_Zero())) 856 return Op0; 857 858 // fsub X, -0 ==> X, when we know X is not -0 859 if (match(Op1, m_NegZero()) && 860 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI))) 861 return Op0; 862 863 // fsub -0.0, (fsub -0.0, X) ==> X 864 Value *X; 865 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X)))) 866 return X; 867 868 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored. 869 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) && 870 match(Op1, m_FSub(m_AnyZero(), m_Value(X)))) 871 return X; 872 873 // fsub nnan x, x ==> 0.0 874 if (FMF.noNaNs() && Op0 == Op1) 875 return Constant::getNullValue(Op0->getType()); 876 877 return nullptr; 878 } 879 880 /// Given the operands for an FMul, see if we can fold the result 881 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, 882 FastMathFlags FMF, 883 const Query &Q, 884 unsigned MaxRecurse) { 885 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 886 if (Constant *CRHS = dyn_cast<Constant>(Op1)) 887 return ConstantFoldBinaryOpOperands(Instruction::FMul, CLHS, CRHS, Q.DL); 888 889 // Canonicalize the constant to the RHS. 890 std::swap(Op0, Op1); 891 } 892 893 // fmul X, 1.0 ==> X 894 if (match(Op1, m_FPOne())) 895 return Op0; 896 897 // fmul nnan nsz X, 0 ==> 0 898 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero())) 899 return Op1; 900 901 return nullptr; 902 } 903 904 /// Given operands for a Mul, see if we can fold the result. 905 /// If not, this returns null. 906 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q, 907 unsigned MaxRecurse) { 908 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 909 if (Constant *CRHS = dyn_cast<Constant>(Op1)) 910 return ConstantFoldBinaryOpOperands(Instruction::Mul, CLHS, CRHS, Q.DL); 911 912 // Canonicalize the constant to the RHS. 913 std::swap(Op0, Op1); 914 } 915 916 // X * undef -> 0 917 if (match(Op1, m_Undef())) 918 return Constant::getNullValue(Op0->getType()); 919 920 // X * 0 -> 0 921 if (match(Op1, m_Zero())) 922 return Op1; 923 924 // X * 1 -> X 925 if (match(Op1, m_One())) 926 return Op0; 927 928 // (X / Y) * Y -> X if the division is exact. 929 Value *X = nullptr; 930 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y 931 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y) 932 return X; 933 934 // i1 mul -> and. 935 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 936 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1)) 937 return V; 938 939 // Try some generic simplifications for associative operations. 940 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q, 941 MaxRecurse)) 942 return V; 943 944 // Mul distributes over Add. Try some generic simplifications based on this. 945 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add, 946 Q, MaxRecurse)) 947 return V; 948 949 // If the operation is with the result of a select instruction, check whether 950 // operating on either branch of the select always yields the same value. 951 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 952 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q, 953 MaxRecurse)) 954 return V; 955 956 // If the operation is with the result of a phi instruction, check whether 957 // operating on all incoming values of the phi always yields the same value. 958 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 959 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q, 960 MaxRecurse)) 961 return V; 962 963 return nullptr; 964 } 965 966 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 967 const DataLayout &DL, 968 const TargetLibraryInfo *TLI, 969 const DominatorTree *DT, AssumptionCache *AC, 970 const Instruction *CxtI) { 971 return ::SimplifyFAddInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 972 RecursionLimit); 973 } 974 975 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 976 const DataLayout &DL, 977 const TargetLibraryInfo *TLI, 978 const DominatorTree *DT, AssumptionCache *AC, 979 const Instruction *CxtI) { 980 return ::SimplifyFSubInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 981 RecursionLimit); 982 } 983 984 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF, 985 const DataLayout &DL, 986 const TargetLibraryInfo *TLI, 987 const DominatorTree *DT, AssumptionCache *AC, 988 const Instruction *CxtI) { 989 return ::SimplifyFMulInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 990 RecursionLimit); 991 } 992 993 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout &DL, 994 const TargetLibraryInfo *TLI, 995 const DominatorTree *DT, AssumptionCache *AC, 996 const Instruction *CxtI) { 997 return ::SimplifyMulInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 998 RecursionLimit); 999 } 1000 1001 /// Given operands for an SDiv or UDiv, see if we can fold the result. 1002 /// If not, this returns null. 1003 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1004 const Query &Q, unsigned MaxRecurse) { 1005 if (Constant *C0 = dyn_cast<Constant>(Op0)) 1006 if (Constant *C1 = dyn_cast<Constant>(Op1)) 1007 return ConstantFoldBinaryOpOperands(Opcode, C0, C1, Q.DL); 1008 1009 bool isSigned = Opcode == Instruction::SDiv; 1010 1011 // X / undef -> undef 1012 if (match(Op1, m_Undef())) 1013 return Op1; 1014 1015 // X / 0 -> undef, we don't need to preserve faults! 1016 if (match(Op1, m_Zero())) 1017 return UndefValue::get(Op1->getType()); 1018 1019 // undef / X -> 0 1020 if (match(Op0, m_Undef())) 1021 return Constant::getNullValue(Op0->getType()); 1022 1023 // 0 / X -> 0, we don't need to preserve faults! 1024 if (match(Op0, m_Zero())) 1025 return Op0; 1026 1027 // X / 1 -> X 1028 if (match(Op1, m_One())) 1029 return Op0; 1030 1031 if (Op0->getType()->isIntegerTy(1)) 1032 // It can't be division by zero, hence it must be division by one. 1033 return Op0; 1034 1035 // X / X -> 1 1036 if (Op0 == Op1) 1037 return ConstantInt::get(Op0->getType(), 1); 1038 1039 // (X * Y) / Y -> X if the multiplication does not overflow. 1040 Value *X = nullptr, *Y = nullptr; 1041 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) { 1042 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1 1043 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0); 1044 // If the Mul knows it does not overflow, then we are good to go. 1045 if ((isSigned && Mul->hasNoSignedWrap()) || 1046 (!isSigned && Mul->hasNoUnsignedWrap())) 1047 return X; 1048 // If X has the form X = A / Y then X * Y cannot overflow. 1049 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X)) 1050 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y) 1051 return X; 1052 } 1053 1054 // (X rem Y) / Y -> 0 1055 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1056 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1057 return Constant::getNullValue(Op0->getType()); 1058 1059 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow 1060 ConstantInt *C1, *C2; 1061 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) && 1062 match(Op1, m_ConstantInt(C2))) { 1063 bool Overflow; 1064 C1->getValue().umul_ov(C2->getValue(), Overflow); 1065 if (Overflow) 1066 return Constant::getNullValue(Op0->getType()); 1067 } 1068 1069 // If the operation is with the result of a select instruction, check whether 1070 // operating on either branch of the select always yields the same value. 1071 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1072 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1073 return V; 1074 1075 // If the operation is with the result of a phi instruction, check whether 1076 // operating on all incoming values of the phi always yields the same value. 1077 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1078 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1079 return V; 1080 1081 return nullptr; 1082 } 1083 1084 /// Given operands for an SDiv, see if we can fold the result. 1085 /// If not, this returns null. 1086 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q, 1087 unsigned MaxRecurse) { 1088 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse)) 1089 return V; 1090 1091 return nullptr; 1092 } 1093 1094 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout &DL, 1095 const TargetLibraryInfo *TLI, 1096 const DominatorTree *DT, AssumptionCache *AC, 1097 const Instruction *CxtI) { 1098 return ::SimplifySDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1099 RecursionLimit); 1100 } 1101 1102 /// Given operands for a UDiv, see if we can fold the result. 1103 /// If not, this returns null. 1104 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q, 1105 unsigned MaxRecurse) { 1106 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse)) 1107 return V; 1108 1109 return nullptr; 1110 } 1111 1112 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout &DL, 1113 const TargetLibraryInfo *TLI, 1114 const DominatorTree *DT, AssumptionCache *AC, 1115 const Instruction *CxtI) { 1116 return ::SimplifyUDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1117 RecursionLimit); 1118 } 1119 1120 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1121 const Query &Q, unsigned) { 1122 // undef / X -> undef (the undef could be a snan). 1123 if (match(Op0, m_Undef())) 1124 return Op0; 1125 1126 // X / undef -> undef 1127 if (match(Op1, m_Undef())) 1128 return Op1; 1129 1130 // 0 / X -> 0 1131 // Requires that NaNs are off (X could be zero) and signed zeroes are 1132 // ignored (X could be positive or negative, so the output sign is unknown). 1133 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1134 return Op0; 1135 1136 if (FMF.noNaNs()) { 1137 // X / X -> 1.0 is legal when NaNs are ignored. 1138 if (Op0 == Op1) 1139 return ConstantFP::get(Op0->getType(), 1.0); 1140 1141 // -X / X -> -1.0 and 1142 // X / -X -> -1.0 are legal when NaNs are ignored. 1143 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored. 1144 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) && 1145 BinaryOperator::getFNegArgument(Op0) == Op1) || 1146 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) && 1147 BinaryOperator::getFNegArgument(Op1) == Op0)) 1148 return ConstantFP::get(Op0->getType(), -1.0); 1149 } 1150 1151 return nullptr; 1152 } 1153 1154 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1155 const DataLayout &DL, 1156 const TargetLibraryInfo *TLI, 1157 const DominatorTree *DT, AssumptionCache *AC, 1158 const Instruction *CxtI) { 1159 return ::SimplifyFDivInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 1160 RecursionLimit); 1161 } 1162 1163 /// Given operands for an SRem or URem, see if we can fold the result. 1164 /// If not, this returns null. 1165 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1166 const Query &Q, unsigned MaxRecurse) { 1167 if (Constant *C0 = dyn_cast<Constant>(Op0)) 1168 if (Constant *C1 = dyn_cast<Constant>(Op1)) 1169 return ConstantFoldBinaryOpOperands(Opcode, C0, C1, Q.DL); 1170 1171 // X % undef -> undef 1172 if (match(Op1, m_Undef())) 1173 return Op1; 1174 1175 // undef % X -> 0 1176 if (match(Op0, m_Undef())) 1177 return Constant::getNullValue(Op0->getType()); 1178 1179 // 0 % X -> 0, we don't need to preserve faults! 1180 if (match(Op0, m_Zero())) 1181 return Op0; 1182 1183 // X % 0 -> undef, we don't need to preserve faults! 1184 if (match(Op1, m_Zero())) 1185 return UndefValue::get(Op0->getType()); 1186 1187 // X % 1 -> 0 1188 if (match(Op1, m_One())) 1189 return Constant::getNullValue(Op0->getType()); 1190 1191 if (Op0->getType()->isIntegerTy(1)) 1192 // It can't be remainder by zero, hence it must be remainder by one. 1193 return Constant::getNullValue(Op0->getType()); 1194 1195 // X % X -> 0 1196 if (Op0 == Op1) 1197 return Constant::getNullValue(Op0->getType()); 1198 1199 // (X % Y) % Y -> X % Y 1200 if ((Opcode == Instruction::SRem && 1201 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1202 (Opcode == Instruction::URem && 1203 match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1204 return Op0; 1205 1206 // If the operation is with the result of a select instruction, check whether 1207 // operating on either branch of the select always yields the same value. 1208 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1209 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1210 return V; 1211 1212 // If the operation is with the result of a phi instruction, check whether 1213 // operating on all incoming values of the phi always yields the same value. 1214 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1215 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1216 return V; 1217 1218 return nullptr; 1219 } 1220 1221 /// Given operands for an SRem, see if we can fold the result. 1222 /// If not, this returns null. 1223 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q, 1224 unsigned MaxRecurse) { 1225 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse)) 1226 return V; 1227 1228 return nullptr; 1229 } 1230 1231 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout &DL, 1232 const TargetLibraryInfo *TLI, 1233 const DominatorTree *DT, AssumptionCache *AC, 1234 const Instruction *CxtI) { 1235 return ::SimplifySRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1236 RecursionLimit); 1237 } 1238 1239 /// Given operands for a URem, see if we can fold the result. 1240 /// If not, this returns null. 1241 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q, 1242 unsigned MaxRecurse) { 1243 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse)) 1244 return V; 1245 1246 return nullptr; 1247 } 1248 1249 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout &DL, 1250 const TargetLibraryInfo *TLI, 1251 const DominatorTree *DT, AssumptionCache *AC, 1252 const Instruction *CxtI) { 1253 return ::SimplifyURemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1254 RecursionLimit); 1255 } 1256 1257 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1258 const Query &, unsigned) { 1259 // undef % X -> undef (the undef could be a snan). 1260 if (match(Op0, m_Undef())) 1261 return Op0; 1262 1263 // X % undef -> undef 1264 if (match(Op1, m_Undef())) 1265 return Op1; 1266 1267 // 0 % X -> 0 1268 // Requires that NaNs are off (X could be zero) and signed zeroes are 1269 // ignored (X could be positive or negative, so the output sign is unknown). 1270 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1271 return Op0; 1272 1273 return nullptr; 1274 } 1275 1276 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1277 const DataLayout &DL, 1278 const TargetLibraryInfo *TLI, 1279 const DominatorTree *DT, AssumptionCache *AC, 1280 const Instruction *CxtI) { 1281 return ::SimplifyFRemInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 1282 RecursionLimit); 1283 } 1284 1285 /// Returns true if a shift by \c Amount always yields undef. 1286 static bool isUndefShift(Value *Amount) { 1287 Constant *C = dyn_cast<Constant>(Amount); 1288 if (!C) 1289 return false; 1290 1291 // X shift by undef -> undef because it may shift by the bitwidth. 1292 if (isa<UndefValue>(C)) 1293 return true; 1294 1295 // Shifting by the bitwidth or more is undefined. 1296 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) 1297 if (CI->getValue().getLimitedValue() >= 1298 CI->getType()->getScalarSizeInBits()) 1299 return true; 1300 1301 // If all lanes of a vector shift are undefined the whole shift is. 1302 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) { 1303 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I) 1304 if (!isUndefShift(C->getAggregateElement(I))) 1305 return false; 1306 return true; 1307 } 1308 1309 return false; 1310 } 1311 1312 /// Given operands for an Shl, LShr or AShr, see if we can fold the result. 1313 /// If not, this returns null. 1314 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1, 1315 const Query &Q, unsigned MaxRecurse) { 1316 if (Constant *C0 = dyn_cast<Constant>(Op0)) 1317 if (Constant *C1 = dyn_cast<Constant>(Op1)) 1318 return ConstantFoldBinaryOpOperands(Opcode, C0, C1, Q.DL); 1319 1320 // 0 shift by X -> 0 1321 if (match(Op0, m_Zero())) 1322 return Op0; 1323 1324 // X shift by 0 -> X 1325 if (match(Op1, m_Zero())) 1326 return Op0; 1327 1328 // Fold undefined shifts. 1329 if (isUndefShift(Op1)) 1330 return UndefValue::get(Op0->getType()); 1331 1332 // If the operation is with the result of a select instruction, check whether 1333 // operating on either branch of the select always yields the same value. 1334 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1335 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1336 return V; 1337 1338 // If the operation is with the result of a phi instruction, check whether 1339 // operating on all incoming values of the phi always yields the same value. 1340 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1341 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1342 return V; 1343 1344 // If any bits in the shift amount make that value greater than or equal to 1345 // the number of bits in the type, the shift is undefined. 1346 unsigned BitWidth = Op1->getType()->getScalarSizeInBits(); 1347 APInt KnownZero(BitWidth, 0); 1348 APInt KnownOne(BitWidth, 0); 1349 computeKnownBits(Op1, KnownZero, KnownOne, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 1350 if (KnownOne.getLimitedValue() >= BitWidth) 1351 return UndefValue::get(Op0->getType()); 1352 1353 // If all valid bits in the shift amount are known zero, the first operand is 1354 // unchanged. 1355 unsigned NumValidShiftBits = Log2_32_Ceil(BitWidth); 1356 APInt ShiftAmountMask = APInt::getLowBitsSet(BitWidth, NumValidShiftBits); 1357 if ((KnownZero & ShiftAmountMask) == ShiftAmountMask) 1358 return Op0; 1359 1360 return nullptr; 1361 } 1362 1363 /// \brief Given operands for an Shl, LShr or AShr, see if we can 1364 /// fold the result. If not, this returns null. 1365 static Value *SimplifyRightShift(unsigned Opcode, Value *Op0, Value *Op1, 1366 bool isExact, const Query &Q, 1367 unsigned MaxRecurse) { 1368 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse)) 1369 return V; 1370 1371 // X >> X -> 0 1372 if (Op0 == Op1) 1373 return Constant::getNullValue(Op0->getType()); 1374 1375 // undef >> X -> 0 1376 // undef >> X -> undef (if it's exact) 1377 if (match(Op0, m_Undef())) 1378 return isExact ? Op0 : Constant::getNullValue(Op0->getType()); 1379 1380 // The low bit cannot be shifted out of an exact shift if it is set. 1381 if (isExact) { 1382 unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); 1383 APInt Op0KnownZero(BitWidth, 0); 1384 APInt Op0KnownOne(BitWidth, 0); 1385 computeKnownBits(Op0, Op0KnownZero, Op0KnownOne, Q.DL, /*Depth=*/0, Q.AC, 1386 Q.CxtI, Q.DT); 1387 if (Op0KnownOne[0]) 1388 return Op0; 1389 } 1390 1391 return nullptr; 1392 } 1393 1394 /// Given operands for an Shl, see if we can fold the result. 1395 /// If not, this returns null. 1396 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1397 const Query &Q, unsigned MaxRecurse) { 1398 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse)) 1399 return V; 1400 1401 // undef << X -> 0 1402 // undef << X -> undef if (if it's NSW/NUW) 1403 if (match(Op0, m_Undef())) 1404 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType()); 1405 1406 // (X >> A) << A -> X 1407 Value *X; 1408 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1))))) 1409 return X; 1410 return nullptr; 1411 } 1412 1413 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1414 const DataLayout &DL, const TargetLibraryInfo *TLI, 1415 const DominatorTree *DT, AssumptionCache *AC, 1416 const Instruction *CxtI) { 1417 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 1418 RecursionLimit); 1419 } 1420 1421 /// Given operands for an LShr, see if we can fold the result. 1422 /// If not, this returns null. 1423 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1424 const Query &Q, unsigned MaxRecurse) { 1425 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q, 1426 MaxRecurse)) 1427 return V; 1428 1429 // (X << A) >> A -> X 1430 Value *X; 1431 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1)))) 1432 return X; 1433 1434 return nullptr; 1435 } 1436 1437 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1438 const DataLayout &DL, 1439 const TargetLibraryInfo *TLI, 1440 const DominatorTree *DT, AssumptionCache *AC, 1441 const Instruction *CxtI) { 1442 return ::SimplifyLShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1443 RecursionLimit); 1444 } 1445 1446 /// Given operands for an AShr, see if we can fold the result. 1447 /// If not, this returns null. 1448 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1449 const Query &Q, unsigned MaxRecurse) { 1450 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q, 1451 MaxRecurse)) 1452 return V; 1453 1454 // all ones >>a X -> all ones 1455 if (match(Op0, m_AllOnes())) 1456 return Op0; 1457 1458 // (X << A) >> A -> X 1459 Value *X; 1460 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1)))) 1461 return X; 1462 1463 // Arithmetic shifting an all-sign-bit value is a no-op. 1464 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 1465 if (NumSignBits == Op0->getType()->getScalarSizeInBits()) 1466 return Op0; 1467 1468 return nullptr; 1469 } 1470 1471 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1472 const DataLayout &DL, 1473 const TargetLibraryInfo *TLI, 1474 const DominatorTree *DT, AssumptionCache *AC, 1475 const Instruction *CxtI) { 1476 return ::SimplifyAShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1477 RecursionLimit); 1478 } 1479 1480 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp, 1481 ICmpInst *UnsignedICmp, bool IsAnd) { 1482 Value *X, *Y; 1483 1484 ICmpInst::Predicate EqPred; 1485 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) || 1486 !ICmpInst::isEquality(EqPred)) 1487 return nullptr; 1488 1489 ICmpInst::Predicate UnsignedPred; 1490 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) && 1491 ICmpInst::isUnsigned(UnsignedPred)) 1492 ; 1493 else if (match(UnsignedICmp, 1494 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) && 1495 ICmpInst::isUnsigned(UnsignedPred)) 1496 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred); 1497 else 1498 return nullptr; 1499 1500 // X < Y && Y != 0 --> X < Y 1501 // X < Y || Y != 0 --> Y != 0 1502 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE) 1503 return IsAnd ? UnsignedICmp : ZeroICmp; 1504 1505 // X >= Y || Y != 0 --> true 1506 // X >= Y || Y == 0 --> X >= Y 1507 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) { 1508 if (EqPred == ICmpInst::ICMP_NE) 1509 return getTrue(UnsignedICmp->getType()); 1510 return UnsignedICmp; 1511 } 1512 1513 // X < Y && Y == 0 --> false 1514 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ && 1515 IsAnd) 1516 return getFalse(UnsignedICmp->getType()); 1517 1518 return nullptr; 1519 } 1520 1521 static Value *SimplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1522 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true)) 1523 return X; 1524 1525 // Look for this pattern: (icmp V, C0) & (icmp V, C1)). 1526 Type *ITy = Op0->getType(); 1527 ICmpInst::Predicate Pred0, Pred1; 1528 const APInt *C0, *C1; 1529 Value *V; 1530 if (match(Op0, m_ICmp(Pred0, m_Value(V), m_APInt(C0))) && 1531 match(Op1, m_ICmp(Pred1, m_Specific(V), m_APInt(C1)))) { 1532 // Make a constant range that's the intersection of the two icmp ranges. 1533 // If the intersection is empty, we know that the result is false. 1534 auto Range0 = ConstantRange::makeAllowedICmpRegion(Pred0, *C0); 1535 auto Range1 = ConstantRange::makeAllowedICmpRegion(Pred1, *C1); 1536 if (Range0.intersectWith(Range1).isEmptySet()) 1537 return getFalse(ITy); 1538 } 1539 1540 // (icmp (add V, C0), C1) & (icmp V, C0) 1541 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1)))) 1542 return nullptr; 1543 1544 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value()))) 1545 return nullptr; 1546 1547 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1548 if (AddInst->getOperand(1) != Op1->getOperand(1)) 1549 return nullptr; 1550 1551 bool isNSW = AddInst->hasNoSignedWrap(); 1552 bool isNUW = AddInst->hasNoUnsignedWrap(); 1553 1554 const APInt Delta = *C1 - *C0; 1555 if (C0->isStrictlyPositive()) { 1556 if (Delta == 2) { 1557 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT) 1558 return getFalse(ITy); 1559 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1560 return getFalse(ITy); 1561 } 1562 if (Delta == 1) { 1563 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT) 1564 return getFalse(ITy); 1565 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1566 return getFalse(ITy); 1567 } 1568 } 1569 if (C0->getBoolValue() && isNUW) { 1570 if (Delta == 2) 1571 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT) 1572 return getFalse(ITy); 1573 if (Delta == 1) 1574 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT) 1575 return getFalse(ITy); 1576 } 1577 1578 return nullptr; 1579 } 1580 1581 /// Given operands for an And, see if we can fold the result. 1582 /// If not, this returns null. 1583 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q, 1584 unsigned MaxRecurse) { 1585 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1586 if (Constant *CRHS = dyn_cast<Constant>(Op1)) 1587 return ConstantFoldBinaryOpOperands(Instruction::And, CLHS, CRHS, Q.DL); 1588 1589 // Canonicalize the constant to the RHS. 1590 std::swap(Op0, Op1); 1591 } 1592 1593 // X & undef -> 0 1594 if (match(Op1, m_Undef())) 1595 return Constant::getNullValue(Op0->getType()); 1596 1597 // X & X = X 1598 if (Op0 == Op1) 1599 return Op0; 1600 1601 // X & 0 = 0 1602 if (match(Op1, m_Zero())) 1603 return Op1; 1604 1605 // X & -1 = X 1606 if (match(Op1, m_AllOnes())) 1607 return Op0; 1608 1609 // A & ~A = ~A & A = 0 1610 if (match(Op0, m_Not(m_Specific(Op1))) || 1611 match(Op1, m_Not(m_Specific(Op0)))) 1612 return Constant::getNullValue(Op0->getType()); 1613 1614 // (A | ?) & A = A 1615 Value *A = nullptr, *B = nullptr; 1616 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1617 (A == Op1 || B == Op1)) 1618 return Op1; 1619 1620 // A & (A | ?) = A 1621 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1622 (A == Op0 || B == Op0)) 1623 return Op0; 1624 1625 // A & (-A) = A if A is a power of two or zero. 1626 if (match(Op0, m_Neg(m_Specific(Op1))) || 1627 match(Op1, m_Neg(m_Specific(Op0)))) { 1628 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, 1629 Q.DT)) 1630 return Op0; 1631 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, 1632 Q.DT)) 1633 return Op1; 1634 } 1635 1636 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1637 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1638 if (Value *V = SimplifyAndOfICmps(ICILHS, ICIRHS)) 1639 return V; 1640 if (Value *V = SimplifyAndOfICmps(ICIRHS, ICILHS)) 1641 return V; 1642 } 1643 } 1644 1645 // The compares may be hidden behind casts. Look through those and try the 1646 // same folds as above. 1647 auto *Cast0 = dyn_cast<CastInst>(Op0); 1648 auto *Cast1 = dyn_cast<CastInst>(Op1); 1649 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() && 1650 Cast0->getSrcTy() == Cast1->getSrcTy()) { 1651 auto *Cmp0 = dyn_cast<ICmpInst>(Cast0->getOperand(0)); 1652 auto *Cmp1 = dyn_cast<ICmpInst>(Cast1->getOperand(0)); 1653 if (Cmp0 && Cmp1) { 1654 Instruction::CastOps CastOpc = Cast0->getOpcode(); 1655 Type *ResultType = Cast0->getType(); 1656 if (auto *V = dyn_cast_or_null<Constant>(SimplifyAndOfICmps(Cmp0, Cmp1))) 1657 return ConstantExpr::getCast(CastOpc, V, ResultType); 1658 if (auto *V = dyn_cast_or_null<Constant>(SimplifyAndOfICmps(Cmp1, Cmp0))) 1659 return ConstantExpr::getCast(CastOpc, V, ResultType); 1660 } 1661 } 1662 1663 // Try some generic simplifications for associative operations. 1664 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q, 1665 MaxRecurse)) 1666 return V; 1667 1668 // And distributes over Or. Try some generic simplifications based on this. 1669 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or, 1670 Q, MaxRecurse)) 1671 return V; 1672 1673 // And distributes over Xor. Try some generic simplifications based on this. 1674 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor, 1675 Q, MaxRecurse)) 1676 return V; 1677 1678 // If the operation is with the result of a select instruction, check whether 1679 // operating on either branch of the select always yields the same value. 1680 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1681 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q, 1682 MaxRecurse)) 1683 return V; 1684 1685 // If the operation is with the result of a phi instruction, check whether 1686 // operating on all incoming values of the phi always yields the same value. 1687 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1688 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q, 1689 MaxRecurse)) 1690 return V; 1691 1692 return nullptr; 1693 } 1694 1695 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout &DL, 1696 const TargetLibraryInfo *TLI, 1697 const DominatorTree *DT, AssumptionCache *AC, 1698 const Instruction *CxtI) { 1699 return ::SimplifyAndInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1700 RecursionLimit); 1701 } 1702 1703 /// Simplify (or (icmp ...) (icmp ...)) to true when we can tell that the union 1704 /// contains all possible values. 1705 static Value *SimplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1706 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false)) 1707 return X; 1708 1709 // (icmp (add V, C0), C1) | (icmp V, C0) 1710 ICmpInst::Predicate Pred0, Pred1; 1711 const APInt *C0, *C1; 1712 Value *V; 1713 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1)))) 1714 return nullptr; 1715 1716 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value()))) 1717 return nullptr; 1718 1719 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1720 if (AddInst->getOperand(1) != Op1->getOperand(1)) 1721 return nullptr; 1722 1723 Type *ITy = Op0->getType(); 1724 bool isNSW = AddInst->hasNoSignedWrap(); 1725 bool isNUW = AddInst->hasNoUnsignedWrap(); 1726 1727 const APInt Delta = *C1 - *C0; 1728 if (C0->isStrictlyPositive()) { 1729 if (Delta == 2) { 1730 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE) 1731 return getTrue(ITy); 1732 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1733 return getTrue(ITy); 1734 } 1735 if (Delta == 1) { 1736 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE) 1737 return getTrue(ITy); 1738 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1739 return getTrue(ITy); 1740 } 1741 } 1742 if (C0->getBoolValue() && isNUW) { 1743 if (Delta == 2) 1744 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE) 1745 return getTrue(ITy); 1746 if (Delta == 1) 1747 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE) 1748 return getTrue(ITy); 1749 } 1750 1751 return nullptr; 1752 } 1753 1754 /// Given operands for an Or, see if we can fold the result. 1755 /// If not, this returns null. 1756 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q, 1757 unsigned MaxRecurse) { 1758 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1759 if (Constant *CRHS = dyn_cast<Constant>(Op1)) 1760 return ConstantFoldBinaryOpOperands(Instruction::Or, CLHS, CRHS, Q.DL); 1761 1762 // Canonicalize the constant to the RHS. 1763 std::swap(Op0, Op1); 1764 } 1765 1766 // X | undef -> -1 1767 if (match(Op1, m_Undef())) 1768 return Constant::getAllOnesValue(Op0->getType()); 1769 1770 // X | X = X 1771 if (Op0 == Op1) 1772 return Op0; 1773 1774 // X | 0 = X 1775 if (match(Op1, m_Zero())) 1776 return Op0; 1777 1778 // X | -1 = -1 1779 if (match(Op1, m_AllOnes())) 1780 return Op1; 1781 1782 // A | ~A = ~A | A = -1 1783 if (match(Op0, m_Not(m_Specific(Op1))) || 1784 match(Op1, m_Not(m_Specific(Op0)))) 1785 return Constant::getAllOnesValue(Op0->getType()); 1786 1787 // (A & ?) | A = A 1788 Value *A = nullptr, *B = nullptr; 1789 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 1790 (A == Op1 || B == Op1)) 1791 return Op1; 1792 1793 // A | (A & ?) = A 1794 if (match(Op1, m_And(m_Value(A), m_Value(B))) && 1795 (A == Op0 || B == Op0)) 1796 return Op0; 1797 1798 // ~(A & ?) | A = -1 1799 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) && 1800 (A == Op1 || B == Op1)) 1801 return Constant::getAllOnesValue(Op1->getType()); 1802 1803 // A | ~(A & ?) = -1 1804 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) && 1805 (A == Op0 || B == Op0)) 1806 return Constant::getAllOnesValue(Op0->getType()); 1807 1808 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1809 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1810 if (Value *V = SimplifyOrOfICmps(ICILHS, ICIRHS)) 1811 return V; 1812 if (Value *V = SimplifyOrOfICmps(ICIRHS, ICILHS)) 1813 return V; 1814 } 1815 } 1816 1817 // Try some generic simplifications for associative operations. 1818 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q, 1819 MaxRecurse)) 1820 return V; 1821 1822 // Or distributes over And. Try some generic simplifications based on this. 1823 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q, 1824 MaxRecurse)) 1825 return V; 1826 1827 // If the operation is with the result of a select instruction, check whether 1828 // operating on either branch of the select always yields the same value. 1829 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1830 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q, 1831 MaxRecurse)) 1832 return V; 1833 1834 // (A & C)|(B & D) 1835 Value *C = nullptr, *D = nullptr; 1836 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 1837 match(Op1, m_And(m_Value(B), m_Value(D)))) { 1838 ConstantInt *C1 = dyn_cast<ConstantInt>(C); 1839 ConstantInt *C2 = dyn_cast<ConstantInt>(D); 1840 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) { 1841 // (A & C1)|(B & C2) 1842 // If we have: ((V + N) & C1) | (V & C2) 1843 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0 1844 // replace with V+N. 1845 Value *V1, *V2; 1846 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+ 1847 match(A, m_Add(m_Value(V1), m_Value(V2)))) { 1848 // Add commutes, try both ways. 1849 if (V1 == B && 1850 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1851 return A; 1852 if (V2 == B && 1853 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1854 return A; 1855 } 1856 // Or commutes, try both ways. 1857 if ((C1->getValue() & (C1->getValue() + 1)) == 0 && 1858 match(B, m_Add(m_Value(V1), m_Value(V2)))) { 1859 // Add commutes, try both ways. 1860 if (V1 == A && 1861 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1862 return B; 1863 if (V2 == A && 1864 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1865 return B; 1866 } 1867 } 1868 } 1869 1870 // If the operation is with the result of a phi instruction, check whether 1871 // operating on all incoming values of the phi always yields the same value. 1872 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1873 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse)) 1874 return V; 1875 1876 return nullptr; 1877 } 1878 1879 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout &DL, 1880 const TargetLibraryInfo *TLI, 1881 const DominatorTree *DT, AssumptionCache *AC, 1882 const Instruction *CxtI) { 1883 return ::SimplifyOrInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1884 RecursionLimit); 1885 } 1886 1887 /// Given operands for a Xor, see if we can fold the result. 1888 /// If not, this returns null. 1889 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q, 1890 unsigned MaxRecurse) { 1891 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1892 if (Constant *CRHS = dyn_cast<Constant>(Op1)) 1893 return ConstantFoldBinaryOpOperands(Instruction::Xor, CLHS, CRHS, Q.DL); 1894 1895 // Canonicalize the constant to the RHS. 1896 std::swap(Op0, Op1); 1897 } 1898 1899 // A ^ undef -> undef 1900 if (match(Op1, m_Undef())) 1901 return Op1; 1902 1903 // A ^ 0 = A 1904 if (match(Op1, m_Zero())) 1905 return Op0; 1906 1907 // A ^ A = 0 1908 if (Op0 == Op1) 1909 return Constant::getNullValue(Op0->getType()); 1910 1911 // A ^ ~A = ~A ^ A = -1 1912 if (match(Op0, m_Not(m_Specific(Op1))) || 1913 match(Op1, m_Not(m_Specific(Op0)))) 1914 return Constant::getAllOnesValue(Op0->getType()); 1915 1916 // Try some generic simplifications for associative operations. 1917 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q, 1918 MaxRecurse)) 1919 return V; 1920 1921 // Threading Xor over selects and phi nodes is pointless, so don't bother. 1922 // Threading over the select in "A ^ select(cond, B, C)" means evaluating 1923 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and 1924 // only if B and C are equal. If B and C are equal then (since we assume 1925 // that operands have already been simplified) "select(cond, B, C)" should 1926 // have been simplified to the common value of B and C already. Analysing 1927 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly 1928 // for threading over phi nodes. 1929 1930 return nullptr; 1931 } 1932 1933 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout &DL, 1934 const TargetLibraryInfo *TLI, 1935 const DominatorTree *DT, AssumptionCache *AC, 1936 const Instruction *CxtI) { 1937 return ::SimplifyXorInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1938 RecursionLimit); 1939 } 1940 1941 static Type *GetCompareTy(Value *Op) { 1942 return CmpInst::makeCmpResultType(Op->getType()); 1943 } 1944 1945 /// Rummage around inside V looking for something equivalent to the comparison 1946 /// "LHS Pred RHS". Return such a value if found, otherwise return null. 1947 /// Helper function for analyzing max/min idioms. 1948 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred, 1949 Value *LHS, Value *RHS) { 1950 SelectInst *SI = dyn_cast<SelectInst>(V); 1951 if (!SI) 1952 return nullptr; 1953 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); 1954 if (!Cmp) 1955 return nullptr; 1956 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1); 1957 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS) 1958 return Cmp; 1959 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) && 1960 LHS == CmpRHS && RHS == CmpLHS) 1961 return Cmp; 1962 return nullptr; 1963 } 1964 1965 // A significant optimization not implemented here is assuming that alloca 1966 // addresses are not equal to incoming argument values. They don't *alias*, 1967 // as we say, but that doesn't mean they aren't equal, so we take a 1968 // conservative approach. 1969 // 1970 // This is inspired in part by C++11 5.10p1: 1971 // "Two pointers of the same type compare equal if and only if they are both 1972 // null, both point to the same function, or both represent the same 1973 // address." 1974 // 1975 // This is pretty permissive. 1976 // 1977 // It's also partly due to C11 6.5.9p6: 1978 // "Two pointers compare equal if and only if both are null pointers, both are 1979 // pointers to the same object (including a pointer to an object and a 1980 // subobject at its beginning) or function, both are pointers to one past the 1981 // last element of the same array object, or one is a pointer to one past the 1982 // end of one array object and the other is a pointer to the start of a 1983 // different array object that happens to immediately follow the first array 1984 // object in the address space.) 1985 // 1986 // C11's version is more restrictive, however there's no reason why an argument 1987 // couldn't be a one-past-the-end value for a stack object in the caller and be 1988 // equal to the beginning of a stack object in the callee. 1989 // 1990 // If the C and C++ standards are ever made sufficiently restrictive in this 1991 // area, it may be possible to update LLVM's semantics accordingly and reinstate 1992 // this optimization. 1993 static Constant * 1994 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI, 1995 const DominatorTree *DT, CmpInst::Predicate Pred, 1996 const Instruction *CxtI, Value *LHS, Value *RHS) { 1997 // First, skip past any trivial no-ops. 1998 LHS = LHS->stripPointerCasts(); 1999 RHS = RHS->stripPointerCasts(); 2000 2001 // A non-null pointer is not equal to a null pointer. 2002 if (llvm::isKnownNonNull(LHS) && isa<ConstantPointerNull>(RHS) && 2003 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE)) 2004 return ConstantInt::get(GetCompareTy(LHS), 2005 !CmpInst::isTrueWhenEqual(Pred)); 2006 2007 // We can only fold certain predicates on pointer comparisons. 2008 switch (Pred) { 2009 default: 2010 return nullptr; 2011 2012 // Equality comaprisons are easy to fold. 2013 case CmpInst::ICMP_EQ: 2014 case CmpInst::ICMP_NE: 2015 break; 2016 2017 // We can only handle unsigned relational comparisons because 'inbounds' on 2018 // a GEP only protects against unsigned wrapping. 2019 case CmpInst::ICMP_UGT: 2020 case CmpInst::ICMP_UGE: 2021 case CmpInst::ICMP_ULT: 2022 case CmpInst::ICMP_ULE: 2023 // However, we have to switch them to their signed variants to handle 2024 // negative indices from the base pointer. 2025 Pred = ICmpInst::getSignedPredicate(Pred); 2026 break; 2027 } 2028 2029 // Strip off any constant offsets so that we can reason about them. 2030 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets 2031 // here and compare base addresses like AliasAnalysis does, however there are 2032 // numerous hazards. AliasAnalysis and its utilities rely on special rules 2033 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis 2034 // doesn't need to guarantee pointer inequality when it says NoAlias. 2035 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 2036 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 2037 2038 // If LHS and RHS are related via constant offsets to the same base 2039 // value, we can replace it with an icmp which just compares the offsets. 2040 if (LHS == RHS) 2041 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset); 2042 2043 // Various optimizations for (in)equality comparisons. 2044 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) { 2045 // Different non-empty allocations that exist at the same time have 2046 // different addresses (if the program can tell). Global variables always 2047 // exist, so they always exist during the lifetime of each other and all 2048 // allocas. Two different allocas usually have different addresses... 2049 // 2050 // However, if there's an @llvm.stackrestore dynamically in between two 2051 // allocas, they may have the same address. It's tempting to reduce the 2052 // scope of the problem by only looking at *static* allocas here. That would 2053 // cover the majority of allocas while significantly reducing the likelihood 2054 // of having an @llvm.stackrestore pop up in the middle. However, it's not 2055 // actually impossible for an @llvm.stackrestore to pop up in the middle of 2056 // an entry block. Also, if we have a block that's not attached to a 2057 // function, we can't tell if it's "static" under the current definition. 2058 // Theoretically, this problem could be fixed by creating a new kind of 2059 // instruction kind specifically for static allocas. Such a new instruction 2060 // could be required to be at the top of the entry block, thus preventing it 2061 // from being subject to a @llvm.stackrestore. Instcombine could even 2062 // convert regular allocas into these special allocas. It'd be nifty. 2063 // However, until then, this problem remains open. 2064 // 2065 // So, we'll assume that two non-empty allocas have different addresses 2066 // for now. 2067 // 2068 // With all that, if the offsets are within the bounds of their allocations 2069 // (and not one-past-the-end! so we can't use inbounds!), and their 2070 // allocations aren't the same, the pointers are not equal. 2071 // 2072 // Note that it's not necessary to check for LHS being a global variable 2073 // address, due to canonicalization and constant folding. 2074 if (isa<AllocaInst>(LHS) && 2075 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) { 2076 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset); 2077 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset); 2078 uint64_t LHSSize, RHSSize; 2079 if (LHSOffsetCI && RHSOffsetCI && 2080 getObjectSize(LHS, LHSSize, DL, TLI) && 2081 getObjectSize(RHS, RHSSize, DL, TLI)) { 2082 const APInt &LHSOffsetValue = LHSOffsetCI->getValue(); 2083 const APInt &RHSOffsetValue = RHSOffsetCI->getValue(); 2084 if (!LHSOffsetValue.isNegative() && 2085 !RHSOffsetValue.isNegative() && 2086 LHSOffsetValue.ult(LHSSize) && 2087 RHSOffsetValue.ult(RHSSize)) { 2088 return ConstantInt::get(GetCompareTy(LHS), 2089 !CmpInst::isTrueWhenEqual(Pred)); 2090 } 2091 } 2092 2093 // Repeat the above check but this time without depending on DataLayout 2094 // or being able to compute a precise size. 2095 if (!cast<PointerType>(LHS->getType())->isEmptyTy() && 2096 !cast<PointerType>(RHS->getType())->isEmptyTy() && 2097 LHSOffset->isNullValue() && 2098 RHSOffset->isNullValue()) 2099 return ConstantInt::get(GetCompareTy(LHS), 2100 !CmpInst::isTrueWhenEqual(Pred)); 2101 } 2102 2103 // Even if an non-inbounds GEP occurs along the path we can still optimize 2104 // equality comparisons concerning the result. We avoid walking the whole 2105 // chain again by starting where the last calls to 2106 // stripAndComputeConstantOffsets left off and accumulate the offsets. 2107 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true); 2108 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true); 2109 if (LHS == RHS) 2110 return ConstantExpr::getICmp(Pred, 2111 ConstantExpr::getAdd(LHSOffset, LHSNoBound), 2112 ConstantExpr::getAdd(RHSOffset, RHSNoBound)); 2113 2114 // If one side of the equality comparison must come from a noalias call 2115 // (meaning a system memory allocation function), and the other side must 2116 // come from a pointer that cannot overlap with dynamically-allocated 2117 // memory within the lifetime of the current function (allocas, byval 2118 // arguments, globals), then determine the comparison result here. 2119 SmallVector<Value *, 8> LHSUObjs, RHSUObjs; 2120 GetUnderlyingObjects(LHS, LHSUObjs, DL); 2121 GetUnderlyingObjects(RHS, RHSUObjs, DL); 2122 2123 // Is the set of underlying objects all noalias calls? 2124 auto IsNAC = [](ArrayRef<Value *> Objects) { 2125 return all_of(Objects, isNoAliasCall); 2126 }; 2127 2128 // Is the set of underlying objects all things which must be disjoint from 2129 // noalias calls. For allocas, we consider only static ones (dynamic 2130 // allocas might be transformed into calls to malloc not simultaneously 2131 // live with the compared-to allocation). For globals, we exclude symbols 2132 // that might be resolve lazily to symbols in another dynamically-loaded 2133 // library (and, thus, could be malloc'ed by the implementation). 2134 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) { 2135 return all_of(Objects, [](Value *V) { 2136 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) 2137 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca(); 2138 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 2139 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() || 2140 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) && 2141 !GV->isThreadLocal(); 2142 if (const Argument *A = dyn_cast<Argument>(V)) 2143 return A->hasByValAttr(); 2144 return false; 2145 }); 2146 }; 2147 2148 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) || 2149 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs))) 2150 return ConstantInt::get(GetCompareTy(LHS), 2151 !CmpInst::isTrueWhenEqual(Pred)); 2152 2153 // Fold comparisons for non-escaping pointer even if the allocation call 2154 // cannot be elided. We cannot fold malloc comparison to null. Also, the 2155 // dynamic allocation call could be either of the operands. 2156 Value *MI = nullptr; 2157 if (isAllocLikeFn(LHS, TLI) && llvm::isKnownNonNullAt(RHS, CxtI, DT)) 2158 MI = LHS; 2159 else if (isAllocLikeFn(RHS, TLI) && llvm::isKnownNonNullAt(LHS, CxtI, DT)) 2160 MI = RHS; 2161 // FIXME: We should also fold the compare when the pointer escapes, but the 2162 // compare dominates the pointer escape 2163 if (MI && !PointerMayBeCaptured(MI, true, true)) 2164 return ConstantInt::get(GetCompareTy(LHS), 2165 CmpInst::isFalseWhenEqual(Pred)); 2166 } 2167 2168 // Otherwise, fail. 2169 return nullptr; 2170 } 2171 2172 /// Fold an icmp when its operands have i1 scalar type. 2173 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS, 2174 Value *RHS, const Query &Q) { 2175 Type *ITy = GetCompareTy(LHS); // The return type. 2176 Type *OpTy = LHS->getType(); // The operand type. 2177 if (!OpTy->getScalarType()->isIntegerTy(1)) 2178 return nullptr; 2179 2180 switch (Pred) { 2181 default: 2182 break; 2183 case ICmpInst::ICMP_EQ: 2184 // X == 1 -> X 2185 if (match(RHS, m_One())) 2186 return LHS; 2187 break; 2188 case ICmpInst::ICMP_NE: 2189 // X != 0 -> X 2190 if (match(RHS, m_Zero())) 2191 return LHS; 2192 break; 2193 case ICmpInst::ICMP_UGT: 2194 // X >u 0 -> X 2195 if (match(RHS, m_Zero())) 2196 return LHS; 2197 break; 2198 case ICmpInst::ICMP_UGE: 2199 // X >=u 1 -> X 2200 if (match(RHS, m_One())) 2201 return LHS; 2202 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false)) 2203 return getTrue(ITy); 2204 break; 2205 case ICmpInst::ICMP_SGE: 2206 /// For signed comparison, the values for an i1 are 0 and -1 2207 /// respectively. This maps into a truth table of: 2208 /// LHS | RHS | LHS >=s RHS | LHS implies RHS 2209 /// 0 | 0 | 1 (0 >= 0) | 1 2210 /// 0 | 1 | 1 (0 >= -1) | 1 2211 /// 1 | 0 | 0 (-1 >= 0) | 0 2212 /// 1 | 1 | 1 (-1 >= -1) | 1 2213 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false)) 2214 return getTrue(ITy); 2215 break; 2216 case ICmpInst::ICMP_SLT: 2217 // X <s 0 -> X 2218 if (match(RHS, m_Zero())) 2219 return LHS; 2220 break; 2221 case ICmpInst::ICMP_SLE: 2222 // X <=s -1 -> X 2223 if (match(RHS, m_One())) 2224 return LHS; 2225 break; 2226 case ICmpInst::ICMP_ULE: 2227 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false)) 2228 return getTrue(ITy); 2229 break; 2230 } 2231 2232 return nullptr; 2233 } 2234 2235 /// Try hard to fold icmp with zero RHS because this is a common case. 2236 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS, 2237 Value *RHS, const Query &Q) { 2238 if (!match(RHS, m_Zero())) 2239 return nullptr; 2240 2241 Type *ITy = GetCompareTy(LHS); // The return type. 2242 bool LHSKnownNonNegative, LHSKnownNegative; 2243 switch (Pred) { 2244 default: 2245 llvm_unreachable("Unknown ICmp predicate!"); 2246 case ICmpInst::ICMP_ULT: 2247 return getFalse(ITy); 2248 case ICmpInst::ICMP_UGE: 2249 return getTrue(ITy); 2250 case ICmpInst::ICMP_EQ: 2251 case ICmpInst::ICMP_ULE: 2252 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2253 return getFalse(ITy); 2254 break; 2255 case ICmpInst::ICMP_NE: 2256 case ICmpInst::ICMP_UGT: 2257 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2258 return getTrue(ITy); 2259 break; 2260 case ICmpInst::ICMP_SLT: 2261 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2262 Q.CxtI, Q.DT); 2263 if (LHSKnownNegative) 2264 return getTrue(ITy); 2265 if (LHSKnownNonNegative) 2266 return getFalse(ITy); 2267 break; 2268 case ICmpInst::ICMP_SLE: 2269 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2270 Q.CxtI, Q.DT); 2271 if (LHSKnownNegative) 2272 return getTrue(ITy); 2273 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2274 return getFalse(ITy); 2275 break; 2276 case ICmpInst::ICMP_SGE: 2277 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2278 Q.CxtI, Q.DT); 2279 if (LHSKnownNegative) 2280 return getFalse(ITy); 2281 if (LHSKnownNonNegative) 2282 return getTrue(ITy); 2283 break; 2284 case ICmpInst::ICMP_SGT: 2285 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2286 Q.CxtI, Q.DT); 2287 if (LHSKnownNegative) 2288 return getFalse(ITy); 2289 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2290 return getTrue(ITy); 2291 break; 2292 } 2293 2294 return nullptr; 2295 } 2296 2297 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS, 2298 Value *RHS) { 2299 const APInt *C; 2300 if (!match(RHS, m_APInt(C))) 2301 return nullptr; 2302 2303 // Rule out tautological comparisons (eg., ult 0 or uge 0). 2304 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C); 2305 if (RHS_CR.isEmptySet()) 2306 return ConstantInt::getFalse(GetCompareTy(RHS)); 2307 if (RHS_CR.isFullSet()) 2308 return ConstantInt::getTrue(GetCompareTy(RHS)); 2309 2310 // Many binary operators with constant RHS have easy to compute constant 2311 // range. Use them to check whether the comparison is a tautology. 2312 unsigned Width = C->getBitWidth(); 2313 APInt Lower = APInt(Width, 0); 2314 APInt Upper = APInt(Width, 0); 2315 const APInt *C2; 2316 if (match(LHS, m_URem(m_Value(), m_APInt(C2)))) { 2317 // 'urem x, C2' produces [0, C2). 2318 Upper = *C2; 2319 } else if (match(LHS, m_SRem(m_Value(), m_APInt(C2)))) { 2320 // 'srem x, C2' produces (-|C2|, |C2|). 2321 Upper = C2->abs(); 2322 Lower = (-Upper) + 1; 2323 } else if (match(LHS, m_UDiv(m_APInt(C2), m_Value()))) { 2324 // 'udiv C2, x' produces [0, C2]. 2325 Upper = *C2 + 1; 2326 } else if (match(LHS, m_UDiv(m_Value(), m_APInt(C2)))) { 2327 // 'udiv x, C2' produces [0, UINT_MAX / C2]. 2328 APInt NegOne = APInt::getAllOnesValue(Width); 2329 if (*C2 != 0) 2330 Upper = NegOne.udiv(*C2) + 1; 2331 } else if (match(LHS, m_SDiv(m_APInt(C2), m_Value()))) { 2332 if (C2->isMinSignedValue()) { 2333 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. 2334 Lower = *C2; 2335 Upper = Lower.lshr(1) + 1; 2336 } else { 2337 // 'sdiv C2, x' produces [-|C2|, |C2|]. 2338 Upper = C2->abs() + 1; 2339 Lower = (-Upper) + 1; 2340 } 2341 } else if (match(LHS, m_SDiv(m_Value(), m_APInt(C2)))) { 2342 APInt IntMin = APInt::getSignedMinValue(Width); 2343 APInt IntMax = APInt::getSignedMaxValue(Width); 2344 if (C2->isAllOnesValue()) { 2345 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] 2346 // where C2 != -1 and C2 != 0 and C2 != 1 2347 Lower = IntMin + 1; 2348 Upper = IntMax + 1; 2349 } else if (C2->countLeadingZeros() < Width - 1) { 2350 // 'sdiv x, C2' produces [INT_MIN / C2, INT_MAX / C2] 2351 // where C2 != -1 and C2 != 0 and C2 != 1 2352 Lower = IntMin.sdiv(*C2); 2353 Upper = IntMax.sdiv(*C2); 2354 if (Lower.sgt(Upper)) 2355 std::swap(Lower, Upper); 2356 Upper = Upper + 1; 2357 assert(Upper != Lower && "Upper part of range has wrapped!"); 2358 } 2359 } else if (match(LHS, m_NUWShl(m_APInt(C2), m_Value()))) { 2360 // 'shl nuw C2, x' produces [C2, C2 << CLZ(C2)] 2361 Lower = *C2; 2362 Upper = Lower.shl(Lower.countLeadingZeros()) + 1; 2363 } else if (match(LHS, m_NSWShl(m_APInt(C2), m_Value()))) { 2364 if (C2->isNegative()) { 2365 // 'shl nsw C2, x' produces [C2 << CLO(C2)-1, C2] 2366 unsigned ShiftAmount = C2->countLeadingOnes() - 1; 2367 Lower = C2->shl(ShiftAmount); 2368 Upper = *C2 + 1; 2369 } else { 2370 // 'shl nsw C2, x' produces [C2, C2 << CLZ(C2)-1] 2371 unsigned ShiftAmount = C2->countLeadingZeros() - 1; 2372 Lower = *C2; 2373 Upper = C2->shl(ShiftAmount) + 1; 2374 } 2375 } else if (match(LHS, m_LShr(m_Value(), m_APInt(C2)))) { 2376 // 'lshr x, C2' produces [0, UINT_MAX >> C2]. 2377 APInt NegOne = APInt::getAllOnesValue(Width); 2378 if (C2->ult(Width)) 2379 Upper = NegOne.lshr(*C2) + 1; 2380 } else if (match(LHS, m_LShr(m_APInt(C2), m_Value()))) { 2381 // 'lshr C2, x' produces [C2 >> (Width-1), C2]. 2382 unsigned ShiftAmount = Width - 1; 2383 if (*C2 != 0 && cast<BinaryOperator>(LHS)->isExact()) 2384 ShiftAmount = C2->countTrailingZeros(); 2385 Lower = C2->lshr(ShiftAmount); 2386 Upper = *C2 + 1; 2387 } else if (match(LHS, m_AShr(m_Value(), m_APInt(C2)))) { 2388 // 'ashr x, C2' produces [INT_MIN >> C2, INT_MAX >> C2]. 2389 APInt IntMin = APInt::getSignedMinValue(Width); 2390 APInt IntMax = APInt::getSignedMaxValue(Width); 2391 if (C2->ult(Width)) { 2392 Lower = IntMin.ashr(*C2); 2393 Upper = IntMax.ashr(*C2) + 1; 2394 } 2395 } else if (match(LHS, m_AShr(m_APInt(C2), m_Value()))) { 2396 unsigned ShiftAmount = Width - 1; 2397 if (*C2 != 0 && cast<BinaryOperator>(LHS)->isExact()) 2398 ShiftAmount = C2->countTrailingZeros(); 2399 if (C2->isNegative()) { 2400 // 'ashr C2, x' produces [C2, C2 >> (Width-1)] 2401 Lower = *C2; 2402 Upper = C2->ashr(ShiftAmount) + 1; 2403 } else { 2404 // 'ashr C2, x' produces [C2 >> (Width-1), C2] 2405 Lower = C2->ashr(ShiftAmount); 2406 Upper = *C2 + 1; 2407 } 2408 } else if (match(LHS, m_Or(m_Value(), m_APInt(C2)))) { 2409 // 'or x, C2' produces [C2, UINT_MAX]. 2410 Lower = *C2; 2411 } else if (match(LHS, m_And(m_Value(), m_APInt(C2)))) { 2412 // 'and x, C2' produces [0, C2]. 2413 Upper = *C2 + 1; 2414 } else if (match(LHS, m_NUWAdd(m_Value(), m_APInt(C2)))) { 2415 // 'add nuw x, C2' produces [C2, UINT_MAX]. 2416 Lower = *C2; 2417 } 2418 2419 ConstantRange LHS_CR = 2420 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true); 2421 2422 if (auto *I = dyn_cast<Instruction>(LHS)) 2423 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range)) 2424 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges)); 2425 2426 if (!LHS_CR.isFullSet()) { 2427 if (RHS_CR.contains(LHS_CR)) 2428 return ConstantInt::getTrue(GetCompareTy(RHS)); 2429 if (RHS_CR.inverse().contains(LHS_CR)) 2430 return ConstantInt::getFalse(GetCompareTy(RHS)); 2431 } 2432 2433 return nullptr; 2434 } 2435 2436 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS, 2437 Value *RHS, const Query &Q, 2438 unsigned MaxRecurse) { 2439 Type *ITy = GetCompareTy(LHS); // The return type. 2440 2441 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS); 2442 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS); 2443 if (MaxRecurse && (LBO || RBO)) { 2444 // Analyze the case when either LHS or RHS is an add instruction. 2445 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 2446 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null). 2447 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false; 2448 if (LBO && LBO->getOpcode() == Instruction::Add) { 2449 A = LBO->getOperand(0); 2450 B = LBO->getOperand(1); 2451 NoLHSWrapProblem = 2452 ICmpInst::isEquality(Pred) || 2453 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) || 2454 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap()); 2455 } 2456 if (RBO && RBO->getOpcode() == Instruction::Add) { 2457 C = RBO->getOperand(0); 2458 D = RBO->getOperand(1); 2459 NoRHSWrapProblem = 2460 ICmpInst::isEquality(Pred) || 2461 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) || 2462 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap()); 2463 } 2464 2465 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2466 if ((A == RHS || B == RHS) && NoLHSWrapProblem) 2467 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A, 2468 Constant::getNullValue(RHS->getType()), Q, 2469 MaxRecurse - 1)) 2470 return V; 2471 2472 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2473 if ((C == LHS || D == LHS) && NoRHSWrapProblem) 2474 if (Value *V = 2475 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()), 2476 C == LHS ? D : C, Q, MaxRecurse - 1)) 2477 return V; 2478 2479 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow. 2480 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem && 2481 NoRHSWrapProblem) { 2482 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2483 Value *Y, *Z; 2484 if (A == C) { 2485 // C + B == C + D -> B == D 2486 Y = B; 2487 Z = D; 2488 } else if (A == D) { 2489 // D + B == C + D -> B == C 2490 Y = B; 2491 Z = C; 2492 } else if (B == C) { 2493 // A + C == C + D -> A == D 2494 Y = A; 2495 Z = D; 2496 } else { 2497 assert(B == D); 2498 // A + D == C + D -> A == C 2499 Y = A; 2500 Z = C; 2501 } 2502 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1)) 2503 return V; 2504 } 2505 } 2506 2507 { 2508 Value *Y = nullptr; 2509 // icmp pred (or X, Y), X 2510 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) { 2511 if (Pred == ICmpInst::ICMP_ULT) 2512 return getFalse(ITy); 2513 if (Pred == ICmpInst::ICMP_UGE) 2514 return getTrue(ITy); 2515 2516 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) { 2517 bool RHSKnownNonNegative, RHSKnownNegative; 2518 bool YKnownNonNegative, YKnownNegative; 2519 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, Q.DL, 0, 2520 Q.AC, Q.CxtI, Q.DT); 2521 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Q.DL, 0, Q.AC, 2522 Q.CxtI, Q.DT); 2523 if (RHSKnownNonNegative && YKnownNegative) 2524 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy); 2525 if (RHSKnownNegative || YKnownNonNegative) 2526 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy); 2527 } 2528 } 2529 // icmp pred X, (or X, Y) 2530 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) { 2531 if (Pred == ICmpInst::ICMP_ULE) 2532 return getTrue(ITy); 2533 if (Pred == ICmpInst::ICMP_UGT) 2534 return getFalse(ITy); 2535 2536 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) { 2537 bool LHSKnownNonNegative, LHSKnownNegative; 2538 bool YKnownNonNegative, YKnownNegative; 2539 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, 2540 Q.AC, Q.CxtI, Q.DT); 2541 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Q.DL, 0, Q.AC, 2542 Q.CxtI, Q.DT); 2543 if (LHSKnownNonNegative && YKnownNegative) 2544 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy); 2545 if (LHSKnownNegative || YKnownNonNegative) 2546 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy); 2547 } 2548 } 2549 } 2550 2551 // icmp pred (and X, Y), X 2552 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)), 2553 m_And(m_Specific(RHS), m_Value())))) { 2554 if (Pred == ICmpInst::ICMP_UGT) 2555 return getFalse(ITy); 2556 if (Pred == ICmpInst::ICMP_ULE) 2557 return getTrue(ITy); 2558 } 2559 // icmp pred X, (and X, Y) 2560 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)), 2561 m_And(m_Specific(LHS), m_Value())))) { 2562 if (Pred == ICmpInst::ICMP_UGE) 2563 return getTrue(ITy); 2564 if (Pred == ICmpInst::ICMP_ULT) 2565 return getFalse(ITy); 2566 } 2567 2568 // 0 - (zext X) pred C 2569 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) { 2570 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 2571 if (RHSC->getValue().isStrictlyPositive()) { 2572 if (Pred == ICmpInst::ICMP_SLT) 2573 return ConstantInt::getTrue(RHSC->getContext()); 2574 if (Pred == ICmpInst::ICMP_SGE) 2575 return ConstantInt::getFalse(RHSC->getContext()); 2576 if (Pred == ICmpInst::ICMP_EQ) 2577 return ConstantInt::getFalse(RHSC->getContext()); 2578 if (Pred == ICmpInst::ICMP_NE) 2579 return ConstantInt::getTrue(RHSC->getContext()); 2580 } 2581 if (RHSC->getValue().isNonNegative()) { 2582 if (Pred == ICmpInst::ICMP_SLE) 2583 return ConstantInt::getTrue(RHSC->getContext()); 2584 if (Pred == ICmpInst::ICMP_SGT) 2585 return ConstantInt::getFalse(RHSC->getContext()); 2586 } 2587 } 2588 } 2589 2590 // icmp pred (urem X, Y), Y 2591 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) { 2592 bool KnownNonNegative, KnownNegative; 2593 switch (Pred) { 2594 default: 2595 break; 2596 case ICmpInst::ICMP_SGT: 2597 case ICmpInst::ICMP_SGE: 2598 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2599 Q.CxtI, Q.DT); 2600 if (!KnownNonNegative) 2601 break; 2602 LLVM_FALLTHROUGH; 2603 case ICmpInst::ICMP_EQ: 2604 case ICmpInst::ICMP_UGT: 2605 case ICmpInst::ICMP_UGE: 2606 return getFalse(ITy); 2607 case ICmpInst::ICMP_SLT: 2608 case ICmpInst::ICMP_SLE: 2609 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2610 Q.CxtI, Q.DT); 2611 if (!KnownNonNegative) 2612 break; 2613 LLVM_FALLTHROUGH; 2614 case ICmpInst::ICMP_NE: 2615 case ICmpInst::ICMP_ULT: 2616 case ICmpInst::ICMP_ULE: 2617 return getTrue(ITy); 2618 } 2619 } 2620 2621 // icmp pred X, (urem Y, X) 2622 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) { 2623 bool KnownNonNegative, KnownNegative; 2624 switch (Pred) { 2625 default: 2626 break; 2627 case ICmpInst::ICMP_SGT: 2628 case ICmpInst::ICMP_SGE: 2629 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2630 Q.CxtI, Q.DT); 2631 if (!KnownNonNegative) 2632 break; 2633 LLVM_FALLTHROUGH; 2634 case ICmpInst::ICMP_NE: 2635 case ICmpInst::ICMP_UGT: 2636 case ICmpInst::ICMP_UGE: 2637 return getTrue(ITy); 2638 case ICmpInst::ICMP_SLT: 2639 case ICmpInst::ICMP_SLE: 2640 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2641 Q.CxtI, Q.DT); 2642 if (!KnownNonNegative) 2643 break; 2644 LLVM_FALLTHROUGH; 2645 case ICmpInst::ICMP_EQ: 2646 case ICmpInst::ICMP_ULT: 2647 case ICmpInst::ICMP_ULE: 2648 return getFalse(ITy); 2649 } 2650 } 2651 2652 // x >> y <=u x 2653 // x udiv y <=u x. 2654 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) || 2655 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) { 2656 // icmp pred (X op Y), X 2657 if (Pred == ICmpInst::ICMP_UGT) 2658 return getFalse(ITy); 2659 if (Pred == ICmpInst::ICMP_ULE) 2660 return getTrue(ITy); 2661 } 2662 2663 // x >=u x >> y 2664 // x >=u x udiv y. 2665 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) || 2666 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) { 2667 // icmp pred X, (X op Y) 2668 if (Pred == ICmpInst::ICMP_ULT) 2669 return getFalse(ITy); 2670 if (Pred == ICmpInst::ICMP_UGE) 2671 return getTrue(ITy); 2672 } 2673 2674 // handle: 2675 // CI2 << X == CI 2676 // CI2 << X != CI 2677 // 2678 // where CI2 is a power of 2 and CI isn't 2679 if (auto *CI = dyn_cast<ConstantInt>(RHS)) { 2680 const APInt *CI2Val, *CIVal = &CI->getValue(); 2681 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) && 2682 CI2Val->isPowerOf2()) { 2683 if (!CIVal->isPowerOf2()) { 2684 // CI2 << X can equal zero in some circumstances, 2685 // this simplification is unsafe if CI is zero. 2686 // 2687 // We know it is safe if: 2688 // - The shift is nsw, we can't shift out the one bit. 2689 // - The shift is nuw, we can't shift out the one bit. 2690 // - CI2 is one 2691 // - CI isn't zero 2692 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() || 2693 *CI2Val == 1 || !CI->isZero()) { 2694 if (Pred == ICmpInst::ICMP_EQ) 2695 return ConstantInt::getFalse(RHS->getContext()); 2696 if (Pred == ICmpInst::ICMP_NE) 2697 return ConstantInt::getTrue(RHS->getContext()); 2698 } 2699 } 2700 if (CIVal->isSignBit() && *CI2Val == 1) { 2701 if (Pred == ICmpInst::ICMP_UGT) 2702 return ConstantInt::getFalse(RHS->getContext()); 2703 if (Pred == ICmpInst::ICMP_ULE) 2704 return ConstantInt::getTrue(RHS->getContext()); 2705 } 2706 } 2707 } 2708 2709 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() && 2710 LBO->getOperand(1) == RBO->getOperand(1)) { 2711 switch (LBO->getOpcode()) { 2712 default: 2713 break; 2714 case Instruction::UDiv: 2715 case Instruction::LShr: 2716 if (ICmpInst::isSigned(Pred)) 2717 break; 2718 LLVM_FALLTHROUGH; 2719 case Instruction::SDiv: 2720 case Instruction::AShr: 2721 if (!LBO->isExact() || !RBO->isExact()) 2722 break; 2723 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2724 RBO->getOperand(0), Q, MaxRecurse - 1)) 2725 return V; 2726 break; 2727 case Instruction::Shl: { 2728 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap(); 2729 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap(); 2730 if (!NUW && !NSW) 2731 break; 2732 if (!NSW && ICmpInst::isSigned(Pred)) 2733 break; 2734 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2735 RBO->getOperand(0), Q, MaxRecurse - 1)) 2736 return V; 2737 break; 2738 } 2739 } 2740 } 2741 return nullptr; 2742 } 2743 2744 /// Simplify comparisons corresponding to integer min/max idioms. 2745 static Value *simplifyMinMax(CmpInst::Predicate Pred, Value *LHS, Value *RHS, 2746 const Query &Q, unsigned MaxRecurse) { 2747 Type *ITy = GetCompareTy(LHS); // The return type. 2748 Value *A, *B; 2749 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE; 2750 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B". 2751 2752 // Signed variants on "max(a,b)>=a -> true". 2753 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2754 if (A != RHS) 2755 std::swap(A, B); // smax(A, B) pred A. 2756 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2757 // We analyze this as smax(A, B) pred A. 2758 P = Pred; 2759 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) && 2760 (A == LHS || B == LHS)) { 2761 if (A != LHS) 2762 std::swap(A, B); // A pred smax(A, B). 2763 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2764 // We analyze this as smax(A, B) swapped-pred A. 2765 P = CmpInst::getSwappedPredicate(Pred); 2766 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2767 (A == RHS || B == RHS)) { 2768 if (A != RHS) 2769 std::swap(A, B); // smin(A, B) pred A. 2770 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2771 // We analyze this as smax(-A, -B) swapped-pred -A. 2772 // Note that we do not need to actually form -A or -B thanks to EqP. 2773 P = CmpInst::getSwappedPredicate(Pred); 2774 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) && 2775 (A == LHS || B == LHS)) { 2776 if (A != LHS) 2777 std::swap(A, B); // A pred smin(A, B). 2778 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2779 // We analyze this as smax(-A, -B) pred -A. 2780 // Note that we do not need to actually form -A or -B thanks to EqP. 2781 P = Pred; 2782 } 2783 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2784 // Cases correspond to "max(A, B) p A". 2785 switch (P) { 2786 default: 2787 break; 2788 case CmpInst::ICMP_EQ: 2789 case CmpInst::ICMP_SLE: 2790 // Equivalent to "A EqP B". This may be the same as the condition tested 2791 // in the max/min; if so, we can just return that. 2792 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2793 return V; 2794 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2795 return V; 2796 // Otherwise, see if "A EqP B" simplifies. 2797 if (MaxRecurse) 2798 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1)) 2799 return V; 2800 break; 2801 case CmpInst::ICMP_NE: 2802 case CmpInst::ICMP_SGT: { 2803 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2804 // Equivalent to "A InvEqP B". This may be the same as the condition 2805 // tested in the max/min; if so, we can just return that. 2806 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2807 return V; 2808 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2809 return V; 2810 // Otherwise, see if "A InvEqP B" simplifies. 2811 if (MaxRecurse) 2812 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1)) 2813 return V; 2814 break; 2815 } 2816 case CmpInst::ICMP_SGE: 2817 // Always true. 2818 return getTrue(ITy); 2819 case CmpInst::ICMP_SLT: 2820 // Always false. 2821 return getFalse(ITy); 2822 } 2823 } 2824 2825 // Unsigned variants on "max(a,b)>=a -> true". 2826 P = CmpInst::BAD_ICMP_PREDICATE; 2827 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2828 if (A != RHS) 2829 std::swap(A, B); // umax(A, B) pred A. 2830 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2831 // We analyze this as umax(A, B) pred A. 2832 P = Pred; 2833 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) && 2834 (A == LHS || B == LHS)) { 2835 if (A != LHS) 2836 std::swap(A, B); // A pred umax(A, B). 2837 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2838 // We analyze this as umax(A, B) swapped-pred A. 2839 P = CmpInst::getSwappedPredicate(Pred); 2840 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2841 (A == RHS || B == RHS)) { 2842 if (A != RHS) 2843 std::swap(A, B); // umin(A, B) pred A. 2844 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2845 // We analyze this as umax(-A, -B) swapped-pred -A. 2846 // Note that we do not need to actually form -A or -B thanks to EqP. 2847 P = CmpInst::getSwappedPredicate(Pred); 2848 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) && 2849 (A == LHS || B == LHS)) { 2850 if (A != LHS) 2851 std::swap(A, B); // A pred umin(A, B). 2852 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2853 // We analyze this as umax(-A, -B) pred -A. 2854 // Note that we do not need to actually form -A or -B thanks to EqP. 2855 P = Pred; 2856 } 2857 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2858 // Cases correspond to "max(A, B) p A". 2859 switch (P) { 2860 default: 2861 break; 2862 case CmpInst::ICMP_EQ: 2863 case CmpInst::ICMP_ULE: 2864 // Equivalent to "A EqP B". This may be the same as the condition tested 2865 // in the max/min; if so, we can just return that. 2866 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2867 return V; 2868 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2869 return V; 2870 // Otherwise, see if "A EqP B" simplifies. 2871 if (MaxRecurse) 2872 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1)) 2873 return V; 2874 break; 2875 case CmpInst::ICMP_NE: 2876 case CmpInst::ICMP_UGT: { 2877 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2878 // Equivalent to "A InvEqP B". This may be the same as the condition 2879 // tested in the max/min; if so, we can just return that. 2880 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2881 return V; 2882 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2883 return V; 2884 // Otherwise, see if "A InvEqP B" simplifies. 2885 if (MaxRecurse) 2886 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1)) 2887 return V; 2888 break; 2889 } 2890 case CmpInst::ICMP_UGE: 2891 // Always true. 2892 return getTrue(ITy); 2893 case CmpInst::ICMP_ULT: 2894 // Always false. 2895 return getFalse(ITy); 2896 } 2897 } 2898 2899 // Variants on "max(x,y) >= min(x,z)". 2900 Value *C, *D; 2901 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && 2902 match(RHS, m_SMin(m_Value(C), m_Value(D))) && 2903 (A == C || A == D || B == C || B == D)) { 2904 // max(x, ?) pred min(x, ?). 2905 if (Pred == CmpInst::ICMP_SGE) 2906 // Always true. 2907 return getTrue(ITy); 2908 if (Pred == CmpInst::ICMP_SLT) 2909 // Always false. 2910 return getFalse(ITy); 2911 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2912 match(RHS, m_SMax(m_Value(C), m_Value(D))) && 2913 (A == C || A == D || B == C || B == D)) { 2914 // min(x, ?) pred max(x, ?). 2915 if (Pred == CmpInst::ICMP_SLE) 2916 // Always true. 2917 return getTrue(ITy); 2918 if (Pred == CmpInst::ICMP_SGT) 2919 // Always false. 2920 return getFalse(ITy); 2921 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && 2922 match(RHS, m_UMin(m_Value(C), m_Value(D))) && 2923 (A == C || A == D || B == C || B == D)) { 2924 // max(x, ?) pred min(x, ?). 2925 if (Pred == CmpInst::ICMP_UGE) 2926 // Always true. 2927 return getTrue(ITy); 2928 if (Pred == CmpInst::ICMP_ULT) 2929 // Always false. 2930 return getFalse(ITy); 2931 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2932 match(RHS, m_UMax(m_Value(C), m_Value(D))) && 2933 (A == C || A == D || B == C || B == D)) { 2934 // min(x, ?) pred max(x, ?). 2935 if (Pred == CmpInst::ICMP_ULE) 2936 // Always true. 2937 return getTrue(ITy); 2938 if (Pred == CmpInst::ICMP_UGT) 2939 // Always false. 2940 return getFalse(ITy); 2941 } 2942 2943 return nullptr; 2944 } 2945 2946 /// Given operands for an ICmpInst, see if we can fold the result. 2947 /// If not, this returns null. 2948 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2949 const Query &Q, unsigned MaxRecurse) { 2950 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 2951 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!"); 2952 2953 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 2954 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 2955 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 2956 2957 // If we have a constant, make sure it is on the RHS. 2958 std::swap(LHS, RHS); 2959 Pred = CmpInst::getSwappedPredicate(Pred); 2960 } 2961 2962 Type *ITy = GetCompareTy(LHS); // The return type. 2963 2964 // icmp X, X -> true/false 2965 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false 2966 // because X could be 0. 2967 if (LHS == RHS || isa<UndefValue>(RHS)) 2968 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred)); 2969 2970 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q)) 2971 return V; 2972 2973 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q)) 2974 return V; 2975 2976 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS)) 2977 return V; 2978 2979 // If both operands have range metadata, use the metadata 2980 // to simplify the comparison. 2981 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) { 2982 auto RHS_Instr = dyn_cast<Instruction>(RHS); 2983 auto LHS_Instr = dyn_cast<Instruction>(LHS); 2984 2985 if (RHS_Instr->getMetadata(LLVMContext::MD_range) && 2986 LHS_Instr->getMetadata(LLVMContext::MD_range)) { 2987 auto RHS_CR = getConstantRangeFromMetadata( 2988 *RHS_Instr->getMetadata(LLVMContext::MD_range)); 2989 auto LHS_CR = getConstantRangeFromMetadata( 2990 *LHS_Instr->getMetadata(LLVMContext::MD_range)); 2991 2992 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR); 2993 if (Satisfied_CR.contains(LHS_CR)) 2994 return ConstantInt::getTrue(RHS->getContext()); 2995 2996 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion( 2997 CmpInst::getInversePredicate(Pred), RHS_CR); 2998 if (InversedSatisfied_CR.contains(LHS_CR)) 2999 return ConstantInt::getFalse(RHS->getContext()); 3000 } 3001 } 3002 3003 // Compare of cast, for example (zext X) != 0 -> X != 0 3004 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) { 3005 Instruction *LI = cast<CastInst>(LHS); 3006 Value *SrcOp = LI->getOperand(0); 3007 Type *SrcTy = SrcOp->getType(); 3008 Type *DstTy = LI->getType(); 3009 3010 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input 3011 // if the integer type is the same size as the pointer type. 3012 if (MaxRecurse && isa<PtrToIntInst>(LI) && 3013 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) { 3014 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 3015 // Transfer the cast to the constant. 3016 if (Value *V = SimplifyICmpInst(Pred, SrcOp, 3017 ConstantExpr::getIntToPtr(RHSC, SrcTy), 3018 Q, MaxRecurse-1)) 3019 return V; 3020 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) { 3021 if (RI->getOperand(0)->getType() == SrcTy) 3022 // Compare without the cast. 3023 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 3024 Q, MaxRecurse-1)) 3025 return V; 3026 } 3027 } 3028 3029 if (isa<ZExtInst>(LHS)) { 3030 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the 3031 // same type. 3032 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) { 3033 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 3034 // Compare X and Y. Note that signed predicates become unsigned. 3035 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 3036 SrcOp, RI->getOperand(0), Q, 3037 MaxRecurse-1)) 3038 return V; 3039 } 3040 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended 3041 // too. If not, then try to deduce the result of the comparison. 3042 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 3043 // Compute the constant that would happen if we truncated to SrcTy then 3044 // reextended to DstTy. 3045 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 3046 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy); 3047 3048 // If the re-extended constant didn't change then this is effectively 3049 // also a case of comparing two zero-extended values. 3050 if (RExt == CI && MaxRecurse) 3051 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 3052 SrcOp, Trunc, Q, MaxRecurse-1)) 3053 return V; 3054 3055 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit 3056 // there. Use this to work out the result of the comparison. 3057 if (RExt != CI) { 3058 switch (Pred) { 3059 default: llvm_unreachable("Unknown ICmp predicate!"); 3060 // LHS <u RHS. 3061 case ICmpInst::ICMP_EQ: 3062 case ICmpInst::ICMP_UGT: 3063 case ICmpInst::ICMP_UGE: 3064 return ConstantInt::getFalse(CI->getContext()); 3065 3066 case ICmpInst::ICMP_NE: 3067 case ICmpInst::ICMP_ULT: 3068 case ICmpInst::ICMP_ULE: 3069 return ConstantInt::getTrue(CI->getContext()); 3070 3071 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS 3072 // is non-negative then LHS <s RHS. 3073 case ICmpInst::ICMP_SGT: 3074 case ICmpInst::ICMP_SGE: 3075 return CI->getValue().isNegative() ? 3076 ConstantInt::getTrue(CI->getContext()) : 3077 ConstantInt::getFalse(CI->getContext()); 3078 3079 case ICmpInst::ICMP_SLT: 3080 case ICmpInst::ICMP_SLE: 3081 return CI->getValue().isNegative() ? 3082 ConstantInt::getFalse(CI->getContext()) : 3083 ConstantInt::getTrue(CI->getContext()); 3084 } 3085 } 3086 } 3087 } 3088 3089 if (isa<SExtInst>(LHS)) { 3090 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the 3091 // same type. 3092 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) { 3093 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 3094 // Compare X and Y. Note that the predicate does not change. 3095 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 3096 Q, MaxRecurse-1)) 3097 return V; 3098 } 3099 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended 3100 // too. If not, then try to deduce the result of the comparison. 3101 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 3102 // Compute the constant that would happen if we truncated to SrcTy then 3103 // reextended to DstTy. 3104 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 3105 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy); 3106 3107 // If the re-extended constant didn't change then this is effectively 3108 // also a case of comparing two sign-extended values. 3109 if (RExt == CI && MaxRecurse) 3110 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1)) 3111 return V; 3112 3113 // Otherwise the upper bits of LHS are all equal, while RHS has varying 3114 // bits there. Use this to work out the result of the comparison. 3115 if (RExt != CI) { 3116 switch (Pred) { 3117 default: llvm_unreachable("Unknown ICmp predicate!"); 3118 case ICmpInst::ICMP_EQ: 3119 return ConstantInt::getFalse(CI->getContext()); 3120 case ICmpInst::ICMP_NE: 3121 return ConstantInt::getTrue(CI->getContext()); 3122 3123 // If RHS is non-negative then LHS <s RHS. If RHS is negative then 3124 // LHS >s RHS. 3125 case ICmpInst::ICMP_SGT: 3126 case ICmpInst::ICMP_SGE: 3127 return CI->getValue().isNegative() ? 3128 ConstantInt::getTrue(CI->getContext()) : 3129 ConstantInt::getFalse(CI->getContext()); 3130 case ICmpInst::ICMP_SLT: 3131 case ICmpInst::ICMP_SLE: 3132 return CI->getValue().isNegative() ? 3133 ConstantInt::getFalse(CI->getContext()) : 3134 ConstantInt::getTrue(CI->getContext()); 3135 3136 // If LHS is non-negative then LHS <u RHS. If LHS is negative then 3137 // LHS >u RHS. 3138 case ICmpInst::ICMP_UGT: 3139 case ICmpInst::ICMP_UGE: 3140 // Comparison is true iff the LHS <s 0. 3141 if (MaxRecurse) 3142 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp, 3143 Constant::getNullValue(SrcTy), 3144 Q, MaxRecurse-1)) 3145 return V; 3146 break; 3147 case ICmpInst::ICMP_ULT: 3148 case ICmpInst::ICMP_ULE: 3149 // Comparison is true iff the LHS >=s 0. 3150 if (MaxRecurse) 3151 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp, 3152 Constant::getNullValue(SrcTy), 3153 Q, MaxRecurse-1)) 3154 return V; 3155 break; 3156 } 3157 } 3158 } 3159 } 3160 } 3161 3162 // icmp eq|ne X, Y -> false|true if X != Y 3163 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) && 3164 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) { 3165 LLVMContext &Ctx = LHS->getType()->getContext(); 3166 return Pred == ICmpInst::ICMP_NE ? 3167 ConstantInt::getTrue(Ctx) : ConstantInt::getFalse(Ctx); 3168 } 3169 3170 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse)) 3171 return V; 3172 3173 if (Value *V = simplifyMinMax(Pred, LHS, RHS, Q, MaxRecurse)) 3174 return V; 3175 3176 // Simplify comparisons of related pointers using a powerful, recursive 3177 // GEP-walk when we have target data available.. 3178 if (LHS->getType()->isPointerTy()) 3179 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, LHS, RHS)) 3180 return C; 3181 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS)) 3182 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS)) 3183 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) == 3184 Q.DL.getTypeSizeInBits(CLHS->getType()) && 3185 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) == 3186 Q.DL.getTypeSizeInBits(CRHS->getType())) 3187 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, 3188 CLHS->getPointerOperand(), 3189 CRHS->getPointerOperand())) 3190 return C; 3191 3192 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) { 3193 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) { 3194 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() && 3195 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() && 3196 (ICmpInst::isEquality(Pred) || 3197 (GLHS->isInBounds() && GRHS->isInBounds() && 3198 Pred == ICmpInst::getSignedPredicate(Pred)))) { 3199 // The bases are equal and the indices are constant. Build a constant 3200 // expression GEP with the same indices and a null base pointer to see 3201 // what constant folding can make out of it. 3202 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType()); 3203 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end()); 3204 Constant *NewLHS = ConstantExpr::getGetElementPtr( 3205 GLHS->getSourceElementType(), Null, IndicesLHS); 3206 3207 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end()); 3208 Constant *NewRHS = ConstantExpr::getGetElementPtr( 3209 GLHS->getSourceElementType(), Null, IndicesRHS); 3210 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS); 3211 } 3212 } 3213 } 3214 3215 // If a bit is known to be zero for A and known to be one for B, 3216 // then A and B cannot be equal. 3217 if (ICmpInst::isEquality(Pred)) { 3218 const APInt *RHSVal; 3219 if (match(RHS, m_APInt(RHSVal))) { 3220 unsigned BitWidth = RHSVal->getBitWidth(); 3221 APInt LHSKnownZero(BitWidth, 0); 3222 APInt LHSKnownOne(BitWidth, 0); 3223 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC, 3224 Q.CxtI, Q.DT); 3225 if (((LHSKnownZero & *RHSVal) != 0) || ((LHSKnownOne & ~(*RHSVal)) != 0)) 3226 return Pred == ICmpInst::ICMP_EQ ? ConstantInt::getFalse(ITy) 3227 : ConstantInt::getTrue(ITy); 3228 } 3229 } 3230 3231 // If the comparison is with the result of a select instruction, check whether 3232 // comparing with either branch of the select always yields the same value. 3233 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3234 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3235 return V; 3236 3237 // If the comparison is with the result of a phi instruction, check whether 3238 // doing the compare with each incoming phi value yields a common result. 3239 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3240 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3241 return V; 3242 3243 return nullptr; 3244 } 3245 3246 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3247 const DataLayout &DL, 3248 const TargetLibraryInfo *TLI, 3249 const DominatorTree *DT, AssumptionCache *AC, 3250 const Instruction *CxtI) { 3251 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3252 RecursionLimit); 3253 } 3254 3255 /// Given operands for an FCmpInst, see if we can fold the result. 3256 /// If not, this returns null. 3257 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3258 FastMathFlags FMF, const Query &Q, 3259 unsigned MaxRecurse) { 3260 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 3261 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!"); 3262 3263 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 3264 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 3265 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 3266 3267 // If we have a constant, make sure it is on the RHS. 3268 std::swap(LHS, RHS); 3269 Pred = CmpInst::getSwappedPredicate(Pred); 3270 } 3271 3272 // Fold trivial predicates. 3273 Type *RetTy = GetCompareTy(LHS); 3274 if (Pred == FCmpInst::FCMP_FALSE) 3275 return getFalse(RetTy); 3276 if (Pred == FCmpInst::FCMP_TRUE) 3277 return getTrue(RetTy); 3278 3279 // UNO/ORD predicates can be trivially folded if NaNs are ignored. 3280 if (FMF.noNaNs()) { 3281 if (Pred == FCmpInst::FCMP_UNO) 3282 return getFalse(RetTy); 3283 if (Pred == FCmpInst::FCMP_ORD) 3284 return getTrue(RetTy); 3285 } 3286 3287 // fcmp pred x, undef and fcmp pred undef, x 3288 // fold to true if unordered, false if ordered 3289 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) { 3290 // Choosing NaN for the undef will always make unordered comparison succeed 3291 // and ordered comparison fail. 3292 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred)); 3293 } 3294 3295 // fcmp x,x -> true/false. Not all compares are foldable. 3296 if (LHS == RHS) { 3297 if (CmpInst::isTrueWhenEqual(Pred)) 3298 return getTrue(RetTy); 3299 if (CmpInst::isFalseWhenEqual(Pred)) 3300 return getFalse(RetTy); 3301 } 3302 3303 // Handle fcmp with constant RHS 3304 const ConstantFP *CFP = nullptr; 3305 if (const auto *RHSC = dyn_cast<Constant>(RHS)) { 3306 if (RHS->getType()->isVectorTy()) 3307 CFP = dyn_cast_or_null<ConstantFP>(RHSC->getSplatValue()); 3308 else 3309 CFP = dyn_cast<ConstantFP>(RHSC); 3310 } 3311 if (CFP) { 3312 // If the constant is a nan, see if we can fold the comparison based on it. 3313 if (CFP->getValueAPF().isNaN()) { 3314 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo" 3315 return getFalse(RetTy); 3316 assert(FCmpInst::isUnordered(Pred) && 3317 "Comparison must be either ordered or unordered!"); 3318 // True if unordered. 3319 return getTrue(RetTy); 3320 } 3321 // Check whether the constant is an infinity. 3322 if (CFP->getValueAPF().isInfinity()) { 3323 if (CFP->getValueAPF().isNegative()) { 3324 switch (Pred) { 3325 case FCmpInst::FCMP_OLT: 3326 // No value is ordered and less than negative infinity. 3327 return getFalse(RetTy); 3328 case FCmpInst::FCMP_UGE: 3329 // All values are unordered with or at least negative infinity. 3330 return getTrue(RetTy); 3331 default: 3332 break; 3333 } 3334 } else { 3335 switch (Pred) { 3336 case FCmpInst::FCMP_OGT: 3337 // No value is ordered and greater than infinity. 3338 return getFalse(RetTy); 3339 case FCmpInst::FCMP_ULE: 3340 // All values are unordered with and at most infinity. 3341 return getTrue(RetTy); 3342 default: 3343 break; 3344 } 3345 } 3346 } 3347 if (CFP->getValueAPF().isZero()) { 3348 switch (Pred) { 3349 case FCmpInst::FCMP_UGE: 3350 if (CannotBeOrderedLessThanZero(LHS, Q.TLI)) 3351 return getTrue(RetTy); 3352 break; 3353 case FCmpInst::FCMP_OLT: 3354 // X < 0 3355 if (CannotBeOrderedLessThanZero(LHS, Q.TLI)) 3356 return getFalse(RetTy); 3357 break; 3358 default: 3359 break; 3360 } 3361 } 3362 } 3363 3364 // If the comparison is with the result of a select instruction, check whether 3365 // comparing with either branch of the select always yields the same value. 3366 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3367 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3368 return V; 3369 3370 // If the comparison is with the result of a phi instruction, check whether 3371 // doing the compare with each incoming phi value yields a common result. 3372 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3373 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3374 return V; 3375 3376 return nullptr; 3377 } 3378 3379 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3380 FastMathFlags FMF, const DataLayout &DL, 3381 const TargetLibraryInfo *TLI, 3382 const DominatorTree *DT, AssumptionCache *AC, 3383 const Instruction *CxtI) { 3384 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, 3385 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3386 } 3387 3388 /// See if V simplifies when its operand Op is replaced with RepOp. 3389 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp, 3390 const Query &Q, 3391 unsigned MaxRecurse) { 3392 // Trivial replacement. 3393 if (V == Op) 3394 return RepOp; 3395 3396 auto *I = dyn_cast<Instruction>(V); 3397 if (!I) 3398 return nullptr; 3399 3400 // If this is a binary operator, try to simplify it with the replaced op. 3401 if (auto *B = dyn_cast<BinaryOperator>(I)) { 3402 // Consider: 3403 // %cmp = icmp eq i32 %x, 2147483647 3404 // %add = add nsw i32 %x, 1 3405 // %sel = select i1 %cmp, i32 -2147483648, i32 %add 3406 // 3407 // We can't replace %sel with %add unless we strip away the flags. 3408 if (isa<OverflowingBinaryOperator>(B)) 3409 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap()) 3410 return nullptr; 3411 if (isa<PossiblyExactOperator>(B)) 3412 if (B->isExact()) 3413 return nullptr; 3414 3415 if (MaxRecurse) { 3416 if (B->getOperand(0) == Op) 3417 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q, 3418 MaxRecurse - 1); 3419 if (B->getOperand(1) == Op) 3420 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q, 3421 MaxRecurse - 1); 3422 } 3423 } 3424 3425 // Same for CmpInsts. 3426 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 3427 if (MaxRecurse) { 3428 if (C->getOperand(0) == Op) 3429 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q, 3430 MaxRecurse - 1); 3431 if (C->getOperand(1) == Op) 3432 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q, 3433 MaxRecurse - 1); 3434 } 3435 } 3436 3437 // TODO: We could hand off more cases to instsimplify here. 3438 3439 // If all operands are constant after substituting Op for RepOp then we can 3440 // constant fold the instruction. 3441 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) { 3442 // Build a list of all constant operands. 3443 SmallVector<Constant *, 8> ConstOps; 3444 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 3445 if (I->getOperand(i) == Op) 3446 ConstOps.push_back(CRepOp); 3447 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i))) 3448 ConstOps.push_back(COp); 3449 else 3450 break; 3451 } 3452 3453 // All operands were constants, fold it. 3454 if (ConstOps.size() == I->getNumOperands()) { 3455 if (CmpInst *C = dyn_cast<CmpInst>(I)) 3456 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0], 3457 ConstOps[1], Q.DL, Q.TLI); 3458 3459 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 3460 if (!LI->isVolatile()) 3461 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL); 3462 3463 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI); 3464 } 3465 } 3466 3467 return nullptr; 3468 } 3469 3470 /// Try to simplify a select instruction when its condition operand is an 3471 /// integer comparison where one operand of the compare is a constant. 3472 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X, 3473 const APInt *Y, bool TrueWhenUnset) { 3474 const APInt *C; 3475 3476 // (X & Y) == 0 ? X & ~Y : X --> X 3477 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y 3478 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) && 3479 *Y == ~*C) 3480 return TrueWhenUnset ? FalseVal : TrueVal; 3481 3482 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y 3483 // (X & Y) != 0 ? X : X & ~Y --> X 3484 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) && 3485 *Y == ~*C) 3486 return TrueWhenUnset ? FalseVal : TrueVal; 3487 3488 if (Y->isPowerOf2()) { 3489 // (X & Y) == 0 ? X | Y : X --> X | Y 3490 // (X & Y) != 0 ? X | Y : X --> X 3491 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) && 3492 *Y == *C) 3493 return TrueWhenUnset ? TrueVal : FalseVal; 3494 3495 // (X & Y) == 0 ? X : X | Y --> X 3496 // (X & Y) != 0 ? X : X | Y --> X | Y 3497 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) && 3498 *Y == *C) 3499 return TrueWhenUnset ? TrueVal : FalseVal; 3500 } 3501 3502 return nullptr; 3503 } 3504 3505 /// An alternative way to test if a bit is set or not uses sgt/slt instead of 3506 /// eq/ne. 3507 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *TrueVal, 3508 Value *FalseVal, 3509 bool TrueWhenUnset) { 3510 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); 3511 if (!BitWidth) 3512 return nullptr; 3513 3514 APInt MinSignedValue; 3515 Value *X; 3516 if (match(CmpLHS, m_Trunc(m_Value(X))) && (X == TrueVal || X == FalseVal)) { 3517 // icmp slt (trunc X), 0 <--> icmp ne (and X, C), 0 3518 // icmp sgt (trunc X), -1 <--> icmp eq (and X, C), 0 3519 unsigned DestSize = CmpLHS->getType()->getScalarSizeInBits(); 3520 MinSignedValue = APInt::getSignedMinValue(DestSize).zext(BitWidth); 3521 } else { 3522 // icmp slt X, 0 <--> icmp ne (and X, C), 0 3523 // icmp sgt X, -1 <--> icmp eq (and X, C), 0 3524 X = CmpLHS; 3525 MinSignedValue = APInt::getSignedMinValue(BitWidth); 3526 } 3527 3528 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, &MinSignedValue, 3529 TrueWhenUnset)) 3530 return V; 3531 3532 return nullptr; 3533 } 3534 3535 /// Try to simplify a select instruction when its condition operand is an 3536 /// integer comparison. 3537 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal, 3538 Value *FalseVal, const Query &Q, 3539 unsigned MaxRecurse) { 3540 ICmpInst::Predicate Pred; 3541 Value *CmpLHS, *CmpRHS; 3542 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)))) 3543 return nullptr; 3544 3545 // FIXME: This code is nearly duplicated in InstCombine. Using/refactoring 3546 // decomposeBitTestICmp() might help. 3547 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) { 3548 Value *X; 3549 const APInt *Y; 3550 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y)))) 3551 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y, 3552 Pred == ICmpInst::ICMP_EQ)) 3553 return V; 3554 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { 3555 // Comparing signed-less-than 0 checks if the sign bit is set. 3556 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal, 3557 false)) 3558 return V; 3559 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { 3560 // Comparing signed-greater-than -1 checks if the sign bit is not set. 3561 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal, 3562 true)) 3563 return V; 3564 } 3565 3566 if (CondVal->hasOneUse()) { 3567 const APInt *C; 3568 if (match(CmpRHS, m_APInt(C))) { 3569 // X < MIN ? T : F --> F 3570 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue()) 3571 return FalseVal; 3572 // X < MIN ? T : F --> F 3573 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue()) 3574 return FalseVal; 3575 // X > MAX ? T : F --> F 3576 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue()) 3577 return FalseVal; 3578 // X > MAX ? T : F --> F 3579 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue()) 3580 return FalseVal; 3581 } 3582 } 3583 3584 // If we have an equality comparison, then we know the value in one of the 3585 // arms of the select. See if substituting this value into the arm and 3586 // simplifying the result yields the same value as the other arm. 3587 if (Pred == ICmpInst::ICMP_EQ) { 3588 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3589 TrueVal || 3590 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3591 TrueVal) 3592 return FalseVal; 3593 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3594 FalseVal || 3595 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3596 FalseVal) 3597 return FalseVal; 3598 } else if (Pred == ICmpInst::ICMP_NE) { 3599 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3600 FalseVal || 3601 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3602 FalseVal) 3603 return TrueVal; 3604 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3605 TrueVal || 3606 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3607 TrueVal) 3608 return TrueVal; 3609 } 3610 3611 return nullptr; 3612 } 3613 3614 /// Given operands for a SelectInst, see if we can fold the result. 3615 /// If not, this returns null. 3616 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal, 3617 Value *FalseVal, const Query &Q, 3618 unsigned MaxRecurse) { 3619 // select true, X, Y -> X 3620 // select false, X, Y -> Y 3621 if (Constant *CB = dyn_cast<Constant>(CondVal)) { 3622 if (CB->isAllOnesValue()) 3623 return TrueVal; 3624 if (CB->isNullValue()) 3625 return FalseVal; 3626 } 3627 3628 // select C, X, X -> X 3629 if (TrueVal == FalseVal) 3630 return TrueVal; 3631 3632 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y 3633 if (isa<Constant>(TrueVal)) 3634 return TrueVal; 3635 return FalseVal; 3636 } 3637 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X 3638 return FalseVal; 3639 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X 3640 return TrueVal; 3641 3642 if (Value *V = 3643 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse)) 3644 return V; 3645 3646 return nullptr; 3647 } 3648 3649 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, 3650 const DataLayout &DL, 3651 const TargetLibraryInfo *TLI, 3652 const DominatorTree *DT, AssumptionCache *AC, 3653 const Instruction *CxtI) { 3654 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, 3655 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3656 } 3657 3658 /// Given operands for an GetElementPtrInst, see if we can fold the result. 3659 /// If not, this returns null. 3660 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, 3661 const Query &Q, unsigned) { 3662 // The type of the GEP pointer operand. 3663 unsigned AS = 3664 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace(); 3665 3666 // getelementptr P -> P. 3667 if (Ops.size() == 1) 3668 return Ops[0]; 3669 3670 // Compute the (pointer) type returned by the GEP instruction. 3671 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1)); 3672 Type *GEPTy = PointerType::get(LastType, AS); 3673 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType())) 3674 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 3675 3676 if (isa<UndefValue>(Ops[0])) 3677 return UndefValue::get(GEPTy); 3678 3679 if (Ops.size() == 2) { 3680 // getelementptr P, 0 -> P. 3681 if (match(Ops[1], m_Zero())) 3682 return Ops[0]; 3683 3684 Type *Ty = SrcTy; 3685 if (Ty->isSized()) { 3686 Value *P; 3687 uint64_t C; 3688 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty); 3689 // getelementptr P, N -> P if P points to a type of zero size. 3690 if (TyAllocSize == 0) 3691 return Ops[0]; 3692 3693 // The following transforms are only safe if the ptrtoint cast 3694 // doesn't truncate the pointers. 3695 if (Ops[1]->getType()->getScalarSizeInBits() == 3696 Q.DL.getPointerSizeInBits(AS)) { 3697 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * { 3698 if (match(P, m_Zero())) 3699 return Constant::getNullValue(GEPTy); 3700 Value *Temp; 3701 if (match(P, m_PtrToInt(m_Value(Temp)))) 3702 if (Temp->getType() == GEPTy) 3703 return Temp; 3704 return nullptr; 3705 }; 3706 3707 // getelementptr V, (sub P, V) -> P if P points to a type of size 1. 3708 if (TyAllocSize == 1 && 3709 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))))) 3710 if (Value *R = PtrToIntOrZero(P)) 3711 return R; 3712 3713 // getelementptr V, (ashr (sub P, V), C) -> Q 3714 // if P points to a type of size 1 << C. 3715 if (match(Ops[1], 3716 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3717 m_ConstantInt(C))) && 3718 TyAllocSize == 1ULL << C) 3719 if (Value *R = PtrToIntOrZero(P)) 3720 return R; 3721 3722 // getelementptr V, (sdiv (sub P, V), C) -> Q 3723 // if P points to a type of size C. 3724 if (match(Ops[1], 3725 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3726 m_SpecificInt(TyAllocSize)))) 3727 if (Value *R = PtrToIntOrZero(P)) 3728 return R; 3729 } 3730 } 3731 } 3732 3733 if (Q.DL.getTypeAllocSize(LastType) == 1 && 3734 all_of(Ops.slice(1).drop_back(1), 3735 [](Value *Idx) { return match(Idx, m_Zero()); })) { 3736 unsigned PtrWidth = 3737 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace()); 3738 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) { 3739 APInt BasePtrOffset(PtrWidth, 0); 3740 Value *StrippedBasePtr = 3741 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL, 3742 BasePtrOffset); 3743 3744 // gep (gep V, C), (sub 0, V) -> C 3745 if (match(Ops.back(), 3746 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) { 3747 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset); 3748 return ConstantExpr::getIntToPtr(CI, GEPTy); 3749 } 3750 // gep (gep V, C), (xor V, -1) -> C-1 3751 if (match(Ops.back(), 3752 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) { 3753 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1); 3754 return ConstantExpr::getIntToPtr(CI, GEPTy); 3755 } 3756 } 3757 } 3758 3759 // Check to see if this is constant foldable. 3760 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 3761 if (!isa<Constant>(Ops[i])) 3762 return nullptr; 3763 3764 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]), 3765 Ops.slice(1)); 3766 } 3767 3768 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, 3769 const DataLayout &DL, 3770 const TargetLibraryInfo *TLI, 3771 const DominatorTree *DT, AssumptionCache *AC, 3772 const Instruction *CxtI) { 3773 return ::SimplifyGEPInst(SrcTy, Ops, 3774 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3775 } 3776 3777 /// Given operands for an InsertValueInst, see if we can fold the result. 3778 /// If not, this returns null. 3779 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val, 3780 ArrayRef<unsigned> Idxs, const Query &Q, 3781 unsigned) { 3782 if (Constant *CAgg = dyn_cast<Constant>(Agg)) 3783 if (Constant *CVal = dyn_cast<Constant>(Val)) 3784 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs); 3785 3786 // insertvalue x, undef, n -> x 3787 if (match(Val, m_Undef())) 3788 return Agg; 3789 3790 // insertvalue x, (extractvalue y, n), n 3791 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val)) 3792 if (EV->getAggregateOperand()->getType() == Agg->getType() && 3793 EV->getIndices() == Idxs) { 3794 // insertvalue undef, (extractvalue y, n), n -> y 3795 if (match(Agg, m_Undef())) 3796 return EV->getAggregateOperand(); 3797 3798 // insertvalue y, (extractvalue y, n), n -> y 3799 if (Agg == EV->getAggregateOperand()) 3800 return Agg; 3801 } 3802 3803 return nullptr; 3804 } 3805 3806 Value *llvm::SimplifyInsertValueInst( 3807 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout &DL, 3808 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC, 3809 const Instruction *CxtI) { 3810 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI), 3811 RecursionLimit); 3812 } 3813 3814 /// Given operands for an ExtractValueInst, see if we can fold the result. 3815 /// If not, this returns null. 3816 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, 3817 const Query &, unsigned) { 3818 if (auto *CAgg = dyn_cast<Constant>(Agg)) 3819 return ConstantFoldExtractValueInstruction(CAgg, Idxs); 3820 3821 // extractvalue x, (insertvalue y, elt, n), n -> elt 3822 unsigned NumIdxs = Idxs.size(); 3823 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr; 3824 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) { 3825 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices(); 3826 unsigned NumInsertValueIdxs = InsertValueIdxs.size(); 3827 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs); 3828 if (InsertValueIdxs.slice(0, NumCommonIdxs) == 3829 Idxs.slice(0, NumCommonIdxs)) { 3830 if (NumIdxs == NumInsertValueIdxs) 3831 return IVI->getInsertedValueOperand(); 3832 break; 3833 } 3834 } 3835 3836 return nullptr; 3837 } 3838 3839 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, 3840 const DataLayout &DL, 3841 const TargetLibraryInfo *TLI, 3842 const DominatorTree *DT, 3843 AssumptionCache *AC, 3844 const Instruction *CxtI) { 3845 return ::SimplifyExtractValueInst(Agg, Idxs, Query(DL, TLI, DT, AC, CxtI), 3846 RecursionLimit); 3847 } 3848 3849 /// Given operands for an ExtractElementInst, see if we can fold the result. 3850 /// If not, this returns null. 3851 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const Query &, 3852 unsigned) { 3853 if (auto *CVec = dyn_cast<Constant>(Vec)) { 3854 if (auto *CIdx = dyn_cast<Constant>(Idx)) 3855 return ConstantFoldExtractElementInstruction(CVec, CIdx); 3856 3857 // The index is not relevant if our vector is a splat. 3858 if (auto *Splat = CVec->getSplatValue()) 3859 return Splat; 3860 3861 if (isa<UndefValue>(Vec)) 3862 return UndefValue::get(Vec->getType()->getVectorElementType()); 3863 } 3864 3865 // If extracting a specified index from the vector, see if we can recursively 3866 // find a previously computed scalar that was inserted into the vector. 3867 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) 3868 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue())) 3869 return Elt; 3870 3871 return nullptr; 3872 } 3873 3874 Value *llvm::SimplifyExtractElementInst( 3875 Value *Vec, Value *Idx, const DataLayout &DL, const TargetLibraryInfo *TLI, 3876 const DominatorTree *DT, AssumptionCache *AC, const Instruction *CxtI) { 3877 return ::SimplifyExtractElementInst(Vec, Idx, Query(DL, TLI, DT, AC, CxtI), 3878 RecursionLimit); 3879 } 3880 3881 /// See if we can fold the given phi. If not, returns null. 3882 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) { 3883 // If all of the PHI's incoming values are the same then replace the PHI node 3884 // with the common value. 3885 Value *CommonValue = nullptr; 3886 bool HasUndefInput = false; 3887 for (Value *Incoming : PN->incoming_values()) { 3888 // If the incoming value is the phi node itself, it can safely be skipped. 3889 if (Incoming == PN) continue; 3890 if (isa<UndefValue>(Incoming)) { 3891 // Remember that we saw an undef value, but otherwise ignore them. 3892 HasUndefInput = true; 3893 continue; 3894 } 3895 if (CommonValue && Incoming != CommonValue) 3896 return nullptr; // Not the same, bail out. 3897 CommonValue = Incoming; 3898 } 3899 3900 // If CommonValue is null then all of the incoming values were either undef or 3901 // equal to the phi node itself. 3902 if (!CommonValue) 3903 return UndefValue::get(PN->getType()); 3904 3905 // If we have a PHI node like phi(X, undef, X), where X is defined by some 3906 // instruction, we cannot return X as the result of the PHI node unless it 3907 // dominates the PHI block. 3908 if (HasUndefInput) 3909 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr; 3910 3911 return CommonValue; 3912 } 3913 3914 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op, 3915 Type *Ty, const Query &Q, unsigned MaxRecurse) { 3916 if (auto *C = dyn_cast<Constant>(Op)) 3917 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL); 3918 3919 if (auto *CI = dyn_cast<CastInst>(Op)) { 3920 auto *Src = CI->getOperand(0); 3921 Type *SrcTy = Src->getType(); 3922 Type *MidTy = CI->getType(); 3923 Type *DstTy = Ty; 3924 if (Src->getType() == Ty) { 3925 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode()); 3926 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc); 3927 Type *SrcIntPtrTy = 3928 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr; 3929 Type *MidIntPtrTy = 3930 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr; 3931 Type *DstIntPtrTy = 3932 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr; 3933 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy, 3934 SrcIntPtrTy, MidIntPtrTy, 3935 DstIntPtrTy) == Instruction::BitCast) 3936 return Src; 3937 } 3938 } 3939 3940 // bitcast x -> x 3941 if (CastOpc == Instruction::BitCast) 3942 if (Op->getType() == Ty) 3943 return Op; 3944 3945 return nullptr; 3946 } 3947 3948 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty, 3949 const DataLayout &DL, 3950 const TargetLibraryInfo *TLI, 3951 const DominatorTree *DT, AssumptionCache *AC, 3952 const Instruction *CxtI) { 3953 return ::SimplifyCastInst(CastOpc, Op, Ty, Query(DL, TLI, DT, AC, CxtI), 3954 RecursionLimit); 3955 } 3956 3957 //=== Helper functions for higher up the class hierarchy. 3958 3959 /// Given operands for a BinaryOperator, see if we can fold the result. 3960 /// If not, this returns null. 3961 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3962 const Query &Q, unsigned MaxRecurse) { 3963 switch (Opcode) { 3964 case Instruction::Add: 3965 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3966 Q, MaxRecurse); 3967 case Instruction::FAdd: 3968 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3969 3970 case Instruction::Sub: 3971 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3972 Q, MaxRecurse); 3973 case Instruction::FSub: 3974 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3975 3976 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse); 3977 case Instruction::FMul: 3978 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3979 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse); 3980 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse); 3981 case Instruction::FDiv: 3982 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3983 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse); 3984 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse); 3985 case Instruction::FRem: 3986 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3987 case Instruction::Shl: 3988 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3989 Q, MaxRecurse); 3990 case Instruction::LShr: 3991 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3992 case Instruction::AShr: 3993 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3994 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse); 3995 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse); 3996 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse); 3997 default: 3998 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 3999 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 4000 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL); 4001 4002 // If the operation is associative, try some generic simplifications. 4003 if (Instruction::isAssociative(Opcode)) 4004 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse)) 4005 return V; 4006 4007 // If the operation is with the result of a select instruction check whether 4008 // operating on either branch of the select always yields the same value. 4009 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 4010 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse)) 4011 return V; 4012 4013 // If the operation is with the result of a phi instruction, check whether 4014 // operating on all incoming values of the phi always yields the same value. 4015 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 4016 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse)) 4017 return V; 4018 4019 return nullptr; 4020 } 4021 } 4022 4023 /// Given operands for a BinaryOperator, see if we can fold the result. 4024 /// If not, this returns null. 4025 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the 4026 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp. 4027 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 4028 const FastMathFlags &FMF, const Query &Q, 4029 unsigned MaxRecurse) { 4030 switch (Opcode) { 4031 case Instruction::FAdd: 4032 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse); 4033 case Instruction::FSub: 4034 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse); 4035 case Instruction::FMul: 4036 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse); 4037 default: 4038 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse); 4039 } 4040 } 4041 4042 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 4043 const DataLayout &DL, const TargetLibraryInfo *TLI, 4044 const DominatorTree *DT, AssumptionCache *AC, 4045 const Instruction *CxtI) { 4046 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 4047 RecursionLimit); 4048 } 4049 4050 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 4051 const FastMathFlags &FMF, const DataLayout &DL, 4052 const TargetLibraryInfo *TLI, 4053 const DominatorTree *DT, AssumptionCache *AC, 4054 const Instruction *CxtI) { 4055 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI), 4056 RecursionLimit); 4057 } 4058 4059 /// Given operands for a CmpInst, see if we can fold the result. 4060 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 4061 const Query &Q, unsigned MaxRecurse) { 4062 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate)) 4063 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 4064 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse); 4065 } 4066 4067 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 4068 const DataLayout &DL, const TargetLibraryInfo *TLI, 4069 const DominatorTree *DT, AssumptionCache *AC, 4070 const Instruction *CxtI) { 4071 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 4072 RecursionLimit); 4073 } 4074 4075 static bool IsIdempotent(Intrinsic::ID ID) { 4076 switch (ID) { 4077 default: return false; 4078 4079 // Unary idempotent: f(f(x)) = f(x) 4080 case Intrinsic::fabs: 4081 case Intrinsic::floor: 4082 case Intrinsic::ceil: 4083 case Intrinsic::trunc: 4084 case Intrinsic::rint: 4085 case Intrinsic::nearbyint: 4086 case Intrinsic::round: 4087 return true; 4088 } 4089 } 4090 4091 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset, 4092 const DataLayout &DL) { 4093 GlobalValue *PtrSym; 4094 APInt PtrOffset; 4095 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL)) 4096 return nullptr; 4097 4098 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext()); 4099 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext()); 4100 Type *Int32PtrTy = Int32Ty->getPointerTo(); 4101 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext()); 4102 4103 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset); 4104 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64) 4105 return nullptr; 4106 4107 uint64_t OffsetInt = OffsetConstInt->getSExtValue(); 4108 if (OffsetInt % 4 != 0) 4109 return nullptr; 4110 4111 Constant *C = ConstantExpr::getGetElementPtr( 4112 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy), 4113 ConstantInt::get(Int64Ty, OffsetInt / 4)); 4114 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL); 4115 if (!Loaded) 4116 return nullptr; 4117 4118 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded); 4119 if (!LoadedCE) 4120 return nullptr; 4121 4122 if (LoadedCE->getOpcode() == Instruction::Trunc) { 4123 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0)); 4124 if (!LoadedCE) 4125 return nullptr; 4126 } 4127 4128 if (LoadedCE->getOpcode() != Instruction::Sub) 4129 return nullptr; 4130 4131 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0)); 4132 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt) 4133 return nullptr; 4134 auto *LoadedLHSPtr = LoadedLHS->getOperand(0); 4135 4136 Constant *LoadedRHS = LoadedCE->getOperand(1); 4137 GlobalValue *LoadedRHSSym; 4138 APInt LoadedRHSOffset; 4139 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset, 4140 DL) || 4141 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset) 4142 return nullptr; 4143 4144 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy); 4145 } 4146 4147 static bool maskIsAllZeroOrUndef(Value *Mask) { 4148 auto *ConstMask = dyn_cast<Constant>(Mask); 4149 if (!ConstMask) 4150 return false; 4151 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask)) 4152 return true; 4153 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E; 4154 ++I) { 4155 if (auto *MaskElt = ConstMask->getAggregateElement(I)) 4156 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt)) 4157 continue; 4158 return false; 4159 } 4160 return true; 4161 } 4162 4163 template <typename IterTy> 4164 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd, 4165 const Query &Q, unsigned MaxRecurse) { 4166 Intrinsic::ID IID = F->getIntrinsicID(); 4167 unsigned NumOperands = std::distance(ArgBegin, ArgEnd); 4168 Type *ReturnType = F->getReturnType(); 4169 4170 // Binary Ops 4171 if (NumOperands == 2) { 4172 Value *LHS = *ArgBegin; 4173 Value *RHS = *(ArgBegin + 1); 4174 if (IID == Intrinsic::usub_with_overflow || 4175 IID == Intrinsic::ssub_with_overflow) { 4176 // X - X -> { 0, false } 4177 if (LHS == RHS) 4178 return Constant::getNullValue(ReturnType); 4179 4180 // X - undef -> undef 4181 // undef - X -> undef 4182 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) 4183 return UndefValue::get(ReturnType); 4184 } 4185 4186 if (IID == Intrinsic::uadd_with_overflow || 4187 IID == Intrinsic::sadd_with_overflow) { 4188 // X + undef -> undef 4189 if (isa<UndefValue>(RHS)) 4190 return UndefValue::get(ReturnType); 4191 } 4192 4193 if (IID == Intrinsic::umul_with_overflow || 4194 IID == Intrinsic::smul_with_overflow) { 4195 // X * 0 -> { 0, false } 4196 if (match(RHS, m_Zero())) 4197 return Constant::getNullValue(ReturnType); 4198 4199 // X * undef -> { 0, false } 4200 if (match(RHS, m_Undef())) 4201 return Constant::getNullValue(ReturnType); 4202 } 4203 4204 if (IID == Intrinsic::load_relative && isa<Constant>(LHS) && 4205 isa<Constant>(RHS)) 4206 return SimplifyRelativeLoad(cast<Constant>(LHS), cast<Constant>(RHS), 4207 Q.DL); 4208 } 4209 4210 // Simplify calls to llvm.masked.load.* 4211 if (IID == Intrinsic::masked_load) { 4212 Value *MaskArg = ArgBegin[2]; 4213 Value *PassthruArg = ArgBegin[3]; 4214 // If the mask is all zeros or undef, the "passthru" argument is the result. 4215 if (maskIsAllZeroOrUndef(MaskArg)) 4216 return PassthruArg; 4217 } 4218 4219 // Perform idempotent optimizations 4220 if (!IsIdempotent(IID)) 4221 return nullptr; 4222 4223 // Unary Ops 4224 if (NumOperands == 1) 4225 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) 4226 if (II->getIntrinsicID() == IID) 4227 return II; 4228 4229 return nullptr; 4230 } 4231 4232 template <typename IterTy> 4233 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd, 4234 const Query &Q, unsigned MaxRecurse) { 4235 Type *Ty = V->getType(); 4236 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 4237 Ty = PTy->getElementType(); 4238 FunctionType *FTy = cast<FunctionType>(Ty); 4239 4240 // call undef -> undef 4241 // call null -> undef 4242 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V)) 4243 return UndefValue::get(FTy->getReturnType()); 4244 4245 Function *F = dyn_cast<Function>(V); 4246 if (!F) 4247 return nullptr; 4248 4249 if (F->isIntrinsic()) 4250 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse)) 4251 return Ret; 4252 4253 if (!canConstantFoldCallTo(F)) 4254 return nullptr; 4255 4256 SmallVector<Constant *, 4> ConstantArgs; 4257 ConstantArgs.reserve(ArgEnd - ArgBegin); 4258 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) { 4259 Constant *C = dyn_cast<Constant>(*I); 4260 if (!C) 4261 return nullptr; 4262 ConstantArgs.push_back(C); 4263 } 4264 4265 return ConstantFoldCall(F, ConstantArgs, Q.TLI); 4266 } 4267 4268 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin, 4269 User::op_iterator ArgEnd, const DataLayout &DL, 4270 const TargetLibraryInfo *TLI, const DominatorTree *DT, 4271 AssumptionCache *AC, const Instruction *CxtI) { 4272 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI), 4273 RecursionLimit); 4274 } 4275 4276 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args, 4277 const DataLayout &DL, const TargetLibraryInfo *TLI, 4278 const DominatorTree *DT, AssumptionCache *AC, 4279 const Instruction *CxtI) { 4280 return ::SimplifyCall(V, Args.begin(), Args.end(), 4281 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 4282 } 4283 4284 /// See if we can compute a simplified version of this instruction. 4285 /// If not, this returns null. 4286 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout &DL, 4287 const TargetLibraryInfo *TLI, 4288 const DominatorTree *DT, AssumptionCache *AC) { 4289 Value *Result; 4290 4291 switch (I->getOpcode()) { 4292 default: 4293 Result = ConstantFoldInstruction(I, DL, TLI); 4294 break; 4295 case Instruction::FAdd: 4296 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1), 4297 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4298 break; 4299 case Instruction::Add: 4300 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1), 4301 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4302 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 4303 TLI, DT, AC, I); 4304 break; 4305 case Instruction::FSub: 4306 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1), 4307 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4308 break; 4309 case Instruction::Sub: 4310 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1), 4311 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4312 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 4313 TLI, DT, AC, I); 4314 break; 4315 case Instruction::FMul: 4316 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1), 4317 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4318 break; 4319 case Instruction::Mul: 4320 Result = 4321 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4322 break; 4323 case Instruction::SDiv: 4324 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4325 AC, I); 4326 break; 4327 case Instruction::UDiv: 4328 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4329 AC, I); 4330 break; 4331 case Instruction::FDiv: 4332 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), 4333 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4334 break; 4335 case Instruction::SRem: 4336 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4337 AC, I); 4338 break; 4339 case Instruction::URem: 4340 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4341 AC, I); 4342 break; 4343 case Instruction::FRem: 4344 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), 4345 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4346 break; 4347 case Instruction::Shl: 4348 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1), 4349 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4350 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 4351 TLI, DT, AC, I); 4352 break; 4353 case Instruction::LShr: 4354 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1), 4355 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 4356 AC, I); 4357 break; 4358 case Instruction::AShr: 4359 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1), 4360 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 4361 AC, I); 4362 break; 4363 case Instruction::And: 4364 Result = 4365 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4366 break; 4367 case Instruction::Or: 4368 Result = 4369 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4370 break; 4371 case Instruction::Xor: 4372 Result = 4373 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4374 break; 4375 case Instruction::ICmp: 4376 Result = 4377 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0), 4378 I->getOperand(1), DL, TLI, DT, AC, I); 4379 break; 4380 case Instruction::FCmp: 4381 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), 4382 I->getOperand(0), I->getOperand(1), 4383 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4384 break; 4385 case Instruction::Select: 4386 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1), 4387 I->getOperand(2), DL, TLI, DT, AC, I); 4388 break; 4389 case Instruction::GetElementPtr: { 4390 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 4391 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(), 4392 Ops, DL, TLI, DT, AC, I); 4393 break; 4394 } 4395 case Instruction::InsertValue: { 4396 InsertValueInst *IV = cast<InsertValueInst>(I); 4397 Result = SimplifyInsertValueInst(IV->getAggregateOperand(), 4398 IV->getInsertedValueOperand(), 4399 IV->getIndices(), DL, TLI, DT, AC, I); 4400 break; 4401 } 4402 case Instruction::ExtractValue: { 4403 auto *EVI = cast<ExtractValueInst>(I); 4404 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(), 4405 EVI->getIndices(), DL, TLI, DT, AC, I); 4406 break; 4407 } 4408 case Instruction::ExtractElement: { 4409 auto *EEI = cast<ExtractElementInst>(I); 4410 Result = SimplifyExtractElementInst( 4411 EEI->getVectorOperand(), EEI->getIndexOperand(), DL, TLI, DT, AC, I); 4412 break; 4413 } 4414 case Instruction::PHI: 4415 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I)); 4416 break; 4417 case Instruction::Call: { 4418 CallSite CS(cast<CallInst>(I)); 4419 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL, 4420 TLI, DT, AC, I); 4421 break; 4422 } 4423 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc: 4424 #include "llvm/IR/Instruction.def" 4425 #undef HANDLE_CAST_INST 4426 Result = SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), 4427 DL, TLI, DT, AC, I); 4428 break; 4429 } 4430 4431 // In general, it is possible for computeKnownBits to determine all bits in a 4432 // value even when the operands are not all constants. 4433 if (!Result && I->getType()->isIntOrIntVectorTy()) { 4434 unsigned BitWidth = I->getType()->getScalarSizeInBits(); 4435 APInt KnownZero(BitWidth, 0); 4436 APInt KnownOne(BitWidth, 0); 4437 computeKnownBits(I, KnownZero, KnownOne, DL, /*Depth*/0, AC, I, DT); 4438 if ((KnownZero | KnownOne).isAllOnesValue()) 4439 Result = ConstantInt::get(I->getType(), KnownOne); 4440 } 4441 4442 /// If called on unreachable code, the above logic may report that the 4443 /// instruction simplified to itself. Make life easier for users by 4444 /// detecting that case here, returning a safe value instead. 4445 return Result == I ? UndefValue::get(I->getType()) : Result; 4446 } 4447 4448 /// \brief Implementation of recursive simplification through an instruction's 4449 /// uses. 4450 /// 4451 /// This is the common implementation of the recursive simplification routines. 4452 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to 4453 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of 4454 /// instructions to process and attempt to simplify it using 4455 /// InstructionSimplify. 4456 /// 4457 /// This routine returns 'true' only when *it* simplifies something. The passed 4458 /// in simplified value does not count toward this. 4459 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV, 4460 const TargetLibraryInfo *TLI, 4461 const DominatorTree *DT, 4462 AssumptionCache *AC) { 4463 bool Simplified = false; 4464 SmallSetVector<Instruction *, 8> Worklist; 4465 const DataLayout &DL = I->getModule()->getDataLayout(); 4466 4467 // If we have an explicit value to collapse to, do that round of the 4468 // simplification loop by hand initially. 4469 if (SimpleV) { 4470 for (User *U : I->users()) 4471 if (U != I) 4472 Worklist.insert(cast<Instruction>(U)); 4473 4474 // Replace the instruction with its simplified value. 4475 I->replaceAllUsesWith(SimpleV); 4476 4477 // Gracefully handle edge cases where the instruction is not wired into any 4478 // parent block. 4479 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) && 4480 !I->mayHaveSideEffects()) 4481 I->eraseFromParent(); 4482 } else { 4483 Worklist.insert(I); 4484 } 4485 4486 // Note that we must test the size on each iteration, the worklist can grow. 4487 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { 4488 I = Worklist[Idx]; 4489 4490 // See if this instruction simplifies. 4491 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC); 4492 if (!SimpleV) 4493 continue; 4494 4495 Simplified = true; 4496 4497 // Stash away all the uses of the old instruction so we can check them for 4498 // recursive simplifications after a RAUW. This is cheaper than checking all 4499 // uses of To on the recursive step in most cases. 4500 for (User *U : I->users()) 4501 Worklist.insert(cast<Instruction>(U)); 4502 4503 // Replace the instruction with its simplified value. 4504 I->replaceAllUsesWith(SimpleV); 4505 4506 // Gracefully handle edge cases where the instruction is not wired into any 4507 // parent block. 4508 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) && 4509 !I->mayHaveSideEffects()) 4510 I->eraseFromParent(); 4511 } 4512 return Simplified; 4513 } 4514 4515 bool llvm::recursivelySimplifyInstruction(Instruction *I, 4516 const TargetLibraryInfo *TLI, 4517 const DominatorTree *DT, 4518 AssumptionCache *AC) { 4519 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC); 4520 } 4521 4522 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV, 4523 const TargetLibraryInfo *TLI, 4524 const DominatorTree *DT, 4525 AssumptionCache *AC) { 4526 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!"); 4527 assert(SimpleV && "Must provide a simplified value."); 4528 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC); 4529 } 4530