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/Analysis/VectorUtils.h" 28 #include "llvm/IR/ConstantRange.h" 29 #include "llvm/IR/DataLayout.h" 30 #include "llvm/IR/Dominators.h" 31 #include "llvm/IR/GetElementPtrTypeIterator.h" 32 #include "llvm/IR/GlobalAlias.h" 33 #include "llvm/IR/Operator.h" 34 #include "llvm/IR/PatternMatch.h" 35 #include "llvm/IR/ValueHandle.h" 36 #include <algorithm> 37 using namespace llvm; 38 using namespace llvm::PatternMatch; 39 40 #define DEBUG_TYPE "instsimplify" 41 42 enum { RecursionLimit = 3 }; 43 44 STATISTIC(NumExpand, "Number of expansions"); 45 STATISTIC(NumReassoc, "Number of reassociations"); 46 47 namespace { 48 struct Query { 49 const DataLayout &DL; 50 const TargetLibraryInfo *TLI; 51 const DominatorTree *DT; 52 AssumptionCache *AC; 53 const Instruction *CxtI; 54 55 Query(const DataLayout &DL, const TargetLibraryInfo *tli, 56 const DominatorTree *dt, AssumptionCache *ac = nullptr, 57 const Instruction *cxti = nullptr) 58 : DL(DL), TLI(tli), DT(dt), AC(ac), CxtI(cxti) {} 59 }; 60 } // end anonymous namespace 61 62 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned); 63 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &, 64 unsigned); 65 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &, 66 const Query &, unsigned); 67 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &, 68 unsigned); 69 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned); 70 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned); 71 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned); 72 73 /// getFalse - For a boolean type, or a vector of boolean type, return false, or 74 /// a vector with every element false, as appropriate for the type. 75 static Constant *getFalse(Type *Ty) { 76 assert(Ty->getScalarType()->isIntegerTy(1) && 77 "Expected i1 type or a vector of i1!"); 78 return Constant::getNullValue(Ty); 79 } 80 81 /// getTrue - For a boolean type, or a vector of boolean type, return true, or 82 /// a vector with every element true, as appropriate for the type. 83 static Constant *getTrue(Type *Ty) { 84 assert(Ty->getScalarType()->isIntegerTy(1) && 85 "Expected i1 type or a vector of i1!"); 86 return Constant::getAllOnesValue(Ty); 87 } 88 89 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"? 90 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS, 91 Value *RHS) { 92 CmpInst *Cmp = dyn_cast<CmpInst>(V); 93 if (!Cmp) 94 return false; 95 CmpInst::Predicate CPred = Cmp->getPredicate(); 96 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1); 97 if (CPred == Pred && CLHS == LHS && CRHS == RHS) 98 return true; 99 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS && 100 CRHS == LHS; 101 } 102 103 /// ValueDominatesPHI - Does the given value dominate the specified phi node? 104 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) { 105 Instruction *I = dyn_cast<Instruction>(V); 106 if (!I) 107 // Arguments and constants dominate all instructions. 108 return true; 109 110 // If we are processing instructions (and/or basic blocks) that have not been 111 // fully added to a function, the parent nodes may still be null. Simply 112 // return the conservative answer in these cases. 113 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent()) 114 return false; 115 116 // If we have a DominatorTree then do a precise test. 117 if (DT) { 118 if (!DT->isReachableFromEntry(P->getParent())) 119 return true; 120 if (!DT->isReachableFromEntry(I->getParent())) 121 return false; 122 return DT->dominates(I, P); 123 } 124 125 // Otherwise, if the instruction is in the entry block, and is not an invoke, 126 // and is not a catchpad, then it obviously dominates all phi nodes. 127 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() && 128 !isa<InvokeInst>(I) && !isa<CatchPadInst>(I)) 129 return true; 130 131 return false; 132 } 133 134 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning 135 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is 136 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS. 137 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)". 138 /// Returns the simplified value, or null if no simplification was performed. 139 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS, 140 unsigned OpcToExpand, const Query &Q, 141 unsigned MaxRecurse) { 142 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand; 143 // Recursion is always used, so bail out at once if we already hit the limit. 144 if (!MaxRecurse--) 145 return nullptr; 146 147 // Check whether the expression has the form "(A op' B) op C". 148 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS)) 149 if (Op0->getOpcode() == OpcodeToExpand) { 150 // It does! Try turning it into "(A op C) op' (B op C)". 151 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 152 // Do "A op C" and "B op C" both simplify? 153 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) 154 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 155 // They do! Return "L op' R" if it simplifies or is already available. 156 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 157 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand) 158 && L == B && R == A)) { 159 ++NumExpand; 160 return LHS; 161 } 162 // Otherwise return "L op' R" if it simplifies. 163 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 164 ++NumExpand; 165 return V; 166 } 167 } 168 } 169 170 // Check whether the expression has the form "A op (B op' C)". 171 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS)) 172 if (Op1->getOpcode() == OpcodeToExpand) { 173 // It does! Try turning it into "(A op B) op' (A op C)". 174 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 175 // Do "A op B" and "A op C" both simplify? 176 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) 177 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) { 178 // They do! Return "L op' R" if it simplifies or is already available. 179 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 180 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand) 181 && L == C && R == B)) { 182 ++NumExpand; 183 return RHS; 184 } 185 // Otherwise return "L op' R" if it simplifies. 186 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 187 ++NumExpand; 188 return V; 189 } 190 } 191 } 192 193 return nullptr; 194 } 195 196 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary 197 /// operations. Returns the simpler value, or null if none was found. 198 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS, 199 const Query &Q, unsigned MaxRecurse) { 200 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc; 201 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!"); 202 203 // Recursion is always used, so bail out at once if we already hit the limit. 204 if (!MaxRecurse--) 205 return nullptr; 206 207 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 208 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 209 210 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely. 211 if (Op0 && Op0->getOpcode() == Opcode) { 212 Value *A = Op0->getOperand(0); 213 Value *B = Op0->getOperand(1); 214 Value *C = RHS; 215 216 // Does "B op C" simplify? 217 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 218 // It does! Return "A op V" if it simplifies or is already available. 219 // If V equals B then "A op V" is just the LHS. 220 if (V == B) return LHS; 221 // Otherwise return "A op V" if it simplifies. 222 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) { 223 ++NumReassoc; 224 return W; 225 } 226 } 227 } 228 229 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely. 230 if (Op1 && Op1->getOpcode() == Opcode) { 231 Value *A = LHS; 232 Value *B = Op1->getOperand(0); 233 Value *C = Op1->getOperand(1); 234 235 // Does "A op B" simplify? 236 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) { 237 // It does! Return "V op C" if it simplifies or is already available. 238 // If V equals B then "V op C" is just the RHS. 239 if (V == B) return RHS; 240 // Otherwise return "V op C" if it simplifies. 241 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) { 242 ++NumReassoc; 243 return W; 244 } 245 } 246 } 247 248 // The remaining transforms require commutativity as well as associativity. 249 if (!Instruction::isCommutative(Opcode)) 250 return nullptr; 251 252 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely. 253 if (Op0 && Op0->getOpcode() == Opcode) { 254 Value *A = Op0->getOperand(0); 255 Value *B = Op0->getOperand(1); 256 Value *C = RHS; 257 258 // Does "C op A" simplify? 259 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 260 // It does! Return "V op B" if it simplifies or is already available. 261 // If V equals A then "V op B" is just the LHS. 262 if (V == A) return LHS; 263 // Otherwise return "V op B" if it simplifies. 264 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) { 265 ++NumReassoc; 266 return W; 267 } 268 } 269 } 270 271 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely. 272 if (Op1 && Op1->getOpcode() == Opcode) { 273 Value *A = LHS; 274 Value *B = Op1->getOperand(0); 275 Value *C = Op1->getOperand(1); 276 277 // Does "C op A" simplify? 278 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 279 // It does! Return "B op V" if it simplifies or is already available. 280 // If V equals C then "B op V" is just the RHS. 281 if (V == C) return RHS; 282 // Otherwise return "B op V" if it simplifies. 283 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) { 284 ++NumReassoc; 285 return W; 286 } 287 } 288 } 289 290 return nullptr; 291 } 292 293 /// ThreadBinOpOverSelect - In the case of a binary operation with a select 294 /// instruction as an operand, try to simplify the binop by seeing whether 295 /// evaluating it on both branches of the select results in the same value. 296 /// Returns the common value if so, otherwise returns null. 297 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS, 298 const Query &Q, unsigned MaxRecurse) { 299 // Recursion is always used, so bail out at once if we already hit the limit. 300 if (!MaxRecurse--) 301 return nullptr; 302 303 SelectInst *SI; 304 if (isa<SelectInst>(LHS)) { 305 SI = cast<SelectInst>(LHS); 306 } else { 307 assert(isa<SelectInst>(RHS) && "No select instruction operand!"); 308 SI = cast<SelectInst>(RHS); 309 } 310 311 // Evaluate the BinOp on the true and false branches of the select. 312 Value *TV; 313 Value *FV; 314 if (SI == LHS) { 315 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse); 316 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse); 317 } else { 318 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse); 319 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse); 320 } 321 322 // If they simplified to the same value, then return the common value. 323 // If they both failed to simplify then return null. 324 if (TV == FV) 325 return TV; 326 327 // If one branch simplified to undef, return the other one. 328 if (TV && isa<UndefValue>(TV)) 329 return FV; 330 if (FV && isa<UndefValue>(FV)) 331 return TV; 332 333 // If applying the operation did not change the true and false select values, 334 // then the result of the binop is the select itself. 335 if (TV == SI->getTrueValue() && FV == SI->getFalseValue()) 336 return SI; 337 338 // If one branch simplified and the other did not, and the simplified 339 // value is equal to the unsimplified one, return the simplified value. 340 // For example, select (cond, X, X & Z) & Z -> X & Z. 341 if ((FV && !TV) || (TV && !FV)) { 342 // Check that the simplified value has the form "X op Y" where "op" is the 343 // same as the original operation. 344 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV); 345 if (Simplified && Simplified->getOpcode() == Opcode) { 346 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS". 347 // We already know that "op" is the same as for the simplified value. See 348 // if the operands match too. If so, return the simplified value. 349 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue(); 350 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS; 351 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch; 352 if (Simplified->getOperand(0) == UnsimplifiedLHS && 353 Simplified->getOperand(1) == UnsimplifiedRHS) 354 return Simplified; 355 if (Simplified->isCommutative() && 356 Simplified->getOperand(1) == UnsimplifiedLHS && 357 Simplified->getOperand(0) == UnsimplifiedRHS) 358 return Simplified; 359 } 360 } 361 362 return nullptr; 363 } 364 365 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction, 366 /// try to simplify the comparison by seeing whether both branches of the select 367 /// result in the same value. Returns the common value if so, otherwise returns 368 /// null. 369 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS, 370 Value *RHS, const Query &Q, 371 unsigned MaxRecurse) { 372 // Recursion is always used, so bail out at once if we already hit the limit. 373 if (!MaxRecurse--) 374 return nullptr; 375 376 // Make sure the select is on the LHS. 377 if (!isa<SelectInst>(LHS)) { 378 std::swap(LHS, RHS); 379 Pred = CmpInst::getSwappedPredicate(Pred); 380 } 381 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!"); 382 SelectInst *SI = cast<SelectInst>(LHS); 383 Value *Cond = SI->getCondition(); 384 Value *TV = SI->getTrueValue(); 385 Value *FV = SI->getFalseValue(); 386 387 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it. 388 // Does "cmp TV, RHS" simplify? 389 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse); 390 if (TCmp == Cond) { 391 // It not only simplified, it simplified to the select condition. Replace 392 // it with 'true'. 393 TCmp = getTrue(Cond->getType()); 394 } else if (!TCmp) { 395 // It didn't simplify. However if "cmp TV, RHS" is equal to the select 396 // condition then we can replace it with 'true'. Otherwise give up. 397 if (!isSameCompare(Cond, Pred, TV, RHS)) 398 return nullptr; 399 TCmp = getTrue(Cond->getType()); 400 } 401 402 // Does "cmp FV, RHS" simplify? 403 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse); 404 if (FCmp == Cond) { 405 // It not only simplified, it simplified to the select condition. Replace 406 // it with 'false'. 407 FCmp = getFalse(Cond->getType()); 408 } else if (!FCmp) { 409 // It didn't simplify. However if "cmp FV, RHS" is equal to the select 410 // condition then we can replace it with 'false'. Otherwise give up. 411 if (!isSameCompare(Cond, Pred, FV, RHS)) 412 return nullptr; 413 FCmp = getFalse(Cond->getType()); 414 } 415 416 // If both sides simplified to the same value, then use it as the result of 417 // the original comparison. 418 if (TCmp == FCmp) 419 return TCmp; 420 421 // The remaining cases only make sense if the select condition has the same 422 // type as the result of the comparison, so bail out if this is not so. 423 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy()) 424 return nullptr; 425 // If the false value simplified to false, then the result of the compare 426 // is equal to "Cond && TCmp". This also catches the case when the false 427 // value simplified to false and the true value to true, returning "Cond". 428 if (match(FCmp, m_Zero())) 429 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse)) 430 return V; 431 // If the true value simplified to true, then the result of the compare 432 // is equal to "Cond || FCmp". 433 if (match(TCmp, m_One())) 434 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse)) 435 return V; 436 // Finally, if the false value simplified to true and the true value to 437 // false, then the result of the compare is equal to "!Cond". 438 if (match(FCmp, m_One()) && match(TCmp, m_Zero())) 439 if (Value *V = 440 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()), 441 Q, MaxRecurse)) 442 return V; 443 444 return nullptr; 445 } 446 447 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that 448 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating 449 /// it on the incoming phi values yields the same result for every value. If so 450 /// returns the common value, otherwise returns null. 451 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS, 452 const Query &Q, unsigned MaxRecurse) { 453 // Recursion is always used, so bail out at once if we already hit the limit. 454 if (!MaxRecurse--) 455 return nullptr; 456 457 PHINode *PI; 458 if (isa<PHINode>(LHS)) { 459 PI = cast<PHINode>(LHS); 460 // Bail out if RHS and the phi may be mutually interdependent due to a loop. 461 if (!ValueDominatesPHI(RHS, PI, Q.DT)) 462 return nullptr; 463 } else { 464 assert(isa<PHINode>(RHS) && "No PHI instruction operand!"); 465 PI = cast<PHINode>(RHS); 466 // Bail out if LHS and the phi may be mutually interdependent due to a loop. 467 if (!ValueDominatesPHI(LHS, PI, Q.DT)) 468 return nullptr; 469 } 470 471 // Evaluate the BinOp on the incoming phi values. 472 Value *CommonValue = nullptr; 473 for (Value *Incoming : PI->incoming_values()) { 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 (Value *Incoming : PI->incoming_values()) { 514 // If the incoming value is the phi node itself, it can safely be skipped. 515 if (Incoming == PI) continue; 516 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse); 517 // If the operation failed to simplify, or simplified to a different value 518 // to previously, then give up. 519 if (!V || (CommonValue && V != CommonValue)) 520 return nullptr; 521 CommonValue = V; 522 } 523 524 return CommonValue; 525 } 526 527 /// SimplifyAddInst - Given operands for an Add, see if we can 528 /// fold the result. If not, this returns null. 529 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 530 const Query &Q, unsigned MaxRecurse) { 531 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 532 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 533 Constant *Ops[] = { CLHS, CRHS }; 534 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops, 535 Q.DL, Q.TLI); 536 } 537 538 // Canonicalize the constant to the RHS. 539 std::swap(Op0, Op1); 540 } 541 542 // X + undef -> undef 543 if (match(Op1, m_Undef())) 544 return Op1; 545 546 // X + 0 -> X 547 if (match(Op1, m_Zero())) 548 return Op0; 549 550 // X + (Y - X) -> Y 551 // (Y - X) + X -> Y 552 // Eg: X + -X -> 0 553 Value *Y = nullptr; 554 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) || 555 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1)))) 556 return Y; 557 558 // X + ~X -> -1 since ~X = -X-1 559 if (match(Op0, m_Not(m_Specific(Op1))) || 560 match(Op1, m_Not(m_Specific(Op0)))) 561 return Constant::getAllOnesValue(Op0->getType()); 562 563 /// i1 add -> xor. 564 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 565 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 566 return V; 567 568 // Try some generic simplifications for associative operations. 569 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q, 570 MaxRecurse)) 571 return V; 572 573 // Threading Add over selects and phi nodes is pointless, so don't bother. 574 // Threading over the select in "A + select(cond, B, C)" means evaluating 575 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and 576 // only if B and C are equal. If B and C are equal then (since we assume 577 // that operands have already been simplified) "select(cond, B, C)" should 578 // have been simplified to the common value of B and C already. Analysing 579 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly 580 // for threading over phi nodes. 581 582 return nullptr; 583 } 584 585 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 586 const DataLayout &DL, const TargetLibraryInfo *TLI, 587 const DominatorTree *DT, AssumptionCache *AC, 588 const Instruction *CxtI) { 589 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 590 RecursionLimit); 591 } 592 593 /// \brief Compute the base pointer and cumulative constant offsets for V. 594 /// 595 /// This strips all constant offsets off of V, leaving it the base pointer, and 596 /// accumulates the total constant offset applied in the returned constant. It 597 /// returns 0 if V is not a pointer, and returns the constant '0' if there are 598 /// no constant offsets applied. 599 /// 600 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't 601 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc. 602 /// folding. 603 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V, 604 bool AllowNonInbounds = false) { 605 assert(V->getType()->getScalarType()->isPointerTy()); 606 607 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType(); 608 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth()); 609 610 // Even though we don't look through PHI nodes, we could be called on an 611 // instruction in an unreachable block, which may be on a cycle. 612 SmallPtrSet<Value *, 4> Visited; 613 Visited.insert(V); 614 do { 615 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 616 if ((!AllowNonInbounds && !GEP->isInBounds()) || 617 !GEP->accumulateConstantOffset(DL, Offset)) 618 break; 619 V = GEP->getPointerOperand(); 620 } else if (Operator::getOpcode(V) == Instruction::BitCast) { 621 V = cast<Operator>(V)->getOperand(0); 622 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 623 if (GA->mayBeOverridden()) 624 break; 625 V = GA->getAliasee(); 626 } else { 627 break; 628 } 629 assert(V->getType()->getScalarType()->isPointerTy() && 630 "Unexpected operand type!"); 631 } while (Visited.insert(V).second); 632 633 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset); 634 if (V->getType()->isVectorTy()) 635 return ConstantVector::getSplat(V->getType()->getVectorNumElements(), 636 OffsetIntPtr); 637 return OffsetIntPtr; 638 } 639 640 /// \brief Compute the constant difference between two pointer values. 641 /// If the difference is not a constant, returns zero. 642 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS, 643 Value *RHS) { 644 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 645 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 646 647 // If LHS and RHS are not related via constant offsets to the same base 648 // value, there is nothing we can do here. 649 if (LHS != RHS) 650 return nullptr; 651 652 // Otherwise, the difference of LHS - RHS can be computed as: 653 // LHS - RHS 654 // = (LHSOffset + Base) - (RHSOffset + Base) 655 // = LHSOffset - RHSOffset 656 return ConstantExpr::getSub(LHSOffset, RHSOffset); 657 } 658 659 /// SimplifySubInst - Given operands for a Sub, see if we can 660 /// fold the result. If not, this returns null. 661 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 662 const Query &Q, unsigned MaxRecurse) { 663 if (Constant *CLHS = dyn_cast<Constant>(Op0)) 664 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 665 Constant *Ops[] = { CLHS, CRHS }; 666 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(), 667 Ops, Q.DL, Q.TLI); 668 } 669 670 // X - undef -> undef 671 // undef - X -> undef 672 if (match(Op0, m_Undef()) || match(Op1, m_Undef())) 673 return UndefValue::get(Op0->getType()); 674 675 // X - 0 -> X 676 if (match(Op1, m_Zero())) 677 return Op0; 678 679 // X - X -> 0 680 if (Op0 == Op1) 681 return Constant::getNullValue(Op0->getType()); 682 683 // 0 - X -> 0 if the sub is NUW. 684 if (isNUW && match(Op0, m_Zero())) 685 return Op0; 686 687 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies. 688 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X 689 Value *X = nullptr, *Y = nullptr, *Z = Op1; 690 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z 691 // See if "V === Y - Z" simplifies. 692 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1)) 693 // It does! Now see if "X + V" simplifies. 694 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) { 695 // It does, we successfully reassociated! 696 ++NumReassoc; 697 return W; 698 } 699 // See if "V === X - Z" simplifies. 700 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 701 // It does! Now see if "Y + V" simplifies. 702 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) { 703 // It does, we successfully reassociated! 704 ++NumReassoc; 705 return W; 706 } 707 } 708 709 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies. 710 // For example, X - (X + 1) -> -1 711 X = Op0; 712 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z) 713 // See if "V === X - Y" simplifies. 714 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 715 // It does! Now see if "V - Z" simplifies. 716 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) { 717 // It does, we successfully reassociated! 718 ++NumReassoc; 719 return W; 720 } 721 // See if "V === X - Z" simplifies. 722 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 723 // It does! Now see if "V - Y" simplifies. 724 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) { 725 // It does, we successfully reassociated! 726 ++NumReassoc; 727 return W; 728 } 729 } 730 731 // Z - (X - Y) -> (Z - X) + Y if everything simplifies. 732 // For example, X - (X - Y) -> Y. 733 Z = Op0; 734 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y) 735 // See if "V === Z - X" simplifies. 736 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1)) 737 // It does! Now see if "V + Y" simplifies. 738 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) { 739 // It does, we successfully reassociated! 740 ++NumReassoc; 741 return W; 742 } 743 744 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies. 745 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) && 746 match(Op1, m_Trunc(m_Value(Y)))) 747 if (X->getType() == Y->getType()) 748 // See if "V === X - Y" simplifies. 749 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 750 // It does! Now see if "trunc V" simplifies. 751 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1)) 752 // It does, return the simplified "trunc V". 753 return W; 754 755 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...). 756 if (match(Op0, m_PtrToInt(m_Value(X))) && 757 match(Op1, m_PtrToInt(m_Value(Y)))) 758 if (Constant *Result = computePointerDifference(Q.DL, X, Y)) 759 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true); 760 761 // i1 sub -> xor. 762 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 763 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 764 return V; 765 766 // Threading Sub over selects and phi nodes is pointless, so don't bother. 767 // Threading over the select in "A - select(cond, B, C)" means evaluating 768 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and 769 // only if B and C are equal. If B and C are equal then (since we assume 770 // that operands have already been simplified) "select(cond, B, C)" should 771 // have been simplified to the common value of B and C already. Analysing 772 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly 773 // for threading over phi nodes. 774 775 return nullptr; 776 } 777 778 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 779 const DataLayout &DL, const TargetLibraryInfo *TLI, 780 const DominatorTree *DT, AssumptionCache *AC, 781 const Instruction *CxtI) { 782 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 783 RecursionLimit); 784 } 785 786 /// Given operands for an FAdd, see if we can fold the result. If not, this 787 /// returns null. 788 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 789 const Query &Q, unsigned MaxRecurse) { 790 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 791 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 792 Constant *Ops[] = { CLHS, CRHS }; 793 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(), 794 Ops, Q.DL, Q.TLI); 795 } 796 797 // Canonicalize the constant to the RHS. 798 std::swap(Op0, Op1); 799 } 800 801 // fadd X, -0 ==> X 802 if (match(Op1, m_NegZero())) 803 return Op0; 804 805 // fadd X, 0 ==> X, when we know X is not -0 806 if (match(Op1, m_Zero()) && 807 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0))) 808 return Op0; 809 810 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0 811 // where nnan and ninf have to occur at least once somewhere in this 812 // expression 813 Value *SubOp = nullptr; 814 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0)))) 815 SubOp = Op1; 816 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1)))) 817 SubOp = Op0; 818 if (SubOp) { 819 Instruction *FSub = cast<Instruction>(SubOp); 820 if ((FMF.noNaNs() || FSub->hasNoNaNs()) && 821 (FMF.noInfs() || FSub->hasNoInfs())) 822 return Constant::getNullValue(Op0->getType()); 823 } 824 825 return nullptr; 826 } 827 828 /// Given operands for an FSub, see if we can fold the result. If not, this 829 /// returns null. 830 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 831 const Query &Q, unsigned MaxRecurse) { 832 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 833 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 834 Constant *Ops[] = { CLHS, CRHS }; 835 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(), 836 Ops, Q.DL, Q.TLI); 837 } 838 } 839 840 // fsub X, 0 ==> X 841 if (match(Op1, m_Zero())) 842 return Op0; 843 844 // fsub X, -0 ==> X, when we know X is not -0 845 if (match(Op1, m_NegZero()) && 846 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0))) 847 return Op0; 848 849 // fsub 0, (fsub -0.0, X) ==> X 850 Value *X; 851 if (match(Op0, m_AnyZero())) { 852 if (match(Op1, m_FSub(m_NegZero(), m_Value(X)))) 853 return X; 854 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X)))) 855 return X; 856 } 857 858 // fsub nnan x, x ==> 0.0 859 if (FMF.noNaNs() && Op0 == Op1) 860 return Constant::getNullValue(Op0->getType()); 861 862 return nullptr; 863 } 864 865 /// Given the operands for an FMul, see if we can fold the result 866 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, 867 FastMathFlags FMF, 868 const Query &Q, 869 unsigned MaxRecurse) { 870 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 871 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 872 Constant *Ops[] = { CLHS, CRHS }; 873 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(), 874 Ops, Q.DL, Q.TLI); 875 } 876 877 // Canonicalize the constant to the RHS. 878 std::swap(Op0, Op1); 879 } 880 881 // fmul X, 1.0 ==> X 882 if (match(Op1, m_FPOne())) 883 return Op0; 884 885 // fmul nnan nsz X, 0 ==> 0 886 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero())) 887 return Op1; 888 889 return nullptr; 890 } 891 892 /// SimplifyMulInst - Given operands for a Mul, see if we can 893 /// fold the result. If not, this returns null. 894 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q, 895 unsigned MaxRecurse) { 896 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 897 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 898 Constant *Ops[] = { CLHS, CRHS }; 899 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(), 900 Ops, Q.DL, Q.TLI); 901 } 902 903 // Canonicalize the constant to the RHS. 904 std::swap(Op0, Op1); 905 } 906 907 // X * undef -> 0 908 if (match(Op1, m_Undef())) 909 return Constant::getNullValue(Op0->getType()); 910 911 // X * 0 -> 0 912 if (match(Op1, m_Zero())) 913 return Op1; 914 915 // X * 1 -> X 916 if (match(Op1, m_One())) 917 return Op0; 918 919 // (X / Y) * Y -> X if the division is exact. 920 Value *X = nullptr; 921 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y 922 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y) 923 return X; 924 925 // i1 mul -> and. 926 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 927 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1)) 928 return V; 929 930 // Try some generic simplifications for associative operations. 931 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q, 932 MaxRecurse)) 933 return V; 934 935 // Mul distributes over Add. Try some generic simplifications based on this. 936 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add, 937 Q, MaxRecurse)) 938 return V; 939 940 // If the operation is with the result of a select instruction, check whether 941 // operating on either branch of the select always yields the same value. 942 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 943 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q, 944 MaxRecurse)) 945 return V; 946 947 // If the operation is with the result of a phi instruction, check whether 948 // operating on all incoming values of the phi always yields the same value. 949 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 950 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q, 951 MaxRecurse)) 952 return V; 953 954 return nullptr; 955 } 956 957 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 958 const DataLayout &DL, 959 const TargetLibraryInfo *TLI, 960 const DominatorTree *DT, AssumptionCache *AC, 961 const Instruction *CxtI) { 962 return ::SimplifyFAddInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 963 RecursionLimit); 964 } 965 966 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 967 const DataLayout &DL, 968 const TargetLibraryInfo *TLI, 969 const DominatorTree *DT, AssumptionCache *AC, 970 const Instruction *CxtI) { 971 return ::SimplifyFSubInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 972 RecursionLimit); 973 } 974 975 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF, 976 const DataLayout &DL, 977 const TargetLibraryInfo *TLI, 978 const DominatorTree *DT, AssumptionCache *AC, 979 const Instruction *CxtI) { 980 return ::SimplifyFMulInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 981 RecursionLimit); 982 } 983 984 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout &DL, 985 const TargetLibraryInfo *TLI, 986 const DominatorTree *DT, AssumptionCache *AC, 987 const Instruction *CxtI) { 988 return ::SimplifyMulInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 989 RecursionLimit); 990 } 991 992 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can 993 /// fold the result. If not, this returns null. 994 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 995 const Query &Q, unsigned MaxRecurse) { 996 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 997 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 998 Constant *Ops[] = { C0, C1 }; 999 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1000 } 1001 } 1002 1003 bool isSigned = Opcode == Instruction::SDiv; 1004 1005 // X / undef -> undef 1006 if (match(Op1, m_Undef())) 1007 return Op1; 1008 1009 // X / 0 -> undef, we don't need to preserve faults! 1010 if (match(Op1, m_Zero())) 1011 return UndefValue::get(Op1->getType()); 1012 1013 // undef / X -> 0 1014 if (match(Op0, m_Undef())) 1015 return Constant::getNullValue(Op0->getType()); 1016 1017 // 0 / X -> 0, we don't need to preserve faults! 1018 if (match(Op0, m_Zero())) 1019 return Op0; 1020 1021 // X / 1 -> X 1022 if (match(Op1, m_One())) 1023 return Op0; 1024 1025 if (Op0->getType()->isIntegerTy(1)) 1026 // It can't be division by zero, hence it must be division by one. 1027 return Op0; 1028 1029 // X / X -> 1 1030 if (Op0 == Op1) 1031 return ConstantInt::get(Op0->getType(), 1); 1032 1033 // (X * Y) / Y -> X if the multiplication does not overflow. 1034 Value *X = nullptr, *Y = nullptr; 1035 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) { 1036 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1 1037 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0); 1038 // If the Mul knows it does not overflow, then we are good to go. 1039 if ((isSigned && Mul->hasNoSignedWrap()) || 1040 (!isSigned && Mul->hasNoUnsignedWrap())) 1041 return X; 1042 // If X has the form X = A / Y then X * Y cannot overflow. 1043 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X)) 1044 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y) 1045 return X; 1046 } 1047 1048 // (X rem Y) / Y -> 0 1049 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1050 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1051 return Constant::getNullValue(Op0->getType()); 1052 1053 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow 1054 ConstantInt *C1, *C2; 1055 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) && 1056 match(Op1, m_ConstantInt(C2))) { 1057 bool Overflow; 1058 C1->getValue().umul_ov(C2->getValue(), Overflow); 1059 if (Overflow) 1060 return Constant::getNullValue(Op0->getType()); 1061 } 1062 1063 // If the operation is with the result of a select instruction, check whether 1064 // operating on either branch of the select always yields the same value. 1065 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1066 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1067 return V; 1068 1069 // If the operation is with the result of a phi instruction, check whether 1070 // operating on all incoming values of the phi always yields the same value. 1071 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1072 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1073 return V; 1074 1075 return nullptr; 1076 } 1077 1078 /// SimplifySDivInst - Given operands for an SDiv, see if we can 1079 /// fold the result. If not, this returns null. 1080 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q, 1081 unsigned MaxRecurse) { 1082 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse)) 1083 return V; 1084 1085 return nullptr; 1086 } 1087 1088 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout &DL, 1089 const TargetLibraryInfo *TLI, 1090 const DominatorTree *DT, AssumptionCache *AC, 1091 const Instruction *CxtI) { 1092 return ::SimplifySDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1093 RecursionLimit); 1094 } 1095 1096 /// SimplifyUDivInst - Given operands for a UDiv, see if we can 1097 /// fold the result. If not, this returns null. 1098 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q, 1099 unsigned MaxRecurse) { 1100 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse)) 1101 return V; 1102 1103 return nullptr; 1104 } 1105 1106 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout &DL, 1107 const TargetLibraryInfo *TLI, 1108 const DominatorTree *DT, AssumptionCache *AC, 1109 const Instruction *CxtI) { 1110 return ::SimplifyUDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1111 RecursionLimit); 1112 } 1113 1114 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1115 const Query &Q, unsigned) { 1116 // undef / X -> undef (the undef could be a snan). 1117 if (match(Op0, m_Undef())) 1118 return Op0; 1119 1120 // X / undef -> undef 1121 if (match(Op1, m_Undef())) 1122 return Op1; 1123 1124 // 0 / X -> 0 1125 // Requires that NaNs are off (X could be zero) and signed zeroes are 1126 // ignored (X could be positive or negative, so the output sign is unknown). 1127 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1128 return Op0; 1129 1130 if (FMF.noNaNs()) { 1131 // X / X -> 1.0 is legal when NaNs are ignored. 1132 if (Op0 == Op1) 1133 return ConstantFP::get(Op0->getType(), 1.0); 1134 1135 // -X / X -> -1.0 and 1136 // X / -X -> -1.0 are legal when NaNs are ignored. 1137 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored. 1138 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) && 1139 BinaryOperator::getFNegArgument(Op0) == Op1) || 1140 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) && 1141 BinaryOperator::getFNegArgument(Op1) == Op0)) 1142 return ConstantFP::get(Op0->getType(), -1.0); 1143 } 1144 1145 return nullptr; 1146 } 1147 1148 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1149 const DataLayout &DL, 1150 const TargetLibraryInfo *TLI, 1151 const DominatorTree *DT, AssumptionCache *AC, 1152 const Instruction *CxtI) { 1153 return ::SimplifyFDivInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 1154 RecursionLimit); 1155 } 1156 1157 /// SimplifyRem - Given operands for an SRem or URem, see if we can 1158 /// fold the result. If not, this returns null. 1159 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1160 const Query &Q, unsigned MaxRecurse) { 1161 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1162 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1163 Constant *Ops[] = { C0, C1 }; 1164 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1165 } 1166 } 1167 1168 // X % undef -> undef 1169 if (match(Op1, m_Undef())) 1170 return Op1; 1171 1172 // undef % X -> 0 1173 if (match(Op0, m_Undef())) 1174 return Constant::getNullValue(Op0->getType()); 1175 1176 // 0 % X -> 0, we don't need to preserve faults! 1177 if (match(Op0, m_Zero())) 1178 return Op0; 1179 1180 // X % 0 -> undef, we don't need to preserve faults! 1181 if (match(Op1, m_Zero())) 1182 return UndefValue::get(Op0->getType()); 1183 1184 // X % 1 -> 0 1185 if (match(Op1, m_One())) 1186 return Constant::getNullValue(Op0->getType()); 1187 1188 if (Op0->getType()->isIntegerTy(1)) 1189 // It can't be remainder by zero, hence it must be remainder by one. 1190 return Constant::getNullValue(Op0->getType()); 1191 1192 // X % X -> 0 1193 if (Op0 == Op1) 1194 return Constant::getNullValue(Op0->getType()); 1195 1196 // (X % Y) % Y -> X % Y 1197 if ((Opcode == Instruction::SRem && 1198 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1199 (Opcode == Instruction::URem && 1200 match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1201 return Op0; 1202 1203 // If the operation is with the result of a select instruction, check whether 1204 // operating on either branch of the select always yields the same value. 1205 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1206 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1207 return V; 1208 1209 // If the operation is with the result of a phi instruction, check whether 1210 // operating on all incoming values of the phi always yields the same value. 1211 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1212 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1213 return V; 1214 1215 return nullptr; 1216 } 1217 1218 /// SimplifySRemInst - Given operands for an SRem, see if we can 1219 /// fold the result. If not, this returns null. 1220 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q, 1221 unsigned MaxRecurse) { 1222 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse)) 1223 return V; 1224 1225 return nullptr; 1226 } 1227 1228 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout &DL, 1229 const TargetLibraryInfo *TLI, 1230 const DominatorTree *DT, AssumptionCache *AC, 1231 const Instruction *CxtI) { 1232 return ::SimplifySRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1233 RecursionLimit); 1234 } 1235 1236 /// SimplifyURemInst - Given operands for a URem, see if we can 1237 /// fold the result. If not, this returns null. 1238 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q, 1239 unsigned MaxRecurse) { 1240 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse)) 1241 return V; 1242 1243 return nullptr; 1244 } 1245 1246 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout &DL, 1247 const TargetLibraryInfo *TLI, 1248 const DominatorTree *DT, AssumptionCache *AC, 1249 const Instruction *CxtI) { 1250 return ::SimplifyURemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1251 RecursionLimit); 1252 } 1253 1254 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1255 const Query &, unsigned) { 1256 // undef % X -> undef (the undef could be a snan). 1257 if (match(Op0, m_Undef())) 1258 return Op0; 1259 1260 // X % undef -> undef 1261 if (match(Op1, m_Undef())) 1262 return Op1; 1263 1264 // 0 % X -> 0 1265 // Requires that NaNs are off (X could be zero) and signed zeroes are 1266 // ignored (X could be positive or negative, so the output sign is unknown). 1267 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1268 return Op0; 1269 1270 return nullptr; 1271 } 1272 1273 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1274 const DataLayout &DL, 1275 const TargetLibraryInfo *TLI, 1276 const DominatorTree *DT, AssumptionCache *AC, 1277 const Instruction *CxtI) { 1278 return ::SimplifyFRemInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 1279 RecursionLimit); 1280 } 1281 1282 /// isUndefShift - Returns true if a shift by \c Amount always yields undef. 1283 static bool isUndefShift(Value *Amount) { 1284 Constant *C = dyn_cast<Constant>(Amount); 1285 if (!C) 1286 return false; 1287 1288 // X shift by undef -> undef because it may shift by the bitwidth. 1289 if (isa<UndefValue>(C)) 1290 return true; 1291 1292 // Shifting by the bitwidth or more is undefined. 1293 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) 1294 if (CI->getValue().getLimitedValue() >= 1295 CI->getType()->getScalarSizeInBits()) 1296 return true; 1297 1298 // If all lanes of a vector shift are undefined the whole shift is. 1299 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) { 1300 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I) 1301 if (!isUndefShift(C->getAggregateElement(I))) 1302 return false; 1303 return true; 1304 } 1305 1306 return false; 1307 } 1308 1309 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can 1310 /// fold the result. If not, this returns null. 1311 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1, 1312 const Query &Q, unsigned MaxRecurse) { 1313 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1314 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1315 Constant *Ops[] = { C0, C1 }; 1316 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1317 } 1318 } 1319 1320 // 0 shift by X -> 0 1321 if (match(Op0, m_Zero())) 1322 return Op0; 1323 1324 // X shift by 0 -> X 1325 if (match(Op1, m_Zero())) 1326 return Op0; 1327 1328 // Fold undefined shifts. 1329 if (isUndefShift(Op1)) 1330 return UndefValue::get(Op0->getType()); 1331 1332 // If the operation is with the result of a select instruction, check whether 1333 // operating on either branch of the select always yields the same value. 1334 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1335 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1336 return V; 1337 1338 // If the operation is with the result of a phi instruction, check whether 1339 // operating on all incoming values of the phi always yields the same value. 1340 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1341 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1342 return V; 1343 1344 return nullptr; 1345 } 1346 1347 /// \brief Given operands for an Shl, LShr or AShr, see if we can 1348 /// fold the result. If not, this returns null. 1349 static Value *SimplifyRightShift(unsigned Opcode, Value *Op0, Value *Op1, 1350 bool isExact, const Query &Q, 1351 unsigned MaxRecurse) { 1352 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse)) 1353 return V; 1354 1355 // X >> X -> 0 1356 if (Op0 == Op1) 1357 return Constant::getNullValue(Op0->getType()); 1358 1359 // undef >> X -> 0 1360 // undef >> X -> undef (if it's exact) 1361 if (match(Op0, m_Undef())) 1362 return isExact ? Op0 : Constant::getNullValue(Op0->getType()); 1363 1364 // The low bit cannot be shifted out of an exact shift if it is set. 1365 if (isExact) { 1366 unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); 1367 APInt Op0KnownZero(BitWidth, 0); 1368 APInt Op0KnownOne(BitWidth, 0); 1369 computeKnownBits(Op0, Op0KnownZero, Op0KnownOne, Q.DL, /*Depth=*/0, Q.AC, 1370 Q.CxtI, Q.DT); 1371 if (Op0KnownOne[0]) 1372 return Op0; 1373 } 1374 1375 return nullptr; 1376 } 1377 1378 /// SimplifyShlInst - Given operands for an Shl, see if we can 1379 /// fold the result. If not, this returns null. 1380 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1381 const Query &Q, unsigned MaxRecurse) { 1382 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse)) 1383 return V; 1384 1385 // undef << X -> 0 1386 // undef << X -> undef if (if it's NSW/NUW) 1387 if (match(Op0, m_Undef())) 1388 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType()); 1389 1390 // (X >> A) << A -> X 1391 Value *X; 1392 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1))))) 1393 return X; 1394 return nullptr; 1395 } 1396 1397 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1398 const DataLayout &DL, const TargetLibraryInfo *TLI, 1399 const DominatorTree *DT, AssumptionCache *AC, 1400 const Instruction *CxtI) { 1401 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 1402 RecursionLimit); 1403 } 1404 1405 /// SimplifyLShrInst - Given operands for an LShr, see if we can 1406 /// fold the result. If not, this returns null. 1407 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1408 const Query &Q, unsigned MaxRecurse) { 1409 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q, 1410 MaxRecurse)) 1411 return V; 1412 1413 // (X << A) >> A -> X 1414 Value *X; 1415 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1)))) 1416 return X; 1417 1418 return nullptr; 1419 } 1420 1421 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1422 const DataLayout &DL, 1423 const TargetLibraryInfo *TLI, 1424 const DominatorTree *DT, AssumptionCache *AC, 1425 const Instruction *CxtI) { 1426 return ::SimplifyLShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1427 RecursionLimit); 1428 } 1429 1430 /// SimplifyAShrInst - Given operands for an AShr, see if we can 1431 /// fold the result. If not, this returns null. 1432 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1433 const Query &Q, unsigned MaxRecurse) { 1434 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q, 1435 MaxRecurse)) 1436 return V; 1437 1438 // all ones >>a X -> all ones 1439 if (match(Op0, m_AllOnes())) 1440 return Op0; 1441 1442 // (X << A) >> A -> X 1443 Value *X; 1444 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1)))) 1445 return X; 1446 1447 // Arithmetic shifting an all-sign-bit value is a no-op. 1448 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 1449 if (NumSignBits == Op0->getType()->getScalarSizeInBits()) 1450 return Op0; 1451 1452 return nullptr; 1453 } 1454 1455 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1456 const DataLayout &DL, 1457 const TargetLibraryInfo *TLI, 1458 const DominatorTree *DT, AssumptionCache *AC, 1459 const Instruction *CxtI) { 1460 return ::SimplifyAShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1461 RecursionLimit); 1462 } 1463 1464 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp, 1465 ICmpInst *UnsignedICmp, bool IsAnd) { 1466 Value *X, *Y; 1467 1468 ICmpInst::Predicate EqPred; 1469 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) || 1470 !ICmpInst::isEquality(EqPred)) 1471 return nullptr; 1472 1473 ICmpInst::Predicate UnsignedPred; 1474 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) && 1475 ICmpInst::isUnsigned(UnsignedPred)) 1476 ; 1477 else if (match(UnsignedICmp, 1478 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) && 1479 ICmpInst::isUnsigned(UnsignedPred)) 1480 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred); 1481 else 1482 return nullptr; 1483 1484 // X < Y && Y != 0 --> X < Y 1485 // X < Y || Y != 0 --> Y != 0 1486 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE) 1487 return IsAnd ? UnsignedICmp : ZeroICmp; 1488 1489 // X >= Y || Y != 0 --> true 1490 // X >= Y || Y == 0 --> X >= Y 1491 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) { 1492 if (EqPred == ICmpInst::ICMP_NE) 1493 return getTrue(UnsignedICmp->getType()); 1494 return UnsignedICmp; 1495 } 1496 1497 // X < Y && Y == 0 --> false 1498 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ && 1499 IsAnd) 1500 return getFalse(UnsignedICmp->getType()); 1501 1502 return nullptr; 1503 } 1504 1505 // Simplify (and (icmp ...) (icmp ...)) to true when we can tell that the range 1506 // of possible values cannot be satisfied. 1507 static Value *SimplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1508 ICmpInst::Predicate Pred0, Pred1; 1509 ConstantInt *CI1, *CI2; 1510 Value *V; 1511 1512 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true)) 1513 return X; 1514 1515 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)), 1516 m_ConstantInt(CI2)))) 1517 return nullptr; 1518 1519 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1)))) 1520 return nullptr; 1521 1522 Type *ITy = Op0->getType(); 1523 1524 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1525 bool isNSW = AddInst->hasNoSignedWrap(); 1526 bool isNUW = AddInst->hasNoUnsignedWrap(); 1527 1528 const APInt &CI1V = CI1->getValue(); 1529 const APInt &CI2V = CI2->getValue(); 1530 const APInt Delta = CI2V - CI1V; 1531 if (CI1V.isStrictlyPositive()) { 1532 if (Delta == 2) { 1533 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT) 1534 return getFalse(ITy); 1535 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1536 return getFalse(ITy); 1537 } 1538 if (Delta == 1) { 1539 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT) 1540 return getFalse(ITy); 1541 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1542 return getFalse(ITy); 1543 } 1544 } 1545 if (CI1V.getBoolValue() && isNUW) { 1546 if (Delta == 2) 1547 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT) 1548 return getFalse(ITy); 1549 if (Delta == 1) 1550 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT) 1551 return getFalse(ITy); 1552 } 1553 1554 return nullptr; 1555 } 1556 1557 /// SimplifyAndInst - Given operands for an And, see if we can 1558 /// fold the result. If not, this returns null. 1559 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q, 1560 unsigned MaxRecurse) { 1561 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1562 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1563 Constant *Ops[] = { CLHS, CRHS }; 1564 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(), 1565 Ops, Q.DL, Q.TLI); 1566 } 1567 1568 // Canonicalize the constant to the RHS. 1569 std::swap(Op0, Op1); 1570 } 1571 1572 // X & undef -> 0 1573 if (match(Op1, m_Undef())) 1574 return Constant::getNullValue(Op0->getType()); 1575 1576 // X & X = X 1577 if (Op0 == Op1) 1578 return Op0; 1579 1580 // X & 0 = 0 1581 if (match(Op1, m_Zero())) 1582 return Op1; 1583 1584 // X & -1 = X 1585 if (match(Op1, m_AllOnes())) 1586 return Op0; 1587 1588 // A & ~A = ~A & A = 0 1589 if (match(Op0, m_Not(m_Specific(Op1))) || 1590 match(Op1, m_Not(m_Specific(Op0)))) 1591 return Constant::getNullValue(Op0->getType()); 1592 1593 // (A | ?) & A = A 1594 Value *A = nullptr, *B = nullptr; 1595 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1596 (A == Op1 || B == Op1)) 1597 return Op1; 1598 1599 // A & (A | ?) = A 1600 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1601 (A == Op0 || B == Op0)) 1602 return Op0; 1603 1604 // A & (-A) = A if A is a power of two or zero. 1605 if (match(Op0, m_Neg(m_Specific(Op1))) || 1606 match(Op1, m_Neg(m_Specific(Op0)))) { 1607 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, 1608 Q.DT)) 1609 return Op0; 1610 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, 1611 Q.DT)) 1612 return Op1; 1613 } 1614 1615 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1616 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1617 if (Value *V = SimplifyAndOfICmps(ICILHS, ICIRHS)) 1618 return V; 1619 if (Value *V = SimplifyAndOfICmps(ICIRHS, ICILHS)) 1620 return V; 1621 } 1622 } 1623 1624 // Try some generic simplifications for associative operations. 1625 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q, 1626 MaxRecurse)) 1627 return V; 1628 1629 // And distributes over Or. Try some generic simplifications based on this. 1630 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or, 1631 Q, MaxRecurse)) 1632 return V; 1633 1634 // And distributes over Xor. Try some generic simplifications based on this. 1635 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor, 1636 Q, MaxRecurse)) 1637 return V; 1638 1639 // If the operation is with the result of a select instruction, check whether 1640 // operating on either branch of the select always yields the same value. 1641 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1642 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q, 1643 MaxRecurse)) 1644 return V; 1645 1646 // If the operation is with the result of a phi instruction, check whether 1647 // operating on all incoming values of the phi always yields the same value. 1648 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1649 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q, 1650 MaxRecurse)) 1651 return V; 1652 1653 return nullptr; 1654 } 1655 1656 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout &DL, 1657 const TargetLibraryInfo *TLI, 1658 const DominatorTree *DT, AssumptionCache *AC, 1659 const Instruction *CxtI) { 1660 return ::SimplifyAndInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1661 RecursionLimit); 1662 } 1663 1664 // Simplify (or (icmp ...) (icmp ...)) to true when we can tell that the union 1665 // contains all possible values. 1666 static Value *SimplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1667 ICmpInst::Predicate Pred0, Pred1; 1668 ConstantInt *CI1, *CI2; 1669 Value *V; 1670 1671 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false)) 1672 return X; 1673 1674 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)), 1675 m_ConstantInt(CI2)))) 1676 return nullptr; 1677 1678 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1)))) 1679 return nullptr; 1680 1681 Type *ITy = Op0->getType(); 1682 1683 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1684 bool isNSW = AddInst->hasNoSignedWrap(); 1685 bool isNUW = AddInst->hasNoUnsignedWrap(); 1686 1687 const APInt &CI1V = CI1->getValue(); 1688 const APInt &CI2V = CI2->getValue(); 1689 const APInt Delta = CI2V - CI1V; 1690 if (CI1V.isStrictlyPositive()) { 1691 if (Delta == 2) { 1692 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE) 1693 return getTrue(ITy); 1694 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1695 return getTrue(ITy); 1696 } 1697 if (Delta == 1) { 1698 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE) 1699 return getTrue(ITy); 1700 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1701 return getTrue(ITy); 1702 } 1703 } 1704 if (CI1V.getBoolValue() && isNUW) { 1705 if (Delta == 2) 1706 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE) 1707 return getTrue(ITy); 1708 if (Delta == 1) 1709 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE) 1710 return getTrue(ITy); 1711 } 1712 1713 return nullptr; 1714 } 1715 1716 /// SimplifyOrInst - Given operands for an Or, see if we can 1717 /// fold the result. If not, this returns null. 1718 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q, 1719 unsigned MaxRecurse) { 1720 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1721 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1722 Constant *Ops[] = { CLHS, CRHS }; 1723 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(), 1724 Ops, Q.DL, Q.TLI); 1725 } 1726 1727 // Canonicalize the constant to the RHS. 1728 std::swap(Op0, Op1); 1729 } 1730 1731 // X | undef -> -1 1732 if (match(Op1, m_Undef())) 1733 return Constant::getAllOnesValue(Op0->getType()); 1734 1735 // X | X = X 1736 if (Op0 == Op1) 1737 return Op0; 1738 1739 // X | 0 = X 1740 if (match(Op1, m_Zero())) 1741 return Op0; 1742 1743 // X | -1 = -1 1744 if (match(Op1, m_AllOnes())) 1745 return Op1; 1746 1747 // A | ~A = ~A | A = -1 1748 if (match(Op0, m_Not(m_Specific(Op1))) || 1749 match(Op1, m_Not(m_Specific(Op0)))) 1750 return Constant::getAllOnesValue(Op0->getType()); 1751 1752 // (A & ?) | A = A 1753 Value *A = nullptr, *B = nullptr; 1754 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 1755 (A == Op1 || B == Op1)) 1756 return Op1; 1757 1758 // A | (A & ?) = A 1759 if (match(Op1, m_And(m_Value(A), m_Value(B))) && 1760 (A == Op0 || B == Op0)) 1761 return Op0; 1762 1763 // ~(A & ?) | A = -1 1764 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) && 1765 (A == Op1 || B == Op1)) 1766 return Constant::getAllOnesValue(Op1->getType()); 1767 1768 // A | ~(A & ?) = -1 1769 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) && 1770 (A == Op0 || B == Op0)) 1771 return Constant::getAllOnesValue(Op0->getType()); 1772 1773 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1774 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1775 if (Value *V = SimplifyOrOfICmps(ICILHS, ICIRHS)) 1776 return V; 1777 if (Value *V = SimplifyOrOfICmps(ICIRHS, ICILHS)) 1778 return V; 1779 } 1780 } 1781 1782 // Try some generic simplifications for associative operations. 1783 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q, 1784 MaxRecurse)) 1785 return V; 1786 1787 // Or distributes over And. Try some generic simplifications based on this. 1788 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q, 1789 MaxRecurse)) 1790 return V; 1791 1792 // If the operation is with the result of a select instruction, check whether 1793 // operating on either branch of the select always yields the same value. 1794 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1795 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q, 1796 MaxRecurse)) 1797 return V; 1798 1799 // (A & C)|(B & D) 1800 Value *C = nullptr, *D = nullptr; 1801 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 1802 match(Op1, m_And(m_Value(B), m_Value(D)))) { 1803 ConstantInt *C1 = dyn_cast<ConstantInt>(C); 1804 ConstantInt *C2 = dyn_cast<ConstantInt>(D); 1805 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) { 1806 // (A & C1)|(B & C2) 1807 // If we have: ((V + N) & C1) | (V & C2) 1808 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0 1809 // replace with V+N. 1810 Value *V1, *V2; 1811 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+ 1812 match(A, m_Add(m_Value(V1), m_Value(V2)))) { 1813 // Add commutes, try both ways. 1814 if (V1 == B && 1815 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1816 return A; 1817 if (V2 == B && 1818 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1819 return A; 1820 } 1821 // Or commutes, try both ways. 1822 if ((C1->getValue() & (C1->getValue() + 1)) == 0 && 1823 match(B, m_Add(m_Value(V1), m_Value(V2)))) { 1824 // Add commutes, try both ways. 1825 if (V1 == A && 1826 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1827 return B; 1828 if (V2 == A && 1829 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1830 return B; 1831 } 1832 } 1833 } 1834 1835 // If the operation is with the result of a phi instruction, check whether 1836 // operating on all incoming values of the phi always yields the same value. 1837 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1838 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse)) 1839 return V; 1840 1841 return nullptr; 1842 } 1843 1844 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout &DL, 1845 const TargetLibraryInfo *TLI, 1846 const DominatorTree *DT, AssumptionCache *AC, 1847 const Instruction *CxtI) { 1848 return ::SimplifyOrInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1849 RecursionLimit); 1850 } 1851 1852 /// SimplifyXorInst - Given operands for a Xor, see if we can 1853 /// fold the result. If not, this returns null. 1854 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q, 1855 unsigned MaxRecurse) { 1856 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1857 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1858 Constant *Ops[] = { CLHS, CRHS }; 1859 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(), 1860 Ops, Q.DL, Q.TLI); 1861 } 1862 1863 // Canonicalize the constant to the RHS. 1864 std::swap(Op0, Op1); 1865 } 1866 1867 // A ^ undef -> undef 1868 if (match(Op1, m_Undef())) 1869 return Op1; 1870 1871 // A ^ 0 = A 1872 if (match(Op1, m_Zero())) 1873 return Op0; 1874 1875 // A ^ A = 0 1876 if (Op0 == Op1) 1877 return Constant::getNullValue(Op0->getType()); 1878 1879 // A ^ ~A = ~A ^ A = -1 1880 if (match(Op0, m_Not(m_Specific(Op1))) || 1881 match(Op1, m_Not(m_Specific(Op0)))) 1882 return Constant::getAllOnesValue(Op0->getType()); 1883 1884 // Try some generic simplifications for associative operations. 1885 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q, 1886 MaxRecurse)) 1887 return V; 1888 1889 // Threading Xor over selects and phi nodes is pointless, so don't bother. 1890 // Threading over the select in "A ^ select(cond, B, C)" means evaluating 1891 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and 1892 // only if B and C are equal. If B and C are equal then (since we assume 1893 // that operands have already been simplified) "select(cond, B, C)" should 1894 // have been simplified to the common value of B and C already. Analysing 1895 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly 1896 // for threading over phi nodes. 1897 1898 return nullptr; 1899 } 1900 1901 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout &DL, 1902 const TargetLibraryInfo *TLI, 1903 const DominatorTree *DT, AssumptionCache *AC, 1904 const Instruction *CxtI) { 1905 return ::SimplifyXorInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1906 RecursionLimit); 1907 } 1908 1909 static Type *GetCompareTy(Value *Op) { 1910 return CmpInst::makeCmpResultType(Op->getType()); 1911 } 1912 1913 /// ExtractEquivalentCondition - Rummage around inside V looking for something 1914 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found, 1915 /// otherwise return null. Helper function for analyzing max/min idioms. 1916 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred, 1917 Value *LHS, Value *RHS) { 1918 SelectInst *SI = dyn_cast<SelectInst>(V); 1919 if (!SI) 1920 return nullptr; 1921 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); 1922 if (!Cmp) 1923 return nullptr; 1924 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1); 1925 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS) 1926 return Cmp; 1927 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) && 1928 LHS == CmpRHS && RHS == CmpLHS) 1929 return Cmp; 1930 return nullptr; 1931 } 1932 1933 // A significant optimization not implemented here is assuming that alloca 1934 // addresses are not equal to incoming argument values. They don't *alias*, 1935 // as we say, but that doesn't mean they aren't equal, so we take a 1936 // conservative approach. 1937 // 1938 // This is inspired in part by C++11 5.10p1: 1939 // "Two pointers of the same type compare equal if and only if they are both 1940 // null, both point to the same function, or both represent the same 1941 // address." 1942 // 1943 // This is pretty permissive. 1944 // 1945 // It's also partly due to C11 6.5.9p6: 1946 // "Two pointers compare equal if and only if both are null pointers, both are 1947 // pointers to the same object (including a pointer to an object and a 1948 // subobject at its beginning) or function, both are pointers to one past the 1949 // last element of the same array object, or one is a pointer to one past the 1950 // end of one array object and the other is a pointer to the start of a 1951 // different array object that happens to immediately follow the first array 1952 // object in the address space.) 1953 // 1954 // C11's version is more restrictive, however there's no reason why an argument 1955 // couldn't be a one-past-the-end value for a stack object in the caller and be 1956 // equal to the beginning of a stack object in the callee. 1957 // 1958 // If the C and C++ standards are ever made sufficiently restrictive in this 1959 // area, it may be possible to update LLVM's semantics accordingly and reinstate 1960 // this optimization. 1961 static Constant *computePointerICmp(const DataLayout &DL, 1962 const TargetLibraryInfo *TLI, 1963 CmpInst::Predicate Pred, Value *LHS, 1964 Value *RHS) { 1965 // First, skip past any trivial no-ops. 1966 LHS = LHS->stripPointerCasts(); 1967 RHS = RHS->stripPointerCasts(); 1968 1969 // A non-null pointer is not equal to a null pointer. 1970 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) && 1971 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE)) 1972 return ConstantInt::get(GetCompareTy(LHS), 1973 !CmpInst::isTrueWhenEqual(Pred)); 1974 1975 // We can only fold certain predicates on pointer comparisons. 1976 switch (Pred) { 1977 default: 1978 return nullptr; 1979 1980 // Equality comaprisons are easy to fold. 1981 case CmpInst::ICMP_EQ: 1982 case CmpInst::ICMP_NE: 1983 break; 1984 1985 // We can only handle unsigned relational comparisons because 'inbounds' on 1986 // a GEP only protects against unsigned wrapping. 1987 case CmpInst::ICMP_UGT: 1988 case CmpInst::ICMP_UGE: 1989 case CmpInst::ICMP_ULT: 1990 case CmpInst::ICMP_ULE: 1991 // However, we have to switch them to their signed variants to handle 1992 // negative indices from the base pointer. 1993 Pred = ICmpInst::getSignedPredicate(Pred); 1994 break; 1995 } 1996 1997 // Strip off any constant offsets so that we can reason about them. 1998 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets 1999 // here and compare base addresses like AliasAnalysis does, however there are 2000 // numerous hazards. AliasAnalysis and its utilities rely on special rules 2001 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis 2002 // doesn't need to guarantee pointer inequality when it says NoAlias. 2003 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 2004 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 2005 2006 // If LHS and RHS are related via constant offsets to the same base 2007 // value, we can replace it with an icmp which just compares the offsets. 2008 if (LHS == RHS) 2009 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset); 2010 2011 // Various optimizations for (in)equality comparisons. 2012 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) { 2013 // Different non-empty allocations that exist at the same time have 2014 // different addresses (if the program can tell). Global variables always 2015 // exist, so they always exist during the lifetime of each other and all 2016 // allocas. Two different allocas usually have different addresses... 2017 // 2018 // However, if there's an @llvm.stackrestore dynamically in between two 2019 // allocas, they may have the same address. It's tempting to reduce the 2020 // scope of the problem by only looking at *static* allocas here. That would 2021 // cover the majority of allocas while significantly reducing the likelihood 2022 // of having an @llvm.stackrestore pop up in the middle. However, it's not 2023 // actually impossible for an @llvm.stackrestore to pop up in the middle of 2024 // an entry block. Also, if we have a block that's not attached to a 2025 // function, we can't tell if it's "static" under the current definition. 2026 // Theoretically, this problem could be fixed by creating a new kind of 2027 // instruction kind specifically for static allocas. Such a new instruction 2028 // could be required to be at the top of the entry block, thus preventing it 2029 // from being subject to a @llvm.stackrestore. Instcombine could even 2030 // convert regular allocas into these special allocas. It'd be nifty. 2031 // However, until then, this problem remains open. 2032 // 2033 // So, we'll assume that two non-empty allocas have different addresses 2034 // for now. 2035 // 2036 // With all that, if the offsets are within the bounds of their allocations 2037 // (and not one-past-the-end! so we can't use inbounds!), and their 2038 // allocations aren't the same, the pointers are not equal. 2039 // 2040 // Note that it's not necessary to check for LHS being a global variable 2041 // address, due to canonicalization and constant folding. 2042 if (isa<AllocaInst>(LHS) && 2043 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) { 2044 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset); 2045 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset); 2046 uint64_t LHSSize, RHSSize; 2047 if (LHSOffsetCI && RHSOffsetCI && 2048 getObjectSize(LHS, LHSSize, DL, TLI) && 2049 getObjectSize(RHS, RHSSize, DL, TLI)) { 2050 const APInt &LHSOffsetValue = LHSOffsetCI->getValue(); 2051 const APInt &RHSOffsetValue = RHSOffsetCI->getValue(); 2052 if (!LHSOffsetValue.isNegative() && 2053 !RHSOffsetValue.isNegative() && 2054 LHSOffsetValue.ult(LHSSize) && 2055 RHSOffsetValue.ult(RHSSize)) { 2056 return ConstantInt::get(GetCompareTy(LHS), 2057 !CmpInst::isTrueWhenEqual(Pred)); 2058 } 2059 } 2060 2061 // Repeat the above check but this time without depending on DataLayout 2062 // or being able to compute a precise size. 2063 if (!cast<PointerType>(LHS->getType())->isEmptyTy() && 2064 !cast<PointerType>(RHS->getType())->isEmptyTy() && 2065 LHSOffset->isNullValue() && 2066 RHSOffset->isNullValue()) 2067 return ConstantInt::get(GetCompareTy(LHS), 2068 !CmpInst::isTrueWhenEqual(Pred)); 2069 } 2070 2071 // Even if an non-inbounds GEP occurs along the path we can still optimize 2072 // equality comparisons concerning the result. We avoid walking the whole 2073 // chain again by starting where the last calls to 2074 // stripAndComputeConstantOffsets left off and accumulate the offsets. 2075 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true); 2076 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true); 2077 if (LHS == RHS) 2078 return ConstantExpr::getICmp(Pred, 2079 ConstantExpr::getAdd(LHSOffset, LHSNoBound), 2080 ConstantExpr::getAdd(RHSOffset, RHSNoBound)); 2081 2082 // If one side of the equality comparison must come from a noalias call 2083 // (meaning a system memory allocation function), and the other side must 2084 // come from a pointer that cannot overlap with dynamically-allocated 2085 // memory within the lifetime of the current function (allocas, byval 2086 // arguments, globals), then determine the comparison result here. 2087 SmallVector<Value *, 8> LHSUObjs, RHSUObjs; 2088 GetUnderlyingObjects(LHS, LHSUObjs, DL); 2089 GetUnderlyingObjects(RHS, RHSUObjs, DL); 2090 2091 // Is the set of underlying objects all noalias calls? 2092 auto IsNAC = [](SmallVectorImpl<Value *> &Objects) { 2093 return std::all_of(Objects.begin(), Objects.end(), 2094 [](Value *V){ return isNoAliasCall(V); }); 2095 }; 2096 2097 // Is the set of underlying objects all things which must be disjoint from 2098 // noalias calls. For allocas, we consider only static ones (dynamic 2099 // allocas might be transformed into calls to malloc not simultaneously 2100 // live with the compared-to allocation). For globals, we exclude symbols 2101 // that might be resolve lazily to symbols in another dynamically-loaded 2102 // library (and, thus, could be malloc'ed by the implementation). 2103 auto IsAllocDisjoint = [](SmallVectorImpl<Value *> &Objects) { 2104 return std::all_of(Objects.begin(), Objects.end(), 2105 [](Value *V){ 2106 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) 2107 return AI->getParent() && AI->getParent()->getParent() && 2108 AI->isStaticAlloca(); 2109 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 2110 return (GV->hasLocalLinkage() || 2111 GV->hasHiddenVisibility() || 2112 GV->hasProtectedVisibility() || 2113 GV->hasUnnamedAddr()) && 2114 !GV->isThreadLocal(); 2115 if (const Argument *A = dyn_cast<Argument>(V)) 2116 return A->hasByValAttr(); 2117 return false; 2118 }); 2119 }; 2120 2121 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) || 2122 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs))) 2123 return ConstantInt::get(GetCompareTy(LHS), 2124 !CmpInst::isTrueWhenEqual(Pred)); 2125 } 2126 2127 // Otherwise, fail. 2128 return nullptr; 2129 } 2130 2131 /// Return true if B is known to be implied by A. A & B must be i1 (boolean) 2132 /// values or a vector of such values. Note that the truth table for 2133 /// implication is the same as <=u on i1 values (but not <=s!). The truth 2134 /// table for both is: 2135 /// | T | F (B) 2136 /// T | T | F 2137 /// F | T | T 2138 /// (A) 2139 static bool implies(Value *A, Value *B) { 2140 assert(A->getType() == B->getType() && "mismatched type"); 2141 Type *OpTy = A->getType(); 2142 assert(OpTy->getScalarType()->isIntegerTy(1)); 2143 2144 // A ==> A by definition 2145 if (A == B) return true; 2146 2147 if (OpTy->isVectorTy()) 2148 // TODO: extending the code below to handle vectors 2149 return false; 2150 assert(OpTy->isIntegerTy(1) && "implied by above"); 2151 2152 ICmpInst::Predicate APred, BPred; 2153 Value *I; 2154 Value *L; 2155 ConstantInt *CI; 2156 // i +_{nsw} C_{>0} <s L ==> i <s L 2157 if (match(A, m_ICmp(APred, 2158 m_NSWAdd(m_Value(I), m_ConstantInt(CI)), 2159 m_Value(L))) && 2160 APred == ICmpInst::ICMP_SLT && 2161 !CI->isNegative() && 2162 match(B, m_ICmp(BPred, m_Specific(I), m_Specific(L))) && 2163 BPred == ICmpInst::ICMP_SLT) 2164 return true; 2165 2166 // i +_{nuw} C_{>0} <u L ==> i <u L 2167 if (match(A, m_ICmp(APred, 2168 m_NUWAdd(m_Value(I), m_ConstantInt(CI)), 2169 m_Value(L))) && 2170 APred == ICmpInst::ICMP_ULT && 2171 !CI->isNegative() && 2172 match(B, m_ICmp(BPred, m_Specific(I), m_Specific(L))) && 2173 BPred == ICmpInst::ICMP_ULT) 2174 return true; 2175 2176 return false; 2177 } 2178 2179 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can 2180 /// fold the result. If not, this returns null. 2181 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2182 const Query &Q, unsigned MaxRecurse) { 2183 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 2184 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!"); 2185 2186 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 2187 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 2188 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 2189 2190 // If we have a constant, make sure it is on the RHS. 2191 std::swap(LHS, RHS); 2192 Pred = CmpInst::getSwappedPredicate(Pred); 2193 } 2194 2195 Type *ITy = GetCompareTy(LHS); // The return type. 2196 Type *OpTy = LHS->getType(); // The operand type. 2197 2198 // icmp X, X -> true/false 2199 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false 2200 // because X could be 0. 2201 if (LHS == RHS || isa<UndefValue>(RHS)) 2202 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred)); 2203 2204 // Special case logic when the operands have i1 type. 2205 if (OpTy->getScalarType()->isIntegerTy(1)) { 2206 switch (Pred) { 2207 default: break; 2208 case ICmpInst::ICMP_EQ: 2209 // X == 1 -> X 2210 if (match(RHS, m_One())) 2211 return LHS; 2212 break; 2213 case ICmpInst::ICMP_NE: 2214 // X != 0 -> X 2215 if (match(RHS, m_Zero())) 2216 return LHS; 2217 break; 2218 case ICmpInst::ICMP_UGT: 2219 // X >u 0 -> X 2220 if (match(RHS, m_Zero())) 2221 return LHS; 2222 break; 2223 case ICmpInst::ICMP_UGE: 2224 // X >=u 1 -> X 2225 if (match(RHS, m_One())) 2226 return LHS; 2227 if (implies(RHS, LHS)) 2228 return getTrue(ITy); 2229 break; 2230 case ICmpInst::ICMP_SLT: 2231 // X <s 0 -> X 2232 if (match(RHS, m_Zero())) 2233 return LHS; 2234 break; 2235 case ICmpInst::ICMP_SLE: 2236 // X <=s -1 -> X 2237 if (match(RHS, m_One())) 2238 return LHS; 2239 break; 2240 case ICmpInst::ICMP_ULE: 2241 if (implies(LHS, RHS)) 2242 return getTrue(ITy); 2243 break; 2244 } 2245 } 2246 2247 // If we are comparing with zero then try hard since this is a common case. 2248 if (match(RHS, m_Zero())) { 2249 bool LHSKnownNonNegative, LHSKnownNegative; 2250 switch (Pred) { 2251 default: llvm_unreachable("Unknown ICmp predicate!"); 2252 case ICmpInst::ICMP_ULT: 2253 return getFalse(ITy); 2254 case ICmpInst::ICMP_UGE: 2255 return getTrue(ITy); 2256 case ICmpInst::ICMP_EQ: 2257 case ICmpInst::ICMP_ULE: 2258 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2259 return getFalse(ITy); 2260 break; 2261 case ICmpInst::ICMP_NE: 2262 case ICmpInst::ICMP_UGT: 2263 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2264 return getTrue(ITy); 2265 break; 2266 case ICmpInst::ICMP_SLT: 2267 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2268 Q.CxtI, Q.DT); 2269 if (LHSKnownNegative) 2270 return getTrue(ITy); 2271 if (LHSKnownNonNegative) 2272 return getFalse(ITy); 2273 break; 2274 case ICmpInst::ICMP_SLE: 2275 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2276 Q.CxtI, Q.DT); 2277 if (LHSKnownNegative) 2278 return getTrue(ITy); 2279 if (LHSKnownNonNegative && 2280 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2281 return getFalse(ITy); 2282 break; 2283 case ICmpInst::ICMP_SGE: 2284 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2285 Q.CxtI, Q.DT); 2286 if (LHSKnownNegative) 2287 return getFalse(ITy); 2288 if (LHSKnownNonNegative) 2289 return getTrue(ITy); 2290 break; 2291 case ICmpInst::ICMP_SGT: 2292 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2293 Q.CxtI, Q.DT); 2294 if (LHSKnownNegative) 2295 return getFalse(ITy); 2296 if (LHSKnownNonNegative && 2297 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2298 return getTrue(ITy); 2299 break; 2300 } 2301 } 2302 2303 // See if we are doing a comparison with a constant integer. 2304 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2305 // Rule out tautological comparisons (eg., ult 0 or uge 0). 2306 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue()); 2307 if (RHS_CR.isEmptySet()) 2308 return ConstantInt::getFalse(CI->getContext()); 2309 if (RHS_CR.isFullSet()) 2310 return ConstantInt::getTrue(CI->getContext()); 2311 2312 // Many binary operators with constant RHS have easy to compute constant 2313 // range. Use them to check whether the comparison is a tautology. 2314 unsigned Width = CI->getBitWidth(); 2315 APInt Lower = APInt(Width, 0); 2316 APInt Upper = APInt(Width, 0); 2317 ConstantInt *CI2; 2318 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) { 2319 // 'urem x, CI2' produces [0, CI2). 2320 Upper = CI2->getValue(); 2321 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) { 2322 // 'srem x, CI2' produces (-|CI2|, |CI2|). 2323 Upper = CI2->getValue().abs(); 2324 Lower = (-Upper) + 1; 2325 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) { 2326 // 'udiv CI2, x' produces [0, CI2]. 2327 Upper = CI2->getValue() + 1; 2328 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) { 2329 // 'udiv x, CI2' produces [0, UINT_MAX / CI2]. 2330 APInt NegOne = APInt::getAllOnesValue(Width); 2331 if (!CI2->isZero()) 2332 Upper = NegOne.udiv(CI2->getValue()) + 1; 2333 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) { 2334 if (CI2->isMinSignedValue()) { 2335 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. 2336 Lower = CI2->getValue(); 2337 Upper = Lower.lshr(1) + 1; 2338 } else { 2339 // 'sdiv CI2, x' produces [-|CI2|, |CI2|]. 2340 Upper = CI2->getValue().abs() + 1; 2341 Lower = (-Upper) + 1; 2342 } 2343 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) { 2344 APInt IntMin = APInt::getSignedMinValue(Width); 2345 APInt IntMax = APInt::getSignedMaxValue(Width); 2346 APInt Val = CI2->getValue(); 2347 if (Val.isAllOnesValue()) { 2348 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] 2349 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2350 Lower = IntMin + 1; 2351 Upper = IntMax + 1; 2352 } else if (Val.countLeadingZeros() < Width - 1) { 2353 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2] 2354 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2355 Lower = IntMin.sdiv(Val); 2356 Upper = IntMax.sdiv(Val); 2357 if (Lower.sgt(Upper)) 2358 std::swap(Lower, Upper); 2359 Upper = Upper + 1; 2360 assert(Upper != Lower && "Upper part of range has wrapped!"); 2361 } 2362 } else if (match(LHS, m_NUWShl(m_ConstantInt(CI2), m_Value()))) { 2363 // 'shl nuw CI2, x' produces [CI2, CI2 << CLZ(CI2)] 2364 Lower = CI2->getValue(); 2365 Upper = Lower.shl(Lower.countLeadingZeros()) + 1; 2366 } else if (match(LHS, m_NSWShl(m_ConstantInt(CI2), m_Value()))) { 2367 if (CI2->isNegative()) { 2368 // 'shl nsw CI2, x' produces [CI2 << CLO(CI2)-1, CI2] 2369 unsigned ShiftAmount = CI2->getValue().countLeadingOnes() - 1; 2370 Lower = CI2->getValue().shl(ShiftAmount); 2371 Upper = CI2->getValue() + 1; 2372 } else { 2373 // 'shl nsw CI2, x' produces [CI2, CI2 << CLZ(CI2)-1] 2374 unsigned ShiftAmount = CI2->getValue().countLeadingZeros() - 1; 2375 Lower = CI2->getValue(); 2376 Upper = CI2->getValue().shl(ShiftAmount) + 1; 2377 } 2378 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) { 2379 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2]. 2380 APInt NegOne = APInt::getAllOnesValue(Width); 2381 if (CI2->getValue().ult(Width)) 2382 Upper = NegOne.lshr(CI2->getValue()) + 1; 2383 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) { 2384 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2]. 2385 unsigned ShiftAmount = Width - 1; 2386 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2387 ShiftAmount = CI2->getValue().countTrailingZeros(); 2388 Lower = CI2->getValue().lshr(ShiftAmount); 2389 Upper = CI2->getValue() + 1; 2390 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) { 2391 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2]. 2392 APInt IntMin = APInt::getSignedMinValue(Width); 2393 APInt IntMax = APInt::getSignedMaxValue(Width); 2394 if (CI2->getValue().ult(Width)) { 2395 Lower = IntMin.ashr(CI2->getValue()); 2396 Upper = IntMax.ashr(CI2->getValue()) + 1; 2397 } 2398 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) { 2399 unsigned ShiftAmount = Width - 1; 2400 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2401 ShiftAmount = CI2->getValue().countTrailingZeros(); 2402 if (CI2->isNegative()) { 2403 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)] 2404 Lower = CI2->getValue(); 2405 Upper = CI2->getValue().ashr(ShiftAmount) + 1; 2406 } else { 2407 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2] 2408 Lower = CI2->getValue().ashr(ShiftAmount); 2409 Upper = CI2->getValue() + 1; 2410 } 2411 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) { 2412 // 'or x, CI2' produces [CI2, UINT_MAX]. 2413 Lower = CI2->getValue(); 2414 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) { 2415 // 'and x, CI2' produces [0, CI2]. 2416 Upper = CI2->getValue() + 1; 2417 } else if (match(LHS, m_NUWAdd(m_Value(), m_ConstantInt(CI2)))) { 2418 // 'add nuw x, CI2' produces [CI2, UINT_MAX]. 2419 Lower = CI2->getValue(); 2420 } 2421 2422 ConstantRange LHS_CR = Lower != Upper ? ConstantRange(Lower, Upper) 2423 : ConstantRange(Width, true); 2424 2425 if (auto *I = dyn_cast<Instruction>(LHS)) 2426 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range)) 2427 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges)); 2428 2429 if (!LHS_CR.isFullSet()) { 2430 if (RHS_CR.contains(LHS_CR)) 2431 return ConstantInt::getTrue(RHS->getContext()); 2432 if (RHS_CR.inverse().contains(LHS_CR)) 2433 return ConstantInt::getFalse(RHS->getContext()); 2434 } 2435 } 2436 2437 // If both operands have range metadata, use the metadata 2438 // to simplify the comparison. 2439 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) { 2440 auto RHS_Instr = dyn_cast<Instruction>(RHS); 2441 auto LHS_Instr = dyn_cast<Instruction>(LHS); 2442 2443 if (RHS_Instr->getMetadata(LLVMContext::MD_range) && 2444 LHS_Instr->getMetadata(LLVMContext::MD_range)) { 2445 auto RHS_CR = getConstantRangeFromMetadata( 2446 *RHS_Instr->getMetadata(LLVMContext::MD_range)); 2447 auto LHS_CR = getConstantRangeFromMetadata( 2448 *LHS_Instr->getMetadata(LLVMContext::MD_range)); 2449 2450 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR); 2451 if (Satisfied_CR.contains(LHS_CR)) 2452 return ConstantInt::getTrue(RHS->getContext()); 2453 2454 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion( 2455 CmpInst::getInversePredicate(Pred), RHS_CR); 2456 if (InversedSatisfied_CR.contains(LHS_CR)) 2457 return ConstantInt::getFalse(RHS->getContext()); 2458 } 2459 } 2460 2461 // Compare of cast, for example (zext X) != 0 -> X != 0 2462 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) { 2463 Instruction *LI = cast<CastInst>(LHS); 2464 Value *SrcOp = LI->getOperand(0); 2465 Type *SrcTy = SrcOp->getType(); 2466 Type *DstTy = LI->getType(); 2467 2468 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input 2469 // if the integer type is the same size as the pointer type. 2470 if (MaxRecurse && isa<PtrToIntInst>(LI) && 2471 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) { 2472 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 2473 // Transfer the cast to the constant. 2474 if (Value *V = SimplifyICmpInst(Pred, SrcOp, 2475 ConstantExpr::getIntToPtr(RHSC, SrcTy), 2476 Q, MaxRecurse-1)) 2477 return V; 2478 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) { 2479 if (RI->getOperand(0)->getType() == SrcTy) 2480 // Compare without the cast. 2481 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2482 Q, MaxRecurse-1)) 2483 return V; 2484 } 2485 } 2486 2487 if (isa<ZExtInst>(LHS)) { 2488 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the 2489 // same type. 2490 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) { 2491 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2492 // Compare X and Y. Note that signed predicates become unsigned. 2493 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2494 SrcOp, RI->getOperand(0), Q, 2495 MaxRecurse-1)) 2496 return V; 2497 } 2498 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended 2499 // too. If not, then try to deduce the result of the comparison. 2500 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2501 // Compute the constant that would happen if we truncated to SrcTy then 2502 // reextended to DstTy. 2503 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2504 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy); 2505 2506 // If the re-extended constant didn't change then this is effectively 2507 // also a case of comparing two zero-extended values. 2508 if (RExt == CI && MaxRecurse) 2509 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2510 SrcOp, Trunc, Q, MaxRecurse-1)) 2511 return V; 2512 2513 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit 2514 // there. Use this to work out the result of the comparison. 2515 if (RExt != CI) { 2516 switch (Pred) { 2517 default: llvm_unreachable("Unknown ICmp predicate!"); 2518 // LHS <u RHS. 2519 case ICmpInst::ICMP_EQ: 2520 case ICmpInst::ICMP_UGT: 2521 case ICmpInst::ICMP_UGE: 2522 return ConstantInt::getFalse(CI->getContext()); 2523 2524 case ICmpInst::ICMP_NE: 2525 case ICmpInst::ICMP_ULT: 2526 case ICmpInst::ICMP_ULE: 2527 return ConstantInt::getTrue(CI->getContext()); 2528 2529 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS 2530 // is non-negative then LHS <s RHS. 2531 case ICmpInst::ICMP_SGT: 2532 case ICmpInst::ICMP_SGE: 2533 return CI->getValue().isNegative() ? 2534 ConstantInt::getTrue(CI->getContext()) : 2535 ConstantInt::getFalse(CI->getContext()); 2536 2537 case ICmpInst::ICMP_SLT: 2538 case ICmpInst::ICMP_SLE: 2539 return CI->getValue().isNegative() ? 2540 ConstantInt::getFalse(CI->getContext()) : 2541 ConstantInt::getTrue(CI->getContext()); 2542 } 2543 } 2544 } 2545 } 2546 2547 if (isa<SExtInst>(LHS)) { 2548 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the 2549 // same type. 2550 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) { 2551 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2552 // Compare X and Y. Note that the predicate does not change. 2553 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2554 Q, MaxRecurse-1)) 2555 return V; 2556 } 2557 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended 2558 // too. If not, then try to deduce the result of the comparison. 2559 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2560 // Compute the constant that would happen if we truncated to SrcTy then 2561 // reextended to DstTy. 2562 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2563 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy); 2564 2565 // If the re-extended constant didn't change then this is effectively 2566 // also a case of comparing two sign-extended values. 2567 if (RExt == CI && MaxRecurse) 2568 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1)) 2569 return V; 2570 2571 // Otherwise the upper bits of LHS are all equal, while RHS has varying 2572 // bits there. Use this to work out the result of the comparison. 2573 if (RExt != CI) { 2574 switch (Pred) { 2575 default: llvm_unreachable("Unknown ICmp predicate!"); 2576 case ICmpInst::ICMP_EQ: 2577 return ConstantInt::getFalse(CI->getContext()); 2578 case ICmpInst::ICMP_NE: 2579 return ConstantInt::getTrue(CI->getContext()); 2580 2581 // If RHS is non-negative then LHS <s RHS. If RHS is negative then 2582 // LHS >s RHS. 2583 case ICmpInst::ICMP_SGT: 2584 case ICmpInst::ICMP_SGE: 2585 return CI->getValue().isNegative() ? 2586 ConstantInt::getTrue(CI->getContext()) : 2587 ConstantInt::getFalse(CI->getContext()); 2588 case ICmpInst::ICMP_SLT: 2589 case ICmpInst::ICMP_SLE: 2590 return CI->getValue().isNegative() ? 2591 ConstantInt::getFalse(CI->getContext()) : 2592 ConstantInt::getTrue(CI->getContext()); 2593 2594 // If LHS is non-negative then LHS <u RHS. If LHS is negative then 2595 // LHS >u RHS. 2596 case ICmpInst::ICMP_UGT: 2597 case ICmpInst::ICMP_UGE: 2598 // Comparison is true iff the LHS <s 0. 2599 if (MaxRecurse) 2600 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp, 2601 Constant::getNullValue(SrcTy), 2602 Q, MaxRecurse-1)) 2603 return V; 2604 break; 2605 case ICmpInst::ICMP_ULT: 2606 case ICmpInst::ICMP_ULE: 2607 // Comparison is true iff the LHS >=s 0. 2608 if (MaxRecurse) 2609 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp, 2610 Constant::getNullValue(SrcTy), 2611 Q, MaxRecurse-1)) 2612 return V; 2613 break; 2614 } 2615 } 2616 } 2617 } 2618 } 2619 2620 // icmp eq|ne X, Y -> false|true if X != Y 2621 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) && 2622 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) { 2623 LLVMContext &Ctx = LHS->getType()->getContext(); 2624 return Pred == ICmpInst::ICMP_NE ? 2625 ConstantInt::getTrue(Ctx) : ConstantInt::getFalse(Ctx); 2626 } 2627 2628 // Special logic for binary operators. 2629 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS); 2630 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS); 2631 if (MaxRecurse && (LBO || RBO)) { 2632 // Analyze the case when either LHS or RHS is an add instruction. 2633 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 2634 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null). 2635 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false; 2636 if (LBO && LBO->getOpcode() == Instruction::Add) { 2637 A = LBO->getOperand(0); B = LBO->getOperand(1); 2638 NoLHSWrapProblem = ICmpInst::isEquality(Pred) || 2639 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) || 2640 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap()); 2641 } 2642 if (RBO && RBO->getOpcode() == Instruction::Add) { 2643 C = RBO->getOperand(0); D = RBO->getOperand(1); 2644 NoRHSWrapProblem = ICmpInst::isEquality(Pred) || 2645 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) || 2646 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap()); 2647 } 2648 2649 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2650 if ((A == RHS || B == RHS) && NoLHSWrapProblem) 2651 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A, 2652 Constant::getNullValue(RHS->getType()), 2653 Q, MaxRecurse-1)) 2654 return V; 2655 2656 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2657 if ((C == LHS || D == LHS) && NoRHSWrapProblem) 2658 if (Value *V = SimplifyICmpInst(Pred, 2659 Constant::getNullValue(LHS->getType()), 2660 C == LHS ? D : C, Q, MaxRecurse-1)) 2661 return V; 2662 2663 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow. 2664 if (A && C && (A == C || A == D || B == C || B == D) && 2665 NoLHSWrapProblem && NoRHSWrapProblem) { 2666 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2667 Value *Y, *Z; 2668 if (A == C) { 2669 // C + B == C + D -> B == D 2670 Y = B; 2671 Z = D; 2672 } else if (A == D) { 2673 // D + B == C + D -> B == C 2674 Y = B; 2675 Z = C; 2676 } else if (B == C) { 2677 // A + C == C + D -> A == D 2678 Y = A; 2679 Z = D; 2680 } else { 2681 assert(B == D); 2682 // A + D == C + D -> A == C 2683 Y = A; 2684 Z = C; 2685 } 2686 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1)) 2687 return V; 2688 } 2689 } 2690 2691 // icmp pred (or X, Y), X 2692 if (LBO && match(LBO, m_CombineOr(m_Or(m_Value(), m_Specific(RHS)), 2693 m_Or(m_Specific(RHS), m_Value())))) { 2694 if (Pred == ICmpInst::ICMP_ULT) 2695 return getFalse(ITy); 2696 if (Pred == ICmpInst::ICMP_UGE) 2697 return getTrue(ITy); 2698 } 2699 // icmp pred X, (or X, Y) 2700 if (RBO && match(RBO, m_CombineOr(m_Or(m_Value(), m_Specific(LHS)), 2701 m_Or(m_Specific(LHS), m_Value())))) { 2702 if (Pred == ICmpInst::ICMP_ULE) 2703 return getTrue(ITy); 2704 if (Pred == ICmpInst::ICMP_UGT) 2705 return getFalse(ITy); 2706 } 2707 2708 // icmp pred (and X, Y), X 2709 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)), 2710 m_And(m_Specific(RHS), m_Value())))) { 2711 if (Pred == ICmpInst::ICMP_UGT) 2712 return getFalse(ITy); 2713 if (Pred == ICmpInst::ICMP_ULE) 2714 return getTrue(ITy); 2715 } 2716 // icmp pred X, (and X, Y) 2717 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)), 2718 m_And(m_Specific(LHS), m_Value())))) { 2719 if (Pred == ICmpInst::ICMP_UGE) 2720 return getTrue(ITy); 2721 if (Pred == ICmpInst::ICMP_ULT) 2722 return getFalse(ITy); 2723 } 2724 2725 // 0 - (zext X) pred C 2726 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) { 2727 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 2728 if (RHSC->getValue().isStrictlyPositive()) { 2729 if (Pred == ICmpInst::ICMP_SLT) 2730 return ConstantInt::getTrue(RHSC->getContext()); 2731 if (Pred == ICmpInst::ICMP_SGE) 2732 return ConstantInt::getFalse(RHSC->getContext()); 2733 if (Pred == ICmpInst::ICMP_EQ) 2734 return ConstantInt::getFalse(RHSC->getContext()); 2735 if (Pred == ICmpInst::ICMP_NE) 2736 return ConstantInt::getTrue(RHSC->getContext()); 2737 } 2738 if (RHSC->getValue().isNonNegative()) { 2739 if (Pred == ICmpInst::ICMP_SLE) 2740 return ConstantInt::getTrue(RHSC->getContext()); 2741 if (Pred == ICmpInst::ICMP_SGT) 2742 return ConstantInt::getFalse(RHSC->getContext()); 2743 } 2744 } 2745 } 2746 2747 // icmp pred (urem X, Y), Y 2748 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) { 2749 bool KnownNonNegative, KnownNegative; 2750 switch (Pred) { 2751 default: 2752 break; 2753 case ICmpInst::ICMP_SGT: 2754 case ICmpInst::ICMP_SGE: 2755 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2756 Q.CxtI, Q.DT); 2757 if (!KnownNonNegative) 2758 break; 2759 // fall-through 2760 case ICmpInst::ICMP_EQ: 2761 case ICmpInst::ICMP_UGT: 2762 case ICmpInst::ICMP_UGE: 2763 return getFalse(ITy); 2764 case ICmpInst::ICMP_SLT: 2765 case ICmpInst::ICMP_SLE: 2766 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2767 Q.CxtI, Q.DT); 2768 if (!KnownNonNegative) 2769 break; 2770 // fall-through 2771 case ICmpInst::ICMP_NE: 2772 case ICmpInst::ICMP_ULT: 2773 case ICmpInst::ICMP_ULE: 2774 return getTrue(ITy); 2775 } 2776 } 2777 2778 // icmp pred X, (urem Y, X) 2779 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) { 2780 bool KnownNonNegative, KnownNegative; 2781 switch (Pred) { 2782 default: 2783 break; 2784 case ICmpInst::ICMP_SGT: 2785 case ICmpInst::ICMP_SGE: 2786 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2787 Q.CxtI, Q.DT); 2788 if (!KnownNonNegative) 2789 break; 2790 // fall-through 2791 case ICmpInst::ICMP_NE: 2792 case ICmpInst::ICMP_UGT: 2793 case ICmpInst::ICMP_UGE: 2794 return getTrue(ITy); 2795 case ICmpInst::ICMP_SLT: 2796 case ICmpInst::ICMP_SLE: 2797 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2798 Q.CxtI, Q.DT); 2799 if (!KnownNonNegative) 2800 break; 2801 // fall-through 2802 case ICmpInst::ICMP_EQ: 2803 case ICmpInst::ICMP_ULT: 2804 case ICmpInst::ICMP_ULE: 2805 return getFalse(ITy); 2806 } 2807 } 2808 2809 // x udiv y <=u x. 2810 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) { 2811 // icmp pred (X /u Y), X 2812 if (Pred == ICmpInst::ICMP_UGT) 2813 return getFalse(ITy); 2814 if (Pred == ICmpInst::ICMP_ULE) 2815 return getTrue(ITy); 2816 } 2817 2818 // handle: 2819 // CI2 << X == CI 2820 // CI2 << X != CI 2821 // 2822 // where CI2 is a power of 2 and CI isn't 2823 if (auto *CI = dyn_cast<ConstantInt>(RHS)) { 2824 const APInt *CI2Val, *CIVal = &CI->getValue(); 2825 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) && 2826 CI2Val->isPowerOf2()) { 2827 if (!CIVal->isPowerOf2()) { 2828 // CI2 << X can equal zero in some circumstances, 2829 // this simplification is unsafe if CI is zero. 2830 // 2831 // We know it is safe if: 2832 // - The shift is nsw, we can't shift out the one bit. 2833 // - The shift is nuw, we can't shift out the one bit. 2834 // - CI2 is one 2835 // - CI isn't zero 2836 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() || 2837 *CI2Val == 1 || !CI->isZero()) { 2838 if (Pred == ICmpInst::ICMP_EQ) 2839 return ConstantInt::getFalse(RHS->getContext()); 2840 if (Pred == ICmpInst::ICMP_NE) 2841 return ConstantInt::getTrue(RHS->getContext()); 2842 } 2843 } 2844 if (CIVal->isSignBit() && *CI2Val == 1) { 2845 if (Pred == ICmpInst::ICMP_UGT) 2846 return ConstantInt::getFalse(RHS->getContext()); 2847 if (Pred == ICmpInst::ICMP_ULE) 2848 return ConstantInt::getTrue(RHS->getContext()); 2849 } 2850 } 2851 } 2852 2853 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() && 2854 LBO->getOperand(1) == RBO->getOperand(1)) { 2855 switch (LBO->getOpcode()) { 2856 default: break; 2857 case Instruction::UDiv: 2858 case Instruction::LShr: 2859 if (ICmpInst::isSigned(Pred)) 2860 break; 2861 // fall-through 2862 case Instruction::SDiv: 2863 case Instruction::AShr: 2864 if (!LBO->isExact() || !RBO->isExact()) 2865 break; 2866 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2867 RBO->getOperand(0), Q, MaxRecurse-1)) 2868 return V; 2869 break; 2870 case Instruction::Shl: { 2871 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap(); 2872 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap(); 2873 if (!NUW && !NSW) 2874 break; 2875 if (!NSW && ICmpInst::isSigned(Pred)) 2876 break; 2877 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2878 RBO->getOperand(0), Q, MaxRecurse-1)) 2879 return V; 2880 break; 2881 } 2882 } 2883 } 2884 2885 // Simplify comparisons involving max/min. 2886 Value *A, *B; 2887 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE; 2888 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B". 2889 2890 // Signed variants on "max(a,b)>=a -> true". 2891 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2892 if (A != RHS) std::swap(A, B); // smax(A, B) pred A. 2893 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2894 // We analyze this as smax(A, B) pred A. 2895 P = Pred; 2896 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) && 2897 (A == LHS || B == LHS)) { 2898 if (A != LHS) std::swap(A, B); // A pred smax(A, B). 2899 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2900 // We analyze this as smax(A, B) swapped-pred A. 2901 P = CmpInst::getSwappedPredicate(Pred); 2902 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2903 (A == RHS || B == RHS)) { 2904 if (A != RHS) std::swap(A, B); // smin(A, B) pred A. 2905 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2906 // We analyze this as smax(-A, -B) swapped-pred -A. 2907 // Note that we do not need to actually form -A or -B thanks to EqP. 2908 P = CmpInst::getSwappedPredicate(Pred); 2909 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) && 2910 (A == LHS || B == LHS)) { 2911 if (A != LHS) std::swap(A, B); // A pred smin(A, B). 2912 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2913 // We analyze this as smax(-A, -B) pred -A. 2914 // Note that we do not need to actually form -A or -B thanks to EqP. 2915 P = Pred; 2916 } 2917 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2918 // Cases correspond to "max(A, B) p A". 2919 switch (P) { 2920 default: 2921 break; 2922 case CmpInst::ICMP_EQ: 2923 case CmpInst::ICMP_SLE: 2924 // Equivalent to "A EqP B". This may be the same as the condition tested 2925 // in the max/min; if so, we can just return that. 2926 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2927 return V; 2928 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2929 return V; 2930 // Otherwise, see if "A EqP B" simplifies. 2931 if (MaxRecurse) 2932 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2933 return V; 2934 break; 2935 case CmpInst::ICMP_NE: 2936 case CmpInst::ICMP_SGT: { 2937 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2938 // Equivalent to "A InvEqP B". This may be the same as the condition 2939 // tested in the max/min; if so, we can just return that. 2940 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2941 return V; 2942 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2943 return V; 2944 // Otherwise, see if "A InvEqP B" simplifies. 2945 if (MaxRecurse) 2946 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2947 return V; 2948 break; 2949 } 2950 case CmpInst::ICMP_SGE: 2951 // Always true. 2952 return getTrue(ITy); 2953 case CmpInst::ICMP_SLT: 2954 // Always false. 2955 return getFalse(ITy); 2956 } 2957 } 2958 2959 // Unsigned variants on "max(a,b)>=a -> true". 2960 P = CmpInst::BAD_ICMP_PREDICATE; 2961 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2962 if (A != RHS) std::swap(A, B); // umax(A, B) pred A. 2963 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2964 // We analyze this as umax(A, B) pred A. 2965 P = Pred; 2966 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) && 2967 (A == LHS || B == LHS)) { 2968 if (A != LHS) std::swap(A, B); // A pred umax(A, B). 2969 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2970 // We analyze this as umax(A, B) swapped-pred A. 2971 P = CmpInst::getSwappedPredicate(Pred); 2972 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2973 (A == RHS || B == RHS)) { 2974 if (A != RHS) std::swap(A, B); // umin(A, B) pred A. 2975 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2976 // We analyze this as umax(-A, -B) swapped-pred -A. 2977 // Note that we do not need to actually form -A or -B thanks to EqP. 2978 P = CmpInst::getSwappedPredicate(Pred); 2979 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) && 2980 (A == LHS || B == LHS)) { 2981 if (A != LHS) std::swap(A, B); // A pred umin(A, B). 2982 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2983 // We analyze this as umax(-A, -B) pred -A. 2984 // Note that we do not need to actually form -A or -B thanks to EqP. 2985 P = Pred; 2986 } 2987 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2988 // Cases correspond to "max(A, B) p A". 2989 switch (P) { 2990 default: 2991 break; 2992 case CmpInst::ICMP_EQ: 2993 case CmpInst::ICMP_ULE: 2994 // Equivalent to "A EqP B". This may be the same as the condition tested 2995 // in the max/min; if so, we can just return that. 2996 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2997 return V; 2998 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2999 return V; 3000 // Otherwise, see if "A EqP B" simplifies. 3001 if (MaxRecurse) 3002 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 3003 return V; 3004 break; 3005 case CmpInst::ICMP_NE: 3006 case CmpInst::ICMP_UGT: { 3007 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 3008 // Equivalent to "A InvEqP B". This may be the same as the condition 3009 // tested in the max/min; if so, we can just return that. 3010 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 3011 return V; 3012 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 3013 return V; 3014 // Otherwise, see if "A InvEqP B" simplifies. 3015 if (MaxRecurse) 3016 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 3017 return V; 3018 break; 3019 } 3020 case CmpInst::ICMP_UGE: 3021 // Always true. 3022 return getTrue(ITy); 3023 case CmpInst::ICMP_ULT: 3024 // Always false. 3025 return getFalse(ITy); 3026 } 3027 } 3028 3029 // Variants on "max(x,y) >= min(x,z)". 3030 Value *C, *D; 3031 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && 3032 match(RHS, m_SMin(m_Value(C), m_Value(D))) && 3033 (A == C || A == D || B == C || B == D)) { 3034 // max(x, ?) pred min(x, ?). 3035 if (Pred == CmpInst::ICMP_SGE) 3036 // Always true. 3037 return getTrue(ITy); 3038 if (Pred == CmpInst::ICMP_SLT) 3039 // Always false. 3040 return getFalse(ITy); 3041 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 3042 match(RHS, m_SMax(m_Value(C), m_Value(D))) && 3043 (A == C || A == D || B == C || B == D)) { 3044 // min(x, ?) pred max(x, ?). 3045 if (Pred == CmpInst::ICMP_SLE) 3046 // Always true. 3047 return getTrue(ITy); 3048 if (Pred == CmpInst::ICMP_SGT) 3049 // Always false. 3050 return getFalse(ITy); 3051 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && 3052 match(RHS, m_UMin(m_Value(C), m_Value(D))) && 3053 (A == C || A == D || B == C || B == D)) { 3054 // max(x, ?) pred min(x, ?). 3055 if (Pred == CmpInst::ICMP_UGE) 3056 // Always true. 3057 return getTrue(ITy); 3058 if (Pred == CmpInst::ICMP_ULT) 3059 // Always false. 3060 return getFalse(ITy); 3061 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 3062 match(RHS, m_UMax(m_Value(C), m_Value(D))) && 3063 (A == C || A == D || B == C || B == D)) { 3064 // min(x, ?) pred max(x, ?). 3065 if (Pred == CmpInst::ICMP_ULE) 3066 // Always true. 3067 return getTrue(ITy); 3068 if (Pred == CmpInst::ICMP_UGT) 3069 // Always false. 3070 return getFalse(ITy); 3071 } 3072 3073 // Simplify comparisons of related pointers using a powerful, recursive 3074 // GEP-walk when we have target data available.. 3075 if (LHS->getType()->isPointerTy()) 3076 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS)) 3077 return C; 3078 3079 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) { 3080 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) { 3081 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() && 3082 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() && 3083 (ICmpInst::isEquality(Pred) || 3084 (GLHS->isInBounds() && GRHS->isInBounds() && 3085 Pred == ICmpInst::getSignedPredicate(Pred)))) { 3086 // The bases are equal and the indices are constant. Build a constant 3087 // expression GEP with the same indices and a null base pointer to see 3088 // what constant folding can make out of it. 3089 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType()); 3090 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end()); 3091 Constant *NewLHS = ConstantExpr::getGetElementPtr( 3092 GLHS->getSourceElementType(), Null, IndicesLHS); 3093 3094 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end()); 3095 Constant *NewRHS = ConstantExpr::getGetElementPtr( 3096 GLHS->getSourceElementType(), Null, IndicesRHS); 3097 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS); 3098 } 3099 } 3100 } 3101 3102 // If a bit is known to be zero for A and known to be one for B, 3103 // then A and B cannot be equal. 3104 if (ICmpInst::isEquality(Pred)) { 3105 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 3106 uint32_t BitWidth = CI->getBitWidth(); 3107 APInt LHSKnownZero(BitWidth, 0); 3108 APInt LHSKnownOne(BitWidth, 0); 3109 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC, 3110 Q.CxtI, Q.DT); 3111 const APInt &RHSVal = CI->getValue(); 3112 if (((LHSKnownZero & RHSVal) != 0) || ((LHSKnownOne & ~RHSVal) != 0)) 3113 return Pred == ICmpInst::ICMP_EQ 3114 ? ConstantInt::getFalse(CI->getContext()) 3115 : ConstantInt::getTrue(CI->getContext()); 3116 } 3117 } 3118 3119 // If the comparison is with the result of a select instruction, check whether 3120 // comparing with either branch of the select always yields the same value. 3121 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3122 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3123 return V; 3124 3125 // If the comparison is with the result of a phi instruction, check whether 3126 // doing the compare with each incoming phi value yields a common result. 3127 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3128 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3129 return V; 3130 3131 return nullptr; 3132 } 3133 3134 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3135 const DataLayout &DL, 3136 const TargetLibraryInfo *TLI, 3137 const DominatorTree *DT, AssumptionCache *AC, 3138 Instruction *CxtI) { 3139 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3140 RecursionLimit); 3141 } 3142 3143 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can 3144 /// fold the result. If not, this returns null. 3145 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3146 FastMathFlags FMF, const Query &Q, 3147 unsigned MaxRecurse) { 3148 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 3149 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!"); 3150 3151 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 3152 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 3153 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 3154 3155 // If we have a constant, make sure it is on the RHS. 3156 std::swap(LHS, RHS); 3157 Pred = CmpInst::getSwappedPredicate(Pred); 3158 } 3159 3160 // Fold trivial predicates. 3161 if (Pred == FCmpInst::FCMP_FALSE) 3162 return ConstantInt::get(GetCompareTy(LHS), 0); 3163 if (Pred == FCmpInst::FCMP_TRUE) 3164 return ConstantInt::get(GetCompareTy(LHS), 1); 3165 3166 // UNO/ORD predicates can be trivially folded if NaNs are ignored. 3167 if (FMF.noNaNs()) { 3168 if (Pred == FCmpInst::FCMP_UNO) 3169 return ConstantInt::get(GetCompareTy(LHS), 0); 3170 if (Pred == FCmpInst::FCMP_ORD) 3171 return ConstantInt::get(GetCompareTy(LHS), 1); 3172 } 3173 3174 // fcmp pred x, undef and fcmp pred undef, x 3175 // fold to true if unordered, false if ordered 3176 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) { 3177 // Choosing NaN for the undef will always make unordered comparison succeed 3178 // and ordered comparison fail. 3179 return ConstantInt::get(GetCompareTy(LHS), CmpInst::isUnordered(Pred)); 3180 } 3181 3182 // fcmp x,x -> true/false. Not all compares are foldable. 3183 if (LHS == RHS) { 3184 if (CmpInst::isTrueWhenEqual(Pred)) 3185 return ConstantInt::get(GetCompareTy(LHS), 1); 3186 if (CmpInst::isFalseWhenEqual(Pred)) 3187 return ConstantInt::get(GetCompareTy(LHS), 0); 3188 } 3189 3190 // Handle fcmp with constant RHS 3191 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) { 3192 // If the constant is a nan, see if we can fold the comparison based on it. 3193 if (CFP->getValueAPF().isNaN()) { 3194 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo" 3195 return ConstantInt::getFalse(CFP->getContext()); 3196 assert(FCmpInst::isUnordered(Pred) && 3197 "Comparison must be either ordered or unordered!"); 3198 // True if unordered. 3199 return ConstantInt::getTrue(CFP->getContext()); 3200 } 3201 // Check whether the constant is an infinity. 3202 if (CFP->getValueAPF().isInfinity()) { 3203 if (CFP->getValueAPF().isNegative()) { 3204 switch (Pred) { 3205 case FCmpInst::FCMP_OLT: 3206 // No value is ordered and less than negative infinity. 3207 return ConstantInt::getFalse(CFP->getContext()); 3208 case FCmpInst::FCMP_UGE: 3209 // All values are unordered with or at least negative infinity. 3210 return ConstantInt::getTrue(CFP->getContext()); 3211 default: 3212 break; 3213 } 3214 } else { 3215 switch (Pred) { 3216 case FCmpInst::FCMP_OGT: 3217 // No value is ordered and greater than infinity. 3218 return ConstantInt::getFalse(CFP->getContext()); 3219 case FCmpInst::FCMP_ULE: 3220 // All values are unordered with and at most infinity. 3221 return ConstantInt::getTrue(CFP->getContext()); 3222 default: 3223 break; 3224 } 3225 } 3226 } 3227 if (CFP->getValueAPF().isZero()) { 3228 switch (Pred) { 3229 case FCmpInst::FCMP_UGE: 3230 if (CannotBeOrderedLessThanZero(LHS)) 3231 return ConstantInt::getTrue(CFP->getContext()); 3232 break; 3233 case FCmpInst::FCMP_OLT: 3234 // X < 0 3235 if (CannotBeOrderedLessThanZero(LHS)) 3236 return ConstantInt::getFalse(CFP->getContext()); 3237 break; 3238 default: 3239 break; 3240 } 3241 } 3242 } 3243 3244 // If the comparison is with the result of a select instruction, check whether 3245 // comparing with either branch of the select always yields the same value. 3246 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3247 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3248 return V; 3249 3250 // If the comparison is with the result of a phi instruction, check whether 3251 // doing the compare with each incoming phi value yields a common result. 3252 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3253 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3254 return V; 3255 3256 return nullptr; 3257 } 3258 3259 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3260 FastMathFlags FMF, const DataLayout &DL, 3261 const TargetLibraryInfo *TLI, 3262 const DominatorTree *DT, AssumptionCache *AC, 3263 const Instruction *CxtI) { 3264 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, 3265 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3266 } 3267 3268 /// SimplifyWithOpReplaced - See if V simplifies when its operand Op is 3269 /// replaced with RepOp. 3270 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp, 3271 const Query &Q, 3272 unsigned MaxRecurse) { 3273 // Trivial replacement. 3274 if (V == Op) 3275 return RepOp; 3276 3277 auto *I = dyn_cast<Instruction>(V); 3278 if (!I) 3279 return nullptr; 3280 3281 // If this is a binary operator, try to simplify it with the replaced op. 3282 if (auto *B = dyn_cast<BinaryOperator>(I)) { 3283 // Consider: 3284 // %cmp = icmp eq i32 %x, 2147483647 3285 // %add = add nsw i32 %x, 1 3286 // %sel = select i1 %cmp, i32 -2147483648, i32 %add 3287 // 3288 // We can't replace %sel with %add unless we strip away the flags. 3289 if (isa<OverflowingBinaryOperator>(B)) 3290 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap()) 3291 return nullptr; 3292 if (isa<PossiblyExactOperator>(B)) 3293 if (B->isExact()) 3294 return nullptr; 3295 3296 if (MaxRecurse) { 3297 if (B->getOperand(0) == Op) 3298 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q, 3299 MaxRecurse - 1); 3300 if (B->getOperand(1) == Op) 3301 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q, 3302 MaxRecurse - 1); 3303 } 3304 } 3305 3306 // Same for CmpInsts. 3307 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 3308 if (MaxRecurse) { 3309 if (C->getOperand(0) == Op) 3310 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q, 3311 MaxRecurse - 1); 3312 if (C->getOperand(1) == Op) 3313 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q, 3314 MaxRecurse - 1); 3315 } 3316 } 3317 3318 // TODO: We could hand off more cases to instsimplify here. 3319 3320 // If all operands are constant after substituting Op for RepOp then we can 3321 // constant fold the instruction. 3322 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) { 3323 // Build a list of all constant operands. 3324 SmallVector<Constant *, 8> ConstOps; 3325 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 3326 if (I->getOperand(i) == Op) 3327 ConstOps.push_back(CRepOp); 3328 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i))) 3329 ConstOps.push_back(COp); 3330 else 3331 break; 3332 } 3333 3334 // All operands were constants, fold it. 3335 if (ConstOps.size() == I->getNumOperands()) { 3336 if (CmpInst *C = dyn_cast<CmpInst>(I)) 3337 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0], 3338 ConstOps[1], Q.DL, Q.TLI); 3339 3340 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 3341 if (!LI->isVolatile()) 3342 return ConstantFoldLoadFromConstPtr(ConstOps[0], Q.DL); 3343 3344 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), ConstOps, 3345 Q.DL, Q.TLI); 3346 } 3347 } 3348 3349 return nullptr; 3350 } 3351 3352 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold 3353 /// the result. If not, this returns null. 3354 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal, 3355 Value *FalseVal, const Query &Q, 3356 unsigned MaxRecurse) { 3357 // select true, X, Y -> X 3358 // select false, X, Y -> Y 3359 if (Constant *CB = dyn_cast<Constant>(CondVal)) { 3360 if (CB->isAllOnesValue()) 3361 return TrueVal; 3362 if (CB->isNullValue()) 3363 return FalseVal; 3364 } 3365 3366 // select C, X, X -> X 3367 if (TrueVal == FalseVal) 3368 return TrueVal; 3369 3370 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y 3371 if (isa<Constant>(TrueVal)) 3372 return TrueVal; 3373 return FalseVal; 3374 } 3375 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X 3376 return FalseVal; 3377 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X 3378 return TrueVal; 3379 3380 if (const auto *ICI = dyn_cast<ICmpInst>(CondVal)) { 3381 unsigned BitWidth = Q.DL.getTypeSizeInBits(TrueVal->getType()); 3382 ICmpInst::Predicate Pred = ICI->getPredicate(); 3383 Value *CmpLHS = ICI->getOperand(0); 3384 Value *CmpRHS = ICI->getOperand(1); 3385 APInt MinSignedValue = APInt::getSignBit(BitWidth); 3386 Value *X; 3387 const APInt *Y; 3388 bool TrueWhenUnset; 3389 bool IsBitTest = false; 3390 if (ICmpInst::isEquality(Pred) && 3391 match(CmpLHS, m_And(m_Value(X), m_APInt(Y))) && 3392 match(CmpRHS, m_Zero())) { 3393 IsBitTest = true; 3394 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 3395 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { 3396 X = CmpLHS; 3397 Y = &MinSignedValue; 3398 IsBitTest = true; 3399 TrueWhenUnset = false; 3400 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { 3401 X = CmpLHS; 3402 Y = &MinSignedValue; 3403 IsBitTest = true; 3404 TrueWhenUnset = true; 3405 } 3406 if (IsBitTest) { 3407 const APInt *C; 3408 // (X & Y) == 0 ? X & ~Y : X --> X 3409 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y 3410 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) && 3411 *Y == ~*C) 3412 return TrueWhenUnset ? FalseVal : TrueVal; 3413 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y 3414 // (X & Y) != 0 ? X : X & ~Y --> X 3415 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) && 3416 *Y == ~*C) 3417 return TrueWhenUnset ? FalseVal : TrueVal; 3418 3419 if (Y->isPowerOf2()) { 3420 // (X & Y) == 0 ? X | Y : X --> X | Y 3421 // (X & Y) != 0 ? X | Y : X --> X 3422 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) && 3423 *Y == *C) 3424 return TrueWhenUnset ? TrueVal : FalseVal; 3425 // (X & Y) == 0 ? X : X | Y --> X 3426 // (X & Y) != 0 ? X : X | Y --> X | Y 3427 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) && 3428 *Y == *C) 3429 return TrueWhenUnset ? TrueVal : FalseVal; 3430 } 3431 } 3432 if (ICI->hasOneUse()) { 3433 const APInt *C; 3434 if (match(CmpRHS, m_APInt(C))) { 3435 // X < MIN ? T : F --> F 3436 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue()) 3437 return FalseVal; 3438 // X < MIN ? T : F --> F 3439 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue()) 3440 return FalseVal; 3441 // X > MAX ? T : F --> F 3442 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue()) 3443 return FalseVal; 3444 // X > MAX ? T : F --> F 3445 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue()) 3446 return FalseVal; 3447 } 3448 } 3449 3450 // If we have an equality comparison then we know the value in one of the 3451 // arms of the select. See if substituting this value into the arm and 3452 // simplifying the result yields the same value as the other arm. 3453 if (Pred == ICmpInst::ICMP_EQ) { 3454 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3455 TrueVal || 3456 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3457 TrueVal) 3458 return FalseVal; 3459 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3460 FalseVal || 3461 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3462 FalseVal) 3463 return FalseVal; 3464 } else if (Pred == ICmpInst::ICMP_NE) { 3465 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3466 FalseVal || 3467 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3468 FalseVal) 3469 return TrueVal; 3470 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3471 TrueVal || 3472 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3473 TrueVal) 3474 return TrueVal; 3475 } 3476 } 3477 3478 return nullptr; 3479 } 3480 3481 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, 3482 const DataLayout &DL, 3483 const TargetLibraryInfo *TLI, 3484 const DominatorTree *DT, AssumptionCache *AC, 3485 const Instruction *CxtI) { 3486 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, 3487 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3488 } 3489 3490 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can 3491 /// fold the result. If not, this returns null. 3492 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, 3493 const Query &Q, unsigned) { 3494 // The type of the GEP pointer operand. 3495 unsigned AS = 3496 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace(); 3497 3498 // getelementptr P -> P. 3499 if (Ops.size() == 1) 3500 return Ops[0]; 3501 3502 // Compute the (pointer) type returned by the GEP instruction. 3503 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1)); 3504 Type *GEPTy = PointerType::get(LastType, AS); 3505 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType())) 3506 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 3507 3508 if (isa<UndefValue>(Ops[0])) 3509 return UndefValue::get(GEPTy); 3510 3511 if (Ops.size() == 2) { 3512 // getelementptr P, 0 -> P. 3513 if (match(Ops[1], m_Zero())) 3514 return Ops[0]; 3515 3516 Type *Ty = SrcTy; 3517 if (Ty->isSized()) { 3518 Value *P; 3519 uint64_t C; 3520 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty); 3521 // getelementptr P, N -> P if P points to a type of zero size. 3522 if (TyAllocSize == 0) 3523 return Ops[0]; 3524 3525 // The following transforms are only safe if the ptrtoint cast 3526 // doesn't truncate the pointers. 3527 if (Ops[1]->getType()->getScalarSizeInBits() == 3528 Q.DL.getPointerSizeInBits(AS)) { 3529 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * { 3530 if (match(P, m_Zero())) 3531 return Constant::getNullValue(GEPTy); 3532 Value *Temp; 3533 if (match(P, m_PtrToInt(m_Value(Temp)))) 3534 if (Temp->getType() == GEPTy) 3535 return Temp; 3536 return nullptr; 3537 }; 3538 3539 // getelementptr V, (sub P, V) -> P if P points to a type of size 1. 3540 if (TyAllocSize == 1 && 3541 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))))) 3542 if (Value *R = PtrToIntOrZero(P)) 3543 return R; 3544 3545 // getelementptr V, (ashr (sub P, V), C) -> Q 3546 // if P points to a type of size 1 << C. 3547 if (match(Ops[1], 3548 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3549 m_ConstantInt(C))) && 3550 TyAllocSize == 1ULL << C) 3551 if (Value *R = PtrToIntOrZero(P)) 3552 return R; 3553 3554 // getelementptr V, (sdiv (sub P, V), C) -> Q 3555 // if P points to a type of size C. 3556 if (match(Ops[1], 3557 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3558 m_SpecificInt(TyAllocSize)))) 3559 if (Value *R = PtrToIntOrZero(P)) 3560 return R; 3561 } 3562 } 3563 } 3564 3565 // Check to see if this is constant foldable. 3566 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 3567 if (!isa<Constant>(Ops[i])) 3568 return nullptr; 3569 3570 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]), 3571 Ops.slice(1)); 3572 } 3573 3574 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout &DL, 3575 const TargetLibraryInfo *TLI, 3576 const DominatorTree *DT, AssumptionCache *AC, 3577 const Instruction *CxtI) { 3578 return ::SimplifyGEPInst( 3579 cast<PointerType>(Ops[0]->getType()->getScalarType())->getElementType(), 3580 Ops, Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3581 } 3582 3583 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we 3584 /// can fold the result. If not, this returns null. 3585 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val, 3586 ArrayRef<unsigned> Idxs, const Query &Q, 3587 unsigned) { 3588 if (Constant *CAgg = dyn_cast<Constant>(Agg)) 3589 if (Constant *CVal = dyn_cast<Constant>(Val)) 3590 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs); 3591 3592 // insertvalue x, undef, n -> x 3593 if (match(Val, m_Undef())) 3594 return Agg; 3595 3596 // insertvalue x, (extractvalue y, n), n 3597 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val)) 3598 if (EV->getAggregateOperand()->getType() == Agg->getType() && 3599 EV->getIndices() == Idxs) { 3600 // insertvalue undef, (extractvalue y, n), n -> y 3601 if (match(Agg, m_Undef())) 3602 return EV->getAggregateOperand(); 3603 3604 // insertvalue y, (extractvalue y, n), n -> y 3605 if (Agg == EV->getAggregateOperand()) 3606 return Agg; 3607 } 3608 3609 return nullptr; 3610 } 3611 3612 Value *llvm::SimplifyInsertValueInst( 3613 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout &DL, 3614 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC, 3615 const Instruction *CxtI) { 3616 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI), 3617 RecursionLimit); 3618 } 3619 3620 /// SimplifyExtractValueInst - Given operands for an ExtractValueInst, see if we 3621 /// can fold the result. If not, this returns null. 3622 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, 3623 const Query &, unsigned) { 3624 if (auto *CAgg = dyn_cast<Constant>(Agg)) 3625 return ConstantFoldExtractValueInstruction(CAgg, Idxs); 3626 3627 // extractvalue x, (insertvalue y, elt, n), n -> elt 3628 unsigned NumIdxs = Idxs.size(); 3629 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr; 3630 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) { 3631 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices(); 3632 unsigned NumInsertValueIdxs = InsertValueIdxs.size(); 3633 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs); 3634 if (InsertValueIdxs.slice(0, NumCommonIdxs) == 3635 Idxs.slice(0, NumCommonIdxs)) { 3636 if (NumIdxs == NumInsertValueIdxs) 3637 return IVI->getInsertedValueOperand(); 3638 break; 3639 } 3640 } 3641 3642 return nullptr; 3643 } 3644 3645 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, 3646 const DataLayout &DL, 3647 const TargetLibraryInfo *TLI, 3648 const DominatorTree *DT, 3649 AssumptionCache *AC, 3650 const Instruction *CxtI) { 3651 return ::SimplifyExtractValueInst(Agg, Idxs, Query(DL, TLI, DT, AC, CxtI), 3652 RecursionLimit); 3653 } 3654 3655 /// SimplifyExtractElementInst - Given operands for an ExtractElementInst, see if we 3656 /// can fold the result. If not, this returns null. 3657 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const Query &, 3658 unsigned) { 3659 if (auto *CVec = dyn_cast<Constant>(Vec)) { 3660 if (auto *CIdx = dyn_cast<Constant>(Idx)) 3661 return ConstantFoldExtractElementInstruction(CVec, CIdx); 3662 3663 // The index is not relevant if our vector is a splat. 3664 if (auto *Splat = CVec->getSplatValue()) 3665 return Splat; 3666 3667 if (isa<UndefValue>(Vec)) 3668 return UndefValue::get(Vec->getType()->getVectorElementType()); 3669 } 3670 3671 // If extracting a specified index from the vector, see if we can recursively 3672 // find a previously computed scalar that was inserted into the vector. 3673 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) 3674 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue())) 3675 return Elt; 3676 3677 return nullptr; 3678 } 3679 3680 Value *llvm::SimplifyExtractElementInst( 3681 Value *Vec, Value *Idx, const DataLayout &DL, const TargetLibraryInfo *TLI, 3682 const DominatorTree *DT, AssumptionCache *AC, const Instruction *CxtI) { 3683 return ::SimplifyExtractElementInst(Vec, Idx, Query(DL, TLI, DT, AC, CxtI), 3684 RecursionLimit); 3685 } 3686 3687 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null. 3688 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) { 3689 // If all of the PHI's incoming values are the same then replace the PHI node 3690 // with the common value. 3691 Value *CommonValue = nullptr; 3692 bool HasUndefInput = false; 3693 for (Value *Incoming : PN->incoming_values()) { 3694 // If the incoming value is the phi node itself, it can safely be skipped. 3695 if (Incoming == PN) continue; 3696 if (isa<UndefValue>(Incoming)) { 3697 // Remember that we saw an undef value, but otherwise ignore them. 3698 HasUndefInput = true; 3699 continue; 3700 } 3701 if (CommonValue && Incoming != CommonValue) 3702 return nullptr; // Not the same, bail out. 3703 CommonValue = Incoming; 3704 } 3705 3706 // If CommonValue is null then all of the incoming values were either undef or 3707 // equal to the phi node itself. 3708 if (!CommonValue) 3709 return UndefValue::get(PN->getType()); 3710 3711 // If we have a PHI node like phi(X, undef, X), where X is defined by some 3712 // instruction, we cannot return X as the result of the PHI node unless it 3713 // dominates the PHI block. 3714 if (HasUndefInput) 3715 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr; 3716 3717 return CommonValue; 3718 } 3719 3720 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) { 3721 if (Constant *C = dyn_cast<Constant>(Op)) 3722 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI); 3723 3724 return nullptr; 3725 } 3726 3727 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout &DL, 3728 const TargetLibraryInfo *TLI, 3729 const DominatorTree *DT, AssumptionCache *AC, 3730 const Instruction *CxtI) { 3731 return ::SimplifyTruncInst(Op, Ty, Query(DL, TLI, DT, AC, CxtI), 3732 RecursionLimit); 3733 } 3734 3735 //=== Helper functions for higher up the class hierarchy. 3736 3737 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can 3738 /// fold the result. If not, this returns null. 3739 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3740 const Query &Q, unsigned MaxRecurse) { 3741 switch (Opcode) { 3742 case Instruction::Add: 3743 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3744 Q, MaxRecurse); 3745 case Instruction::FAdd: 3746 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3747 3748 case Instruction::Sub: 3749 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3750 Q, MaxRecurse); 3751 case Instruction::FSub: 3752 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3753 3754 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse); 3755 case Instruction::FMul: 3756 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3757 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse); 3758 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse); 3759 case Instruction::FDiv: 3760 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3761 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse); 3762 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse); 3763 case Instruction::FRem: 3764 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3765 case Instruction::Shl: 3766 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3767 Q, MaxRecurse); 3768 case Instruction::LShr: 3769 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3770 case Instruction::AShr: 3771 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3772 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse); 3773 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse); 3774 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse); 3775 default: 3776 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 3777 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 3778 Constant *COps[] = {CLHS, CRHS}; 3779 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL, 3780 Q.TLI); 3781 } 3782 3783 // If the operation is associative, try some generic simplifications. 3784 if (Instruction::isAssociative(Opcode)) 3785 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse)) 3786 return V; 3787 3788 // If the operation is with the result of a select instruction check whether 3789 // operating on either branch of the select always yields the same value. 3790 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3791 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse)) 3792 return V; 3793 3794 // If the operation is with the result of a phi instruction, check whether 3795 // operating on all incoming values of the phi always yields the same value. 3796 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3797 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse)) 3798 return V; 3799 3800 return nullptr; 3801 } 3802 } 3803 3804 /// SimplifyFPBinOp - Given operands for a BinaryOperator, see if we can 3805 /// fold the result. If not, this returns null. 3806 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the 3807 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp. 3808 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3809 const FastMathFlags &FMF, const Query &Q, 3810 unsigned MaxRecurse) { 3811 switch (Opcode) { 3812 case Instruction::FAdd: 3813 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse); 3814 case Instruction::FSub: 3815 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse); 3816 case Instruction::FMul: 3817 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse); 3818 default: 3819 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse); 3820 } 3821 } 3822 3823 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3824 const DataLayout &DL, const TargetLibraryInfo *TLI, 3825 const DominatorTree *DT, AssumptionCache *AC, 3826 const Instruction *CxtI) { 3827 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3828 RecursionLimit); 3829 } 3830 3831 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3832 const FastMathFlags &FMF, const DataLayout &DL, 3833 const TargetLibraryInfo *TLI, 3834 const DominatorTree *DT, AssumptionCache *AC, 3835 const Instruction *CxtI) { 3836 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI), 3837 RecursionLimit); 3838 } 3839 3840 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can 3841 /// fold the result. 3842 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3843 const Query &Q, unsigned MaxRecurse) { 3844 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate)) 3845 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 3846 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3847 } 3848 3849 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3850 const DataLayout &DL, const TargetLibraryInfo *TLI, 3851 const DominatorTree *DT, AssumptionCache *AC, 3852 const Instruction *CxtI) { 3853 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3854 RecursionLimit); 3855 } 3856 3857 static bool IsIdempotent(Intrinsic::ID ID) { 3858 switch (ID) { 3859 default: return false; 3860 3861 // Unary idempotent: f(f(x)) = f(x) 3862 case Intrinsic::fabs: 3863 case Intrinsic::floor: 3864 case Intrinsic::ceil: 3865 case Intrinsic::trunc: 3866 case Intrinsic::rint: 3867 case Intrinsic::nearbyint: 3868 case Intrinsic::round: 3869 return true; 3870 } 3871 } 3872 3873 template <typename IterTy> 3874 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd, 3875 const Query &Q, unsigned MaxRecurse) { 3876 Intrinsic::ID IID = F->getIntrinsicID(); 3877 unsigned NumOperands = std::distance(ArgBegin, ArgEnd); 3878 Type *ReturnType = F->getReturnType(); 3879 3880 // Binary Ops 3881 if (NumOperands == 2) { 3882 Value *LHS = *ArgBegin; 3883 Value *RHS = *(ArgBegin + 1); 3884 if (IID == Intrinsic::usub_with_overflow || 3885 IID == Intrinsic::ssub_with_overflow) { 3886 // X - X -> { 0, false } 3887 if (LHS == RHS) 3888 return Constant::getNullValue(ReturnType); 3889 3890 // X - undef -> undef 3891 // undef - X -> undef 3892 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) 3893 return UndefValue::get(ReturnType); 3894 } 3895 3896 if (IID == Intrinsic::uadd_with_overflow || 3897 IID == Intrinsic::sadd_with_overflow) { 3898 // X + undef -> undef 3899 if (isa<UndefValue>(RHS)) 3900 return UndefValue::get(ReturnType); 3901 } 3902 3903 if (IID == Intrinsic::umul_with_overflow || 3904 IID == Intrinsic::smul_with_overflow) { 3905 // X * 0 -> { 0, false } 3906 if (match(RHS, m_Zero())) 3907 return Constant::getNullValue(ReturnType); 3908 3909 // X * undef -> { 0, false } 3910 if (match(RHS, m_Undef())) 3911 return Constant::getNullValue(ReturnType); 3912 } 3913 } 3914 3915 // Perform idempotent optimizations 3916 if (!IsIdempotent(IID)) 3917 return nullptr; 3918 3919 // Unary Ops 3920 if (NumOperands == 1) 3921 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) 3922 if (II->getIntrinsicID() == IID) 3923 return II; 3924 3925 return nullptr; 3926 } 3927 3928 template <typename IterTy> 3929 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd, 3930 const Query &Q, unsigned MaxRecurse) { 3931 Type *Ty = V->getType(); 3932 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 3933 Ty = PTy->getElementType(); 3934 FunctionType *FTy = cast<FunctionType>(Ty); 3935 3936 // call undef -> undef 3937 if (isa<UndefValue>(V)) 3938 return UndefValue::get(FTy->getReturnType()); 3939 3940 Function *F = dyn_cast<Function>(V); 3941 if (!F) 3942 return nullptr; 3943 3944 if (F->isIntrinsic()) 3945 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse)) 3946 return Ret; 3947 3948 if (!canConstantFoldCallTo(F)) 3949 return nullptr; 3950 3951 SmallVector<Constant *, 4> ConstantArgs; 3952 ConstantArgs.reserve(ArgEnd - ArgBegin); 3953 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) { 3954 Constant *C = dyn_cast<Constant>(*I); 3955 if (!C) 3956 return nullptr; 3957 ConstantArgs.push_back(C); 3958 } 3959 3960 return ConstantFoldCall(F, ConstantArgs, Q.TLI); 3961 } 3962 3963 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin, 3964 User::op_iterator ArgEnd, const DataLayout &DL, 3965 const TargetLibraryInfo *TLI, const DominatorTree *DT, 3966 AssumptionCache *AC, const Instruction *CxtI) { 3967 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI), 3968 RecursionLimit); 3969 } 3970 3971 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args, 3972 const DataLayout &DL, const TargetLibraryInfo *TLI, 3973 const DominatorTree *DT, AssumptionCache *AC, 3974 const Instruction *CxtI) { 3975 return ::SimplifyCall(V, Args.begin(), Args.end(), 3976 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3977 } 3978 3979 /// SimplifyInstruction - See if we can compute a simplified version of this 3980 /// instruction. If not, this returns null. 3981 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout &DL, 3982 const TargetLibraryInfo *TLI, 3983 const DominatorTree *DT, AssumptionCache *AC) { 3984 Value *Result; 3985 3986 switch (I->getOpcode()) { 3987 default: 3988 Result = ConstantFoldInstruction(I, DL, TLI); 3989 break; 3990 case Instruction::FAdd: 3991 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1), 3992 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3993 break; 3994 case Instruction::Add: 3995 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1), 3996 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3997 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3998 TLI, DT, AC, I); 3999 break; 4000 case Instruction::FSub: 4001 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1), 4002 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4003 break; 4004 case Instruction::Sub: 4005 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1), 4006 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4007 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 4008 TLI, DT, AC, I); 4009 break; 4010 case Instruction::FMul: 4011 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1), 4012 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4013 break; 4014 case Instruction::Mul: 4015 Result = 4016 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4017 break; 4018 case Instruction::SDiv: 4019 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4020 AC, I); 4021 break; 4022 case Instruction::UDiv: 4023 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4024 AC, I); 4025 break; 4026 case Instruction::FDiv: 4027 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), 4028 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4029 break; 4030 case Instruction::SRem: 4031 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4032 AC, I); 4033 break; 4034 case Instruction::URem: 4035 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4036 AC, I); 4037 break; 4038 case Instruction::FRem: 4039 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), 4040 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4041 break; 4042 case Instruction::Shl: 4043 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1), 4044 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4045 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 4046 TLI, DT, AC, I); 4047 break; 4048 case Instruction::LShr: 4049 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1), 4050 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 4051 AC, I); 4052 break; 4053 case Instruction::AShr: 4054 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1), 4055 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 4056 AC, I); 4057 break; 4058 case Instruction::And: 4059 Result = 4060 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4061 break; 4062 case Instruction::Or: 4063 Result = 4064 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4065 break; 4066 case Instruction::Xor: 4067 Result = 4068 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4069 break; 4070 case Instruction::ICmp: 4071 Result = 4072 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0), 4073 I->getOperand(1), DL, TLI, DT, AC, I); 4074 break; 4075 case Instruction::FCmp: 4076 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), 4077 I->getOperand(0), I->getOperand(1), 4078 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4079 break; 4080 case Instruction::Select: 4081 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1), 4082 I->getOperand(2), DL, TLI, DT, AC, I); 4083 break; 4084 case Instruction::GetElementPtr: { 4085 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 4086 Result = SimplifyGEPInst(Ops, DL, TLI, DT, AC, I); 4087 break; 4088 } 4089 case Instruction::InsertValue: { 4090 InsertValueInst *IV = cast<InsertValueInst>(I); 4091 Result = SimplifyInsertValueInst(IV->getAggregateOperand(), 4092 IV->getInsertedValueOperand(), 4093 IV->getIndices(), DL, TLI, DT, AC, I); 4094 break; 4095 } 4096 case Instruction::ExtractValue: { 4097 auto *EVI = cast<ExtractValueInst>(I); 4098 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(), 4099 EVI->getIndices(), DL, TLI, DT, AC, I); 4100 break; 4101 } 4102 case Instruction::ExtractElement: { 4103 auto *EEI = cast<ExtractElementInst>(I); 4104 Result = SimplifyExtractElementInst( 4105 EEI->getVectorOperand(), EEI->getIndexOperand(), DL, TLI, DT, AC, I); 4106 break; 4107 } 4108 case Instruction::PHI: 4109 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I)); 4110 break; 4111 case Instruction::Call: { 4112 CallSite CS(cast<CallInst>(I)); 4113 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL, 4114 TLI, DT, AC, I); 4115 break; 4116 } 4117 case Instruction::Trunc: 4118 Result = 4119 SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT, AC, I); 4120 break; 4121 } 4122 4123 // In general, it is possible for computeKnownBits to determine all bits in a 4124 // value even when the operands are not all constants. 4125 if (!Result && I->getType()->isIntegerTy()) { 4126 unsigned BitWidth = I->getType()->getScalarSizeInBits(); 4127 APInt KnownZero(BitWidth, 0); 4128 APInt KnownOne(BitWidth, 0); 4129 computeKnownBits(I, KnownZero, KnownOne, DL, /*Depth*/0, AC, I, DT); 4130 if ((KnownZero | KnownOne).isAllOnesValue()) 4131 Result = ConstantInt::get(I->getContext(), KnownOne); 4132 } 4133 4134 /// If called on unreachable code, the above logic may report that the 4135 /// instruction simplified to itself. Make life easier for users by 4136 /// detecting that case here, returning a safe value instead. 4137 return Result == I ? UndefValue::get(I->getType()) : Result; 4138 } 4139 4140 /// \brief Implementation of recursive simplification through an instructions 4141 /// uses. 4142 /// 4143 /// This is the common implementation of the recursive simplification routines. 4144 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to 4145 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of 4146 /// instructions to process and attempt to simplify it using 4147 /// InstructionSimplify. 4148 /// 4149 /// This routine returns 'true' only when *it* simplifies something. The passed 4150 /// in simplified value does not count toward this. 4151 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV, 4152 const TargetLibraryInfo *TLI, 4153 const DominatorTree *DT, 4154 AssumptionCache *AC) { 4155 bool Simplified = false; 4156 SmallSetVector<Instruction *, 8> Worklist; 4157 const DataLayout &DL = I->getModule()->getDataLayout(); 4158 4159 // If we have an explicit value to collapse to, do that round of the 4160 // simplification loop by hand initially. 4161 if (SimpleV) { 4162 for (User *U : I->users()) 4163 if (U != I) 4164 Worklist.insert(cast<Instruction>(U)); 4165 4166 // Replace the instruction with its simplified value. 4167 I->replaceAllUsesWith(SimpleV); 4168 4169 // Gracefully handle edge cases where the instruction is not wired into any 4170 // parent block. 4171 if (I->getParent()) 4172 I->eraseFromParent(); 4173 } else { 4174 Worklist.insert(I); 4175 } 4176 4177 // Note that we must test the size on each iteration, the worklist can grow. 4178 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { 4179 I = Worklist[Idx]; 4180 4181 // See if this instruction simplifies. 4182 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC); 4183 if (!SimpleV) 4184 continue; 4185 4186 Simplified = true; 4187 4188 // Stash away all the uses of the old instruction so we can check them for 4189 // recursive simplifications after a RAUW. This is cheaper than checking all 4190 // uses of To on the recursive step in most cases. 4191 for (User *U : I->users()) 4192 Worklist.insert(cast<Instruction>(U)); 4193 4194 // Replace the instruction with its simplified value. 4195 I->replaceAllUsesWith(SimpleV); 4196 4197 // Gracefully handle edge cases where the instruction is not wired into any 4198 // parent block. 4199 if (I->getParent()) 4200 I->eraseFromParent(); 4201 } 4202 return Simplified; 4203 } 4204 4205 bool llvm::recursivelySimplifyInstruction(Instruction *I, 4206 const TargetLibraryInfo *TLI, 4207 const DominatorTree *DT, 4208 AssumptionCache *AC) { 4209 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC); 4210 } 4211 4212 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV, 4213 const TargetLibraryInfo *TLI, 4214 const DominatorTree *DT, 4215 AssumptionCache *AC) { 4216 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!"); 4217 assert(SimpleV && "Must provide a simplified value."); 4218 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC); 4219 } 4220