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 static ConstantRange GetConstantRangeFromMetadata(MDNode *Ranges, uint32_t BitWidth) { 2180 const unsigned NumRanges = Ranges->getNumOperands() / 2; 2181 assert(NumRanges >= 1); 2182 2183 ConstantRange CR(BitWidth, false); 2184 for (unsigned i = 0; i < NumRanges; ++i) { 2185 auto *Low = 2186 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0)); 2187 auto *High = 2188 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1)); 2189 2190 // Union will merge two ranges to one and potentially introduce a range 2191 // not covered by the original two ranges. For example, [1, 5) and [8, 10) 2192 // will become [1, 10). In this case, we can not fold comparison between 2193 // constant 6 and a value of the above ranges. In practice, most values 2194 // have only one range, so it might not be worth handling this by 2195 // introducing additional complexity. 2196 CR = CR.unionWith(ConstantRange(Low->getValue(), High->getValue())); 2197 } 2198 2199 return CR; 2200 } 2201 2202 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can 2203 /// fold the result. If not, this returns null. 2204 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2205 const Query &Q, unsigned MaxRecurse) { 2206 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 2207 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!"); 2208 2209 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 2210 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 2211 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 2212 2213 // If we have a constant, make sure it is on the RHS. 2214 std::swap(LHS, RHS); 2215 Pred = CmpInst::getSwappedPredicate(Pred); 2216 } 2217 2218 Type *ITy = GetCompareTy(LHS); // The return type. 2219 Type *OpTy = LHS->getType(); // The operand type. 2220 2221 // icmp X, X -> true/false 2222 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false 2223 // because X could be 0. 2224 if (LHS == RHS || isa<UndefValue>(RHS)) 2225 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred)); 2226 2227 // Special case logic when the operands have i1 type. 2228 if (OpTy->getScalarType()->isIntegerTy(1)) { 2229 switch (Pred) { 2230 default: break; 2231 case ICmpInst::ICMP_EQ: 2232 // X == 1 -> X 2233 if (match(RHS, m_One())) 2234 return LHS; 2235 break; 2236 case ICmpInst::ICMP_NE: 2237 // X != 0 -> X 2238 if (match(RHS, m_Zero())) 2239 return LHS; 2240 break; 2241 case ICmpInst::ICMP_UGT: 2242 // X >u 0 -> X 2243 if (match(RHS, m_Zero())) 2244 return LHS; 2245 break; 2246 case ICmpInst::ICMP_UGE: 2247 // X >=u 1 -> X 2248 if (match(RHS, m_One())) 2249 return LHS; 2250 if (implies(RHS, LHS)) 2251 return getTrue(ITy); 2252 break; 2253 case ICmpInst::ICMP_SLT: 2254 // X <s 0 -> X 2255 if (match(RHS, m_Zero())) 2256 return LHS; 2257 break; 2258 case ICmpInst::ICMP_SLE: 2259 // X <=s -1 -> X 2260 if (match(RHS, m_One())) 2261 return LHS; 2262 break; 2263 case ICmpInst::ICMP_ULE: 2264 if (implies(LHS, RHS)) 2265 return getTrue(ITy); 2266 break; 2267 } 2268 } 2269 2270 // If we are comparing with zero then try hard since this is a common case. 2271 if (match(RHS, m_Zero())) { 2272 bool LHSKnownNonNegative, LHSKnownNegative; 2273 switch (Pred) { 2274 default: llvm_unreachable("Unknown ICmp predicate!"); 2275 case ICmpInst::ICMP_ULT: 2276 return getFalse(ITy); 2277 case ICmpInst::ICMP_UGE: 2278 return getTrue(ITy); 2279 case ICmpInst::ICMP_EQ: 2280 case ICmpInst::ICMP_ULE: 2281 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2282 return getFalse(ITy); 2283 break; 2284 case ICmpInst::ICMP_NE: 2285 case ICmpInst::ICMP_UGT: 2286 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2287 return getTrue(ITy); 2288 break; 2289 case ICmpInst::ICMP_SLT: 2290 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2291 Q.CxtI, Q.DT); 2292 if (LHSKnownNegative) 2293 return getTrue(ITy); 2294 if (LHSKnownNonNegative) 2295 return getFalse(ITy); 2296 break; 2297 case ICmpInst::ICMP_SLE: 2298 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2299 Q.CxtI, Q.DT); 2300 if (LHSKnownNegative) 2301 return getTrue(ITy); 2302 if (LHSKnownNonNegative && 2303 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2304 return getFalse(ITy); 2305 break; 2306 case ICmpInst::ICMP_SGE: 2307 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2308 Q.CxtI, Q.DT); 2309 if (LHSKnownNegative) 2310 return getFalse(ITy); 2311 if (LHSKnownNonNegative) 2312 return getTrue(ITy); 2313 break; 2314 case ICmpInst::ICMP_SGT: 2315 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2316 Q.CxtI, Q.DT); 2317 if (LHSKnownNegative) 2318 return getFalse(ITy); 2319 if (LHSKnownNonNegative && 2320 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2321 return getTrue(ITy); 2322 break; 2323 } 2324 } 2325 2326 // See if we are doing a comparison with a constant integer. 2327 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2328 // Rule out tautological comparisons (eg., ult 0 or uge 0). 2329 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue()); 2330 if (RHS_CR.isEmptySet()) 2331 return ConstantInt::getFalse(CI->getContext()); 2332 if (RHS_CR.isFullSet()) 2333 return ConstantInt::getTrue(CI->getContext()); 2334 2335 // Many binary operators with constant RHS have easy to compute constant 2336 // range. Use them to check whether the comparison is a tautology. 2337 unsigned Width = CI->getBitWidth(); 2338 APInt Lower = APInt(Width, 0); 2339 APInt Upper = APInt(Width, 0); 2340 ConstantInt *CI2; 2341 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) { 2342 // 'urem x, CI2' produces [0, CI2). 2343 Upper = CI2->getValue(); 2344 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) { 2345 // 'srem x, CI2' produces (-|CI2|, |CI2|). 2346 Upper = CI2->getValue().abs(); 2347 Lower = (-Upper) + 1; 2348 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) { 2349 // 'udiv CI2, x' produces [0, CI2]. 2350 Upper = CI2->getValue() + 1; 2351 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) { 2352 // 'udiv x, CI2' produces [0, UINT_MAX / CI2]. 2353 APInt NegOne = APInt::getAllOnesValue(Width); 2354 if (!CI2->isZero()) 2355 Upper = NegOne.udiv(CI2->getValue()) + 1; 2356 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) { 2357 if (CI2->isMinSignedValue()) { 2358 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. 2359 Lower = CI2->getValue(); 2360 Upper = Lower.lshr(1) + 1; 2361 } else { 2362 // 'sdiv CI2, x' produces [-|CI2|, |CI2|]. 2363 Upper = CI2->getValue().abs() + 1; 2364 Lower = (-Upper) + 1; 2365 } 2366 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) { 2367 APInt IntMin = APInt::getSignedMinValue(Width); 2368 APInt IntMax = APInt::getSignedMaxValue(Width); 2369 APInt Val = CI2->getValue(); 2370 if (Val.isAllOnesValue()) { 2371 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] 2372 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2373 Lower = IntMin + 1; 2374 Upper = IntMax + 1; 2375 } else if (Val.countLeadingZeros() < Width - 1) { 2376 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2] 2377 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2378 Lower = IntMin.sdiv(Val); 2379 Upper = IntMax.sdiv(Val); 2380 if (Lower.sgt(Upper)) 2381 std::swap(Lower, Upper); 2382 Upper = Upper + 1; 2383 assert(Upper != Lower && "Upper part of range has wrapped!"); 2384 } 2385 } else if (match(LHS, m_NUWShl(m_ConstantInt(CI2), m_Value()))) { 2386 // 'shl nuw CI2, x' produces [CI2, CI2 << CLZ(CI2)] 2387 Lower = CI2->getValue(); 2388 Upper = Lower.shl(Lower.countLeadingZeros()) + 1; 2389 } else if (match(LHS, m_NSWShl(m_ConstantInt(CI2), m_Value()))) { 2390 if (CI2->isNegative()) { 2391 // 'shl nsw CI2, x' produces [CI2 << CLO(CI2)-1, CI2] 2392 unsigned ShiftAmount = CI2->getValue().countLeadingOnes() - 1; 2393 Lower = CI2->getValue().shl(ShiftAmount); 2394 Upper = CI2->getValue() + 1; 2395 } else { 2396 // 'shl nsw CI2, x' produces [CI2, CI2 << CLZ(CI2)-1] 2397 unsigned ShiftAmount = CI2->getValue().countLeadingZeros() - 1; 2398 Lower = CI2->getValue(); 2399 Upper = CI2->getValue().shl(ShiftAmount) + 1; 2400 } 2401 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) { 2402 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2]. 2403 APInt NegOne = APInt::getAllOnesValue(Width); 2404 if (CI2->getValue().ult(Width)) 2405 Upper = NegOne.lshr(CI2->getValue()) + 1; 2406 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) { 2407 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2]. 2408 unsigned ShiftAmount = Width - 1; 2409 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2410 ShiftAmount = CI2->getValue().countTrailingZeros(); 2411 Lower = CI2->getValue().lshr(ShiftAmount); 2412 Upper = CI2->getValue() + 1; 2413 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) { 2414 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2]. 2415 APInt IntMin = APInt::getSignedMinValue(Width); 2416 APInt IntMax = APInt::getSignedMaxValue(Width); 2417 if (CI2->getValue().ult(Width)) { 2418 Lower = IntMin.ashr(CI2->getValue()); 2419 Upper = IntMax.ashr(CI2->getValue()) + 1; 2420 } 2421 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) { 2422 unsigned ShiftAmount = Width - 1; 2423 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2424 ShiftAmount = CI2->getValue().countTrailingZeros(); 2425 if (CI2->isNegative()) { 2426 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)] 2427 Lower = CI2->getValue(); 2428 Upper = CI2->getValue().ashr(ShiftAmount) + 1; 2429 } else { 2430 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2] 2431 Lower = CI2->getValue().ashr(ShiftAmount); 2432 Upper = CI2->getValue() + 1; 2433 } 2434 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) { 2435 // 'or x, CI2' produces [CI2, UINT_MAX]. 2436 Lower = CI2->getValue(); 2437 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) { 2438 // 'and x, CI2' produces [0, CI2]. 2439 Upper = CI2->getValue() + 1; 2440 } else if (match(LHS, m_NUWAdd(m_Value(), m_ConstantInt(CI2)))) { 2441 // 'add nuw x, CI2' produces [CI2, UINT_MAX]. 2442 Lower = CI2->getValue(); 2443 } 2444 2445 ConstantRange LHS_CR = Lower != Upper ? ConstantRange(Lower, Upper) 2446 : ConstantRange(Width, true); 2447 2448 if (auto *I = dyn_cast<Instruction>(LHS)) 2449 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range)) 2450 LHS_CR = LHS_CR.intersectWith(GetConstantRangeFromMetadata(Ranges, Width)); 2451 2452 if (!LHS_CR.isFullSet()) { 2453 if (RHS_CR.contains(LHS_CR)) 2454 return ConstantInt::getTrue(RHS->getContext()); 2455 if (RHS_CR.inverse().contains(LHS_CR)) 2456 return ConstantInt::getFalse(RHS->getContext()); 2457 } 2458 } 2459 2460 // If both operands have range metadata, use the metadata 2461 // to simplify the comparison. 2462 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) { 2463 auto RHS_Instr = dyn_cast<Instruction>(RHS); 2464 auto LHS_Instr = dyn_cast<Instruction>(LHS); 2465 2466 if (RHS_Instr->getMetadata(LLVMContext::MD_range) && 2467 LHS_Instr->getMetadata(LLVMContext::MD_range)) { 2468 uint32_t BitWidth = Q.DL.getTypeSizeInBits(RHS->getType()); 2469 2470 auto RHS_CR = GetConstantRangeFromMetadata( 2471 RHS_Instr->getMetadata(LLVMContext::MD_range), BitWidth); 2472 auto LHS_CR = GetConstantRangeFromMetadata( 2473 LHS_Instr->getMetadata(LLVMContext::MD_range), BitWidth); 2474 2475 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR); 2476 if (Satisfied_CR.contains(LHS_CR)) 2477 return ConstantInt::getTrue(RHS->getContext()); 2478 2479 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion( 2480 CmpInst::getInversePredicate(Pred), RHS_CR); 2481 if (InversedSatisfied_CR.contains(LHS_CR)) 2482 return ConstantInt::getFalse(RHS->getContext()); 2483 } 2484 } 2485 2486 // Compare of cast, for example (zext X) != 0 -> X != 0 2487 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) { 2488 Instruction *LI = cast<CastInst>(LHS); 2489 Value *SrcOp = LI->getOperand(0); 2490 Type *SrcTy = SrcOp->getType(); 2491 Type *DstTy = LI->getType(); 2492 2493 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input 2494 // if the integer type is the same size as the pointer type. 2495 if (MaxRecurse && isa<PtrToIntInst>(LI) && 2496 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) { 2497 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 2498 // Transfer the cast to the constant. 2499 if (Value *V = SimplifyICmpInst(Pred, SrcOp, 2500 ConstantExpr::getIntToPtr(RHSC, SrcTy), 2501 Q, MaxRecurse-1)) 2502 return V; 2503 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) { 2504 if (RI->getOperand(0)->getType() == SrcTy) 2505 // Compare without the cast. 2506 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2507 Q, MaxRecurse-1)) 2508 return V; 2509 } 2510 } 2511 2512 if (isa<ZExtInst>(LHS)) { 2513 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the 2514 // same type. 2515 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) { 2516 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2517 // Compare X and Y. Note that signed predicates become unsigned. 2518 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2519 SrcOp, RI->getOperand(0), Q, 2520 MaxRecurse-1)) 2521 return V; 2522 } 2523 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended 2524 // too. If not, then try to deduce the result of the comparison. 2525 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2526 // Compute the constant that would happen if we truncated to SrcTy then 2527 // reextended to DstTy. 2528 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2529 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy); 2530 2531 // If the re-extended constant didn't change then this is effectively 2532 // also a case of comparing two zero-extended values. 2533 if (RExt == CI && MaxRecurse) 2534 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2535 SrcOp, Trunc, Q, MaxRecurse-1)) 2536 return V; 2537 2538 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit 2539 // there. Use this to work out the result of the comparison. 2540 if (RExt != CI) { 2541 switch (Pred) { 2542 default: llvm_unreachable("Unknown ICmp predicate!"); 2543 // LHS <u RHS. 2544 case ICmpInst::ICMP_EQ: 2545 case ICmpInst::ICMP_UGT: 2546 case ICmpInst::ICMP_UGE: 2547 return ConstantInt::getFalse(CI->getContext()); 2548 2549 case ICmpInst::ICMP_NE: 2550 case ICmpInst::ICMP_ULT: 2551 case ICmpInst::ICMP_ULE: 2552 return ConstantInt::getTrue(CI->getContext()); 2553 2554 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS 2555 // is non-negative then LHS <s RHS. 2556 case ICmpInst::ICMP_SGT: 2557 case ICmpInst::ICMP_SGE: 2558 return CI->getValue().isNegative() ? 2559 ConstantInt::getTrue(CI->getContext()) : 2560 ConstantInt::getFalse(CI->getContext()); 2561 2562 case ICmpInst::ICMP_SLT: 2563 case ICmpInst::ICMP_SLE: 2564 return CI->getValue().isNegative() ? 2565 ConstantInt::getFalse(CI->getContext()) : 2566 ConstantInt::getTrue(CI->getContext()); 2567 } 2568 } 2569 } 2570 } 2571 2572 if (isa<SExtInst>(LHS)) { 2573 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the 2574 // same type. 2575 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) { 2576 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2577 // Compare X and Y. Note that the predicate does not change. 2578 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2579 Q, MaxRecurse-1)) 2580 return V; 2581 } 2582 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended 2583 // too. If not, then try to deduce the result of the comparison. 2584 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2585 // Compute the constant that would happen if we truncated to SrcTy then 2586 // reextended to DstTy. 2587 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2588 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy); 2589 2590 // If the re-extended constant didn't change then this is effectively 2591 // also a case of comparing two sign-extended values. 2592 if (RExt == CI && MaxRecurse) 2593 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1)) 2594 return V; 2595 2596 // Otherwise the upper bits of LHS are all equal, while RHS has varying 2597 // bits there. Use this to work out the result of the comparison. 2598 if (RExt != CI) { 2599 switch (Pred) { 2600 default: llvm_unreachable("Unknown ICmp predicate!"); 2601 case ICmpInst::ICMP_EQ: 2602 return ConstantInt::getFalse(CI->getContext()); 2603 case ICmpInst::ICMP_NE: 2604 return ConstantInt::getTrue(CI->getContext()); 2605 2606 // If RHS is non-negative then LHS <s RHS. If RHS is negative then 2607 // LHS >s RHS. 2608 case ICmpInst::ICMP_SGT: 2609 case ICmpInst::ICMP_SGE: 2610 return CI->getValue().isNegative() ? 2611 ConstantInt::getTrue(CI->getContext()) : 2612 ConstantInt::getFalse(CI->getContext()); 2613 case ICmpInst::ICMP_SLT: 2614 case ICmpInst::ICMP_SLE: 2615 return CI->getValue().isNegative() ? 2616 ConstantInt::getFalse(CI->getContext()) : 2617 ConstantInt::getTrue(CI->getContext()); 2618 2619 // If LHS is non-negative then LHS <u RHS. If LHS is negative then 2620 // LHS >u RHS. 2621 case ICmpInst::ICMP_UGT: 2622 case ICmpInst::ICMP_UGE: 2623 // Comparison is true iff the LHS <s 0. 2624 if (MaxRecurse) 2625 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp, 2626 Constant::getNullValue(SrcTy), 2627 Q, MaxRecurse-1)) 2628 return V; 2629 break; 2630 case ICmpInst::ICMP_ULT: 2631 case ICmpInst::ICMP_ULE: 2632 // Comparison is true iff the LHS >=s 0. 2633 if (MaxRecurse) 2634 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp, 2635 Constant::getNullValue(SrcTy), 2636 Q, MaxRecurse-1)) 2637 return V; 2638 break; 2639 } 2640 } 2641 } 2642 } 2643 } 2644 2645 // Special logic for binary operators. 2646 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS); 2647 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS); 2648 if (MaxRecurse && (LBO || RBO)) { 2649 // Analyze the case when either LHS or RHS is an add instruction. 2650 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 2651 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null). 2652 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false; 2653 if (LBO && LBO->getOpcode() == Instruction::Add) { 2654 A = LBO->getOperand(0); B = LBO->getOperand(1); 2655 NoLHSWrapProblem = ICmpInst::isEquality(Pred) || 2656 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) || 2657 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap()); 2658 } 2659 if (RBO && RBO->getOpcode() == Instruction::Add) { 2660 C = RBO->getOperand(0); D = RBO->getOperand(1); 2661 NoRHSWrapProblem = ICmpInst::isEquality(Pred) || 2662 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) || 2663 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap()); 2664 } 2665 2666 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2667 if ((A == RHS || B == RHS) && NoLHSWrapProblem) 2668 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A, 2669 Constant::getNullValue(RHS->getType()), 2670 Q, MaxRecurse-1)) 2671 return V; 2672 2673 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2674 if ((C == LHS || D == LHS) && NoRHSWrapProblem) 2675 if (Value *V = SimplifyICmpInst(Pred, 2676 Constant::getNullValue(LHS->getType()), 2677 C == LHS ? D : C, Q, MaxRecurse-1)) 2678 return V; 2679 2680 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow. 2681 if (A && C && (A == C || A == D || B == C || B == D) && 2682 NoLHSWrapProblem && NoRHSWrapProblem) { 2683 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2684 Value *Y, *Z; 2685 if (A == C) { 2686 // C + B == C + D -> B == D 2687 Y = B; 2688 Z = D; 2689 } else if (A == D) { 2690 // D + B == C + D -> B == C 2691 Y = B; 2692 Z = C; 2693 } else if (B == C) { 2694 // A + C == C + D -> A == D 2695 Y = A; 2696 Z = D; 2697 } else { 2698 assert(B == D); 2699 // A + D == C + D -> A == C 2700 Y = A; 2701 Z = C; 2702 } 2703 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1)) 2704 return V; 2705 } 2706 } 2707 2708 // icmp pred (or X, Y), X 2709 if (LBO && match(LBO, m_CombineOr(m_Or(m_Value(), m_Specific(RHS)), 2710 m_Or(m_Specific(RHS), m_Value())))) { 2711 if (Pred == ICmpInst::ICMP_ULT) 2712 return getFalse(ITy); 2713 if (Pred == ICmpInst::ICMP_UGE) 2714 return getTrue(ITy); 2715 } 2716 // icmp pred X, (or X, Y) 2717 if (RBO && match(RBO, m_CombineOr(m_Or(m_Value(), m_Specific(LHS)), 2718 m_Or(m_Specific(LHS), m_Value())))) { 2719 if (Pred == ICmpInst::ICMP_ULE) 2720 return getTrue(ITy); 2721 if (Pred == ICmpInst::ICMP_UGT) 2722 return getFalse(ITy); 2723 } 2724 2725 // icmp pred (and X, Y), X 2726 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)), 2727 m_And(m_Specific(RHS), m_Value())))) { 2728 if (Pred == ICmpInst::ICMP_UGT) 2729 return getFalse(ITy); 2730 if (Pred == ICmpInst::ICMP_ULE) 2731 return getTrue(ITy); 2732 } 2733 // icmp pred X, (and X, Y) 2734 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)), 2735 m_And(m_Specific(LHS), m_Value())))) { 2736 if (Pred == ICmpInst::ICMP_UGE) 2737 return getTrue(ITy); 2738 if (Pred == ICmpInst::ICMP_ULT) 2739 return getFalse(ITy); 2740 } 2741 2742 // 0 - (zext X) pred C 2743 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) { 2744 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 2745 if (RHSC->getValue().isStrictlyPositive()) { 2746 if (Pred == ICmpInst::ICMP_SLT) 2747 return ConstantInt::getTrue(RHSC->getContext()); 2748 if (Pred == ICmpInst::ICMP_SGE) 2749 return ConstantInt::getFalse(RHSC->getContext()); 2750 if (Pred == ICmpInst::ICMP_EQ) 2751 return ConstantInt::getFalse(RHSC->getContext()); 2752 if (Pred == ICmpInst::ICMP_NE) 2753 return ConstantInt::getTrue(RHSC->getContext()); 2754 } 2755 if (RHSC->getValue().isNonNegative()) { 2756 if (Pred == ICmpInst::ICMP_SLE) 2757 return ConstantInt::getTrue(RHSC->getContext()); 2758 if (Pred == ICmpInst::ICMP_SGT) 2759 return ConstantInt::getFalse(RHSC->getContext()); 2760 } 2761 } 2762 } 2763 2764 // icmp pred (urem X, Y), Y 2765 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) { 2766 bool KnownNonNegative, KnownNegative; 2767 switch (Pred) { 2768 default: 2769 break; 2770 case ICmpInst::ICMP_SGT: 2771 case ICmpInst::ICMP_SGE: 2772 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2773 Q.CxtI, Q.DT); 2774 if (!KnownNonNegative) 2775 break; 2776 // fall-through 2777 case ICmpInst::ICMP_EQ: 2778 case ICmpInst::ICMP_UGT: 2779 case ICmpInst::ICMP_UGE: 2780 return getFalse(ITy); 2781 case ICmpInst::ICMP_SLT: 2782 case ICmpInst::ICMP_SLE: 2783 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2784 Q.CxtI, Q.DT); 2785 if (!KnownNonNegative) 2786 break; 2787 // fall-through 2788 case ICmpInst::ICMP_NE: 2789 case ICmpInst::ICMP_ULT: 2790 case ICmpInst::ICMP_ULE: 2791 return getTrue(ITy); 2792 } 2793 } 2794 2795 // icmp pred X, (urem Y, X) 2796 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) { 2797 bool KnownNonNegative, KnownNegative; 2798 switch (Pred) { 2799 default: 2800 break; 2801 case ICmpInst::ICMP_SGT: 2802 case ICmpInst::ICMP_SGE: 2803 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2804 Q.CxtI, Q.DT); 2805 if (!KnownNonNegative) 2806 break; 2807 // fall-through 2808 case ICmpInst::ICMP_NE: 2809 case ICmpInst::ICMP_UGT: 2810 case ICmpInst::ICMP_UGE: 2811 return getTrue(ITy); 2812 case ICmpInst::ICMP_SLT: 2813 case ICmpInst::ICMP_SLE: 2814 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2815 Q.CxtI, Q.DT); 2816 if (!KnownNonNegative) 2817 break; 2818 // fall-through 2819 case ICmpInst::ICMP_EQ: 2820 case ICmpInst::ICMP_ULT: 2821 case ICmpInst::ICMP_ULE: 2822 return getFalse(ITy); 2823 } 2824 } 2825 2826 // x udiv y <=u x. 2827 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) { 2828 // icmp pred (X /u Y), X 2829 if (Pred == ICmpInst::ICMP_UGT) 2830 return getFalse(ITy); 2831 if (Pred == ICmpInst::ICMP_ULE) 2832 return getTrue(ITy); 2833 } 2834 2835 // handle: 2836 // CI2 << X == CI 2837 // CI2 << X != CI 2838 // 2839 // where CI2 is a power of 2 and CI isn't 2840 if (auto *CI = dyn_cast<ConstantInt>(RHS)) { 2841 const APInt *CI2Val, *CIVal = &CI->getValue(); 2842 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) && 2843 CI2Val->isPowerOf2()) { 2844 if (!CIVal->isPowerOf2()) { 2845 // CI2 << X can equal zero in some circumstances, 2846 // this simplification is unsafe if CI is zero. 2847 // 2848 // We know it is safe if: 2849 // - The shift is nsw, we can't shift out the one bit. 2850 // - The shift is nuw, we can't shift out the one bit. 2851 // - CI2 is one 2852 // - CI isn't zero 2853 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() || 2854 *CI2Val == 1 || !CI->isZero()) { 2855 if (Pred == ICmpInst::ICMP_EQ) 2856 return ConstantInt::getFalse(RHS->getContext()); 2857 if (Pred == ICmpInst::ICMP_NE) 2858 return ConstantInt::getTrue(RHS->getContext()); 2859 } 2860 } 2861 if (CIVal->isSignBit() && *CI2Val == 1) { 2862 if (Pred == ICmpInst::ICMP_UGT) 2863 return ConstantInt::getFalse(RHS->getContext()); 2864 if (Pred == ICmpInst::ICMP_ULE) 2865 return ConstantInt::getTrue(RHS->getContext()); 2866 } 2867 } 2868 } 2869 2870 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() && 2871 LBO->getOperand(1) == RBO->getOperand(1)) { 2872 switch (LBO->getOpcode()) { 2873 default: break; 2874 case Instruction::UDiv: 2875 case Instruction::LShr: 2876 if (ICmpInst::isSigned(Pred)) 2877 break; 2878 // fall-through 2879 case Instruction::SDiv: 2880 case Instruction::AShr: 2881 if (!LBO->isExact() || !RBO->isExact()) 2882 break; 2883 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2884 RBO->getOperand(0), Q, MaxRecurse-1)) 2885 return V; 2886 break; 2887 case Instruction::Shl: { 2888 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap(); 2889 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap(); 2890 if (!NUW && !NSW) 2891 break; 2892 if (!NSW && ICmpInst::isSigned(Pred)) 2893 break; 2894 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2895 RBO->getOperand(0), Q, MaxRecurse-1)) 2896 return V; 2897 break; 2898 } 2899 } 2900 } 2901 2902 // Simplify comparisons involving max/min. 2903 Value *A, *B; 2904 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE; 2905 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B". 2906 2907 // Signed variants on "max(a,b)>=a -> true". 2908 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2909 if (A != RHS) std::swap(A, B); // smax(A, B) pred A. 2910 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2911 // We analyze this as smax(A, B) pred A. 2912 P = Pred; 2913 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) && 2914 (A == LHS || B == LHS)) { 2915 if (A != LHS) std::swap(A, B); // A pred smax(A, B). 2916 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2917 // We analyze this as smax(A, B) swapped-pred A. 2918 P = CmpInst::getSwappedPredicate(Pred); 2919 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2920 (A == RHS || B == RHS)) { 2921 if (A != RHS) std::swap(A, B); // smin(A, B) pred A. 2922 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2923 // We analyze this as smax(-A, -B) swapped-pred -A. 2924 // Note that we do not need to actually form -A or -B thanks to EqP. 2925 P = CmpInst::getSwappedPredicate(Pred); 2926 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) && 2927 (A == LHS || B == LHS)) { 2928 if (A != LHS) std::swap(A, B); // A pred smin(A, B). 2929 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2930 // We analyze this as smax(-A, -B) pred -A. 2931 // Note that we do not need to actually form -A or -B thanks to EqP. 2932 P = Pred; 2933 } 2934 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2935 // Cases correspond to "max(A, B) p A". 2936 switch (P) { 2937 default: 2938 break; 2939 case CmpInst::ICMP_EQ: 2940 case CmpInst::ICMP_SLE: 2941 // Equivalent to "A EqP B". This may be the same as the condition tested 2942 // in the max/min; if so, we can just return that. 2943 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2944 return V; 2945 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2946 return V; 2947 // Otherwise, see if "A EqP B" simplifies. 2948 if (MaxRecurse) 2949 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2950 return V; 2951 break; 2952 case CmpInst::ICMP_NE: 2953 case CmpInst::ICMP_SGT: { 2954 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2955 // Equivalent to "A InvEqP B". This may be the same as the condition 2956 // tested in the max/min; if so, we can just return that. 2957 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2958 return V; 2959 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2960 return V; 2961 // Otherwise, see if "A InvEqP B" simplifies. 2962 if (MaxRecurse) 2963 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2964 return V; 2965 break; 2966 } 2967 case CmpInst::ICMP_SGE: 2968 // Always true. 2969 return getTrue(ITy); 2970 case CmpInst::ICMP_SLT: 2971 // Always false. 2972 return getFalse(ITy); 2973 } 2974 } 2975 2976 // Unsigned variants on "max(a,b)>=a -> true". 2977 P = CmpInst::BAD_ICMP_PREDICATE; 2978 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2979 if (A != RHS) std::swap(A, B); // umax(A, B) pred A. 2980 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2981 // We analyze this as umax(A, B) pred A. 2982 P = Pred; 2983 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) && 2984 (A == LHS || B == LHS)) { 2985 if (A != LHS) std::swap(A, B); // A pred umax(A, B). 2986 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2987 // We analyze this as umax(A, B) swapped-pred A. 2988 P = CmpInst::getSwappedPredicate(Pred); 2989 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2990 (A == RHS || B == RHS)) { 2991 if (A != RHS) std::swap(A, B); // umin(A, B) pred A. 2992 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2993 // We analyze this as umax(-A, -B) swapped-pred -A. 2994 // Note that we do not need to actually form -A or -B thanks to EqP. 2995 P = CmpInst::getSwappedPredicate(Pred); 2996 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) && 2997 (A == LHS || B == LHS)) { 2998 if (A != LHS) std::swap(A, B); // A pred umin(A, B). 2999 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 3000 // We analyze this as umax(-A, -B) pred -A. 3001 // Note that we do not need to actually form -A or -B thanks to EqP. 3002 P = Pred; 3003 } 3004 if (P != CmpInst::BAD_ICMP_PREDICATE) { 3005 // Cases correspond to "max(A, B) p A". 3006 switch (P) { 3007 default: 3008 break; 3009 case CmpInst::ICMP_EQ: 3010 case CmpInst::ICMP_ULE: 3011 // Equivalent to "A EqP B". This may be the same as the condition tested 3012 // in the max/min; if so, we can just return that. 3013 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 3014 return V; 3015 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 3016 return V; 3017 // Otherwise, see if "A EqP B" simplifies. 3018 if (MaxRecurse) 3019 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 3020 return V; 3021 break; 3022 case CmpInst::ICMP_NE: 3023 case CmpInst::ICMP_UGT: { 3024 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 3025 // Equivalent to "A InvEqP B". This may be the same as the condition 3026 // tested in the max/min; if so, we can just return that. 3027 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 3028 return V; 3029 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 3030 return V; 3031 // Otherwise, see if "A InvEqP B" simplifies. 3032 if (MaxRecurse) 3033 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 3034 return V; 3035 break; 3036 } 3037 case CmpInst::ICMP_UGE: 3038 // Always true. 3039 return getTrue(ITy); 3040 case CmpInst::ICMP_ULT: 3041 // Always false. 3042 return getFalse(ITy); 3043 } 3044 } 3045 3046 // Variants on "max(x,y) >= min(x,z)". 3047 Value *C, *D; 3048 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && 3049 match(RHS, m_SMin(m_Value(C), m_Value(D))) && 3050 (A == C || A == D || B == C || B == D)) { 3051 // max(x, ?) pred min(x, ?). 3052 if (Pred == CmpInst::ICMP_SGE) 3053 // Always true. 3054 return getTrue(ITy); 3055 if (Pred == CmpInst::ICMP_SLT) 3056 // Always false. 3057 return getFalse(ITy); 3058 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 3059 match(RHS, m_SMax(m_Value(C), m_Value(D))) && 3060 (A == C || A == D || B == C || B == D)) { 3061 // min(x, ?) pred max(x, ?). 3062 if (Pred == CmpInst::ICMP_SLE) 3063 // Always true. 3064 return getTrue(ITy); 3065 if (Pred == CmpInst::ICMP_SGT) 3066 // Always false. 3067 return getFalse(ITy); 3068 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && 3069 match(RHS, m_UMin(m_Value(C), m_Value(D))) && 3070 (A == C || A == D || B == C || B == D)) { 3071 // max(x, ?) pred min(x, ?). 3072 if (Pred == CmpInst::ICMP_UGE) 3073 // Always true. 3074 return getTrue(ITy); 3075 if (Pred == CmpInst::ICMP_ULT) 3076 // Always false. 3077 return getFalse(ITy); 3078 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 3079 match(RHS, m_UMax(m_Value(C), m_Value(D))) && 3080 (A == C || A == D || B == C || B == D)) { 3081 // min(x, ?) pred max(x, ?). 3082 if (Pred == CmpInst::ICMP_ULE) 3083 // Always true. 3084 return getTrue(ITy); 3085 if (Pred == CmpInst::ICMP_UGT) 3086 // Always false. 3087 return getFalse(ITy); 3088 } 3089 3090 // Simplify comparisons of related pointers using a powerful, recursive 3091 // GEP-walk when we have target data available.. 3092 if (LHS->getType()->isPointerTy()) 3093 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS)) 3094 return C; 3095 3096 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) { 3097 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) { 3098 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() && 3099 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() && 3100 (ICmpInst::isEquality(Pred) || 3101 (GLHS->isInBounds() && GRHS->isInBounds() && 3102 Pred == ICmpInst::getSignedPredicate(Pred)))) { 3103 // The bases are equal and the indices are constant. Build a constant 3104 // expression GEP with the same indices and a null base pointer to see 3105 // what constant folding can make out of it. 3106 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType()); 3107 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end()); 3108 Constant *NewLHS = ConstantExpr::getGetElementPtr( 3109 GLHS->getSourceElementType(), Null, IndicesLHS); 3110 3111 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end()); 3112 Constant *NewRHS = ConstantExpr::getGetElementPtr( 3113 GLHS->getSourceElementType(), Null, IndicesRHS); 3114 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS); 3115 } 3116 } 3117 } 3118 3119 // If a bit is known to be zero for A and known to be one for B, 3120 // then A and B cannot be equal. 3121 if (ICmpInst::isEquality(Pred)) { 3122 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 3123 uint32_t BitWidth = CI->getBitWidth(); 3124 APInt LHSKnownZero(BitWidth, 0); 3125 APInt LHSKnownOne(BitWidth, 0); 3126 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC, 3127 Q.CxtI, Q.DT); 3128 const APInt &RHSVal = CI->getValue(); 3129 if (((LHSKnownZero & RHSVal) != 0) || ((LHSKnownOne & ~RHSVal) != 0)) 3130 return Pred == ICmpInst::ICMP_EQ 3131 ? ConstantInt::getFalse(CI->getContext()) 3132 : ConstantInt::getTrue(CI->getContext()); 3133 } 3134 } 3135 3136 // If the comparison is with the result of a select instruction, check whether 3137 // comparing with either branch of the select always yields the same value. 3138 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3139 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3140 return V; 3141 3142 // If the comparison is with the result of a phi instruction, check whether 3143 // doing the compare with each incoming phi value yields a common result. 3144 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3145 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3146 return V; 3147 3148 return nullptr; 3149 } 3150 3151 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3152 const DataLayout &DL, 3153 const TargetLibraryInfo *TLI, 3154 const DominatorTree *DT, AssumptionCache *AC, 3155 Instruction *CxtI) { 3156 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3157 RecursionLimit); 3158 } 3159 3160 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can 3161 /// fold the result. If not, this returns null. 3162 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3163 FastMathFlags FMF, const Query &Q, 3164 unsigned MaxRecurse) { 3165 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 3166 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!"); 3167 3168 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 3169 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 3170 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 3171 3172 // If we have a constant, make sure it is on the RHS. 3173 std::swap(LHS, RHS); 3174 Pred = CmpInst::getSwappedPredicate(Pred); 3175 } 3176 3177 // Fold trivial predicates. 3178 if (Pred == FCmpInst::FCMP_FALSE) 3179 return ConstantInt::get(GetCompareTy(LHS), 0); 3180 if (Pred == FCmpInst::FCMP_TRUE) 3181 return ConstantInt::get(GetCompareTy(LHS), 1); 3182 3183 // UNO/ORD predicates can be trivially folded if NaNs are ignored. 3184 if (FMF.noNaNs()) { 3185 if (Pred == FCmpInst::FCMP_UNO) 3186 return ConstantInt::get(GetCompareTy(LHS), 0); 3187 if (Pred == FCmpInst::FCMP_ORD) 3188 return ConstantInt::get(GetCompareTy(LHS), 1); 3189 } 3190 3191 // fcmp pred x, undef and fcmp pred undef, x 3192 // fold to true if unordered, false if ordered 3193 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) { 3194 // Choosing NaN for the undef will always make unordered comparison succeed 3195 // and ordered comparison fail. 3196 return ConstantInt::get(GetCompareTy(LHS), CmpInst::isUnordered(Pred)); 3197 } 3198 3199 // fcmp x,x -> true/false. Not all compares are foldable. 3200 if (LHS == RHS) { 3201 if (CmpInst::isTrueWhenEqual(Pred)) 3202 return ConstantInt::get(GetCompareTy(LHS), 1); 3203 if (CmpInst::isFalseWhenEqual(Pred)) 3204 return ConstantInt::get(GetCompareTy(LHS), 0); 3205 } 3206 3207 // Handle fcmp with constant RHS 3208 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) { 3209 // If the constant is a nan, see if we can fold the comparison based on it. 3210 if (CFP->getValueAPF().isNaN()) { 3211 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo" 3212 return ConstantInt::getFalse(CFP->getContext()); 3213 assert(FCmpInst::isUnordered(Pred) && 3214 "Comparison must be either ordered or unordered!"); 3215 // True if unordered. 3216 return ConstantInt::getTrue(CFP->getContext()); 3217 } 3218 // Check whether the constant is an infinity. 3219 if (CFP->getValueAPF().isInfinity()) { 3220 if (CFP->getValueAPF().isNegative()) { 3221 switch (Pred) { 3222 case FCmpInst::FCMP_OLT: 3223 // No value is ordered and less than negative infinity. 3224 return ConstantInt::getFalse(CFP->getContext()); 3225 case FCmpInst::FCMP_UGE: 3226 // All values are unordered with or at least negative infinity. 3227 return ConstantInt::getTrue(CFP->getContext()); 3228 default: 3229 break; 3230 } 3231 } else { 3232 switch (Pred) { 3233 case FCmpInst::FCMP_OGT: 3234 // No value is ordered and greater than infinity. 3235 return ConstantInt::getFalse(CFP->getContext()); 3236 case FCmpInst::FCMP_ULE: 3237 // All values are unordered with and at most infinity. 3238 return ConstantInt::getTrue(CFP->getContext()); 3239 default: 3240 break; 3241 } 3242 } 3243 } 3244 if (CFP->getValueAPF().isZero()) { 3245 switch (Pred) { 3246 case FCmpInst::FCMP_UGE: 3247 if (CannotBeOrderedLessThanZero(LHS)) 3248 return ConstantInt::getTrue(CFP->getContext()); 3249 break; 3250 case FCmpInst::FCMP_OLT: 3251 // X < 0 3252 if (CannotBeOrderedLessThanZero(LHS)) 3253 return ConstantInt::getFalse(CFP->getContext()); 3254 break; 3255 default: 3256 break; 3257 } 3258 } 3259 } 3260 3261 // If the comparison is with the result of a select instruction, check whether 3262 // comparing with either branch of the select always yields the same value. 3263 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3264 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3265 return V; 3266 3267 // If the comparison is with the result of a phi instruction, check whether 3268 // doing the compare with each incoming phi value yields a common result. 3269 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3270 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3271 return V; 3272 3273 return nullptr; 3274 } 3275 3276 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3277 FastMathFlags FMF, const DataLayout &DL, 3278 const TargetLibraryInfo *TLI, 3279 const DominatorTree *DT, AssumptionCache *AC, 3280 const Instruction *CxtI) { 3281 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, 3282 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3283 } 3284 3285 /// SimplifyWithOpReplaced - See if V simplifies when its operand Op is 3286 /// replaced with RepOp. 3287 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp, 3288 const Query &Q, 3289 unsigned MaxRecurse) { 3290 // Trivial replacement. 3291 if (V == Op) 3292 return RepOp; 3293 3294 auto *I = dyn_cast<Instruction>(V); 3295 if (!I) 3296 return nullptr; 3297 3298 // If this is a binary operator, try to simplify it with the replaced op. 3299 if (auto *B = dyn_cast<BinaryOperator>(I)) { 3300 // Consider: 3301 // %cmp = icmp eq i32 %x, 2147483647 3302 // %add = add nsw i32 %x, 1 3303 // %sel = select i1 %cmp, i32 -2147483648, i32 %add 3304 // 3305 // We can't replace %sel with %add unless we strip away the flags. 3306 if (isa<OverflowingBinaryOperator>(B)) 3307 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap()) 3308 return nullptr; 3309 if (isa<PossiblyExactOperator>(B)) 3310 if (B->isExact()) 3311 return nullptr; 3312 3313 if (MaxRecurse) { 3314 if (B->getOperand(0) == Op) 3315 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q, 3316 MaxRecurse - 1); 3317 if (B->getOperand(1) == Op) 3318 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q, 3319 MaxRecurse - 1); 3320 } 3321 } 3322 3323 // Same for CmpInsts. 3324 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 3325 if (MaxRecurse) { 3326 if (C->getOperand(0) == Op) 3327 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q, 3328 MaxRecurse - 1); 3329 if (C->getOperand(1) == Op) 3330 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q, 3331 MaxRecurse - 1); 3332 } 3333 } 3334 3335 // TODO: We could hand off more cases to instsimplify here. 3336 3337 // If all operands are constant after substituting Op for RepOp then we can 3338 // constant fold the instruction. 3339 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) { 3340 // Build a list of all constant operands. 3341 SmallVector<Constant *, 8> ConstOps; 3342 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 3343 if (I->getOperand(i) == Op) 3344 ConstOps.push_back(CRepOp); 3345 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i))) 3346 ConstOps.push_back(COp); 3347 else 3348 break; 3349 } 3350 3351 // All operands were constants, fold it. 3352 if (ConstOps.size() == I->getNumOperands()) { 3353 if (CmpInst *C = dyn_cast<CmpInst>(I)) 3354 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0], 3355 ConstOps[1], Q.DL, Q.TLI); 3356 3357 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 3358 if (!LI->isVolatile()) 3359 return ConstantFoldLoadFromConstPtr(ConstOps[0], Q.DL); 3360 3361 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), ConstOps, 3362 Q.DL, Q.TLI); 3363 } 3364 } 3365 3366 return nullptr; 3367 } 3368 3369 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold 3370 /// the result. If not, this returns null. 3371 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal, 3372 Value *FalseVal, const Query &Q, 3373 unsigned MaxRecurse) { 3374 // select true, X, Y -> X 3375 // select false, X, Y -> Y 3376 if (Constant *CB = dyn_cast<Constant>(CondVal)) { 3377 if (CB->isAllOnesValue()) 3378 return TrueVal; 3379 if (CB->isNullValue()) 3380 return FalseVal; 3381 } 3382 3383 // select C, X, X -> X 3384 if (TrueVal == FalseVal) 3385 return TrueVal; 3386 3387 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y 3388 if (isa<Constant>(TrueVal)) 3389 return TrueVal; 3390 return FalseVal; 3391 } 3392 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X 3393 return FalseVal; 3394 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X 3395 return TrueVal; 3396 3397 if (const auto *ICI = dyn_cast<ICmpInst>(CondVal)) { 3398 unsigned BitWidth = Q.DL.getTypeSizeInBits(TrueVal->getType()); 3399 ICmpInst::Predicate Pred = ICI->getPredicate(); 3400 Value *CmpLHS = ICI->getOperand(0); 3401 Value *CmpRHS = ICI->getOperand(1); 3402 APInt MinSignedValue = APInt::getSignBit(BitWidth); 3403 Value *X; 3404 const APInt *Y; 3405 bool TrueWhenUnset; 3406 bool IsBitTest = false; 3407 if (ICmpInst::isEquality(Pred) && 3408 match(CmpLHS, m_And(m_Value(X), m_APInt(Y))) && 3409 match(CmpRHS, m_Zero())) { 3410 IsBitTest = true; 3411 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 3412 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { 3413 X = CmpLHS; 3414 Y = &MinSignedValue; 3415 IsBitTest = true; 3416 TrueWhenUnset = false; 3417 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { 3418 X = CmpLHS; 3419 Y = &MinSignedValue; 3420 IsBitTest = true; 3421 TrueWhenUnset = true; 3422 } 3423 if (IsBitTest) { 3424 const APInt *C; 3425 // (X & Y) == 0 ? X & ~Y : X --> X 3426 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y 3427 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) && 3428 *Y == ~*C) 3429 return TrueWhenUnset ? FalseVal : TrueVal; 3430 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y 3431 // (X & Y) != 0 ? X : X & ~Y --> X 3432 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) && 3433 *Y == ~*C) 3434 return TrueWhenUnset ? FalseVal : TrueVal; 3435 3436 if (Y->isPowerOf2()) { 3437 // (X & Y) == 0 ? X | Y : X --> X | Y 3438 // (X & Y) != 0 ? X | Y : X --> X 3439 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) && 3440 *Y == *C) 3441 return TrueWhenUnset ? TrueVal : FalseVal; 3442 // (X & Y) == 0 ? X : X | Y --> X 3443 // (X & Y) != 0 ? X : X | Y --> X | Y 3444 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) && 3445 *Y == *C) 3446 return TrueWhenUnset ? TrueVal : FalseVal; 3447 } 3448 } 3449 if (ICI->hasOneUse()) { 3450 const APInt *C; 3451 if (match(CmpRHS, m_APInt(C))) { 3452 // X < MIN ? T : F --> F 3453 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue()) 3454 return FalseVal; 3455 // X < MIN ? T : F --> F 3456 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue()) 3457 return FalseVal; 3458 // X > MAX ? T : F --> F 3459 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue()) 3460 return FalseVal; 3461 // X > MAX ? T : F --> F 3462 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue()) 3463 return FalseVal; 3464 } 3465 } 3466 3467 // If we have an equality comparison then we know the value in one of the 3468 // arms of the select. See if substituting this value into the arm and 3469 // simplifying the result yields the same value as the other arm. 3470 if (Pred == ICmpInst::ICMP_EQ) { 3471 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3472 TrueVal || 3473 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3474 TrueVal) 3475 return FalseVal; 3476 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3477 FalseVal || 3478 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3479 FalseVal) 3480 return FalseVal; 3481 } else if (Pred == ICmpInst::ICMP_NE) { 3482 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3483 FalseVal || 3484 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3485 FalseVal) 3486 return TrueVal; 3487 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == 3488 TrueVal || 3489 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == 3490 TrueVal) 3491 return TrueVal; 3492 } 3493 } 3494 3495 return nullptr; 3496 } 3497 3498 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, 3499 const DataLayout &DL, 3500 const TargetLibraryInfo *TLI, 3501 const DominatorTree *DT, AssumptionCache *AC, 3502 const Instruction *CxtI) { 3503 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, 3504 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3505 } 3506 3507 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can 3508 /// fold the result. If not, this returns null. 3509 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, 3510 const Query &Q, unsigned) { 3511 // The type of the GEP pointer operand. 3512 unsigned AS = 3513 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace(); 3514 3515 // getelementptr P -> P. 3516 if (Ops.size() == 1) 3517 return Ops[0]; 3518 3519 // Compute the (pointer) type returned by the GEP instruction. 3520 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1)); 3521 Type *GEPTy = PointerType::get(LastType, AS); 3522 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType())) 3523 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 3524 3525 if (isa<UndefValue>(Ops[0])) 3526 return UndefValue::get(GEPTy); 3527 3528 if (Ops.size() == 2) { 3529 // getelementptr P, 0 -> P. 3530 if (match(Ops[1], m_Zero())) 3531 return Ops[0]; 3532 3533 Type *Ty = SrcTy; 3534 if (Ty->isSized()) { 3535 Value *P; 3536 uint64_t C; 3537 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty); 3538 // getelementptr P, N -> P if P points to a type of zero size. 3539 if (TyAllocSize == 0) 3540 return Ops[0]; 3541 3542 // The following transforms are only safe if the ptrtoint cast 3543 // doesn't truncate the pointers. 3544 if (Ops[1]->getType()->getScalarSizeInBits() == 3545 Q.DL.getPointerSizeInBits(AS)) { 3546 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * { 3547 if (match(P, m_Zero())) 3548 return Constant::getNullValue(GEPTy); 3549 Value *Temp; 3550 if (match(P, m_PtrToInt(m_Value(Temp)))) 3551 if (Temp->getType() == GEPTy) 3552 return Temp; 3553 return nullptr; 3554 }; 3555 3556 // getelementptr V, (sub P, V) -> P if P points to a type of size 1. 3557 if (TyAllocSize == 1 && 3558 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))))) 3559 if (Value *R = PtrToIntOrZero(P)) 3560 return R; 3561 3562 // getelementptr V, (ashr (sub P, V), C) -> Q 3563 // if P points to a type of size 1 << C. 3564 if (match(Ops[1], 3565 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3566 m_ConstantInt(C))) && 3567 TyAllocSize == 1ULL << C) 3568 if (Value *R = PtrToIntOrZero(P)) 3569 return R; 3570 3571 // getelementptr V, (sdiv (sub P, V), C) -> Q 3572 // if P points to a type of size C. 3573 if (match(Ops[1], 3574 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3575 m_SpecificInt(TyAllocSize)))) 3576 if (Value *R = PtrToIntOrZero(P)) 3577 return R; 3578 } 3579 } 3580 } 3581 3582 // Check to see if this is constant foldable. 3583 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 3584 if (!isa<Constant>(Ops[i])) 3585 return nullptr; 3586 3587 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]), 3588 Ops.slice(1)); 3589 } 3590 3591 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout &DL, 3592 const TargetLibraryInfo *TLI, 3593 const DominatorTree *DT, AssumptionCache *AC, 3594 const Instruction *CxtI) { 3595 return ::SimplifyGEPInst( 3596 cast<PointerType>(Ops[0]->getType()->getScalarType())->getElementType(), 3597 Ops, Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3598 } 3599 3600 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we 3601 /// can fold the result. If not, this returns null. 3602 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val, 3603 ArrayRef<unsigned> Idxs, const Query &Q, 3604 unsigned) { 3605 if (Constant *CAgg = dyn_cast<Constant>(Agg)) 3606 if (Constant *CVal = dyn_cast<Constant>(Val)) 3607 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs); 3608 3609 // insertvalue x, undef, n -> x 3610 if (match(Val, m_Undef())) 3611 return Agg; 3612 3613 // insertvalue x, (extractvalue y, n), n 3614 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val)) 3615 if (EV->getAggregateOperand()->getType() == Agg->getType() && 3616 EV->getIndices() == Idxs) { 3617 // insertvalue undef, (extractvalue y, n), n -> y 3618 if (match(Agg, m_Undef())) 3619 return EV->getAggregateOperand(); 3620 3621 // insertvalue y, (extractvalue y, n), n -> y 3622 if (Agg == EV->getAggregateOperand()) 3623 return Agg; 3624 } 3625 3626 return nullptr; 3627 } 3628 3629 Value *llvm::SimplifyInsertValueInst( 3630 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout &DL, 3631 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC, 3632 const Instruction *CxtI) { 3633 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI), 3634 RecursionLimit); 3635 } 3636 3637 /// SimplifyExtractValueInst - Given operands for an ExtractValueInst, see if we 3638 /// can fold the result. If not, this returns null. 3639 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, 3640 const Query &, unsigned) { 3641 if (auto *CAgg = dyn_cast<Constant>(Agg)) 3642 return ConstantFoldExtractValueInstruction(CAgg, Idxs); 3643 3644 // extractvalue x, (insertvalue y, elt, n), n -> elt 3645 unsigned NumIdxs = Idxs.size(); 3646 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr; 3647 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) { 3648 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices(); 3649 unsigned NumInsertValueIdxs = InsertValueIdxs.size(); 3650 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs); 3651 if (InsertValueIdxs.slice(0, NumCommonIdxs) == 3652 Idxs.slice(0, NumCommonIdxs)) { 3653 if (NumIdxs == NumInsertValueIdxs) 3654 return IVI->getInsertedValueOperand(); 3655 break; 3656 } 3657 } 3658 3659 return nullptr; 3660 } 3661 3662 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, 3663 const DataLayout &DL, 3664 const TargetLibraryInfo *TLI, 3665 const DominatorTree *DT, 3666 AssumptionCache *AC, 3667 const Instruction *CxtI) { 3668 return ::SimplifyExtractValueInst(Agg, Idxs, Query(DL, TLI, DT, AC, CxtI), 3669 RecursionLimit); 3670 } 3671 3672 /// SimplifyExtractElementInst - Given operands for an ExtractElementInst, see if we 3673 /// can fold the result. If not, this returns null. 3674 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const Query &, 3675 unsigned) { 3676 if (auto *CVec = dyn_cast<Constant>(Vec)) { 3677 if (auto *CIdx = dyn_cast<Constant>(Idx)) 3678 return ConstantFoldExtractElementInstruction(CVec, CIdx); 3679 3680 // The index is not relevant if our vector is a splat. 3681 if (auto *Splat = CVec->getSplatValue()) 3682 return Splat; 3683 3684 if (isa<UndefValue>(Vec)) 3685 return UndefValue::get(Vec->getType()->getVectorElementType()); 3686 } 3687 3688 // If extracting a specified index from the vector, see if we can recursively 3689 // find a previously computed scalar that was inserted into the vector. 3690 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) 3691 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue())) 3692 return Elt; 3693 3694 return nullptr; 3695 } 3696 3697 Value *llvm::SimplifyExtractElementInst( 3698 Value *Vec, Value *Idx, const DataLayout &DL, const TargetLibraryInfo *TLI, 3699 const DominatorTree *DT, AssumptionCache *AC, const Instruction *CxtI) { 3700 return ::SimplifyExtractElementInst(Vec, Idx, Query(DL, TLI, DT, AC, CxtI), 3701 RecursionLimit); 3702 } 3703 3704 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null. 3705 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) { 3706 // If all of the PHI's incoming values are the same then replace the PHI node 3707 // with the common value. 3708 Value *CommonValue = nullptr; 3709 bool HasUndefInput = false; 3710 for (Value *Incoming : PN->incoming_values()) { 3711 // If the incoming value is the phi node itself, it can safely be skipped. 3712 if (Incoming == PN) continue; 3713 if (isa<UndefValue>(Incoming)) { 3714 // Remember that we saw an undef value, but otherwise ignore them. 3715 HasUndefInput = true; 3716 continue; 3717 } 3718 if (CommonValue && Incoming != CommonValue) 3719 return nullptr; // Not the same, bail out. 3720 CommonValue = Incoming; 3721 } 3722 3723 // If CommonValue is null then all of the incoming values were either undef or 3724 // equal to the phi node itself. 3725 if (!CommonValue) 3726 return UndefValue::get(PN->getType()); 3727 3728 // If we have a PHI node like phi(X, undef, X), where X is defined by some 3729 // instruction, we cannot return X as the result of the PHI node unless it 3730 // dominates the PHI block. 3731 if (HasUndefInput) 3732 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr; 3733 3734 return CommonValue; 3735 } 3736 3737 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) { 3738 if (Constant *C = dyn_cast<Constant>(Op)) 3739 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI); 3740 3741 return nullptr; 3742 } 3743 3744 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout &DL, 3745 const TargetLibraryInfo *TLI, 3746 const DominatorTree *DT, AssumptionCache *AC, 3747 const Instruction *CxtI) { 3748 return ::SimplifyTruncInst(Op, Ty, Query(DL, TLI, DT, AC, CxtI), 3749 RecursionLimit); 3750 } 3751 3752 //=== Helper functions for higher up the class hierarchy. 3753 3754 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can 3755 /// fold the result. If not, this returns null. 3756 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3757 const Query &Q, unsigned MaxRecurse) { 3758 switch (Opcode) { 3759 case Instruction::Add: 3760 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3761 Q, MaxRecurse); 3762 case Instruction::FAdd: 3763 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3764 3765 case Instruction::Sub: 3766 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3767 Q, MaxRecurse); 3768 case Instruction::FSub: 3769 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3770 3771 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse); 3772 case Instruction::FMul: 3773 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3774 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse); 3775 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse); 3776 case Instruction::FDiv: 3777 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3778 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse); 3779 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse); 3780 case Instruction::FRem: 3781 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3782 case Instruction::Shl: 3783 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3784 Q, MaxRecurse); 3785 case Instruction::LShr: 3786 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3787 case Instruction::AShr: 3788 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3789 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse); 3790 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse); 3791 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse); 3792 default: 3793 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 3794 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 3795 Constant *COps[] = {CLHS, CRHS}; 3796 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL, 3797 Q.TLI); 3798 } 3799 3800 // If the operation is associative, try some generic simplifications. 3801 if (Instruction::isAssociative(Opcode)) 3802 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse)) 3803 return V; 3804 3805 // If the operation is with the result of a select instruction check whether 3806 // operating on either branch of the select always yields the same value. 3807 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3808 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse)) 3809 return V; 3810 3811 // If the operation is with the result of a phi instruction, check whether 3812 // operating on all incoming values of the phi always yields the same value. 3813 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3814 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse)) 3815 return V; 3816 3817 return nullptr; 3818 } 3819 } 3820 3821 /// SimplifyFPBinOp - Given operands for a BinaryOperator, see if we can 3822 /// fold the result. If not, this returns null. 3823 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the 3824 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp. 3825 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3826 const FastMathFlags &FMF, const Query &Q, 3827 unsigned MaxRecurse) { 3828 switch (Opcode) { 3829 case Instruction::FAdd: 3830 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse); 3831 case Instruction::FSub: 3832 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse); 3833 case Instruction::FMul: 3834 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse); 3835 default: 3836 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse); 3837 } 3838 } 3839 3840 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3841 const DataLayout &DL, const TargetLibraryInfo *TLI, 3842 const DominatorTree *DT, AssumptionCache *AC, 3843 const Instruction *CxtI) { 3844 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3845 RecursionLimit); 3846 } 3847 3848 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3849 const FastMathFlags &FMF, const DataLayout &DL, 3850 const TargetLibraryInfo *TLI, 3851 const DominatorTree *DT, AssumptionCache *AC, 3852 const Instruction *CxtI) { 3853 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI), 3854 RecursionLimit); 3855 } 3856 3857 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can 3858 /// fold the result. 3859 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3860 const Query &Q, unsigned MaxRecurse) { 3861 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate)) 3862 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 3863 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3864 } 3865 3866 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3867 const DataLayout &DL, const TargetLibraryInfo *TLI, 3868 const DominatorTree *DT, AssumptionCache *AC, 3869 const Instruction *CxtI) { 3870 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3871 RecursionLimit); 3872 } 3873 3874 static bool IsIdempotent(Intrinsic::ID ID) { 3875 switch (ID) { 3876 default: return false; 3877 3878 // Unary idempotent: f(f(x)) = f(x) 3879 case Intrinsic::fabs: 3880 case Intrinsic::floor: 3881 case Intrinsic::ceil: 3882 case Intrinsic::trunc: 3883 case Intrinsic::rint: 3884 case Intrinsic::nearbyint: 3885 case Intrinsic::round: 3886 return true; 3887 } 3888 } 3889 3890 template <typename IterTy> 3891 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd, 3892 const Query &Q, unsigned MaxRecurse) { 3893 Intrinsic::ID IID = F->getIntrinsicID(); 3894 unsigned NumOperands = std::distance(ArgBegin, ArgEnd); 3895 Type *ReturnType = F->getReturnType(); 3896 3897 // Binary Ops 3898 if (NumOperands == 2) { 3899 Value *LHS = *ArgBegin; 3900 Value *RHS = *(ArgBegin + 1); 3901 if (IID == Intrinsic::usub_with_overflow || 3902 IID == Intrinsic::ssub_with_overflow) { 3903 // X - X -> { 0, false } 3904 if (LHS == RHS) 3905 return Constant::getNullValue(ReturnType); 3906 3907 // X - undef -> undef 3908 // undef - X -> undef 3909 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) 3910 return UndefValue::get(ReturnType); 3911 } 3912 3913 if (IID == Intrinsic::uadd_with_overflow || 3914 IID == Intrinsic::sadd_with_overflow) { 3915 // X + undef -> undef 3916 if (isa<UndefValue>(RHS)) 3917 return UndefValue::get(ReturnType); 3918 } 3919 3920 if (IID == Intrinsic::umul_with_overflow || 3921 IID == Intrinsic::smul_with_overflow) { 3922 // X * 0 -> { 0, false } 3923 if (match(RHS, m_Zero())) 3924 return Constant::getNullValue(ReturnType); 3925 3926 // X * undef -> { 0, false } 3927 if (match(RHS, m_Undef())) 3928 return Constant::getNullValue(ReturnType); 3929 } 3930 } 3931 3932 // Perform idempotent optimizations 3933 if (!IsIdempotent(IID)) 3934 return nullptr; 3935 3936 // Unary Ops 3937 if (NumOperands == 1) 3938 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) 3939 if (II->getIntrinsicID() == IID) 3940 return II; 3941 3942 return nullptr; 3943 } 3944 3945 template <typename IterTy> 3946 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd, 3947 const Query &Q, unsigned MaxRecurse) { 3948 Type *Ty = V->getType(); 3949 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 3950 Ty = PTy->getElementType(); 3951 FunctionType *FTy = cast<FunctionType>(Ty); 3952 3953 // call undef -> undef 3954 if (isa<UndefValue>(V)) 3955 return UndefValue::get(FTy->getReturnType()); 3956 3957 Function *F = dyn_cast<Function>(V); 3958 if (!F) 3959 return nullptr; 3960 3961 if (F->isIntrinsic()) 3962 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse)) 3963 return Ret; 3964 3965 if (!canConstantFoldCallTo(F)) 3966 return nullptr; 3967 3968 SmallVector<Constant *, 4> ConstantArgs; 3969 ConstantArgs.reserve(ArgEnd - ArgBegin); 3970 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) { 3971 Constant *C = dyn_cast<Constant>(*I); 3972 if (!C) 3973 return nullptr; 3974 ConstantArgs.push_back(C); 3975 } 3976 3977 return ConstantFoldCall(F, ConstantArgs, Q.TLI); 3978 } 3979 3980 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin, 3981 User::op_iterator ArgEnd, const DataLayout &DL, 3982 const TargetLibraryInfo *TLI, const DominatorTree *DT, 3983 AssumptionCache *AC, const Instruction *CxtI) { 3984 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI), 3985 RecursionLimit); 3986 } 3987 3988 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args, 3989 const DataLayout &DL, const TargetLibraryInfo *TLI, 3990 const DominatorTree *DT, AssumptionCache *AC, 3991 const Instruction *CxtI) { 3992 return ::SimplifyCall(V, Args.begin(), Args.end(), 3993 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3994 } 3995 3996 /// SimplifyInstruction - See if we can compute a simplified version of this 3997 /// instruction. If not, this returns null. 3998 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout &DL, 3999 const TargetLibraryInfo *TLI, 4000 const DominatorTree *DT, AssumptionCache *AC) { 4001 Value *Result; 4002 4003 switch (I->getOpcode()) { 4004 default: 4005 Result = ConstantFoldInstruction(I, DL, TLI); 4006 break; 4007 case Instruction::FAdd: 4008 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1), 4009 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4010 break; 4011 case Instruction::Add: 4012 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1), 4013 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4014 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 4015 TLI, DT, AC, I); 4016 break; 4017 case Instruction::FSub: 4018 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1), 4019 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4020 break; 4021 case Instruction::Sub: 4022 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1), 4023 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4024 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 4025 TLI, DT, AC, I); 4026 break; 4027 case Instruction::FMul: 4028 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1), 4029 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4030 break; 4031 case Instruction::Mul: 4032 Result = 4033 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4034 break; 4035 case Instruction::SDiv: 4036 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4037 AC, I); 4038 break; 4039 case Instruction::UDiv: 4040 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4041 AC, I); 4042 break; 4043 case Instruction::FDiv: 4044 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), 4045 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4046 break; 4047 case Instruction::SRem: 4048 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4049 AC, I); 4050 break; 4051 case Instruction::URem: 4052 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 4053 AC, I); 4054 break; 4055 case Instruction::FRem: 4056 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), 4057 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4058 break; 4059 case Instruction::Shl: 4060 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1), 4061 cast<BinaryOperator>(I)->hasNoSignedWrap(), 4062 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 4063 TLI, DT, AC, I); 4064 break; 4065 case Instruction::LShr: 4066 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1), 4067 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 4068 AC, I); 4069 break; 4070 case Instruction::AShr: 4071 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1), 4072 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 4073 AC, I); 4074 break; 4075 case Instruction::And: 4076 Result = 4077 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4078 break; 4079 case Instruction::Or: 4080 Result = 4081 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4082 break; 4083 case Instruction::Xor: 4084 Result = 4085 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 4086 break; 4087 case Instruction::ICmp: 4088 Result = 4089 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0), 4090 I->getOperand(1), DL, TLI, DT, AC, I); 4091 break; 4092 case Instruction::FCmp: 4093 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), 4094 I->getOperand(0), I->getOperand(1), 4095 I->getFastMathFlags(), DL, TLI, DT, AC, I); 4096 break; 4097 case Instruction::Select: 4098 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1), 4099 I->getOperand(2), DL, TLI, DT, AC, I); 4100 break; 4101 case Instruction::GetElementPtr: { 4102 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 4103 Result = SimplifyGEPInst(Ops, DL, TLI, DT, AC, I); 4104 break; 4105 } 4106 case Instruction::InsertValue: { 4107 InsertValueInst *IV = cast<InsertValueInst>(I); 4108 Result = SimplifyInsertValueInst(IV->getAggregateOperand(), 4109 IV->getInsertedValueOperand(), 4110 IV->getIndices(), DL, TLI, DT, AC, I); 4111 break; 4112 } 4113 case Instruction::ExtractValue: { 4114 auto *EVI = cast<ExtractValueInst>(I); 4115 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(), 4116 EVI->getIndices(), DL, TLI, DT, AC, I); 4117 break; 4118 } 4119 case Instruction::ExtractElement: { 4120 auto *EEI = cast<ExtractElementInst>(I); 4121 Result = SimplifyExtractElementInst( 4122 EEI->getVectorOperand(), EEI->getIndexOperand(), DL, TLI, DT, AC, I); 4123 break; 4124 } 4125 case Instruction::PHI: 4126 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I)); 4127 break; 4128 case Instruction::Call: { 4129 CallSite CS(cast<CallInst>(I)); 4130 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL, 4131 TLI, DT, AC, I); 4132 break; 4133 } 4134 case Instruction::Trunc: 4135 Result = 4136 SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT, AC, I); 4137 break; 4138 } 4139 4140 /// If called on unreachable code, the above logic may report that the 4141 /// instruction simplified to itself. Make life easier for users by 4142 /// detecting that case here, returning a safe value instead. 4143 return Result == I ? UndefValue::get(I->getType()) : Result; 4144 } 4145 4146 /// \brief Implementation of recursive simplification through an instructions 4147 /// uses. 4148 /// 4149 /// This is the common implementation of the recursive simplification routines. 4150 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to 4151 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of 4152 /// instructions to process and attempt to simplify it using 4153 /// InstructionSimplify. 4154 /// 4155 /// This routine returns 'true' only when *it* simplifies something. The passed 4156 /// in simplified value does not count toward this. 4157 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV, 4158 const TargetLibraryInfo *TLI, 4159 const DominatorTree *DT, 4160 AssumptionCache *AC) { 4161 bool Simplified = false; 4162 SmallSetVector<Instruction *, 8> Worklist; 4163 const DataLayout &DL = I->getModule()->getDataLayout(); 4164 4165 // If we have an explicit value to collapse to, do that round of the 4166 // simplification loop by hand initially. 4167 if (SimpleV) { 4168 for (User *U : I->users()) 4169 if (U != I) 4170 Worklist.insert(cast<Instruction>(U)); 4171 4172 // Replace the instruction with its simplified value. 4173 I->replaceAllUsesWith(SimpleV); 4174 4175 // Gracefully handle edge cases where the instruction is not wired into any 4176 // parent block. 4177 if (I->getParent()) 4178 I->eraseFromParent(); 4179 } else { 4180 Worklist.insert(I); 4181 } 4182 4183 // Note that we must test the size on each iteration, the worklist can grow. 4184 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { 4185 I = Worklist[Idx]; 4186 4187 // See if this instruction simplifies. 4188 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC); 4189 if (!SimpleV) 4190 continue; 4191 4192 Simplified = true; 4193 4194 // Stash away all the uses of the old instruction so we can check them for 4195 // recursive simplifications after a RAUW. This is cheaper than checking all 4196 // uses of To on the recursive step in most cases. 4197 for (User *U : I->users()) 4198 Worklist.insert(cast<Instruction>(U)); 4199 4200 // Replace the instruction with its simplified value. 4201 I->replaceAllUsesWith(SimpleV); 4202 4203 // Gracefully handle edge cases where the instruction is not wired into any 4204 // parent block. 4205 if (I->getParent()) 4206 I->eraseFromParent(); 4207 } 4208 return Simplified; 4209 } 4210 4211 bool llvm::recursivelySimplifyInstruction(Instruction *I, 4212 const TargetLibraryInfo *TLI, 4213 const DominatorTree *DT, 4214 AssumptionCache *AC) { 4215 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC); 4216 } 4217 4218 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV, 4219 const TargetLibraryInfo *TLI, 4220 const DominatorTree *DT, 4221 AssumptionCache *AC) { 4222 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!"); 4223 assert(SimpleV && "Must provide a simplified value."); 4224 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC); 4225 } 4226