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/AssumptionCache.h" 25 #include "llvm/Analysis/CaptureTracking.h" 26 #include "llvm/Analysis/ConstantFolding.h" 27 #include "llvm/Analysis/LoopAnalysisManager.h" 28 #include "llvm/Analysis/MemoryBuiltins.h" 29 #include "llvm/Analysis/OptimizationDiagnosticInfo.h" 30 #include "llvm/Analysis/ValueTracking.h" 31 #include "llvm/Analysis/VectorUtils.h" 32 #include "llvm/IR/ConstantRange.h" 33 #include "llvm/IR/DataLayout.h" 34 #include "llvm/IR/Dominators.h" 35 #include "llvm/IR/GetElementPtrTypeIterator.h" 36 #include "llvm/IR/GlobalAlias.h" 37 #include "llvm/IR/Operator.h" 38 #include "llvm/IR/PatternMatch.h" 39 #include "llvm/IR/ValueHandle.h" 40 #include "llvm/Support/KnownBits.h" 41 #include <algorithm> 42 using namespace llvm; 43 using namespace llvm::PatternMatch; 44 45 #define DEBUG_TYPE "instsimplify" 46 47 enum { RecursionLimit = 3 }; 48 49 STATISTIC(NumExpand, "Number of expansions"); 50 STATISTIC(NumReassoc, "Number of reassociations"); 51 52 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned); 53 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &, 54 unsigned); 55 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &, 56 const SimplifyQuery &, unsigned); 57 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &, 58 unsigned); 59 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 60 const SimplifyQuery &Q, unsigned MaxRecurse); 61 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned); 62 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned); 63 static Value *SimplifyCastInst(unsigned, Value *, Type *, 64 const SimplifyQuery &, unsigned); 65 66 /// For a boolean type or a vector of boolean type, return false or a vector 67 /// with every element false. 68 static Constant *getFalse(Type *Ty) { 69 return ConstantInt::getFalse(Ty); 70 } 71 72 /// For a boolean type or a vector of boolean type, return true or a vector 73 /// with every element true. 74 static Constant *getTrue(Type *Ty) { 75 return ConstantInt::getTrue(Ty); 76 } 77 78 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"? 79 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS, 80 Value *RHS) { 81 CmpInst *Cmp = dyn_cast<CmpInst>(V); 82 if (!Cmp) 83 return false; 84 CmpInst::Predicate CPred = Cmp->getPredicate(); 85 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1); 86 if (CPred == Pred && CLHS == LHS && CRHS == RHS) 87 return true; 88 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS && 89 CRHS == LHS; 90 } 91 92 /// Does the given value dominate the specified phi node? 93 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) { 94 Instruction *I = dyn_cast<Instruction>(V); 95 if (!I) 96 // Arguments and constants dominate all instructions. 97 return true; 98 99 // If we are processing instructions (and/or basic blocks) that have not been 100 // fully added to a function, the parent nodes may still be null. Simply 101 // return the conservative answer in these cases. 102 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent()) 103 return false; 104 105 // If we have a DominatorTree then do a precise test. 106 if (DT) { 107 if (!DT->isReachableFromEntry(P->getParent())) 108 return true; 109 if (!DT->isReachableFromEntry(I->getParent())) 110 return false; 111 return DT->dominates(I, P); 112 } 113 114 // Otherwise, if the instruction is in the entry block and is not an invoke, 115 // then it obviously dominates all phi nodes. 116 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() && 117 !isa<InvokeInst>(I)) 118 return true; 119 120 return false; 121 } 122 123 /// Simplify "A op (B op' C)" by distributing op over op', turning it into 124 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is 125 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS. 126 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)". 127 /// Returns the simplified value, or null if no simplification was performed. 128 static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS, 129 Instruction::BinaryOps OpcodeToExpand, const SimplifyQuery &Q, 130 unsigned MaxRecurse) { 131 // Recursion is always used, so bail out at once if we already hit the limit. 132 if (!MaxRecurse--) 133 return nullptr; 134 135 // Check whether the expression has the form "(A op' B) op C". 136 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS)) 137 if (Op0->getOpcode() == OpcodeToExpand) { 138 // It does! Try turning it into "(A op C) op' (B op C)". 139 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 140 // Do "A op C" and "B op C" both simplify? 141 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) 142 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 143 // They do! Return "L op' R" if it simplifies or is already available. 144 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 145 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand) 146 && L == B && R == A)) { 147 ++NumExpand; 148 return LHS; 149 } 150 // Otherwise return "L op' R" if it simplifies. 151 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 152 ++NumExpand; 153 return V; 154 } 155 } 156 } 157 158 // Check whether the expression has the form "A op (B op' C)". 159 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS)) 160 if (Op1->getOpcode() == OpcodeToExpand) { 161 // It does! Try turning it into "(A op B) op' (A op C)". 162 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 163 // Do "A op B" and "A op C" both simplify? 164 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) 165 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) { 166 // They do! Return "L op' R" if it simplifies or is already available. 167 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 168 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand) 169 && L == C && R == B)) { 170 ++NumExpand; 171 return RHS; 172 } 173 // Otherwise return "L op' R" if it simplifies. 174 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 175 ++NumExpand; 176 return V; 177 } 178 } 179 } 180 181 return nullptr; 182 } 183 184 /// Generic simplifications for associative binary operations. 185 /// Returns the simpler value, or null if none was found. 186 static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode, 187 Value *LHS, Value *RHS, const SimplifyQuery &Q, 188 unsigned MaxRecurse) { 189 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!"); 190 191 // Recursion is always used, so bail out at once if we already hit the limit. 192 if (!MaxRecurse--) 193 return nullptr; 194 195 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 196 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 197 198 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely. 199 if (Op0 && Op0->getOpcode() == Opcode) { 200 Value *A = Op0->getOperand(0); 201 Value *B = Op0->getOperand(1); 202 Value *C = RHS; 203 204 // Does "B op C" simplify? 205 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 206 // It does! Return "A op V" if it simplifies or is already available. 207 // If V equals B then "A op V" is just the LHS. 208 if (V == B) return LHS; 209 // Otherwise return "A op V" if it simplifies. 210 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) { 211 ++NumReassoc; 212 return W; 213 } 214 } 215 } 216 217 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely. 218 if (Op1 && Op1->getOpcode() == Opcode) { 219 Value *A = LHS; 220 Value *B = Op1->getOperand(0); 221 Value *C = Op1->getOperand(1); 222 223 // Does "A op B" simplify? 224 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) { 225 // It does! Return "V op C" if it simplifies or is already available. 226 // If V equals B then "V op C" is just the RHS. 227 if (V == B) return RHS; 228 // Otherwise return "V op C" if it simplifies. 229 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) { 230 ++NumReassoc; 231 return W; 232 } 233 } 234 } 235 236 // The remaining transforms require commutativity as well as associativity. 237 if (!Instruction::isCommutative(Opcode)) 238 return nullptr; 239 240 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely. 241 if (Op0 && Op0->getOpcode() == Opcode) { 242 Value *A = Op0->getOperand(0); 243 Value *B = Op0->getOperand(1); 244 Value *C = RHS; 245 246 // Does "C op A" simplify? 247 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 248 // It does! Return "V op B" if it simplifies or is already available. 249 // If V equals A then "V op B" is just the LHS. 250 if (V == A) return LHS; 251 // Otherwise return "V op B" if it simplifies. 252 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) { 253 ++NumReassoc; 254 return W; 255 } 256 } 257 } 258 259 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely. 260 if (Op1 && Op1->getOpcode() == Opcode) { 261 Value *A = LHS; 262 Value *B = Op1->getOperand(0); 263 Value *C = Op1->getOperand(1); 264 265 // Does "C op A" simplify? 266 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 267 // It does! Return "B op V" if it simplifies or is already available. 268 // If V equals C then "B op V" is just the RHS. 269 if (V == C) return RHS; 270 // Otherwise return "B op V" if it simplifies. 271 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) { 272 ++NumReassoc; 273 return W; 274 } 275 } 276 } 277 278 return nullptr; 279 } 280 281 /// In the case of a binary operation with a select instruction as an operand, 282 /// try to simplify the binop by seeing whether evaluating it on both branches 283 /// of the select results in the same value. Returns the common value if so, 284 /// otherwise returns null. 285 static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS, 286 Value *RHS, const SimplifyQuery &Q, 287 unsigned MaxRecurse) { 288 // Recursion is always used, so bail out at once if we already hit the limit. 289 if (!MaxRecurse--) 290 return nullptr; 291 292 SelectInst *SI; 293 if (isa<SelectInst>(LHS)) { 294 SI = cast<SelectInst>(LHS); 295 } else { 296 assert(isa<SelectInst>(RHS) && "No select instruction operand!"); 297 SI = cast<SelectInst>(RHS); 298 } 299 300 // Evaluate the BinOp on the true and false branches of the select. 301 Value *TV; 302 Value *FV; 303 if (SI == LHS) { 304 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse); 305 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse); 306 } else { 307 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse); 308 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse); 309 } 310 311 // If they simplified to the same value, then return the common value. 312 // If they both failed to simplify then return null. 313 if (TV == FV) 314 return TV; 315 316 // If one branch simplified to undef, return the other one. 317 if (TV && isa<UndefValue>(TV)) 318 return FV; 319 if (FV && isa<UndefValue>(FV)) 320 return TV; 321 322 // If applying the operation did not change the true and false select values, 323 // then the result of the binop is the select itself. 324 if (TV == SI->getTrueValue() && FV == SI->getFalseValue()) 325 return SI; 326 327 // If one branch simplified and the other did not, and the simplified 328 // value is equal to the unsimplified one, return the simplified value. 329 // For example, select (cond, X, X & Z) & Z -> X & Z. 330 if ((FV && !TV) || (TV && !FV)) { 331 // Check that the simplified value has the form "X op Y" where "op" is the 332 // same as the original operation. 333 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV); 334 if (Simplified && Simplified->getOpcode() == Opcode) { 335 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS". 336 // We already know that "op" is the same as for the simplified value. See 337 // if the operands match too. If so, return the simplified value. 338 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue(); 339 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS; 340 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch; 341 if (Simplified->getOperand(0) == UnsimplifiedLHS && 342 Simplified->getOperand(1) == UnsimplifiedRHS) 343 return Simplified; 344 if (Simplified->isCommutative() && 345 Simplified->getOperand(1) == UnsimplifiedLHS && 346 Simplified->getOperand(0) == UnsimplifiedRHS) 347 return Simplified; 348 } 349 } 350 351 return nullptr; 352 } 353 354 /// In the case of a comparison with a select instruction, try to simplify the 355 /// comparison by seeing whether both branches of the select result in the same 356 /// value. Returns the common value if so, otherwise returns null. 357 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS, 358 Value *RHS, const SimplifyQuery &Q, 359 unsigned MaxRecurse) { 360 // Recursion is always used, so bail out at once if we already hit the limit. 361 if (!MaxRecurse--) 362 return nullptr; 363 364 // Make sure the select is on the LHS. 365 if (!isa<SelectInst>(LHS)) { 366 std::swap(LHS, RHS); 367 Pred = CmpInst::getSwappedPredicate(Pred); 368 } 369 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!"); 370 SelectInst *SI = cast<SelectInst>(LHS); 371 Value *Cond = SI->getCondition(); 372 Value *TV = SI->getTrueValue(); 373 Value *FV = SI->getFalseValue(); 374 375 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it. 376 // Does "cmp TV, RHS" simplify? 377 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse); 378 if (TCmp == Cond) { 379 // It not only simplified, it simplified to the select condition. Replace 380 // it with 'true'. 381 TCmp = getTrue(Cond->getType()); 382 } else if (!TCmp) { 383 // It didn't simplify. However if "cmp TV, RHS" is equal to the select 384 // condition then we can replace it with 'true'. Otherwise give up. 385 if (!isSameCompare(Cond, Pred, TV, RHS)) 386 return nullptr; 387 TCmp = getTrue(Cond->getType()); 388 } 389 390 // Does "cmp FV, RHS" simplify? 391 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse); 392 if (FCmp == Cond) { 393 // It not only simplified, it simplified to the select condition. Replace 394 // it with 'false'. 395 FCmp = getFalse(Cond->getType()); 396 } else if (!FCmp) { 397 // It didn't simplify. However if "cmp FV, RHS" is equal to the select 398 // condition then we can replace it with 'false'. Otherwise give up. 399 if (!isSameCompare(Cond, Pred, FV, RHS)) 400 return nullptr; 401 FCmp = getFalse(Cond->getType()); 402 } 403 404 // If both sides simplified to the same value, then use it as the result of 405 // the original comparison. 406 if (TCmp == FCmp) 407 return TCmp; 408 409 // The remaining cases only make sense if the select condition has the same 410 // type as the result of the comparison, so bail out if this is not so. 411 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy()) 412 return nullptr; 413 // If the false value simplified to false, then the result of the compare 414 // is equal to "Cond && TCmp". This also catches the case when the false 415 // value simplified to false and the true value to true, returning "Cond". 416 if (match(FCmp, m_Zero())) 417 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse)) 418 return V; 419 // If the true value simplified to true, then the result of the compare 420 // is equal to "Cond || FCmp". 421 if (match(TCmp, m_One())) 422 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse)) 423 return V; 424 // Finally, if the false value simplified to true and the true value to 425 // false, then the result of the compare is equal to "!Cond". 426 if (match(FCmp, m_One()) && match(TCmp, m_Zero())) 427 if (Value *V = 428 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()), 429 Q, MaxRecurse)) 430 return V; 431 432 return nullptr; 433 } 434 435 /// In the case of a binary operation with an operand that is a PHI instruction, 436 /// try to simplify the binop by seeing whether evaluating it on the incoming 437 /// phi values yields the same result for every value. If so returns the common 438 /// value, otherwise returns null. 439 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS, 440 Value *RHS, const SimplifyQuery &Q, 441 unsigned MaxRecurse) { 442 // Recursion is always used, so bail out at once if we already hit the limit. 443 if (!MaxRecurse--) 444 return nullptr; 445 446 PHINode *PI; 447 if (isa<PHINode>(LHS)) { 448 PI = cast<PHINode>(LHS); 449 // Bail out if RHS and the phi may be mutually interdependent due to a loop. 450 if (!ValueDominatesPHI(RHS, PI, Q.DT)) 451 return nullptr; 452 } else { 453 assert(isa<PHINode>(RHS) && "No PHI instruction operand!"); 454 PI = cast<PHINode>(RHS); 455 // Bail out if LHS and the phi may be mutually interdependent due to a loop. 456 if (!ValueDominatesPHI(LHS, PI, Q.DT)) 457 return nullptr; 458 } 459 460 // Evaluate the BinOp on the incoming phi values. 461 Value *CommonValue = nullptr; 462 for (Value *Incoming : PI->incoming_values()) { 463 // If the incoming value is the phi node itself, it can safely be skipped. 464 if (Incoming == PI) continue; 465 Value *V = PI == LHS ? 466 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) : 467 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse); 468 // If the operation failed to simplify, or simplified to a different value 469 // to previously, then give up. 470 if (!V || (CommonValue && V != CommonValue)) 471 return nullptr; 472 CommonValue = V; 473 } 474 475 return CommonValue; 476 } 477 478 /// In the case of a comparison with a PHI instruction, try to simplify the 479 /// comparison by seeing whether comparing with all of the incoming phi values 480 /// yields the same result every time. If so returns the common result, 481 /// otherwise returns null. 482 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS, 483 const SimplifyQuery &Q, unsigned MaxRecurse) { 484 // Recursion is always used, so bail out at once if we already hit the limit. 485 if (!MaxRecurse--) 486 return nullptr; 487 488 // Make sure the phi is on the LHS. 489 if (!isa<PHINode>(LHS)) { 490 std::swap(LHS, RHS); 491 Pred = CmpInst::getSwappedPredicate(Pred); 492 } 493 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!"); 494 PHINode *PI = cast<PHINode>(LHS); 495 496 // Bail out if RHS and the phi may be mutually interdependent due to a loop. 497 if (!ValueDominatesPHI(RHS, PI, Q.DT)) 498 return nullptr; 499 500 // Evaluate the BinOp on the incoming phi values. 501 Value *CommonValue = nullptr; 502 for (Value *Incoming : PI->incoming_values()) { 503 // If the incoming value is the phi node itself, it can safely be skipped. 504 if (Incoming == PI) continue; 505 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse); 506 // If the operation failed to simplify, or simplified to a different value 507 // to previously, then give up. 508 if (!V || (CommonValue && V != CommonValue)) 509 return nullptr; 510 CommonValue = V; 511 } 512 513 return CommonValue; 514 } 515 516 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode, 517 Value *&Op0, Value *&Op1, 518 const SimplifyQuery &Q) { 519 if (auto *CLHS = dyn_cast<Constant>(Op0)) { 520 if (auto *CRHS = dyn_cast<Constant>(Op1)) 521 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL); 522 523 // Canonicalize the constant to the RHS if this is a commutative operation. 524 if (Instruction::isCommutative(Opcode)) 525 std::swap(Op0, Op1); 526 } 527 return nullptr; 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 SimplifyQuery &Q, unsigned MaxRecurse) { 534 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q)) 535 return C; 536 537 // X + undef -> undef 538 if (match(Op1, m_Undef())) 539 return Op1; 540 541 // X + 0 -> X 542 if (match(Op1, m_Zero())) 543 return Op0; 544 545 // X + (Y - X) -> Y 546 // (Y - X) + X -> Y 547 // Eg: X + -X -> 0 548 Value *Y = nullptr; 549 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) || 550 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1)))) 551 return Y; 552 553 // X + ~X -> -1 since ~X = -X-1 554 Type *Ty = Op0->getType(); 555 if (match(Op0, m_Not(m_Specific(Op1))) || 556 match(Op1, m_Not(m_Specific(Op0)))) 557 return Constant::getAllOnesValue(Ty); 558 559 // add nsw/nuw (xor Y, signmask), signmask --> Y 560 // The no-wrapping add guarantees that the top bit will be set by the add. 561 // Therefore, the xor must be clearing the already set sign bit of Y. 562 if ((isNSW || isNUW) && match(Op1, m_SignMask()) && 563 match(Op0, m_Xor(m_Value(Y), m_SignMask()))) 564 return Y; 565 566 /// i1 add -> xor. 567 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1)) 568 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 569 return V; 570 571 // Try some generic simplifications for associative operations. 572 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q, 573 MaxRecurse)) 574 return V; 575 576 // Threading Add over selects and phi nodes is pointless, so don't bother. 577 // Threading over the select in "A + select(cond, B, C)" means evaluating 578 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and 579 // only if B and C are equal. If B and C are equal then (since we assume 580 // that operands have already been simplified) "select(cond, B, C)" should 581 // have been simplified to the common value of B and C already. Analysing 582 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly 583 // for threading over phi nodes. 584 585 return nullptr; 586 } 587 588 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 589 const SimplifyQuery &Query) { 590 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query, 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 SimplifyQuery &Q, unsigned MaxRecurse) { 668 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q)) 669 return C; 670 671 // X - undef -> undef 672 // undef - X -> undef 673 if (match(Op0, m_Undef()) || match(Op1, m_Undef())) 674 return UndefValue::get(Op0->getType()); 675 676 // X - 0 -> X 677 if (match(Op1, m_Zero())) 678 return Op0; 679 680 // X - X -> 0 681 if (Op0 == Op1) 682 return Constant::getNullValue(Op0->getType()); 683 684 // Is this a negation? 685 if (match(Op0, m_Zero())) { 686 // 0 - X -> 0 if the sub is NUW. 687 if (isNUW) 688 return Op0; 689 690 unsigned BitWidth = Op1->getType()->getScalarSizeInBits(); 691 KnownBits Known(BitWidth); 692 computeKnownBits(Op1, Known, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 693 if (Known.Zero.isMaxSignedValue()) { 694 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then 695 // Op1 must be 0 because negating the minimum signed value is undefined. 696 if (isNSW) 697 return Op0; 698 699 // 0 - X -> X if X is 0 or the minimum signed value. 700 return Op1; 701 } 702 } 703 704 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies. 705 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X 706 Value *X = nullptr, *Y = nullptr, *Z = Op1; 707 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z 708 // See if "V === Y - Z" simplifies. 709 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1)) 710 // It does! Now see if "X + V" simplifies. 711 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) { 712 // It does, we successfully reassociated! 713 ++NumReassoc; 714 return W; 715 } 716 // See if "V === X - Z" simplifies. 717 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 718 // It does! Now see if "Y + V" simplifies. 719 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) { 720 // It does, we successfully reassociated! 721 ++NumReassoc; 722 return W; 723 } 724 } 725 726 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies. 727 // For example, X - (X + 1) -> -1 728 X = Op0; 729 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z) 730 // See if "V === X - Y" simplifies. 731 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 732 // It does! Now see if "V - Z" simplifies. 733 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) { 734 // It does, we successfully reassociated! 735 ++NumReassoc; 736 return W; 737 } 738 // See if "V === X - Z" simplifies. 739 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 740 // It does! Now see if "V - Y" simplifies. 741 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) { 742 // It does, we successfully reassociated! 743 ++NumReassoc; 744 return W; 745 } 746 } 747 748 // Z - (X - Y) -> (Z - X) + Y if everything simplifies. 749 // For example, X - (X - Y) -> Y. 750 Z = Op0; 751 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y) 752 // See if "V === Z - X" simplifies. 753 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1)) 754 // It does! Now see if "V + Y" simplifies. 755 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) { 756 // It does, we successfully reassociated! 757 ++NumReassoc; 758 return W; 759 } 760 761 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies. 762 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) && 763 match(Op1, m_Trunc(m_Value(Y)))) 764 if (X->getType() == Y->getType()) 765 // See if "V === X - Y" simplifies. 766 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 767 // It does! Now see if "trunc V" simplifies. 768 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(), 769 Q, MaxRecurse - 1)) 770 // It does, return the simplified "trunc V". 771 return W; 772 773 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...). 774 if (match(Op0, m_PtrToInt(m_Value(X))) && 775 match(Op1, m_PtrToInt(m_Value(Y)))) 776 if (Constant *Result = computePointerDifference(Q.DL, X, Y)) 777 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true); 778 779 // i1 sub -> xor. 780 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1)) 781 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 782 return V; 783 784 // Threading Sub over selects and phi nodes is pointless, so don't bother. 785 // Threading over the select in "A - select(cond, B, C)" means evaluating 786 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and 787 // only if B and C are equal. If B and C are equal then (since we assume 788 // that operands have already been simplified) "select(cond, B, C)" should 789 // have been simplified to the common value of B and C already. Analysing 790 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly 791 // for threading over phi nodes. 792 793 return nullptr; 794 } 795 796 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 797 const SimplifyQuery &Q) { 798 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit); 799 } 800 801 /// Given operands for an FAdd, see if we can fold the result. If not, this 802 /// returns null. 803 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 804 const SimplifyQuery &Q, unsigned MaxRecurse) { 805 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q)) 806 return C; 807 808 // fadd X, -0 ==> X 809 if (match(Op1, m_NegZero())) 810 return Op0; 811 812 // fadd X, 0 ==> X, when we know X is not -0 813 if (match(Op1, m_Zero()) && 814 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI))) 815 return Op0; 816 817 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0 818 // where nnan and ninf have to occur at least once somewhere in this 819 // expression 820 Value *SubOp = nullptr; 821 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0)))) 822 SubOp = Op1; 823 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1)))) 824 SubOp = Op0; 825 if (SubOp) { 826 Instruction *FSub = cast<Instruction>(SubOp); 827 if ((FMF.noNaNs() || FSub->hasNoNaNs()) && 828 (FMF.noInfs() || FSub->hasNoInfs())) 829 return Constant::getNullValue(Op0->getType()); 830 } 831 832 return nullptr; 833 } 834 835 /// Given operands for an FSub, see if we can fold the result. If not, this 836 /// returns null. 837 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 838 const SimplifyQuery &Q, unsigned MaxRecurse) { 839 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q)) 840 return C; 841 842 // fsub X, 0 ==> X 843 if (match(Op1, m_Zero())) 844 return Op0; 845 846 // fsub X, -0 ==> X, when we know X is not -0 847 if (match(Op1, m_NegZero()) && 848 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI))) 849 return Op0; 850 851 // fsub -0.0, (fsub -0.0, X) ==> X 852 Value *X; 853 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X)))) 854 return X; 855 856 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored. 857 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) && 858 match(Op1, m_FSub(m_AnyZero(), m_Value(X)))) 859 return X; 860 861 // fsub nnan x, x ==> 0.0 862 if (FMF.noNaNs() && Op0 == Op1) 863 return Constant::getNullValue(Op0->getType()); 864 865 return nullptr; 866 } 867 868 /// Given the operands for an FMul, see if we can fold the result 869 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF, 870 const SimplifyQuery &Q, unsigned MaxRecurse) { 871 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q)) 872 return C; 873 874 // fmul X, 1.0 ==> X 875 if (match(Op1, m_FPOne())) 876 return Op0; 877 878 // fmul nnan nsz X, 0 ==> 0 879 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero())) 880 return Op1; 881 882 return nullptr; 883 } 884 885 /// Given operands for a Mul, see if we can fold the result. 886 /// If not, this returns null. 887 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, 888 unsigned MaxRecurse) { 889 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q)) 890 return C; 891 892 // X * undef -> 0 893 if (match(Op1, m_Undef())) 894 return Constant::getNullValue(Op0->getType()); 895 896 // X * 0 -> 0 897 if (match(Op1, m_Zero())) 898 return Op1; 899 900 // X * 1 -> X 901 if (match(Op1, m_One())) 902 return Op0; 903 904 // (X / Y) * Y -> X if the division is exact. 905 Value *X = nullptr; 906 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y 907 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y) 908 return X; 909 910 // i1 mul -> and. 911 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1)) 912 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1)) 913 return V; 914 915 // Try some generic simplifications for associative operations. 916 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q, 917 MaxRecurse)) 918 return V; 919 920 // Mul distributes over Add. Try some generic simplifications based on this. 921 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add, 922 Q, MaxRecurse)) 923 return V; 924 925 // If the operation is with the result of a select instruction, check whether 926 // operating on either branch of the select always yields the same value. 927 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 928 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q, 929 MaxRecurse)) 930 return V; 931 932 // If the operation is with the result of a phi instruction, check whether 933 // operating on all incoming values of the phi always yields the same value. 934 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 935 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q, 936 MaxRecurse)) 937 return V; 938 939 return nullptr; 940 } 941 942 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 943 const SimplifyQuery &Q) { 944 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit); 945 } 946 947 948 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 949 const SimplifyQuery &Q) { 950 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit); 951 } 952 953 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF, 954 const SimplifyQuery &Q) { 955 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit); 956 } 957 958 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) { 959 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit); 960 } 961 962 /// Check for common or similar folds of integer division or integer remainder. 963 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) { 964 Type *Ty = Op0->getType(); 965 966 // X / undef -> undef 967 // X % undef -> undef 968 if (match(Op1, m_Undef())) 969 return Op1; 970 971 // X / 0 -> undef 972 // X % 0 -> undef 973 // We don't need to preserve faults! 974 if (match(Op1, m_Zero())) 975 return UndefValue::get(Ty); 976 977 // If any element of a constant divisor vector is zero, the whole op is undef. 978 auto *Op1C = dyn_cast<Constant>(Op1); 979 if (Op1C && Ty->isVectorTy()) { 980 unsigned NumElts = Ty->getVectorNumElements(); 981 for (unsigned i = 0; i != NumElts; ++i) { 982 Constant *Elt = Op1C->getAggregateElement(i); 983 if (Elt && Elt->isNullValue()) 984 return UndefValue::get(Ty); 985 } 986 } 987 988 // undef / X -> 0 989 // undef % X -> 0 990 if (match(Op0, m_Undef())) 991 return Constant::getNullValue(Ty); 992 993 // 0 / X -> 0 994 // 0 % X -> 0 995 if (match(Op0, m_Zero())) 996 return Op0; 997 998 // X / X -> 1 999 // X % X -> 0 1000 if (Op0 == Op1) 1001 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty); 1002 1003 // X / 1 -> X 1004 // X % 1 -> 0 1005 // If this is a boolean op (single-bit element type), we can't have 1006 // division-by-zero or remainder-by-zero, so assume the divisor is 1. 1007 if (match(Op1, m_One()) || Ty->getScalarType()->isIntegerTy(1)) 1008 return IsDiv ? Op0 : Constant::getNullValue(Ty); 1009 1010 return nullptr; 1011 } 1012 1013 /// Given operands for an SDiv or UDiv, see if we can fold the result. 1014 /// If not, this returns null. 1015 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1016 const SimplifyQuery &Q, unsigned MaxRecurse) { 1017 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q)) 1018 return C; 1019 1020 if (Value *V = simplifyDivRem(Op0, Op1, true)) 1021 return V; 1022 1023 bool isSigned = Opcode == Instruction::SDiv; 1024 1025 // (X * Y) / Y -> X if the multiplication does not overflow. 1026 Value *X = nullptr, *Y = nullptr; 1027 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) { 1028 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1 1029 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0); 1030 // If the Mul knows it does not overflow, then we are good to go. 1031 if ((isSigned && Mul->hasNoSignedWrap()) || 1032 (!isSigned && Mul->hasNoUnsignedWrap())) 1033 return X; 1034 // If X has the form X = A / Y then X * Y cannot overflow. 1035 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X)) 1036 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y) 1037 return X; 1038 } 1039 1040 // (X rem Y) / Y -> 0 1041 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1042 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1043 return Constant::getNullValue(Op0->getType()); 1044 1045 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow 1046 ConstantInt *C1, *C2; 1047 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) && 1048 match(Op1, m_ConstantInt(C2))) { 1049 bool Overflow; 1050 (void)C1->getValue().umul_ov(C2->getValue(), Overflow); 1051 if (Overflow) 1052 return Constant::getNullValue(Op0->getType()); 1053 } 1054 1055 // If the operation is with the result of a select instruction, check whether 1056 // operating on either branch of the select always yields the same value. 1057 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1058 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1059 return V; 1060 1061 // If the operation is with the result of a phi instruction, check whether 1062 // operating on all incoming values of the phi always yields the same value. 1063 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1064 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1065 return V; 1066 1067 return nullptr; 1068 } 1069 1070 /// Given operands for an SDiv, see if we can fold the result. 1071 /// If not, this returns null. 1072 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, 1073 unsigned MaxRecurse) { 1074 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse)) 1075 return V; 1076 1077 return nullptr; 1078 } 1079 1080 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) { 1081 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit); 1082 } 1083 1084 /// Given operands for a UDiv, see if we can fold the result. 1085 /// If not, this returns null. 1086 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, 1087 unsigned MaxRecurse) { 1088 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse)) 1089 return V; 1090 1091 // udiv %V, C -> 0 if %V < C 1092 if (MaxRecurse) { 1093 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst( 1094 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) { 1095 if (C->isAllOnesValue()) { 1096 return Constant::getNullValue(Op0->getType()); 1097 } 1098 } 1099 } 1100 1101 return nullptr; 1102 } 1103 1104 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) { 1105 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit); 1106 } 1107 1108 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1109 const SimplifyQuery &Q, unsigned) { 1110 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q)) 1111 return C; 1112 1113 // undef / X -> undef (the undef could be a snan). 1114 if (match(Op0, m_Undef())) 1115 return Op0; 1116 1117 // X / undef -> undef 1118 if (match(Op1, m_Undef())) 1119 return Op1; 1120 1121 // X / 1.0 -> X 1122 if (match(Op1, m_FPOne())) 1123 return Op0; 1124 1125 // 0 / X -> 0 1126 // Requires that NaNs are off (X could be zero) and signed zeroes are 1127 // ignored (X could be positive or negative, so the output sign is unknown). 1128 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1129 return Op0; 1130 1131 if (FMF.noNaNs()) { 1132 // X / X -> 1.0 is legal when NaNs are ignored. 1133 if (Op0 == Op1) 1134 return ConstantFP::get(Op0->getType(), 1.0); 1135 1136 // -X / X -> -1.0 and 1137 // X / -X -> -1.0 are legal when NaNs are ignored. 1138 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored. 1139 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) && 1140 BinaryOperator::getFNegArgument(Op0) == Op1) || 1141 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) && 1142 BinaryOperator::getFNegArgument(Op1) == Op0)) 1143 return ConstantFP::get(Op0->getType(), -1.0); 1144 } 1145 1146 return nullptr; 1147 } 1148 1149 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1150 const SimplifyQuery &Q) { 1151 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit); 1152 } 1153 1154 /// Given operands for an SRem or URem, see if we can fold the result. 1155 /// If not, this returns null. 1156 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1157 const SimplifyQuery &Q, unsigned MaxRecurse) { 1158 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q)) 1159 return C; 1160 1161 if (Value *V = simplifyDivRem(Op0, Op1, false)) 1162 return V; 1163 1164 // (X % Y) % Y -> X % Y 1165 if ((Opcode == Instruction::SRem && 1166 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1167 (Opcode == Instruction::URem && 1168 match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1169 return Op0; 1170 1171 // If the operation is with the result of a select instruction, check whether 1172 // operating on either branch of the select always yields the same value. 1173 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1174 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1175 return V; 1176 1177 // If the operation is with the result of a phi instruction, check whether 1178 // operating on all incoming values of the phi always yields the same value. 1179 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1180 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1181 return V; 1182 1183 return nullptr; 1184 } 1185 1186 /// Given operands for an SRem, see if we can fold the result. 1187 /// If not, this returns null. 1188 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, 1189 unsigned MaxRecurse) { 1190 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse)) 1191 return V; 1192 1193 return nullptr; 1194 } 1195 1196 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) { 1197 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit); 1198 } 1199 1200 /// Given operands for a URem, see if we can fold the result. 1201 /// If not, this returns null. 1202 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, 1203 unsigned MaxRecurse) { 1204 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse)) 1205 return V; 1206 1207 // urem %V, C -> %V if %V < C 1208 if (MaxRecurse) { 1209 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst( 1210 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) { 1211 if (C->isAllOnesValue()) { 1212 return Op0; 1213 } 1214 } 1215 } 1216 1217 return nullptr; 1218 } 1219 1220 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) { 1221 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit); 1222 } 1223 1224 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1225 const SimplifyQuery &Q, unsigned) { 1226 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q)) 1227 return C; 1228 1229 // undef % X -> undef (the undef could be a snan). 1230 if (match(Op0, m_Undef())) 1231 return Op0; 1232 1233 // X % undef -> undef 1234 if (match(Op1, m_Undef())) 1235 return Op1; 1236 1237 // 0 % X -> 0 1238 // Requires that NaNs are off (X could be zero) and signed zeroes are 1239 // ignored (X could be positive or negative, so the output sign is unknown). 1240 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1241 return Op0; 1242 1243 return nullptr; 1244 } 1245 1246 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1247 const SimplifyQuery &Q) { 1248 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit); 1249 } 1250 1251 /// Returns true if a shift by \c Amount always yields undef. 1252 static bool isUndefShift(Value *Amount) { 1253 Constant *C = dyn_cast<Constant>(Amount); 1254 if (!C) 1255 return false; 1256 1257 // X shift by undef -> undef because it may shift by the bitwidth. 1258 if (isa<UndefValue>(C)) 1259 return true; 1260 1261 // Shifting by the bitwidth or more is undefined. 1262 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) 1263 if (CI->getValue().getLimitedValue() >= 1264 CI->getType()->getScalarSizeInBits()) 1265 return true; 1266 1267 // If all lanes of a vector shift are undefined the whole shift is. 1268 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) { 1269 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I) 1270 if (!isUndefShift(C->getAggregateElement(I))) 1271 return false; 1272 return true; 1273 } 1274 1275 return false; 1276 } 1277 1278 /// Given operands for an Shl, LShr or AShr, see if we can fold the result. 1279 /// If not, this returns null. 1280 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0, 1281 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) { 1282 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q)) 1283 return C; 1284 1285 // 0 shift by X -> 0 1286 if (match(Op0, m_Zero())) 1287 return Op0; 1288 1289 // X shift by 0 -> X 1290 if (match(Op1, m_Zero())) 1291 return Op0; 1292 1293 // Fold undefined shifts. 1294 if (isUndefShift(Op1)) 1295 return UndefValue::get(Op0->getType()); 1296 1297 // If the operation is with the result of a select instruction, check whether 1298 // operating on either branch of the select always yields the same value. 1299 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1300 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1301 return V; 1302 1303 // If the operation is with the result of a phi instruction, check whether 1304 // operating on all incoming values of the phi always yields the same value. 1305 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1306 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1307 return V; 1308 1309 // If any bits in the shift amount make that value greater than or equal to 1310 // the number of bits in the type, the shift is undefined. 1311 unsigned BitWidth = Op1->getType()->getScalarSizeInBits(); 1312 KnownBits Known(BitWidth); 1313 computeKnownBits(Op1, Known, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 1314 if (Known.One.getLimitedValue() >= BitWidth) 1315 return UndefValue::get(Op0->getType()); 1316 1317 // If all valid bits in the shift amount are known zero, the first operand is 1318 // unchanged. 1319 unsigned NumValidShiftBits = Log2_32_Ceil(BitWidth); 1320 if (Known.Zero.countTrailingOnes() >= NumValidShiftBits) 1321 return Op0; 1322 1323 return nullptr; 1324 } 1325 1326 /// \brief Given operands for an Shl, LShr or AShr, see if we can 1327 /// fold the result. If not, this returns null. 1328 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0, 1329 Value *Op1, bool isExact, const SimplifyQuery &Q, 1330 unsigned MaxRecurse) { 1331 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse)) 1332 return V; 1333 1334 // X >> X -> 0 1335 if (Op0 == Op1) 1336 return Constant::getNullValue(Op0->getType()); 1337 1338 // undef >> X -> 0 1339 // undef >> X -> undef (if it's exact) 1340 if (match(Op0, m_Undef())) 1341 return isExact ? Op0 : Constant::getNullValue(Op0->getType()); 1342 1343 // The low bit cannot be shifted out of an exact shift if it is set. 1344 if (isExact) { 1345 unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); 1346 KnownBits Op0Known(BitWidth); 1347 computeKnownBits(Op0, Op0Known, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT); 1348 if (Op0Known.One[0]) 1349 return Op0; 1350 } 1351 1352 return nullptr; 1353 } 1354 1355 /// Given operands for an Shl, see if we can fold the result. 1356 /// If not, this returns null. 1357 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1358 const SimplifyQuery &Q, unsigned MaxRecurse) { 1359 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse)) 1360 return V; 1361 1362 // undef << X -> 0 1363 // undef << X -> undef if (if it's NSW/NUW) 1364 if (match(Op0, m_Undef())) 1365 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType()); 1366 1367 // (X >> A) << A -> X 1368 Value *X; 1369 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1))))) 1370 return X; 1371 return nullptr; 1372 } 1373 1374 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1375 const SimplifyQuery &Q) { 1376 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit); 1377 } 1378 1379 /// Given operands for an LShr, see if we can fold the result. 1380 /// If not, this returns null. 1381 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1382 const SimplifyQuery &Q, unsigned MaxRecurse) { 1383 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q, 1384 MaxRecurse)) 1385 return V; 1386 1387 // (X << A) >> A -> X 1388 Value *X; 1389 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1)))) 1390 return X; 1391 1392 return nullptr; 1393 } 1394 1395 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1396 const SimplifyQuery &Q) { 1397 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit); 1398 } 1399 1400 /// Given operands for an AShr, see if we can fold the result. 1401 /// If not, this returns null. 1402 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1403 const SimplifyQuery &Q, unsigned MaxRecurse) { 1404 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q, 1405 MaxRecurse)) 1406 return V; 1407 1408 // all ones >>a X -> all ones 1409 if (match(Op0, m_AllOnes())) 1410 return Op0; 1411 1412 // (X << A) >> A -> X 1413 Value *X; 1414 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1)))) 1415 return X; 1416 1417 // Arithmetic shifting an all-sign-bit value is a no-op. 1418 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 1419 if (NumSignBits == Op0->getType()->getScalarSizeInBits()) 1420 return Op0; 1421 1422 return nullptr; 1423 } 1424 1425 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1426 const SimplifyQuery &Q) { 1427 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit); 1428 } 1429 1430 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp, 1431 ICmpInst *UnsignedICmp, bool IsAnd) { 1432 Value *X, *Y; 1433 1434 ICmpInst::Predicate EqPred; 1435 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) || 1436 !ICmpInst::isEquality(EqPred)) 1437 return nullptr; 1438 1439 ICmpInst::Predicate UnsignedPred; 1440 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) && 1441 ICmpInst::isUnsigned(UnsignedPred)) 1442 ; 1443 else if (match(UnsignedICmp, 1444 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) && 1445 ICmpInst::isUnsigned(UnsignedPred)) 1446 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred); 1447 else 1448 return nullptr; 1449 1450 // X < Y && Y != 0 --> X < Y 1451 // X < Y || Y != 0 --> Y != 0 1452 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE) 1453 return IsAnd ? UnsignedICmp : ZeroICmp; 1454 1455 // X >= Y || Y != 0 --> true 1456 // X >= Y || Y == 0 --> X >= Y 1457 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) { 1458 if (EqPred == ICmpInst::ICMP_NE) 1459 return getTrue(UnsignedICmp->getType()); 1460 return UnsignedICmp; 1461 } 1462 1463 // X < Y && Y == 0 --> false 1464 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ && 1465 IsAnd) 1466 return getFalse(UnsignedICmp->getType()); 1467 1468 return nullptr; 1469 } 1470 1471 /// Commuted variants are assumed to be handled by calling this function again 1472 /// with the parameters swapped. 1473 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) { 1474 ICmpInst::Predicate Pred0, Pred1; 1475 Value *A ,*B; 1476 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) || 1477 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B)))) 1478 return nullptr; 1479 1480 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B). 1481 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we 1482 // can eliminate Op1 from this 'and'. 1483 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1)) 1484 return Op0; 1485 1486 // Check for any combination of predicates that are guaranteed to be disjoint. 1487 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) || 1488 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) || 1489 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) || 1490 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)) 1491 return getFalse(Op0->getType()); 1492 1493 return nullptr; 1494 } 1495 1496 /// Commuted variants are assumed to be handled by calling this function again 1497 /// with the parameters swapped. 1498 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) { 1499 ICmpInst::Predicate Pred0, Pred1; 1500 Value *A ,*B; 1501 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) || 1502 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B)))) 1503 return nullptr; 1504 1505 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B). 1506 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we 1507 // can eliminate Op0 from this 'or'. 1508 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1)) 1509 return Op1; 1510 1511 // Check for any combination of predicates that cover the entire range of 1512 // possibilities. 1513 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) || 1514 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) || 1515 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) || 1516 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE)) 1517 return getTrue(Op0->getType()); 1518 1519 return nullptr; 1520 } 1521 1522 /// Test if a pair of compares with a shared operand and 2 constants has an 1523 /// empty set intersection, full set union, or if one compare is a superset of 1524 /// the other. 1525 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1, 1526 bool IsAnd) { 1527 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)). 1528 if (Cmp0->getOperand(0) != Cmp1->getOperand(0)) 1529 return nullptr; 1530 1531 const APInt *C0, *C1; 1532 if (!match(Cmp0->getOperand(1), m_APInt(C0)) || 1533 !match(Cmp1->getOperand(1), m_APInt(C1))) 1534 return nullptr; 1535 1536 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0); 1537 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1); 1538 1539 // For and-of-compares, check if the intersection is empty: 1540 // (icmp X, C0) && (icmp X, C1) --> empty set --> false 1541 if (IsAnd && Range0.intersectWith(Range1).isEmptySet()) 1542 return getFalse(Cmp0->getType()); 1543 1544 // For or-of-compares, check if the union is full: 1545 // (icmp X, C0) || (icmp X, C1) --> full set --> true 1546 if (!IsAnd && Range0.unionWith(Range1).isFullSet()) 1547 return getTrue(Cmp0->getType()); 1548 1549 // Is one range a superset of the other? 1550 // If this is and-of-compares, take the smaller set: 1551 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42 1552 // If this is or-of-compares, take the larger set: 1553 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4 1554 if (Range0.contains(Range1)) 1555 return IsAnd ? Cmp1 : Cmp0; 1556 if (Range1.contains(Range0)) 1557 return IsAnd ? Cmp0 : Cmp1; 1558 1559 return nullptr; 1560 } 1561 1562 /// Commuted variants are assumed to be handled by calling this function again 1563 /// with the parameters swapped. 1564 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1565 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true)) 1566 return X; 1567 1568 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1)) 1569 return X; 1570 1571 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true)) 1572 return X; 1573 1574 // (icmp (add V, C0), C1) & (icmp V, C0) 1575 Type *ITy = Op0->getType(); 1576 ICmpInst::Predicate Pred0, Pred1; 1577 const APInt *C0, *C1; 1578 Value *V; 1579 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1)))) 1580 return nullptr; 1581 1582 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value()))) 1583 return nullptr; 1584 1585 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1586 if (AddInst->getOperand(1) != Op1->getOperand(1)) 1587 return nullptr; 1588 1589 bool isNSW = AddInst->hasNoSignedWrap(); 1590 bool isNUW = AddInst->hasNoUnsignedWrap(); 1591 1592 const APInt Delta = *C1 - *C0; 1593 if (C0->isStrictlyPositive()) { 1594 if (Delta == 2) { 1595 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT) 1596 return getFalse(ITy); 1597 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1598 return getFalse(ITy); 1599 } 1600 if (Delta == 1) { 1601 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT) 1602 return getFalse(ITy); 1603 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1604 return getFalse(ITy); 1605 } 1606 } 1607 if (C0->getBoolValue() && isNUW) { 1608 if (Delta == 2) 1609 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT) 1610 return getFalse(ITy); 1611 if (Delta == 1) 1612 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT) 1613 return getFalse(ITy); 1614 } 1615 1616 return nullptr; 1617 } 1618 1619 /// Commuted variants are assumed to be handled by calling this function again 1620 /// with the parameters swapped. 1621 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1622 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false)) 1623 return X; 1624 1625 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1)) 1626 return X; 1627 1628 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false)) 1629 return X; 1630 1631 // (icmp (add V, C0), C1) | (icmp V, C0) 1632 ICmpInst::Predicate Pred0, Pred1; 1633 const APInt *C0, *C1; 1634 Value *V; 1635 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1)))) 1636 return nullptr; 1637 1638 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value()))) 1639 return nullptr; 1640 1641 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1642 if (AddInst->getOperand(1) != Op1->getOperand(1)) 1643 return nullptr; 1644 1645 Type *ITy = Op0->getType(); 1646 bool isNSW = AddInst->hasNoSignedWrap(); 1647 bool isNUW = AddInst->hasNoUnsignedWrap(); 1648 1649 const APInt Delta = *C1 - *C0; 1650 if (C0->isStrictlyPositive()) { 1651 if (Delta == 2) { 1652 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE) 1653 return getTrue(ITy); 1654 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1655 return getTrue(ITy); 1656 } 1657 if (Delta == 1) { 1658 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE) 1659 return getTrue(ITy); 1660 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1661 return getTrue(ITy); 1662 } 1663 } 1664 if (C0->getBoolValue() && isNUW) { 1665 if (Delta == 2) 1666 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE) 1667 return getTrue(ITy); 1668 if (Delta == 1) 1669 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE) 1670 return getTrue(ITy); 1671 } 1672 1673 return nullptr; 1674 } 1675 1676 static Value *simplifyPossiblyCastedAndOrOfICmps(ICmpInst *Cmp0, ICmpInst *Cmp1, 1677 bool IsAnd, CastInst *Cast) { 1678 Value *V = 1679 IsAnd ? simplifyAndOfICmps(Cmp0, Cmp1) : simplifyOrOfICmps(Cmp0, Cmp1); 1680 if (!V) 1681 return nullptr; 1682 if (!Cast) 1683 return V; 1684 1685 // If we looked through casts, we can only handle a constant simplification 1686 // because we are not allowed to create a cast instruction here. 1687 if (auto *C = dyn_cast<Constant>(V)) 1688 return ConstantExpr::getCast(Cast->getOpcode(), C, Cast->getType()); 1689 1690 return nullptr; 1691 } 1692 1693 static Value *simplifyAndOrOfICmps(Value *Op0, Value *Op1, bool IsAnd) { 1694 // Look through casts of the 'and' operands to find compares. 1695 auto *Cast0 = dyn_cast<CastInst>(Op0); 1696 auto *Cast1 = dyn_cast<CastInst>(Op1); 1697 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() && 1698 Cast0->getSrcTy() == Cast1->getSrcTy()) { 1699 Op0 = Cast0->getOperand(0); 1700 Op1 = Cast1->getOperand(0); 1701 } 1702 1703 auto *Cmp0 = dyn_cast<ICmpInst>(Op0); 1704 auto *Cmp1 = dyn_cast<ICmpInst>(Op1); 1705 if (!Cmp0 || !Cmp1) 1706 return nullptr; 1707 1708 if (Value *V = simplifyPossiblyCastedAndOrOfICmps(Cmp0, Cmp1, IsAnd, Cast0)) 1709 return V; 1710 if (Value *V = simplifyPossiblyCastedAndOrOfICmps(Cmp1, Cmp0, IsAnd, Cast0)) 1711 return V; 1712 1713 return nullptr; 1714 } 1715 1716 /// Given operands for an And, see if we can fold the result. 1717 /// If not, this returns null. 1718 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, 1719 unsigned MaxRecurse) { 1720 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q)) 1721 return C; 1722 1723 // X & undef -> 0 1724 if (match(Op1, m_Undef())) 1725 return Constant::getNullValue(Op0->getType()); 1726 1727 // X & X = X 1728 if (Op0 == Op1) 1729 return Op0; 1730 1731 // X & 0 = 0 1732 if (match(Op1, m_Zero())) 1733 return Op1; 1734 1735 // X & -1 = X 1736 if (match(Op1, m_AllOnes())) 1737 return Op0; 1738 1739 // A & ~A = ~A & A = 0 1740 if (match(Op0, m_Not(m_Specific(Op1))) || 1741 match(Op1, m_Not(m_Specific(Op0)))) 1742 return Constant::getNullValue(Op0->getType()); 1743 1744 // (A | ?) & A = A 1745 Value *A = nullptr, *B = nullptr; 1746 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1747 (A == Op1 || B == Op1)) 1748 return Op1; 1749 1750 // A & (A | ?) = A 1751 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1752 (A == Op0 || B == Op0)) 1753 return Op0; 1754 1755 // A & (-A) = A if A is a power of two or zero. 1756 if (match(Op0, m_Neg(m_Specific(Op1))) || 1757 match(Op1, m_Neg(m_Specific(Op0)))) { 1758 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, 1759 Q.DT)) 1760 return Op0; 1761 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, 1762 Q.DT)) 1763 return Op1; 1764 } 1765 1766 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, true)) 1767 return V; 1768 1769 // Try some generic simplifications for associative operations. 1770 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q, 1771 MaxRecurse)) 1772 return V; 1773 1774 // And distributes over Or. Try some generic simplifications based on this. 1775 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or, 1776 Q, MaxRecurse)) 1777 return V; 1778 1779 // And distributes over Xor. Try some generic simplifications based on this. 1780 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor, 1781 Q, MaxRecurse)) 1782 return V; 1783 1784 // If the operation is with the result of a select instruction, check whether 1785 // operating on either branch of the select always yields the same value. 1786 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1787 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q, 1788 MaxRecurse)) 1789 return V; 1790 1791 // If the operation is with the result of a phi instruction, check whether 1792 // operating on all incoming values of the phi always yields the same value. 1793 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1794 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q, 1795 MaxRecurse)) 1796 return V; 1797 1798 return nullptr; 1799 } 1800 1801 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) { 1802 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit); 1803 } 1804 1805 /// Given operands for an Or, see if we can fold the result. 1806 /// If not, this returns null. 1807 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, 1808 unsigned MaxRecurse) { 1809 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q)) 1810 return C; 1811 1812 // X | undef -> -1 1813 if (match(Op1, m_Undef())) 1814 return Constant::getAllOnesValue(Op0->getType()); 1815 1816 // X | X = X 1817 if (Op0 == Op1) 1818 return Op0; 1819 1820 // X | 0 = X 1821 if (match(Op1, m_Zero())) 1822 return Op0; 1823 1824 // X | -1 = -1 1825 if (match(Op1, m_AllOnes())) 1826 return Op1; 1827 1828 // A | ~A = ~A | A = -1 1829 if (match(Op0, m_Not(m_Specific(Op1))) || 1830 match(Op1, m_Not(m_Specific(Op0)))) 1831 return Constant::getAllOnesValue(Op0->getType()); 1832 1833 // (A & ?) | A = A 1834 Value *A = nullptr, *B = nullptr; 1835 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 1836 (A == Op1 || B == Op1)) 1837 return Op1; 1838 1839 // A | (A & ?) = A 1840 if (match(Op1, m_And(m_Value(A), m_Value(B))) && 1841 (A == Op0 || B == Op0)) 1842 return Op0; 1843 1844 // ~(A & ?) | A = -1 1845 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) && 1846 (A == Op1 || B == Op1)) 1847 return Constant::getAllOnesValue(Op1->getType()); 1848 1849 // A | ~(A & ?) = -1 1850 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) && 1851 (A == Op0 || B == Op0)) 1852 return Constant::getAllOnesValue(Op0->getType()); 1853 1854 // (A & ~B) | (A ^ B) -> (A ^ B) 1855 // (~B & A) | (A ^ B) -> (A ^ B) 1856 // (A & ~B) | (B ^ A) -> (B ^ A) 1857 // (~B & A) | (B ^ A) -> (B ^ A) 1858 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && 1859 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) || 1860 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B))))) 1861 return Op1; 1862 1863 // Commute the 'or' operands. 1864 // (A ^ B) | (A & ~B) -> (A ^ B) 1865 // (A ^ B) | (~B & A) -> (A ^ B) 1866 // (B ^ A) | (A & ~B) -> (B ^ A) 1867 // (B ^ A) | (~B & A) -> (B ^ A) 1868 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) && 1869 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) || 1870 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B))))) 1871 return Op0; 1872 1873 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, false)) 1874 return V; 1875 1876 // Try some generic simplifications for associative operations. 1877 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q, 1878 MaxRecurse)) 1879 return V; 1880 1881 // Or distributes over And. Try some generic simplifications based on this. 1882 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q, 1883 MaxRecurse)) 1884 return V; 1885 1886 // If the operation is with the result of a select instruction, check whether 1887 // operating on either branch of the select always yields the same value. 1888 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1889 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q, 1890 MaxRecurse)) 1891 return V; 1892 1893 // (A & C)|(B & D) 1894 Value *C = nullptr, *D = nullptr; 1895 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 1896 match(Op1, m_And(m_Value(B), m_Value(D)))) { 1897 ConstantInt *C1 = dyn_cast<ConstantInt>(C); 1898 ConstantInt *C2 = dyn_cast<ConstantInt>(D); 1899 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) { 1900 // (A & C1)|(B & C2) 1901 // If we have: ((V + N) & C1) | (V & C2) 1902 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0 1903 // replace with V+N. 1904 Value *V1, *V2; 1905 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+ 1906 match(A, m_Add(m_Value(V1), m_Value(V2)))) { 1907 // Add commutes, try both ways. 1908 if (V1 == B && 1909 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1910 return A; 1911 if (V2 == B && 1912 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1913 return A; 1914 } 1915 // Or commutes, try both ways. 1916 if ((C1->getValue() & (C1->getValue() + 1)) == 0 && 1917 match(B, m_Add(m_Value(V1), m_Value(V2)))) { 1918 // Add commutes, try both ways. 1919 if (V1 == A && 1920 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1921 return B; 1922 if (V2 == A && 1923 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1924 return B; 1925 } 1926 } 1927 } 1928 1929 // If the operation is with the result of a phi instruction, check whether 1930 // operating on all incoming values of the phi always yields the same value. 1931 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1932 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse)) 1933 return V; 1934 1935 return nullptr; 1936 } 1937 1938 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) { 1939 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit); 1940 } 1941 1942 /// Given operands for a Xor, see if we can fold the result. 1943 /// If not, this returns null. 1944 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, 1945 unsigned MaxRecurse) { 1946 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q)) 1947 return C; 1948 1949 // A ^ undef -> undef 1950 if (match(Op1, m_Undef())) 1951 return Op1; 1952 1953 // A ^ 0 = A 1954 if (match(Op1, m_Zero())) 1955 return Op0; 1956 1957 // A ^ A = 0 1958 if (Op0 == Op1) 1959 return Constant::getNullValue(Op0->getType()); 1960 1961 // A ^ ~A = ~A ^ A = -1 1962 if (match(Op0, m_Not(m_Specific(Op1))) || 1963 match(Op1, m_Not(m_Specific(Op0)))) 1964 return Constant::getAllOnesValue(Op0->getType()); 1965 1966 // Try some generic simplifications for associative operations. 1967 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q, 1968 MaxRecurse)) 1969 return V; 1970 1971 // Threading Xor over selects and phi nodes is pointless, so don't bother. 1972 // Threading over the select in "A ^ select(cond, B, C)" means evaluating 1973 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and 1974 // only if B and C are equal. If B and C are equal then (since we assume 1975 // that operands have already been simplified) "select(cond, B, C)" should 1976 // have been simplified to the common value of B and C already. Analysing 1977 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly 1978 // for threading over phi nodes. 1979 1980 return nullptr; 1981 } 1982 1983 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) { 1984 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit); 1985 } 1986 1987 1988 static Type *GetCompareTy(Value *Op) { 1989 return CmpInst::makeCmpResultType(Op->getType()); 1990 } 1991 1992 /// Rummage around inside V looking for something equivalent to the comparison 1993 /// "LHS Pred RHS". Return such a value if found, otherwise return null. 1994 /// Helper function for analyzing max/min idioms. 1995 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred, 1996 Value *LHS, Value *RHS) { 1997 SelectInst *SI = dyn_cast<SelectInst>(V); 1998 if (!SI) 1999 return nullptr; 2000 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); 2001 if (!Cmp) 2002 return nullptr; 2003 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1); 2004 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS) 2005 return Cmp; 2006 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) && 2007 LHS == CmpRHS && RHS == CmpLHS) 2008 return Cmp; 2009 return nullptr; 2010 } 2011 2012 // A significant optimization not implemented here is assuming that alloca 2013 // addresses are not equal to incoming argument values. They don't *alias*, 2014 // as we say, but that doesn't mean they aren't equal, so we take a 2015 // conservative approach. 2016 // 2017 // This is inspired in part by C++11 5.10p1: 2018 // "Two pointers of the same type compare equal if and only if they are both 2019 // null, both point to the same function, or both represent the same 2020 // address." 2021 // 2022 // This is pretty permissive. 2023 // 2024 // It's also partly due to C11 6.5.9p6: 2025 // "Two pointers compare equal if and only if both are null pointers, both are 2026 // pointers to the same object (including a pointer to an object and a 2027 // subobject at its beginning) or function, both are pointers to one past the 2028 // last element of the same array object, or one is a pointer to one past the 2029 // end of one array object and the other is a pointer to the start of a 2030 // different array object that happens to immediately follow the first array 2031 // object in the address space.) 2032 // 2033 // C11's version is more restrictive, however there's no reason why an argument 2034 // couldn't be a one-past-the-end value for a stack object in the caller and be 2035 // equal to the beginning of a stack object in the callee. 2036 // 2037 // If the C and C++ standards are ever made sufficiently restrictive in this 2038 // area, it may be possible to update LLVM's semantics accordingly and reinstate 2039 // this optimization. 2040 static Constant * 2041 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI, 2042 const DominatorTree *DT, CmpInst::Predicate Pred, 2043 const Instruction *CxtI, Value *LHS, Value *RHS) { 2044 // First, skip past any trivial no-ops. 2045 LHS = LHS->stripPointerCasts(); 2046 RHS = RHS->stripPointerCasts(); 2047 2048 // A non-null pointer is not equal to a null pointer. 2049 if (llvm::isKnownNonNull(LHS) && isa<ConstantPointerNull>(RHS) && 2050 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE)) 2051 return ConstantInt::get(GetCompareTy(LHS), 2052 !CmpInst::isTrueWhenEqual(Pred)); 2053 2054 // We can only fold certain predicates on pointer comparisons. 2055 switch (Pred) { 2056 default: 2057 return nullptr; 2058 2059 // Equality comaprisons are easy to fold. 2060 case CmpInst::ICMP_EQ: 2061 case CmpInst::ICMP_NE: 2062 break; 2063 2064 // We can only handle unsigned relational comparisons because 'inbounds' on 2065 // a GEP only protects against unsigned wrapping. 2066 case CmpInst::ICMP_UGT: 2067 case CmpInst::ICMP_UGE: 2068 case CmpInst::ICMP_ULT: 2069 case CmpInst::ICMP_ULE: 2070 // However, we have to switch them to their signed variants to handle 2071 // negative indices from the base pointer. 2072 Pred = ICmpInst::getSignedPredicate(Pred); 2073 break; 2074 } 2075 2076 // Strip off any constant offsets so that we can reason about them. 2077 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets 2078 // here and compare base addresses like AliasAnalysis does, however there are 2079 // numerous hazards. AliasAnalysis and its utilities rely on special rules 2080 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis 2081 // doesn't need to guarantee pointer inequality when it says NoAlias. 2082 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 2083 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 2084 2085 // If LHS and RHS are related via constant offsets to the same base 2086 // value, we can replace it with an icmp which just compares the offsets. 2087 if (LHS == RHS) 2088 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset); 2089 2090 // Various optimizations for (in)equality comparisons. 2091 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) { 2092 // Different non-empty allocations that exist at the same time have 2093 // different addresses (if the program can tell). Global variables always 2094 // exist, so they always exist during the lifetime of each other and all 2095 // allocas. Two different allocas usually have different addresses... 2096 // 2097 // However, if there's an @llvm.stackrestore dynamically in between two 2098 // allocas, they may have the same address. It's tempting to reduce the 2099 // scope of the problem by only looking at *static* allocas here. That would 2100 // cover the majority of allocas while significantly reducing the likelihood 2101 // of having an @llvm.stackrestore pop up in the middle. However, it's not 2102 // actually impossible for an @llvm.stackrestore to pop up in the middle of 2103 // an entry block. Also, if we have a block that's not attached to a 2104 // function, we can't tell if it's "static" under the current definition. 2105 // Theoretically, this problem could be fixed by creating a new kind of 2106 // instruction kind specifically for static allocas. Such a new instruction 2107 // could be required to be at the top of the entry block, thus preventing it 2108 // from being subject to a @llvm.stackrestore. Instcombine could even 2109 // convert regular allocas into these special allocas. It'd be nifty. 2110 // However, until then, this problem remains open. 2111 // 2112 // So, we'll assume that two non-empty allocas have different addresses 2113 // for now. 2114 // 2115 // With all that, if the offsets are within the bounds of their allocations 2116 // (and not one-past-the-end! so we can't use inbounds!), and their 2117 // allocations aren't the same, the pointers are not equal. 2118 // 2119 // Note that it's not necessary to check for LHS being a global variable 2120 // address, due to canonicalization and constant folding. 2121 if (isa<AllocaInst>(LHS) && 2122 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) { 2123 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset); 2124 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset); 2125 uint64_t LHSSize, RHSSize; 2126 if (LHSOffsetCI && RHSOffsetCI && 2127 getObjectSize(LHS, LHSSize, DL, TLI) && 2128 getObjectSize(RHS, RHSSize, DL, TLI)) { 2129 const APInt &LHSOffsetValue = LHSOffsetCI->getValue(); 2130 const APInt &RHSOffsetValue = RHSOffsetCI->getValue(); 2131 if (!LHSOffsetValue.isNegative() && 2132 !RHSOffsetValue.isNegative() && 2133 LHSOffsetValue.ult(LHSSize) && 2134 RHSOffsetValue.ult(RHSSize)) { 2135 return ConstantInt::get(GetCompareTy(LHS), 2136 !CmpInst::isTrueWhenEqual(Pred)); 2137 } 2138 } 2139 2140 // Repeat the above check but this time without depending on DataLayout 2141 // or being able to compute a precise size. 2142 if (!cast<PointerType>(LHS->getType())->isEmptyTy() && 2143 !cast<PointerType>(RHS->getType())->isEmptyTy() && 2144 LHSOffset->isNullValue() && 2145 RHSOffset->isNullValue()) 2146 return ConstantInt::get(GetCompareTy(LHS), 2147 !CmpInst::isTrueWhenEqual(Pred)); 2148 } 2149 2150 // Even if an non-inbounds GEP occurs along the path we can still optimize 2151 // equality comparisons concerning the result. We avoid walking the whole 2152 // chain again by starting where the last calls to 2153 // stripAndComputeConstantOffsets left off and accumulate the offsets. 2154 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true); 2155 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true); 2156 if (LHS == RHS) 2157 return ConstantExpr::getICmp(Pred, 2158 ConstantExpr::getAdd(LHSOffset, LHSNoBound), 2159 ConstantExpr::getAdd(RHSOffset, RHSNoBound)); 2160 2161 // If one side of the equality comparison must come from a noalias call 2162 // (meaning a system memory allocation function), and the other side must 2163 // come from a pointer that cannot overlap with dynamically-allocated 2164 // memory within the lifetime of the current function (allocas, byval 2165 // arguments, globals), then determine the comparison result here. 2166 SmallVector<Value *, 8> LHSUObjs, RHSUObjs; 2167 GetUnderlyingObjects(LHS, LHSUObjs, DL); 2168 GetUnderlyingObjects(RHS, RHSUObjs, DL); 2169 2170 // Is the set of underlying objects all noalias calls? 2171 auto IsNAC = [](ArrayRef<Value *> Objects) { 2172 return all_of(Objects, isNoAliasCall); 2173 }; 2174 2175 // Is the set of underlying objects all things which must be disjoint from 2176 // noalias calls. For allocas, we consider only static ones (dynamic 2177 // allocas might be transformed into calls to malloc not simultaneously 2178 // live with the compared-to allocation). For globals, we exclude symbols 2179 // that might be resolve lazily to symbols in another dynamically-loaded 2180 // library (and, thus, could be malloc'ed by the implementation). 2181 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) { 2182 return all_of(Objects, [](Value *V) { 2183 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) 2184 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca(); 2185 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 2186 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() || 2187 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) && 2188 !GV->isThreadLocal(); 2189 if (const Argument *A = dyn_cast<Argument>(V)) 2190 return A->hasByValAttr(); 2191 return false; 2192 }); 2193 }; 2194 2195 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) || 2196 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs))) 2197 return ConstantInt::get(GetCompareTy(LHS), 2198 !CmpInst::isTrueWhenEqual(Pred)); 2199 2200 // Fold comparisons for non-escaping pointer even if the allocation call 2201 // cannot be elided. We cannot fold malloc comparison to null. Also, the 2202 // dynamic allocation call could be either of the operands. 2203 Value *MI = nullptr; 2204 if (isAllocLikeFn(LHS, TLI) && llvm::isKnownNonNullAt(RHS, CxtI, DT)) 2205 MI = LHS; 2206 else if (isAllocLikeFn(RHS, TLI) && llvm::isKnownNonNullAt(LHS, CxtI, DT)) 2207 MI = RHS; 2208 // FIXME: We should also fold the compare when the pointer escapes, but the 2209 // compare dominates the pointer escape 2210 if (MI && !PointerMayBeCaptured(MI, true, true)) 2211 return ConstantInt::get(GetCompareTy(LHS), 2212 CmpInst::isFalseWhenEqual(Pred)); 2213 } 2214 2215 // Otherwise, fail. 2216 return nullptr; 2217 } 2218 2219 /// Fold an icmp when its operands have i1 scalar type. 2220 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS, 2221 Value *RHS, const SimplifyQuery &Q) { 2222 Type *ITy = GetCompareTy(LHS); // The return type. 2223 Type *OpTy = LHS->getType(); // The operand type. 2224 if (!OpTy->getScalarType()->isIntegerTy(1)) 2225 return nullptr; 2226 2227 switch (Pred) { 2228 default: 2229 break; 2230 case ICmpInst::ICMP_EQ: 2231 // X == 1 -> X 2232 if (match(RHS, m_One())) 2233 return LHS; 2234 break; 2235 case ICmpInst::ICMP_NE: 2236 // X != 0 -> X 2237 if (match(RHS, m_Zero())) 2238 return LHS; 2239 break; 2240 case ICmpInst::ICMP_UGT: 2241 // X >u 0 -> X 2242 if (match(RHS, m_Zero())) 2243 return LHS; 2244 break; 2245 case ICmpInst::ICMP_UGE: 2246 // X >=u 1 -> X 2247 if (match(RHS, m_One())) 2248 return LHS; 2249 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false)) 2250 return getTrue(ITy); 2251 break; 2252 case ICmpInst::ICMP_SGE: 2253 /// For signed comparison, the values for an i1 are 0 and -1 2254 /// respectively. This maps into a truth table of: 2255 /// LHS | RHS | LHS >=s RHS | LHS implies RHS 2256 /// 0 | 0 | 1 (0 >= 0) | 1 2257 /// 0 | 1 | 1 (0 >= -1) | 1 2258 /// 1 | 0 | 0 (-1 >= 0) | 0 2259 /// 1 | 1 | 1 (-1 >= -1) | 1 2260 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false)) 2261 return getTrue(ITy); 2262 break; 2263 case ICmpInst::ICMP_SLT: 2264 // X <s 0 -> X 2265 if (match(RHS, m_Zero())) 2266 return LHS; 2267 break; 2268 case ICmpInst::ICMP_SLE: 2269 // X <=s -1 -> X 2270 if (match(RHS, m_One())) 2271 return LHS; 2272 break; 2273 case ICmpInst::ICMP_ULE: 2274 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false)) 2275 return getTrue(ITy); 2276 break; 2277 } 2278 2279 return nullptr; 2280 } 2281 2282 /// Try hard to fold icmp with zero RHS because this is a common case. 2283 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS, 2284 Value *RHS, const SimplifyQuery &Q) { 2285 if (!match(RHS, m_Zero())) 2286 return nullptr; 2287 2288 Type *ITy = GetCompareTy(LHS); // The return type. 2289 bool LHSKnownNonNegative, LHSKnownNegative; 2290 switch (Pred) { 2291 default: 2292 llvm_unreachable("Unknown ICmp predicate!"); 2293 case ICmpInst::ICMP_ULT: 2294 return getFalse(ITy); 2295 case ICmpInst::ICMP_UGE: 2296 return getTrue(ITy); 2297 case ICmpInst::ICMP_EQ: 2298 case ICmpInst::ICMP_ULE: 2299 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2300 return getFalse(ITy); 2301 break; 2302 case ICmpInst::ICMP_NE: 2303 case ICmpInst::ICMP_UGT: 2304 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2305 return getTrue(ITy); 2306 break; 2307 case ICmpInst::ICMP_SLT: 2308 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2309 Q.CxtI, Q.DT); 2310 if (LHSKnownNegative) 2311 return getTrue(ITy); 2312 if (LHSKnownNonNegative) 2313 return getFalse(ITy); 2314 break; 2315 case ICmpInst::ICMP_SLE: 2316 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2317 Q.CxtI, Q.DT); 2318 if (LHSKnownNegative) 2319 return getTrue(ITy); 2320 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2321 return getFalse(ITy); 2322 break; 2323 case ICmpInst::ICMP_SGE: 2324 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2325 Q.CxtI, Q.DT); 2326 if (LHSKnownNegative) 2327 return getFalse(ITy); 2328 if (LHSKnownNonNegative) 2329 return getTrue(ITy); 2330 break; 2331 case ICmpInst::ICMP_SGT: 2332 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2333 Q.CxtI, Q.DT); 2334 if (LHSKnownNegative) 2335 return getFalse(ITy); 2336 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2337 return getTrue(ITy); 2338 break; 2339 } 2340 2341 return nullptr; 2342 } 2343 2344 /// Many binary operators with a constant operand have an easy-to-compute 2345 /// range of outputs. This can be used to fold a comparison to always true or 2346 /// always false. 2347 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) { 2348 unsigned Width = Lower.getBitWidth(); 2349 const APInt *C; 2350 switch (BO.getOpcode()) { 2351 case Instruction::Add: 2352 if (match(BO.getOperand(1), m_APInt(C)) && *C != 0) { 2353 // FIXME: If we have both nuw and nsw, we should reduce the range further. 2354 if (BO.hasNoUnsignedWrap()) { 2355 // 'add nuw x, C' produces [C, UINT_MAX]. 2356 Lower = *C; 2357 } else if (BO.hasNoSignedWrap()) { 2358 if (C->isNegative()) { 2359 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C]. 2360 Lower = APInt::getSignedMinValue(Width); 2361 Upper = APInt::getSignedMaxValue(Width) + *C + 1; 2362 } else { 2363 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX]. 2364 Lower = APInt::getSignedMinValue(Width) + *C; 2365 Upper = APInt::getSignedMaxValue(Width) + 1; 2366 } 2367 } 2368 } 2369 break; 2370 2371 case Instruction::And: 2372 if (match(BO.getOperand(1), m_APInt(C))) 2373 // 'and x, C' produces [0, C]. 2374 Upper = *C + 1; 2375 break; 2376 2377 case Instruction::Or: 2378 if (match(BO.getOperand(1), m_APInt(C))) 2379 // 'or x, C' produces [C, UINT_MAX]. 2380 Lower = *C; 2381 break; 2382 2383 case Instruction::AShr: 2384 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { 2385 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C]. 2386 Lower = APInt::getSignedMinValue(Width).ashr(*C); 2387 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1; 2388 } else if (match(BO.getOperand(0), m_APInt(C))) { 2389 unsigned ShiftAmount = Width - 1; 2390 if (*C != 0 && BO.isExact()) 2391 ShiftAmount = C->countTrailingZeros(); 2392 if (C->isNegative()) { 2393 // 'ashr C, x' produces [C, C >> (Width-1)] 2394 Lower = *C; 2395 Upper = C->ashr(ShiftAmount) + 1; 2396 } else { 2397 // 'ashr C, x' produces [C >> (Width-1), C] 2398 Lower = C->ashr(ShiftAmount); 2399 Upper = *C + 1; 2400 } 2401 } 2402 break; 2403 2404 case Instruction::LShr: 2405 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { 2406 // 'lshr x, C' produces [0, UINT_MAX >> C]. 2407 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1; 2408 } else if (match(BO.getOperand(0), m_APInt(C))) { 2409 // 'lshr C, x' produces [C >> (Width-1), C]. 2410 unsigned ShiftAmount = Width - 1; 2411 if (*C != 0 && BO.isExact()) 2412 ShiftAmount = C->countTrailingZeros(); 2413 Lower = C->lshr(ShiftAmount); 2414 Upper = *C + 1; 2415 } 2416 break; 2417 2418 case Instruction::Shl: 2419 if (match(BO.getOperand(0), m_APInt(C))) { 2420 if (BO.hasNoUnsignedWrap()) { 2421 // 'shl nuw C, x' produces [C, C << CLZ(C)] 2422 Lower = *C; 2423 Upper = Lower.shl(Lower.countLeadingZeros()) + 1; 2424 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw? 2425 if (C->isNegative()) { 2426 // 'shl nsw C, x' produces [C << CLO(C)-1, C] 2427 unsigned ShiftAmount = C->countLeadingOnes() - 1; 2428 Lower = C->shl(ShiftAmount); 2429 Upper = *C + 1; 2430 } else { 2431 // 'shl nsw C, x' produces [C, C << CLZ(C)-1] 2432 unsigned ShiftAmount = C->countLeadingZeros() - 1; 2433 Lower = *C; 2434 Upper = C->shl(ShiftAmount) + 1; 2435 } 2436 } 2437 } 2438 break; 2439 2440 case Instruction::SDiv: 2441 if (match(BO.getOperand(1), m_APInt(C))) { 2442 APInt IntMin = APInt::getSignedMinValue(Width); 2443 APInt IntMax = APInt::getSignedMaxValue(Width); 2444 if (C->isAllOnesValue()) { 2445 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] 2446 // where C != -1 and C != 0 and C != 1 2447 Lower = IntMin + 1; 2448 Upper = IntMax + 1; 2449 } else if (C->countLeadingZeros() < Width - 1) { 2450 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C] 2451 // where C != -1 and C != 0 and C != 1 2452 Lower = IntMin.sdiv(*C); 2453 Upper = IntMax.sdiv(*C); 2454 if (Lower.sgt(Upper)) 2455 std::swap(Lower, Upper); 2456 Upper = Upper + 1; 2457 assert(Upper != Lower && "Upper part of range has wrapped!"); 2458 } 2459 } else if (match(BO.getOperand(0), m_APInt(C))) { 2460 if (C->isMinSignedValue()) { 2461 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. 2462 Lower = *C; 2463 Upper = Lower.lshr(1) + 1; 2464 } else { 2465 // 'sdiv C, x' produces [-|C|, |C|]. 2466 Upper = C->abs() + 1; 2467 Lower = (-Upper) + 1; 2468 } 2469 } 2470 break; 2471 2472 case Instruction::UDiv: 2473 if (match(BO.getOperand(1), m_APInt(C)) && *C != 0) { 2474 // 'udiv x, C' produces [0, UINT_MAX / C]. 2475 Upper = APInt::getMaxValue(Width).udiv(*C) + 1; 2476 } else if (match(BO.getOperand(0), m_APInt(C))) { 2477 // 'udiv C, x' produces [0, C]. 2478 Upper = *C + 1; 2479 } 2480 break; 2481 2482 case Instruction::SRem: 2483 if (match(BO.getOperand(1), m_APInt(C))) { 2484 // 'srem x, C' produces (-|C|, |C|). 2485 Upper = C->abs(); 2486 Lower = (-Upper) + 1; 2487 } 2488 break; 2489 2490 case Instruction::URem: 2491 if (match(BO.getOperand(1), m_APInt(C))) 2492 // 'urem x, C' produces [0, C). 2493 Upper = *C; 2494 break; 2495 2496 default: 2497 break; 2498 } 2499 } 2500 2501 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS, 2502 Value *RHS) { 2503 const APInt *C; 2504 if (!match(RHS, m_APInt(C))) 2505 return nullptr; 2506 2507 // Rule out tautological comparisons (eg., ult 0 or uge 0). 2508 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C); 2509 if (RHS_CR.isEmptySet()) 2510 return ConstantInt::getFalse(GetCompareTy(RHS)); 2511 if (RHS_CR.isFullSet()) 2512 return ConstantInt::getTrue(GetCompareTy(RHS)); 2513 2514 // Find the range of possible values for binary operators. 2515 unsigned Width = C->getBitWidth(); 2516 APInt Lower = APInt(Width, 0); 2517 APInt Upper = APInt(Width, 0); 2518 if (auto *BO = dyn_cast<BinaryOperator>(LHS)) 2519 setLimitsForBinOp(*BO, Lower, Upper); 2520 2521 ConstantRange LHS_CR = 2522 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true); 2523 2524 if (auto *I = dyn_cast<Instruction>(LHS)) 2525 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range)) 2526 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges)); 2527 2528 if (!LHS_CR.isFullSet()) { 2529 if (RHS_CR.contains(LHS_CR)) 2530 return ConstantInt::getTrue(GetCompareTy(RHS)); 2531 if (RHS_CR.inverse().contains(LHS_CR)) 2532 return ConstantInt::getFalse(GetCompareTy(RHS)); 2533 } 2534 2535 return nullptr; 2536 } 2537 2538 /// TODO: A large part of this logic is duplicated in InstCombine's 2539 /// foldICmpBinOp(). We should be able to share that and avoid the code 2540 /// duplication. 2541 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS, 2542 Value *RHS, const SimplifyQuery &Q, 2543 unsigned MaxRecurse) { 2544 Type *ITy = GetCompareTy(LHS); // The return type. 2545 2546 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS); 2547 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS); 2548 if (MaxRecurse && (LBO || RBO)) { 2549 // Analyze the case when either LHS or RHS is an add instruction. 2550 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 2551 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null). 2552 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false; 2553 if (LBO && LBO->getOpcode() == Instruction::Add) { 2554 A = LBO->getOperand(0); 2555 B = LBO->getOperand(1); 2556 NoLHSWrapProblem = 2557 ICmpInst::isEquality(Pred) || 2558 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) || 2559 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap()); 2560 } 2561 if (RBO && RBO->getOpcode() == Instruction::Add) { 2562 C = RBO->getOperand(0); 2563 D = RBO->getOperand(1); 2564 NoRHSWrapProblem = 2565 ICmpInst::isEquality(Pred) || 2566 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) || 2567 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap()); 2568 } 2569 2570 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2571 if ((A == RHS || B == RHS) && NoLHSWrapProblem) 2572 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A, 2573 Constant::getNullValue(RHS->getType()), Q, 2574 MaxRecurse - 1)) 2575 return V; 2576 2577 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2578 if ((C == LHS || D == LHS) && NoRHSWrapProblem) 2579 if (Value *V = 2580 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()), 2581 C == LHS ? D : C, Q, MaxRecurse - 1)) 2582 return V; 2583 2584 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow. 2585 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem && 2586 NoRHSWrapProblem) { 2587 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2588 Value *Y, *Z; 2589 if (A == C) { 2590 // C + B == C + D -> B == D 2591 Y = B; 2592 Z = D; 2593 } else if (A == D) { 2594 // D + B == C + D -> B == C 2595 Y = B; 2596 Z = C; 2597 } else if (B == C) { 2598 // A + C == C + D -> A == D 2599 Y = A; 2600 Z = D; 2601 } else { 2602 assert(B == D); 2603 // A + D == C + D -> A == C 2604 Y = A; 2605 Z = C; 2606 } 2607 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1)) 2608 return V; 2609 } 2610 } 2611 2612 { 2613 Value *Y = nullptr; 2614 // icmp pred (or X, Y), X 2615 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) { 2616 if (Pred == ICmpInst::ICMP_ULT) 2617 return getFalse(ITy); 2618 if (Pred == ICmpInst::ICMP_UGE) 2619 return getTrue(ITy); 2620 2621 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) { 2622 bool RHSKnownNonNegative, RHSKnownNegative; 2623 bool YKnownNonNegative, YKnownNegative; 2624 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, Q.DL, 0, 2625 Q.AC, Q.CxtI, Q.DT); 2626 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Q.DL, 0, Q.AC, 2627 Q.CxtI, Q.DT); 2628 if (RHSKnownNonNegative && YKnownNegative) 2629 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy); 2630 if (RHSKnownNegative || YKnownNonNegative) 2631 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy); 2632 } 2633 } 2634 // icmp pred X, (or X, Y) 2635 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) { 2636 if (Pred == ICmpInst::ICMP_ULE) 2637 return getTrue(ITy); 2638 if (Pred == ICmpInst::ICMP_UGT) 2639 return getFalse(ITy); 2640 2641 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) { 2642 bool LHSKnownNonNegative, LHSKnownNegative; 2643 bool YKnownNonNegative, YKnownNegative; 2644 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, 2645 Q.AC, Q.CxtI, Q.DT); 2646 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Q.DL, 0, Q.AC, 2647 Q.CxtI, Q.DT); 2648 if (LHSKnownNonNegative && YKnownNegative) 2649 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy); 2650 if (LHSKnownNegative || YKnownNonNegative) 2651 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy); 2652 } 2653 } 2654 } 2655 2656 // icmp pred (and X, Y), X 2657 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)), 2658 m_And(m_Specific(RHS), m_Value())))) { 2659 if (Pred == ICmpInst::ICMP_UGT) 2660 return getFalse(ITy); 2661 if (Pred == ICmpInst::ICMP_ULE) 2662 return getTrue(ITy); 2663 } 2664 // icmp pred X, (and X, Y) 2665 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)), 2666 m_And(m_Specific(LHS), m_Value())))) { 2667 if (Pred == ICmpInst::ICMP_UGE) 2668 return getTrue(ITy); 2669 if (Pred == ICmpInst::ICMP_ULT) 2670 return getFalse(ITy); 2671 } 2672 2673 // 0 - (zext X) pred C 2674 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) { 2675 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 2676 if (RHSC->getValue().isStrictlyPositive()) { 2677 if (Pred == ICmpInst::ICMP_SLT) 2678 return ConstantInt::getTrue(RHSC->getContext()); 2679 if (Pred == ICmpInst::ICMP_SGE) 2680 return ConstantInt::getFalse(RHSC->getContext()); 2681 if (Pred == ICmpInst::ICMP_EQ) 2682 return ConstantInt::getFalse(RHSC->getContext()); 2683 if (Pred == ICmpInst::ICMP_NE) 2684 return ConstantInt::getTrue(RHSC->getContext()); 2685 } 2686 if (RHSC->getValue().isNonNegative()) { 2687 if (Pred == ICmpInst::ICMP_SLE) 2688 return ConstantInt::getTrue(RHSC->getContext()); 2689 if (Pred == ICmpInst::ICMP_SGT) 2690 return ConstantInt::getFalse(RHSC->getContext()); 2691 } 2692 } 2693 } 2694 2695 // icmp pred (urem X, Y), Y 2696 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) { 2697 bool KnownNonNegative, KnownNegative; 2698 switch (Pred) { 2699 default: 2700 break; 2701 case ICmpInst::ICMP_SGT: 2702 case ICmpInst::ICMP_SGE: 2703 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2704 Q.CxtI, Q.DT); 2705 if (!KnownNonNegative) 2706 break; 2707 LLVM_FALLTHROUGH; 2708 case ICmpInst::ICMP_EQ: 2709 case ICmpInst::ICMP_UGT: 2710 case ICmpInst::ICMP_UGE: 2711 return getFalse(ITy); 2712 case ICmpInst::ICMP_SLT: 2713 case ICmpInst::ICMP_SLE: 2714 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2715 Q.CxtI, Q.DT); 2716 if (!KnownNonNegative) 2717 break; 2718 LLVM_FALLTHROUGH; 2719 case ICmpInst::ICMP_NE: 2720 case ICmpInst::ICMP_ULT: 2721 case ICmpInst::ICMP_ULE: 2722 return getTrue(ITy); 2723 } 2724 } 2725 2726 // icmp pred X, (urem Y, X) 2727 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) { 2728 bool KnownNonNegative, KnownNegative; 2729 switch (Pred) { 2730 default: 2731 break; 2732 case ICmpInst::ICMP_SGT: 2733 case ICmpInst::ICMP_SGE: 2734 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2735 Q.CxtI, Q.DT); 2736 if (!KnownNonNegative) 2737 break; 2738 LLVM_FALLTHROUGH; 2739 case ICmpInst::ICMP_NE: 2740 case ICmpInst::ICMP_UGT: 2741 case ICmpInst::ICMP_UGE: 2742 return getTrue(ITy); 2743 case ICmpInst::ICMP_SLT: 2744 case ICmpInst::ICMP_SLE: 2745 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2746 Q.CxtI, Q.DT); 2747 if (!KnownNonNegative) 2748 break; 2749 LLVM_FALLTHROUGH; 2750 case ICmpInst::ICMP_EQ: 2751 case ICmpInst::ICMP_ULT: 2752 case ICmpInst::ICMP_ULE: 2753 return getFalse(ITy); 2754 } 2755 } 2756 2757 // x >> y <=u x 2758 // x udiv y <=u x. 2759 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) || 2760 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) { 2761 // icmp pred (X op Y), X 2762 if (Pred == ICmpInst::ICMP_UGT) 2763 return getFalse(ITy); 2764 if (Pred == ICmpInst::ICMP_ULE) 2765 return getTrue(ITy); 2766 } 2767 2768 // x >=u x >> y 2769 // x >=u x udiv y. 2770 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) || 2771 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) { 2772 // icmp pred X, (X op Y) 2773 if (Pred == ICmpInst::ICMP_ULT) 2774 return getFalse(ITy); 2775 if (Pred == ICmpInst::ICMP_UGE) 2776 return getTrue(ITy); 2777 } 2778 2779 // handle: 2780 // CI2 << X == CI 2781 // CI2 << X != CI 2782 // 2783 // where CI2 is a power of 2 and CI isn't 2784 if (auto *CI = dyn_cast<ConstantInt>(RHS)) { 2785 const APInt *CI2Val, *CIVal = &CI->getValue(); 2786 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) && 2787 CI2Val->isPowerOf2()) { 2788 if (!CIVal->isPowerOf2()) { 2789 // CI2 << X can equal zero in some circumstances, 2790 // this simplification is unsafe if CI is zero. 2791 // 2792 // We know it is safe if: 2793 // - The shift is nsw, we can't shift out the one bit. 2794 // - The shift is nuw, we can't shift out the one bit. 2795 // - CI2 is one 2796 // - CI isn't zero 2797 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() || 2798 *CI2Val == 1 || !CI->isZero()) { 2799 if (Pred == ICmpInst::ICMP_EQ) 2800 return ConstantInt::getFalse(RHS->getContext()); 2801 if (Pred == ICmpInst::ICMP_NE) 2802 return ConstantInt::getTrue(RHS->getContext()); 2803 } 2804 } 2805 if (CIVal->isSignMask() && *CI2Val == 1) { 2806 if (Pred == ICmpInst::ICMP_UGT) 2807 return ConstantInt::getFalse(RHS->getContext()); 2808 if (Pred == ICmpInst::ICMP_ULE) 2809 return ConstantInt::getTrue(RHS->getContext()); 2810 } 2811 } 2812 } 2813 2814 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() && 2815 LBO->getOperand(1) == RBO->getOperand(1)) { 2816 switch (LBO->getOpcode()) { 2817 default: 2818 break; 2819 case Instruction::UDiv: 2820 case Instruction::LShr: 2821 if (ICmpInst::isSigned(Pred)) 2822 break; 2823 LLVM_FALLTHROUGH; 2824 case Instruction::SDiv: 2825 case Instruction::AShr: 2826 if (!LBO->isExact() || !RBO->isExact()) 2827 break; 2828 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2829 RBO->getOperand(0), Q, MaxRecurse - 1)) 2830 return V; 2831 break; 2832 case Instruction::Shl: { 2833 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap(); 2834 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap(); 2835 if (!NUW && !NSW) 2836 break; 2837 if (!NSW && ICmpInst::isSigned(Pred)) 2838 break; 2839 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2840 RBO->getOperand(0), Q, MaxRecurse - 1)) 2841 return V; 2842 break; 2843 } 2844 } 2845 } 2846 return nullptr; 2847 } 2848 2849 /// Simplify integer comparisons where at least one operand of the compare 2850 /// matches an integer min/max idiom. 2851 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS, 2852 Value *RHS, const SimplifyQuery &Q, 2853 unsigned MaxRecurse) { 2854 Type *ITy = GetCompareTy(LHS); // The return type. 2855 Value *A, *B; 2856 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE; 2857 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B". 2858 2859 // Signed variants on "max(a,b)>=a -> true". 2860 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2861 if (A != RHS) 2862 std::swap(A, B); // smax(A, B) pred A. 2863 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2864 // We analyze this as smax(A, B) pred A. 2865 P = Pred; 2866 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) && 2867 (A == LHS || B == LHS)) { 2868 if (A != LHS) 2869 std::swap(A, B); // A pred smax(A, B). 2870 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2871 // We analyze this as smax(A, B) swapped-pred A. 2872 P = CmpInst::getSwappedPredicate(Pred); 2873 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2874 (A == RHS || B == RHS)) { 2875 if (A != RHS) 2876 std::swap(A, B); // smin(A, B) pred A. 2877 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2878 // We analyze this as smax(-A, -B) swapped-pred -A. 2879 // Note that we do not need to actually form -A or -B thanks to EqP. 2880 P = CmpInst::getSwappedPredicate(Pred); 2881 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) && 2882 (A == LHS || B == LHS)) { 2883 if (A != LHS) 2884 std::swap(A, B); // A pred smin(A, B). 2885 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2886 // We analyze this as smax(-A, -B) pred -A. 2887 // Note that we do not need to actually form -A or -B thanks to EqP. 2888 P = Pred; 2889 } 2890 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2891 // Cases correspond to "max(A, B) p A". 2892 switch (P) { 2893 default: 2894 break; 2895 case CmpInst::ICMP_EQ: 2896 case CmpInst::ICMP_SLE: 2897 // Equivalent to "A EqP B". This may be the same as the condition tested 2898 // in the max/min; if so, we can just return that. 2899 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2900 return V; 2901 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2902 return V; 2903 // Otherwise, see if "A EqP B" simplifies. 2904 if (MaxRecurse) 2905 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1)) 2906 return V; 2907 break; 2908 case CmpInst::ICMP_NE: 2909 case CmpInst::ICMP_SGT: { 2910 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2911 // Equivalent to "A InvEqP B". This may be the same as the condition 2912 // tested in the max/min; if so, we can just return that. 2913 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2914 return V; 2915 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2916 return V; 2917 // Otherwise, see if "A InvEqP B" simplifies. 2918 if (MaxRecurse) 2919 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1)) 2920 return V; 2921 break; 2922 } 2923 case CmpInst::ICMP_SGE: 2924 // Always true. 2925 return getTrue(ITy); 2926 case CmpInst::ICMP_SLT: 2927 // Always false. 2928 return getFalse(ITy); 2929 } 2930 } 2931 2932 // Unsigned variants on "max(a,b)>=a -> true". 2933 P = CmpInst::BAD_ICMP_PREDICATE; 2934 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2935 if (A != RHS) 2936 std::swap(A, B); // umax(A, B) pred A. 2937 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2938 // We analyze this as umax(A, B) pred A. 2939 P = Pred; 2940 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) && 2941 (A == LHS || B == LHS)) { 2942 if (A != LHS) 2943 std::swap(A, B); // A pred umax(A, B). 2944 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2945 // We analyze this as umax(A, B) swapped-pred A. 2946 P = CmpInst::getSwappedPredicate(Pred); 2947 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2948 (A == RHS || B == RHS)) { 2949 if (A != RHS) 2950 std::swap(A, B); // umin(A, B) pred A. 2951 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2952 // We analyze this as umax(-A, -B) swapped-pred -A. 2953 // Note that we do not need to actually form -A or -B thanks to EqP. 2954 P = CmpInst::getSwappedPredicate(Pred); 2955 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) && 2956 (A == LHS || B == LHS)) { 2957 if (A != LHS) 2958 std::swap(A, B); // A pred umin(A, B). 2959 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2960 // We analyze this as umax(-A, -B) pred -A. 2961 // Note that we do not need to actually form -A or -B thanks to EqP. 2962 P = Pred; 2963 } 2964 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2965 // Cases correspond to "max(A, B) p A". 2966 switch (P) { 2967 default: 2968 break; 2969 case CmpInst::ICMP_EQ: 2970 case CmpInst::ICMP_ULE: 2971 // Equivalent to "A EqP B". This may be the same as the condition tested 2972 // in the max/min; if so, we can just return that. 2973 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2974 return V; 2975 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2976 return V; 2977 // Otherwise, see if "A EqP B" simplifies. 2978 if (MaxRecurse) 2979 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1)) 2980 return V; 2981 break; 2982 case CmpInst::ICMP_NE: 2983 case CmpInst::ICMP_UGT: { 2984 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2985 // Equivalent to "A InvEqP B". This may be the same as the condition 2986 // tested in the max/min; if so, we can just return that. 2987 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2988 return V; 2989 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2990 return V; 2991 // Otherwise, see if "A InvEqP B" simplifies. 2992 if (MaxRecurse) 2993 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1)) 2994 return V; 2995 break; 2996 } 2997 case CmpInst::ICMP_UGE: 2998 // Always true. 2999 return getTrue(ITy); 3000 case CmpInst::ICMP_ULT: 3001 // Always false. 3002 return getFalse(ITy); 3003 } 3004 } 3005 3006 // Variants on "max(x,y) >= min(x,z)". 3007 Value *C, *D; 3008 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && 3009 match(RHS, m_SMin(m_Value(C), m_Value(D))) && 3010 (A == C || A == D || B == C || B == D)) { 3011 // max(x, ?) pred min(x, ?). 3012 if (Pred == CmpInst::ICMP_SGE) 3013 // Always true. 3014 return getTrue(ITy); 3015 if (Pred == CmpInst::ICMP_SLT) 3016 // Always false. 3017 return getFalse(ITy); 3018 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 3019 match(RHS, m_SMax(m_Value(C), m_Value(D))) && 3020 (A == C || A == D || B == C || B == D)) { 3021 // min(x, ?) pred max(x, ?). 3022 if (Pred == CmpInst::ICMP_SLE) 3023 // Always true. 3024 return getTrue(ITy); 3025 if (Pred == CmpInst::ICMP_SGT) 3026 // Always false. 3027 return getFalse(ITy); 3028 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && 3029 match(RHS, m_UMin(m_Value(C), m_Value(D))) && 3030 (A == C || A == D || B == C || B == D)) { 3031 // max(x, ?) pred min(x, ?). 3032 if (Pred == CmpInst::ICMP_UGE) 3033 // Always true. 3034 return getTrue(ITy); 3035 if (Pred == CmpInst::ICMP_ULT) 3036 // Always false. 3037 return getFalse(ITy); 3038 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 3039 match(RHS, m_UMax(m_Value(C), m_Value(D))) && 3040 (A == C || A == D || B == C || B == D)) { 3041 // min(x, ?) pred max(x, ?). 3042 if (Pred == CmpInst::ICMP_ULE) 3043 // Always true. 3044 return getTrue(ITy); 3045 if (Pred == CmpInst::ICMP_UGT) 3046 // Always false. 3047 return getFalse(ITy); 3048 } 3049 3050 return nullptr; 3051 } 3052 3053 /// Given operands for an ICmpInst, see if we can fold the result. 3054 /// If not, this returns null. 3055 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3056 const SimplifyQuery &Q, unsigned MaxRecurse) { 3057 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 3058 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!"); 3059 3060 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 3061 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 3062 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 3063 3064 // If we have a constant, make sure it is on the RHS. 3065 std::swap(LHS, RHS); 3066 Pred = CmpInst::getSwappedPredicate(Pred); 3067 } 3068 3069 Type *ITy = GetCompareTy(LHS); // The return type. 3070 3071 // icmp X, X -> true/false 3072 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false 3073 // because X could be 0. 3074 if (LHS == RHS || isa<UndefValue>(RHS)) 3075 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred)); 3076 3077 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q)) 3078 return V; 3079 3080 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q)) 3081 return V; 3082 3083 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS)) 3084 return V; 3085 3086 // If both operands have range metadata, use the metadata 3087 // to simplify the comparison. 3088 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) { 3089 auto RHS_Instr = cast<Instruction>(RHS); 3090 auto LHS_Instr = cast<Instruction>(LHS); 3091 3092 if (RHS_Instr->getMetadata(LLVMContext::MD_range) && 3093 LHS_Instr->getMetadata(LLVMContext::MD_range)) { 3094 auto RHS_CR = getConstantRangeFromMetadata( 3095 *RHS_Instr->getMetadata(LLVMContext::MD_range)); 3096 auto LHS_CR = getConstantRangeFromMetadata( 3097 *LHS_Instr->getMetadata(LLVMContext::MD_range)); 3098 3099 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR); 3100 if (Satisfied_CR.contains(LHS_CR)) 3101 return ConstantInt::getTrue(RHS->getContext()); 3102 3103 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion( 3104 CmpInst::getInversePredicate(Pred), RHS_CR); 3105 if (InversedSatisfied_CR.contains(LHS_CR)) 3106 return ConstantInt::getFalse(RHS->getContext()); 3107 } 3108 } 3109 3110 // Compare of cast, for example (zext X) != 0 -> X != 0 3111 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) { 3112 Instruction *LI = cast<CastInst>(LHS); 3113 Value *SrcOp = LI->getOperand(0); 3114 Type *SrcTy = SrcOp->getType(); 3115 Type *DstTy = LI->getType(); 3116 3117 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input 3118 // if the integer type is the same size as the pointer type. 3119 if (MaxRecurse && isa<PtrToIntInst>(LI) && 3120 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) { 3121 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 3122 // Transfer the cast to the constant. 3123 if (Value *V = SimplifyICmpInst(Pred, SrcOp, 3124 ConstantExpr::getIntToPtr(RHSC, SrcTy), 3125 Q, MaxRecurse-1)) 3126 return V; 3127 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) { 3128 if (RI->getOperand(0)->getType() == SrcTy) 3129 // Compare without the cast. 3130 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 3131 Q, MaxRecurse-1)) 3132 return V; 3133 } 3134 } 3135 3136 if (isa<ZExtInst>(LHS)) { 3137 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the 3138 // same type. 3139 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) { 3140 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 3141 // Compare X and Y. Note that signed predicates become unsigned. 3142 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 3143 SrcOp, RI->getOperand(0), Q, 3144 MaxRecurse-1)) 3145 return V; 3146 } 3147 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended 3148 // too. If not, then try to deduce the result of the comparison. 3149 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 3150 // Compute the constant that would happen if we truncated to SrcTy then 3151 // reextended to DstTy. 3152 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 3153 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy); 3154 3155 // If the re-extended constant didn't change then this is effectively 3156 // also a case of comparing two zero-extended values. 3157 if (RExt == CI && MaxRecurse) 3158 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 3159 SrcOp, Trunc, Q, MaxRecurse-1)) 3160 return V; 3161 3162 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit 3163 // there. Use this to work out the result of the comparison. 3164 if (RExt != CI) { 3165 switch (Pred) { 3166 default: llvm_unreachable("Unknown ICmp predicate!"); 3167 // LHS <u RHS. 3168 case ICmpInst::ICMP_EQ: 3169 case ICmpInst::ICMP_UGT: 3170 case ICmpInst::ICMP_UGE: 3171 return ConstantInt::getFalse(CI->getContext()); 3172 3173 case ICmpInst::ICMP_NE: 3174 case ICmpInst::ICMP_ULT: 3175 case ICmpInst::ICMP_ULE: 3176 return ConstantInt::getTrue(CI->getContext()); 3177 3178 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS 3179 // is non-negative then LHS <s RHS. 3180 case ICmpInst::ICMP_SGT: 3181 case ICmpInst::ICMP_SGE: 3182 return CI->getValue().isNegative() ? 3183 ConstantInt::getTrue(CI->getContext()) : 3184 ConstantInt::getFalse(CI->getContext()); 3185 3186 case ICmpInst::ICMP_SLT: 3187 case ICmpInst::ICMP_SLE: 3188 return CI->getValue().isNegative() ? 3189 ConstantInt::getFalse(CI->getContext()) : 3190 ConstantInt::getTrue(CI->getContext()); 3191 } 3192 } 3193 } 3194 } 3195 3196 if (isa<SExtInst>(LHS)) { 3197 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the 3198 // same type. 3199 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) { 3200 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 3201 // Compare X and Y. Note that the predicate does not change. 3202 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 3203 Q, MaxRecurse-1)) 3204 return V; 3205 } 3206 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended 3207 // too. If not, then try to deduce the result of the comparison. 3208 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 3209 // Compute the constant that would happen if we truncated to SrcTy then 3210 // reextended to DstTy. 3211 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 3212 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy); 3213 3214 // If the re-extended constant didn't change then this is effectively 3215 // also a case of comparing two sign-extended values. 3216 if (RExt == CI && MaxRecurse) 3217 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1)) 3218 return V; 3219 3220 // Otherwise the upper bits of LHS are all equal, while RHS has varying 3221 // bits there. Use this to work out the result of the comparison. 3222 if (RExt != CI) { 3223 switch (Pred) { 3224 default: llvm_unreachable("Unknown ICmp predicate!"); 3225 case ICmpInst::ICMP_EQ: 3226 return ConstantInt::getFalse(CI->getContext()); 3227 case ICmpInst::ICMP_NE: 3228 return ConstantInt::getTrue(CI->getContext()); 3229 3230 // If RHS is non-negative then LHS <s RHS. If RHS is negative then 3231 // LHS >s RHS. 3232 case ICmpInst::ICMP_SGT: 3233 case ICmpInst::ICMP_SGE: 3234 return CI->getValue().isNegative() ? 3235 ConstantInt::getTrue(CI->getContext()) : 3236 ConstantInt::getFalse(CI->getContext()); 3237 case ICmpInst::ICMP_SLT: 3238 case ICmpInst::ICMP_SLE: 3239 return CI->getValue().isNegative() ? 3240 ConstantInt::getFalse(CI->getContext()) : 3241 ConstantInt::getTrue(CI->getContext()); 3242 3243 // If LHS is non-negative then LHS <u RHS. If LHS is negative then 3244 // LHS >u RHS. 3245 case ICmpInst::ICMP_UGT: 3246 case ICmpInst::ICMP_UGE: 3247 // Comparison is true iff the LHS <s 0. 3248 if (MaxRecurse) 3249 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp, 3250 Constant::getNullValue(SrcTy), 3251 Q, MaxRecurse-1)) 3252 return V; 3253 break; 3254 case ICmpInst::ICMP_ULT: 3255 case ICmpInst::ICMP_ULE: 3256 // Comparison is true iff the LHS >=s 0. 3257 if (MaxRecurse) 3258 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp, 3259 Constant::getNullValue(SrcTy), 3260 Q, MaxRecurse-1)) 3261 return V; 3262 break; 3263 } 3264 } 3265 } 3266 } 3267 } 3268 3269 // icmp eq|ne X, Y -> false|true if X != Y 3270 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) && 3271 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) { 3272 LLVMContext &Ctx = LHS->getType()->getContext(); 3273 return Pred == ICmpInst::ICMP_NE ? 3274 ConstantInt::getTrue(Ctx) : ConstantInt::getFalse(Ctx); 3275 } 3276 3277 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse)) 3278 return V; 3279 3280 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse)) 3281 return V; 3282 3283 // Simplify comparisons of related pointers using a powerful, recursive 3284 // GEP-walk when we have target data available.. 3285 if (LHS->getType()->isPointerTy()) 3286 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, LHS, RHS)) 3287 return C; 3288 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS)) 3289 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS)) 3290 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) == 3291 Q.DL.getTypeSizeInBits(CLHS->getType()) && 3292 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) == 3293 Q.DL.getTypeSizeInBits(CRHS->getType())) 3294 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, 3295 CLHS->getPointerOperand(), 3296 CRHS->getPointerOperand())) 3297 return C; 3298 3299 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) { 3300 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) { 3301 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() && 3302 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() && 3303 (ICmpInst::isEquality(Pred) || 3304 (GLHS->isInBounds() && GRHS->isInBounds() && 3305 Pred == ICmpInst::getSignedPredicate(Pred)))) { 3306 // The bases are equal and the indices are constant. Build a constant 3307 // expression GEP with the same indices and a null base pointer to see 3308 // what constant folding can make out of it. 3309 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType()); 3310 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end()); 3311 Constant *NewLHS = ConstantExpr::getGetElementPtr( 3312 GLHS->getSourceElementType(), Null, IndicesLHS); 3313 3314 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end()); 3315 Constant *NewRHS = ConstantExpr::getGetElementPtr( 3316 GLHS->getSourceElementType(), Null, IndicesRHS); 3317 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS); 3318 } 3319 } 3320 } 3321 3322 // If a bit is known to be zero for A and known to be one for B, 3323 // then A and B cannot be equal. 3324 if (ICmpInst::isEquality(Pred)) { 3325 const APInt *RHSVal; 3326 if (match(RHS, m_APInt(RHSVal))) { 3327 unsigned BitWidth = RHSVal->getBitWidth(); 3328 KnownBits LHSKnown(BitWidth); 3329 computeKnownBits(LHS, LHSKnown, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT); 3330 if (LHSKnown.Zero.intersects(*RHSVal) || 3331 !LHSKnown.One.isSubsetOf(*RHSVal)) 3332 return Pred == ICmpInst::ICMP_EQ ? ConstantInt::getFalse(ITy) 3333 : ConstantInt::getTrue(ITy); 3334 } 3335 } 3336 3337 // If the comparison is with the result of a select instruction, check whether 3338 // comparing with either branch of the select always yields the same value. 3339 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3340 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3341 return V; 3342 3343 // If the comparison is with the result of a phi instruction, check whether 3344 // doing the compare with each incoming phi value yields a common result. 3345 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3346 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3347 return V; 3348 3349 return nullptr; 3350 } 3351 3352 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3353 const SimplifyQuery &Q) { 3354 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit); 3355 } 3356 3357 /// Given operands for an FCmpInst, see if we can fold the result. 3358 /// If not, this returns null. 3359 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3360 FastMathFlags FMF, const SimplifyQuery &Q, 3361 unsigned MaxRecurse) { 3362 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 3363 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!"); 3364 3365 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 3366 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 3367 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 3368 3369 // If we have a constant, make sure it is on the RHS. 3370 std::swap(LHS, RHS); 3371 Pred = CmpInst::getSwappedPredicate(Pred); 3372 } 3373 3374 // Fold trivial predicates. 3375 Type *RetTy = GetCompareTy(LHS); 3376 if (Pred == FCmpInst::FCMP_FALSE) 3377 return getFalse(RetTy); 3378 if (Pred == FCmpInst::FCMP_TRUE) 3379 return getTrue(RetTy); 3380 3381 // UNO/ORD predicates can be trivially folded if NaNs are ignored. 3382 if (FMF.noNaNs()) { 3383 if (Pred == FCmpInst::FCMP_UNO) 3384 return getFalse(RetTy); 3385 if (Pred == FCmpInst::FCMP_ORD) 3386 return getTrue(RetTy); 3387 } 3388 3389 // fcmp pred x, undef and fcmp pred undef, x 3390 // fold to true if unordered, false if ordered 3391 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) { 3392 // Choosing NaN for the undef will always make unordered comparison succeed 3393 // and ordered comparison fail. 3394 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred)); 3395 } 3396 3397 // fcmp x,x -> true/false. Not all compares are foldable. 3398 if (LHS == RHS) { 3399 if (CmpInst::isTrueWhenEqual(Pred)) 3400 return getTrue(RetTy); 3401 if (CmpInst::isFalseWhenEqual(Pred)) 3402 return getFalse(RetTy); 3403 } 3404 3405 // Handle fcmp with constant RHS 3406 const ConstantFP *CFP = nullptr; 3407 if (const auto *RHSC = dyn_cast<Constant>(RHS)) { 3408 if (RHS->getType()->isVectorTy()) 3409 CFP = dyn_cast_or_null<ConstantFP>(RHSC->getSplatValue()); 3410 else 3411 CFP = dyn_cast<ConstantFP>(RHSC); 3412 } 3413 if (CFP) { 3414 // If the constant is a nan, see if we can fold the comparison based on it. 3415 if (CFP->getValueAPF().isNaN()) { 3416 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo" 3417 return getFalse(RetTy); 3418 assert(FCmpInst::isUnordered(Pred) && 3419 "Comparison must be either ordered or unordered!"); 3420 // True if unordered. 3421 return getTrue(RetTy); 3422 } 3423 // Check whether the constant is an infinity. 3424 if (CFP->getValueAPF().isInfinity()) { 3425 if (CFP->getValueAPF().isNegative()) { 3426 switch (Pred) { 3427 case FCmpInst::FCMP_OLT: 3428 // No value is ordered and less than negative infinity. 3429 return getFalse(RetTy); 3430 case FCmpInst::FCMP_UGE: 3431 // All values are unordered with or at least negative infinity. 3432 return getTrue(RetTy); 3433 default: 3434 break; 3435 } 3436 } else { 3437 switch (Pred) { 3438 case FCmpInst::FCMP_OGT: 3439 // No value is ordered and greater than infinity. 3440 return getFalse(RetTy); 3441 case FCmpInst::FCMP_ULE: 3442 // All values are unordered with and at most infinity. 3443 return getTrue(RetTy); 3444 default: 3445 break; 3446 } 3447 } 3448 } 3449 if (CFP->getValueAPF().isZero()) { 3450 switch (Pred) { 3451 case FCmpInst::FCMP_UGE: 3452 if (CannotBeOrderedLessThanZero(LHS, Q.TLI)) 3453 return getTrue(RetTy); 3454 break; 3455 case FCmpInst::FCMP_OLT: 3456 // X < 0 3457 if (CannotBeOrderedLessThanZero(LHS, Q.TLI)) 3458 return getFalse(RetTy); 3459 break; 3460 default: 3461 break; 3462 } 3463 } 3464 } 3465 3466 // If the comparison is with the result of a select instruction, check whether 3467 // comparing with either branch of the select always yields the same value. 3468 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3469 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3470 return V; 3471 3472 // If the comparison is with the result of a phi instruction, check whether 3473 // doing the compare with each incoming phi value yields a common result. 3474 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3475 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3476 return V; 3477 3478 return nullptr; 3479 } 3480 3481 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3482 FastMathFlags FMF, const SimplifyQuery &Q) { 3483 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit); 3484 } 3485 3486 /// See if V simplifies when its operand Op is replaced with RepOp. 3487 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp, 3488 const SimplifyQuery &Q, 3489 unsigned MaxRecurse) { 3490 // Trivial replacement. 3491 if (V == Op) 3492 return RepOp; 3493 3494 auto *I = dyn_cast<Instruction>(V); 3495 if (!I) 3496 return nullptr; 3497 3498 // If this is a binary operator, try to simplify it with the replaced op. 3499 if (auto *B = dyn_cast<BinaryOperator>(I)) { 3500 // Consider: 3501 // %cmp = icmp eq i32 %x, 2147483647 3502 // %add = add nsw i32 %x, 1 3503 // %sel = select i1 %cmp, i32 -2147483648, i32 %add 3504 // 3505 // We can't replace %sel with %add unless we strip away the flags. 3506 if (isa<OverflowingBinaryOperator>(B)) 3507 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap()) 3508 return nullptr; 3509 if (isa<PossiblyExactOperator>(B)) 3510 if (B->isExact()) 3511 return nullptr; 3512 3513 if (MaxRecurse) { 3514 if (B->getOperand(0) == Op) 3515 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q, 3516 MaxRecurse - 1); 3517 if (B->getOperand(1) == Op) 3518 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q, 3519 MaxRecurse - 1); 3520 } 3521 } 3522 3523 // Same for CmpInsts. 3524 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 3525 if (MaxRecurse) { 3526 if (C->getOperand(0) == Op) 3527 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q, 3528 MaxRecurse - 1); 3529 if (C->getOperand(1) == Op) 3530 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q, 3531 MaxRecurse - 1); 3532 } 3533 } 3534 3535 // TODO: We could hand off more cases to instsimplify here. 3536 3537 // If all operands are constant after substituting Op for RepOp then we can 3538 // constant fold the instruction. 3539 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) { 3540 // Build a list of all constant operands. 3541 SmallVector<Constant *, 8> ConstOps; 3542 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 3543 if (I->getOperand(i) == Op) 3544 ConstOps.push_back(CRepOp); 3545 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i))) 3546 ConstOps.push_back(COp); 3547 else 3548 break; 3549 } 3550 3551 // All operands were constants, fold it. 3552 if (ConstOps.size() == I->getNumOperands()) { 3553 if (CmpInst *C = dyn_cast<CmpInst>(I)) 3554 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0], 3555 ConstOps[1], Q.DL, Q.TLI); 3556 3557 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 3558 if (!LI->isVolatile()) 3559 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL); 3560 3561 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI); 3562 } 3563 } 3564 3565 return nullptr; 3566 } 3567 3568 /// Try to simplify a select instruction when its condition operand is an 3569 /// integer comparison where one operand of the compare is a constant. 3570 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X, 3571 const APInt *Y, bool TrueWhenUnset) { 3572 const APInt *C; 3573 3574 // (X & Y) == 0 ? X & ~Y : X --> X 3575 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y 3576 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) && 3577 *Y == ~*C) 3578 return TrueWhenUnset ? FalseVal : TrueVal; 3579 3580 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y 3581 // (X & Y) != 0 ? X : X & ~Y --> X 3582 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) && 3583 *Y == ~*C) 3584 return TrueWhenUnset ? FalseVal : TrueVal; 3585 3586 if (Y->isPowerOf2()) { 3587 // (X & Y) == 0 ? X | Y : X --> X | Y 3588 // (X & Y) != 0 ? X | Y : X --> X 3589 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) && 3590 *Y == *C) 3591 return TrueWhenUnset ? TrueVal : FalseVal; 3592 3593 // (X & Y) == 0 ? X : X | Y --> X 3594 // (X & Y) != 0 ? X : X | Y --> X | Y 3595 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) && 3596 *Y == *C) 3597 return TrueWhenUnset ? TrueVal : FalseVal; 3598 } 3599 3600 return nullptr; 3601 } 3602 3603 /// An alternative way to test if a bit is set or not uses sgt/slt instead of 3604 /// eq/ne. 3605 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *TrueVal, 3606 Value *FalseVal, 3607 bool TrueWhenUnset) { 3608 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); 3609 if (!BitWidth) 3610 return nullptr; 3611 3612 APInt MinSignedValue; 3613 Value *X; 3614 if (match(CmpLHS, m_Trunc(m_Value(X))) && (X == TrueVal || X == FalseVal)) { 3615 // icmp slt (trunc X), 0 <--> icmp ne (and X, C), 0 3616 // icmp sgt (trunc X), -1 <--> icmp eq (and X, C), 0 3617 unsigned DestSize = CmpLHS->getType()->getScalarSizeInBits(); 3618 MinSignedValue = APInt::getSignedMinValue(DestSize).zext(BitWidth); 3619 } else { 3620 // icmp slt X, 0 <--> icmp ne (and X, C), 0 3621 // icmp sgt X, -1 <--> icmp eq (and X, C), 0 3622 X = CmpLHS; 3623 MinSignedValue = APInt::getSignedMinValue(BitWidth); 3624 } 3625 3626 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, &MinSignedValue, 3627 TrueWhenUnset)) 3628 return V; 3629 3630 return nullptr; 3631 } 3632 3633 /// Try to simplify a select instruction when its condition operand is an 3634 /// integer comparison. 3635 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal, 3636 Value *FalseVal, const SimplifyQuery &Q, 3637 unsigned MaxRecurse) { 3638 ICmpInst::Predicate Pred; 3639 Value *CmpLHS, *CmpRHS; 3640 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)))) 3641 return nullptr; 3642 3643 // FIXME: This code is nearly duplicated in InstCombine. Using/refactoring 3644 // decomposeBitTestICmp() might help. 3645 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) { 3646 Value *X; 3647 const APInt *Y; 3648 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y)))) 3649 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y, 3650 Pred == ICmpInst::ICMP_EQ)) 3651 return V; 3652 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { 3653 // Comparing signed-less-than 0 checks if the sign bit is set. 3654 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal, 3655 false)) 3656 return V; 3657 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { 3658 // Comparing signed-greater-than -1 checks if the sign bit is not set. 3659 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal, 3660 true)) 3661 return V; 3662 } 3663 3664 if (CondVal->hasOneUse()) { 3665 const APInt *C; 3666 if (match(CmpRHS, m_APInt(C))) { 3667 // X < MIN ? T : F --> F 3668 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue()) 3669 return FalseVal; 3670 // X < MIN ? T : F --> F 3671 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue()) 3672 return FalseVal; 3673 // X > MAX ? T : F --> F 3674 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue()) 3675 return FalseVal; 3676 // X > MAX ? T : F --> F 3677 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue()) 3678 return FalseVal; 3679 } 3680 } 3681 3682 // If we have an equality comparison, then we know the value in one of the 3683 // arms of the select. See if substituting this value into the arm and 3684 // simplifying the result yields the same value as the other arm. 3685 if (Pred == ICmpInst::ICMP_EQ) { 3686 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3687 TrueVal || 3688 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3689 TrueVal) 3690 return FalseVal; 3691 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3692 FalseVal || 3693 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3694 FalseVal) 3695 return FalseVal; 3696 } else if (Pred == ICmpInst::ICMP_NE) { 3697 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3698 FalseVal || 3699 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3700 FalseVal) 3701 return TrueVal; 3702 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3703 TrueVal || 3704 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3705 TrueVal) 3706 return TrueVal; 3707 } 3708 3709 return nullptr; 3710 } 3711 3712 /// Given operands for a SelectInst, see if we can fold the result. 3713 /// If not, this returns null. 3714 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal, 3715 Value *FalseVal, const SimplifyQuery &Q, 3716 unsigned MaxRecurse) { 3717 // select true, X, Y -> X 3718 // select false, X, Y -> Y 3719 if (Constant *CB = dyn_cast<Constant>(CondVal)) { 3720 if (CB->isAllOnesValue()) 3721 return TrueVal; 3722 if (CB->isNullValue()) 3723 return FalseVal; 3724 } 3725 3726 // select C, X, X -> X 3727 if (TrueVal == FalseVal) 3728 return TrueVal; 3729 3730 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y 3731 if (isa<Constant>(FalseVal)) 3732 return FalseVal; 3733 return TrueVal; 3734 } 3735 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X 3736 return FalseVal; 3737 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X 3738 return TrueVal; 3739 3740 if (Value *V = 3741 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse)) 3742 return V; 3743 3744 return nullptr; 3745 } 3746 3747 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, 3748 const SimplifyQuery &Q) { 3749 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit); 3750 } 3751 3752 /// Given operands for an GetElementPtrInst, see if we can fold the result. 3753 /// If not, this returns null. 3754 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, 3755 const SimplifyQuery &Q, unsigned) { 3756 // The type of the GEP pointer operand. 3757 unsigned AS = 3758 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace(); 3759 3760 // getelementptr P -> P. 3761 if (Ops.size() == 1) 3762 return Ops[0]; 3763 3764 // Compute the (pointer) type returned by the GEP instruction. 3765 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1)); 3766 Type *GEPTy = PointerType::get(LastType, AS); 3767 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType())) 3768 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 3769 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType())) 3770 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 3771 3772 if (isa<UndefValue>(Ops[0])) 3773 return UndefValue::get(GEPTy); 3774 3775 if (Ops.size() == 2) { 3776 // getelementptr P, 0 -> P. 3777 if (match(Ops[1], m_Zero())) 3778 return Ops[0]; 3779 3780 Type *Ty = SrcTy; 3781 if (Ty->isSized()) { 3782 Value *P; 3783 uint64_t C; 3784 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty); 3785 // getelementptr P, N -> P if P points to a type of zero size. 3786 if (TyAllocSize == 0) 3787 return Ops[0]; 3788 3789 // The following transforms are only safe if the ptrtoint cast 3790 // doesn't truncate the pointers. 3791 if (Ops[1]->getType()->getScalarSizeInBits() == 3792 Q.DL.getPointerSizeInBits(AS)) { 3793 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * { 3794 if (match(P, m_Zero())) 3795 return Constant::getNullValue(GEPTy); 3796 Value *Temp; 3797 if (match(P, m_PtrToInt(m_Value(Temp)))) 3798 if (Temp->getType() == GEPTy) 3799 return Temp; 3800 return nullptr; 3801 }; 3802 3803 // getelementptr V, (sub P, V) -> P if P points to a type of size 1. 3804 if (TyAllocSize == 1 && 3805 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))))) 3806 if (Value *R = PtrToIntOrZero(P)) 3807 return R; 3808 3809 // getelementptr V, (ashr (sub P, V), C) -> Q 3810 // if P points to a type of size 1 << C. 3811 if (match(Ops[1], 3812 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3813 m_ConstantInt(C))) && 3814 TyAllocSize == 1ULL << C) 3815 if (Value *R = PtrToIntOrZero(P)) 3816 return R; 3817 3818 // getelementptr V, (sdiv (sub P, V), C) -> Q 3819 // if P points to a type of size C. 3820 if (match(Ops[1], 3821 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3822 m_SpecificInt(TyAllocSize)))) 3823 if (Value *R = PtrToIntOrZero(P)) 3824 return R; 3825 } 3826 } 3827 } 3828 3829 if (Q.DL.getTypeAllocSize(LastType) == 1 && 3830 all_of(Ops.slice(1).drop_back(1), 3831 [](Value *Idx) { return match(Idx, m_Zero()); })) { 3832 unsigned PtrWidth = 3833 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace()); 3834 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) { 3835 APInt BasePtrOffset(PtrWidth, 0); 3836 Value *StrippedBasePtr = 3837 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL, 3838 BasePtrOffset); 3839 3840 // gep (gep V, C), (sub 0, V) -> C 3841 if (match(Ops.back(), 3842 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) { 3843 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset); 3844 return ConstantExpr::getIntToPtr(CI, GEPTy); 3845 } 3846 // gep (gep V, C), (xor V, -1) -> C-1 3847 if (match(Ops.back(), 3848 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) { 3849 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1); 3850 return ConstantExpr::getIntToPtr(CI, GEPTy); 3851 } 3852 } 3853 } 3854 3855 // Check to see if this is constant foldable. 3856 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 3857 if (!isa<Constant>(Ops[i])) 3858 return nullptr; 3859 3860 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]), 3861 Ops.slice(1)); 3862 } 3863 3864 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, 3865 const SimplifyQuery &Q) { 3866 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit); 3867 } 3868 3869 /// Given operands for an InsertValueInst, see if we can fold the result. 3870 /// If not, this returns null. 3871 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val, 3872 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q, 3873 unsigned) { 3874 if (Constant *CAgg = dyn_cast<Constant>(Agg)) 3875 if (Constant *CVal = dyn_cast<Constant>(Val)) 3876 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs); 3877 3878 // insertvalue x, undef, n -> x 3879 if (match(Val, m_Undef())) 3880 return Agg; 3881 3882 // insertvalue x, (extractvalue y, n), n 3883 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val)) 3884 if (EV->getAggregateOperand()->getType() == Agg->getType() && 3885 EV->getIndices() == Idxs) { 3886 // insertvalue undef, (extractvalue y, n), n -> y 3887 if (match(Agg, m_Undef())) 3888 return EV->getAggregateOperand(); 3889 3890 // insertvalue y, (extractvalue y, n), n -> y 3891 if (Agg == EV->getAggregateOperand()) 3892 return Agg; 3893 } 3894 3895 return nullptr; 3896 } 3897 3898 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val, 3899 ArrayRef<unsigned> Idxs, 3900 const SimplifyQuery &Q) { 3901 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit); 3902 } 3903 3904 /// Given operands for an ExtractValueInst, see if we can fold the result. 3905 /// If not, this returns null. 3906 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, 3907 const SimplifyQuery &, unsigned) { 3908 if (auto *CAgg = dyn_cast<Constant>(Agg)) 3909 return ConstantFoldExtractValueInstruction(CAgg, Idxs); 3910 3911 // extractvalue x, (insertvalue y, elt, n), n -> elt 3912 unsigned NumIdxs = Idxs.size(); 3913 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr; 3914 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) { 3915 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices(); 3916 unsigned NumInsertValueIdxs = InsertValueIdxs.size(); 3917 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs); 3918 if (InsertValueIdxs.slice(0, NumCommonIdxs) == 3919 Idxs.slice(0, NumCommonIdxs)) { 3920 if (NumIdxs == NumInsertValueIdxs) 3921 return IVI->getInsertedValueOperand(); 3922 break; 3923 } 3924 } 3925 3926 return nullptr; 3927 } 3928 3929 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, 3930 const SimplifyQuery &Q) { 3931 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit); 3932 } 3933 3934 /// Given operands for an ExtractElementInst, see if we can fold the result. 3935 /// If not, this returns null. 3936 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &, 3937 unsigned) { 3938 if (auto *CVec = dyn_cast<Constant>(Vec)) { 3939 if (auto *CIdx = dyn_cast<Constant>(Idx)) 3940 return ConstantFoldExtractElementInstruction(CVec, CIdx); 3941 3942 // The index is not relevant if our vector is a splat. 3943 if (auto *Splat = CVec->getSplatValue()) 3944 return Splat; 3945 3946 if (isa<UndefValue>(Vec)) 3947 return UndefValue::get(Vec->getType()->getVectorElementType()); 3948 } 3949 3950 // If extracting a specified index from the vector, see if we can recursively 3951 // find a previously computed scalar that was inserted into the vector. 3952 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) 3953 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue())) 3954 return Elt; 3955 3956 return nullptr; 3957 } 3958 3959 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx, 3960 const SimplifyQuery &Q) { 3961 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit); 3962 } 3963 3964 /// See if we can fold the given phi. If not, returns null. 3965 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) { 3966 // If all of the PHI's incoming values are the same then replace the PHI node 3967 // with the common value. 3968 Value *CommonValue = nullptr; 3969 bool HasUndefInput = false; 3970 for (Value *Incoming : PN->incoming_values()) { 3971 // If the incoming value is the phi node itself, it can safely be skipped. 3972 if (Incoming == PN) continue; 3973 if (isa<UndefValue>(Incoming)) { 3974 // Remember that we saw an undef value, but otherwise ignore them. 3975 HasUndefInput = true; 3976 continue; 3977 } 3978 if (CommonValue && Incoming != CommonValue) 3979 return nullptr; // Not the same, bail out. 3980 CommonValue = Incoming; 3981 } 3982 3983 // If CommonValue is null then all of the incoming values were either undef or 3984 // equal to the phi node itself. 3985 if (!CommonValue) 3986 return UndefValue::get(PN->getType()); 3987 3988 // If we have a PHI node like phi(X, undef, X), where X is defined by some 3989 // instruction, we cannot return X as the result of the PHI node unless it 3990 // dominates the PHI block. 3991 if (HasUndefInput) 3992 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr; 3993 3994 return CommonValue; 3995 } 3996 3997 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op, 3998 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) { 3999 if (auto *C = dyn_cast<Constant>(Op)) 4000 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL); 4001 4002 if (auto *CI = dyn_cast<CastInst>(Op)) { 4003 auto *Src = CI->getOperand(0); 4004 Type *SrcTy = Src->getType(); 4005 Type *MidTy = CI->getType(); 4006 Type *DstTy = Ty; 4007 if (Src->getType() == Ty) { 4008 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode()); 4009 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc); 4010 Type *SrcIntPtrTy = 4011 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr; 4012 Type *MidIntPtrTy = 4013 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr; 4014 Type *DstIntPtrTy = 4015 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr; 4016 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy, 4017 SrcIntPtrTy, MidIntPtrTy, 4018 DstIntPtrTy) == Instruction::BitCast) 4019 return Src; 4020 } 4021 } 4022 4023 // bitcast x -> x 4024 if (CastOpc == Instruction::BitCast) 4025 if (Op->getType() == Ty) 4026 return Op; 4027 4028 return nullptr; 4029 } 4030 4031 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty, 4032 const SimplifyQuery &Q) { 4033 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit); 4034 } 4035 4036 /// For the given destination element of a shuffle, peek through shuffles to 4037 /// match a root vector source operand that contains that element in the same 4038 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s). 4039 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1, 4040 int MaskVal, Value *RootVec, 4041 unsigned MaxRecurse) { 4042 if (!MaxRecurse--) 4043 return nullptr; 4044 4045 // Bail out if any mask value is undefined. That kind of shuffle may be 4046 // simplified further based on demanded bits or other folds. 4047 if (MaskVal == -1) 4048 return nullptr; 4049 4050 // The mask value chooses which source operand we need to look at next. 4051 int InVecNumElts = Op0->getType()->getVectorNumElements(); 4052 int RootElt = MaskVal; 4053 Value *SourceOp = Op0; 4054 if (MaskVal >= InVecNumElts) { 4055 RootElt = MaskVal - InVecNumElts; 4056 SourceOp = Op1; 4057 } 4058 4059 // If the source operand is a shuffle itself, look through it to find the 4060 // matching root vector. 4061 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) { 4062 return foldIdentityShuffles( 4063 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1), 4064 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse); 4065 } 4066 4067 // TODO: Look through bitcasts? What if the bitcast changes the vector element 4068 // size? 4069 4070 // The source operand is not a shuffle. Initialize the root vector value for 4071 // this shuffle if that has not been done yet. 4072 if (!RootVec) 4073 RootVec = SourceOp; 4074 4075 // Give up as soon as a source operand does not match the existing root value. 4076 if (RootVec != SourceOp) 4077 return nullptr; 4078 4079 // The element must be coming from the same lane in the source vector 4080 // (although it may have crossed lanes in intermediate shuffles). 4081 if (RootElt != DestElt) 4082 return nullptr; 4083 4084 return RootVec; 4085 } 4086 4087 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask, 4088 Type *RetTy, const SimplifyQuery &Q, 4089 unsigned MaxRecurse) { 4090 if (isa<UndefValue>(Mask)) 4091 return UndefValue::get(RetTy); 4092 4093 Type *InVecTy = Op0->getType(); 4094 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 4095 unsigned InVecNumElts = InVecTy->getVectorNumElements(); 4096 4097 SmallVector<int, 32> Indices; 4098 ShuffleVectorInst::getShuffleMask(Mask, Indices); 4099 assert(MaskNumElts == Indices.size() && 4100 "Size of Indices not same as number of mask elements?"); 4101 4102 // Canonicalization: If mask does not select elements from an input vector, 4103 // replace that input vector with undef. 4104 bool MaskSelects0 = false, MaskSelects1 = false; 4105 for (unsigned i = 0; i != MaskNumElts; ++i) { 4106 if (Indices[i] == -1) 4107 continue; 4108 if ((unsigned)Indices[i] < InVecNumElts) 4109 MaskSelects0 = true; 4110 else 4111 MaskSelects1 = true; 4112 } 4113 if (!MaskSelects0) 4114 Op0 = UndefValue::get(InVecTy); 4115 if (!MaskSelects1) 4116 Op1 = UndefValue::get(InVecTy); 4117 4118 auto *Op0Const = dyn_cast<Constant>(Op0); 4119 auto *Op1Const = dyn_cast<Constant>(Op1); 4120 4121 // If all operands are constant, constant fold the shuffle. 4122 if (Op0Const && Op1Const) 4123 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask); 4124 4125 // Canonicalization: if only one input vector is constant, it shall be the 4126 // second one. 4127 if (Op0Const && !Op1Const) { 4128 std::swap(Op0, Op1); 4129 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts); 4130 } 4131 4132 // A shuffle of a splat is always the splat itself. Legal if the shuffle's 4133 // value type is same as the input vectors' type. 4134 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0)) 4135 if (isa<UndefValue>(Op1) && RetTy == InVecTy && 4136 OpShuf->getMask()->getSplatValue()) 4137 return Op0; 4138 4139 // Don't fold a shuffle with undef mask elements. This may get folded in a 4140 // better way using demanded bits or other analysis. 4141 // TODO: Should we allow this? 4142 if (find(Indices, -1) != Indices.end()) 4143 return nullptr; 4144 4145 // Check if every element of this shuffle can be mapped back to the 4146 // corresponding element of a single root vector. If so, we don't need this 4147 // shuffle. This handles simple identity shuffles as well as chains of 4148 // shuffles that may widen/narrow and/or move elements across lanes and back. 4149 Value *RootVec = nullptr; 4150 for (unsigned i = 0; i != MaskNumElts; ++i) { 4151 // Note that recursion is limited for each vector element, so if any element 4152 // exceeds the limit, this will fail to simplify. 4153 RootVec = 4154 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse); 4155 4156 // We can't replace a widening/narrowing shuffle with one of its operands. 4157 if (!RootVec || RootVec->getType() != RetTy) 4158 return nullptr; 4159 } 4160 return RootVec; 4161 } 4162 4163 /// Given operands for a ShuffleVectorInst, fold the result or return null. 4164 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask, 4165 Type *RetTy, const SimplifyQuery &Q) { 4166 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit); 4167 } 4168 4169 //=== Helper functions for higher up the class hierarchy. 4170 4171 /// Given operands for a BinaryOperator, see if we can fold the result. 4172 /// If not, this returns null. 4173 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 4174 const SimplifyQuery &Q, unsigned MaxRecurse) { 4175 switch (Opcode) { 4176 case Instruction::Add: 4177 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse); 4178 case Instruction::FAdd: 4179 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 4180 case Instruction::Sub: 4181 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse); 4182 case Instruction::FSub: 4183 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 4184 case Instruction::Mul: 4185 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse); 4186 case Instruction::FMul: 4187 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 4188 case Instruction::SDiv: 4189 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse); 4190 case Instruction::UDiv: 4191 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse); 4192 case Instruction::FDiv: 4193 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 4194 case Instruction::SRem: 4195 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse); 4196 case Instruction::URem: 4197 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse); 4198 case Instruction::FRem: 4199 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 4200 case Instruction::Shl: 4201 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse); 4202 case Instruction::LShr: 4203 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse); 4204 case Instruction::AShr: 4205 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse); 4206 case Instruction::And: 4207 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse); 4208 case Instruction::Or: 4209 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse); 4210 case Instruction::Xor: 4211 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse); 4212 default: 4213 llvm_unreachable("Unexpected opcode"); 4214 } 4215 } 4216 4217 /// Given operands for a BinaryOperator, see if we can fold the result. 4218 /// If not, this returns null. 4219 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the 4220 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp. 4221 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 4222 const FastMathFlags &FMF, const SimplifyQuery &Q, 4223 unsigned MaxRecurse) { 4224 switch (Opcode) { 4225 case Instruction::FAdd: 4226 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse); 4227 case Instruction::FSub: 4228 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse); 4229 case Instruction::FMul: 4230 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse); 4231 case Instruction::FDiv: 4232 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse); 4233 default: 4234 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse); 4235 } 4236 } 4237 4238 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 4239 const SimplifyQuery &Q) { 4240 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit); 4241 } 4242 4243 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 4244 FastMathFlags FMF, const SimplifyQuery &Q) { 4245 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit); 4246 } 4247 4248 /// Given operands for a CmpInst, see if we can fold the result. 4249 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 4250 const SimplifyQuery &Q, unsigned MaxRecurse) { 4251 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate)) 4252 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 4253 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse); 4254 } 4255 4256 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 4257 const SimplifyQuery &Q) { 4258 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit); 4259 } 4260 4261 static bool IsIdempotent(Intrinsic::ID ID) { 4262 switch (ID) { 4263 default: return false; 4264 4265 // Unary idempotent: f(f(x)) = f(x) 4266 case Intrinsic::fabs: 4267 case Intrinsic::floor: 4268 case Intrinsic::ceil: 4269 case Intrinsic::trunc: 4270 case Intrinsic::rint: 4271 case Intrinsic::nearbyint: 4272 case Intrinsic::round: 4273 return true; 4274 } 4275 } 4276 4277 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset, 4278 const DataLayout &DL) { 4279 GlobalValue *PtrSym; 4280 APInt PtrOffset; 4281 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL)) 4282 return nullptr; 4283 4284 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext()); 4285 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext()); 4286 Type *Int32PtrTy = Int32Ty->getPointerTo(); 4287 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext()); 4288 4289 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset); 4290 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64) 4291 return nullptr; 4292 4293 uint64_t OffsetInt = OffsetConstInt->getSExtValue(); 4294 if (OffsetInt % 4 != 0) 4295 return nullptr; 4296 4297 Constant *C = ConstantExpr::getGetElementPtr( 4298 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy), 4299 ConstantInt::get(Int64Ty, OffsetInt / 4)); 4300 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL); 4301 if (!Loaded) 4302 return nullptr; 4303 4304 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded); 4305 if (!LoadedCE) 4306 return nullptr; 4307 4308 if (LoadedCE->getOpcode() == Instruction::Trunc) { 4309 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0)); 4310 if (!LoadedCE) 4311 return nullptr; 4312 } 4313 4314 if (LoadedCE->getOpcode() != Instruction::Sub) 4315 return nullptr; 4316 4317 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0)); 4318 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt) 4319 return nullptr; 4320 auto *LoadedLHSPtr = LoadedLHS->getOperand(0); 4321 4322 Constant *LoadedRHS = LoadedCE->getOperand(1); 4323 GlobalValue *LoadedRHSSym; 4324 APInt LoadedRHSOffset; 4325 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset, 4326 DL) || 4327 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset) 4328 return nullptr; 4329 4330 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy); 4331 } 4332 4333 static bool maskIsAllZeroOrUndef(Value *Mask) { 4334 auto *ConstMask = dyn_cast<Constant>(Mask); 4335 if (!ConstMask) 4336 return false; 4337 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask)) 4338 return true; 4339 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E; 4340 ++I) { 4341 if (auto *MaskElt = ConstMask->getAggregateElement(I)) 4342 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt)) 4343 continue; 4344 return false; 4345 } 4346 return true; 4347 } 4348 4349 template <typename IterTy> 4350 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd, 4351 const SimplifyQuery &Q, unsigned MaxRecurse) { 4352 Intrinsic::ID IID = F->getIntrinsicID(); 4353 unsigned NumOperands = std::distance(ArgBegin, ArgEnd); 4354 4355 // Unary Ops 4356 if (NumOperands == 1) { 4357 // Perform idempotent optimizations 4358 if (IsIdempotent(IID)) { 4359 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) { 4360 if (II->getIntrinsicID() == IID) 4361 return II; 4362 } 4363 } 4364 4365 switch (IID) { 4366 case Intrinsic::fabs: { 4367 if (SignBitMustBeZero(*ArgBegin, Q.TLI)) 4368 return *ArgBegin; 4369 return nullptr; 4370 } 4371 default: 4372 return nullptr; 4373 } 4374 } 4375 4376 // Binary Ops 4377 if (NumOperands == 2) { 4378 Value *LHS = *ArgBegin; 4379 Value *RHS = *(ArgBegin + 1); 4380 Type *ReturnType = F->getReturnType(); 4381 4382 switch (IID) { 4383 case Intrinsic::usub_with_overflow: 4384 case Intrinsic::ssub_with_overflow: { 4385 // X - X -> { 0, false } 4386 if (LHS == RHS) 4387 return Constant::getNullValue(ReturnType); 4388 4389 // X - undef -> undef 4390 // undef - X -> undef 4391 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) 4392 return UndefValue::get(ReturnType); 4393 4394 return nullptr; 4395 } 4396 case Intrinsic::uadd_with_overflow: 4397 case Intrinsic::sadd_with_overflow: { 4398 // X + undef -> undef 4399 if (isa<UndefValue>(RHS)) 4400 return UndefValue::get(ReturnType); 4401 4402 return nullptr; 4403 } 4404 case Intrinsic::umul_with_overflow: 4405 case Intrinsic::smul_with_overflow: { 4406 // X * 0 -> { 0, false } 4407 if (match(RHS, m_Zero())) 4408 return Constant::getNullValue(ReturnType); 4409 4410 // X * undef -> { 0, false } 4411 if (match(RHS, m_Undef())) 4412 return Constant::getNullValue(ReturnType); 4413 4414 return nullptr; 4415 } 4416 case Intrinsic::load_relative: { 4417 Constant *C0 = dyn_cast<Constant>(LHS); 4418 Constant *C1 = dyn_cast<Constant>(RHS); 4419 if (C0 && C1) 4420 return SimplifyRelativeLoad(C0, C1, Q.DL); 4421 return nullptr; 4422 } 4423 default: 4424 return nullptr; 4425 } 4426 } 4427 4428 // Simplify calls to llvm.masked.load.* 4429 switch (IID) { 4430 case Intrinsic::masked_load: { 4431 Value *MaskArg = ArgBegin[2]; 4432 Value *PassthruArg = ArgBegin[3]; 4433 // If the mask is all zeros or undef, the "passthru" argument is the result. 4434 if (maskIsAllZeroOrUndef(MaskArg)) 4435 return PassthruArg; 4436 return nullptr; 4437 } 4438 default: 4439 return nullptr; 4440 } 4441 } 4442 4443 template <typename IterTy> 4444 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd, 4445 const SimplifyQuery &Q, unsigned MaxRecurse) { 4446 Type *Ty = V->getType(); 4447 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 4448 Ty = PTy->getElementType(); 4449 FunctionType *FTy = cast<FunctionType>(Ty); 4450 4451 // call undef -> undef 4452 // call null -> undef 4453 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V)) 4454 return UndefValue::get(FTy->getReturnType()); 4455 4456 Function *F = dyn_cast<Function>(V); 4457 if (!F) 4458 return nullptr; 4459 4460 if (F->isIntrinsic()) 4461 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse)) 4462 return Ret; 4463 4464 if (!canConstantFoldCallTo(F)) 4465 return nullptr; 4466 4467 SmallVector<Constant *, 4> ConstantArgs; 4468 ConstantArgs.reserve(ArgEnd - ArgBegin); 4469 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) { 4470 Constant *C = dyn_cast<Constant>(*I); 4471 if (!C) 4472 return nullptr; 4473 ConstantArgs.push_back(C); 4474 } 4475 4476 return ConstantFoldCall(F, ConstantArgs, Q.TLI); 4477 } 4478 4479 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin, 4480 User::op_iterator ArgEnd, const SimplifyQuery &Q) { 4481 return ::SimplifyCall(V, ArgBegin, ArgEnd, Q, RecursionLimit); 4482 } 4483 4484 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args, 4485 const SimplifyQuery &Q) { 4486 return ::SimplifyCall(V, Args.begin(), Args.end(), Q, RecursionLimit); 4487 } 4488 4489 /// See if we can compute a simplified version of this instruction. 4490 /// If not, this returns null. 4491 4492 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ, 4493 OptimizationRemarkEmitter *ORE) { 4494 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I); 4495 Value *Result; 4496 4497 switch (I->getOpcode()) { 4498 default: 4499 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI); 4500 break; 4501 case Instruction::FAdd: 4502 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1), 4503 I->getFastMathFlags(), Q); 4504 break; 4505 case Instruction::Add: 4506 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1), 4507 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4508 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q); 4509 break; 4510 case Instruction::FSub: 4511 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1), 4512 I->getFastMathFlags(), Q); 4513 break; 4514 case Instruction::Sub: 4515 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1), 4516 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4517 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q); 4518 break; 4519 case Instruction::FMul: 4520 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1), 4521 I->getFastMathFlags(), Q); 4522 break; 4523 case Instruction::Mul: 4524 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q); 4525 break; 4526 case Instruction::SDiv: 4527 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q); 4528 break; 4529 case Instruction::UDiv: 4530 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q); 4531 break; 4532 case Instruction::FDiv: 4533 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), 4534 I->getFastMathFlags(), Q); 4535 break; 4536 case Instruction::SRem: 4537 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q); 4538 break; 4539 case Instruction::URem: 4540 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q); 4541 break; 4542 case Instruction::FRem: 4543 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), 4544 I->getFastMathFlags(), Q); 4545 break; 4546 case Instruction::Shl: 4547 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1), 4548 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4549 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q); 4550 break; 4551 case Instruction::LShr: 4552 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1), 4553 cast<BinaryOperator>(I)->isExact(), Q); 4554 break; 4555 case Instruction::AShr: 4556 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1), 4557 cast<BinaryOperator>(I)->isExact(), Q); 4558 break; 4559 case Instruction::And: 4560 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q); 4561 break; 4562 case Instruction::Or: 4563 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q); 4564 break; 4565 case Instruction::Xor: 4566 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q); 4567 break; 4568 case Instruction::ICmp: 4569 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), 4570 I->getOperand(0), I->getOperand(1), Q); 4571 break; 4572 case Instruction::FCmp: 4573 Result = 4574 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0), 4575 I->getOperand(1), I->getFastMathFlags(), Q); 4576 break; 4577 case Instruction::Select: 4578 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1), 4579 I->getOperand(2), Q); 4580 break; 4581 case Instruction::GetElementPtr: { 4582 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end()); 4583 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(), 4584 Ops, Q); 4585 break; 4586 } 4587 case Instruction::InsertValue: { 4588 InsertValueInst *IV = cast<InsertValueInst>(I); 4589 Result = SimplifyInsertValueInst(IV->getAggregateOperand(), 4590 IV->getInsertedValueOperand(), 4591 IV->getIndices(), Q); 4592 break; 4593 } 4594 case Instruction::ExtractValue: { 4595 auto *EVI = cast<ExtractValueInst>(I); 4596 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(), 4597 EVI->getIndices(), Q); 4598 break; 4599 } 4600 case Instruction::ExtractElement: { 4601 auto *EEI = cast<ExtractElementInst>(I); 4602 Result = SimplifyExtractElementInst(EEI->getVectorOperand(), 4603 EEI->getIndexOperand(), Q); 4604 break; 4605 } 4606 case Instruction::ShuffleVector: { 4607 auto *SVI = cast<ShuffleVectorInst>(I); 4608 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1), 4609 SVI->getMask(), SVI->getType(), Q); 4610 break; 4611 } 4612 case Instruction::PHI: 4613 Result = SimplifyPHINode(cast<PHINode>(I), Q); 4614 break; 4615 case Instruction::Call: { 4616 CallSite CS(cast<CallInst>(I)); 4617 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), Q); 4618 break; 4619 } 4620 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc: 4621 #include "llvm/IR/Instruction.def" 4622 #undef HANDLE_CAST_INST 4623 Result = 4624 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q); 4625 break; 4626 case Instruction::Alloca: 4627 // No simplifications for Alloca and it can't be constant folded. 4628 Result = nullptr; 4629 break; 4630 } 4631 4632 // In general, it is possible for computeKnownBits to determine all bits in a 4633 // value even when the operands are not all constants. 4634 if (!Result && I->getType()->isIntOrIntVectorTy()) { 4635 unsigned BitWidth = I->getType()->getScalarSizeInBits(); 4636 KnownBits Known(BitWidth); 4637 computeKnownBits(I, Known, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE); 4638 if (Known.isConstant()) 4639 Result = ConstantInt::get(I->getType(), Known.getConstant()); 4640 } 4641 4642 /// If called on unreachable code, the above logic may report that the 4643 /// instruction simplified to itself. Make life easier for users by 4644 /// detecting that case here, returning a safe value instead. 4645 return Result == I ? UndefValue::get(I->getType()) : Result; 4646 } 4647 4648 /// \brief Implementation of recursive simplification through an instruction's 4649 /// uses. 4650 /// 4651 /// This is the common implementation of the recursive simplification routines. 4652 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to 4653 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of 4654 /// instructions to process and attempt to simplify it using 4655 /// InstructionSimplify. 4656 /// 4657 /// This routine returns 'true' only when *it* simplifies something. The passed 4658 /// in simplified value does not count toward this. 4659 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV, 4660 const TargetLibraryInfo *TLI, 4661 const DominatorTree *DT, 4662 AssumptionCache *AC) { 4663 bool Simplified = false; 4664 SmallSetVector<Instruction *, 8> Worklist; 4665 const DataLayout &DL = I->getModule()->getDataLayout(); 4666 4667 // If we have an explicit value to collapse to, do that round of the 4668 // simplification loop by hand initially. 4669 if (SimpleV) { 4670 for (User *U : I->users()) 4671 if (U != I) 4672 Worklist.insert(cast<Instruction>(U)); 4673 4674 // Replace the instruction with its simplified value. 4675 I->replaceAllUsesWith(SimpleV); 4676 4677 // Gracefully handle edge cases where the instruction is not wired into any 4678 // parent block. 4679 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) && 4680 !I->mayHaveSideEffects()) 4681 I->eraseFromParent(); 4682 } else { 4683 Worklist.insert(I); 4684 } 4685 4686 // Note that we must test the size on each iteration, the worklist can grow. 4687 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { 4688 I = Worklist[Idx]; 4689 4690 // See if this instruction simplifies. 4691 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC}); 4692 if (!SimpleV) 4693 continue; 4694 4695 Simplified = true; 4696 4697 // Stash away all the uses of the old instruction so we can check them for 4698 // recursive simplifications after a RAUW. This is cheaper than checking all 4699 // uses of To on the recursive step in most cases. 4700 for (User *U : I->users()) 4701 Worklist.insert(cast<Instruction>(U)); 4702 4703 // Replace the instruction with its simplified value. 4704 I->replaceAllUsesWith(SimpleV); 4705 4706 // Gracefully handle edge cases where the instruction is not wired into any 4707 // parent block. 4708 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) && 4709 !I->mayHaveSideEffects()) 4710 I->eraseFromParent(); 4711 } 4712 return Simplified; 4713 } 4714 4715 bool llvm::recursivelySimplifyInstruction(Instruction *I, 4716 const TargetLibraryInfo *TLI, 4717 const DominatorTree *DT, 4718 AssumptionCache *AC) { 4719 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC); 4720 } 4721 4722 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV, 4723 const TargetLibraryInfo *TLI, 4724 const DominatorTree *DT, 4725 AssumptionCache *AC) { 4726 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!"); 4727 assert(SimpleV && "Must provide a simplified value."); 4728 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC); 4729 } 4730 4731 namespace llvm { 4732 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) { 4733 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 4734 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr; 4735 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); 4736 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr; 4737 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>(); 4738 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr; 4739 return {F.getParent()->getDataLayout(), TLI, DT, AC}; 4740 } 4741 4742 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR, 4743 const DataLayout &DL) { 4744 return {DL, &AR.TLI, &AR.DT, &AR.AC}; 4745 } 4746 4747 template <class T, class... TArgs> 4748 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM, 4749 Function &F) { 4750 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F); 4751 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F); 4752 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F); 4753 return {F.getParent()->getDataLayout(), TLI, DT, AC}; 4754 } 4755 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &, 4756 Function &); 4757 } 4758