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, FastMathFlags FMF, 1122 const Query &Q, 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 // 0 / X -> 0 1132 // Requires that NaNs are off (X could be zero) and signed zeroes are 1133 // ignored (X could be positive or negative, so the output sign is unknown). 1134 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1135 return Op0; 1136 1137 return nullptr; 1138 } 1139 1140 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1141 const DataLayout *DL, 1142 const TargetLibraryInfo *TLI, 1143 const DominatorTree *DT, AssumptionCache *AC, 1144 const Instruction *CxtI) { 1145 return ::SimplifyFDivInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 1146 RecursionLimit); 1147 } 1148 1149 /// SimplifyRem - Given operands for an SRem or URem, see if we can 1150 /// fold the result. If not, this returns null. 1151 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1152 const Query &Q, unsigned MaxRecurse) { 1153 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1154 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1155 Constant *Ops[] = { C0, C1 }; 1156 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1157 } 1158 } 1159 1160 // X % undef -> undef 1161 if (match(Op1, m_Undef())) 1162 return Op1; 1163 1164 // undef % X -> 0 1165 if (match(Op0, m_Undef())) 1166 return Constant::getNullValue(Op0->getType()); 1167 1168 // 0 % X -> 0, we don't need to preserve faults! 1169 if (match(Op0, m_Zero())) 1170 return Op0; 1171 1172 // X % 0 -> undef, we don't need to preserve faults! 1173 if (match(Op1, m_Zero())) 1174 return UndefValue::get(Op0->getType()); 1175 1176 // X % 1 -> 0 1177 if (match(Op1, m_One())) 1178 return Constant::getNullValue(Op0->getType()); 1179 1180 if (Op0->getType()->isIntegerTy(1)) 1181 // It can't be remainder by zero, hence it must be remainder by one. 1182 return Constant::getNullValue(Op0->getType()); 1183 1184 // X % X -> 0 1185 if (Op0 == Op1) 1186 return Constant::getNullValue(Op0->getType()); 1187 1188 // (X % Y) % Y -> X % Y 1189 if ((Opcode == Instruction::SRem && 1190 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1191 (Opcode == Instruction::URem && 1192 match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1193 return Op0; 1194 1195 // If the operation is with the result of a select instruction, check whether 1196 // operating on either branch of the select always yields the same value. 1197 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1198 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1199 return V; 1200 1201 // If the operation is with the result of a phi instruction, check whether 1202 // operating on all incoming values of the phi always yields the same value. 1203 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1204 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1205 return V; 1206 1207 return nullptr; 1208 } 1209 1210 /// SimplifySRemInst - Given operands for an SRem, see if we can 1211 /// fold the result. If not, this returns null. 1212 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q, 1213 unsigned MaxRecurse) { 1214 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse)) 1215 return V; 1216 1217 return nullptr; 1218 } 1219 1220 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout *DL, 1221 const TargetLibraryInfo *TLI, 1222 const DominatorTree *DT, AssumptionCache *AC, 1223 const Instruction *CxtI) { 1224 return ::SimplifySRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1225 RecursionLimit); 1226 } 1227 1228 /// SimplifyURemInst - Given operands for a URem, see if we can 1229 /// fold the result. If not, this returns null. 1230 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q, 1231 unsigned MaxRecurse) { 1232 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse)) 1233 return V; 1234 1235 return nullptr; 1236 } 1237 1238 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout *DL, 1239 const TargetLibraryInfo *TLI, 1240 const DominatorTree *DT, AssumptionCache *AC, 1241 const Instruction *CxtI) { 1242 return ::SimplifyURemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1243 RecursionLimit); 1244 } 1245 1246 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1247 const Query &, unsigned) { 1248 // undef % X -> undef (the undef could be a snan). 1249 if (match(Op0, m_Undef())) 1250 return Op0; 1251 1252 // X % undef -> undef 1253 if (match(Op1, m_Undef())) 1254 return Op1; 1255 1256 // 0 % X -> 0 1257 // Requires that NaNs are off (X could be zero) and signed zeroes are 1258 // ignored (X could be positive or negative, so the output sign is unknown). 1259 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1260 return Op0; 1261 1262 return nullptr; 1263 } 1264 1265 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1266 const DataLayout *DL, 1267 const TargetLibraryInfo *TLI, 1268 const DominatorTree *DT, AssumptionCache *AC, 1269 const Instruction *CxtI) { 1270 return ::SimplifyFRemInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 1271 RecursionLimit); 1272 } 1273 1274 /// isUndefShift - Returns true if a shift by \c Amount always yields undef. 1275 static bool isUndefShift(Value *Amount) { 1276 Constant *C = dyn_cast<Constant>(Amount); 1277 if (!C) 1278 return false; 1279 1280 // X shift by undef -> undef because it may shift by the bitwidth. 1281 if (isa<UndefValue>(C)) 1282 return true; 1283 1284 // Shifting by the bitwidth or more is undefined. 1285 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) 1286 if (CI->getValue().getLimitedValue() >= 1287 CI->getType()->getScalarSizeInBits()) 1288 return true; 1289 1290 // If all lanes of a vector shift are undefined the whole shift is. 1291 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) { 1292 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I) 1293 if (!isUndefShift(C->getAggregateElement(I))) 1294 return false; 1295 return true; 1296 } 1297 1298 return false; 1299 } 1300 1301 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can 1302 /// fold the result. If not, this returns null. 1303 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1, 1304 const Query &Q, unsigned MaxRecurse) { 1305 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1306 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1307 Constant *Ops[] = { C0, C1 }; 1308 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1309 } 1310 } 1311 1312 // 0 shift by X -> 0 1313 if (match(Op0, m_Zero())) 1314 return Op0; 1315 1316 // X shift by 0 -> X 1317 if (match(Op1, m_Zero())) 1318 return Op0; 1319 1320 // Fold undefined shifts. 1321 if (isUndefShift(Op1)) 1322 return UndefValue::get(Op0->getType()); 1323 1324 // If the operation is with the result of a select instruction, check whether 1325 // operating on either branch of the select always yields the same value. 1326 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1327 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1328 return V; 1329 1330 // If the operation is with the result of a phi instruction, check whether 1331 // operating on all incoming values of the phi always yields the same value. 1332 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1333 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1334 return V; 1335 1336 return nullptr; 1337 } 1338 1339 /// \brief Given operands for an Shl, LShr or AShr, see if we can 1340 /// fold the result. If not, this returns null. 1341 static Value *SimplifyRightShift(unsigned Opcode, Value *Op0, Value *Op1, 1342 bool isExact, const Query &Q, 1343 unsigned MaxRecurse) { 1344 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse)) 1345 return V; 1346 1347 // X >> X -> 0 1348 if (Op0 == Op1) 1349 return Constant::getNullValue(Op0->getType()); 1350 1351 // undef >> X -> 0 1352 // undef >> X -> undef (if it's exact) 1353 if (match(Op0, m_Undef())) 1354 return isExact ? Op0 : Constant::getNullValue(Op0->getType()); 1355 1356 // The low bit cannot be shifted out of an exact shift if it is set. 1357 if (isExact) { 1358 unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); 1359 APInt Op0KnownZero(BitWidth, 0); 1360 APInt Op0KnownOne(BitWidth, 0); 1361 computeKnownBits(Op0, Op0KnownZero, Op0KnownOne, Q.DL, /*Depth=*/0, Q.AC, 1362 Q.CxtI, Q.DT); 1363 if (Op0KnownOne[0]) 1364 return Op0; 1365 } 1366 1367 return nullptr; 1368 } 1369 1370 /// SimplifyShlInst - Given operands for an Shl, see if we can 1371 /// fold the result. If not, this returns null. 1372 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1373 const Query &Q, unsigned MaxRecurse) { 1374 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse)) 1375 return V; 1376 1377 // undef << X -> 0 1378 // undef << X -> undef if (if it's NSW/NUW) 1379 if (match(Op0, m_Undef())) 1380 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType()); 1381 1382 // (X >> A) << A -> X 1383 Value *X; 1384 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1))))) 1385 return X; 1386 return nullptr; 1387 } 1388 1389 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1390 const DataLayout *DL, const TargetLibraryInfo *TLI, 1391 const DominatorTree *DT, AssumptionCache *AC, 1392 const Instruction *CxtI) { 1393 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 1394 RecursionLimit); 1395 } 1396 1397 /// SimplifyLShrInst - Given operands for an LShr, see if we can 1398 /// fold the result. If not, this returns null. 1399 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1400 const Query &Q, unsigned MaxRecurse) { 1401 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q, 1402 MaxRecurse)) 1403 return V; 1404 1405 // (X << A) >> A -> X 1406 Value *X; 1407 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1)))) 1408 return X; 1409 1410 return nullptr; 1411 } 1412 1413 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1414 const DataLayout *DL, 1415 const TargetLibraryInfo *TLI, 1416 const DominatorTree *DT, AssumptionCache *AC, 1417 const Instruction *CxtI) { 1418 return ::SimplifyLShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1419 RecursionLimit); 1420 } 1421 1422 /// SimplifyAShrInst - Given operands for an AShr, see if we can 1423 /// fold the result. If not, this returns null. 1424 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1425 const Query &Q, unsigned MaxRecurse) { 1426 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q, 1427 MaxRecurse)) 1428 return V; 1429 1430 // all ones >>a X -> all ones 1431 if (match(Op0, m_AllOnes())) 1432 return Op0; 1433 1434 // (X << A) >> A -> X 1435 Value *X; 1436 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1)))) 1437 return X; 1438 1439 // Arithmetic shifting an all-sign-bit value is a no-op. 1440 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 1441 if (NumSignBits == Op0->getType()->getScalarSizeInBits()) 1442 return Op0; 1443 1444 return nullptr; 1445 } 1446 1447 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1448 const DataLayout *DL, 1449 const TargetLibraryInfo *TLI, 1450 const DominatorTree *DT, AssumptionCache *AC, 1451 const Instruction *CxtI) { 1452 return ::SimplifyAShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1453 RecursionLimit); 1454 } 1455 1456 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp, 1457 ICmpInst *UnsignedICmp, bool IsAnd) { 1458 Value *X, *Y; 1459 1460 ICmpInst::Predicate EqPred; 1461 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) || 1462 !ICmpInst::isEquality(EqPred)) 1463 return nullptr; 1464 1465 ICmpInst::Predicate UnsignedPred; 1466 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) && 1467 ICmpInst::isUnsigned(UnsignedPred)) 1468 ; 1469 else if (match(UnsignedICmp, 1470 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) && 1471 ICmpInst::isUnsigned(UnsignedPred)) 1472 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred); 1473 else 1474 return nullptr; 1475 1476 // X < Y && Y != 0 --> X < Y 1477 // X < Y || Y != 0 --> Y != 0 1478 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE) 1479 return IsAnd ? UnsignedICmp : ZeroICmp; 1480 1481 // X >= Y || Y != 0 --> true 1482 // X >= Y || Y == 0 --> X >= Y 1483 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) { 1484 if (EqPred == ICmpInst::ICMP_NE) 1485 return getTrue(UnsignedICmp->getType()); 1486 return UnsignedICmp; 1487 } 1488 1489 // X < Y && Y == 0 --> false 1490 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ && 1491 IsAnd) 1492 return getFalse(UnsignedICmp->getType()); 1493 1494 return nullptr; 1495 } 1496 1497 // Simplify (and (icmp ...) (icmp ...)) to true when we can tell that the range 1498 // of possible values cannot be satisfied. 1499 static Value *SimplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1500 ICmpInst::Predicate Pred0, Pred1; 1501 ConstantInt *CI1, *CI2; 1502 Value *V; 1503 1504 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true)) 1505 return X; 1506 1507 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)), 1508 m_ConstantInt(CI2)))) 1509 return nullptr; 1510 1511 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1)))) 1512 return nullptr; 1513 1514 Type *ITy = Op0->getType(); 1515 1516 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1517 bool isNSW = AddInst->hasNoSignedWrap(); 1518 bool isNUW = AddInst->hasNoUnsignedWrap(); 1519 1520 const APInt &CI1V = CI1->getValue(); 1521 const APInt &CI2V = CI2->getValue(); 1522 const APInt Delta = CI2V - CI1V; 1523 if (CI1V.isStrictlyPositive()) { 1524 if (Delta == 2) { 1525 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT) 1526 return getFalse(ITy); 1527 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1528 return getFalse(ITy); 1529 } 1530 if (Delta == 1) { 1531 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT) 1532 return getFalse(ITy); 1533 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1534 return getFalse(ITy); 1535 } 1536 } 1537 if (CI1V.getBoolValue() && isNUW) { 1538 if (Delta == 2) 1539 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT) 1540 return getFalse(ITy); 1541 if (Delta == 1) 1542 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT) 1543 return getFalse(ITy); 1544 } 1545 1546 return nullptr; 1547 } 1548 1549 /// SimplifyAndInst - Given operands for an And, see if we can 1550 /// fold the result. If not, this returns null. 1551 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q, 1552 unsigned MaxRecurse) { 1553 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1554 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1555 Constant *Ops[] = { CLHS, CRHS }; 1556 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(), 1557 Ops, Q.DL, Q.TLI); 1558 } 1559 1560 // Canonicalize the constant to the RHS. 1561 std::swap(Op0, Op1); 1562 } 1563 1564 // X & undef -> 0 1565 if (match(Op1, m_Undef())) 1566 return Constant::getNullValue(Op0->getType()); 1567 1568 // X & X = X 1569 if (Op0 == Op1) 1570 return Op0; 1571 1572 // X & 0 = 0 1573 if (match(Op1, m_Zero())) 1574 return Op1; 1575 1576 // X & -1 = X 1577 if (match(Op1, m_AllOnes())) 1578 return Op0; 1579 1580 // A & ~A = ~A & A = 0 1581 if (match(Op0, m_Not(m_Specific(Op1))) || 1582 match(Op1, m_Not(m_Specific(Op0)))) 1583 return Constant::getNullValue(Op0->getType()); 1584 1585 // (A | ?) & A = A 1586 Value *A = nullptr, *B = nullptr; 1587 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1588 (A == Op1 || B == Op1)) 1589 return Op1; 1590 1591 // A & (A | ?) = A 1592 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1593 (A == Op0 || B == Op0)) 1594 return Op0; 1595 1596 // A & (-A) = A if A is a power of two or zero. 1597 if (match(Op0, m_Neg(m_Specific(Op1))) || 1598 match(Op1, m_Neg(m_Specific(Op0)))) { 1599 if (isKnownToBeAPowerOfTwo(Op0, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT)) 1600 return Op0; 1601 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT)) 1602 return Op1; 1603 } 1604 1605 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1606 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1607 if (Value *V = SimplifyAndOfICmps(ICILHS, ICIRHS)) 1608 return V; 1609 if (Value *V = SimplifyAndOfICmps(ICIRHS, ICILHS)) 1610 return V; 1611 } 1612 } 1613 1614 // Try some generic simplifications for associative operations. 1615 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q, 1616 MaxRecurse)) 1617 return V; 1618 1619 // And distributes over Or. Try some generic simplifications based on this. 1620 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or, 1621 Q, MaxRecurse)) 1622 return V; 1623 1624 // And distributes over Xor. Try some generic simplifications based on this. 1625 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor, 1626 Q, MaxRecurse)) 1627 return V; 1628 1629 // If the operation is with the result of a select instruction, check whether 1630 // operating on either branch of the select always yields the same value. 1631 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1632 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q, 1633 MaxRecurse)) 1634 return V; 1635 1636 // If the operation is with the result of a phi instruction, check whether 1637 // operating on all incoming values of the phi always yields the same value. 1638 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1639 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q, 1640 MaxRecurse)) 1641 return V; 1642 1643 return nullptr; 1644 } 1645 1646 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout *DL, 1647 const TargetLibraryInfo *TLI, 1648 const DominatorTree *DT, AssumptionCache *AC, 1649 const Instruction *CxtI) { 1650 return ::SimplifyAndInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1651 RecursionLimit); 1652 } 1653 1654 // Simplify (or (icmp ...) (icmp ...)) to true when we can tell that the union 1655 // contains all possible values. 1656 static Value *SimplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1657 ICmpInst::Predicate Pred0, Pred1; 1658 ConstantInt *CI1, *CI2; 1659 Value *V; 1660 1661 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false)) 1662 return X; 1663 1664 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)), 1665 m_ConstantInt(CI2)))) 1666 return nullptr; 1667 1668 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1)))) 1669 return nullptr; 1670 1671 Type *ITy = Op0->getType(); 1672 1673 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1674 bool isNSW = AddInst->hasNoSignedWrap(); 1675 bool isNUW = AddInst->hasNoUnsignedWrap(); 1676 1677 const APInt &CI1V = CI1->getValue(); 1678 const APInt &CI2V = CI2->getValue(); 1679 const APInt Delta = CI2V - CI1V; 1680 if (CI1V.isStrictlyPositive()) { 1681 if (Delta == 2) { 1682 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE) 1683 return getTrue(ITy); 1684 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1685 return getTrue(ITy); 1686 } 1687 if (Delta == 1) { 1688 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE) 1689 return getTrue(ITy); 1690 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1691 return getTrue(ITy); 1692 } 1693 } 1694 if (CI1V.getBoolValue() && isNUW) { 1695 if (Delta == 2) 1696 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE) 1697 return getTrue(ITy); 1698 if (Delta == 1) 1699 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE) 1700 return getTrue(ITy); 1701 } 1702 1703 return nullptr; 1704 } 1705 1706 /// SimplifyOrInst - Given operands for an Or, see if we can 1707 /// fold the result. If not, this returns null. 1708 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q, 1709 unsigned MaxRecurse) { 1710 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1711 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1712 Constant *Ops[] = { CLHS, CRHS }; 1713 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(), 1714 Ops, Q.DL, Q.TLI); 1715 } 1716 1717 // Canonicalize the constant to the RHS. 1718 std::swap(Op0, Op1); 1719 } 1720 1721 // X | undef -> -1 1722 if (match(Op1, m_Undef())) 1723 return Constant::getAllOnesValue(Op0->getType()); 1724 1725 // X | X = X 1726 if (Op0 == Op1) 1727 return Op0; 1728 1729 // X | 0 = X 1730 if (match(Op1, m_Zero())) 1731 return Op0; 1732 1733 // X | -1 = -1 1734 if (match(Op1, m_AllOnes())) 1735 return Op1; 1736 1737 // A | ~A = ~A | A = -1 1738 if (match(Op0, m_Not(m_Specific(Op1))) || 1739 match(Op1, m_Not(m_Specific(Op0)))) 1740 return Constant::getAllOnesValue(Op0->getType()); 1741 1742 // (A & ?) | A = A 1743 Value *A = nullptr, *B = nullptr; 1744 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 1745 (A == Op1 || B == Op1)) 1746 return Op1; 1747 1748 // A | (A & ?) = A 1749 if (match(Op1, m_And(m_Value(A), m_Value(B))) && 1750 (A == Op0 || B == Op0)) 1751 return Op0; 1752 1753 // ~(A & ?) | A = -1 1754 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) && 1755 (A == Op1 || B == Op1)) 1756 return Constant::getAllOnesValue(Op1->getType()); 1757 1758 // A | ~(A & ?) = -1 1759 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) && 1760 (A == Op0 || B == Op0)) 1761 return Constant::getAllOnesValue(Op0->getType()); 1762 1763 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1764 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1765 if (Value *V = SimplifyOrOfICmps(ICILHS, ICIRHS)) 1766 return V; 1767 if (Value *V = SimplifyOrOfICmps(ICIRHS, ICILHS)) 1768 return V; 1769 } 1770 } 1771 1772 // Try some generic simplifications for associative operations. 1773 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q, 1774 MaxRecurse)) 1775 return V; 1776 1777 // Or distributes over And. Try some generic simplifications based on this. 1778 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q, 1779 MaxRecurse)) 1780 return V; 1781 1782 // If the operation is with the result of a select instruction, check whether 1783 // operating on either branch of the select always yields the same value. 1784 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1785 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q, 1786 MaxRecurse)) 1787 return V; 1788 1789 // (A & C)|(B & D) 1790 Value *C = nullptr, *D = nullptr; 1791 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 1792 match(Op1, m_And(m_Value(B), m_Value(D)))) { 1793 ConstantInt *C1 = dyn_cast<ConstantInt>(C); 1794 ConstantInt *C2 = dyn_cast<ConstantInt>(D); 1795 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) { 1796 // (A & C1)|(B & C2) 1797 // If we have: ((V + N) & C1) | (V & C2) 1798 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0 1799 // replace with V+N. 1800 Value *V1, *V2; 1801 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+ 1802 match(A, m_Add(m_Value(V1), m_Value(V2)))) { 1803 // Add commutes, try both ways. 1804 if (V1 == B && 1805 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1806 return A; 1807 if (V2 == B && 1808 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1809 return A; 1810 } 1811 // Or commutes, try both ways. 1812 if ((C1->getValue() & (C1->getValue() + 1)) == 0 && 1813 match(B, m_Add(m_Value(V1), m_Value(V2)))) { 1814 // Add commutes, try both ways. 1815 if (V1 == A && 1816 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1817 return B; 1818 if (V2 == A && 1819 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1820 return B; 1821 } 1822 } 1823 } 1824 1825 // If the operation is with the result of a phi instruction, check whether 1826 // operating on all incoming values of the phi always yields the same value. 1827 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1828 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse)) 1829 return V; 1830 1831 return nullptr; 1832 } 1833 1834 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout *DL, 1835 const TargetLibraryInfo *TLI, 1836 const DominatorTree *DT, AssumptionCache *AC, 1837 const Instruction *CxtI) { 1838 return ::SimplifyOrInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1839 RecursionLimit); 1840 } 1841 1842 /// SimplifyXorInst - Given operands for a Xor, see if we can 1843 /// fold the result. If not, this returns null. 1844 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q, 1845 unsigned MaxRecurse) { 1846 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1847 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1848 Constant *Ops[] = { CLHS, CRHS }; 1849 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(), 1850 Ops, Q.DL, Q.TLI); 1851 } 1852 1853 // Canonicalize the constant to the RHS. 1854 std::swap(Op0, Op1); 1855 } 1856 1857 // A ^ undef -> undef 1858 if (match(Op1, m_Undef())) 1859 return Op1; 1860 1861 // A ^ 0 = A 1862 if (match(Op1, m_Zero())) 1863 return Op0; 1864 1865 // A ^ A = 0 1866 if (Op0 == Op1) 1867 return Constant::getNullValue(Op0->getType()); 1868 1869 // A ^ ~A = ~A ^ A = -1 1870 if (match(Op0, m_Not(m_Specific(Op1))) || 1871 match(Op1, m_Not(m_Specific(Op0)))) 1872 return Constant::getAllOnesValue(Op0->getType()); 1873 1874 // Try some generic simplifications for associative operations. 1875 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q, 1876 MaxRecurse)) 1877 return V; 1878 1879 // Threading Xor over selects and phi nodes is pointless, so don't bother. 1880 // Threading over the select in "A ^ select(cond, B, C)" means evaluating 1881 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and 1882 // only if B and C are equal. If B and C are equal then (since we assume 1883 // that operands have already been simplified) "select(cond, B, C)" should 1884 // have been simplified to the common value of B and C already. Analysing 1885 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly 1886 // for threading over phi nodes. 1887 1888 return nullptr; 1889 } 1890 1891 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout *DL, 1892 const TargetLibraryInfo *TLI, 1893 const DominatorTree *DT, AssumptionCache *AC, 1894 const Instruction *CxtI) { 1895 return ::SimplifyXorInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1896 RecursionLimit); 1897 } 1898 1899 static Type *GetCompareTy(Value *Op) { 1900 return CmpInst::makeCmpResultType(Op->getType()); 1901 } 1902 1903 /// ExtractEquivalentCondition - Rummage around inside V looking for something 1904 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found, 1905 /// otherwise return null. Helper function for analyzing max/min idioms. 1906 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred, 1907 Value *LHS, Value *RHS) { 1908 SelectInst *SI = dyn_cast<SelectInst>(V); 1909 if (!SI) 1910 return nullptr; 1911 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); 1912 if (!Cmp) 1913 return nullptr; 1914 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1); 1915 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS) 1916 return Cmp; 1917 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) && 1918 LHS == CmpRHS && RHS == CmpLHS) 1919 return Cmp; 1920 return nullptr; 1921 } 1922 1923 // A significant optimization not implemented here is assuming that alloca 1924 // addresses are not equal to incoming argument values. They don't *alias*, 1925 // as we say, but that doesn't mean they aren't equal, so we take a 1926 // conservative approach. 1927 // 1928 // This is inspired in part by C++11 5.10p1: 1929 // "Two pointers of the same type compare equal if and only if they are both 1930 // null, both point to the same function, or both represent the same 1931 // address." 1932 // 1933 // This is pretty permissive. 1934 // 1935 // It's also partly due to C11 6.5.9p6: 1936 // "Two pointers compare equal if and only if both are null pointers, both are 1937 // pointers to the same object (including a pointer to an object and a 1938 // subobject at its beginning) or function, both are pointers to one past the 1939 // last element of the same array object, or one is a pointer to one past the 1940 // end of one array object and the other is a pointer to the start of a 1941 // different array object that happens to immediately follow the first array 1942 // object in the address space.) 1943 // 1944 // C11's version is more restrictive, however there's no reason why an argument 1945 // couldn't be a one-past-the-end value for a stack object in the caller and be 1946 // equal to the beginning of a stack object in the callee. 1947 // 1948 // If the C and C++ standards are ever made sufficiently restrictive in this 1949 // area, it may be possible to update LLVM's semantics accordingly and reinstate 1950 // this optimization. 1951 static Constant *computePointerICmp(const DataLayout *DL, 1952 const TargetLibraryInfo *TLI, 1953 CmpInst::Predicate Pred, 1954 Value *LHS, Value *RHS) { 1955 // First, skip past any trivial no-ops. 1956 LHS = LHS->stripPointerCasts(); 1957 RHS = RHS->stripPointerCasts(); 1958 1959 // A non-null pointer is not equal to a null pointer. 1960 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) && 1961 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE)) 1962 return ConstantInt::get(GetCompareTy(LHS), 1963 !CmpInst::isTrueWhenEqual(Pred)); 1964 1965 // We can only fold certain predicates on pointer comparisons. 1966 switch (Pred) { 1967 default: 1968 return nullptr; 1969 1970 // Equality comaprisons are easy to fold. 1971 case CmpInst::ICMP_EQ: 1972 case CmpInst::ICMP_NE: 1973 break; 1974 1975 // We can only handle unsigned relational comparisons because 'inbounds' on 1976 // a GEP only protects against unsigned wrapping. 1977 case CmpInst::ICMP_UGT: 1978 case CmpInst::ICMP_UGE: 1979 case CmpInst::ICMP_ULT: 1980 case CmpInst::ICMP_ULE: 1981 // However, we have to switch them to their signed variants to handle 1982 // negative indices from the base pointer. 1983 Pred = ICmpInst::getSignedPredicate(Pred); 1984 break; 1985 } 1986 1987 // Strip off any constant offsets so that we can reason about them. 1988 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets 1989 // here and compare base addresses like AliasAnalysis does, however there are 1990 // numerous hazards. AliasAnalysis and its utilities rely on special rules 1991 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis 1992 // doesn't need to guarantee pointer inequality when it says NoAlias. 1993 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 1994 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 1995 1996 // If LHS and RHS are related via constant offsets to the same base 1997 // value, we can replace it with an icmp which just compares the offsets. 1998 if (LHS == RHS) 1999 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset); 2000 2001 // Various optimizations for (in)equality comparisons. 2002 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) { 2003 // Different non-empty allocations that exist at the same time have 2004 // different addresses (if the program can tell). Global variables always 2005 // exist, so they always exist during the lifetime of each other and all 2006 // allocas. Two different allocas usually have different addresses... 2007 // 2008 // However, if there's an @llvm.stackrestore dynamically in between two 2009 // allocas, they may have the same address. It's tempting to reduce the 2010 // scope of the problem by only looking at *static* allocas here. That would 2011 // cover the majority of allocas while significantly reducing the likelihood 2012 // of having an @llvm.stackrestore pop up in the middle. However, it's not 2013 // actually impossible for an @llvm.stackrestore to pop up in the middle of 2014 // an entry block. Also, if we have a block that's not attached to a 2015 // function, we can't tell if it's "static" under the current definition. 2016 // Theoretically, this problem could be fixed by creating a new kind of 2017 // instruction kind specifically for static allocas. Such a new instruction 2018 // could be required to be at the top of the entry block, thus preventing it 2019 // from being subject to a @llvm.stackrestore. Instcombine could even 2020 // convert regular allocas into these special allocas. It'd be nifty. 2021 // However, until then, this problem remains open. 2022 // 2023 // So, we'll assume that two non-empty allocas have different addresses 2024 // for now. 2025 // 2026 // With all that, if the offsets are within the bounds of their allocations 2027 // (and not one-past-the-end! so we can't use inbounds!), and their 2028 // allocations aren't the same, the pointers are not equal. 2029 // 2030 // Note that it's not necessary to check for LHS being a global variable 2031 // address, due to canonicalization and constant folding. 2032 if (isa<AllocaInst>(LHS) && 2033 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) { 2034 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset); 2035 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset); 2036 uint64_t LHSSize, RHSSize; 2037 if (LHSOffsetCI && RHSOffsetCI && 2038 getObjectSize(LHS, LHSSize, DL, TLI) && 2039 getObjectSize(RHS, RHSSize, DL, TLI)) { 2040 const APInt &LHSOffsetValue = LHSOffsetCI->getValue(); 2041 const APInt &RHSOffsetValue = RHSOffsetCI->getValue(); 2042 if (!LHSOffsetValue.isNegative() && 2043 !RHSOffsetValue.isNegative() && 2044 LHSOffsetValue.ult(LHSSize) && 2045 RHSOffsetValue.ult(RHSSize)) { 2046 return ConstantInt::get(GetCompareTy(LHS), 2047 !CmpInst::isTrueWhenEqual(Pred)); 2048 } 2049 } 2050 2051 // Repeat the above check but this time without depending on DataLayout 2052 // or being able to compute a precise size. 2053 if (!cast<PointerType>(LHS->getType())->isEmptyTy() && 2054 !cast<PointerType>(RHS->getType())->isEmptyTy() && 2055 LHSOffset->isNullValue() && 2056 RHSOffset->isNullValue()) 2057 return ConstantInt::get(GetCompareTy(LHS), 2058 !CmpInst::isTrueWhenEqual(Pred)); 2059 } 2060 2061 // Even if an non-inbounds GEP occurs along the path we can still optimize 2062 // equality comparisons concerning the result. We avoid walking the whole 2063 // chain again by starting where the last calls to 2064 // stripAndComputeConstantOffsets left off and accumulate the offsets. 2065 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true); 2066 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true); 2067 if (LHS == RHS) 2068 return ConstantExpr::getICmp(Pred, 2069 ConstantExpr::getAdd(LHSOffset, LHSNoBound), 2070 ConstantExpr::getAdd(RHSOffset, RHSNoBound)); 2071 2072 // If one side of the equality comparison must come from a noalias call 2073 // (meaning a system memory allocation function), and the other side must 2074 // come from a pointer that cannot overlap with dynamically-allocated 2075 // memory within the lifetime of the current function (allocas, byval 2076 // arguments, globals), then determine the comparison result here. 2077 SmallVector<Value *, 8> LHSUObjs, RHSUObjs; 2078 GetUnderlyingObjects(LHS, LHSUObjs, DL); 2079 GetUnderlyingObjects(RHS, RHSUObjs, DL); 2080 2081 // Is the set of underlying objects all noalias calls? 2082 auto IsNAC = [](SmallVectorImpl<Value *> &Objects) { 2083 return std::all_of(Objects.begin(), Objects.end(), 2084 [](Value *V){ return isNoAliasCall(V); }); 2085 }; 2086 2087 // Is the set of underlying objects all things which must be disjoint from 2088 // noalias calls. For allocas, we consider only static ones (dynamic 2089 // allocas might be transformed into calls to malloc not simultaneously 2090 // live with the compared-to allocation). For globals, we exclude symbols 2091 // that might be resolve lazily to symbols in another dynamically-loaded 2092 // library (and, thus, could be malloc'ed by the implementation). 2093 auto IsAllocDisjoint = [](SmallVectorImpl<Value *> &Objects) { 2094 return std::all_of(Objects.begin(), Objects.end(), 2095 [](Value *V){ 2096 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) 2097 return AI->getParent() && AI->getParent()->getParent() && 2098 AI->isStaticAlloca(); 2099 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 2100 return (GV->hasLocalLinkage() || 2101 GV->hasHiddenVisibility() || 2102 GV->hasProtectedVisibility() || 2103 GV->hasUnnamedAddr()) && 2104 !GV->isThreadLocal(); 2105 if (const Argument *A = dyn_cast<Argument>(V)) 2106 return A->hasByValAttr(); 2107 return false; 2108 }); 2109 }; 2110 2111 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) || 2112 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs))) 2113 return ConstantInt::get(GetCompareTy(LHS), 2114 !CmpInst::isTrueWhenEqual(Pred)); 2115 } 2116 2117 // Otherwise, fail. 2118 return nullptr; 2119 } 2120 2121 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can 2122 /// fold the result. If not, this returns null. 2123 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2124 const Query &Q, unsigned MaxRecurse) { 2125 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 2126 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!"); 2127 2128 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 2129 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 2130 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 2131 2132 // If we have a constant, make sure it is on the RHS. 2133 std::swap(LHS, RHS); 2134 Pred = CmpInst::getSwappedPredicate(Pred); 2135 } 2136 2137 Type *ITy = GetCompareTy(LHS); // The return type. 2138 Type *OpTy = LHS->getType(); // The operand type. 2139 2140 // icmp X, X -> true/false 2141 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false 2142 // because X could be 0. 2143 if (LHS == RHS || isa<UndefValue>(RHS)) 2144 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred)); 2145 2146 // Special case logic when the operands have i1 type. 2147 if (OpTy->getScalarType()->isIntegerTy(1)) { 2148 switch (Pred) { 2149 default: break; 2150 case ICmpInst::ICMP_EQ: 2151 // X == 1 -> X 2152 if (match(RHS, m_One())) 2153 return LHS; 2154 break; 2155 case ICmpInst::ICMP_NE: 2156 // X != 0 -> X 2157 if (match(RHS, m_Zero())) 2158 return LHS; 2159 break; 2160 case ICmpInst::ICMP_UGT: 2161 // X >u 0 -> X 2162 if (match(RHS, m_Zero())) 2163 return LHS; 2164 break; 2165 case ICmpInst::ICMP_UGE: 2166 // X >=u 1 -> X 2167 if (match(RHS, m_One())) 2168 return LHS; 2169 break; 2170 case ICmpInst::ICMP_SLT: 2171 // X <s 0 -> X 2172 if (match(RHS, m_Zero())) 2173 return LHS; 2174 break; 2175 case ICmpInst::ICMP_SLE: 2176 // X <=s -1 -> X 2177 if (match(RHS, m_One())) 2178 return LHS; 2179 break; 2180 } 2181 } 2182 2183 // If we are comparing with zero then try hard since this is a common case. 2184 if (match(RHS, m_Zero())) { 2185 bool LHSKnownNonNegative, LHSKnownNegative; 2186 switch (Pred) { 2187 default: llvm_unreachable("Unknown ICmp predicate!"); 2188 case ICmpInst::ICMP_ULT: 2189 return getFalse(ITy); 2190 case ICmpInst::ICMP_UGE: 2191 return getTrue(ITy); 2192 case ICmpInst::ICMP_EQ: 2193 case ICmpInst::ICMP_ULE: 2194 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2195 return getFalse(ITy); 2196 break; 2197 case ICmpInst::ICMP_NE: 2198 case ICmpInst::ICMP_UGT: 2199 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2200 return getTrue(ITy); 2201 break; 2202 case ICmpInst::ICMP_SLT: 2203 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2204 Q.CxtI, Q.DT); 2205 if (LHSKnownNegative) 2206 return getTrue(ITy); 2207 if (LHSKnownNonNegative) 2208 return getFalse(ITy); 2209 break; 2210 case ICmpInst::ICMP_SLE: 2211 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2212 Q.CxtI, Q.DT); 2213 if (LHSKnownNegative) 2214 return getTrue(ITy); 2215 if (LHSKnownNonNegative && 2216 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2217 return getFalse(ITy); 2218 break; 2219 case ICmpInst::ICMP_SGE: 2220 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2221 Q.CxtI, Q.DT); 2222 if (LHSKnownNegative) 2223 return getFalse(ITy); 2224 if (LHSKnownNonNegative) 2225 return getTrue(ITy); 2226 break; 2227 case ICmpInst::ICMP_SGT: 2228 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2229 Q.CxtI, Q.DT); 2230 if (LHSKnownNegative) 2231 return getFalse(ITy); 2232 if (LHSKnownNonNegative && 2233 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2234 return getTrue(ITy); 2235 break; 2236 } 2237 } 2238 2239 // See if we are doing a comparison with a constant integer. 2240 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2241 // Rule out tautological comparisons (eg., ult 0 or uge 0). 2242 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue()); 2243 if (RHS_CR.isEmptySet()) 2244 return ConstantInt::getFalse(CI->getContext()); 2245 if (RHS_CR.isFullSet()) 2246 return ConstantInt::getTrue(CI->getContext()); 2247 2248 // Many binary operators with constant RHS have easy to compute constant 2249 // range. Use them to check whether the comparison is a tautology. 2250 unsigned Width = CI->getBitWidth(); 2251 APInt Lower = APInt(Width, 0); 2252 APInt Upper = APInt(Width, 0); 2253 ConstantInt *CI2; 2254 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) { 2255 // 'urem x, CI2' produces [0, CI2). 2256 Upper = CI2->getValue(); 2257 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) { 2258 // 'srem x, CI2' produces (-|CI2|, |CI2|). 2259 Upper = CI2->getValue().abs(); 2260 Lower = (-Upper) + 1; 2261 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) { 2262 // 'udiv CI2, x' produces [0, CI2]. 2263 Upper = CI2->getValue() + 1; 2264 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) { 2265 // 'udiv x, CI2' produces [0, UINT_MAX / CI2]. 2266 APInt NegOne = APInt::getAllOnesValue(Width); 2267 if (!CI2->isZero()) 2268 Upper = NegOne.udiv(CI2->getValue()) + 1; 2269 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) { 2270 if (CI2->isMinSignedValue()) { 2271 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. 2272 Lower = CI2->getValue(); 2273 Upper = Lower.lshr(1) + 1; 2274 } else { 2275 // 'sdiv CI2, x' produces [-|CI2|, |CI2|]. 2276 Upper = CI2->getValue().abs() + 1; 2277 Lower = (-Upper) + 1; 2278 } 2279 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) { 2280 APInt IntMin = APInt::getSignedMinValue(Width); 2281 APInt IntMax = APInt::getSignedMaxValue(Width); 2282 APInt Val = CI2->getValue(); 2283 if (Val.isAllOnesValue()) { 2284 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] 2285 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2286 Lower = IntMin + 1; 2287 Upper = IntMax + 1; 2288 } else if (Val.countLeadingZeros() < Width - 1) { 2289 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2] 2290 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2291 Lower = IntMin.sdiv(Val); 2292 Upper = IntMax.sdiv(Val); 2293 if (Lower.sgt(Upper)) 2294 std::swap(Lower, Upper); 2295 Upper = Upper + 1; 2296 assert(Upper != Lower && "Upper part of range has wrapped!"); 2297 } 2298 } else if (match(LHS, m_NUWShl(m_ConstantInt(CI2), m_Value()))) { 2299 // 'shl nuw CI2, x' produces [CI2, CI2 << CLZ(CI2)] 2300 Lower = CI2->getValue(); 2301 Upper = Lower.shl(Lower.countLeadingZeros()) + 1; 2302 } else if (match(LHS, m_NSWShl(m_ConstantInt(CI2), m_Value()))) { 2303 if (CI2->isNegative()) { 2304 // 'shl nsw CI2, x' produces [CI2 << CLO(CI2)-1, CI2] 2305 unsigned ShiftAmount = CI2->getValue().countLeadingOnes() - 1; 2306 Lower = CI2->getValue().shl(ShiftAmount); 2307 Upper = CI2->getValue() + 1; 2308 } else { 2309 // 'shl nsw CI2, x' produces [CI2, CI2 << CLZ(CI2)-1] 2310 unsigned ShiftAmount = CI2->getValue().countLeadingZeros() - 1; 2311 Lower = CI2->getValue(); 2312 Upper = CI2->getValue().shl(ShiftAmount) + 1; 2313 } 2314 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) { 2315 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2]. 2316 APInt NegOne = APInt::getAllOnesValue(Width); 2317 if (CI2->getValue().ult(Width)) 2318 Upper = NegOne.lshr(CI2->getValue()) + 1; 2319 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) { 2320 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2]. 2321 unsigned ShiftAmount = Width - 1; 2322 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2323 ShiftAmount = CI2->getValue().countTrailingZeros(); 2324 Lower = CI2->getValue().lshr(ShiftAmount); 2325 Upper = CI2->getValue() + 1; 2326 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) { 2327 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2]. 2328 APInt IntMin = APInt::getSignedMinValue(Width); 2329 APInt IntMax = APInt::getSignedMaxValue(Width); 2330 if (CI2->getValue().ult(Width)) { 2331 Lower = IntMin.ashr(CI2->getValue()); 2332 Upper = IntMax.ashr(CI2->getValue()) + 1; 2333 } 2334 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) { 2335 unsigned ShiftAmount = Width - 1; 2336 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2337 ShiftAmount = CI2->getValue().countTrailingZeros(); 2338 if (CI2->isNegative()) { 2339 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)] 2340 Lower = CI2->getValue(); 2341 Upper = CI2->getValue().ashr(ShiftAmount) + 1; 2342 } else { 2343 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2] 2344 Lower = CI2->getValue().ashr(ShiftAmount); 2345 Upper = CI2->getValue() + 1; 2346 } 2347 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) { 2348 // 'or x, CI2' produces [CI2, UINT_MAX]. 2349 Lower = CI2->getValue(); 2350 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) { 2351 // 'and x, CI2' produces [0, CI2]. 2352 Upper = CI2->getValue() + 1; 2353 } 2354 if (Lower != Upper) { 2355 ConstantRange LHS_CR = ConstantRange(Lower, Upper); 2356 if (RHS_CR.contains(LHS_CR)) 2357 return ConstantInt::getTrue(RHS->getContext()); 2358 if (RHS_CR.inverse().contains(LHS_CR)) 2359 return ConstantInt::getFalse(RHS->getContext()); 2360 } 2361 } 2362 2363 // Compare of cast, for example (zext X) != 0 -> X != 0 2364 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) { 2365 Instruction *LI = cast<CastInst>(LHS); 2366 Value *SrcOp = LI->getOperand(0); 2367 Type *SrcTy = SrcOp->getType(); 2368 Type *DstTy = LI->getType(); 2369 2370 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input 2371 // if the integer type is the same size as the pointer type. 2372 if (MaxRecurse && Q.DL && isa<PtrToIntInst>(LI) && 2373 Q.DL->getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) { 2374 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 2375 // Transfer the cast to the constant. 2376 if (Value *V = SimplifyICmpInst(Pred, SrcOp, 2377 ConstantExpr::getIntToPtr(RHSC, SrcTy), 2378 Q, MaxRecurse-1)) 2379 return V; 2380 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) { 2381 if (RI->getOperand(0)->getType() == SrcTy) 2382 // Compare without the cast. 2383 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2384 Q, MaxRecurse-1)) 2385 return V; 2386 } 2387 } 2388 2389 if (isa<ZExtInst>(LHS)) { 2390 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the 2391 // same type. 2392 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) { 2393 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2394 // Compare X and Y. Note that signed predicates become unsigned. 2395 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2396 SrcOp, RI->getOperand(0), Q, 2397 MaxRecurse-1)) 2398 return V; 2399 } 2400 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended 2401 // too. If not, then try to deduce the result of the comparison. 2402 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2403 // Compute the constant that would happen if we truncated to SrcTy then 2404 // reextended to DstTy. 2405 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2406 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy); 2407 2408 // If the re-extended constant didn't change then this is effectively 2409 // also a case of comparing two zero-extended values. 2410 if (RExt == CI && MaxRecurse) 2411 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2412 SrcOp, Trunc, Q, MaxRecurse-1)) 2413 return V; 2414 2415 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit 2416 // there. Use this to work out the result of the comparison. 2417 if (RExt != CI) { 2418 switch (Pred) { 2419 default: llvm_unreachable("Unknown ICmp predicate!"); 2420 // LHS <u RHS. 2421 case ICmpInst::ICMP_EQ: 2422 case ICmpInst::ICMP_UGT: 2423 case ICmpInst::ICMP_UGE: 2424 return ConstantInt::getFalse(CI->getContext()); 2425 2426 case ICmpInst::ICMP_NE: 2427 case ICmpInst::ICMP_ULT: 2428 case ICmpInst::ICMP_ULE: 2429 return ConstantInt::getTrue(CI->getContext()); 2430 2431 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS 2432 // is non-negative then LHS <s RHS. 2433 case ICmpInst::ICMP_SGT: 2434 case ICmpInst::ICMP_SGE: 2435 return CI->getValue().isNegative() ? 2436 ConstantInt::getTrue(CI->getContext()) : 2437 ConstantInt::getFalse(CI->getContext()); 2438 2439 case ICmpInst::ICMP_SLT: 2440 case ICmpInst::ICMP_SLE: 2441 return CI->getValue().isNegative() ? 2442 ConstantInt::getFalse(CI->getContext()) : 2443 ConstantInt::getTrue(CI->getContext()); 2444 } 2445 } 2446 } 2447 } 2448 2449 if (isa<SExtInst>(LHS)) { 2450 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the 2451 // same type. 2452 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) { 2453 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2454 // Compare X and Y. Note that the predicate does not change. 2455 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2456 Q, MaxRecurse-1)) 2457 return V; 2458 } 2459 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended 2460 // too. If not, then try to deduce the result of the comparison. 2461 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2462 // Compute the constant that would happen if we truncated to SrcTy then 2463 // reextended to DstTy. 2464 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2465 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy); 2466 2467 // If the re-extended constant didn't change then this is effectively 2468 // also a case of comparing two sign-extended values. 2469 if (RExt == CI && MaxRecurse) 2470 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1)) 2471 return V; 2472 2473 // Otherwise the upper bits of LHS are all equal, while RHS has varying 2474 // bits there. Use this to work out the result of the comparison. 2475 if (RExt != CI) { 2476 switch (Pred) { 2477 default: llvm_unreachable("Unknown ICmp predicate!"); 2478 case ICmpInst::ICMP_EQ: 2479 return ConstantInt::getFalse(CI->getContext()); 2480 case ICmpInst::ICMP_NE: 2481 return ConstantInt::getTrue(CI->getContext()); 2482 2483 // If RHS is non-negative then LHS <s RHS. If RHS is negative then 2484 // LHS >s RHS. 2485 case ICmpInst::ICMP_SGT: 2486 case ICmpInst::ICMP_SGE: 2487 return CI->getValue().isNegative() ? 2488 ConstantInt::getTrue(CI->getContext()) : 2489 ConstantInt::getFalse(CI->getContext()); 2490 case ICmpInst::ICMP_SLT: 2491 case ICmpInst::ICMP_SLE: 2492 return CI->getValue().isNegative() ? 2493 ConstantInt::getFalse(CI->getContext()) : 2494 ConstantInt::getTrue(CI->getContext()); 2495 2496 // If LHS is non-negative then LHS <u RHS. If LHS is negative then 2497 // LHS >u RHS. 2498 case ICmpInst::ICMP_UGT: 2499 case ICmpInst::ICMP_UGE: 2500 // Comparison is true iff the LHS <s 0. 2501 if (MaxRecurse) 2502 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp, 2503 Constant::getNullValue(SrcTy), 2504 Q, MaxRecurse-1)) 2505 return V; 2506 break; 2507 case ICmpInst::ICMP_ULT: 2508 case ICmpInst::ICMP_ULE: 2509 // Comparison is true iff the LHS >=s 0. 2510 if (MaxRecurse) 2511 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp, 2512 Constant::getNullValue(SrcTy), 2513 Q, MaxRecurse-1)) 2514 return V; 2515 break; 2516 } 2517 } 2518 } 2519 } 2520 } 2521 2522 // Special logic for binary operators. 2523 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS); 2524 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS); 2525 if (MaxRecurse && (LBO || RBO)) { 2526 // Analyze the case when either LHS or RHS is an add instruction. 2527 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 2528 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null). 2529 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false; 2530 if (LBO && LBO->getOpcode() == Instruction::Add) { 2531 A = LBO->getOperand(0); B = LBO->getOperand(1); 2532 NoLHSWrapProblem = ICmpInst::isEquality(Pred) || 2533 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) || 2534 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap()); 2535 } 2536 if (RBO && RBO->getOpcode() == Instruction::Add) { 2537 C = RBO->getOperand(0); D = RBO->getOperand(1); 2538 NoRHSWrapProblem = ICmpInst::isEquality(Pred) || 2539 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) || 2540 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap()); 2541 } 2542 2543 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2544 if ((A == RHS || B == RHS) && NoLHSWrapProblem) 2545 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A, 2546 Constant::getNullValue(RHS->getType()), 2547 Q, MaxRecurse-1)) 2548 return V; 2549 2550 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2551 if ((C == LHS || D == LHS) && NoRHSWrapProblem) 2552 if (Value *V = SimplifyICmpInst(Pred, 2553 Constant::getNullValue(LHS->getType()), 2554 C == LHS ? D : C, Q, MaxRecurse-1)) 2555 return V; 2556 2557 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow. 2558 if (A && C && (A == C || A == D || B == C || B == D) && 2559 NoLHSWrapProblem && NoRHSWrapProblem) { 2560 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2561 Value *Y, *Z; 2562 if (A == C) { 2563 // C + B == C + D -> B == D 2564 Y = B; 2565 Z = D; 2566 } else if (A == D) { 2567 // D + B == C + D -> B == C 2568 Y = B; 2569 Z = C; 2570 } else if (B == C) { 2571 // A + C == C + D -> A == D 2572 Y = A; 2573 Z = D; 2574 } else { 2575 assert(B == D); 2576 // A + D == C + D -> A == C 2577 Y = A; 2578 Z = C; 2579 } 2580 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1)) 2581 return V; 2582 } 2583 } 2584 2585 // icmp pred (or X, Y), X 2586 if (LBO && match(LBO, m_CombineOr(m_Or(m_Value(), m_Specific(RHS)), 2587 m_Or(m_Specific(RHS), m_Value())))) { 2588 if (Pred == ICmpInst::ICMP_ULT) 2589 return getFalse(ITy); 2590 if (Pred == ICmpInst::ICMP_UGE) 2591 return getTrue(ITy); 2592 } 2593 // icmp pred X, (or X, Y) 2594 if (RBO && match(RBO, m_CombineOr(m_Or(m_Value(), m_Specific(LHS)), 2595 m_Or(m_Specific(LHS), m_Value())))) { 2596 if (Pred == ICmpInst::ICMP_ULE) 2597 return getTrue(ITy); 2598 if (Pred == ICmpInst::ICMP_UGT) 2599 return getFalse(ITy); 2600 } 2601 2602 // icmp pred (and X, Y), X 2603 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)), 2604 m_And(m_Specific(RHS), m_Value())))) { 2605 if (Pred == ICmpInst::ICMP_UGT) 2606 return getFalse(ITy); 2607 if (Pred == ICmpInst::ICMP_ULE) 2608 return getTrue(ITy); 2609 } 2610 // icmp pred X, (and X, Y) 2611 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)), 2612 m_And(m_Specific(LHS), m_Value())))) { 2613 if (Pred == ICmpInst::ICMP_UGE) 2614 return getTrue(ITy); 2615 if (Pred == ICmpInst::ICMP_ULT) 2616 return getFalse(ITy); 2617 } 2618 2619 // 0 - (zext X) pred C 2620 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) { 2621 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 2622 if (RHSC->getValue().isStrictlyPositive()) { 2623 if (Pred == ICmpInst::ICMP_SLT) 2624 return ConstantInt::getTrue(RHSC->getContext()); 2625 if (Pred == ICmpInst::ICMP_SGE) 2626 return ConstantInt::getFalse(RHSC->getContext()); 2627 if (Pred == ICmpInst::ICMP_EQ) 2628 return ConstantInt::getFalse(RHSC->getContext()); 2629 if (Pred == ICmpInst::ICMP_NE) 2630 return ConstantInt::getTrue(RHSC->getContext()); 2631 } 2632 if (RHSC->getValue().isNonNegative()) { 2633 if (Pred == ICmpInst::ICMP_SLE) 2634 return ConstantInt::getTrue(RHSC->getContext()); 2635 if (Pred == ICmpInst::ICMP_SGT) 2636 return ConstantInt::getFalse(RHSC->getContext()); 2637 } 2638 } 2639 } 2640 2641 // icmp pred (urem X, Y), Y 2642 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) { 2643 bool KnownNonNegative, KnownNegative; 2644 switch (Pred) { 2645 default: 2646 break; 2647 case ICmpInst::ICMP_SGT: 2648 case ICmpInst::ICMP_SGE: 2649 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2650 Q.CxtI, Q.DT); 2651 if (!KnownNonNegative) 2652 break; 2653 // fall-through 2654 case ICmpInst::ICMP_EQ: 2655 case ICmpInst::ICMP_UGT: 2656 case ICmpInst::ICMP_UGE: 2657 return getFalse(ITy); 2658 case ICmpInst::ICMP_SLT: 2659 case ICmpInst::ICMP_SLE: 2660 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2661 Q.CxtI, Q.DT); 2662 if (!KnownNonNegative) 2663 break; 2664 // fall-through 2665 case ICmpInst::ICMP_NE: 2666 case ICmpInst::ICMP_ULT: 2667 case ICmpInst::ICMP_ULE: 2668 return getTrue(ITy); 2669 } 2670 } 2671 2672 // icmp pred X, (urem Y, X) 2673 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) { 2674 bool KnownNonNegative, KnownNegative; 2675 switch (Pred) { 2676 default: 2677 break; 2678 case ICmpInst::ICMP_SGT: 2679 case ICmpInst::ICMP_SGE: 2680 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2681 Q.CxtI, Q.DT); 2682 if (!KnownNonNegative) 2683 break; 2684 // fall-through 2685 case ICmpInst::ICMP_NE: 2686 case ICmpInst::ICMP_UGT: 2687 case ICmpInst::ICMP_UGE: 2688 return getTrue(ITy); 2689 case ICmpInst::ICMP_SLT: 2690 case ICmpInst::ICMP_SLE: 2691 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2692 Q.CxtI, Q.DT); 2693 if (!KnownNonNegative) 2694 break; 2695 // fall-through 2696 case ICmpInst::ICMP_EQ: 2697 case ICmpInst::ICMP_ULT: 2698 case ICmpInst::ICMP_ULE: 2699 return getFalse(ITy); 2700 } 2701 } 2702 2703 // x udiv y <=u x. 2704 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) { 2705 // icmp pred (X /u Y), X 2706 if (Pred == ICmpInst::ICMP_UGT) 2707 return getFalse(ITy); 2708 if (Pred == ICmpInst::ICMP_ULE) 2709 return getTrue(ITy); 2710 } 2711 2712 // handle: 2713 // CI2 << X == CI 2714 // CI2 << X != CI 2715 // 2716 // where CI2 is a power of 2 and CI isn't 2717 if (auto *CI = dyn_cast<ConstantInt>(RHS)) { 2718 const APInt *CI2Val, *CIVal = &CI->getValue(); 2719 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) && 2720 CI2Val->isPowerOf2()) { 2721 if (!CIVal->isPowerOf2()) { 2722 // CI2 << X can equal zero in some circumstances, 2723 // this simplification is unsafe if CI is zero. 2724 // 2725 // We know it is safe if: 2726 // - The shift is nsw, we can't shift out the one bit. 2727 // - The shift is nuw, we can't shift out the one bit. 2728 // - CI2 is one 2729 // - CI isn't zero 2730 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() || 2731 *CI2Val == 1 || !CI->isZero()) { 2732 if (Pred == ICmpInst::ICMP_EQ) 2733 return ConstantInt::getFalse(RHS->getContext()); 2734 if (Pred == ICmpInst::ICMP_NE) 2735 return ConstantInt::getTrue(RHS->getContext()); 2736 } 2737 } 2738 if (CIVal->isSignBit() && *CI2Val == 1) { 2739 if (Pred == ICmpInst::ICMP_UGT) 2740 return ConstantInt::getFalse(RHS->getContext()); 2741 if (Pred == ICmpInst::ICMP_ULE) 2742 return ConstantInt::getTrue(RHS->getContext()); 2743 } 2744 } 2745 } 2746 2747 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() && 2748 LBO->getOperand(1) == RBO->getOperand(1)) { 2749 switch (LBO->getOpcode()) { 2750 default: break; 2751 case Instruction::UDiv: 2752 case Instruction::LShr: 2753 if (ICmpInst::isSigned(Pred)) 2754 break; 2755 // fall-through 2756 case Instruction::SDiv: 2757 case Instruction::AShr: 2758 if (!LBO->isExact() || !RBO->isExact()) 2759 break; 2760 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2761 RBO->getOperand(0), Q, MaxRecurse-1)) 2762 return V; 2763 break; 2764 case Instruction::Shl: { 2765 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap(); 2766 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap(); 2767 if (!NUW && !NSW) 2768 break; 2769 if (!NSW && ICmpInst::isSigned(Pred)) 2770 break; 2771 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2772 RBO->getOperand(0), Q, MaxRecurse-1)) 2773 return V; 2774 break; 2775 } 2776 } 2777 } 2778 2779 // Simplify comparisons involving max/min. 2780 Value *A, *B; 2781 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE; 2782 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B". 2783 2784 // Signed variants on "max(a,b)>=a -> true". 2785 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2786 if (A != RHS) std::swap(A, B); // smax(A, B) pred A. 2787 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2788 // We analyze this as smax(A, B) pred A. 2789 P = Pred; 2790 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) && 2791 (A == LHS || B == LHS)) { 2792 if (A != LHS) std::swap(A, B); // A pred smax(A, B). 2793 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2794 // We analyze this as smax(A, B) swapped-pred A. 2795 P = CmpInst::getSwappedPredicate(Pred); 2796 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2797 (A == RHS || B == RHS)) { 2798 if (A != RHS) std::swap(A, B); // smin(A, B) pred A. 2799 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2800 // We analyze this as smax(-A, -B) swapped-pred -A. 2801 // Note that we do not need to actually form -A or -B thanks to EqP. 2802 P = CmpInst::getSwappedPredicate(Pred); 2803 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) && 2804 (A == LHS || B == LHS)) { 2805 if (A != LHS) std::swap(A, B); // A pred smin(A, B). 2806 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2807 // We analyze this as smax(-A, -B) pred -A. 2808 // Note that we do not need to actually form -A or -B thanks to EqP. 2809 P = Pred; 2810 } 2811 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2812 // Cases correspond to "max(A, B) p A". 2813 switch (P) { 2814 default: 2815 break; 2816 case CmpInst::ICMP_EQ: 2817 case CmpInst::ICMP_SLE: 2818 // Equivalent to "A EqP B". This may be the same as the condition tested 2819 // in the max/min; if so, we can just return that. 2820 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2821 return V; 2822 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2823 return V; 2824 // Otherwise, see if "A EqP B" simplifies. 2825 if (MaxRecurse) 2826 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2827 return V; 2828 break; 2829 case CmpInst::ICMP_NE: 2830 case CmpInst::ICMP_SGT: { 2831 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2832 // Equivalent to "A InvEqP B". This may be the same as the condition 2833 // tested in the max/min; if so, we can just return that. 2834 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2835 return V; 2836 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2837 return V; 2838 // Otherwise, see if "A InvEqP B" simplifies. 2839 if (MaxRecurse) 2840 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2841 return V; 2842 break; 2843 } 2844 case CmpInst::ICMP_SGE: 2845 // Always true. 2846 return getTrue(ITy); 2847 case CmpInst::ICMP_SLT: 2848 // Always false. 2849 return getFalse(ITy); 2850 } 2851 } 2852 2853 // Unsigned variants on "max(a,b)>=a -> true". 2854 P = CmpInst::BAD_ICMP_PREDICATE; 2855 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2856 if (A != RHS) std::swap(A, B); // umax(A, B) pred A. 2857 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2858 // We analyze this as umax(A, B) pred A. 2859 P = Pred; 2860 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) && 2861 (A == LHS || B == LHS)) { 2862 if (A != LHS) std::swap(A, B); // A pred umax(A, B). 2863 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2864 // We analyze this as umax(A, B) swapped-pred A. 2865 P = CmpInst::getSwappedPredicate(Pred); 2866 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2867 (A == RHS || B == RHS)) { 2868 if (A != RHS) std::swap(A, B); // umin(A, B) pred A. 2869 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2870 // We analyze this as umax(-A, -B) swapped-pred -A. 2871 // Note that we do not need to actually form -A or -B thanks to EqP. 2872 P = CmpInst::getSwappedPredicate(Pred); 2873 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) && 2874 (A == LHS || B == LHS)) { 2875 if (A != LHS) std::swap(A, B); // A pred umin(A, B). 2876 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2877 // We analyze this as umax(-A, -B) pred -A. 2878 // Note that we do not need to actually form -A or -B thanks to EqP. 2879 P = Pred; 2880 } 2881 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2882 // Cases correspond to "max(A, B) p A". 2883 switch (P) { 2884 default: 2885 break; 2886 case CmpInst::ICMP_EQ: 2887 case CmpInst::ICMP_ULE: 2888 // Equivalent to "A EqP B". This may be the same as the condition tested 2889 // in the max/min; if so, we can just return that. 2890 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2891 return V; 2892 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2893 return V; 2894 // Otherwise, see if "A EqP B" simplifies. 2895 if (MaxRecurse) 2896 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2897 return V; 2898 break; 2899 case CmpInst::ICMP_NE: 2900 case CmpInst::ICMP_UGT: { 2901 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2902 // Equivalent to "A InvEqP B". This may be the same as the condition 2903 // tested in the max/min; if so, we can just return that. 2904 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2905 return V; 2906 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2907 return V; 2908 // Otherwise, see if "A InvEqP B" simplifies. 2909 if (MaxRecurse) 2910 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2911 return V; 2912 break; 2913 } 2914 case CmpInst::ICMP_UGE: 2915 // Always true. 2916 return getTrue(ITy); 2917 case CmpInst::ICMP_ULT: 2918 // Always false. 2919 return getFalse(ITy); 2920 } 2921 } 2922 2923 // Variants on "max(x,y) >= min(x,z)". 2924 Value *C, *D; 2925 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && 2926 match(RHS, m_SMin(m_Value(C), m_Value(D))) && 2927 (A == C || A == D || B == C || B == D)) { 2928 // max(x, ?) pred min(x, ?). 2929 if (Pred == CmpInst::ICMP_SGE) 2930 // Always true. 2931 return getTrue(ITy); 2932 if (Pred == CmpInst::ICMP_SLT) 2933 // Always false. 2934 return getFalse(ITy); 2935 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2936 match(RHS, m_SMax(m_Value(C), m_Value(D))) && 2937 (A == C || A == D || B == C || B == D)) { 2938 // min(x, ?) pred max(x, ?). 2939 if (Pred == CmpInst::ICMP_SLE) 2940 // Always true. 2941 return getTrue(ITy); 2942 if (Pred == CmpInst::ICMP_SGT) 2943 // Always false. 2944 return getFalse(ITy); 2945 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && 2946 match(RHS, m_UMin(m_Value(C), m_Value(D))) && 2947 (A == C || A == D || B == C || B == D)) { 2948 // max(x, ?) pred min(x, ?). 2949 if (Pred == CmpInst::ICMP_UGE) 2950 // Always true. 2951 return getTrue(ITy); 2952 if (Pred == CmpInst::ICMP_ULT) 2953 // Always false. 2954 return getFalse(ITy); 2955 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2956 match(RHS, m_UMax(m_Value(C), m_Value(D))) && 2957 (A == C || A == D || B == C || B == D)) { 2958 // min(x, ?) pred max(x, ?). 2959 if (Pred == CmpInst::ICMP_ULE) 2960 // Always true. 2961 return getTrue(ITy); 2962 if (Pred == CmpInst::ICMP_UGT) 2963 // Always false. 2964 return getFalse(ITy); 2965 } 2966 2967 // Simplify comparisons of related pointers using a powerful, recursive 2968 // GEP-walk when we have target data available.. 2969 if (LHS->getType()->isPointerTy()) 2970 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS)) 2971 return C; 2972 2973 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) { 2974 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) { 2975 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() && 2976 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() && 2977 (ICmpInst::isEquality(Pred) || 2978 (GLHS->isInBounds() && GRHS->isInBounds() && 2979 Pred == ICmpInst::getSignedPredicate(Pred)))) { 2980 // The bases are equal and the indices are constant. Build a constant 2981 // expression GEP with the same indices and a null base pointer to see 2982 // what constant folding can make out of it. 2983 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType()); 2984 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end()); 2985 Constant *NewLHS = ConstantExpr::getGetElementPtr(Null, IndicesLHS); 2986 2987 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end()); 2988 Constant *NewRHS = ConstantExpr::getGetElementPtr(Null, IndicesRHS); 2989 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS); 2990 } 2991 } 2992 } 2993 2994 // If a bit is known to be zero for A and known to be one for B, 2995 // then A and B cannot be equal. 2996 if (ICmpInst::isEquality(Pred)) { 2997 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2998 uint32_t BitWidth = CI->getBitWidth(); 2999 APInt LHSKnownZero(BitWidth, 0); 3000 APInt LHSKnownOne(BitWidth, 0); 3001 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC, 3002 Q.CxtI, Q.DT); 3003 const APInt &RHSVal = CI->getValue(); 3004 if (((LHSKnownZero & RHSVal) != 0) || ((LHSKnownOne & ~RHSVal) != 0)) 3005 return Pred == ICmpInst::ICMP_EQ 3006 ? ConstantInt::getFalse(CI->getContext()) 3007 : ConstantInt::getTrue(CI->getContext()); 3008 } 3009 } 3010 3011 // If the comparison is with the result of a select instruction, check whether 3012 // comparing with either branch of the select always yields the same value. 3013 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3014 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3015 return V; 3016 3017 // If the comparison is with the result of a phi instruction, check whether 3018 // doing the compare with each incoming phi value yields a common result. 3019 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3020 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3021 return V; 3022 3023 return nullptr; 3024 } 3025 3026 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3027 const DataLayout *DL, 3028 const TargetLibraryInfo *TLI, 3029 const DominatorTree *DT, AssumptionCache *AC, 3030 Instruction *CxtI) { 3031 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3032 RecursionLimit); 3033 } 3034 3035 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can 3036 /// fold the result. If not, this returns null. 3037 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3038 const Query &Q, unsigned MaxRecurse) { 3039 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 3040 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!"); 3041 3042 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 3043 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 3044 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 3045 3046 // If we have a constant, make sure it is on the RHS. 3047 std::swap(LHS, RHS); 3048 Pred = CmpInst::getSwappedPredicate(Pred); 3049 } 3050 3051 // Fold trivial predicates. 3052 if (Pred == FCmpInst::FCMP_FALSE) 3053 return ConstantInt::get(GetCompareTy(LHS), 0); 3054 if (Pred == FCmpInst::FCMP_TRUE) 3055 return ConstantInt::get(GetCompareTy(LHS), 1); 3056 3057 if (isa<UndefValue>(RHS)) // fcmp pred X, undef -> undef 3058 return UndefValue::get(GetCompareTy(LHS)); 3059 3060 // fcmp x,x -> true/false. Not all compares are foldable. 3061 if (LHS == RHS) { 3062 if (CmpInst::isTrueWhenEqual(Pred)) 3063 return ConstantInt::get(GetCompareTy(LHS), 1); 3064 if (CmpInst::isFalseWhenEqual(Pred)) 3065 return ConstantInt::get(GetCompareTy(LHS), 0); 3066 } 3067 3068 // Handle fcmp with constant RHS 3069 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) { 3070 // If the constant is a nan, see if we can fold the comparison based on it. 3071 if (CFP->getValueAPF().isNaN()) { 3072 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo" 3073 return ConstantInt::getFalse(CFP->getContext()); 3074 assert(FCmpInst::isUnordered(Pred) && 3075 "Comparison must be either ordered or unordered!"); 3076 // True if unordered. 3077 return ConstantInt::getTrue(CFP->getContext()); 3078 } 3079 // Check whether the constant is an infinity. 3080 if (CFP->getValueAPF().isInfinity()) { 3081 if (CFP->getValueAPF().isNegative()) { 3082 switch (Pred) { 3083 case FCmpInst::FCMP_OLT: 3084 // No value is ordered and less than negative infinity. 3085 return ConstantInt::getFalse(CFP->getContext()); 3086 case FCmpInst::FCMP_UGE: 3087 // All values are unordered with or at least negative infinity. 3088 return ConstantInt::getTrue(CFP->getContext()); 3089 default: 3090 break; 3091 } 3092 } else { 3093 switch (Pred) { 3094 case FCmpInst::FCMP_OGT: 3095 // No value is ordered and greater than infinity. 3096 return ConstantInt::getFalse(CFP->getContext()); 3097 case FCmpInst::FCMP_ULE: 3098 // All values are unordered with and at most infinity. 3099 return ConstantInt::getTrue(CFP->getContext()); 3100 default: 3101 break; 3102 } 3103 } 3104 } 3105 if (CFP->getValueAPF().isZero()) { 3106 switch (Pred) { 3107 case FCmpInst::FCMP_UGE: 3108 if (CannotBeOrderedLessThanZero(LHS)) 3109 return ConstantInt::getTrue(CFP->getContext()); 3110 break; 3111 case FCmpInst::FCMP_OLT: 3112 // X < 0 3113 if (CannotBeOrderedLessThanZero(LHS)) 3114 return ConstantInt::getFalse(CFP->getContext()); 3115 break; 3116 default: 3117 break; 3118 } 3119 } 3120 } 3121 3122 // If the comparison is with the result of a select instruction, check whether 3123 // comparing with either branch of the select always yields the same value. 3124 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3125 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3126 return V; 3127 3128 // If the comparison is with the result of a phi instruction, check whether 3129 // doing the compare with each incoming phi value yields a common result. 3130 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3131 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3132 return V; 3133 3134 return nullptr; 3135 } 3136 3137 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3138 const DataLayout *DL, 3139 const TargetLibraryInfo *TLI, 3140 const DominatorTree *DT, AssumptionCache *AC, 3141 const Instruction *CxtI) { 3142 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3143 RecursionLimit); 3144 } 3145 3146 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold 3147 /// the result. If not, this returns null. 3148 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal, 3149 Value *FalseVal, const Query &Q, 3150 unsigned MaxRecurse) { 3151 // select true, X, Y -> X 3152 // select false, X, Y -> Y 3153 if (Constant *CB = dyn_cast<Constant>(CondVal)) { 3154 if (CB->isAllOnesValue()) 3155 return TrueVal; 3156 if (CB->isNullValue()) 3157 return FalseVal; 3158 } 3159 3160 // select C, X, X -> X 3161 if (TrueVal == FalseVal) 3162 return TrueVal; 3163 3164 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y 3165 if (isa<Constant>(TrueVal)) 3166 return TrueVal; 3167 return FalseVal; 3168 } 3169 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X 3170 return FalseVal; 3171 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X 3172 return TrueVal; 3173 3174 const auto *ICI = dyn_cast<ICmpInst>(CondVal); 3175 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); 3176 if (ICI && BitWidth) { 3177 ICmpInst::Predicate Pred = ICI->getPredicate(); 3178 APInt MinSignedValue = APInt::getSignBit(BitWidth); 3179 Value *X; 3180 const APInt *Y; 3181 bool TrueWhenUnset; 3182 bool IsBitTest = false; 3183 if (ICmpInst::isEquality(Pred) && 3184 match(ICI->getOperand(0), m_And(m_Value(X), m_APInt(Y))) && 3185 match(ICI->getOperand(1), m_Zero())) { 3186 IsBitTest = true; 3187 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 3188 } else if (Pred == ICmpInst::ICMP_SLT && 3189 match(ICI->getOperand(1), m_Zero())) { 3190 X = ICI->getOperand(0); 3191 Y = &MinSignedValue; 3192 IsBitTest = true; 3193 TrueWhenUnset = false; 3194 } else if (Pred == ICmpInst::ICMP_SGT && 3195 match(ICI->getOperand(1), m_AllOnes())) { 3196 X = ICI->getOperand(0); 3197 Y = &MinSignedValue; 3198 IsBitTest = true; 3199 TrueWhenUnset = true; 3200 } 3201 if (IsBitTest) { 3202 const APInt *C; 3203 // (X & Y) == 0 ? X & ~Y : X --> X 3204 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y 3205 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) && 3206 *Y == ~*C) 3207 return TrueWhenUnset ? FalseVal : TrueVal; 3208 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y 3209 // (X & Y) != 0 ? X : X & ~Y --> X 3210 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) && 3211 *Y == ~*C) 3212 return TrueWhenUnset ? FalseVal : TrueVal; 3213 3214 if (Y->isPowerOf2()) { 3215 // (X & Y) == 0 ? X | Y : X --> X | Y 3216 // (X & Y) != 0 ? X | Y : X --> X 3217 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) && 3218 *Y == *C) 3219 return TrueWhenUnset ? TrueVal : FalseVal; 3220 // (X & Y) == 0 ? X : X | Y --> X 3221 // (X & Y) != 0 ? X : X | Y --> X | Y 3222 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) && 3223 *Y == *C) 3224 return TrueWhenUnset ? TrueVal : FalseVal; 3225 } 3226 } 3227 } 3228 3229 return nullptr; 3230 } 3231 3232 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, 3233 const DataLayout *DL, 3234 const TargetLibraryInfo *TLI, 3235 const DominatorTree *DT, AssumptionCache *AC, 3236 const Instruction *CxtI) { 3237 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, 3238 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3239 } 3240 3241 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can 3242 /// fold the result. If not, this returns null. 3243 static Value *SimplifyGEPInst(ArrayRef<Value *> Ops, const Query &Q, unsigned) { 3244 // The type of the GEP pointer operand. 3245 PointerType *PtrTy = cast<PointerType>(Ops[0]->getType()->getScalarType()); 3246 unsigned AS = PtrTy->getAddressSpace(); 3247 3248 // getelementptr P -> P. 3249 if (Ops.size() == 1) 3250 return Ops[0]; 3251 3252 // Compute the (pointer) type returned by the GEP instruction. 3253 Type *LastType = GetElementPtrInst::getIndexedType(PtrTy, Ops.slice(1)); 3254 Type *GEPTy = PointerType::get(LastType, AS); 3255 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType())) 3256 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 3257 3258 if (isa<UndefValue>(Ops[0])) 3259 return UndefValue::get(GEPTy); 3260 3261 if (Ops.size() == 2) { 3262 // getelementptr P, 0 -> P. 3263 if (match(Ops[1], m_Zero())) 3264 return Ops[0]; 3265 3266 Type *Ty = PtrTy->getElementType(); 3267 if (Q.DL && Ty->isSized()) { 3268 Value *P; 3269 uint64_t C; 3270 uint64_t TyAllocSize = Q.DL->getTypeAllocSize(Ty); 3271 // getelementptr P, N -> P if P points to a type of zero size. 3272 if (TyAllocSize == 0) 3273 return Ops[0]; 3274 3275 // The following transforms are only safe if the ptrtoint cast 3276 // doesn't truncate the pointers. 3277 if (Ops[1]->getType()->getScalarSizeInBits() == 3278 Q.DL->getPointerSizeInBits(AS)) { 3279 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * { 3280 if (match(P, m_Zero())) 3281 return Constant::getNullValue(GEPTy); 3282 Value *Temp; 3283 if (match(P, m_PtrToInt(m_Value(Temp)))) 3284 if (Temp->getType() == GEPTy) 3285 return Temp; 3286 return nullptr; 3287 }; 3288 3289 // getelementptr V, (sub P, V) -> P if P points to a type of size 1. 3290 if (TyAllocSize == 1 && 3291 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))))) 3292 if (Value *R = PtrToIntOrZero(P)) 3293 return R; 3294 3295 // getelementptr V, (ashr (sub P, V), C) -> Q 3296 // if P points to a type of size 1 << C. 3297 if (match(Ops[1], 3298 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3299 m_ConstantInt(C))) && 3300 TyAllocSize == 1ULL << C) 3301 if (Value *R = PtrToIntOrZero(P)) 3302 return R; 3303 3304 // getelementptr V, (sdiv (sub P, V), C) -> Q 3305 // if P points to a type of size C. 3306 if (match(Ops[1], 3307 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3308 m_SpecificInt(TyAllocSize)))) 3309 if (Value *R = PtrToIntOrZero(P)) 3310 return R; 3311 } 3312 } 3313 } 3314 3315 // Check to see if this is constant foldable. 3316 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 3317 if (!isa<Constant>(Ops[i])) 3318 return nullptr; 3319 3320 return ConstantExpr::getGetElementPtr(cast<Constant>(Ops[0]), Ops.slice(1)); 3321 } 3322 3323 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout *DL, 3324 const TargetLibraryInfo *TLI, 3325 const DominatorTree *DT, AssumptionCache *AC, 3326 const Instruction *CxtI) { 3327 return ::SimplifyGEPInst(Ops, Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3328 } 3329 3330 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we 3331 /// can fold the result. If not, this returns null. 3332 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val, 3333 ArrayRef<unsigned> Idxs, const Query &Q, 3334 unsigned) { 3335 if (Constant *CAgg = dyn_cast<Constant>(Agg)) 3336 if (Constant *CVal = dyn_cast<Constant>(Val)) 3337 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs); 3338 3339 // insertvalue x, undef, n -> x 3340 if (match(Val, m_Undef())) 3341 return Agg; 3342 3343 // insertvalue x, (extractvalue y, n), n 3344 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val)) 3345 if (EV->getAggregateOperand()->getType() == Agg->getType() && 3346 EV->getIndices() == Idxs) { 3347 // insertvalue undef, (extractvalue y, n), n -> y 3348 if (match(Agg, m_Undef())) 3349 return EV->getAggregateOperand(); 3350 3351 // insertvalue y, (extractvalue y, n), n -> y 3352 if (Agg == EV->getAggregateOperand()) 3353 return Agg; 3354 } 3355 3356 return nullptr; 3357 } 3358 3359 Value *llvm::SimplifyInsertValueInst( 3360 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout *DL, 3361 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC, 3362 const Instruction *CxtI) { 3363 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI), 3364 RecursionLimit); 3365 } 3366 3367 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null. 3368 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) { 3369 // If all of the PHI's incoming values are the same then replace the PHI node 3370 // with the common value. 3371 Value *CommonValue = nullptr; 3372 bool HasUndefInput = false; 3373 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 3374 Value *Incoming = PN->getIncomingValue(i); 3375 // If the incoming value is the phi node itself, it can safely be skipped. 3376 if (Incoming == PN) continue; 3377 if (isa<UndefValue>(Incoming)) { 3378 // Remember that we saw an undef value, but otherwise ignore them. 3379 HasUndefInput = true; 3380 continue; 3381 } 3382 if (CommonValue && Incoming != CommonValue) 3383 return nullptr; // Not the same, bail out. 3384 CommonValue = Incoming; 3385 } 3386 3387 // If CommonValue is null then all of the incoming values were either undef or 3388 // equal to the phi node itself. 3389 if (!CommonValue) 3390 return UndefValue::get(PN->getType()); 3391 3392 // If we have a PHI node like phi(X, undef, X), where X is defined by some 3393 // instruction, we cannot return X as the result of the PHI node unless it 3394 // dominates the PHI block. 3395 if (HasUndefInput) 3396 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr; 3397 3398 return CommonValue; 3399 } 3400 3401 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) { 3402 if (Constant *C = dyn_cast<Constant>(Op)) 3403 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI); 3404 3405 return nullptr; 3406 } 3407 3408 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout *DL, 3409 const TargetLibraryInfo *TLI, 3410 const DominatorTree *DT, AssumptionCache *AC, 3411 const Instruction *CxtI) { 3412 return ::SimplifyTruncInst(Op, Ty, Query(DL, TLI, DT, AC, CxtI), 3413 RecursionLimit); 3414 } 3415 3416 //=== Helper functions for higher up the class hierarchy. 3417 3418 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can 3419 /// fold the result. If not, this returns null. 3420 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3421 const Query &Q, unsigned MaxRecurse) { 3422 switch (Opcode) { 3423 case Instruction::Add: 3424 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3425 Q, MaxRecurse); 3426 case Instruction::FAdd: 3427 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3428 3429 case Instruction::Sub: 3430 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3431 Q, MaxRecurse); 3432 case Instruction::FSub: 3433 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3434 3435 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse); 3436 case Instruction::FMul: 3437 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3438 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse); 3439 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse); 3440 case Instruction::FDiv: 3441 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3442 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse); 3443 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse); 3444 case Instruction::FRem: 3445 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3446 case Instruction::Shl: 3447 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3448 Q, MaxRecurse); 3449 case Instruction::LShr: 3450 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3451 case Instruction::AShr: 3452 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3453 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse); 3454 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse); 3455 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse); 3456 default: 3457 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 3458 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 3459 Constant *COps[] = {CLHS, CRHS}; 3460 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL, 3461 Q.TLI); 3462 } 3463 3464 // If the operation is associative, try some generic simplifications. 3465 if (Instruction::isAssociative(Opcode)) 3466 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse)) 3467 return V; 3468 3469 // If the operation is with the result of a select instruction check whether 3470 // operating on either branch of the select always yields the same value. 3471 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3472 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse)) 3473 return V; 3474 3475 // If the operation is with the result of a phi instruction, check whether 3476 // operating on all incoming values of the phi always yields the same value. 3477 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3478 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse)) 3479 return V; 3480 3481 return nullptr; 3482 } 3483 } 3484 3485 /// SimplifyFPBinOp - Given operands for a BinaryOperator, see if we can 3486 /// fold the result. If not, this returns null. 3487 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the 3488 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp. 3489 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3490 const FastMathFlags &FMF, const Query &Q, 3491 unsigned MaxRecurse) { 3492 switch (Opcode) { 3493 case Instruction::FAdd: 3494 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse); 3495 case Instruction::FSub: 3496 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse); 3497 case Instruction::FMul: 3498 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse); 3499 default: 3500 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse); 3501 } 3502 } 3503 3504 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3505 const DataLayout *DL, const TargetLibraryInfo *TLI, 3506 const DominatorTree *DT, AssumptionCache *AC, 3507 const Instruction *CxtI) { 3508 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3509 RecursionLimit); 3510 } 3511 3512 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3513 const FastMathFlags &FMF, const DataLayout *DL, 3514 const TargetLibraryInfo *TLI, 3515 const DominatorTree *DT, AssumptionCache *AC, 3516 const Instruction *CxtI) { 3517 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI), 3518 RecursionLimit); 3519 } 3520 3521 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can 3522 /// fold the result. 3523 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3524 const Query &Q, unsigned MaxRecurse) { 3525 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate)) 3526 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 3527 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 3528 } 3529 3530 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3531 const DataLayout *DL, const TargetLibraryInfo *TLI, 3532 const DominatorTree *DT, AssumptionCache *AC, 3533 const Instruction *CxtI) { 3534 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3535 RecursionLimit); 3536 } 3537 3538 static bool IsIdempotent(Intrinsic::ID ID) { 3539 switch (ID) { 3540 default: return false; 3541 3542 // Unary idempotent: f(f(x)) = f(x) 3543 case Intrinsic::fabs: 3544 case Intrinsic::floor: 3545 case Intrinsic::ceil: 3546 case Intrinsic::trunc: 3547 case Intrinsic::rint: 3548 case Intrinsic::nearbyint: 3549 case Intrinsic::round: 3550 return true; 3551 } 3552 } 3553 3554 template <typename IterTy> 3555 static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd, 3556 const Query &Q, unsigned MaxRecurse) { 3557 // Perform idempotent optimizations 3558 if (!IsIdempotent(IID)) 3559 return nullptr; 3560 3561 // Unary Ops 3562 if (std::distance(ArgBegin, ArgEnd) == 1) 3563 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) 3564 if (II->getIntrinsicID() == IID) 3565 return II; 3566 3567 return nullptr; 3568 } 3569 3570 template <typename IterTy> 3571 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd, 3572 const Query &Q, unsigned MaxRecurse) { 3573 Type *Ty = V->getType(); 3574 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 3575 Ty = PTy->getElementType(); 3576 FunctionType *FTy = cast<FunctionType>(Ty); 3577 3578 // call undef -> undef 3579 if (isa<UndefValue>(V)) 3580 return UndefValue::get(FTy->getReturnType()); 3581 3582 Function *F = dyn_cast<Function>(V); 3583 if (!F) 3584 return nullptr; 3585 3586 if (unsigned IID = F->getIntrinsicID()) 3587 if (Value *Ret = 3588 SimplifyIntrinsic((Intrinsic::ID) IID, ArgBegin, ArgEnd, Q, MaxRecurse)) 3589 return Ret; 3590 3591 if (!canConstantFoldCallTo(F)) 3592 return nullptr; 3593 3594 SmallVector<Constant *, 4> ConstantArgs; 3595 ConstantArgs.reserve(ArgEnd - ArgBegin); 3596 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) { 3597 Constant *C = dyn_cast<Constant>(*I); 3598 if (!C) 3599 return nullptr; 3600 ConstantArgs.push_back(C); 3601 } 3602 3603 return ConstantFoldCall(F, ConstantArgs, Q.TLI); 3604 } 3605 3606 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin, 3607 User::op_iterator ArgEnd, const DataLayout *DL, 3608 const TargetLibraryInfo *TLI, const DominatorTree *DT, 3609 AssumptionCache *AC, const Instruction *CxtI) { 3610 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI), 3611 RecursionLimit); 3612 } 3613 3614 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args, 3615 const DataLayout *DL, const TargetLibraryInfo *TLI, 3616 const DominatorTree *DT, AssumptionCache *AC, 3617 const Instruction *CxtI) { 3618 return ::SimplifyCall(V, Args.begin(), Args.end(), 3619 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3620 } 3621 3622 /// SimplifyInstruction - See if we can compute a simplified version of this 3623 /// instruction. If not, this returns null. 3624 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout *DL, 3625 const TargetLibraryInfo *TLI, 3626 const DominatorTree *DT, AssumptionCache *AC) { 3627 Value *Result; 3628 3629 switch (I->getOpcode()) { 3630 default: 3631 Result = ConstantFoldInstruction(I, DL, TLI); 3632 break; 3633 case Instruction::FAdd: 3634 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1), 3635 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3636 break; 3637 case Instruction::Add: 3638 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1), 3639 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3640 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3641 TLI, DT, AC, I); 3642 break; 3643 case Instruction::FSub: 3644 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1), 3645 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3646 break; 3647 case Instruction::Sub: 3648 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1), 3649 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3650 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3651 TLI, DT, AC, I); 3652 break; 3653 case Instruction::FMul: 3654 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1), 3655 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3656 break; 3657 case Instruction::Mul: 3658 Result = 3659 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3660 break; 3661 case Instruction::SDiv: 3662 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3663 AC, I); 3664 break; 3665 case Instruction::UDiv: 3666 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3667 AC, I); 3668 break; 3669 case Instruction::FDiv: 3670 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), 3671 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3672 break; 3673 case Instruction::SRem: 3674 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3675 AC, I); 3676 break; 3677 case Instruction::URem: 3678 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3679 AC, I); 3680 break; 3681 case Instruction::FRem: 3682 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), 3683 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3684 break; 3685 case Instruction::Shl: 3686 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1), 3687 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3688 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3689 TLI, DT, AC, I); 3690 break; 3691 case Instruction::LShr: 3692 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1), 3693 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 3694 AC, I); 3695 break; 3696 case Instruction::AShr: 3697 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1), 3698 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 3699 AC, I); 3700 break; 3701 case Instruction::And: 3702 Result = 3703 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3704 break; 3705 case Instruction::Or: 3706 Result = 3707 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3708 break; 3709 case Instruction::Xor: 3710 Result = 3711 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3712 break; 3713 case Instruction::ICmp: 3714 Result = 3715 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0), 3716 I->getOperand(1), DL, TLI, DT, AC, I); 3717 break; 3718 case Instruction::FCmp: 3719 Result = 3720 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0), 3721 I->getOperand(1), DL, TLI, DT, AC, I); 3722 break; 3723 case Instruction::Select: 3724 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1), 3725 I->getOperand(2), DL, TLI, DT, AC, I); 3726 break; 3727 case Instruction::GetElementPtr: { 3728 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 3729 Result = SimplifyGEPInst(Ops, DL, TLI, DT, AC, I); 3730 break; 3731 } 3732 case Instruction::InsertValue: { 3733 InsertValueInst *IV = cast<InsertValueInst>(I); 3734 Result = SimplifyInsertValueInst(IV->getAggregateOperand(), 3735 IV->getInsertedValueOperand(), 3736 IV->getIndices(), DL, TLI, DT, AC, I); 3737 break; 3738 } 3739 case Instruction::PHI: 3740 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I)); 3741 break; 3742 case Instruction::Call: { 3743 CallSite CS(cast<CallInst>(I)); 3744 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL, 3745 TLI, DT, AC, I); 3746 break; 3747 } 3748 case Instruction::Trunc: 3749 Result = 3750 SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT, AC, I); 3751 break; 3752 } 3753 3754 /// If called on unreachable code, the above logic may report that the 3755 /// instruction simplified to itself. Make life easier for users by 3756 /// detecting that case here, returning a safe value instead. 3757 return Result == I ? UndefValue::get(I->getType()) : Result; 3758 } 3759 3760 /// \brief Implementation of recursive simplification through an instructions 3761 /// uses. 3762 /// 3763 /// This is the common implementation of the recursive simplification routines. 3764 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to 3765 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of 3766 /// instructions to process and attempt to simplify it using 3767 /// InstructionSimplify. 3768 /// 3769 /// This routine returns 'true' only when *it* simplifies something. The passed 3770 /// in simplified value does not count toward this. 3771 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV, 3772 const DataLayout *DL, 3773 const TargetLibraryInfo *TLI, 3774 const DominatorTree *DT, 3775 AssumptionCache *AC) { 3776 bool Simplified = false; 3777 SmallSetVector<Instruction *, 8> Worklist; 3778 3779 // If we have an explicit value to collapse to, do that round of the 3780 // simplification loop by hand initially. 3781 if (SimpleV) { 3782 for (User *U : I->users()) 3783 if (U != I) 3784 Worklist.insert(cast<Instruction>(U)); 3785 3786 // Replace the instruction with its simplified value. 3787 I->replaceAllUsesWith(SimpleV); 3788 3789 // Gracefully handle edge cases where the instruction is not wired into any 3790 // parent block. 3791 if (I->getParent()) 3792 I->eraseFromParent(); 3793 } else { 3794 Worklist.insert(I); 3795 } 3796 3797 // Note that we must test the size on each iteration, the worklist can grow. 3798 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { 3799 I = Worklist[Idx]; 3800 3801 // See if this instruction simplifies. 3802 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC); 3803 if (!SimpleV) 3804 continue; 3805 3806 Simplified = true; 3807 3808 // Stash away all the uses of the old instruction so we can check them for 3809 // recursive simplifications after a RAUW. This is cheaper than checking all 3810 // uses of To on the recursive step in most cases. 3811 for (User *U : I->users()) 3812 Worklist.insert(cast<Instruction>(U)); 3813 3814 // Replace the instruction with its simplified value. 3815 I->replaceAllUsesWith(SimpleV); 3816 3817 // Gracefully handle edge cases where the instruction is not wired into any 3818 // parent block. 3819 if (I->getParent()) 3820 I->eraseFromParent(); 3821 } 3822 return Simplified; 3823 } 3824 3825 bool llvm::recursivelySimplifyInstruction(Instruction *I, const DataLayout *DL, 3826 const TargetLibraryInfo *TLI, 3827 const DominatorTree *DT, 3828 AssumptionCache *AC) { 3829 return replaceAndRecursivelySimplifyImpl(I, nullptr, DL, TLI, DT, AC); 3830 } 3831 3832 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV, 3833 const DataLayout *DL, 3834 const TargetLibraryInfo *TLI, 3835 const DominatorTree *DT, 3836 AssumptionCache *AC) { 3837 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!"); 3838 assert(SimpleV && "Must provide a simplified value."); 3839 return replaceAndRecursivelySimplifyImpl(I, SimpleV, DL, TLI, DT, AC); 3840 } 3841