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/ConstantFolding.h" 25 #include "llvm/Analysis/MemoryBuiltins.h" 26 #include "llvm/Analysis/ValueTracking.h" 27 #include "llvm/IR/ConstantRange.h" 28 #include "llvm/IR/DataLayout.h" 29 #include "llvm/IR/Dominators.h" 30 #include "llvm/IR/GetElementPtrTypeIterator.h" 31 #include "llvm/IR/GlobalAlias.h" 32 #include "llvm/IR/Operator.h" 33 #include "llvm/IR/PatternMatch.h" 34 #include "llvm/IR/ValueHandle.h" 35 #include <algorithm> 36 using namespace llvm; 37 using namespace llvm::PatternMatch; 38 39 #define DEBUG_TYPE "instsimplify" 40 41 enum { RecursionLimit = 3 }; 42 43 STATISTIC(NumExpand, "Number of expansions"); 44 STATISTIC(NumReassoc, "Number of reassociations"); 45 46 namespace { 47 struct Query { 48 const DataLayout *DL; 49 const TargetLibraryInfo *TLI; 50 const DominatorTree *DT; 51 AssumptionCache *AC; 52 const Instruction *CxtI; 53 54 Query(const DataLayout *DL, const TargetLibraryInfo *tli, 55 const DominatorTree *dt, AssumptionCache *ac = nullptr, 56 const Instruction *cxti = nullptr) 57 : DL(DL), TLI(tli), DT(dt), AC(ac), CxtI(cxti) {} 58 }; 59 } // end anonymous namespace 60 61 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned); 62 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &, 63 unsigned); 64 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &, 65 const Query &, unsigned); 66 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &, 67 unsigned); 68 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned); 69 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned); 70 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned); 71 72 /// getFalse - For a boolean type, or a vector of boolean type, return false, or 73 /// a vector with every element false, as appropriate for the type. 74 static Constant *getFalse(Type *Ty) { 75 assert(Ty->getScalarType()->isIntegerTy(1) && 76 "Expected i1 type or a vector of i1!"); 77 return Constant::getNullValue(Ty); 78 } 79 80 /// getTrue - For a boolean type, or a vector of boolean type, return true, or 81 /// a vector with every element true, as appropriate for the type. 82 static Constant *getTrue(Type *Ty) { 83 assert(Ty->getScalarType()->isIntegerTy(1) && 84 "Expected i1 type or a vector of i1!"); 85 return Constant::getAllOnesValue(Ty); 86 } 87 88 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"? 89 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS, 90 Value *RHS) { 91 CmpInst *Cmp = dyn_cast<CmpInst>(V); 92 if (!Cmp) 93 return false; 94 CmpInst::Predicate CPred = Cmp->getPredicate(); 95 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1); 96 if (CPred == Pred && CLHS == LHS && CRHS == RHS) 97 return true; 98 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS && 99 CRHS == LHS; 100 } 101 102 /// ValueDominatesPHI - Does the given value dominate the specified phi node? 103 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) { 104 Instruction *I = dyn_cast<Instruction>(V); 105 if (!I) 106 // Arguments and constants dominate all instructions. 107 return true; 108 109 // If we are processing instructions (and/or basic blocks) that have not been 110 // fully added to a function, the parent nodes may still be null. Simply 111 // return the conservative answer in these cases. 112 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent()) 113 return false; 114 115 // If we have a DominatorTree then do a precise test. 116 if (DT) { 117 if (!DT->isReachableFromEntry(P->getParent())) 118 return true; 119 if (!DT->isReachableFromEntry(I->getParent())) 120 return false; 121 return DT->dominates(I, P); 122 } 123 124 // Otherwise, if the instruction is in the entry block, and is not an invoke, 125 // then it obviously dominates all phi nodes. 126 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() && 127 !isa<InvokeInst>(I)) 128 return true; 129 130 return false; 131 } 132 133 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning 134 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is 135 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS. 136 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)". 137 /// Returns the simplified value, or null if no simplification was performed. 138 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS, 139 unsigned OpcToExpand, const Query &Q, 140 unsigned MaxRecurse) { 141 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand; 142 // Recursion is always used, so bail out at once if we already hit the limit. 143 if (!MaxRecurse--) 144 return nullptr; 145 146 // Check whether the expression has the form "(A op' B) op C". 147 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS)) 148 if (Op0->getOpcode() == OpcodeToExpand) { 149 // It does! Try turning it into "(A op C) op' (B op C)". 150 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 151 // Do "A op C" and "B op C" both simplify? 152 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) 153 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 154 // They do! Return "L op' R" if it simplifies or is already available. 155 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 156 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand) 157 && L == B && R == A)) { 158 ++NumExpand; 159 return LHS; 160 } 161 // Otherwise return "L op' R" if it simplifies. 162 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 163 ++NumExpand; 164 return V; 165 } 166 } 167 } 168 169 // Check whether the expression has the form "A op (B op' C)". 170 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS)) 171 if (Op1->getOpcode() == OpcodeToExpand) { 172 // It does! Try turning it into "(A op B) op' (A op C)". 173 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 174 // Do "A op B" and "A op C" both simplify? 175 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) 176 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) { 177 // They do! Return "L op' R" if it simplifies or is already available. 178 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 179 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand) 180 && L == C && R == B)) { 181 ++NumExpand; 182 return RHS; 183 } 184 // Otherwise return "L op' R" if it simplifies. 185 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 186 ++NumExpand; 187 return V; 188 } 189 } 190 } 191 192 return nullptr; 193 } 194 195 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary 196 /// operations. Returns the simpler value, or null if none was found. 197 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS, 198 const Query &Q, unsigned MaxRecurse) { 199 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc; 200 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!"); 201 202 // Recursion is always used, so bail out at once if we already hit the limit. 203 if (!MaxRecurse--) 204 return nullptr; 205 206 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 207 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 208 209 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely. 210 if (Op0 && Op0->getOpcode() == Opcode) { 211 Value *A = Op0->getOperand(0); 212 Value *B = Op0->getOperand(1); 213 Value *C = RHS; 214 215 // Does "B op C" simplify? 216 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 217 // It does! Return "A op V" if it simplifies or is already available. 218 // If V equals B then "A op V" is just the LHS. 219 if (V == B) return LHS; 220 // Otherwise return "A op V" if it simplifies. 221 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) { 222 ++NumReassoc; 223 return W; 224 } 225 } 226 } 227 228 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely. 229 if (Op1 && Op1->getOpcode() == Opcode) { 230 Value *A = LHS; 231 Value *B = Op1->getOperand(0); 232 Value *C = Op1->getOperand(1); 233 234 // Does "A op B" simplify? 235 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) { 236 // It does! Return "V op C" if it simplifies or is already available. 237 // If V equals B then "V op C" is just the RHS. 238 if (V == B) return RHS; 239 // Otherwise return "V op C" if it simplifies. 240 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) { 241 ++NumReassoc; 242 return W; 243 } 244 } 245 } 246 247 // The remaining transforms require commutativity as well as associativity. 248 if (!Instruction::isCommutative(Opcode)) 249 return nullptr; 250 251 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely. 252 if (Op0 && Op0->getOpcode() == Opcode) { 253 Value *A = Op0->getOperand(0); 254 Value *B = Op0->getOperand(1); 255 Value *C = RHS; 256 257 // Does "C op A" simplify? 258 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 259 // It does! Return "V op B" if it simplifies or is already available. 260 // If V equals A then "V op B" is just the LHS. 261 if (V == A) return LHS; 262 // Otherwise return "V op B" if it simplifies. 263 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) { 264 ++NumReassoc; 265 return W; 266 } 267 } 268 } 269 270 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely. 271 if (Op1 && Op1->getOpcode() == Opcode) { 272 Value *A = LHS; 273 Value *B = Op1->getOperand(0); 274 Value *C = Op1->getOperand(1); 275 276 // Does "C op A" simplify? 277 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 278 // It does! Return "B op V" if it simplifies or is already available. 279 // If V equals C then "B op V" is just the RHS. 280 if (V == C) return RHS; 281 // Otherwise return "B op V" if it simplifies. 282 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) { 283 ++NumReassoc; 284 return W; 285 } 286 } 287 } 288 289 return nullptr; 290 } 291 292 /// ThreadBinOpOverSelect - In the case of a binary operation with a select 293 /// instruction as an operand, try to simplify the binop by seeing whether 294 /// evaluating it on both branches of the select results in the same value. 295 /// Returns the common value if so, otherwise returns null. 296 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS, 297 const Query &Q, unsigned MaxRecurse) { 298 // Recursion is always used, so bail out at once if we already hit the limit. 299 if (!MaxRecurse--) 300 return nullptr; 301 302 SelectInst *SI; 303 if (isa<SelectInst>(LHS)) { 304 SI = cast<SelectInst>(LHS); 305 } else { 306 assert(isa<SelectInst>(RHS) && "No select instruction operand!"); 307 SI = cast<SelectInst>(RHS); 308 } 309 310 // Evaluate the BinOp on the true and false branches of the select. 311 Value *TV; 312 Value *FV; 313 if (SI == LHS) { 314 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse); 315 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse); 316 } else { 317 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse); 318 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse); 319 } 320 321 // If they simplified to the same value, then return the common value. 322 // If they both failed to simplify then return null. 323 if (TV == FV) 324 return TV; 325 326 // If one branch simplified to undef, return the other one. 327 if (TV && isa<UndefValue>(TV)) 328 return FV; 329 if (FV && isa<UndefValue>(FV)) 330 return TV; 331 332 // If applying the operation did not change the true and false select values, 333 // then the result of the binop is the select itself. 334 if (TV == SI->getTrueValue() && FV == SI->getFalseValue()) 335 return SI; 336 337 // If one branch simplified and the other did not, and the simplified 338 // value is equal to the unsimplified one, return the simplified value. 339 // For example, select (cond, X, X & Z) & Z -> X & Z. 340 if ((FV && !TV) || (TV && !FV)) { 341 // Check that the simplified value has the form "X op Y" where "op" is the 342 // same as the original operation. 343 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV); 344 if (Simplified && Simplified->getOpcode() == Opcode) { 345 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS". 346 // We already know that "op" is the same as for the simplified value. See 347 // if the operands match too. If so, return the simplified value. 348 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue(); 349 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS; 350 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch; 351 if (Simplified->getOperand(0) == UnsimplifiedLHS && 352 Simplified->getOperand(1) == UnsimplifiedRHS) 353 return Simplified; 354 if (Simplified->isCommutative() && 355 Simplified->getOperand(1) == UnsimplifiedLHS && 356 Simplified->getOperand(0) == UnsimplifiedRHS) 357 return Simplified; 358 } 359 } 360 361 return nullptr; 362 } 363 364 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction, 365 /// try to simplify the comparison by seeing whether both branches of the select 366 /// result in the same value. Returns the common value if so, otherwise returns 367 /// null. 368 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS, 369 Value *RHS, const Query &Q, 370 unsigned MaxRecurse) { 371 // Recursion is always used, so bail out at once if we already hit the limit. 372 if (!MaxRecurse--) 373 return nullptr; 374 375 // Make sure the select is on the LHS. 376 if (!isa<SelectInst>(LHS)) { 377 std::swap(LHS, RHS); 378 Pred = CmpInst::getSwappedPredicate(Pred); 379 } 380 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!"); 381 SelectInst *SI = cast<SelectInst>(LHS); 382 Value *Cond = SI->getCondition(); 383 Value *TV = SI->getTrueValue(); 384 Value *FV = SI->getFalseValue(); 385 386 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it. 387 // Does "cmp TV, RHS" simplify? 388 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse); 389 if (TCmp == Cond) { 390 // It not only simplified, it simplified to the select condition. Replace 391 // it with 'true'. 392 TCmp = getTrue(Cond->getType()); 393 } else if (!TCmp) { 394 // It didn't simplify. However if "cmp TV, RHS" is equal to the select 395 // condition then we can replace it with 'true'. Otherwise give up. 396 if (!isSameCompare(Cond, Pred, TV, RHS)) 397 return nullptr; 398 TCmp = getTrue(Cond->getType()); 399 } 400 401 // Does "cmp FV, RHS" simplify? 402 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse); 403 if (FCmp == Cond) { 404 // It not only simplified, it simplified to the select condition. Replace 405 // it with 'false'. 406 FCmp = getFalse(Cond->getType()); 407 } else if (!FCmp) { 408 // It didn't simplify. However if "cmp FV, RHS" is equal to the select 409 // condition then we can replace it with 'false'. Otherwise give up. 410 if (!isSameCompare(Cond, Pred, FV, RHS)) 411 return nullptr; 412 FCmp = getFalse(Cond->getType()); 413 } 414 415 // If both sides simplified to the same value, then use it as the result of 416 // the original comparison. 417 if (TCmp == FCmp) 418 return TCmp; 419 420 // The remaining cases only make sense if the select condition has the same 421 // type as the result of the comparison, so bail out if this is not so. 422 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy()) 423 return nullptr; 424 // If the false value simplified to false, then the result of the compare 425 // is equal to "Cond && TCmp". This also catches the case when the false 426 // value simplified to false and the true value to true, returning "Cond". 427 if (match(FCmp, m_Zero())) 428 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse)) 429 return V; 430 // If the true value simplified to true, then the result of the compare 431 // is equal to "Cond || FCmp". 432 if (match(TCmp, m_One())) 433 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse)) 434 return V; 435 // Finally, if the false value simplified to true and the true value to 436 // false, then the result of the compare is equal to "!Cond". 437 if (match(FCmp, m_One()) && match(TCmp, m_Zero())) 438 if (Value *V = 439 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()), 440 Q, MaxRecurse)) 441 return V; 442 443 return nullptr; 444 } 445 446 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that 447 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating 448 /// it on the incoming phi values yields the same result for every value. If so 449 /// returns the common value, otherwise returns null. 450 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS, 451 const Query &Q, unsigned MaxRecurse) { 452 // Recursion is always used, so bail out at once if we already hit the limit. 453 if (!MaxRecurse--) 454 return nullptr; 455 456 PHINode *PI; 457 if (isa<PHINode>(LHS)) { 458 PI = cast<PHINode>(LHS); 459 // Bail out if RHS and the phi may be mutually interdependent due to a loop. 460 if (!ValueDominatesPHI(RHS, PI, Q.DT)) 461 return nullptr; 462 } else { 463 assert(isa<PHINode>(RHS) && "No PHI instruction operand!"); 464 PI = cast<PHINode>(RHS); 465 // Bail out if LHS and the phi may be mutually interdependent due to a loop. 466 if (!ValueDominatesPHI(LHS, PI, Q.DT)) 467 return nullptr; 468 } 469 470 // Evaluate the BinOp on the incoming phi values. 471 Value *CommonValue = nullptr; 472 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) { 473 Value *Incoming = PI->getIncomingValue(i); 474 // If the incoming value is the phi node itself, it can safely be skipped. 475 if (Incoming == PI) continue; 476 Value *V = PI == LHS ? 477 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) : 478 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse); 479 // If the operation failed to simplify, or simplified to a different value 480 // to previously, then give up. 481 if (!V || (CommonValue && V != CommonValue)) 482 return nullptr; 483 CommonValue = V; 484 } 485 486 return CommonValue; 487 } 488 489 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try 490 /// try to simplify the comparison by seeing whether comparing with all of the 491 /// incoming phi values yields the same result every time. If so returns the 492 /// common result, otherwise returns null. 493 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS, 494 const Query &Q, unsigned MaxRecurse) { 495 // Recursion is always used, so bail out at once if we already hit the limit. 496 if (!MaxRecurse--) 497 return nullptr; 498 499 // Make sure the phi is on the LHS. 500 if (!isa<PHINode>(LHS)) { 501 std::swap(LHS, RHS); 502 Pred = CmpInst::getSwappedPredicate(Pred); 503 } 504 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!"); 505 PHINode *PI = cast<PHINode>(LHS); 506 507 // Bail out if RHS and the phi may be mutually interdependent due to a loop. 508 if (!ValueDominatesPHI(RHS, PI, Q.DT)) 509 return nullptr; 510 511 // Evaluate the BinOp on the incoming phi values. 512 Value *CommonValue = nullptr; 513 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) { 514 Value *Incoming = PI->getIncomingValue(i); 515 // If the incoming value is the phi node itself, it can safely be skipped. 516 if (Incoming == PI) continue; 517 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse); 518 // If the operation failed to simplify, or simplified to a different value 519 // to previously, then give up. 520 if (!V || (CommonValue && V != CommonValue)) 521 return nullptr; 522 CommonValue = V; 523 } 524 525 return CommonValue; 526 } 527 528 /// SimplifyAddInst - Given operands for an Add, see if we can 529 /// fold the result. If not, this returns null. 530 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 531 const Query &Q, unsigned MaxRecurse) { 532 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 533 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 534 Constant *Ops[] = { CLHS, CRHS }; 535 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops, 536 Q.DL, Q.TLI); 537 } 538 539 // Canonicalize the constant to the RHS. 540 std::swap(Op0, Op1); 541 } 542 543 // X + undef -> undef 544 if (match(Op1, m_Undef())) 545 return Op1; 546 547 // X + 0 -> X 548 if (match(Op1, m_Zero())) 549 return Op0; 550 551 // X + (Y - X) -> Y 552 // (Y - X) + X -> Y 553 // Eg: X + -X -> 0 554 Value *Y = nullptr; 555 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) || 556 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1)))) 557 return Y; 558 559 // X + ~X -> -1 since ~X = -X-1 560 if (match(Op0, m_Not(m_Specific(Op1))) || 561 match(Op1, m_Not(m_Specific(Op0)))) 562 return Constant::getAllOnesValue(Op0->getType()); 563 564 /// i1 add -> xor. 565 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 566 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 567 return V; 568 569 // Try some generic simplifications for associative operations. 570 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q, 571 MaxRecurse)) 572 return V; 573 574 // Threading Add over selects and phi nodes is pointless, so don't bother. 575 // Threading over the select in "A + select(cond, B, C)" means evaluating 576 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and 577 // only if B and C are equal. If B and C are equal then (since we assume 578 // that operands have already been simplified) "select(cond, B, C)" should 579 // have been simplified to the common value of B and C already. Analysing 580 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly 581 // for threading over phi nodes. 582 583 return nullptr; 584 } 585 586 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 587 const DataLayout *DL, const TargetLibraryInfo *TLI, 588 const DominatorTree *DT, AssumptionCache *AC, 589 const Instruction *CxtI) { 590 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 591 RecursionLimit); 592 } 593 594 /// \brief Compute the base pointer and cumulative constant offsets for V. 595 /// 596 /// This strips all constant offsets off of V, leaving it the base pointer, and 597 /// accumulates the total constant offset applied in the returned constant. It 598 /// returns 0 if V is not a pointer, and returns the constant '0' if there are 599 /// no constant offsets applied. 600 /// 601 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't 602 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc. 603 /// folding. 604 static Constant *stripAndComputeConstantOffsets(const DataLayout *DL, 605 Value *&V, 606 bool AllowNonInbounds = false) { 607 assert(V->getType()->getScalarType()->isPointerTy()); 608 609 // Without DataLayout, just be conservative for now. Theoretically, more could 610 // be done in this case. 611 if (!DL) 612 return ConstantInt::get(IntegerType::get(V->getContext(), 64), 0); 613 614 Type *IntPtrTy = DL->getIntPtrType(V->getType())->getScalarType(); 615 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth()); 616 617 // Even though we don't look through PHI nodes, we could be called on an 618 // instruction in an unreachable block, which may be on a cycle. 619 SmallPtrSet<Value *, 4> Visited; 620 Visited.insert(V); 621 do { 622 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 623 if ((!AllowNonInbounds && !GEP->isInBounds()) || 624 !GEP->accumulateConstantOffset(*DL, Offset)) 625 break; 626 V = GEP->getPointerOperand(); 627 } else if (Operator::getOpcode(V) == Instruction::BitCast) { 628 V = cast<Operator>(V)->getOperand(0); 629 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 630 if (GA->mayBeOverridden()) 631 break; 632 V = GA->getAliasee(); 633 } else { 634 break; 635 } 636 assert(V->getType()->getScalarType()->isPointerTy() && 637 "Unexpected operand type!"); 638 } while (Visited.insert(V).second); 639 640 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset); 641 if (V->getType()->isVectorTy()) 642 return ConstantVector::getSplat(V->getType()->getVectorNumElements(), 643 OffsetIntPtr); 644 return OffsetIntPtr; 645 } 646 647 /// \brief Compute the constant difference between two pointer values. 648 /// If the difference is not a constant, returns zero. 649 static Constant *computePointerDifference(const DataLayout *DL, 650 Value *LHS, Value *RHS) { 651 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 652 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 653 654 // If LHS and RHS are not related via constant offsets to the same base 655 // value, there is nothing we can do here. 656 if (LHS != RHS) 657 return nullptr; 658 659 // Otherwise, the difference of LHS - RHS can be computed as: 660 // LHS - RHS 661 // = (LHSOffset + Base) - (RHSOffset + Base) 662 // = LHSOffset - RHSOffset 663 return ConstantExpr::getSub(LHSOffset, RHSOffset); 664 } 665 666 /// SimplifySubInst - Given operands for a Sub, see if we can 667 /// fold the result. If not, this returns null. 668 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 669 const Query &Q, unsigned MaxRecurse) { 670 if (Constant *CLHS = dyn_cast<Constant>(Op0)) 671 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 672 Constant *Ops[] = { CLHS, CRHS }; 673 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(), 674 Ops, Q.DL, Q.TLI); 675 } 676 677 // X - undef -> undef 678 // undef - X -> undef 679 if (match(Op0, m_Undef()) || match(Op1, m_Undef())) 680 return UndefValue::get(Op0->getType()); 681 682 // X - 0 -> X 683 if (match(Op1, m_Zero())) 684 return Op0; 685 686 // X - X -> 0 687 if (Op0 == Op1) 688 return Constant::getNullValue(Op0->getType()); 689 690 // 0 - X -> 0 if the sub is NUW. 691 if (isNUW && match(Op0, m_Zero())) 692 return Op0; 693 694 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies. 695 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X 696 Value *X = nullptr, *Y = nullptr, *Z = Op1; 697 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z 698 // See if "V === Y - Z" simplifies. 699 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1)) 700 // It does! Now see if "X + V" simplifies. 701 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) { 702 // It does, we successfully reassociated! 703 ++NumReassoc; 704 return W; 705 } 706 // See if "V === X - Z" simplifies. 707 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 708 // It does! Now see if "Y + V" simplifies. 709 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) { 710 // It does, we successfully reassociated! 711 ++NumReassoc; 712 return W; 713 } 714 } 715 716 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies. 717 // For example, X - (X + 1) -> -1 718 X = Op0; 719 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z) 720 // See if "V === X - Y" simplifies. 721 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 722 // It does! Now see if "V - Z" simplifies. 723 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) { 724 // It does, we successfully reassociated! 725 ++NumReassoc; 726 return W; 727 } 728 // See if "V === X - Z" simplifies. 729 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 730 // It does! Now see if "V - Y" simplifies. 731 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) { 732 // It does, we successfully reassociated! 733 ++NumReassoc; 734 return W; 735 } 736 } 737 738 // Z - (X - Y) -> (Z - X) + Y if everything simplifies. 739 // For example, X - (X - Y) -> Y. 740 Z = Op0; 741 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y) 742 // See if "V === Z - X" simplifies. 743 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1)) 744 // It does! Now see if "V + Y" simplifies. 745 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) { 746 // It does, we successfully reassociated! 747 ++NumReassoc; 748 return W; 749 } 750 751 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies. 752 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) && 753 match(Op1, m_Trunc(m_Value(Y)))) 754 if (X->getType() == Y->getType()) 755 // See if "V === X - Y" simplifies. 756 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 757 // It does! Now see if "trunc V" simplifies. 758 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1)) 759 // It does, return the simplified "trunc V". 760 return W; 761 762 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...). 763 if (match(Op0, m_PtrToInt(m_Value(X))) && 764 match(Op1, m_PtrToInt(m_Value(Y)))) 765 if (Constant *Result = computePointerDifference(Q.DL, X, Y)) 766 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true); 767 768 // i1 sub -> xor. 769 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 770 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 771 return V; 772 773 // Threading Sub over selects and phi nodes is pointless, so don't bother. 774 // Threading over the select in "A - select(cond, B, C)" means evaluating 775 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and 776 // only if B and C are equal. If B and C are equal then (since we assume 777 // that operands have already been simplified) "select(cond, B, C)" should 778 // have been simplified to the common value of B and C already. Analysing 779 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly 780 // for threading over phi nodes. 781 782 return nullptr; 783 } 784 785 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 786 const DataLayout *DL, const TargetLibraryInfo *TLI, 787 const DominatorTree *DT, AssumptionCache *AC, 788 const Instruction *CxtI) { 789 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 790 RecursionLimit); 791 } 792 793 /// Given operands for an FAdd, see if we can fold the result. If not, this 794 /// returns null. 795 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 796 const Query &Q, unsigned MaxRecurse) { 797 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 798 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 799 Constant *Ops[] = { CLHS, CRHS }; 800 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(), 801 Ops, Q.DL, Q.TLI); 802 } 803 804 // Canonicalize the constant to the RHS. 805 std::swap(Op0, Op1); 806 } 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))) 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 Query &Q, unsigned MaxRecurse) { 839 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 840 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 841 Constant *Ops[] = { CLHS, CRHS }; 842 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(), 843 Ops, Q.DL, Q.TLI); 844 } 845 } 846 847 // fsub X, 0 ==> X 848 if (match(Op1, m_Zero())) 849 return Op0; 850 851 // fsub X, -0 ==> X, when we know X is not -0 852 if (match(Op1, m_NegZero()) && 853 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0))) 854 return Op0; 855 856 // fsub 0, (fsub -0.0, X) ==> X 857 Value *X; 858 if (match(Op0, m_AnyZero())) { 859 if (match(Op1, m_FSub(m_NegZero(), m_Value(X)))) 860 return X; 861 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X)))) 862 return X; 863 } 864 865 // fsub nnan ninf x, x ==> 0.0 866 if (FMF.noNaNs() && FMF.noInfs() && Op0 == Op1) 867 return Constant::getNullValue(Op0->getType()); 868 869 return nullptr; 870 } 871 872 /// Given the operands for an FMul, see if we can fold the result 873 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, 874 FastMathFlags FMF, 875 const Query &Q, 876 unsigned MaxRecurse) { 877 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 878 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 879 Constant *Ops[] = { CLHS, CRHS }; 880 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(), 881 Ops, Q.DL, Q.TLI); 882 } 883 884 // Canonicalize the constant to the RHS. 885 std::swap(Op0, Op1); 886 } 887 888 // fmul X, 1.0 ==> X 889 if (match(Op1, m_FPOne())) 890 return Op0; 891 892 // fmul nnan nsz X, 0 ==> 0 893 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero())) 894 return Op1; 895 896 return nullptr; 897 } 898 899 /// SimplifyMulInst - Given operands for a Mul, see if we can 900 /// fold the result. If not, this returns null. 901 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q, 902 unsigned MaxRecurse) { 903 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 904 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 905 Constant *Ops[] = { CLHS, CRHS }; 906 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(), 907 Ops, Q.DL, Q.TLI); 908 } 909 910 // Canonicalize the constant to the RHS. 911 std::swap(Op0, Op1); 912 } 913 914 // X * undef -> 0 915 if (match(Op1, m_Undef())) 916 return Constant::getNullValue(Op0->getType()); 917 918 // X * 0 -> 0 919 if (match(Op1, m_Zero())) 920 return Op1; 921 922 // X * 1 -> X 923 if (match(Op1, m_One())) 924 return Op0; 925 926 // (X / Y) * Y -> X if the division is exact. 927 Value *X = nullptr; 928 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y 929 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y) 930 return X; 931 932 // i1 mul -> and. 933 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 934 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1)) 935 return V; 936 937 // Try some generic simplifications for associative operations. 938 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q, 939 MaxRecurse)) 940 return V; 941 942 // Mul distributes over Add. Try some generic simplifications based on this. 943 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add, 944 Q, MaxRecurse)) 945 return V; 946 947 // If the operation is with the result of a select instruction, check whether 948 // operating on either branch of the select always yields the same value. 949 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 950 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q, 951 MaxRecurse)) 952 return V; 953 954 // If the operation is with the result of a phi instruction, check whether 955 // operating on all incoming values of the phi always yields the same value. 956 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 957 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q, 958 MaxRecurse)) 959 return V; 960 961 return nullptr; 962 } 963 964 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 965 const DataLayout *DL, 966 const TargetLibraryInfo *TLI, 967 const DominatorTree *DT, AssumptionCache *AC, 968 const Instruction *CxtI) { 969 return ::SimplifyFAddInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 970 RecursionLimit); 971 } 972 973 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 974 const DataLayout *DL, 975 const TargetLibraryInfo *TLI, 976 const DominatorTree *DT, AssumptionCache *AC, 977 const Instruction *CxtI) { 978 return ::SimplifyFSubInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 979 RecursionLimit); 980 } 981 982 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF, 983 const DataLayout *DL, 984 const TargetLibraryInfo *TLI, 985 const DominatorTree *DT, AssumptionCache *AC, 986 const Instruction *CxtI) { 987 return ::SimplifyFMulInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 988 RecursionLimit); 989 } 990 991 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout *DL, 992 const TargetLibraryInfo *TLI, 993 const DominatorTree *DT, AssumptionCache *AC, 994 const Instruction *CxtI) { 995 return ::SimplifyMulInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 996 RecursionLimit); 997 } 998 999 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can 1000 /// fold the result. If not, this returns null. 1001 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1002 const Query &Q, unsigned MaxRecurse) { 1003 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1004 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1005 Constant *Ops[] = { C0, C1 }; 1006 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1007 } 1008 } 1009 1010 bool isSigned = Opcode == Instruction::SDiv; 1011 1012 // X / undef -> undef 1013 if (match(Op1, m_Undef())) 1014 return Op1; 1015 1016 // X / 0 -> undef, we don't need to preserve faults! 1017 if (match(Op1, m_Zero())) 1018 return UndefValue::get(Op1->getType()); 1019 1020 // undef / X -> 0 1021 if (match(Op0, m_Undef())) 1022 return Constant::getNullValue(Op0->getType()); 1023 1024 // 0 / X -> 0, we don't need to preserve faults! 1025 if (match(Op0, m_Zero())) 1026 return Op0; 1027 1028 // X / 1 -> X 1029 if (match(Op1, m_One())) 1030 return Op0; 1031 1032 if (Op0->getType()->isIntegerTy(1)) 1033 // It can't be division by zero, hence it must be division by one. 1034 return Op0; 1035 1036 // X / X -> 1 1037 if (Op0 == Op1) 1038 return ConstantInt::get(Op0->getType(), 1); 1039 1040 // (X * Y) / Y -> X if the multiplication does not overflow. 1041 Value *X = nullptr, *Y = nullptr; 1042 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) { 1043 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1 1044 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0); 1045 // If the Mul knows it does not overflow, then we are good to go. 1046 if ((isSigned && Mul->hasNoSignedWrap()) || 1047 (!isSigned && Mul->hasNoUnsignedWrap())) 1048 return X; 1049 // If X has the form X = A / Y then X * Y cannot overflow. 1050 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X)) 1051 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y) 1052 return X; 1053 } 1054 1055 // (X rem Y) / Y -> 0 1056 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1057 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1058 return Constant::getNullValue(Op0->getType()); 1059 1060 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow 1061 ConstantInt *C1, *C2; 1062 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) && 1063 match(Op1, m_ConstantInt(C2))) { 1064 bool Overflow; 1065 C1->getValue().umul_ov(C2->getValue(), Overflow); 1066 if (Overflow) 1067 return Constant::getNullValue(Op0->getType()); 1068 } 1069 1070 // If the operation is with the result of a select instruction, check whether 1071 // operating on either branch of the select always yields the same value. 1072 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1073 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1074 return V; 1075 1076 // If the operation is with the result of a phi instruction, check whether 1077 // operating on all incoming values of the phi always yields the same value. 1078 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1079 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1080 return V; 1081 1082 return nullptr; 1083 } 1084 1085 /// SimplifySDivInst - Given operands for an SDiv, see if we can 1086 /// fold the result. If not, this returns null. 1087 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q, 1088 unsigned MaxRecurse) { 1089 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse)) 1090 return V; 1091 1092 return nullptr; 1093 } 1094 1095 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout *DL, 1096 const TargetLibraryInfo *TLI, 1097 const DominatorTree *DT, AssumptionCache *AC, 1098 const Instruction *CxtI) { 1099 return ::SimplifySDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1100 RecursionLimit); 1101 } 1102 1103 /// SimplifyUDivInst - Given operands for a UDiv, see if we can 1104 /// fold the result. If not, this returns null. 1105 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q, 1106 unsigned MaxRecurse) { 1107 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse)) 1108 return V; 1109 1110 return nullptr; 1111 } 1112 1113 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout *DL, 1114 const TargetLibraryInfo *TLI, 1115 const DominatorTree *DT, AssumptionCache *AC, 1116 const Instruction *CxtI) { 1117 return ::SimplifyUDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1118 RecursionLimit); 1119 } 1120 1121 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, const Query &Q, 1122 unsigned) { 1123 // undef / X -> undef (the undef could be a snan). 1124 if (match(Op0, m_Undef())) 1125 return Op0; 1126 1127 // X / undef -> undef 1128 if (match(Op1, m_Undef())) 1129 return Op1; 1130 1131 return nullptr; 1132 } 1133 1134 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, const DataLayout *DL, 1135 const TargetLibraryInfo *TLI, 1136 const DominatorTree *DT, AssumptionCache *AC, 1137 const Instruction *CxtI) { 1138 return ::SimplifyFDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1139 RecursionLimit); 1140 } 1141 1142 /// SimplifyRem - Given operands for an SRem or URem, see if we can 1143 /// fold the result. If not, this returns null. 1144 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1145 const Query &Q, unsigned MaxRecurse) { 1146 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1147 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1148 Constant *Ops[] = { C0, C1 }; 1149 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1150 } 1151 } 1152 1153 // X % undef -> undef 1154 if (match(Op1, m_Undef())) 1155 return Op1; 1156 1157 // undef % X -> 0 1158 if (match(Op0, m_Undef())) 1159 return Constant::getNullValue(Op0->getType()); 1160 1161 // 0 % X -> 0, we don't need to preserve faults! 1162 if (match(Op0, m_Zero())) 1163 return Op0; 1164 1165 // X % 0 -> undef, we don't need to preserve faults! 1166 if (match(Op1, m_Zero())) 1167 return UndefValue::get(Op0->getType()); 1168 1169 // X % 1 -> 0 1170 if (match(Op1, m_One())) 1171 return Constant::getNullValue(Op0->getType()); 1172 1173 if (Op0->getType()->isIntegerTy(1)) 1174 // It can't be remainder by zero, hence it must be remainder by one. 1175 return Constant::getNullValue(Op0->getType()); 1176 1177 // X % X -> 0 1178 if (Op0 == Op1) 1179 return Constant::getNullValue(Op0->getType()); 1180 1181 // (X % Y) % Y -> X % Y 1182 if ((Opcode == Instruction::SRem && 1183 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1184 (Opcode == Instruction::URem && 1185 match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1186 return Op0; 1187 1188 // If the operation is with the result of a select instruction, check whether 1189 // operating on either branch of the select always yields the same value. 1190 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1191 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1192 return V; 1193 1194 // If the operation is with the result of a phi instruction, check whether 1195 // operating on all incoming values of the phi always yields the same value. 1196 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1197 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1198 return V; 1199 1200 return nullptr; 1201 } 1202 1203 /// SimplifySRemInst - Given operands for an SRem, see if we can 1204 /// fold the result. If not, this returns null. 1205 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q, 1206 unsigned MaxRecurse) { 1207 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse)) 1208 return V; 1209 1210 return nullptr; 1211 } 1212 1213 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout *DL, 1214 const TargetLibraryInfo *TLI, 1215 const DominatorTree *DT, AssumptionCache *AC, 1216 const Instruction *CxtI) { 1217 return ::SimplifySRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1218 RecursionLimit); 1219 } 1220 1221 /// SimplifyURemInst - Given operands for a URem, see if we can 1222 /// fold the result. If not, this returns null. 1223 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q, 1224 unsigned MaxRecurse) { 1225 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse)) 1226 return V; 1227 1228 return nullptr; 1229 } 1230 1231 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout *DL, 1232 const TargetLibraryInfo *TLI, 1233 const DominatorTree *DT, AssumptionCache *AC, 1234 const Instruction *CxtI) { 1235 return ::SimplifyURemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1236 RecursionLimit); 1237 } 1238 1239 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, const Query &, 1240 unsigned) { 1241 // undef % X -> undef (the undef could be a snan). 1242 if (match(Op0, m_Undef())) 1243 return Op0; 1244 1245 // X % undef -> undef 1246 if (match(Op1, m_Undef())) 1247 return Op1; 1248 1249 return nullptr; 1250 } 1251 1252 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, const DataLayout *DL, 1253 const TargetLibraryInfo *TLI, 1254 const DominatorTree *DT, AssumptionCache *AC, 1255 const Instruction *CxtI) { 1256 return ::SimplifyFRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1257 RecursionLimit); 1258 } 1259 1260 /// isUndefShift - Returns true if a shift by \c Amount always yields undef. 1261 static bool isUndefShift(Value *Amount) { 1262 Constant *C = dyn_cast<Constant>(Amount); 1263 if (!C) 1264 return false; 1265 1266 // X shift by undef -> undef because it may shift by the bitwidth. 1267 if (isa<UndefValue>(C)) 1268 return true; 1269 1270 // Shifting by the bitwidth or more is undefined. 1271 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) 1272 if (CI->getValue().getLimitedValue() >= 1273 CI->getType()->getScalarSizeInBits()) 1274 return true; 1275 1276 // If all lanes of a vector shift are undefined the whole shift is. 1277 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) { 1278 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I) 1279 if (!isUndefShift(C->getAggregateElement(I))) 1280 return false; 1281 return true; 1282 } 1283 1284 return false; 1285 } 1286 1287 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can 1288 /// fold the result. If not, this returns null. 1289 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1, 1290 const Query &Q, unsigned MaxRecurse) { 1291 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1292 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1293 Constant *Ops[] = { C0, C1 }; 1294 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1295 } 1296 } 1297 1298 // 0 shift by X -> 0 1299 if (match(Op0, m_Zero())) 1300 return Op0; 1301 1302 // X shift by 0 -> X 1303 if (match(Op1, m_Zero())) 1304 return Op0; 1305 1306 // Fold undefined shifts. 1307 if (isUndefShift(Op1)) 1308 return UndefValue::get(Op0->getType()); 1309 1310 // If the operation is with the result of a select instruction, check whether 1311 // operating on either branch of the select always yields the same value. 1312 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1313 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1314 return V; 1315 1316 // If the operation is with the result of a phi instruction, check whether 1317 // operating on all incoming values of the phi always yields the same value. 1318 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1319 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1320 return V; 1321 1322 return nullptr; 1323 } 1324 1325 /// \brief Given operands for an Shl, LShr or AShr, see if we can 1326 /// fold the result. If not, this returns null. 1327 static Value *SimplifyRightShift(unsigned Opcode, Value *Op0, Value *Op1, 1328 bool isExact, const Query &Q, 1329 unsigned MaxRecurse) { 1330 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse)) 1331 return V; 1332 1333 // X >> X -> 0 1334 if (Op0 == Op1) 1335 return Constant::getNullValue(Op0->getType()); 1336 1337 // undef >> X -> 0 1338 // undef >> X -> undef (if it's exact) 1339 if (match(Op0, m_Undef())) 1340 return isExact ? Op0 : Constant::getNullValue(Op0->getType()); 1341 1342 // The low bit cannot be shifted out of an exact shift if it is set. 1343 if (isExact) { 1344 unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); 1345 APInt Op0KnownZero(BitWidth, 0); 1346 APInt Op0KnownOne(BitWidth, 0); 1347 computeKnownBits(Op0, Op0KnownZero, Op0KnownOne, Q.DL, /*Depth=*/0, Q.AC, 1348 Q.CxtI, Q.DT); 1349 if (Op0KnownOne[0]) 1350 return Op0; 1351 } 1352 1353 return nullptr; 1354 } 1355 1356 /// SimplifyShlInst - Given operands for an Shl, see if we can 1357 /// fold the result. If not, this returns null. 1358 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1359 const Query &Q, unsigned MaxRecurse) { 1360 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse)) 1361 return V; 1362 1363 // undef << X -> 0 1364 // undef << X -> undef if (if it's NSW/NUW) 1365 if (match(Op0, m_Undef())) 1366 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType()); 1367 1368 // (X >> A) << A -> X 1369 Value *X; 1370 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1))))) 1371 return X; 1372 return nullptr; 1373 } 1374 1375 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1376 const DataLayout *DL, const TargetLibraryInfo *TLI, 1377 const DominatorTree *DT, AssumptionCache *AC, 1378 const Instruction *CxtI) { 1379 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 1380 RecursionLimit); 1381 } 1382 1383 /// SimplifyLShrInst - Given operands for an LShr, see if we can 1384 /// fold the result. If not, this returns null. 1385 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1386 const Query &Q, unsigned MaxRecurse) { 1387 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q, 1388 MaxRecurse)) 1389 return V; 1390 1391 // (X << A) >> A -> X 1392 Value *X; 1393 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1)))) 1394 return X; 1395 1396 return nullptr; 1397 } 1398 1399 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1400 const DataLayout *DL, 1401 const TargetLibraryInfo *TLI, 1402 const DominatorTree *DT, AssumptionCache *AC, 1403 const Instruction *CxtI) { 1404 return ::SimplifyLShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1405 RecursionLimit); 1406 } 1407 1408 /// SimplifyAShrInst - Given operands for an AShr, see if we can 1409 /// fold the result. If not, this returns null. 1410 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1411 const Query &Q, unsigned MaxRecurse) { 1412 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q, 1413 MaxRecurse)) 1414 return V; 1415 1416 // all ones >>a X -> all ones 1417 if (match(Op0, m_AllOnes())) 1418 return Op0; 1419 1420 // (X << A) >> A -> X 1421 Value *X; 1422 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1)))) 1423 return X; 1424 1425 // Arithmetic shifting an all-sign-bit value is a no-op. 1426 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 1427 if (NumSignBits == Op0->getType()->getScalarSizeInBits()) 1428 return Op0; 1429 1430 return nullptr; 1431 } 1432 1433 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1434 const DataLayout *DL, 1435 const TargetLibraryInfo *TLI, 1436 const DominatorTree *DT, AssumptionCache *AC, 1437 const Instruction *CxtI) { 1438 return ::SimplifyAShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1439 RecursionLimit); 1440 } 1441 1442 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp, 1443 ICmpInst *UnsignedICmp, bool IsAnd) { 1444 Value *X, *Y; 1445 1446 ICmpInst::Predicate EqPred; 1447 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) || 1448 !ICmpInst::isEquality(EqPred)) 1449 return nullptr; 1450 1451 ICmpInst::Predicate UnsignedPred; 1452 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) && 1453 ICmpInst::isUnsigned(UnsignedPred)) 1454 ; 1455 else if (match(UnsignedICmp, 1456 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) && 1457 ICmpInst::isUnsigned(UnsignedPred)) 1458 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred); 1459 else 1460 return nullptr; 1461 1462 // X < Y && Y != 0 --> X < Y 1463 // X < Y || Y != 0 --> Y != 0 1464 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE) 1465 return IsAnd ? UnsignedICmp : ZeroICmp; 1466 1467 // X >= Y || Y != 0 --> true 1468 // X >= Y || Y == 0 --> X >= Y 1469 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) { 1470 if (EqPred == ICmpInst::ICMP_NE) 1471 return getTrue(UnsignedICmp->getType()); 1472 return UnsignedICmp; 1473 } 1474 1475 // X < Y && Y == 0 --> false 1476 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ && 1477 IsAnd) 1478 return getFalse(UnsignedICmp->getType()); 1479 1480 return nullptr; 1481 } 1482 1483 // Simplify (and (icmp ...) (icmp ...)) to true when we can tell that the range 1484 // of possible values cannot be satisfied. 1485 static Value *SimplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1486 ICmpInst::Predicate Pred0, Pred1; 1487 ConstantInt *CI1, *CI2; 1488 Value *V; 1489 1490 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true)) 1491 return X; 1492 1493 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)), 1494 m_ConstantInt(CI2)))) 1495 return nullptr; 1496 1497 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1)))) 1498 return nullptr; 1499 1500 Type *ITy = Op0->getType(); 1501 1502 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1503 bool isNSW = AddInst->hasNoSignedWrap(); 1504 bool isNUW = AddInst->hasNoUnsignedWrap(); 1505 1506 const APInt &CI1V = CI1->getValue(); 1507 const APInt &CI2V = CI2->getValue(); 1508 const APInt Delta = CI2V - CI1V; 1509 if (CI1V.isStrictlyPositive()) { 1510 if (Delta == 2) { 1511 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT) 1512 return getFalse(ITy); 1513 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1514 return getFalse(ITy); 1515 } 1516 if (Delta == 1) { 1517 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT) 1518 return getFalse(ITy); 1519 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1520 return getFalse(ITy); 1521 } 1522 } 1523 if (CI1V.getBoolValue() && isNUW) { 1524 if (Delta == 2) 1525 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT) 1526 return getFalse(ITy); 1527 if (Delta == 1) 1528 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT) 1529 return getFalse(ITy); 1530 } 1531 1532 return nullptr; 1533 } 1534 1535 /// SimplifyAndInst - Given operands for an And, see if we can 1536 /// fold the result. If not, this returns null. 1537 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q, 1538 unsigned MaxRecurse) { 1539 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1540 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1541 Constant *Ops[] = { CLHS, CRHS }; 1542 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(), 1543 Ops, Q.DL, Q.TLI); 1544 } 1545 1546 // Canonicalize the constant to the RHS. 1547 std::swap(Op0, Op1); 1548 } 1549 1550 // X & undef -> 0 1551 if (match(Op1, m_Undef())) 1552 return Constant::getNullValue(Op0->getType()); 1553 1554 // X & X = X 1555 if (Op0 == Op1) 1556 return Op0; 1557 1558 // X & 0 = 0 1559 if (match(Op1, m_Zero())) 1560 return Op1; 1561 1562 // X & -1 = X 1563 if (match(Op1, m_AllOnes())) 1564 return Op0; 1565 1566 // A & ~A = ~A & A = 0 1567 if (match(Op0, m_Not(m_Specific(Op1))) || 1568 match(Op1, m_Not(m_Specific(Op0)))) 1569 return Constant::getNullValue(Op0->getType()); 1570 1571 // (A | ?) & A = A 1572 Value *A = nullptr, *B = nullptr; 1573 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1574 (A == Op1 || B == Op1)) 1575 return Op1; 1576 1577 // A & (A | ?) = A 1578 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1579 (A == Op0 || B == Op0)) 1580 return Op0; 1581 1582 // A & (-A) = A if A is a power of two or zero. 1583 if (match(Op0, m_Neg(m_Specific(Op1))) || 1584 match(Op1, m_Neg(m_Specific(Op0)))) { 1585 if (isKnownToBeAPowerOfTwo(Op0, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT)) 1586 return Op0; 1587 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT)) 1588 return Op1; 1589 } 1590 1591 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1592 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1593 if (Value *V = SimplifyAndOfICmps(ICILHS, ICIRHS)) 1594 return V; 1595 if (Value *V = SimplifyAndOfICmps(ICIRHS, ICILHS)) 1596 return V; 1597 } 1598 } 1599 1600 // Try some generic simplifications for associative operations. 1601 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q, 1602 MaxRecurse)) 1603 return V; 1604 1605 // And distributes over Or. Try some generic simplifications based on this. 1606 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or, 1607 Q, MaxRecurse)) 1608 return V; 1609 1610 // And distributes over Xor. Try some generic simplifications based on this. 1611 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor, 1612 Q, MaxRecurse)) 1613 return V; 1614 1615 // If the operation is with the result of a select instruction, check whether 1616 // operating on either branch of the select always yields the same value. 1617 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1618 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q, 1619 MaxRecurse)) 1620 return V; 1621 1622 // If the operation is with the result of a phi instruction, check whether 1623 // operating on all incoming values of the phi always yields the same value. 1624 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1625 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q, 1626 MaxRecurse)) 1627 return V; 1628 1629 return nullptr; 1630 } 1631 1632 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout *DL, 1633 const TargetLibraryInfo *TLI, 1634 const DominatorTree *DT, AssumptionCache *AC, 1635 const Instruction *CxtI) { 1636 return ::SimplifyAndInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1637 RecursionLimit); 1638 } 1639 1640 // Simplify (or (icmp ...) (icmp ...)) to true when we can tell that the union 1641 // contains all possible values. 1642 static Value *SimplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1643 ICmpInst::Predicate Pred0, Pred1; 1644 ConstantInt *CI1, *CI2; 1645 Value *V; 1646 1647 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false)) 1648 return X; 1649 1650 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)), 1651 m_ConstantInt(CI2)))) 1652 return nullptr; 1653 1654 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1)))) 1655 return nullptr; 1656 1657 Type *ITy = Op0->getType(); 1658 1659 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1660 bool isNSW = AddInst->hasNoSignedWrap(); 1661 bool isNUW = AddInst->hasNoUnsignedWrap(); 1662 1663 const APInt &CI1V = CI1->getValue(); 1664 const APInt &CI2V = CI2->getValue(); 1665 const APInt Delta = CI2V - CI1V; 1666 if (CI1V.isStrictlyPositive()) { 1667 if (Delta == 2) { 1668 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE) 1669 return getTrue(ITy); 1670 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1671 return getTrue(ITy); 1672 } 1673 if (Delta == 1) { 1674 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE) 1675 return getTrue(ITy); 1676 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1677 return getTrue(ITy); 1678 } 1679 } 1680 if (CI1V.getBoolValue() && isNUW) { 1681 if (Delta == 2) 1682 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE) 1683 return getTrue(ITy); 1684 if (Delta == 1) 1685 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE) 1686 return getTrue(ITy); 1687 } 1688 1689 return nullptr; 1690 } 1691 1692 /// SimplifyOrInst - Given operands for an Or, see if we can 1693 /// fold the result. If not, this returns null. 1694 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q, 1695 unsigned MaxRecurse) { 1696 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1697 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1698 Constant *Ops[] = { CLHS, CRHS }; 1699 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(), 1700 Ops, Q.DL, Q.TLI); 1701 } 1702 1703 // Canonicalize the constant to the RHS. 1704 std::swap(Op0, Op1); 1705 } 1706 1707 // X | undef -> -1 1708 if (match(Op1, m_Undef())) 1709 return Constant::getAllOnesValue(Op0->getType()); 1710 1711 // X | X = X 1712 if (Op0 == Op1) 1713 return Op0; 1714 1715 // X | 0 = X 1716 if (match(Op1, m_Zero())) 1717 return Op0; 1718 1719 // X | -1 = -1 1720 if (match(Op1, m_AllOnes())) 1721 return Op1; 1722 1723 // A | ~A = ~A | A = -1 1724 if (match(Op0, m_Not(m_Specific(Op1))) || 1725 match(Op1, m_Not(m_Specific(Op0)))) 1726 return Constant::getAllOnesValue(Op0->getType()); 1727 1728 // (A & ?) | A = A 1729 Value *A = nullptr, *B = nullptr; 1730 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 1731 (A == Op1 || B == Op1)) 1732 return Op1; 1733 1734 // A | (A & ?) = A 1735 if (match(Op1, m_And(m_Value(A), m_Value(B))) && 1736 (A == Op0 || B == Op0)) 1737 return Op0; 1738 1739 // ~(A & ?) | A = -1 1740 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) && 1741 (A == Op1 || B == Op1)) 1742 return Constant::getAllOnesValue(Op1->getType()); 1743 1744 // A | ~(A & ?) = -1 1745 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) && 1746 (A == Op0 || B == Op0)) 1747 return Constant::getAllOnesValue(Op0->getType()); 1748 1749 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1750 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1751 if (Value *V = SimplifyOrOfICmps(ICILHS, ICIRHS)) 1752 return V; 1753 if (Value *V = SimplifyOrOfICmps(ICIRHS, ICILHS)) 1754 return V; 1755 } 1756 } 1757 1758 // Try some generic simplifications for associative operations. 1759 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q, 1760 MaxRecurse)) 1761 return V; 1762 1763 // Or distributes over And. Try some generic simplifications based on this. 1764 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q, 1765 MaxRecurse)) 1766 return V; 1767 1768 // If the operation is with the result of a select instruction, check whether 1769 // operating on either branch of the select always yields the same value. 1770 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1771 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q, 1772 MaxRecurse)) 1773 return V; 1774 1775 // (A & C)|(B & D) 1776 Value *C = nullptr, *D = nullptr; 1777 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 1778 match(Op1, m_And(m_Value(B), m_Value(D)))) { 1779 ConstantInt *C1 = dyn_cast<ConstantInt>(C); 1780 ConstantInt *C2 = dyn_cast<ConstantInt>(D); 1781 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) { 1782 // (A & C1)|(B & C2) 1783 // If we have: ((V + N) & C1) | (V & C2) 1784 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0 1785 // replace with V+N. 1786 Value *V1, *V2; 1787 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+ 1788 match(A, m_Add(m_Value(V1), m_Value(V2)))) { 1789 // Add commutes, try both ways. 1790 if (V1 == B && 1791 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1792 return A; 1793 if (V2 == B && 1794 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1795 return A; 1796 } 1797 // Or commutes, try both ways. 1798 if ((C1->getValue() & (C1->getValue() + 1)) == 0 && 1799 match(B, m_Add(m_Value(V1), m_Value(V2)))) { 1800 // Add commutes, try both ways. 1801 if (V1 == A && 1802 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1803 return B; 1804 if (V2 == A && 1805 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1806 return B; 1807 } 1808 } 1809 } 1810 1811 // If the operation is with the result of a phi instruction, check whether 1812 // operating on all incoming values of the phi always yields the same value. 1813 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1814 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse)) 1815 return V; 1816 1817 return nullptr; 1818 } 1819 1820 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout *DL, 1821 const TargetLibraryInfo *TLI, 1822 const DominatorTree *DT, AssumptionCache *AC, 1823 const Instruction *CxtI) { 1824 return ::SimplifyOrInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1825 RecursionLimit); 1826 } 1827 1828 /// SimplifyXorInst - Given operands for a Xor, see if we can 1829 /// fold the result. If not, this returns null. 1830 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q, 1831 unsigned MaxRecurse) { 1832 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1833 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1834 Constant *Ops[] = { CLHS, CRHS }; 1835 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(), 1836 Ops, Q.DL, Q.TLI); 1837 } 1838 1839 // Canonicalize the constant to the RHS. 1840 std::swap(Op0, Op1); 1841 } 1842 1843 // A ^ undef -> undef 1844 if (match(Op1, m_Undef())) 1845 return Op1; 1846 1847 // A ^ 0 = A 1848 if (match(Op1, m_Zero())) 1849 return Op0; 1850 1851 // A ^ A = 0 1852 if (Op0 == Op1) 1853 return Constant::getNullValue(Op0->getType()); 1854 1855 // A ^ ~A = ~A ^ A = -1 1856 if (match(Op0, m_Not(m_Specific(Op1))) || 1857 match(Op1, m_Not(m_Specific(Op0)))) 1858 return Constant::getAllOnesValue(Op0->getType()); 1859 1860 // Try some generic simplifications for associative operations. 1861 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q, 1862 MaxRecurse)) 1863 return V; 1864 1865 // Threading Xor over selects and phi nodes is pointless, so don't bother. 1866 // Threading over the select in "A ^ select(cond, B, C)" means evaluating 1867 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and 1868 // only if B and C are equal. If B and C are equal then (since we assume 1869 // that operands have already been simplified) "select(cond, B, C)" should 1870 // have been simplified to the common value of B and C already. Analysing 1871 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly 1872 // for threading over phi nodes. 1873 1874 return nullptr; 1875 } 1876 1877 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout *DL, 1878 const TargetLibraryInfo *TLI, 1879 const DominatorTree *DT, AssumptionCache *AC, 1880 const Instruction *CxtI) { 1881 return ::SimplifyXorInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1882 RecursionLimit); 1883 } 1884 1885 static Type *GetCompareTy(Value *Op) { 1886 return CmpInst::makeCmpResultType(Op->getType()); 1887 } 1888 1889 /// ExtractEquivalentCondition - Rummage around inside V looking for something 1890 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found, 1891 /// otherwise return null. Helper function for analyzing max/min idioms. 1892 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred, 1893 Value *LHS, Value *RHS) { 1894 SelectInst *SI = dyn_cast<SelectInst>(V); 1895 if (!SI) 1896 return nullptr; 1897 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); 1898 if (!Cmp) 1899 return nullptr; 1900 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1); 1901 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS) 1902 return Cmp; 1903 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) && 1904 LHS == CmpRHS && RHS == CmpLHS) 1905 return Cmp; 1906 return nullptr; 1907 } 1908 1909 // A significant optimization not implemented here is assuming that alloca 1910 // addresses are not equal to incoming argument values. They don't *alias*, 1911 // as we say, but that doesn't mean they aren't equal, so we take a 1912 // conservative approach. 1913 // 1914 // This is inspired in part by C++11 5.10p1: 1915 // "Two pointers of the same type compare equal if and only if they are both 1916 // null, both point to the same function, or both represent the same 1917 // address." 1918 // 1919 // This is pretty permissive. 1920 // 1921 // It's also partly due to C11 6.5.9p6: 1922 // "Two pointers compare equal if and only if both are null pointers, both are 1923 // pointers to the same object (including a pointer to an object and a 1924 // subobject at its beginning) or function, both are pointers to one past the 1925 // last element of the same array object, or one is a pointer to one past the 1926 // end of one array object and the other is a pointer to the start of a 1927 // different array object that happens to immediately follow the first array 1928 // object in the address space.) 1929 // 1930 // C11's version is more restrictive, however there's no reason why an argument 1931 // couldn't be a one-past-the-end value for a stack object in the caller and be 1932 // equal to the beginning of a stack object in the callee. 1933 // 1934 // If the C and C++ standards are ever made sufficiently restrictive in this 1935 // area, it may be possible to update LLVM's semantics accordingly and reinstate 1936 // this optimization. 1937 static Constant *computePointerICmp(const DataLayout *DL, 1938 const TargetLibraryInfo *TLI, 1939 CmpInst::Predicate Pred, 1940 Value *LHS, Value *RHS) { 1941 // First, skip past any trivial no-ops. 1942 LHS = LHS->stripPointerCasts(); 1943 RHS = RHS->stripPointerCasts(); 1944 1945 // A non-null pointer is not equal to a null pointer. 1946 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) && 1947 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE)) 1948 return ConstantInt::get(GetCompareTy(LHS), 1949 !CmpInst::isTrueWhenEqual(Pred)); 1950 1951 // We can only fold certain predicates on pointer comparisons. 1952 switch (Pred) { 1953 default: 1954 return nullptr; 1955 1956 // Equality comaprisons are easy to fold. 1957 case CmpInst::ICMP_EQ: 1958 case CmpInst::ICMP_NE: 1959 break; 1960 1961 // We can only handle unsigned relational comparisons because 'inbounds' on 1962 // a GEP only protects against unsigned wrapping. 1963 case CmpInst::ICMP_UGT: 1964 case CmpInst::ICMP_UGE: 1965 case CmpInst::ICMP_ULT: 1966 case CmpInst::ICMP_ULE: 1967 // However, we have to switch them to their signed variants to handle 1968 // negative indices from the base pointer. 1969 Pred = ICmpInst::getSignedPredicate(Pred); 1970 break; 1971 } 1972 1973 // Strip off any constant offsets so that we can reason about them. 1974 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets 1975 // here and compare base addresses like AliasAnalysis does, however there are 1976 // numerous hazards. AliasAnalysis and its utilities rely on special rules 1977 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis 1978 // doesn't need to guarantee pointer inequality when it says NoAlias. 1979 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 1980 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 1981 1982 // If LHS and RHS are related via constant offsets to the same base 1983 // value, we can replace it with an icmp which just compares the offsets. 1984 if (LHS == RHS) 1985 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset); 1986 1987 // Various optimizations for (in)equality comparisons. 1988 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) { 1989 // Different non-empty allocations that exist at the same time have 1990 // different addresses (if the program can tell). Global variables always 1991 // exist, so they always exist during the lifetime of each other and all 1992 // allocas. Two different allocas usually have different addresses... 1993 // 1994 // However, if there's an @llvm.stackrestore dynamically in between two 1995 // allocas, they may have the same address. It's tempting to reduce the 1996 // scope of the problem by only looking at *static* allocas here. That would 1997 // cover the majority of allocas while significantly reducing the likelihood 1998 // of having an @llvm.stackrestore pop up in the middle. However, it's not 1999 // actually impossible for an @llvm.stackrestore to pop up in the middle of 2000 // an entry block. Also, if we have a block that's not attached to a 2001 // function, we can't tell if it's "static" under the current definition. 2002 // Theoretically, this problem could be fixed by creating a new kind of 2003 // instruction kind specifically for static allocas. Such a new instruction 2004 // could be required to be at the top of the entry block, thus preventing it 2005 // from being subject to a @llvm.stackrestore. Instcombine could even 2006 // convert regular allocas into these special allocas. It'd be nifty. 2007 // However, until then, this problem remains open. 2008 // 2009 // So, we'll assume that two non-empty allocas have different addresses 2010 // for now. 2011 // 2012 // With all that, if the offsets are within the bounds of their allocations 2013 // (and not one-past-the-end! so we can't use inbounds!), and their 2014 // allocations aren't the same, the pointers are not equal. 2015 // 2016 // Note that it's not necessary to check for LHS being a global variable 2017 // address, due to canonicalization and constant folding. 2018 if (isa<AllocaInst>(LHS) && 2019 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) { 2020 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset); 2021 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset); 2022 uint64_t LHSSize, RHSSize; 2023 if (LHSOffsetCI && RHSOffsetCI && 2024 getObjectSize(LHS, LHSSize, DL, TLI) && 2025 getObjectSize(RHS, RHSSize, DL, TLI)) { 2026 const APInt &LHSOffsetValue = LHSOffsetCI->getValue(); 2027 const APInt &RHSOffsetValue = RHSOffsetCI->getValue(); 2028 if (!LHSOffsetValue.isNegative() && 2029 !RHSOffsetValue.isNegative() && 2030 LHSOffsetValue.ult(LHSSize) && 2031 RHSOffsetValue.ult(RHSSize)) { 2032 return ConstantInt::get(GetCompareTy(LHS), 2033 !CmpInst::isTrueWhenEqual(Pred)); 2034 } 2035 } 2036 2037 // Repeat the above check but this time without depending on DataLayout 2038 // or being able to compute a precise size. 2039 if (!cast<PointerType>(LHS->getType())->isEmptyTy() && 2040 !cast<PointerType>(RHS->getType())->isEmptyTy() && 2041 LHSOffset->isNullValue() && 2042 RHSOffset->isNullValue()) 2043 return ConstantInt::get(GetCompareTy(LHS), 2044 !CmpInst::isTrueWhenEqual(Pred)); 2045 } 2046 2047 // Even if an non-inbounds GEP occurs along the path we can still optimize 2048 // equality comparisons concerning the result. We avoid walking the whole 2049 // chain again by starting where the last calls to 2050 // stripAndComputeConstantOffsets left off and accumulate the offsets. 2051 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true); 2052 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true); 2053 if (LHS == RHS) 2054 return ConstantExpr::getICmp(Pred, 2055 ConstantExpr::getAdd(LHSOffset, LHSNoBound), 2056 ConstantExpr::getAdd(RHSOffset, RHSNoBound)); 2057 2058 // If one side of the equality comparison must come from a noalias call 2059 // (meaning a system memory allocation function), and the other side must 2060 // come from a pointer that cannot overlap with dynamically-allocated 2061 // memory within the lifetime of the current function (allocas, byval 2062 // arguments, globals), then determine the comparison result here. 2063 SmallVector<Value *, 8> LHSUObjs, RHSUObjs; 2064 GetUnderlyingObjects(LHS, LHSUObjs, DL); 2065 GetUnderlyingObjects(RHS, RHSUObjs, DL); 2066 2067 // Is the set of underlying objects all noalias calls? 2068 auto IsNAC = [](SmallVectorImpl<Value *> &Objects) { 2069 return std::all_of(Objects.begin(), Objects.end(), 2070 [](Value *V){ return isNoAliasCall(V); }); 2071 }; 2072 2073 // Is the set of underlying objects all things which must be disjoint from 2074 // noalias calls. For allocas, we consider only static ones (dynamic 2075 // allocas might be transformed into calls to malloc not simultaneously 2076 // live with the compared-to allocation). For globals, we exclude symbols 2077 // that might be resolve lazily to symbols in another dynamically-loaded 2078 // library (and, thus, could be malloc'ed by the implementation). 2079 auto IsAllocDisjoint = [](SmallVectorImpl<Value *> &Objects) { 2080 return std::all_of(Objects.begin(), Objects.end(), 2081 [](Value *V){ 2082 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) 2083 return AI->getParent() && AI->getParent()->getParent() && 2084 AI->isStaticAlloca(); 2085 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 2086 return (GV->hasLocalLinkage() || 2087 GV->hasHiddenVisibility() || 2088 GV->hasProtectedVisibility() || 2089 GV->hasUnnamedAddr()) && 2090 !GV->isThreadLocal(); 2091 if (const Argument *A = dyn_cast<Argument>(V)) 2092 return A->hasByValAttr(); 2093 return false; 2094 }); 2095 }; 2096 2097 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) || 2098 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs))) 2099 return ConstantInt::get(GetCompareTy(LHS), 2100 !CmpInst::isTrueWhenEqual(Pred)); 2101 } 2102 2103 // Otherwise, fail. 2104 return nullptr; 2105 } 2106 2107 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can 2108 /// fold the result. If not, this returns null. 2109 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2110 const Query &Q, unsigned MaxRecurse) { 2111 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 2112 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!"); 2113 2114 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 2115 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 2116 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 2117 2118 // If we have a constant, make sure it is on the RHS. 2119 std::swap(LHS, RHS); 2120 Pred = CmpInst::getSwappedPredicate(Pred); 2121 } 2122 2123 Type *ITy = GetCompareTy(LHS); // The return type. 2124 Type *OpTy = LHS->getType(); // The operand type. 2125 2126 // icmp X, X -> true/false 2127 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false 2128 // because X could be 0. 2129 if (LHS == RHS || isa<UndefValue>(RHS)) 2130 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred)); 2131 2132 // Special case logic when the operands have i1 type. 2133 if (OpTy->getScalarType()->isIntegerTy(1)) { 2134 switch (Pred) { 2135 default: break; 2136 case ICmpInst::ICMP_EQ: 2137 // X == 1 -> X 2138 if (match(RHS, m_One())) 2139 return LHS; 2140 break; 2141 case ICmpInst::ICMP_NE: 2142 // X != 0 -> X 2143 if (match(RHS, m_Zero())) 2144 return LHS; 2145 break; 2146 case ICmpInst::ICMP_UGT: 2147 // X >u 0 -> X 2148 if (match(RHS, m_Zero())) 2149 return LHS; 2150 break; 2151 case ICmpInst::ICMP_UGE: 2152 // X >=u 1 -> X 2153 if (match(RHS, m_One())) 2154 return LHS; 2155 break; 2156 case ICmpInst::ICMP_SLT: 2157 // X <s 0 -> X 2158 if (match(RHS, m_Zero())) 2159 return LHS; 2160 break; 2161 case ICmpInst::ICMP_SLE: 2162 // X <=s -1 -> X 2163 if (match(RHS, m_One())) 2164 return LHS; 2165 break; 2166 } 2167 } 2168 2169 // If we are comparing with zero then try hard since this is a common case. 2170 if (match(RHS, m_Zero())) { 2171 bool LHSKnownNonNegative, LHSKnownNegative; 2172 switch (Pred) { 2173 default: llvm_unreachable("Unknown ICmp predicate!"); 2174 case ICmpInst::ICMP_ULT: 2175 return getFalse(ITy); 2176 case ICmpInst::ICMP_UGE: 2177 return getTrue(ITy); 2178 case ICmpInst::ICMP_EQ: 2179 case ICmpInst::ICMP_ULE: 2180 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2181 return getFalse(ITy); 2182 break; 2183 case ICmpInst::ICMP_NE: 2184 case ICmpInst::ICMP_UGT: 2185 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2186 return getTrue(ITy); 2187 break; 2188 case ICmpInst::ICMP_SLT: 2189 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2190 Q.CxtI, Q.DT); 2191 if (LHSKnownNegative) 2192 return getTrue(ITy); 2193 if (LHSKnownNonNegative) 2194 return getFalse(ITy); 2195 break; 2196 case ICmpInst::ICMP_SLE: 2197 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2198 Q.CxtI, Q.DT); 2199 if (LHSKnownNegative) 2200 return getTrue(ITy); 2201 if (LHSKnownNonNegative && 2202 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2203 return getFalse(ITy); 2204 break; 2205 case ICmpInst::ICMP_SGE: 2206 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2207 Q.CxtI, Q.DT); 2208 if (LHSKnownNegative) 2209 return getFalse(ITy); 2210 if (LHSKnownNonNegative) 2211 return getTrue(ITy); 2212 break; 2213 case ICmpInst::ICMP_SGT: 2214 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2215 Q.CxtI, Q.DT); 2216 if (LHSKnownNegative) 2217 return getFalse(ITy); 2218 if (LHSKnownNonNegative && 2219 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2220 return getTrue(ITy); 2221 break; 2222 } 2223 } 2224 2225 // See if we are doing a comparison with a constant integer. 2226 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2227 // Rule out tautological comparisons (eg., ult 0 or uge 0). 2228 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue()); 2229 if (RHS_CR.isEmptySet()) 2230 return ConstantInt::getFalse(CI->getContext()); 2231 if (RHS_CR.isFullSet()) 2232 return ConstantInt::getTrue(CI->getContext()); 2233 2234 // Many binary operators with constant RHS have easy to compute constant 2235 // range. Use them to check whether the comparison is a tautology. 2236 unsigned Width = CI->getBitWidth(); 2237 APInt Lower = APInt(Width, 0); 2238 APInt Upper = APInt(Width, 0); 2239 ConstantInt *CI2; 2240 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) { 2241 // 'urem x, CI2' produces [0, CI2). 2242 Upper = CI2->getValue(); 2243 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) { 2244 // 'srem x, CI2' produces (-|CI2|, |CI2|). 2245 Upper = CI2->getValue().abs(); 2246 Lower = (-Upper) + 1; 2247 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) { 2248 // 'udiv CI2, x' produces [0, CI2]. 2249 Upper = CI2->getValue() + 1; 2250 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) { 2251 // 'udiv x, CI2' produces [0, UINT_MAX / CI2]. 2252 APInt NegOne = APInt::getAllOnesValue(Width); 2253 if (!CI2->isZero()) 2254 Upper = NegOne.udiv(CI2->getValue()) + 1; 2255 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) { 2256 if (CI2->isMinSignedValue()) { 2257 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. 2258 Lower = CI2->getValue(); 2259 Upper = Lower.lshr(1) + 1; 2260 } else { 2261 // 'sdiv CI2, x' produces [-|CI2|, |CI2|]. 2262 Upper = CI2->getValue().abs() + 1; 2263 Lower = (-Upper) + 1; 2264 } 2265 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) { 2266 APInt IntMin = APInt::getSignedMinValue(Width); 2267 APInt IntMax = APInt::getSignedMaxValue(Width); 2268 APInt Val = CI2->getValue(); 2269 if (Val.isAllOnesValue()) { 2270 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] 2271 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2272 Lower = IntMin + 1; 2273 Upper = IntMax + 1; 2274 } else if (Val.countLeadingZeros() < Width - 1) { 2275 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2] 2276 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2277 Lower = IntMin.sdiv(Val); 2278 Upper = IntMax.sdiv(Val); 2279 if (Lower.sgt(Upper)) 2280 std::swap(Lower, Upper); 2281 Upper = Upper + 1; 2282 assert(Upper != Lower && "Upper part of range has wrapped!"); 2283 } 2284 } else if (match(LHS, m_NUWShl(m_ConstantInt(CI2), m_Value()))) { 2285 // 'shl nuw CI2, x' produces [CI2, CI2 << CLZ(CI2)] 2286 Lower = CI2->getValue(); 2287 Upper = Lower.shl(Lower.countLeadingZeros()) + 1; 2288 } else if (match(LHS, m_NSWShl(m_ConstantInt(CI2), m_Value()))) { 2289 if (CI2->isNegative()) { 2290 // 'shl nsw CI2, x' produces [CI2 << CLO(CI2)-1, CI2] 2291 unsigned ShiftAmount = CI2->getValue().countLeadingOnes() - 1; 2292 Lower = CI2->getValue().shl(ShiftAmount); 2293 Upper = CI2->getValue() + 1; 2294 } else { 2295 // 'shl nsw CI2, x' produces [CI2, CI2 << CLZ(CI2)-1] 2296 unsigned ShiftAmount = CI2->getValue().countLeadingZeros() - 1; 2297 Lower = CI2->getValue(); 2298 Upper = CI2->getValue().shl(ShiftAmount) + 1; 2299 } 2300 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) { 2301 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2]. 2302 APInt NegOne = APInt::getAllOnesValue(Width); 2303 if (CI2->getValue().ult(Width)) 2304 Upper = NegOne.lshr(CI2->getValue()) + 1; 2305 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) { 2306 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2]. 2307 unsigned ShiftAmount = Width - 1; 2308 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2309 ShiftAmount = CI2->getValue().countTrailingZeros(); 2310 Lower = CI2->getValue().lshr(ShiftAmount); 2311 Upper = CI2->getValue() + 1; 2312 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) { 2313 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2]. 2314 APInt IntMin = APInt::getSignedMinValue(Width); 2315 APInt IntMax = APInt::getSignedMaxValue(Width); 2316 if (CI2->getValue().ult(Width)) { 2317 Lower = IntMin.ashr(CI2->getValue()); 2318 Upper = IntMax.ashr(CI2->getValue()) + 1; 2319 } 2320 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) { 2321 unsigned ShiftAmount = Width - 1; 2322 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2323 ShiftAmount = CI2->getValue().countTrailingZeros(); 2324 if (CI2->isNegative()) { 2325 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)] 2326 Lower = CI2->getValue(); 2327 Upper = CI2->getValue().ashr(ShiftAmount) + 1; 2328 } else { 2329 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2] 2330 Lower = CI2->getValue().ashr(ShiftAmount); 2331 Upper = CI2->getValue() + 1; 2332 } 2333 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) { 2334 // 'or x, CI2' produces [CI2, UINT_MAX]. 2335 Lower = CI2->getValue(); 2336 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) { 2337 // 'and x, CI2' produces [0, CI2]. 2338 Upper = CI2->getValue() + 1; 2339 } 2340 if (Lower != Upper) { 2341 ConstantRange LHS_CR = ConstantRange(Lower, Upper); 2342 if (RHS_CR.contains(LHS_CR)) 2343 return ConstantInt::getTrue(RHS->getContext()); 2344 if (RHS_CR.inverse().contains(LHS_CR)) 2345 return ConstantInt::getFalse(RHS->getContext()); 2346 } 2347 } 2348 2349 // Compare of cast, for example (zext X) != 0 -> X != 0 2350 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) { 2351 Instruction *LI = cast<CastInst>(LHS); 2352 Value *SrcOp = LI->getOperand(0); 2353 Type *SrcTy = SrcOp->getType(); 2354 Type *DstTy = LI->getType(); 2355 2356 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input 2357 // if the integer type is the same size as the pointer type. 2358 if (MaxRecurse && Q.DL && isa<PtrToIntInst>(LI) && 2359 Q.DL->getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) { 2360 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 2361 // Transfer the cast to the constant. 2362 if (Value *V = SimplifyICmpInst(Pred, SrcOp, 2363 ConstantExpr::getIntToPtr(RHSC, SrcTy), 2364 Q, MaxRecurse-1)) 2365 return V; 2366 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) { 2367 if (RI->getOperand(0)->getType() == SrcTy) 2368 // Compare without the cast. 2369 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2370 Q, MaxRecurse-1)) 2371 return V; 2372 } 2373 } 2374 2375 if (isa<ZExtInst>(LHS)) { 2376 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the 2377 // same type. 2378 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) { 2379 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2380 // Compare X and Y. Note that signed predicates become unsigned. 2381 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2382 SrcOp, RI->getOperand(0), Q, 2383 MaxRecurse-1)) 2384 return V; 2385 } 2386 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended 2387 // too. If not, then try to deduce the result of the comparison. 2388 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2389 // Compute the constant that would happen if we truncated to SrcTy then 2390 // reextended to DstTy. 2391 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2392 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy); 2393 2394 // If the re-extended constant didn't change then this is effectively 2395 // also a case of comparing two zero-extended values. 2396 if (RExt == CI && MaxRecurse) 2397 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2398 SrcOp, Trunc, Q, MaxRecurse-1)) 2399 return V; 2400 2401 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit 2402 // there. Use this to work out the result of the comparison. 2403 if (RExt != CI) { 2404 switch (Pred) { 2405 default: llvm_unreachable("Unknown ICmp predicate!"); 2406 // LHS <u RHS. 2407 case ICmpInst::ICMP_EQ: 2408 case ICmpInst::ICMP_UGT: 2409 case ICmpInst::ICMP_UGE: 2410 return ConstantInt::getFalse(CI->getContext()); 2411 2412 case ICmpInst::ICMP_NE: 2413 case ICmpInst::ICMP_ULT: 2414 case ICmpInst::ICMP_ULE: 2415 return ConstantInt::getTrue(CI->getContext()); 2416 2417 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS 2418 // is non-negative then LHS <s RHS. 2419 case ICmpInst::ICMP_SGT: 2420 case ICmpInst::ICMP_SGE: 2421 return CI->getValue().isNegative() ? 2422 ConstantInt::getTrue(CI->getContext()) : 2423 ConstantInt::getFalse(CI->getContext()); 2424 2425 case ICmpInst::ICMP_SLT: 2426 case ICmpInst::ICMP_SLE: 2427 return CI->getValue().isNegative() ? 2428 ConstantInt::getFalse(CI->getContext()) : 2429 ConstantInt::getTrue(CI->getContext()); 2430 } 2431 } 2432 } 2433 } 2434 2435 if (isa<SExtInst>(LHS)) { 2436 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the 2437 // same type. 2438 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) { 2439 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2440 // Compare X and Y. Note that the predicate does not change. 2441 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2442 Q, MaxRecurse-1)) 2443 return V; 2444 } 2445 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended 2446 // too. If not, then try to deduce the result of the comparison. 2447 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2448 // Compute the constant that would happen if we truncated to SrcTy then 2449 // reextended to DstTy. 2450 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2451 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy); 2452 2453 // If the re-extended constant didn't change then this is effectively 2454 // also a case of comparing two sign-extended values. 2455 if (RExt == CI && MaxRecurse) 2456 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1)) 2457 return V; 2458 2459 // Otherwise the upper bits of LHS are all equal, while RHS has varying 2460 // bits there. Use this to work out the result of the comparison. 2461 if (RExt != CI) { 2462 switch (Pred) { 2463 default: llvm_unreachable("Unknown ICmp predicate!"); 2464 case ICmpInst::ICMP_EQ: 2465 return ConstantInt::getFalse(CI->getContext()); 2466 case ICmpInst::ICMP_NE: 2467 return ConstantInt::getTrue(CI->getContext()); 2468 2469 // If RHS is non-negative then LHS <s RHS. If RHS is negative then 2470 // LHS >s RHS. 2471 case ICmpInst::ICMP_SGT: 2472 case ICmpInst::ICMP_SGE: 2473 return CI->getValue().isNegative() ? 2474 ConstantInt::getTrue(CI->getContext()) : 2475 ConstantInt::getFalse(CI->getContext()); 2476 case ICmpInst::ICMP_SLT: 2477 case ICmpInst::ICMP_SLE: 2478 return CI->getValue().isNegative() ? 2479 ConstantInt::getFalse(CI->getContext()) : 2480 ConstantInt::getTrue(CI->getContext()); 2481 2482 // If LHS is non-negative then LHS <u RHS. If LHS is negative then 2483 // LHS >u RHS. 2484 case ICmpInst::ICMP_UGT: 2485 case ICmpInst::ICMP_UGE: 2486 // Comparison is true iff the LHS <s 0. 2487 if (MaxRecurse) 2488 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp, 2489 Constant::getNullValue(SrcTy), 2490 Q, MaxRecurse-1)) 2491 return V; 2492 break; 2493 case ICmpInst::ICMP_ULT: 2494 case ICmpInst::ICMP_ULE: 2495 // Comparison is true iff the LHS >=s 0. 2496 if (MaxRecurse) 2497 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp, 2498 Constant::getNullValue(SrcTy), 2499 Q, MaxRecurse-1)) 2500 return V; 2501 break; 2502 } 2503 } 2504 } 2505 } 2506 } 2507 2508 // Special logic for binary operators. 2509 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS); 2510 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS); 2511 if (MaxRecurse && (LBO || RBO)) { 2512 // Analyze the case when either LHS or RHS is an add instruction. 2513 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 2514 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null). 2515 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false; 2516 if (LBO && LBO->getOpcode() == Instruction::Add) { 2517 A = LBO->getOperand(0); B = LBO->getOperand(1); 2518 NoLHSWrapProblem = ICmpInst::isEquality(Pred) || 2519 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) || 2520 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap()); 2521 } 2522 if (RBO && RBO->getOpcode() == Instruction::Add) { 2523 C = RBO->getOperand(0); D = RBO->getOperand(1); 2524 NoRHSWrapProblem = ICmpInst::isEquality(Pred) || 2525 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) || 2526 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap()); 2527 } 2528 2529 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2530 if ((A == RHS || B == RHS) && NoLHSWrapProblem) 2531 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A, 2532 Constant::getNullValue(RHS->getType()), 2533 Q, MaxRecurse-1)) 2534 return V; 2535 2536 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2537 if ((C == LHS || D == LHS) && NoRHSWrapProblem) 2538 if (Value *V = SimplifyICmpInst(Pred, 2539 Constant::getNullValue(LHS->getType()), 2540 C == LHS ? D : C, Q, MaxRecurse-1)) 2541 return V; 2542 2543 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow. 2544 if (A && C && (A == C || A == D || B == C || B == D) && 2545 NoLHSWrapProblem && NoRHSWrapProblem) { 2546 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2547 Value *Y, *Z; 2548 if (A == C) { 2549 // C + B == C + D -> B == D 2550 Y = B; 2551 Z = D; 2552 } else if (A == D) { 2553 // D + B == C + D -> B == C 2554 Y = B; 2555 Z = C; 2556 } else if (B == C) { 2557 // A + C == C + D -> A == D 2558 Y = A; 2559 Z = D; 2560 } else { 2561 assert(B == D); 2562 // A + D == C + D -> A == C 2563 Y = A; 2564 Z = C; 2565 } 2566 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1)) 2567 return V; 2568 } 2569 } 2570 2571 // icmp pred (or X, Y), X 2572 if (LBO && match(LBO, m_CombineOr(m_Or(m_Value(), m_Specific(RHS)), 2573 m_Or(m_Specific(RHS), m_Value())))) { 2574 if (Pred == ICmpInst::ICMP_ULT) 2575 return getFalse(ITy); 2576 if (Pred == ICmpInst::ICMP_UGE) 2577 return getTrue(ITy); 2578 } 2579 // icmp pred X, (or X, Y) 2580 if (RBO && match(RBO, m_CombineOr(m_Or(m_Value(), m_Specific(LHS)), 2581 m_Or(m_Specific(LHS), m_Value())))) { 2582 if (Pred == ICmpInst::ICMP_ULE) 2583 return getTrue(ITy); 2584 if (Pred == ICmpInst::ICMP_UGT) 2585 return getFalse(ITy); 2586 } 2587 2588 // icmp pred (and X, Y), X 2589 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)), 2590 m_And(m_Specific(RHS), m_Value())))) { 2591 if (Pred == ICmpInst::ICMP_UGT) 2592 return getFalse(ITy); 2593 if (Pred == ICmpInst::ICMP_ULE) 2594 return getTrue(ITy); 2595 } 2596 // icmp pred X, (and X, Y) 2597 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)), 2598 m_And(m_Specific(LHS), m_Value())))) { 2599 if (Pred == ICmpInst::ICMP_UGE) 2600 return getTrue(ITy); 2601 if (Pred == ICmpInst::ICMP_ULT) 2602 return getFalse(ITy); 2603 } 2604 2605 // 0 - (zext X) pred C 2606 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) { 2607 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 2608 if (RHSC->getValue().isStrictlyPositive()) { 2609 if (Pred == ICmpInst::ICMP_SLT) 2610 return ConstantInt::getTrue(RHSC->getContext()); 2611 if (Pred == ICmpInst::ICMP_SGE) 2612 return ConstantInt::getFalse(RHSC->getContext()); 2613 if (Pred == ICmpInst::ICMP_EQ) 2614 return ConstantInt::getFalse(RHSC->getContext()); 2615 if (Pred == ICmpInst::ICMP_NE) 2616 return ConstantInt::getTrue(RHSC->getContext()); 2617 } 2618 if (RHSC->getValue().isNonNegative()) { 2619 if (Pred == ICmpInst::ICMP_SLE) 2620 return ConstantInt::getTrue(RHSC->getContext()); 2621 if (Pred == ICmpInst::ICMP_SGT) 2622 return ConstantInt::getFalse(RHSC->getContext()); 2623 } 2624 } 2625 } 2626 2627 // icmp pred (urem X, Y), Y 2628 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) { 2629 bool KnownNonNegative, KnownNegative; 2630 switch (Pred) { 2631 default: 2632 break; 2633 case ICmpInst::ICMP_SGT: 2634 case ICmpInst::ICMP_SGE: 2635 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2636 Q.CxtI, Q.DT); 2637 if (!KnownNonNegative) 2638 break; 2639 // fall-through 2640 case ICmpInst::ICMP_EQ: 2641 case ICmpInst::ICMP_UGT: 2642 case ICmpInst::ICMP_UGE: 2643 return getFalse(ITy); 2644 case ICmpInst::ICMP_SLT: 2645 case ICmpInst::ICMP_SLE: 2646 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2647 Q.CxtI, Q.DT); 2648 if (!KnownNonNegative) 2649 break; 2650 // fall-through 2651 case ICmpInst::ICMP_NE: 2652 case ICmpInst::ICMP_ULT: 2653 case ICmpInst::ICMP_ULE: 2654 return getTrue(ITy); 2655 } 2656 } 2657 2658 // icmp pred X, (urem Y, X) 2659 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) { 2660 bool KnownNonNegative, KnownNegative; 2661 switch (Pred) { 2662 default: 2663 break; 2664 case ICmpInst::ICMP_SGT: 2665 case ICmpInst::ICMP_SGE: 2666 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2667 Q.CxtI, Q.DT); 2668 if (!KnownNonNegative) 2669 break; 2670 // fall-through 2671 case ICmpInst::ICMP_NE: 2672 case ICmpInst::ICMP_UGT: 2673 case ICmpInst::ICMP_UGE: 2674 return getTrue(ITy); 2675 case ICmpInst::ICMP_SLT: 2676 case ICmpInst::ICMP_SLE: 2677 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2678 Q.CxtI, Q.DT); 2679 if (!KnownNonNegative) 2680 break; 2681 // fall-through 2682 case ICmpInst::ICMP_EQ: 2683 case ICmpInst::ICMP_ULT: 2684 case ICmpInst::ICMP_ULE: 2685 return getFalse(ITy); 2686 } 2687 } 2688 2689 // x udiv y <=u x. 2690 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) { 2691 // icmp pred (X /u Y), X 2692 if (Pred == ICmpInst::ICMP_UGT) 2693 return getFalse(ITy); 2694 if (Pred == ICmpInst::ICMP_ULE) 2695 return getTrue(ITy); 2696 } 2697 2698 // handle: 2699 // CI2 << X == CI 2700 // CI2 << X != CI 2701 // 2702 // where CI2 is a power of 2 and CI isn't 2703 if (auto *CI = dyn_cast<ConstantInt>(RHS)) { 2704 const APInt *CI2Val, *CIVal = &CI->getValue(); 2705 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) && 2706 CI2Val->isPowerOf2()) { 2707 if (!CIVal->isPowerOf2()) { 2708 // CI2 << X can equal zero in some circumstances, 2709 // this simplification is unsafe if CI is zero. 2710 // 2711 // We know it is safe if: 2712 // - The shift is nsw, we can't shift out the one bit. 2713 // - The shift is nuw, we can't shift out the one bit. 2714 // - CI2 is one 2715 // - CI isn't zero 2716 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() || 2717 *CI2Val == 1 || !CI->isZero()) { 2718 if (Pred == ICmpInst::ICMP_EQ) 2719 return ConstantInt::getFalse(RHS->getContext()); 2720 if (Pred == ICmpInst::ICMP_NE) 2721 return ConstantInt::getTrue(RHS->getContext()); 2722 } 2723 } 2724 if (CIVal->isSignBit() && *CI2Val == 1) { 2725 if (Pred == ICmpInst::ICMP_UGT) 2726 return ConstantInt::getFalse(RHS->getContext()); 2727 if (Pred == ICmpInst::ICMP_ULE) 2728 return ConstantInt::getTrue(RHS->getContext()); 2729 } 2730 } 2731 } 2732 2733 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() && 2734 LBO->getOperand(1) == RBO->getOperand(1)) { 2735 switch (LBO->getOpcode()) { 2736 default: break; 2737 case Instruction::UDiv: 2738 case Instruction::LShr: 2739 if (ICmpInst::isSigned(Pred)) 2740 break; 2741 // fall-through 2742 case Instruction::SDiv: 2743 case Instruction::AShr: 2744 if (!LBO->isExact() || !RBO->isExact()) 2745 break; 2746 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2747 RBO->getOperand(0), Q, MaxRecurse-1)) 2748 return V; 2749 break; 2750 case Instruction::Shl: { 2751 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap(); 2752 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap(); 2753 if (!NUW && !NSW) 2754 break; 2755 if (!NSW && ICmpInst::isSigned(Pred)) 2756 break; 2757 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2758 RBO->getOperand(0), Q, MaxRecurse-1)) 2759 return V; 2760 break; 2761 } 2762 } 2763 } 2764 2765 // Simplify comparisons involving max/min. 2766 Value *A, *B; 2767 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE; 2768 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B". 2769 2770 // Signed variants on "max(a,b)>=a -> true". 2771 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2772 if (A != RHS) std::swap(A, B); // smax(A, B) pred A. 2773 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2774 // We analyze this as smax(A, B) pred A. 2775 P = Pred; 2776 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) && 2777 (A == LHS || B == LHS)) { 2778 if (A != LHS) std::swap(A, B); // A pred smax(A, B). 2779 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2780 // We analyze this as smax(A, B) swapped-pred A. 2781 P = CmpInst::getSwappedPredicate(Pred); 2782 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2783 (A == RHS || B == RHS)) { 2784 if (A != RHS) std::swap(A, B); // smin(A, B) pred A. 2785 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2786 // We analyze this as smax(-A, -B) swapped-pred -A. 2787 // Note that we do not need to actually form -A or -B thanks to EqP. 2788 P = CmpInst::getSwappedPredicate(Pred); 2789 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) && 2790 (A == LHS || B == LHS)) { 2791 if (A != LHS) std::swap(A, B); // A pred smin(A, B). 2792 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2793 // We analyze this as smax(-A, -B) pred -A. 2794 // Note that we do not need to actually form -A or -B thanks to EqP. 2795 P = Pred; 2796 } 2797 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2798 // Cases correspond to "max(A, B) p A". 2799 switch (P) { 2800 default: 2801 break; 2802 case CmpInst::ICMP_EQ: 2803 case CmpInst::ICMP_SLE: 2804 // Equivalent to "A EqP B". This may be the same as the condition tested 2805 // in the max/min; if so, we can just return that. 2806 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2807 return V; 2808 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2809 return V; 2810 // Otherwise, see if "A EqP B" simplifies. 2811 if (MaxRecurse) 2812 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2813 return V; 2814 break; 2815 case CmpInst::ICMP_NE: 2816 case CmpInst::ICMP_SGT: { 2817 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2818 // Equivalent to "A InvEqP B". This may be the same as the condition 2819 // tested in the max/min; if so, we can just return that. 2820 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2821 return V; 2822 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2823 return V; 2824 // Otherwise, see if "A InvEqP B" simplifies. 2825 if (MaxRecurse) 2826 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2827 return V; 2828 break; 2829 } 2830 case CmpInst::ICMP_SGE: 2831 // Always true. 2832 return getTrue(ITy); 2833 case CmpInst::ICMP_SLT: 2834 // Always false. 2835 return getFalse(ITy); 2836 } 2837 } 2838 2839 // Unsigned variants on "max(a,b)>=a -> true". 2840 P = CmpInst::BAD_ICMP_PREDICATE; 2841 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2842 if (A != RHS) std::swap(A, B); // umax(A, B) pred A. 2843 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2844 // We analyze this as umax(A, B) pred A. 2845 P = Pred; 2846 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) && 2847 (A == LHS || B == LHS)) { 2848 if (A != LHS) std::swap(A, B); // A pred umax(A, B). 2849 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2850 // We analyze this as umax(A, B) swapped-pred A. 2851 P = CmpInst::getSwappedPredicate(Pred); 2852 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2853 (A == RHS || B == RHS)) { 2854 if (A != RHS) std::swap(A, B); // umin(A, B) pred A. 2855 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2856 // We analyze this as umax(-A, -B) swapped-pred -A. 2857 // Note that we do not need to actually form -A or -B thanks to EqP. 2858 P = CmpInst::getSwappedPredicate(Pred); 2859 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) && 2860 (A == LHS || B == LHS)) { 2861 if (A != LHS) std::swap(A, B); // A pred umin(A, B). 2862 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2863 // We analyze this as umax(-A, -B) pred -A. 2864 // Note that we do not need to actually form -A or -B thanks to EqP. 2865 P = Pred; 2866 } 2867 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2868 // Cases correspond to "max(A, B) p A". 2869 switch (P) { 2870 default: 2871 break; 2872 case CmpInst::ICMP_EQ: 2873 case CmpInst::ICMP_ULE: 2874 // Equivalent to "A EqP B". This may be the same as the condition tested 2875 // in the max/min; if so, we can just return that. 2876 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2877 return V; 2878 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2879 return V; 2880 // Otherwise, see if "A EqP B" simplifies. 2881 if (MaxRecurse) 2882 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2883 return V; 2884 break; 2885 case CmpInst::ICMP_NE: 2886 case CmpInst::ICMP_UGT: { 2887 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2888 // Equivalent to "A InvEqP B". This may be the same as the condition 2889 // tested in the max/min; if so, we can just return that. 2890 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2891 return V; 2892 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2893 return V; 2894 // Otherwise, see if "A InvEqP B" simplifies. 2895 if (MaxRecurse) 2896 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2897 return V; 2898 break; 2899 } 2900 case CmpInst::ICMP_UGE: 2901 // Always true. 2902 return getTrue(ITy); 2903 case CmpInst::ICMP_ULT: 2904 // Always false. 2905 return getFalse(ITy); 2906 } 2907 } 2908 2909 // Variants on "max(x,y) >= min(x,z)". 2910 Value *C, *D; 2911 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && 2912 match(RHS, m_SMin(m_Value(C), m_Value(D))) && 2913 (A == C || A == D || B == C || B == D)) { 2914 // max(x, ?) pred min(x, ?). 2915 if (Pred == CmpInst::ICMP_SGE) 2916 // Always true. 2917 return getTrue(ITy); 2918 if (Pred == CmpInst::ICMP_SLT) 2919 // Always false. 2920 return getFalse(ITy); 2921 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2922 match(RHS, m_SMax(m_Value(C), m_Value(D))) && 2923 (A == C || A == D || B == C || B == D)) { 2924 // min(x, ?) pred max(x, ?). 2925 if (Pred == CmpInst::ICMP_SLE) 2926 // Always true. 2927 return getTrue(ITy); 2928 if (Pred == CmpInst::ICMP_SGT) 2929 // Always false. 2930 return getFalse(ITy); 2931 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && 2932 match(RHS, m_UMin(m_Value(C), m_Value(D))) && 2933 (A == C || A == D || B == C || B == D)) { 2934 // max(x, ?) pred min(x, ?). 2935 if (Pred == CmpInst::ICMP_UGE) 2936 // Always true. 2937 return getTrue(ITy); 2938 if (Pred == CmpInst::ICMP_ULT) 2939 // Always false. 2940 return getFalse(ITy); 2941 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2942 match(RHS, m_UMax(m_Value(C), m_Value(D))) && 2943 (A == C || A == D || B == C || B == D)) { 2944 // min(x, ?) pred max(x, ?). 2945 if (Pred == CmpInst::ICMP_ULE) 2946 // Always true. 2947 return getTrue(ITy); 2948 if (Pred == CmpInst::ICMP_UGT) 2949 // Always false. 2950 return getFalse(ITy); 2951 } 2952 2953 // Simplify comparisons of related pointers using a powerful, recursive 2954 // GEP-walk when we have target data available.. 2955 if (LHS->getType()->isPointerTy()) 2956 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS)) 2957 return C; 2958 2959 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) { 2960 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) { 2961 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() && 2962 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() && 2963 (ICmpInst::isEquality(Pred) || 2964 (GLHS->isInBounds() && GRHS->isInBounds() && 2965 Pred == ICmpInst::getSignedPredicate(Pred)))) { 2966 // The bases are equal and the indices are constant. Build a constant 2967 // expression GEP with the same indices and a null base pointer to see 2968 // what constant folding can make out of it. 2969 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType()); 2970 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end()); 2971 Constant *NewLHS = ConstantExpr::getGetElementPtr(Null, IndicesLHS); 2972 2973 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end()); 2974 Constant *NewRHS = ConstantExpr::getGetElementPtr(Null, IndicesRHS); 2975 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS); 2976 } 2977 } 2978 } 2979 2980 // If a bit is known to be zero for A and known to be one for B, 2981 // then A and B cannot be equal. 2982 if (ICmpInst::isEquality(Pred)) { 2983 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2984 uint32_t BitWidth = CI->getBitWidth(); 2985 APInt LHSKnownZero(BitWidth, 0); 2986 APInt LHSKnownOne(BitWidth, 0); 2987 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC, 2988 Q.CxtI, Q.DT); 2989 const APInt &RHSVal = CI->getValue(); 2990 if (((LHSKnownZero & RHSVal) != 0) || ((LHSKnownOne & ~RHSVal) != 0)) 2991 return Pred == ICmpInst::ICMP_EQ 2992 ? ConstantInt::getFalse(CI->getContext()) 2993 : ConstantInt::getTrue(CI->getContext()); 2994 } 2995 } 2996 2997 // If the comparison is with the result of a select instruction, check whether 2998 // comparing with either branch of the select always yields the same value. 2999 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3000 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3001 return V; 3002 3003 // If the comparison is with the result of a phi instruction, check whether 3004 // doing the compare with each incoming phi value yields a common result. 3005 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3006 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3007 return V; 3008 3009 return nullptr; 3010 } 3011 3012 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3013 const DataLayout *DL, 3014 const TargetLibraryInfo *TLI, 3015 const DominatorTree *DT, AssumptionCache *AC, 3016 Instruction *CxtI) { 3017 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3018 RecursionLimit); 3019 } 3020 3021 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can 3022 /// fold the result. If not, this returns null. 3023 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3024 const Query &Q, unsigned MaxRecurse) { 3025 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 3026 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!"); 3027 3028 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 3029 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 3030 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 3031 3032 // If we have a constant, make sure it is on the RHS. 3033 std::swap(LHS, RHS); 3034 Pred = CmpInst::getSwappedPredicate(Pred); 3035 } 3036 3037 // Fold trivial predicates. 3038 if (Pred == FCmpInst::FCMP_FALSE) 3039 return ConstantInt::get(GetCompareTy(LHS), 0); 3040 if (Pred == FCmpInst::FCMP_TRUE) 3041 return ConstantInt::get(GetCompareTy(LHS), 1); 3042 3043 if (isa<UndefValue>(RHS)) // fcmp pred X, undef -> undef 3044 return UndefValue::get(GetCompareTy(LHS)); 3045 3046 // fcmp x,x -> true/false. Not all compares are foldable. 3047 if (LHS == RHS) { 3048 if (CmpInst::isTrueWhenEqual(Pred)) 3049 return ConstantInt::get(GetCompareTy(LHS), 1); 3050 if (CmpInst::isFalseWhenEqual(Pred)) 3051 return ConstantInt::get(GetCompareTy(LHS), 0); 3052 } 3053 3054 // Handle fcmp with constant RHS 3055 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 3056 // If the constant is a nan, see if we can fold the comparison based on it. 3057 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) { 3058 if (CFP->getValueAPF().isNaN()) { 3059 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo" 3060 return ConstantInt::getFalse(CFP->getContext()); 3061 assert(FCmpInst::isUnordered(Pred) && 3062 "Comparison must be either ordered or unordered!"); 3063 // True if unordered. 3064 return ConstantInt::getTrue(CFP->getContext()); 3065 } 3066 // Check whether the constant is an infinity. 3067 if (CFP->getValueAPF().isInfinity()) { 3068 if (CFP->getValueAPF().isNegative()) { 3069 switch (Pred) { 3070 case FCmpInst::FCMP_OLT: 3071 // No value is ordered and less than negative infinity. 3072 return ConstantInt::getFalse(CFP->getContext()); 3073 case FCmpInst::FCMP_UGE: 3074 // All values are unordered with or at least negative infinity. 3075 return ConstantInt::getTrue(CFP->getContext()); 3076 default: 3077 break; 3078 } 3079 } else { 3080 switch (Pred) { 3081 case FCmpInst::FCMP_OGT: 3082 // No value is ordered and greater than infinity. 3083 return ConstantInt::getFalse(CFP->getContext()); 3084 case FCmpInst::FCMP_ULE: 3085 // All values are unordered with and at most infinity. 3086 return ConstantInt::getTrue(CFP->getContext()); 3087 default: 3088 break; 3089 } 3090 } 3091 } 3092 if (CFP->getValueAPF().isZero()) { 3093 switch (Pred) { 3094 case FCmpInst::FCMP_UGE: 3095 if (CannotBeOrderedLessThanZero(LHS)) 3096 return ConstantInt::getTrue(CFP->getContext()); 3097 break; 3098 case FCmpInst::FCMP_OLT: 3099 if (CannotBeOrderedLessThanZero(LHS)) 3100 return ConstantInt::getFalse(CFP->getContext()); 3101 break; 3102 default: 3103 break; 3104 } 3105 } 3106 } 3107 } 3108 3109 // If the comparison is with the result of a select instruction, check whether 3110 // comparing with either branch of the select always yields the same value. 3111 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3112 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3113 return V; 3114 3115 // If the comparison is with the result of a phi instruction, check whether 3116 // doing the compare with each incoming phi value yields a common result. 3117 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3118 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3119 return V; 3120 3121 return nullptr; 3122 } 3123 3124 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3125 const DataLayout *DL, 3126 const TargetLibraryInfo *TLI, 3127 const DominatorTree *DT, AssumptionCache *AC, 3128 const Instruction *CxtI) { 3129 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3130 RecursionLimit); 3131 } 3132 3133 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold 3134 /// the result. If not, this returns null. 3135 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal, 3136 Value *FalseVal, const Query &Q, 3137 unsigned MaxRecurse) { 3138 // select true, X, Y -> X 3139 // select false, X, Y -> Y 3140 if (Constant *CB = dyn_cast<Constant>(CondVal)) { 3141 if (CB->isAllOnesValue()) 3142 return TrueVal; 3143 if (CB->isNullValue()) 3144 return FalseVal; 3145 } 3146 3147 // select C, X, X -> X 3148 if (TrueVal == FalseVal) 3149 return TrueVal; 3150 3151 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y 3152 if (isa<Constant>(TrueVal)) 3153 return TrueVal; 3154 return FalseVal; 3155 } 3156 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X 3157 return FalseVal; 3158 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X 3159 return TrueVal; 3160 3161 const auto *ICI = dyn_cast<ICmpInst>(CondVal); 3162 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); 3163 if (ICI && BitWidth) { 3164 ICmpInst::Predicate Pred = ICI->getPredicate(); 3165 APInt MinSignedValue = APInt::getSignBit(BitWidth); 3166 Value *X; 3167 const APInt *Y; 3168 bool TrueWhenUnset; 3169 bool IsBitTest = false; 3170 if (ICmpInst::isEquality(Pred) && 3171 match(ICI->getOperand(0), m_And(m_Value(X), m_APInt(Y))) && 3172 match(ICI->getOperand(1), m_Zero())) { 3173 IsBitTest = true; 3174 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 3175 } else if (Pred == ICmpInst::ICMP_SLT && 3176 match(ICI->getOperand(1), m_Zero())) { 3177 X = ICI->getOperand(0); 3178 Y = &MinSignedValue; 3179 IsBitTest = true; 3180 TrueWhenUnset = false; 3181 } else if (Pred == ICmpInst::ICMP_SGT && 3182 match(ICI->getOperand(1), m_AllOnes())) { 3183 X = ICI->getOperand(0); 3184 Y = &MinSignedValue; 3185 IsBitTest = true; 3186 TrueWhenUnset = true; 3187 } 3188 if (IsBitTest) { 3189 const APInt *C; 3190 // (X & Y) == 0 ? X & ~Y : X --> X 3191 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y 3192 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) && 3193 *Y == ~*C) 3194 return TrueWhenUnset ? FalseVal : TrueVal; 3195 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y 3196 // (X & Y) != 0 ? X : X & ~Y --> X 3197 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) && 3198 *Y == ~*C) 3199 return TrueWhenUnset ? FalseVal : TrueVal; 3200 3201 if (Y->isPowerOf2()) { 3202 // (X & Y) == 0 ? X | Y : X --> X | Y 3203 // (X & Y) != 0 ? X | Y : X --> X 3204 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) && 3205 *Y == *C) 3206 return TrueWhenUnset ? TrueVal : FalseVal; 3207 // (X & Y) == 0 ? X : X | Y --> X 3208 // (X & Y) != 0 ? X : X | Y --> X | Y 3209 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) && 3210 *Y == *C) 3211 return TrueWhenUnset ? TrueVal : FalseVal; 3212 } 3213 } 3214 } 3215 3216 return nullptr; 3217 } 3218 3219 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, 3220 const DataLayout *DL, 3221 const TargetLibraryInfo *TLI, 3222 const DominatorTree *DT, AssumptionCache *AC, 3223 const Instruction *CxtI) { 3224 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, 3225 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3226 } 3227 3228 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can 3229 /// fold the result. If not, this returns null. 3230 static Value *SimplifyGEPInst(ArrayRef<Value *> Ops, const Query &Q, unsigned) { 3231 // The type of the GEP pointer operand. 3232 PointerType *PtrTy = cast<PointerType>(Ops[0]->getType()->getScalarType()); 3233 unsigned AS = PtrTy->getAddressSpace(); 3234 3235 // getelementptr P -> P. 3236 if (Ops.size() == 1) 3237 return Ops[0]; 3238 3239 // Compute the (pointer) type returned by the GEP instruction. 3240 Type *LastType = GetElementPtrInst::getIndexedType(PtrTy, Ops.slice(1)); 3241 Type *GEPTy = PointerType::get(LastType, AS); 3242 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType())) 3243 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 3244 3245 if (isa<UndefValue>(Ops[0])) 3246 return UndefValue::get(GEPTy); 3247 3248 if (Ops.size() == 2) { 3249 // getelementptr P, 0 -> P. 3250 if (match(Ops[1], m_Zero())) 3251 return Ops[0]; 3252 3253 Type *Ty = PtrTy->getElementType(); 3254 if (Q.DL && Ty->isSized()) { 3255 Value *P; 3256 uint64_t C; 3257 uint64_t TyAllocSize = Q.DL->getTypeAllocSize(Ty); 3258 // getelementptr P, N -> P if P points to a type of zero size. 3259 if (TyAllocSize == 0) 3260 return Ops[0]; 3261 3262 // The following transforms are only safe if the ptrtoint cast 3263 // doesn't truncate the pointers. 3264 if (Ops[1]->getType()->getScalarSizeInBits() == 3265 Q.DL->getPointerSizeInBits(AS)) { 3266 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * { 3267 if (match(P, m_Zero())) 3268 return Constant::getNullValue(GEPTy); 3269 Value *Temp; 3270 if (match(P, m_PtrToInt(m_Value(Temp)))) 3271 if (Temp->getType() == GEPTy) 3272 return Temp; 3273 return nullptr; 3274 }; 3275 3276 // getelementptr V, (sub P, V) -> P if P points to a type of size 1. 3277 if (TyAllocSize == 1 && 3278 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))))) 3279 if (Value *R = PtrToIntOrZero(P)) 3280 return R; 3281 3282 // getelementptr V, (ashr (sub P, V), C) -> Q 3283 // if P points to a type of size 1 << C. 3284 if (match(Ops[1], 3285 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3286 m_ConstantInt(C))) && 3287 TyAllocSize == 1ULL << C) 3288 if (Value *R = PtrToIntOrZero(P)) 3289 return R; 3290 3291 // getelementptr V, (sdiv (sub P, V), C) -> Q 3292 // if P points to a type of size C. 3293 if (match(Ops[1], 3294 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3295 m_SpecificInt(TyAllocSize)))) 3296 if (Value *R = PtrToIntOrZero(P)) 3297 return R; 3298 } 3299 } 3300 } 3301 3302 // Check to see if this is constant foldable. 3303 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 3304 if (!isa<Constant>(Ops[i])) 3305 return nullptr; 3306 3307 return ConstantExpr::getGetElementPtr(cast<Constant>(Ops[0]), Ops.slice(1)); 3308 } 3309 3310 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout *DL, 3311 const TargetLibraryInfo *TLI, 3312 const DominatorTree *DT, AssumptionCache *AC, 3313 const Instruction *CxtI) { 3314 return ::SimplifyGEPInst(Ops, Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3315 } 3316 3317 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we 3318 /// can fold the result. If not, this returns null. 3319 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val, 3320 ArrayRef<unsigned> Idxs, const Query &Q, 3321 unsigned) { 3322 if (Constant *CAgg = dyn_cast<Constant>(Agg)) 3323 if (Constant *CVal = dyn_cast<Constant>(Val)) 3324 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs); 3325 3326 // insertvalue x, undef, n -> x 3327 if (match(Val, m_Undef())) 3328 return Agg; 3329 3330 // insertvalue x, (extractvalue y, n), n 3331 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val)) 3332 if (EV->getAggregateOperand()->getType() == Agg->getType() && 3333 EV->getIndices() == Idxs) { 3334 // insertvalue undef, (extractvalue y, n), n -> y 3335 if (match(Agg, m_Undef())) 3336 return EV->getAggregateOperand(); 3337 3338 // insertvalue y, (extractvalue y, n), n -> y 3339 if (Agg == EV->getAggregateOperand()) 3340 return Agg; 3341 } 3342 3343 return nullptr; 3344 } 3345 3346 Value *llvm::SimplifyInsertValueInst( 3347 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout *DL, 3348 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC, 3349 const Instruction *CxtI) { 3350 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI), 3351 RecursionLimit); 3352 } 3353 3354 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null. 3355 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) { 3356 // If all of the PHI's incoming values are the same then replace the PHI node 3357 // with the common value. 3358 Value *CommonValue = nullptr; 3359 bool HasUndefInput = false; 3360 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 3361 Value *Incoming = PN->getIncomingValue(i); 3362 // If the incoming value is the phi node itself, it can safely be skipped. 3363 if (Incoming == PN) continue; 3364 if (isa<UndefValue>(Incoming)) { 3365 // Remember that we saw an undef value, but otherwise ignore them. 3366 HasUndefInput = true; 3367 continue; 3368 } 3369 if (CommonValue && Incoming != CommonValue) 3370 return nullptr; // Not the same, bail out. 3371 CommonValue = Incoming; 3372 } 3373 3374 // If CommonValue is null then all of the incoming values were either undef or 3375 // equal to the phi node itself. 3376 if (!CommonValue) 3377 return UndefValue::get(PN->getType()); 3378 3379 // If we have a PHI node like phi(X, undef, X), where X is defined by some 3380 // instruction, we cannot return X as the result of the PHI node unless it 3381 // dominates the PHI block. 3382 if (HasUndefInput) 3383 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr; 3384 3385 return CommonValue; 3386 } 3387 3388 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) { 3389 if (Constant *C = dyn_cast<Constant>(Op)) 3390 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI); 3391 3392 return nullptr; 3393 } 3394 3395 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout *DL, 3396 const TargetLibraryInfo *TLI, 3397 const DominatorTree *DT, AssumptionCache *AC, 3398 const Instruction *CxtI) { 3399 return ::SimplifyTruncInst(Op, Ty, Query(DL, TLI, DT, AC, CxtI), 3400 RecursionLimit); 3401 } 3402 3403 //=== Helper functions for higher up the class hierarchy. 3404 3405 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can 3406 /// fold the result. If not, this returns null. 3407 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3408 const Query &Q, unsigned MaxRecurse) { 3409 switch (Opcode) { 3410 case Instruction::Add: 3411 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3412 Q, MaxRecurse); 3413 case Instruction::FAdd: 3414 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3415 3416 case Instruction::Sub: 3417 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3418 Q, MaxRecurse); 3419 case Instruction::FSub: 3420 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3421 3422 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse); 3423 case Instruction::FMul: 3424 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3425 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse); 3426 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse); 3427 case Instruction::FDiv: return SimplifyFDivInst(LHS, RHS, Q, MaxRecurse); 3428 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse); 3429 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse); 3430 case Instruction::FRem: return SimplifyFRemInst(LHS, RHS, Q, MaxRecurse); 3431 case Instruction::Shl: 3432 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3433 Q, MaxRecurse); 3434 case Instruction::LShr: 3435 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3436 case Instruction::AShr: 3437 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3438 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse); 3439 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse); 3440 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse); 3441 default: 3442 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 3443 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 3444 Constant *COps[] = {CLHS, CRHS}; 3445 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL, 3446 Q.TLI); 3447 } 3448 3449 // If the operation is associative, try some generic simplifications. 3450 if (Instruction::isAssociative(Opcode)) 3451 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse)) 3452 return V; 3453 3454 // If the operation is with the result of a select instruction check whether 3455 // operating on either branch of the select always yields the same value. 3456 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3457 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse)) 3458 return V; 3459 3460 // If the operation is with the result of a phi instruction, check whether 3461 // operating on all incoming values of the phi always yields the same value. 3462 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3463 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse)) 3464 return V; 3465 3466 return nullptr; 3467 } 3468 } 3469 3470 /// SimplifyFPBinOp - Given operands for a BinaryOperator, see if we can 3471 /// fold the result. If not, this returns null. 3472 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the 3473 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp. 3474 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3475 const FastMathFlags &FMF, const Query &Q, 3476 unsigned MaxRecurse) { 3477 switch (Opcode) { 3478 case Instruction::FAdd: 3479 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse); 3480 case Instruction::FSub: 3481 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse); 3482 case Instruction::FMul: 3483 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse); 3484 default: 3485 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse); 3486 } 3487 } 3488 3489 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3490 const DataLayout *DL, const TargetLibraryInfo *TLI, 3491 const DominatorTree *DT, AssumptionCache *AC, 3492 const Instruction *CxtI) { 3493 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3494 RecursionLimit); 3495 } 3496 3497 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3498 const FastMathFlags &FMF, const DataLayout *DL, 3499 const TargetLibraryInfo *TLI, 3500 const DominatorTree *DT, AssumptionCache *AC, 3501 const Instruction *CxtI) { 3502 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI), 3503 RecursionLimit); 3504 } 3505 3506 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can 3507 /// fold the result. 3508 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3509 const Query &Q, unsigned MaxRecurse) { 3510 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate)) 3511 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 3512 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 3513 } 3514 3515 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3516 const DataLayout *DL, const TargetLibraryInfo *TLI, 3517 const DominatorTree *DT, AssumptionCache *AC, 3518 const Instruction *CxtI) { 3519 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3520 RecursionLimit); 3521 } 3522 3523 static bool IsIdempotent(Intrinsic::ID ID) { 3524 switch (ID) { 3525 default: return false; 3526 3527 // Unary idempotent: f(f(x)) = f(x) 3528 case Intrinsic::fabs: 3529 case Intrinsic::floor: 3530 case Intrinsic::ceil: 3531 case Intrinsic::trunc: 3532 case Intrinsic::rint: 3533 case Intrinsic::nearbyint: 3534 case Intrinsic::round: 3535 return true; 3536 } 3537 } 3538 3539 template <typename IterTy> 3540 static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd, 3541 const Query &Q, unsigned MaxRecurse) { 3542 // Perform idempotent optimizations 3543 if (!IsIdempotent(IID)) 3544 return nullptr; 3545 3546 // Unary Ops 3547 if (std::distance(ArgBegin, ArgEnd) == 1) 3548 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) 3549 if (II->getIntrinsicID() == IID) 3550 return II; 3551 3552 return nullptr; 3553 } 3554 3555 template <typename IterTy> 3556 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd, 3557 const Query &Q, unsigned MaxRecurse) { 3558 Type *Ty = V->getType(); 3559 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 3560 Ty = PTy->getElementType(); 3561 FunctionType *FTy = cast<FunctionType>(Ty); 3562 3563 // call undef -> undef 3564 if (isa<UndefValue>(V)) 3565 return UndefValue::get(FTy->getReturnType()); 3566 3567 Function *F = dyn_cast<Function>(V); 3568 if (!F) 3569 return nullptr; 3570 3571 if (unsigned IID = F->getIntrinsicID()) 3572 if (Value *Ret = 3573 SimplifyIntrinsic((Intrinsic::ID) IID, ArgBegin, ArgEnd, Q, MaxRecurse)) 3574 return Ret; 3575 3576 if (!canConstantFoldCallTo(F)) 3577 return nullptr; 3578 3579 SmallVector<Constant *, 4> ConstantArgs; 3580 ConstantArgs.reserve(ArgEnd - ArgBegin); 3581 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) { 3582 Constant *C = dyn_cast<Constant>(*I); 3583 if (!C) 3584 return nullptr; 3585 ConstantArgs.push_back(C); 3586 } 3587 3588 return ConstantFoldCall(F, ConstantArgs, Q.TLI); 3589 } 3590 3591 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin, 3592 User::op_iterator ArgEnd, const DataLayout *DL, 3593 const TargetLibraryInfo *TLI, const DominatorTree *DT, 3594 AssumptionCache *AC, const Instruction *CxtI) { 3595 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI), 3596 RecursionLimit); 3597 } 3598 3599 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args, 3600 const DataLayout *DL, const TargetLibraryInfo *TLI, 3601 const DominatorTree *DT, AssumptionCache *AC, 3602 const Instruction *CxtI) { 3603 return ::SimplifyCall(V, Args.begin(), Args.end(), 3604 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3605 } 3606 3607 /// SimplifyInstruction - See if we can compute a simplified version of this 3608 /// instruction. If not, this returns null. 3609 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout *DL, 3610 const TargetLibraryInfo *TLI, 3611 const DominatorTree *DT, AssumptionCache *AC) { 3612 Value *Result; 3613 3614 switch (I->getOpcode()) { 3615 default: 3616 Result = ConstantFoldInstruction(I, DL, TLI); 3617 break; 3618 case Instruction::FAdd: 3619 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1), 3620 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3621 break; 3622 case Instruction::Add: 3623 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1), 3624 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3625 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3626 TLI, DT, AC, I); 3627 break; 3628 case Instruction::FSub: 3629 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1), 3630 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3631 break; 3632 case Instruction::Sub: 3633 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1), 3634 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3635 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3636 TLI, DT, AC, I); 3637 break; 3638 case Instruction::FMul: 3639 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1), 3640 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3641 break; 3642 case Instruction::Mul: 3643 Result = 3644 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3645 break; 3646 case Instruction::SDiv: 3647 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3648 AC, I); 3649 break; 3650 case Instruction::UDiv: 3651 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3652 AC, I); 3653 break; 3654 case Instruction::FDiv: 3655 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3656 AC, I); 3657 break; 3658 case Instruction::SRem: 3659 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3660 AC, I); 3661 break; 3662 case Instruction::URem: 3663 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3664 AC, I); 3665 break; 3666 case Instruction::FRem: 3667 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3668 AC, I); 3669 break; 3670 case Instruction::Shl: 3671 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1), 3672 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3673 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3674 TLI, DT, AC, I); 3675 break; 3676 case Instruction::LShr: 3677 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1), 3678 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 3679 AC, I); 3680 break; 3681 case Instruction::AShr: 3682 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1), 3683 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 3684 AC, I); 3685 break; 3686 case Instruction::And: 3687 Result = 3688 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3689 break; 3690 case Instruction::Or: 3691 Result = 3692 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3693 break; 3694 case Instruction::Xor: 3695 Result = 3696 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3697 break; 3698 case Instruction::ICmp: 3699 Result = 3700 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0), 3701 I->getOperand(1), DL, TLI, DT, AC, I); 3702 break; 3703 case Instruction::FCmp: 3704 Result = 3705 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0), 3706 I->getOperand(1), DL, TLI, DT, AC, I); 3707 break; 3708 case Instruction::Select: 3709 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1), 3710 I->getOperand(2), DL, TLI, DT, AC, I); 3711 break; 3712 case Instruction::GetElementPtr: { 3713 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 3714 Result = SimplifyGEPInst(Ops, DL, TLI, DT, AC, I); 3715 break; 3716 } 3717 case Instruction::InsertValue: { 3718 InsertValueInst *IV = cast<InsertValueInst>(I); 3719 Result = SimplifyInsertValueInst(IV->getAggregateOperand(), 3720 IV->getInsertedValueOperand(), 3721 IV->getIndices(), DL, TLI, DT, AC, I); 3722 break; 3723 } 3724 case Instruction::PHI: 3725 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I)); 3726 break; 3727 case Instruction::Call: { 3728 CallSite CS(cast<CallInst>(I)); 3729 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL, 3730 TLI, DT, AC, I); 3731 break; 3732 } 3733 case Instruction::Trunc: 3734 Result = 3735 SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT, AC, I); 3736 break; 3737 } 3738 3739 /// If called on unreachable code, the above logic may report that the 3740 /// instruction simplified to itself. Make life easier for users by 3741 /// detecting that case here, returning a safe value instead. 3742 return Result == I ? UndefValue::get(I->getType()) : Result; 3743 } 3744 3745 /// \brief Implementation of recursive simplification through an instructions 3746 /// uses. 3747 /// 3748 /// This is the common implementation of the recursive simplification routines. 3749 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to 3750 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of 3751 /// instructions to process and attempt to simplify it using 3752 /// InstructionSimplify. 3753 /// 3754 /// This routine returns 'true' only when *it* simplifies something. The passed 3755 /// in simplified value does not count toward this. 3756 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV, 3757 const DataLayout *DL, 3758 const TargetLibraryInfo *TLI, 3759 const DominatorTree *DT, 3760 AssumptionCache *AC) { 3761 bool Simplified = false; 3762 SmallSetVector<Instruction *, 8> Worklist; 3763 3764 // If we have an explicit value to collapse to, do that round of the 3765 // simplification loop by hand initially. 3766 if (SimpleV) { 3767 for (User *U : I->users()) 3768 if (U != I) 3769 Worklist.insert(cast<Instruction>(U)); 3770 3771 // Replace the instruction with its simplified value. 3772 I->replaceAllUsesWith(SimpleV); 3773 3774 // Gracefully handle edge cases where the instruction is not wired into any 3775 // parent block. 3776 if (I->getParent()) 3777 I->eraseFromParent(); 3778 } else { 3779 Worklist.insert(I); 3780 } 3781 3782 // Note that we must test the size on each iteration, the worklist can grow. 3783 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { 3784 I = Worklist[Idx]; 3785 3786 // See if this instruction simplifies. 3787 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC); 3788 if (!SimpleV) 3789 continue; 3790 3791 Simplified = true; 3792 3793 // Stash away all the uses of the old instruction so we can check them for 3794 // recursive simplifications after a RAUW. This is cheaper than checking all 3795 // uses of To on the recursive step in most cases. 3796 for (User *U : I->users()) 3797 Worklist.insert(cast<Instruction>(U)); 3798 3799 // Replace the instruction with its simplified value. 3800 I->replaceAllUsesWith(SimpleV); 3801 3802 // Gracefully handle edge cases where the instruction is not wired into any 3803 // parent block. 3804 if (I->getParent()) 3805 I->eraseFromParent(); 3806 } 3807 return Simplified; 3808 } 3809 3810 bool llvm::recursivelySimplifyInstruction(Instruction *I, const DataLayout *DL, 3811 const TargetLibraryInfo *TLI, 3812 const DominatorTree *DT, 3813 AssumptionCache *AC) { 3814 return replaceAndRecursivelySimplifyImpl(I, nullptr, DL, TLI, DT, AC); 3815 } 3816 3817 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV, 3818 const DataLayout *DL, 3819 const TargetLibraryInfo *TLI, 3820 const DominatorTree *DT, 3821 AssumptionCache *AC) { 3822 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!"); 3823 assert(SimpleV && "Must provide a simplified value."); 3824 return replaceAndRecursivelySimplifyImpl(I, SimpleV, DL, TLI, DT, AC); 3825 } 3826