1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements routines for folding instructions into simpler forms 11 // that do not require creating new instructions. This does constant folding 12 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either 13 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value 14 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been 15 // simplified: This is usually true and assuming it simplifies the logic (if 16 // they have not been simplified then results are correct but maybe suboptimal). 17 // 18 //===----------------------------------------------------------------------===// 19 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/ADT/SetVector.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/Analysis/AliasAnalysis.h" 24 #include "llvm/Analysis/ConstantFolding.h" 25 #include "llvm/Analysis/MemoryBuiltins.h" 26 #include "llvm/Analysis/ValueTracking.h" 27 #include "llvm/IR/ConstantRange.h" 28 #include "llvm/IR/DataLayout.h" 29 #include "llvm/IR/Dominators.h" 30 #include "llvm/IR/GetElementPtrTypeIterator.h" 31 #include "llvm/IR/GlobalAlias.h" 32 #include "llvm/IR/Operator.h" 33 #include "llvm/IR/PatternMatch.h" 34 #include "llvm/IR/ValueHandle.h" 35 #include <algorithm> 36 using namespace llvm; 37 using namespace llvm::PatternMatch; 38 39 #define DEBUG_TYPE "instsimplify" 40 41 enum { RecursionLimit = 3 }; 42 43 STATISTIC(NumExpand, "Number of expansions"); 44 STATISTIC(NumReassoc, "Number of reassociations"); 45 46 namespace { 47 struct Query { 48 const DataLayout &DL; 49 const TargetLibraryInfo *TLI; 50 const DominatorTree *DT; 51 AssumptionCache *AC; 52 const Instruction *CxtI; 53 54 Query(const DataLayout &DL, const TargetLibraryInfo *tli, 55 const DominatorTree *dt, AssumptionCache *ac = nullptr, 56 const Instruction *cxti = nullptr) 57 : DL(DL), TLI(tli), DT(dt), AC(ac), CxtI(cxti) {} 58 }; 59 } // end anonymous namespace 60 61 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned); 62 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &, 63 unsigned); 64 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &, 65 const Query &, unsigned); 66 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &, 67 unsigned); 68 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned); 69 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned); 70 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned); 71 72 /// getFalse - For a boolean type, or a vector of boolean type, return false, or 73 /// a vector with every element false, as appropriate for the type. 74 static Constant *getFalse(Type *Ty) { 75 assert(Ty->getScalarType()->isIntegerTy(1) && 76 "Expected i1 type or a vector of i1!"); 77 return Constant::getNullValue(Ty); 78 } 79 80 /// getTrue - For a boolean type, or a vector of boolean type, return true, or 81 /// a vector with every element true, as appropriate for the type. 82 static Constant *getTrue(Type *Ty) { 83 assert(Ty->getScalarType()->isIntegerTy(1) && 84 "Expected i1 type or a vector of i1!"); 85 return Constant::getAllOnesValue(Ty); 86 } 87 88 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"? 89 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS, 90 Value *RHS) { 91 CmpInst *Cmp = dyn_cast<CmpInst>(V); 92 if (!Cmp) 93 return false; 94 CmpInst::Predicate CPred = Cmp->getPredicate(); 95 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1); 96 if (CPred == Pred && CLHS == LHS && CRHS == RHS) 97 return true; 98 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS && 99 CRHS == LHS; 100 } 101 102 /// ValueDominatesPHI - Does the given value dominate the specified phi node? 103 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) { 104 Instruction *I = dyn_cast<Instruction>(V); 105 if (!I) 106 // Arguments and constants dominate all instructions. 107 return true; 108 109 // If we are processing instructions (and/or basic blocks) that have not been 110 // fully added to a function, the parent nodes may still be null. Simply 111 // return the conservative answer in these cases. 112 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent()) 113 return false; 114 115 // If we have a DominatorTree then do a precise test. 116 if (DT) { 117 if (!DT->isReachableFromEntry(P->getParent())) 118 return true; 119 if (!DT->isReachableFromEntry(I->getParent())) 120 return false; 121 return DT->dominates(I, P); 122 } 123 124 // Otherwise, if the instruction is in the entry block, and is not an invoke, 125 // then it obviously dominates all phi nodes. 126 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() && 127 !isa<InvokeInst>(I)) 128 return true; 129 130 return false; 131 } 132 133 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning 134 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is 135 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS. 136 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)". 137 /// Returns the simplified value, or null if no simplification was performed. 138 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS, 139 unsigned OpcToExpand, const Query &Q, 140 unsigned MaxRecurse) { 141 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand; 142 // Recursion is always used, so bail out at once if we already hit the limit. 143 if (!MaxRecurse--) 144 return nullptr; 145 146 // Check whether the expression has the form "(A op' B) op C". 147 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS)) 148 if (Op0->getOpcode() == OpcodeToExpand) { 149 // It does! Try turning it into "(A op C) op' (B op C)". 150 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS; 151 // Do "A op C" and "B op C" both simplify? 152 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) 153 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 154 // They do! Return "L op' R" if it simplifies or is already available. 155 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS. 156 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand) 157 && L == B && R == A)) { 158 ++NumExpand; 159 return LHS; 160 } 161 // Otherwise return "L op' R" if it simplifies. 162 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 163 ++NumExpand; 164 return V; 165 } 166 } 167 } 168 169 // Check whether the expression has the form "A op (B op' C)". 170 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS)) 171 if (Op1->getOpcode() == OpcodeToExpand) { 172 // It does! Try turning it into "(A op B) op' (A op C)". 173 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1); 174 // Do "A op B" and "A op C" both simplify? 175 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) 176 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) { 177 // They do! Return "L op' R" if it simplifies or is already available. 178 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS. 179 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand) 180 && L == C && R == B)) { 181 ++NumExpand; 182 return RHS; 183 } 184 // Otherwise return "L op' R" if it simplifies. 185 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) { 186 ++NumExpand; 187 return V; 188 } 189 } 190 } 191 192 return nullptr; 193 } 194 195 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary 196 /// operations. Returns the simpler value, or null if none was found. 197 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS, 198 const Query &Q, unsigned MaxRecurse) { 199 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc; 200 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!"); 201 202 // Recursion is always used, so bail out at once if we already hit the limit. 203 if (!MaxRecurse--) 204 return nullptr; 205 206 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS); 207 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS); 208 209 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely. 210 if (Op0 && Op0->getOpcode() == Opcode) { 211 Value *A = Op0->getOperand(0); 212 Value *B = Op0->getOperand(1); 213 Value *C = RHS; 214 215 // Does "B op C" simplify? 216 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) { 217 // It does! Return "A op V" if it simplifies or is already available. 218 // If V equals B then "A op V" is just the LHS. 219 if (V == B) return LHS; 220 // Otherwise return "A op V" if it simplifies. 221 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) { 222 ++NumReassoc; 223 return W; 224 } 225 } 226 } 227 228 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely. 229 if (Op1 && Op1->getOpcode() == Opcode) { 230 Value *A = LHS; 231 Value *B = Op1->getOperand(0); 232 Value *C = Op1->getOperand(1); 233 234 // Does "A op B" simplify? 235 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) { 236 // It does! Return "V op C" if it simplifies or is already available. 237 // If V equals B then "V op C" is just the RHS. 238 if (V == B) return RHS; 239 // Otherwise return "V op C" if it simplifies. 240 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) { 241 ++NumReassoc; 242 return W; 243 } 244 } 245 } 246 247 // The remaining transforms require commutativity as well as associativity. 248 if (!Instruction::isCommutative(Opcode)) 249 return nullptr; 250 251 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely. 252 if (Op0 && Op0->getOpcode() == Opcode) { 253 Value *A = Op0->getOperand(0); 254 Value *B = Op0->getOperand(1); 255 Value *C = RHS; 256 257 // Does "C op A" simplify? 258 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 259 // It does! Return "V op B" if it simplifies or is already available. 260 // If V equals A then "V op B" is just the LHS. 261 if (V == A) return LHS; 262 // Otherwise return "V op B" if it simplifies. 263 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) { 264 ++NumReassoc; 265 return W; 266 } 267 } 268 } 269 270 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely. 271 if (Op1 && Op1->getOpcode() == Opcode) { 272 Value *A = LHS; 273 Value *B = Op1->getOperand(0); 274 Value *C = Op1->getOperand(1); 275 276 // Does "C op A" simplify? 277 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) { 278 // It does! Return "B op V" if it simplifies or is already available. 279 // If V equals C then "B op V" is just the RHS. 280 if (V == C) return RHS; 281 // Otherwise return "B op V" if it simplifies. 282 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) { 283 ++NumReassoc; 284 return W; 285 } 286 } 287 } 288 289 return nullptr; 290 } 291 292 /// ThreadBinOpOverSelect - In the case of a binary operation with a select 293 /// instruction as an operand, try to simplify the binop by seeing whether 294 /// evaluating it on both branches of the select results in the same value. 295 /// Returns the common value if so, otherwise returns null. 296 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS, 297 const Query &Q, unsigned MaxRecurse) { 298 // Recursion is always used, so bail out at once if we already hit the limit. 299 if (!MaxRecurse--) 300 return nullptr; 301 302 SelectInst *SI; 303 if (isa<SelectInst>(LHS)) { 304 SI = cast<SelectInst>(LHS); 305 } else { 306 assert(isa<SelectInst>(RHS) && "No select instruction operand!"); 307 SI = cast<SelectInst>(RHS); 308 } 309 310 // Evaluate the BinOp on the true and false branches of the select. 311 Value *TV; 312 Value *FV; 313 if (SI == LHS) { 314 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse); 315 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse); 316 } else { 317 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse); 318 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse); 319 } 320 321 // If they simplified to the same value, then return the common value. 322 // If they both failed to simplify then return null. 323 if (TV == FV) 324 return TV; 325 326 // If one branch simplified to undef, return the other one. 327 if (TV && isa<UndefValue>(TV)) 328 return FV; 329 if (FV && isa<UndefValue>(FV)) 330 return TV; 331 332 // If applying the operation did not change the true and false select values, 333 // then the result of the binop is the select itself. 334 if (TV == SI->getTrueValue() && FV == SI->getFalseValue()) 335 return SI; 336 337 // If one branch simplified and the other did not, and the simplified 338 // value is equal to the unsimplified one, return the simplified value. 339 // For example, select (cond, X, X & Z) & Z -> X & Z. 340 if ((FV && !TV) || (TV && !FV)) { 341 // Check that the simplified value has the form "X op Y" where "op" is the 342 // same as the original operation. 343 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV); 344 if (Simplified && Simplified->getOpcode() == Opcode) { 345 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS". 346 // We already know that "op" is the same as for the simplified value. See 347 // if the operands match too. If so, return the simplified value. 348 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue(); 349 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS; 350 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch; 351 if (Simplified->getOperand(0) == UnsimplifiedLHS && 352 Simplified->getOperand(1) == UnsimplifiedRHS) 353 return Simplified; 354 if (Simplified->isCommutative() && 355 Simplified->getOperand(1) == UnsimplifiedLHS && 356 Simplified->getOperand(0) == UnsimplifiedRHS) 357 return Simplified; 358 } 359 } 360 361 return nullptr; 362 } 363 364 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction, 365 /// try to simplify the comparison by seeing whether both branches of the select 366 /// result in the same value. Returns the common value if so, otherwise returns 367 /// null. 368 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS, 369 Value *RHS, const Query &Q, 370 unsigned MaxRecurse) { 371 // Recursion is always used, so bail out at once if we already hit the limit. 372 if (!MaxRecurse--) 373 return nullptr; 374 375 // Make sure the select is on the LHS. 376 if (!isa<SelectInst>(LHS)) { 377 std::swap(LHS, RHS); 378 Pred = CmpInst::getSwappedPredicate(Pred); 379 } 380 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!"); 381 SelectInst *SI = cast<SelectInst>(LHS); 382 Value *Cond = SI->getCondition(); 383 Value *TV = SI->getTrueValue(); 384 Value *FV = SI->getFalseValue(); 385 386 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it. 387 // Does "cmp TV, RHS" simplify? 388 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse); 389 if (TCmp == Cond) { 390 // It not only simplified, it simplified to the select condition. Replace 391 // it with 'true'. 392 TCmp = getTrue(Cond->getType()); 393 } else if (!TCmp) { 394 // It didn't simplify. However if "cmp TV, RHS" is equal to the select 395 // condition then we can replace it with 'true'. Otherwise give up. 396 if (!isSameCompare(Cond, Pred, TV, RHS)) 397 return nullptr; 398 TCmp = getTrue(Cond->getType()); 399 } 400 401 // Does "cmp FV, RHS" simplify? 402 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse); 403 if (FCmp == Cond) { 404 // It not only simplified, it simplified to the select condition. Replace 405 // it with 'false'. 406 FCmp = getFalse(Cond->getType()); 407 } else if (!FCmp) { 408 // It didn't simplify. However if "cmp FV, RHS" is equal to the select 409 // condition then we can replace it with 'false'. Otherwise give up. 410 if (!isSameCompare(Cond, Pred, FV, RHS)) 411 return nullptr; 412 FCmp = getFalse(Cond->getType()); 413 } 414 415 // If both sides simplified to the same value, then use it as the result of 416 // the original comparison. 417 if (TCmp == FCmp) 418 return TCmp; 419 420 // The remaining cases only make sense if the select condition has the same 421 // type as the result of the comparison, so bail out if this is not so. 422 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy()) 423 return nullptr; 424 // If the false value simplified to false, then the result of the compare 425 // is equal to "Cond && TCmp". This also catches the case when the false 426 // value simplified to false and the true value to true, returning "Cond". 427 if (match(FCmp, m_Zero())) 428 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse)) 429 return V; 430 // If the true value simplified to true, then the result of the compare 431 // is equal to "Cond || FCmp". 432 if (match(TCmp, m_One())) 433 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse)) 434 return V; 435 // Finally, if the false value simplified to true and the true value to 436 // false, then the result of the compare is equal to "!Cond". 437 if (match(FCmp, m_One()) && match(TCmp, m_Zero())) 438 if (Value *V = 439 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()), 440 Q, MaxRecurse)) 441 return V; 442 443 return nullptr; 444 } 445 446 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that 447 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating 448 /// it on the incoming phi values yields the same result for every value. If so 449 /// returns the common value, otherwise returns null. 450 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS, 451 const Query &Q, unsigned MaxRecurse) { 452 // Recursion is always used, so bail out at once if we already hit the limit. 453 if (!MaxRecurse--) 454 return nullptr; 455 456 PHINode *PI; 457 if (isa<PHINode>(LHS)) { 458 PI = cast<PHINode>(LHS); 459 // Bail out if RHS and the phi may be mutually interdependent due to a loop. 460 if (!ValueDominatesPHI(RHS, PI, Q.DT)) 461 return nullptr; 462 } else { 463 assert(isa<PHINode>(RHS) && "No PHI instruction operand!"); 464 PI = cast<PHINode>(RHS); 465 // Bail out if LHS and the phi may be mutually interdependent due to a loop. 466 if (!ValueDominatesPHI(LHS, PI, Q.DT)) 467 return nullptr; 468 } 469 470 // Evaluate the BinOp on the incoming phi values. 471 Value *CommonValue = nullptr; 472 for (Value *Incoming : PI->incoming_values()) { 473 // If the incoming value is the phi node itself, it can safely be skipped. 474 if (Incoming == PI) continue; 475 Value *V = PI == LHS ? 476 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) : 477 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse); 478 // If the operation failed to simplify, or simplified to a different value 479 // to previously, then give up. 480 if (!V || (CommonValue && V != CommonValue)) 481 return nullptr; 482 CommonValue = V; 483 } 484 485 return CommonValue; 486 } 487 488 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try 489 /// try to simplify the comparison by seeing whether comparing with all of the 490 /// incoming phi values yields the same result every time. If so returns the 491 /// common result, otherwise returns null. 492 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS, 493 const Query &Q, unsigned MaxRecurse) { 494 // Recursion is always used, so bail out at once if we already hit the limit. 495 if (!MaxRecurse--) 496 return nullptr; 497 498 // Make sure the phi is on the LHS. 499 if (!isa<PHINode>(LHS)) { 500 std::swap(LHS, RHS); 501 Pred = CmpInst::getSwappedPredicate(Pred); 502 } 503 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!"); 504 PHINode *PI = cast<PHINode>(LHS); 505 506 // Bail out if RHS and the phi may be mutually interdependent due to a loop. 507 if (!ValueDominatesPHI(RHS, PI, Q.DT)) 508 return nullptr; 509 510 // Evaluate the BinOp on the incoming phi values. 511 Value *CommonValue = nullptr; 512 for (Value *Incoming : PI->incoming_values()) { 513 // If the incoming value is the phi node itself, it can safely be skipped. 514 if (Incoming == PI) continue; 515 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse); 516 // If the operation failed to simplify, or simplified to a different value 517 // to previously, then give up. 518 if (!V || (CommonValue && V != CommonValue)) 519 return nullptr; 520 CommonValue = V; 521 } 522 523 return CommonValue; 524 } 525 526 /// SimplifyAddInst - Given operands for an Add, see if we can 527 /// fold the result. If not, this returns null. 528 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 529 const Query &Q, unsigned MaxRecurse) { 530 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 531 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 532 Constant *Ops[] = { CLHS, CRHS }; 533 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops, 534 Q.DL, Q.TLI); 535 } 536 537 // Canonicalize the constant to the RHS. 538 std::swap(Op0, Op1); 539 } 540 541 // X + undef -> undef 542 if (match(Op1, m_Undef())) 543 return Op1; 544 545 // X + 0 -> X 546 if (match(Op1, m_Zero())) 547 return Op0; 548 549 // X + (Y - X) -> Y 550 // (Y - X) + X -> Y 551 // Eg: X + -X -> 0 552 Value *Y = nullptr; 553 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) || 554 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1)))) 555 return Y; 556 557 // X + ~X -> -1 since ~X = -X-1 558 if (match(Op0, m_Not(m_Specific(Op1))) || 559 match(Op1, m_Not(m_Specific(Op0)))) 560 return Constant::getAllOnesValue(Op0->getType()); 561 562 /// i1 add -> xor. 563 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 564 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 565 return V; 566 567 // Try some generic simplifications for associative operations. 568 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q, 569 MaxRecurse)) 570 return V; 571 572 // Threading Add over selects and phi nodes is pointless, so don't bother. 573 // Threading over the select in "A + select(cond, B, C)" means evaluating 574 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and 575 // only if B and C are equal. If B and C are equal then (since we assume 576 // that operands have already been simplified) "select(cond, B, C)" should 577 // have been simplified to the common value of B and C already. Analysing 578 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly 579 // for threading over phi nodes. 580 581 return nullptr; 582 } 583 584 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 585 const DataLayout &DL, const TargetLibraryInfo *TLI, 586 const DominatorTree *DT, AssumptionCache *AC, 587 const Instruction *CxtI) { 588 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 589 RecursionLimit); 590 } 591 592 /// \brief Compute the base pointer and cumulative constant offsets for V. 593 /// 594 /// This strips all constant offsets off of V, leaving it the base pointer, and 595 /// accumulates the total constant offset applied in the returned constant. It 596 /// returns 0 if V is not a pointer, and returns the constant '0' if there are 597 /// no constant offsets applied. 598 /// 599 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't 600 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc. 601 /// folding. 602 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V, 603 bool AllowNonInbounds = false) { 604 assert(V->getType()->getScalarType()->isPointerTy()); 605 606 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType(); 607 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth()); 608 609 // Even though we don't look through PHI nodes, we could be called on an 610 // instruction in an unreachable block, which may be on a cycle. 611 SmallPtrSet<Value *, 4> Visited; 612 Visited.insert(V); 613 do { 614 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 615 if ((!AllowNonInbounds && !GEP->isInBounds()) || 616 !GEP->accumulateConstantOffset(DL, Offset)) 617 break; 618 V = GEP->getPointerOperand(); 619 } else if (Operator::getOpcode(V) == Instruction::BitCast) { 620 V = cast<Operator>(V)->getOperand(0); 621 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 622 if (GA->mayBeOverridden()) 623 break; 624 V = GA->getAliasee(); 625 } else { 626 break; 627 } 628 assert(V->getType()->getScalarType()->isPointerTy() && 629 "Unexpected operand type!"); 630 } while (Visited.insert(V).second); 631 632 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset); 633 if (V->getType()->isVectorTy()) 634 return ConstantVector::getSplat(V->getType()->getVectorNumElements(), 635 OffsetIntPtr); 636 return OffsetIntPtr; 637 } 638 639 /// \brief Compute the constant difference between two pointer values. 640 /// If the difference is not a constant, returns zero. 641 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS, 642 Value *RHS) { 643 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 644 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 645 646 // If LHS and RHS are not related via constant offsets to the same base 647 // value, there is nothing we can do here. 648 if (LHS != RHS) 649 return nullptr; 650 651 // Otherwise, the difference of LHS - RHS can be computed as: 652 // LHS - RHS 653 // = (LHSOffset + Base) - (RHSOffset + Base) 654 // = LHSOffset - RHSOffset 655 return ConstantExpr::getSub(LHSOffset, RHSOffset); 656 } 657 658 /// SimplifySubInst - Given operands for a Sub, see if we can 659 /// fold the result. If not, this returns null. 660 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 661 const Query &Q, unsigned MaxRecurse) { 662 if (Constant *CLHS = dyn_cast<Constant>(Op0)) 663 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 664 Constant *Ops[] = { CLHS, CRHS }; 665 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(), 666 Ops, Q.DL, Q.TLI); 667 } 668 669 // X - undef -> undef 670 // undef - X -> undef 671 if (match(Op0, m_Undef()) || match(Op1, m_Undef())) 672 return UndefValue::get(Op0->getType()); 673 674 // X - 0 -> X 675 if (match(Op1, m_Zero())) 676 return Op0; 677 678 // X - X -> 0 679 if (Op0 == Op1) 680 return Constant::getNullValue(Op0->getType()); 681 682 // 0 - X -> 0 if the sub is NUW. 683 if (isNUW && match(Op0, m_Zero())) 684 return Op0; 685 686 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies. 687 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X 688 Value *X = nullptr, *Y = nullptr, *Z = Op1; 689 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z 690 // See if "V === Y - Z" simplifies. 691 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1)) 692 // It does! Now see if "X + V" simplifies. 693 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) { 694 // It does, we successfully reassociated! 695 ++NumReassoc; 696 return W; 697 } 698 // See if "V === X - Z" simplifies. 699 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 700 // It does! Now see if "Y + V" simplifies. 701 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) { 702 // It does, we successfully reassociated! 703 ++NumReassoc; 704 return W; 705 } 706 } 707 708 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies. 709 // For example, X - (X + 1) -> -1 710 X = Op0; 711 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z) 712 // See if "V === X - Y" simplifies. 713 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 714 // It does! Now see if "V - Z" simplifies. 715 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) { 716 // It does, we successfully reassociated! 717 ++NumReassoc; 718 return W; 719 } 720 // See if "V === X - Z" simplifies. 721 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 722 // It does! Now see if "V - Y" simplifies. 723 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) { 724 // It does, we successfully reassociated! 725 ++NumReassoc; 726 return W; 727 } 728 } 729 730 // Z - (X - Y) -> (Z - X) + Y if everything simplifies. 731 // For example, X - (X - Y) -> Y. 732 Z = Op0; 733 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y) 734 // See if "V === Z - X" simplifies. 735 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1)) 736 // It does! Now see if "V + Y" simplifies. 737 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) { 738 // It does, we successfully reassociated! 739 ++NumReassoc; 740 return W; 741 } 742 743 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies. 744 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) && 745 match(Op1, m_Trunc(m_Value(Y)))) 746 if (X->getType() == Y->getType()) 747 // See if "V === X - Y" simplifies. 748 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 749 // It does! Now see if "trunc V" simplifies. 750 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1)) 751 // It does, return the simplified "trunc V". 752 return W; 753 754 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...). 755 if (match(Op0, m_PtrToInt(m_Value(X))) && 756 match(Op1, m_PtrToInt(m_Value(Y)))) 757 if (Constant *Result = computePointerDifference(Q.DL, X, Y)) 758 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true); 759 760 // i1 sub -> xor. 761 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 762 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 763 return V; 764 765 // Threading Sub over selects and phi nodes is pointless, so don't bother. 766 // Threading over the select in "A - select(cond, B, C)" means evaluating 767 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and 768 // only if B and C are equal. If B and C are equal then (since we assume 769 // that operands have already been simplified) "select(cond, B, C)" should 770 // have been simplified to the common value of B and C already. Analysing 771 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly 772 // for threading over phi nodes. 773 774 return nullptr; 775 } 776 777 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 778 const DataLayout &DL, const TargetLibraryInfo *TLI, 779 const DominatorTree *DT, AssumptionCache *AC, 780 const Instruction *CxtI) { 781 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 782 RecursionLimit); 783 } 784 785 /// Given operands for an FAdd, see if we can fold the result. If not, this 786 /// returns null. 787 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 788 const Query &Q, unsigned MaxRecurse) { 789 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 790 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 791 Constant *Ops[] = { CLHS, CRHS }; 792 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(), 793 Ops, Q.DL, Q.TLI); 794 } 795 796 // Canonicalize the constant to the RHS. 797 std::swap(Op0, Op1); 798 } 799 800 // fadd X, -0 ==> X 801 if (match(Op1, m_NegZero())) 802 return Op0; 803 804 // fadd X, 0 ==> X, when we know X is not -0 805 if (match(Op1, m_Zero()) && 806 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0))) 807 return Op0; 808 809 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0 810 // where nnan and ninf have to occur at least once somewhere in this 811 // expression 812 Value *SubOp = nullptr; 813 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0)))) 814 SubOp = Op1; 815 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1)))) 816 SubOp = Op0; 817 if (SubOp) { 818 Instruction *FSub = cast<Instruction>(SubOp); 819 if ((FMF.noNaNs() || FSub->hasNoNaNs()) && 820 (FMF.noInfs() || FSub->hasNoInfs())) 821 return Constant::getNullValue(Op0->getType()); 822 } 823 824 return nullptr; 825 } 826 827 /// Given operands for an FSub, see if we can fold the result. If not, this 828 /// returns null. 829 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 830 const Query &Q, unsigned MaxRecurse) { 831 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 832 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 833 Constant *Ops[] = { CLHS, CRHS }; 834 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(), 835 Ops, Q.DL, Q.TLI); 836 } 837 } 838 839 // fsub X, 0 ==> X 840 if (match(Op1, m_Zero())) 841 return Op0; 842 843 // fsub X, -0 ==> X, when we know X is not -0 844 if (match(Op1, m_NegZero()) && 845 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0))) 846 return Op0; 847 848 // fsub 0, (fsub -0.0, X) ==> X 849 Value *X; 850 if (match(Op0, m_AnyZero())) { 851 if (match(Op1, m_FSub(m_NegZero(), m_Value(X)))) 852 return X; 853 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X)))) 854 return X; 855 } 856 857 // fsub nnan ninf x, x ==> 0.0 858 if (FMF.noNaNs() && FMF.noInfs() && Op0 == Op1) 859 return Constant::getNullValue(Op0->getType()); 860 861 return nullptr; 862 } 863 864 /// Given the operands for an FMul, see if we can fold the result 865 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, 866 FastMathFlags FMF, 867 const Query &Q, 868 unsigned MaxRecurse) { 869 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 870 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 871 Constant *Ops[] = { CLHS, CRHS }; 872 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(), 873 Ops, Q.DL, Q.TLI); 874 } 875 876 // Canonicalize the constant to the RHS. 877 std::swap(Op0, Op1); 878 } 879 880 // fmul X, 1.0 ==> X 881 if (match(Op1, m_FPOne())) 882 return Op0; 883 884 // fmul nnan nsz X, 0 ==> 0 885 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero())) 886 return Op1; 887 888 return nullptr; 889 } 890 891 /// SimplifyMulInst - Given operands for a Mul, see if we can 892 /// fold the result. If not, this returns null. 893 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q, 894 unsigned MaxRecurse) { 895 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 896 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 897 Constant *Ops[] = { CLHS, CRHS }; 898 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(), 899 Ops, Q.DL, Q.TLI); 900 } 901 902 // Canonicalize the constant to the RHS. 903 std::swap(Op0, Op1); 904 } 905 906 // X * undef -> 0 907 if (match(Op1, m_Undef())) 908 return Constant::getNullValue(Op0->getType()); 909 910 // X * 0 -> 0 911 if (match(Op1, m_Zero())) 912 return Op1; 913 914 // X * 1 -> X 915 if (match(Op1, m_One())) 916 return Op0; 917 918 // (X / Y) * Y -> X if the division is exact. 919 Value *X = nullptr; 920 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y 921 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y) 922 return X; 923 924 // i1 mul -> and. 925 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 926 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1)) 927 return V; 928 929 // Try some generic simplifications for associative operations. 930 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q, 931 MaxRecurse)) 932 return V; 933 934 // Mul distributes over Add. Try some generic simplifications based on this. 935 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add, 936 Q, MaxRecurse)) 937 return V; 938 939 // If the operation is with the result of a select instruction, check whether 940 // operating on either branch of the select always yields the same value. 941 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 942 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q, 943 MaxRecurse)) 944 return V; 945 946 // If the operation is with the result of a phi instruction, check whether 947 // operating on all incoming values of the phi always yields the same value. 948 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 949 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q, 950 MaxRecurse)) 951 return V; 952 953 return nullptr; 954 } 955 956 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 957 const DataLayout &DL, 958 const TargetLibraryInfo *TLI, 959 const DominatorTree *DT, AssumptionCache *AC, 960 const Instruction *CxtI) { 961 return ::SimplifyFAddInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 962 RecursionLimit); 963 } 964 965 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 966 const DataLayout &DL, 967 const TargetLibraryInfo *TLI, 968 const DominatorTree *DT, AssumptionCache *AC, 969 const Instruction *CxtI) { 970 return ::SimplifyFSubInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 971 RecursionLimit); 972 } 973 974 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF, 975 const DataLayout &DL, 976 const TargetLibraryInfo *TLI, 977 const DominatorTree *DT, AssumptionCache *AC, 978 const Instruction *CxtI) { 979 return ::SimplifyFMulInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 980 RecursionLimit); 981 } 982 983 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout &DL, 984 const TargetLibraryInfo *TLI, 985 const DominatorTree *DT, AssumptionCache *AC, 986 const Instruction *CxtI) { 987 return ::SimplifyMulInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 988 RecursionLimit); 989 } 990 991 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can 992 /// fold the result. If not, this returns null. 993 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 994 const Query &Q, unsigned MaxRecurse) { 995 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 996 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 997 Constant *Ops[] = { C0, C1 }; 998 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 999 } 1000 } 1001 1002 bool isSigned = Opcode == Instruction::SDiv; 1003 1004 // X / undef -> undef 1005 if (match(Op1, m_Undef())) 1006 return Op1; 1007 1008 // X / 0 -> undef, we don't need to preserve faults! 1009 if (match(Op1, m_Zero())) 1010 return UndefValue::get(Op1->getType()); 1011 1012 // undef / X -> 0 1013 if (match(Op0, m_Undef())) 1014 return Constant::getNullValue(Op0->getType()); 1015 1016 // 0 / X -> 0, we don't need to preserve faults! 1017 if (match(Op0, m_Zero())) 1018 return Op0; 1019 1020 // X / 1 -> X 1021 if (match(Op1, m_One())) 1022 return Op0; 1023 1024 if (Op0->getType()->isIntegerTy(1)) 1025 // It can't be division by zero, hence it must be division by one. 1026 return Op0; 1027 1028 // X / X -> 1 1029 if (Op0 == Op1) 1030 return ConstantInt::get(Op0->getType(), 1); 1031 1032 // (X * Y) / Y -> X if the multiplication does not overflow. 1033 Value *X = nullptr, *Y = nullptr; 1034 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) { 1035 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1 1036 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0); 1037 // If the Mul knows it does not overflow, then we are good to go. 1038 if ((isSigned && Mul->hasNoSignedWrap()) || 1039 (!isSigned && Mul->hasNoUnsignedWrap())) 1040 return X; 1041 // If X has the form X = A / Y then X * Y cannot overflow. 1042 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X)) 1043 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y) 1044 return X; 1045 } 1046 1047 // (X rem Y) / Y -> 0 1048 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1049 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1050 return Constant::getNullValue(Op0->getType()); 1051 1052 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow 1053 ConstantInt *C1, *C2; 1054 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) && 1055 match(Op1, m_ConstantInt(C2))) { 1056 bool Overflow; 1057 C1->getValue().umul_ov(C2->getValue(), Overflow); 1058 if (Overflow) 1059 return Constant::getNullValue(Op0->getType()); 1060 } 1061 1062 // If the operation is with the result of a select instruction, check whether 1063 // operating on either branch of the select always yields the same value. 1064 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1065 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1066 return V; 1067 1068 // If the operation is with the result of a phi instruction, check whether 1069 // operating on all incoming values of the phi always yields the same value. 1070 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1071 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1072 return V; 1073 1074 return nullptr; 1075 } 1076 1077 /// SimplifySDivInst - Given operands for an SDiv, see if we can 1078 /// fold the result. If not, this returns null. 1079 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q, 1080 unsigned MaxRecurse) { 1081 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse)) 1082 return V; 1083 1084 return nullptr; 1085 } 1086 1087 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout &DL, 1088 const TargetLibraryInfo *TLI, 1089 const DominatorTree *DT, AssumptionCache *AC, 1090 const Instruction *CxtI) { 1091 return ::SimplifySDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1092 RecursionLimit); 1093 } 1094 1095 /// SimplifyUDivInst - Given operands for a UDiv, see if we can 1096 /// fold the result. If not, this returns null. 1097 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q, 1098 unsigned MaxRecurse) { 1099 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse)) 1100 return V; 1101 1102 return nullptr; 1103 } 1104 1105 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout &DL, 1106 const TargetLibraryInfo *TLI, 1107 const DominatorTree *DT, AssumptionCache *AC, 1108 const Instruction *CxtI) { 1109 return ::SimplifyUDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1110 RecursionLimit); 1111 } 1112 1113 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1114 const Query &Q, unsigned) { 1115 // undef / X -> undef (the undef could be a snan). 1116 if (match(Op0, m_Undef())) 1117 return Op0; 1118 1119 // X / undef -> undef 1120 if (match(Op1, m_Undef())) 1121 return Op1; 1122 1123 // 0 / X -> 0 1124 // Requires that NaNs are off (X could be zero) and signed zeroes are 1125 // ignored (X could be positive or negative, so the output sign is unknown). 1126 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1127 return Op0; 1128 1129 return nullptr; 1130 } 1131 1132 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1133 const DataLayout &DL, 1134 const TargetLibraryInfo *TLI, 1135 const DominatorTree *DT, AssumptionCache *AC, 1136 const Instruction *CxtI) { 1137 return ::SimplifyFDivInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 1138 RecursionLimit); 1139 } 1140 1141 /// SimplifyRem - Given operands for an SRem or URem, see if we can 1142 /// fold the result. If not, this returns null. 1143 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1144 const Query &Q, unsigned MaxRecurse) { 1145 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1146 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1147 Constant *Ops[] = { C0, C1 }; 1148 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1149 } 1150 } 1151 1152 // X % undef -> undef 1153 if (match(Op1, m_Undef())) 1154 return Op1; 1155 1156 // undef % X -> 0 1157 if (match(Op0, m_Undef())) 1158 return Constant::getNullValue(Op0->getType()); 1159 1160 // 0 % X -> 0, we don't need to preserve faults! 1161 if (match(Op0, m_Zero())) 1162 return Op0; 1163 1164 // X % 0 -> undef, we don't need to preserve faults! 1165 if (match(Op1, m_Zero())) 1166 return UndefValue::get(Op0->getType()); 1167 1168 // X % 1 -> 0 1169 if (match(Op1, m_One())) 1170 return Constant::getNullValue(Op0->getType()); 1171 1172 if (Op0->getType()->isIntegerTy(1)) 1173 // It can't be remainder by zero, hence it must be remainder by one. 1174 return Constant::getNullValue(Op0->getType()); 1175 1176 // X % X -> 0 1177 if (Op0 == Op1) 1178 return Constant::getNullValue(Op0->getType()); 1179 1180 // (X % Y) % Y -> X % Y 1181 if ((Opcode == Instruction::SRem && 1182 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1183 (Opcode == Instruction::URem && 1184 match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1185 return Op0; 1186 1187 // If the operation is with the result of a select instruction, check whether 1188 // operating on either branch of the select always yields the same value. 1189 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1190 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1191 return V; 1192 1193 // If the operation is with the result of a phi instruction, check whether 1194 // operating on all incoming values of the phi always yields the same value. 1195 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1196 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1197 return V; 1198 1199 return nullptr; 1200 } 1201 1202 /// SimplifySRemInst - Given operands for an SRem, see if we can 1203 /// fold the result. If not, this returns null. 1204 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q, 1205 unsigned MaxRecurse) { 1206 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse)) 1207 return V; 1208 1209 return nullptr; 1210 } 1211 1212 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout &DL, 1213 const TargetLibraryInfo *TLI, 1214 const DominatorTree *DT, AssumptionCache *AC, 1215 const Instruction *CxtI) { 1216 return ::SimplifySRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1217 RecursionLimit); 1218 } 1219 1220 /// SimplifyURemInst - Given operands for a URem, see if we can 1221 /// fold the result. If not, this returns null. 1222 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q, 1223 unsigned MaxRecurse) { 1224 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse)) 1225 return V; 1226 1227 return nullptr; 1228 } 1229 1230 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout &DL, 1231 const TargetLibraryInfo *TLI, 1232 const DominatorTree *DT, AssumptionCache *AC, 1233 const Instruction *CxtI) { 1234 return ::SimplifyURemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1235 RecursionLimit); 1236 } 1237 1238 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1239 const Query &, unsigned) { 1240 // undef % X -> undef (the undef could be a snan). 1241 if (match(Op0, m_Undef())) 1242 return Op0; 1243 1244 // X % undef -> undef 1245 if (match(Op1, m_Undef())) 1246 return Op1; 1247 1248 // 0 % X -> 0 1249 // Requires that NaNs are off (X could be zero) and signed zeroes are 1250 // ignored (X could be positive or negative, so the output sign is unknown). 1251 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero())) 1252 return Op0; 1253 1254 return nullptr; 1255 } 1256 1257 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1258 const DataLayout &DL, 1259 const TargetLibraryInfo *TLI, 1260 const DominatorTree *DT, AssumptionCache *AC, 1261 const Instruction *CxtI) { 1262 return ::SimplifyFRemInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI), 1263 RecursionLimit); 1264 } 1265 1266 /// isUndefShift - Returns true if a shift by \c Amount always yields undef. 1267 static bool isUndefShift(Value *Amount) { 1268 Constant *C = dyn_cast<Constant>(Amount); 1269 if (!C) 1270 return false; 1271 1272 // X shift by undef -> undef because it may shift by the bitwidth. 1273 if (isa<UndefValue>(C)) 1274 return true; 1275 1276 // Shifting by the bitwidth or more is undefined. 1277 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) 1278 if (CI->getValue().getLimitedValue() >= 1279 CI->getType()->getScalarSizeInBits()) 1280 return true; 1281 1282 // If all lanes of a vector shift are undefined the whole shift is. 1283 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) { 1284 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I) 1285 if (!isUndefShift(C->getAggregateElement(I))) 1286 return false; 1287 return true; 1288 } 1289 1290 return false; 1291 } 1292 1293 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can 1294 /// fold the result. If not, this returns null. 1295 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1, 1296 const Query &Q, unsigned MaxRecurse) { 1297 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1298 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1299 Constant *Ops[] = { C0, C1 }; 1300 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI); 1301 } 1302 } 1303 1304 // 0 shift by X -> 0 1305 if (match(Op0, m_Zero())) 1306 return Op0; 1307 1308 // X shift by 0 -> X 1309 if (match(Op1, m_Zero())) 1310 return Op0; 1311 1312 // Fold undefined shifts. 1313 if (isUndefShift(Op1)) 1314 return UndefValue::get(Op0->getType()); 1315 1316 // If the operation is with the result of a select instruction, check whether 1317 // operating on either branch of the select always yields the same value. 1318 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1319 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1320 return V; 1321 1322 // If the operation is with the result of a phi instruction, check whether 1323 // operating on all incoming values of the phi always yields the same value. 1324 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1325 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1326 return V; 1327 1328 return nullptr; 1329 } 1330 1331 /// \brief Given operands for an Shl, LShr or AShr, see if we can 1332 /// fold the result. If not, this returns null. 1333 static Value *SimplifyRightShift(unsigned Opcode, Value *Op0, Value *Op1, 1334 bool isExact, const Query &Q, 1335 unsigned MaxRecurse) { 1336 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse)) 1337 return V; 1338 1339 // X >> X -> 0 1340 if (Op0 == Op1) 1341 return Constant::getNullValue(Op0->getType()); 1342 1343 // undef >> X -> 0 1344 // undef >> X -> undef (if it's exact) 1345 if (match(Op0, m_Undef())) 1346 return isExact ? Op0 : Constant::getNullValue(Op0->getType()); 1347 1348 // The low bit cannot be shifted out of an exact shift if it is set. 1349 if (isExact) { 1350 unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); 1351 APInt Op0KnownZero(BitWidth, 0); 1352 APInt Op0KnownOne(BitWidth, 0); 1353 computeKnownBits(Op0, Op0KnownZero, Op0KnownOne, Q.DL, /*Depth=*/0, Q.AC, 1354 Q.CxtI, Q.DT); 1355 if (Op0KnownOne[0]) 1356 return Op0; 1357 } 1358 1359 return nullptr; 1360 } 1361 1362 /// SimplifyShlInst - Given operands for an Shl, see if we can 1363 /// fold the result. If not, this returns null. 1364 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1365 const Query &Q, unsigned MaxRecurse) { 1366 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse)) 1367 return V; 1368 1369 // undef << X -> 0 1370 // undef << X -> undef if (if it's NSW/NUW) 1371 if (match(Op0, m_Undef())) 1372 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType()); 1373 1374 // (X >> A) << A -> X 1375 Value *X; 1376 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1))))) 1377 return X; 1378 return nullptr; 1379 } 1380 1381 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1382 const DataLayout &DL, const TargetLibraryInfo *TLI, 1383 const DominatorTree *DT, AssumptionCache *AC, 1384 const Instruction *CxtI) { 1385 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI), 1386 RecursionLimit); 1387 } 1388 1389 /// SimplifyLShrInst - Given operands for an LShr, see if we can 1390 /// fold the result. If not, this returns null. 1391 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1392 const Query &Q, unsigned MaxRecurse) { 1393 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q, 1394 MaxRecurse)) 1395 return V; 1396 1397 // (X << A) >> A -> X 1398 Value *X; 1399 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1)))) 1400 return X; 1401 1402 return nullptr; 1403 } 1404 1405 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1406 const DataLayout &DL, 1407 const TargetLibraryInfo *TLI, 1408 const DominatorTree *DT, AssumptionCache *AC, 1409 const Instruction *CxtI) { 1410 return ::SimplifyLShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1411 RecursionLimit); 1412 } 1413 1414 /// SimplifyAShrInst - Given operands for an AShr, see if we can 1415 /// fold the result. If not, this returns null. 1416 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1417 const Query &Q, unsigned MaxRecurse) { 1418 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q, 1419 MaxRecurse)) 1420 return V; 1421 1422 // all ones >>a X -> all ones 1423 if (match(Op0, m_AllOnes())) 1424 return Op0; 1425 1426 // (X << A) >> A -> X 1427 Value *X; 1428 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1)))) 1429 return X; 1430 1431 // Arithmetic shifting an all-sign-bit value is a no-op. 1432 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); 1433 if (NumSignBits == Op0->getType()->getScalarSizeInBits()) 1434 return Op0; 1435 1436 return nullptr; 1437 } 1438 1439 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1440 const DataLayout &DL, 1441 const TargetLibraryInfo *TLI, 1442 const DominatorTree *DT, AssumptionCache *AC, 1443 const Instruction *CxtI) { 1444 return ::SimplifyAShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI), 1445 RecursionLimit); 1446 } 1447 1448 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp, 1449 ICmpInst *UnsignedICmp, bool IsAnd) { 1450 Value *X, *Y; 1451 1452 ICmpInst::Predicate EqPred; 1453 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) || 1454 !ICmpInst::isEquality(EqPred)) 1455 return nullptr; 1456 1457 ICmpInst::Predicate UnsignedPred; 1458 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) && 1459 ICmpInst::isUnsigned(UnsignedPred)) 1460 ; 1461 else if (match(UnsignedICmp, 1462 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) && 1463 ICmpInst::isUnsigned(UnsignedPred)) 1464 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred); 1465 else 1466 return nullptr; 1467 1468 // X < Y && Y != 0 --> X < Y 1469 // X < Y || Y != 0 --> Y != 0 1470 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE) 1471 return IsAnd ? UnsignedICmp : ZeroICmp; 1472 1473 // X >= Y || Y != 0 --> true 1474 // X >= Y || Y == 0 --> X >= Y 1475 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) { 1476 if (EqPred == ICmpInst::ICMP_NE) 1477 return getTrue(UnsignedICmp->getType()); 1478 return UnsignedICmp; 1479 } 1480 1481 // X < Y && Y == 0 --> false 1482 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ && 1483 IsAnd) 1484 return getFalse(UnsignedICmp->getType()); 1485 1486 return nullptr; 1487 } 1488 1489 // Simplify (and (icmp ...) (icmp ...)) to true when we can tell that the range 1490 // of possible values cannot be satisfied. 1491 static Value *SimplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1492 ICmpInst::Predicate Pred0, Pred1; 1493 ConstantInt *CI1, *CI2; 1494 Value *V; 1495 1496 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true)) 1497 return X; 1498 1499 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)), 1500 m_ConstantInt(CI2)))) 1501 return nullptr; 1502 1503 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1)))) 1504 return nullptr; 1505 1506 Type *ITy = Op0->getType(); 1507 1508 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1509 bool isNSW = AddInst->hasNoSignedWrap(); 1510 bool isNUW = AddInst->hasNoUnsignedWrap(); 1511 1512 const APInt &CI1V = CI1->getValue(); 1513 const APInt &CI2V = CI2->getValue(); 1514 const APInt Delta = CI2V - CI1V; 1515 if (CI1V.isStrictlyPositive()) { 1516 if (Delta == 2) { 1517 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT) 1518 return getFalse(ITy); 1519 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1520 return getFalse(ITy); 1521 } 1522 if (Delta == 1) { 1523 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT) 1524 return getFalse(ITy); 1525 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW) 1526 return getFalse(ITy); 1527 } 1528 } 1529 if (CI1V.getBoolValue() && isNUW) { 1530 if (Delta == 2) 1531 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT) 1532 return getFalse(ITy); 1533 if (Delta == 1) 1534 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT) 1535 return getFalse(ITy); 1536 } 1537 1538 return nullptr; 1539 } 1540 1541 /// SimplifyAndInst - Given operands for an And, see if we can 1542 /// fold the result. If not, this returns null. 1543 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q, 1544 unsigned MaxRecurse) { 1545 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1546 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1547 Constant *Ops[] = { CLHS, CRHS }; 1548 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(), 1549 Ops, Q.DL, Q.TLI); 1550 } 1551 1552 // Canonicalize the constant to the RHS. 1553 std::swap(Op0, Op1); 1554 } 1555 1556 // X & undef -> 0 1557 if (match(Op1, m_Undef())) 1558 return Constant::getNullValue(Op0->getType()); 1559 1560 // X & X = X 1561 if (Op0 == Op1) 1562 return Op0; 1563 1564 // X & 0 = 0 1565 if (match(Op1, m_Zero())) 1566 return Op1; 1567 1568 // X & -1 = X 1569 if (match(Op1, m_AllOnes())) 1570 return Op0; 1571 1572 // A & ~A = ~A & A = 0 1573 if (match(Op0, m_Not(m_Specific(Op1))) || 1574 match(Op1, m_Not(m_Specific(Op0)))) 1575 return Constant::getNullValue(Op0->getType()); 1576 1577 // (A | ?) & A = A 1578 Value *A = nullptr, *B = nullptr; 1579 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1580 (A == Op1 || B == Op1)) 1581 return Op1; 1582 1583 // A & (A | ?) = A 1584 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1585 (A == Op0 || B == Op0)) 1586 return Op0; 1587 1588 // A & (-A) = A if A is a power of two or zero. 1589 if (match(Op0, m_Neg(m_Specific(Op1))) || 1590 match(Op1, m_Neg(m_Specific(Op0)))) { 1591 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, 1592 Q.DT)) 1593 return Op0; 1594 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, 1595 Q.DT)) 1596 return Op1; 1597 } 1598 1599 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1600 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1601 if (Value *V = SimplifyAndOfICmps(ICILHS, ICIRHS)) 1602 return V; 1603 if (Value *V = SimplifyAndOfICmps(ICIRHS, ICILHS)) 1604 return V; 1605 } 1606 } 1607 1608 // Try some generic simplifications for associative operations. 1609 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q, 1610 MaxRecurse)) 1611 return V; 1612 1613 // And distributes over Or. Try some generic simplifications based on this. 1614 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or, 1615 Q, MaxRecurse)) 1616 return V; 1617 1618 // And distributes over Xor. Try some generic simplifications based on this. 1619 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor, 1620 Q, MaxRecurse)) 1621 return V; 1622 1623 // If the operation is with the result of a select instruction, check whether 1624 // operating on either branch of the select always yields the same value. 1625 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1626 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q, 1627 MaxRecurse)) 1628 return V; 1629 1630 // If the operation is with the result of a phi instruction, check whether 1631 // operating on all incoming values of the phi always yields the same value. 1632 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1633 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q, 1634 MaxRecurse)) 1635 return V; 1636 1637 return nullptr; 1638 } 1639 1640 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout &DL, 1641 const TargetLibraryInfo *TLI, 1642 const DominatorTree *DT, AssumptionCache *AC, 1643 const Instruction *CxtI) { 1644 return ::SimplifyAndInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1645 RecursionLimit); 1646 } 1647 1648 // Simplify (or (icmp ...) (icmp ...)) to true when we can tell that the union 1649 // contains all possible values. 1650 static Value *SimplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) { 1651 ICmpInst::Predicate Pred0, Pred1; 1652 ConstantInt *CI1, *CI2; 1653 Value *V; 1654 1655 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false)) 1656 return X; 1657 1658 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)), 1659 m_ConstantInt(CI2)))) 1660 return nullptr; 1661 1662 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1)))) 1663 return nullptr; 1664 1665 Type *ITy = Op0->getType(); 1666 1667 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0)); 1668 bool isNSW = AddInst->hasNoSignedWrap(); 1669 bool isNUW = AddInst->hasNoUnsignedWrap(); 1670 1671 const APInt &CI1V = CI1->getValue(); 1672 const APInt &CI2V = CI2->getValue(); 1673 const APInt Delta = CI2V - CI1V; 1674 if (CI1V.isStrictlyPositive()) { 1675 if (Delta == 2) { 1676 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE) 1677 return getTrue(ITy); 1678 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1679 return getTrue(ITy); 1680 } 1681 if (Delta == 1) { 1682 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE) 1683 return getTrue(ITy); 1684 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW) 1685 return getTrue(ITy); 1686 } 1687 } 1688 if (CI1V.getBoolValue() && isNUW) { 1689 if (Delta == 2) 1690 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE) 1691 return getTrue(ITy); 1692 if (Delta == 1) 1693 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE) 1694 return getTrue(ITy); 1695 } 1696 1697 return nullptr; 1698 } 1699 1700 /// SimplifyOrInst - Given operands for an Or, see if we can 1701 /// fold the result. If not, this returns null. 1702 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q, 1703 unsigned MaxRecurse) { 1704 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1705 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1706 Constant *Ops[] = { CLHS, CRHS }; 1707 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(), 1708 Ops, Q.DL, Q.TLI); 1709 } 1710 1711 // Canonicalize the constant to the RHS. 1712 std::swap(Op0, Op1); 1713 } 1714 1715 // X | undef -> -1 1716 if (match(Op1, m_Undef())) 1717 return Constant::getAllOnesValue(Op0->getType()); 1718 1719 // X | X = X 1720 if (Op0 == Op1) 1721 return Op0; 1722 1723 // X | 0 = X 1724 if (match(Op1, m_Zero())) 1725 return Op0; 1726 1727 // X | -1 = -1 1728 if (match(Op1, m_AllOnes())) 1729 return Op1; 1730 1731 // A | ~A = ~A | A = -1 1732 if (match(Op0, m_Not(m_Specific(Op1))) || 1733 match(Op1, m_Not(m_Specific(Op0)))) 1734 return Constant::getAllOnesValue(Op0->getType()); 1735 1736 // (A & ?) | A = A 1737 Value *A = nullptr, *B = nullptr; 1738 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 1739 (A == Op1 || B == Op1)) 1740 return Op1; 1741 1742 // A | (A & ?) = A 1743 if (match(Op1, m_And(m_Value(A), m_Value(B))) && 1744 (A == Op0 || B == Op0)) 1745 return Op0; 1746 1747 // ~(A & ?) | A = -1 1748 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) && 1749 (A == Op1 || B == Op1)) 1750 return Constant::getAllOnesValue(Op1->getType()); 1751 1752 // A | ~(A & ?) = -1 1753 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) && 1754 (A == Op0 || B == Op0)) 1755 return Constant::getAllOnesValue(Op0->getType()); 1756 1757 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) { 1758 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) { 1759 if (Value *V = SimplifyOrOfICmps(ICILHS, ICIRHS)) 1760 return V; 1761 if (Value *V = SimplifyOrOfICmps(ICIRHS, ICILHS)) 1762 return V; 1763 } 1764 } 1765 1766 // Try some generic simplifications for associative operations. 1767 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q, 1768 MaxRecurse)) 1769 return V; 1770 1771 // Or distributes over And. Try some generic simplifications based on this. 1772 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q, 1773 MaxRecurse)) 1774 return V; 1775 1776 // If the operation is with the result of a select instruction, check whether 1777 // operating on either branch of the select always yields the same value. 1778 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1779 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q, 1780 MaxRecurse)) 1781 return V; 1782 1783 // (A & C)|(B & D) 1784 Value *C = nullptr, *D = nullptr; 1785 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 1786 match(Op1, m_And(m_Value(B), m_Value(D)))) { 1787 ConstantInt *C1 = dyn_cast<ConstantInt>(C); 1788 ConstantInt *C2 = dyn_cast<ConstantInt>(D); 1789 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) { 1790 // (A & C1)|(B & C2) 1791 // If we have: ((V + N) & C1) | (V & C2) 1792 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0 1793 // replace with V+N. 1794 Value *V1, *V2; 1795 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+ 1796 match(A, m_Add(m_Value(V1), m_Value(V2)))) { 1797 // Add commutes, try both ways. 1798 if (V1 == B && 1799 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1800 return A; 1801 if (V2 == B && 1802 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1803 return A; 1804 } 1805 // Or commutes, try both ways. 1806 if ((C1->getValue() & (C1->getValue() + 1)) == 0 && 1807 match(B, m_Add(m_Value(V1), m_Value(V2)))) { 1808 // Add commutes, try both ways. 1809 if (V1 == A && 1810 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1811 return B; 1812 if (V2 == A && 1813 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 1814 return B; 1815 } 1816 } 1817 } 1818 1819 // If the operation is with the result of a phi instruction, check whether 1820 // operating on all incoming values of the phi always yields the same value. 1821 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1822 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse)) 1823 return V; 1824 1825 return nullptr; 1826 } 1827 1828 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout &DL, 1829 const TargetLibraryInfo *TLI, 1830 const DominatorTree *DT, AssumptionCache *AC, 1831 const Instruction *CxtI) { 1832 return ::SimplifyOrInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1833 RecursionLimit); 1834 } 1835 1836 /// SimplifyXorInst - Given operands for a Xor, see if we can 1837 /// fold the result. If not, this returns null. 1838 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q, 1839 unsigned MaxRecurse) { 1840 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1841 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1842 Constant *Ops[] = { CLHS, CRHS }; 1843 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(), 1844 Ops, Q.DL, Q.TLI); 1845 } 1846 1847 // Canonicalize the constant to the RHS. 1848 std::swap(Op0, Op1); 1849 } 1850 1851 // A ^ undef -> undef 1852 if (match(Op1, m_Undef())) 1853 return Op1; 1854 1855 // A ^ 0 = A 1856 if (match(Op1, m_Zero())) 1857 return Op0; 1858 1859 // A ^ A = 0 1860 if (Op0 == Op1) 1861 return Constant::getNullValue(Op0->getType()); 1862 1863 // A ^ ~A = ~A ^ A = -1 1864 if (match(Op0, m_Not(m_Specific(Op1))) || 1865 match(Op1, m_Not(m_Specific(Op0)))) 1866 return Constant::getAllOnesValue(Op0->getType()); 1867 1868 // Try some generic simplifications for associative operations. 1869 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q, 1870 MaxRecurse)) 1871 return V; 1872 1873 // Threading Xor over selects and phi nodes is pointless, so don't bother. 1874 // Threading over the select in "A ^ select(cond, B, C)" means evaluating 1875 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and 1876 // only if B and C are equal. If B and C are equal then (since we assume 1877 // that operands have already been simplified) "select(cond, B, C)" should 1878 // have been simplified to the common value of B and C already. Analysing 1879 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly 1880 // for threading over phi nodes. 1881 1882 return nullptr; 1883 } 1884 1885 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout &DL, 1886 const TargetLibraryInfo *TLI, 1887 const DominatorTree *DT, AssumptionCache *AC, 1888 const Instruction *CxtI) { 1889 return ::SimplifyXorInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI), 1890 RecursionLimit); 1891 } 1892 1893 static Type *GetCompareTy(Value *Op) { 1894 return CmpInst::makeCmpResultType(Op->getType()); 1895 } 1896 1897 /// ExtractEquivalentCondition - Rummage around inside V looking for something 1898 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found, 1899 /// otherwise return null. Helper function for analyzing max/min idioms. 1900 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred, 1901 Value *LHS, Value *RHS) { 1902 SelectInst *SI = dyn_cast<SelectInst>(V); 1903 if (!SI) 1904 return nullptr; 1905 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); 1906 if (!Cmp) 1907 return nullptr; 1908 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1); 1909 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS) 1910 return Cmp; 1911 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) && 1912 LHS == CmpRHS && RHS == CmpLHS) 1913 return Cmp; 1914 return nullptr; 1915 } 1916 1917 // A significant optimization not implemented here is assuming that alloca 1918 // addresses are not equal to incoming argument values. They don't *alias*, 1919 // as we say, but that doesn't mean they aren't equal, so we take a 1920 // conservative approach. 1921 // 1922 // This is inspired in part by C++11 5.10p1: 1923 // "Two pointers of the same type compare equal if and only if they are both 1924 // null, both point to the same function, or both represent the same 1925 // address." 1926 // 1927 // This is pretty permissive. 1928 // 1929 // It's also partly due to C11 6.5.9p6: 1930 // "Two pointers compare equal if and only if both are null pointers, both are 1931 // pointers to the same object (including a pointer to an object and a 1932 // subobject at its beginning) or function, both are pointers to one past the 1933 // last element of the same array object, or one is a pointer to one past the 1934 // end of one array object and the other is a pointer to the start of a 1935 // different array object that happens to immediately follow the first array 1936 // object in the address space.) 1937 // 1938 // C11's version is more restrictive, however there's no reason why an argument 1939 // couldn't be a one-past-the-end value for a stack object in the caller and be 1940 // equal to the beginning of a stack object in the callee. 1941 // 1942 // If the C and C++ standards are ever made sufficiently restrictive in this 1943 // area, it may be possible to update LLVM's semantics accordingly and reinstate 1944 // this optimization. 1945 static Constant *computePointerICmp(const DataLayout &DL, 1946 const TargetLibraryInfo *TLI, 1947 CmpInst::Predicate Pred, Value *LHS, 1948 Value *RHS) { 1949 // First, skip past any trivial no-ops. 1950 LHS = LHS->stripPointerCasts(); 1951 RHS = RHS->stripPointerCasts(); 1952 1953 // A non-null pointer is not equal to a null pointer. 1954 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) && 1955 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE)) 1956 return ConstantInt::get(GetCompareTy(LHS), 1957 !CmpInst::isTrueWhenEqual(Pred)); 1958 1959 // We can only fold certain predicates on pointer comparisons. 1960 switch (Pred) { 1961 default: 1962 return nullptr; 1963 1964 // Equality comaprisons are easy to fold. 1965 case CmpInst::ICMP_EQ: 1966 case CmpInst::ICMP_NE: 1967 break; 1968 1969 // We can only handle unsigned relational comparisons because 'inbounds' on 1970 // a GEP only protects against unsigned wrapping. 1971 case CmpInst::ICMP_UGT: 1972 case CmpInst::ICMP_UGE: 1973 case CmpInst::ICMP_ULT: 1974 case CmpInst::ICMP_ULE: 1975 // However, we have to switch them to their signed variants to handle 1976 // negative indices from the base pointer. 1977 Pred = ICmpInst::getSignedPredicate(Pred); 1978 break; 1979 } 1980 1981 // Strip off any constant offsets so that we can reason about them. 1982 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets 1983 // here and compare base addresses like AliasAnalysis does, however there are 1984 // numerous hazards. AliasAnalysis and its utilities rely on special rules 1985 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis 1986 // doesn't need to guarantee pointer inequality when it says NoAlias. 1987 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); 1988 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); 1989 1990 // If LHS and RHS are related via constant offsets to the same base 1991 // value, we can replace it with an icmp which just compares the offsets. 1992 if (LHS == RHS) 1993 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset); 1994 1995 // Various optimizations for (in)equality comparisons. 1996 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) { 1997 // Different non-empty allocations that exist at the same time have 1998 // different addresses (if the program can tell). Global variables always 1999 // exist, so they always exist during the lifetime of each other and all 2000 // allocas. Two different allocas usually have different addresses... 2001 // 2002 // However, if there's an @llvm.stackrestore dynamically in between two 2003 // allocas, they may have the same address. It's tempting to reduce the 2004 // scope of the problem by only looking at *static* allocas here. That would 2005 // cover the majority of allocas while significantly reducing the likelihood 2006 // of having an @llvm.stackrestore pop up in the middle. However, it's not 2007 // actually impossible for an @llvm.stackrestore to pop up in the middle of 2008 // an entry block. Also, if we have a block that's not attached to a 2009 // function, we can't tell if it's "static" under the current definition. 2010 // Theoretically, this problem could be fixed by creating a new kind of 2011 // instruction kind specifically for static allocas. Such a new instruction 2012 // could be required to be at the top of the entry block, thus preventing it 2013 // from being subject to a @llvm.stackrestore. Instcombine could even 2014 // convert regular allocas into these special allocas. It'd be nifty. 2015 // However, until then, this problem remains open. 2016 // 2017 // So, we'll assume that two non-empty allocas have different addresses 2018 // for now. 2019 // 2020 // With all that, if the offsets are within the bounds of their allocations 2021 // (and not one-past-the-end! so we can't use inbounds!), and their 2022 // allocations aren't the same, the pointers are not equal. 2023 // 2024 // Note that it's not necessary to check for LHS being a global variable 2025 // address, due to canonicalization and constant folding. 2026 if (isa<AllocaInst>(LHS) && 2027 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) { 2028 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset); 2029 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset); 2030 uint64_t LHSSize, RHSSize; 2031 if (LHSOffsetCI && RHSOffsetCI && 2032 getObjectSize(LHS, LHSSize, DL, TLI) && 2033 getObjectSize(RHS, RHSSize, DL, TLI)) { 2034 const APInt &LHSOffsetValue = LHSOffsetCI->getValue(); 2035 const APInt &RHSOffsetValue = RHSOffsetCI->getValue(); 2036 if (!LHSOffsetValue.isNegative() && 2037 !RHSOffsetValue.isNegative() && 2038 LHSOffsetValue.ult(LHSSize) && 2039 RHSOffsetValue.ult(RHSSize)) { 2040 return ConstantInt::get(GetCompareTy(LHS), 2041 !CmpInst::isTrueWhenEqual(Pred)); 2042 } 2043 } 2044 2045 // Repeat the above check but this time without depending on DataLayout 2046 // or being able to compute a precise size. 2047 if (!cast<PointerType>(LHS->getType())->isEmptyTy() && 2048 !cast<PointerType>(RHS->getType())->isEmptyTy() && 2049 LHSOffset->isNullValue() && 2050 RHSOffset->isNullValue()) 2051 return ConstantInt::get(GetCompareTy(LHS), 2052 !CmpInst::isTrueWhenEqual(Pred)); 2053 } 2054 2055 // Even if an non-inbounds GEP occurs along the path we can still optimize 2056 // equality comparisons concerning the result. We avoid walking the whole 2057 // chain again by starting where the last calls to 2058 // stripAndComputeConstantOffsets left off and accumulate the offsets. 2059 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true); 2060 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true); 2061 if (LHS == RHS) 2062 return ConstantExpr::getICmp(Pred, 2063 ConstantExpr::getAdd(LHSOffset, LHSNoBound), 2064 ConstantExpr::getAdd(RHSOffset, RHSNoBound)); 2065 2066 // If one side of the equality comparison must come from a noalias call 2067 // (meaning a system memory allocation function), and the other side must 2068 // come from a pointer that cannot overlap with dynamically-allocated 2069 // memory within the lifetime of the current function (allocas, byval 2070 // arguments, globals), then determine the comparison result here. 2071 SmallVector<Value *, 8> LHSUObjs, RHSUObjs; 2072 GetUnderlyingObjects(LHS, LHSUObjs, DL); 2073 GetUnderlyingObjects(RHS, RHSUObjs, DL); 2074 2075 // Is the set of underlying objects all noalias calls? 2076 auto IsNAC = [](SmallVectorImpl<Value *> &Objects) { 2077 return std::all_of(Objects.begin(), Objects.end(), 2078 [](Value *V){ return isNoAliasCall(V); }); 2079 }; 2080 2081 // Is the set of underlying objects all things which must be disjoint from 2082 // noalias calls. For allocas, we consider only static ones (dynamic 2083 // allocas might be transformed into calls to malloc not simultaneously 2084 // live with the compared-to allocation). For globals, we exclude symbols 2085 // that might be resolve lazily to symbols in another dynamically-loaded 2086 // library (and, thus, could be malloc'ed by the implementation). 2087 auto IsAllocDisjoint = [](SmallVectorImpl<Value *> &Objects) { 2088 return std::all_of(Objects.begin(), Objects.end(), 2089 [](Value *V){ 2090 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) 2091 return AI->getParent() && AI->getParent()->getParent() && 2092 AI->isStaticAlloca(); 2093 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 2094 return (GV->hasLocalLinkage() || 2095 GV->hasHiddenVisibility() || 2096 GV->hasProtectedVisibility() || 2097 GV->hasUnnamedAddr()) && 2098 !GV->isThreadLocal(); 2099 if (const Argument *A = dyn_cast<Argument>(V)) 2100 return A->hasByValAttr(); 2101 return false; 2102 }); 2103 }; 2104 2105 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) || 2106 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs))) 2107 return ConstantInt::get(GetCompareTy(LHS), 2108 !CmpInst::isTrueWhenEqual(Pred)); 2109 } 2110 2111 // Otherwise, fail. 2112 return nullptr; 2113 } 2114 2115 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can 2116 /// fold the result. If not, this returns null. 2117 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2118 const Query &Q, unsigned MaxRecurse) { 2119 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 2120 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!"); 2121 2122 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 2123 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 2124 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 2125 2126 // If we have a constant, make sure it is on the RHS. 2127 std::swap(LHS, RHS); 2128 Pred = CmpInst::getSwappedPredicate(Pred); 2129 } 2130 2131 Type *ITy = GetCompareTy(LHS); // The return type. 2132 Type *OpTy = LHS->getType(); // The operand type. 2133 2134 // icmp X, X -> true/false 2135 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false 2136 // because X could be 0. 2137 if (LHS == RHS || isa<UndefValue>(RHS)) 2138 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred)); 2139 2140 // Special case logic when the operands have i1 type. 2141 if (OpTy->getScalarType()->isIntegerTy(1)) { 2142 switch (Pred) { 2143 default: break; 2144 case ICmpInst::ICMP_EQ: 2145 // X == 1 -> X 2146 if (match(RHS, m_One())) 2147 return LHS; 2148 break; 2149 case ICmpInst::ICMP_NE: 2150 // X != 0 -> X 2151 if (match(RHS, m_Zero())) 2152 return LHS; 2153 break; 2154 case ICmpInst::ICMP_UGT: 2155 // X >u 0 -> X 2156 if (match(RHS, m_Zero())) 2157 return LHS; 2158 break; 2159 case ICmpInst::ICMP_UGE: 2160 // X >=u 1 -> X 2161 if (match(RHS, m_One())) 2162 return LHS; 2163 break; 2164 case ICmpInst::ICMP_SLT: 2165 // X <s 0 -> X 2166 if (match(RHS, m_Zero())) 2167 return LHS; 2168 break; 2169 case ICmpInst::ICMP_SLE: 2170 // X <=s -1 -> X 2171 if (match(RHS, m_One())) 2172 return LHS; 2173 break; 2174 } 2175 } 2176 2177 // If we are comparing with zero then try hard since this is a common case. 2178 if (match(RHS, m_Zero())) { 2179 bool LHSKnownNonNegative, LHSKnownNegative; 2180 switch (Pred) { 2181 default: llvm_unreachable("Unknown ICmp predicate!"); 2182 case ICmpInst::ICMP_ULT: 2183 return getFalse(ITy); 2184 case ICmpInst::ICMP_UGE: 2185 return getTrue(ITy); 2186 case ICmpInst::ICMP_EQ: 2187 case ICmpInst::ICMP_ULE: 2188 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2189 return getFalse(ITy); 2190 break; 2191 case ICmpInst::ICMP_NE: 2192 case ICmpInst::ICMP_UGT: 2193 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2194 return getTrue(ITy); 2195 break; 2196 case ICmpInst::ICMP_SLT: 2197 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2198 Q.CxtI, Q.DT); 2199 if (LHSKnownNegative) 2200 return getTrue(ITy); 2201 if (LHSKnownNonNegative) 2202 return getFalse(ITy); 2203 break; 2204 case ICmpInst::ICMP_SLE: 2205 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2206 Q.CxtI, Q.DT); 2207 if (LHSKnownNegative) 2208 return getTrue(ITy); 2209 if (LHSKnownNonNegative && 2210 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2211 return getFalse(ITy); 2212 break; 2213 case ICmpInst::ICMP_SGE: 2214 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2215 Q.CxtI, Q.DT); 2216 if (LHSKnownNegative) 2217 return getFalse(ITy); 2218 if (LHSKnownNonNegative) 2219 return getTrue(ITy); 2220 break; 2221 case ICmpInst::ICMP_SGT: 2222 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC, 2223 Q.CxtI, Q.DT); 2224 if (LHSKnownNegative) 2225 return getFalse(ITy); 2226 if (LHSKnownNonNegative && 2227 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT)) 2228 return getTrue(ITy); 2229 break; 2230 } 2231 } 2232 2233 // See if we are doing a comparison with a constant integer. 2234 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2235 // Rule out tautological comparisons (eg., ult 0 or uge 0). 2236 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue()); 2237 if (RHS_CR.isEmptySet()) 2238 return ConstantInt::getFalse(CI->getContext()); 2239 if (RHS_CR.isFullSet()) 2240 return ConstantInt::getTrue(CI->getContext()); 2241 2242 // Many binary operators with constant RHS have easy to compute constant 2243 // range. Use them to check whether the comparison is a tautology. 2244 unsigned Width = CI->getBitWidth(); 2245 APInt Lower = APInt(Width, 0); 2246 APInt Upper = APInt(Width, 0); 2247 ConstantInt *CI2; 2248 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) { 2249 // 'urem x, CI2' produces [0, CI2). 2250 Upper = CI2->getValue(); 2251 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) { 2252 // 'srem x, CI2' produces (-|CI2|, |CI2|). 2253 Upper = CI2->getValue().abs(); 2254 Lower = (-Upper) + 1; 2255 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) { 2256 // 'udiv CI2, x' produces [0, CI2]. 2257 Upper = CI2->getValue() + 1; 2258 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) { 2259 // 'udiv x, CI2' produces [0, UINT_MAX / CI2]. 2260 APInt NegOne = APInt::getAllOnesValue(Width); 2261 if (!CI2->isZero()) 2262 Upper = NegOne.udiv(CI2->getValue()) + 1; 2263 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) { 2264 if (CI2->isMinSignedValue()) { 2265 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. 2266 Lower = CI2->getValue(); 2267 Upper = Lower.lshr(1) + 1; 2268 } else { 2269 // 'sdiv CI2, x' produces [-|CI2|, |CI2|]. 2270 Upper = CI2->getValue().abs() + 1; 2271 Lower = (-Upper) + 1; 2272 } 2273 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) { 2274 APInt IntMin = APInt::getSignedMinValue(Width); 2275 APInt IntMax = APInt::getSignedMaxValue(Width); 2276 APInt Val = CI2->getValue(); 2277 if (Val.isAllOnesValue()) { 2278 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] 2279 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2280 Lower = IntMin + 1; 2281 Upper = IntMax + 1; 2282 } else if (Val.countLeadingZeros() < Width - 1) { 2283 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2] 2284 // where CI2 != -1 and CI2 != 0 and CI2 != 1 2285 Lower = IntMin.sdiv(Val); 2286 Upper = IntMax.sdiv(Val); 2287 if (Lower.sgt(Upper)) 2288 std::swap(Lower, Upper); 2289 Upper = Upper + 1; 2290 assert(Upper != Lower && "Upper part of range has wrapped!"); 2291 } 2292 } else if (match(LHS, m_NUWShl(m_ConstantInt(CI2), m_Value()))) { 2293 // 'shl nuw CI2, x' produces [CI2, CI2 << CLZ(CI2)] 2294 Lower = CI2->getValue(); 2295 Upper = Lower.shl(Lower.countLeadingZeros()) + 1; 2296 } else if (match(LHS, m_NSWShl(m_ConstantInt(CI2), m_Value()))) { 2297 if (CI2->isNegative()) { 2298 // 'shl nsw CI2, x' produces [CI2 << CLO(CI2)-1, CI2] 2299 unsigned ShiftAmount = CI2->getValue().countLeadingOnes() - 1; 2300 Lower = CI2->getValue().shl(ShiftAmount); 2301 Upper = CI2->getValue() + 1; 2302 } else { 2303 // 'shl nsw CI2, x' produces [CI2, CI2 << CLZ(CI2)-1] 2304 unsigned ShiftAmount = CI2->getValue().countLeadingZeros() - 1; 2305 Lower = CI2->getValue(); 2306 Upper = CI2->getValue().shl(ShiftAmount) + 1; 2307 } 2308 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) { 2309 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2]. 2310 APInt NegOne = APInt::getAllOnesValue(Width); 2311 if (CI2->getValue().ult(Width)) 2312 Upper = NegOne.lshr(CI2->getValue()) + 1; 2313 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) { 2314 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2]. 2315 unsigned ShiftAmount = Width - 1; 2316 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2317 ShiftAmount = CI2->getValue().countTrailingZeros(); 2318 Lower = CI2->getValue().lshr(ShiftAmount); 2319 Upper = CI2->getValue() + 1; 2320 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) { 2321 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2]. 2322 APInt IntMin = APInt::getSignedMinValue(Width); 2323 APInt IntMax = APInt::getSignedMaxValue(Width); 2324 if (CI2->getValue().ult(Width)) { 2325 Lower = IntMin.ashr(CI2->getValue()); 2326 Upper = IntMax.ashr(CI2->getValue()) + 1; 2327 } 2328 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) { 2329 unsigned ShiftAmount = Width - 1; 2330 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact()) 2331 ShiftAmount = CI2->getValue().countTrailingZeros(); 2332 if (CI2->isNegative()) { 2333 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)] 2334 Lower = CI2->getValue(); 2335 Upper = CI2->getValue().ashr(ShiftAmount) + 1; 2336 } else { 2337 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2] 2338 Lower = CI2->getValue().ashr(ShiftAmount); 2339 Upper = CI2->getValue() + 1; 2340 } 2341 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) { 2342 // 'or x, CI2' produces [CI2, UINT_MAX]. 2343 Lower = CI2->getValue(); 2344 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) { 2345 // 'and x, CI2' produces [0, CI2]. 2346 Upper = CI2->getValue() + 1; 2347 } 2348 if (Lower != Upper) { 2349 ConstantRange LHS_CR = ConstantRange(Lower, Upper); 2350 if (RHS_CR.contains(LHS_CR)) 2351 return ConstantInt::getTrue(RHS->getContext()); 2352 if (RHS_CR.inverse().contains(LHS_CR)) 2353 return ConstantInt::getFalse(RHS->getContext()); 2354 } 2355 } 2356 2357 // Compare of cast, for example (zext X) != 0 -> X != 0 2358 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) { 2359 Instruction *LI = cast<CastInst>(LHS); 2360 Value *SrcOp = LI->getOperand(0); 2361 Type *SrcTy = SrcOp->getType(); 2362 Type *DstTy = LI->getType(); 2363 2364 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input 2365 // if the integer type is the same size as the pointer type. 2366 if (MaxRecurse && isa<PtrToIntInst>(LI) && 2367 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) { 2368 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 2369 // Transfer the cast to the constant. 2370 if (Value *V = SimplifyICmpInst(Pred, SrcOp, 2371 ConstantExpr::getIntToPtr(RHSC, SrcTy), 2372 Q, MaxRecurse-1)) 2373 return V; 2374 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) { 2375 if (RI->getOperand(0)->getType() == SrcTy) 2376 // Compare without the cast. 2377 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2378 Q, MaxRecurse-1)) 2379 return V; 2380 } 2381 } 2382 2383 if (isa<ZExtInst>(LHS)) { 2384 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the 2385 // same type. 2386 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) { 2387 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2388 // Compare X and Y. Note that signed predicates become unsigned. 2389 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2390 SrcOp, RI->getOperand(0), Q, 2391 MaxRecurse-1)) 2392 return V; 2393 } 2394 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended 2395 // too. If not, then try to deduce the result of the comparison. 2396 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2397 // Compute the constant that would happen if we truncated to SrcTy then 2398 // reextended to DstTy. 2399 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2400 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy); 2401 2402 // If the re-extended constant didn't change then this is effectively 2403 // also a case of comparing two zero-extended values. 2404 if (RExt == CI && MaxRecurse) 2405 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2406 SrcOp, Trunc, Q, MaxRecurse-1)) 2407 return V; 2408 2409 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit 2410 // there. Use this to work out the result of the comparison. 2411 if (RExt != CI) { 2412 switch (Pred) { 2413 default: llvm_unreachable("Unknown ICmp predicate!"); 2414 // LHS <u RHS. 2415 case ICmpInst::ICMP_EQ: 2416 case ICmpInst::ICMP_UGT: 2417 case ICmpInst::ICMP_UGE: 2418 return ConstantInt::getFalse(CI->getContext()); 2419 2420 case ICmpInst::ICMP_NE: 2421 case ICmpInst::ICMP_ULT: 2422 case ICmpInst::ICMP_ULE: 2423 return ConstantInt::getTrue(CI->getContext()); 2424 2425 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS 2426 // is non-negative then LHS <s RHS. 2427 case ICmpInst::ICMP_SGT: 2428 case ICmpInst::ICMP_SGE: 2429 return CI->getValue().isNegative() ? 2430 ConstantInt::getTrue(CI->getContext()) : 2431 ConstantInt::getFalse(CI->getContext()); 2432 2433 case ICmpInst::ICMP_SLT: 2434 case ICmpInst::ICMP_SLE: 2435 return CI->getValue().isNegative() ? 2436 ConstantInt::getFalse(CI->getContext()) : 2437 ConstantInt::getTrue(CI->getContext()); 2438 } 2439 } 2440 } 2441 } 2442 2443 if (isa<SExtInst>(LHS)) { 2444 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the 2445 // same type. 2446 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) { 2447 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2448 // Compare X and Y. Note that the predicate does not change. 2449 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2450 Q, MaxRecurse-1)) 2451 return V; 2452 } 2453 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended 2454 // too. If not, then try to deduce the result of the comparison. 2455 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2456 // Compute the constant that would happen if we truncated to SrcTy then 2457 // reextended to DstTy. 2458 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2459 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy); 2460 2461 // If the re-extended constant didn't change then this is effectively 2462 // also a case of comparing two sign-extended values. 2463 if (RExt == CI && MaxRecurse) 2464 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1)) 2465 return V; 2466 2467 // Otherwise the upper bits of LHS are all equal, while RHS has varying 2468 // bits there. Use this to work out the result of the comparison. 2469 if (RExt != CI) { 2470 switch (Pred) { 2471 default: llvm_unreachable("Unknown ICmp predicate!"); 2472 case ICmpInst::ICMP_EQ: 2473 return ConstantInt::getFalse(CI->getContext()); 2474 case ICmpInst::ICMP_NE: 2475 return ConstantInt::getTrue(CI->getContext()); 2476 2477 // If RHS is non-negative then LHS <s RHS. If RHS is negative then 2478 // LHS >s RHS. 2479 case ICmpInst::ICMP_SGT: 2480 case ICmpInst::ICMP_SGE: 2481 return CI->getValue().isNegative() ? 2482 ConstantInt::getTrue(CI->getContext()) : 2483 ConstantInt::getFalse(CI->getContext()); 2484 case ICmpInst::ICMP_SLT: 2485 case ICmpInst::ICMP_SLE: 2486 return CI->getValue().isNegative() ? 2487 ConstantInt::getFalse(CI->getContext()) : 2488 ConstantInt::getTrue(CI->getContext()); 2489 2490 // If LHS is non-negative then LHS <u RHS. If LHS is negative then 2491 // LHS >u RHS. 2492 case ICmpInst::ICMP_UGT: 2493 case ICmpInst::ICMP_UGE: 2494 // Comparison is true iff the LHS <s 0. 2495 if (MaxRecurse) 2496 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp, 2497 Constant::getNullValue(SrcTy), 2498 Q, MaxRecurse-1)) 2499 return V; 2500 break; 2501 case ICmpInst::ICMP_ULT: 2502 case ICmpInst::ICMP_ULE: 2503 // Comparison is true iff the LHS >=s 0. 2504 if (MaxRecurse) 2505 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp, 2506 Constant::getNullValue(SrcTy), 2507 Q, MaxRecurse-1)) 2508 return V; 2509 break; 2510 } 2511 } 2512 } 2513 } 2514 } 2515 2516 // Special logic for binary operators. 2517 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS); 2518 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS); 2519 if (MaxRecurse && (LBO || RBO)) { 2520 // Analyze the case when either LHS or RHS is an add instruction. 2521 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 2522 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null). 2523 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false; 2524 if (LBO && LBO->getOpcode() == Instruction::Add) { 2525 A = LBO->getOperand(0); B = LBO->getOperand(1); 2526 NoLHSWrapProblem = ICmpInst::isEquality(Pred) || 2527 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) || 2528 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap()); 2529 } 2530 if (RBO && RBO->getOpcode() == Instruction::Add) { 2531 C = RBO->getOperand(0); D = RBO->getOperand(1); 2532 NoRHSWrapProblem = ICmpInst::isEquality(Pred) || 2533 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) || 2534 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap()); 2535 } 2536 2537 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2538 if ((A == RHS || B == RHS) && NoLHSWrapProblem) 2539 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A, 2540 Constant::getNullValue(RHS->getType()), 2541 Q, MaxRecurse-1)) 2542 return V; 2543 2544 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2545 if ((C == LHS || D == LHS) && NoRHSWrapProblem) 2546 if (Value *V = SimplifyICmpInst(Pred, 2547 Constant::getNullValue(LHS->getType()), 2548 C == LHS ? D : C, Q, MaxRecurse-1)) 2549 return V; 2550 2551 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow. 2552 if (A && C && (A == C || A == D || B == C || B == D) && 2553 NoLHSWrapProblem && NoRHSWrapProblem) { 2554 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2555 Value *Y, *Z; 2556 if (A == C) { 2557 // C + B == C + D -> B == D 2558 Y = B; 2559 Z = D; 2560 } else if (A == D) { 2561 // D + B == C + D -> B == C 2562 Y = B; 2563 Z = C; 2564 } else if (B == C) { 2565 // A + C == C + D -> A == D 2566 Y = A; 2567 Z = D; 2568 } else { 2569 assert(B == D); 2570 // A + D == C + D -> A == C 2571 Y = A; 2572 Z = C; 2573 } 2574 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1)) 2575 return V; 2576 } 2577 } 2578 2579 // icmp pred (or X, Y), X 2580 if (LBO && match(LBO, m_CombineOr(m_Or(m_Value(), m_Specific(RHS)), 2581 m_Or(m_Specific(RHS), m_Value())))) { 2582 if (Pred == ICmpInst::ICMP_ULT) 2583 return getFalse(ITy); 2584 if (Pred == ICmpInst::ICMP_UGE) 2585 return getTrue(ITy); 2586 } 2587 // icmp pred X, (or X, Y) 2588 if (RBO && match(RBO, m_CombineOr(m_Or(m_Value(), m_Specific(LHS)), 2589 m_Or(m_Specific(LHS), m_Value())))) { 2590 if (Pred == ICmpInst::ICMP_ULE) 2591 return getTrue(ITy); 2592 if (Pred == ICmpInst::ICMP_UGT) 2593 return getFalse(ITy); 2594 } 2595 2596 // icmp pred (and X, Y), X 2597 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)), 2598 m_And(m_Specific(RHS), m_Value())))) { 2599 if (Pred == ICmpInst::ICMP_UGT) 2600 return getFalse(ITy); 2601 if (Pred == ICmpInst::ICMP_ULE) 2602 return getTrue(ITy); 2603 } 2604 // icmp pred X, (and X, Y) 2605 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)), 2606 m_And(m_Specific(LHS), m_Value())))) { 2607 if (Pred == ICmpInst::ICMP_UGE) 2608 return getTrue(ITy); 2609 if (Pred == ICmpInst::ICMP_ULT) 2610 return getFalse(ITy); 2611 } 2612 2613 // 0 - (zext X) pred C 2614 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) { 2615 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { 2616 if (RHSC->getValue().isStrictlyPositive()) { 2617 if (Pred == ICmpInst::ICMP_SLT) 2618 return ConstantInt::getTrue(RHSC->getContext()); 2619 if (Pred == ICmpInst::ICMP_SGE) 2620 return ConstantInt::getFalse(RHSC->getContext()); 2621 if (Pred == ICmpInst::ICMP_EQ) 2622 return ConstantInt::getFalse(RHSC->getContext()); 2623 if (Pred == ICmpInst::ICMP_NE) 2624 return ConstantInt::getTrue(RHSC->getContext()); 2625 } 2626 if (RHSC->getValue().isNonNegative()) { 2627 if (Pred == ICmpInst::ICMP_SLE) 2628 return ConstantInt::getTrue(RHSC->getContext()); 2629 if (Pred == ICmpInst::ICMP_SGT) 2630 return ConstantInt::getFalse(RHSC->getContext()); 2631 } 2632 } 2633 } 2634 2635 // icmp pred (urem X, Y), Y 2636 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) { 2637 bool KnownNonNegative, KnownNegative; 2638 switch (Pred) { 2639 default: 2640 break; 2641 case ICmpInst::ICMP_SGT: 2642 case ICmpInst::ICMP_SGE: 2643 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2644 Q.CxtI, Q.DT); 2645 if (!KnownNonNegative) 2646 break; 2647 // fall-through 2648 case ICmpInst::ICMP_EQ: 2649 case ICmpInst::ICMP_UGT: 2650 case ICmpInst::ICMP_UGE: 2651 return getFalse(ITy); 2652 case ICmpInst::ICMP_SLT: 2653 case ICmpInst::ICMP_SLE: 2654 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2655 Q.CxtI, Q.DT); 2656 if (!KnownNonNegative) 2657 break; 2658 // fall-through 2659 case ICmpInst::ICMP_NE: 2660 case ICmpInst::ICMP_ULT: 2661 case ICmpInst::ICMP_ULE: 2662 return getTrue(ITy); 2663 } 2664 } 2665 2666 // icmp pred X, (urem Y, X) 2667 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) { 2668 bool KnownNonNegative, KnownNegative; 2669 switch (Pred) { 2670 default: 2671 break; 2672 case ICmpInst::ICMP_SGT: 2673 case ICmpInst::ICMP_SGE: 2674 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2675 Q.CxtI, Q.DT); 2676 if (!KnownNonNegative) 2677 break; 2678 // fall-through 2679 case ICmpInst::ICMP_NE: 2680 case ICmpInst::ICMP_UGT: 2681 case ICmpInst::ICMP_UGE: 2682 return getTrue(ITy); 2683 case ICmpInst::ICMP_SLT: 2684 case ICmpInst::ICMP_SLE: 2685 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC, 2686 Q.CxtI, Q.DT); 2687 if (!KnownNonNegative) 2688 break; 2689 // fall-through 2690 case ICmpInst::ICMP_EQ: 2691 case ICmpInst::ICMP_ULT: 2692 case ICmpInst::ICMP_ULE: 2693 return getFalse(ITy); 2694 } 2695 } 2696 2697 // x udiv y <=u x. 2698 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) { 2699 // icmp pred (X /u Y), X 2700 if (Pred == ICmpInst::ICMP_UGT) 2701 return getFalse(ITy); 2702 if (Pred == ICmpInst::ICMP_ULE) 2703 return getTrue(ITy); 2704 } 2705 2706 // handle: 2707 // CI2 << X == CI 2708 // CI2 << X != CI 2709 // 2710 // where CI2 is a power of 2 and CI isn't 2711 if (auto *CI = dyn_cast<ConstantInt>(RHS)) { 2712 const APInt *CI2Val, *CIVal = &CI->getValue(); 2713 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) && 2714 CI2Val->isPowerOf2()) { 2715 if (!CIVal->isPowerOf2()) { 2716 // CI2 << X can equal zero in some circumstances, 2717 // this simplification is unsafe if CI is zero. 2718 // 2719 // We know it is safe if: 2720 // - The shift is nsw, we can't shift out the one bit. 2721 // - The shift is nuw, we can't shift out the one bit. 2722 // - CI2 is one 2723 // - CI isn't zero 2724 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() || 2725 *CI2Val == 1 || !CI->isZero()) { 2726 if (Pred == ICmpInst::ICMP_EQ) 2727 return ConstantInt::getFalse(RHS->getContext()); 2728 if (Pred == ICmpInst::ICMP_NE) 2729 return ConstantInt::getTrue(RHS->getContext()); 2730 } 2731 } 2732 if (CIVal->isSignBit() && *CI2Val == 1) { 2733 if (Pred == ICmpInst::ICMP_UGT) 2734 return ConstantInt::getFalse(RHS->getContext()); 2735 if (Pred == ICmpInst::ICMP_ULE) 2736 return ConstantInt::getTrue(RHS->getContext()); 2737 } 2738 } 2739 } 2740 2741 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() && 2742 LBO->getOperand(1) == RBO->getOperand(1)) { 2743 switch (LBO->getOpcode()) { 2744 default: break; 2745 case Instruction::UDiv: 2746 case Instruction::LShr: 2747 if (ICmpInst::isSigned(Pred)) 2748 break; 2749 // fall-through 2750 case Instruction::SDiv: 2751 case Instruction::AShr: 2752 if (!LBO->isExact() || !RBO->isExact()) 2753 break; 2754 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2755 RBO->getOperand(0), Q, MaxRecurse-1)) 2756 return V; 2757 break; 2758 case Instruction::Shl: { 2759 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap(); 2760 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap(); 2761 if (!NUW && !NSW) 2762 break; 2763 if (!NSW && ICmpInst::isSigned(Pred)) 2764 break; 2765 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2766 RBO->getOperand(0), Q, MaxRecurse-1)) 2767 return V; 2768 break; 2769 } 2770 } 2771 } 2772 2773 // Simplify comparisons involving max/min. 2774 Value *A, *B; 2775 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE; 2776 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B". 2777 2778 // Signed variants on "max(a,b)>=a -> true". 2779 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2780 if (A != RHS) std::swap(A, B); // smax(A, B) pred A. 2781 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2782 // We analyze this as smax(A, B) pred A. 2783 P = Pred; 2784 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) && 2785 (A == LHS || B == LHS)) { 2786 if (A != LHS) std::swap(A, B); // A pred smax(A, B). 2787 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2788 // We analyze this as smax(A, B) swapped-pred A. 2789 P = CmpInst::getSwappedPredicate(Pred); 2790 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2791 (A == RHS || B == RHS)) { 2792 if (A != RHS) std::swap(A, B); // smin(A, B) pred A. 2793 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2794 // We analyze this as smax(-A, -B) swapped-pred -A. 2795 // Note that we do not need to actually form -A or -B thanks to EqP. 2796 P = CmpInst::getSwappedPredicate(Pred); 2797 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) && 2798 (A == LHS || B == LHS)) { 2799 if (A != LHS) std::swap(A, B); // A pred smin(A, B). 2800 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2801 // We analyze this as smax(-A, -B) pred -A. 2802 // Note that we do not need to actually form -A or -B thanks to EqP. 2803 P = Pred; 2804 } 2805 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2806 // Cases correspond to "max(A, B) p A". 2807 switch (P) { 2808 default: 2809 break; 2810 case CmpInst::ICMP_EQ: 2811 case CmpInst::ICMP_SLE: 2812 // Equivalent to "A EqP B". This may be the same as the condition tested 2813 // in the max/min; if so, we can just return that. 2814 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2815 return V; 2816 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2817 return V; 2818 // Otherwise, see if "A EqP B" simplifies. 2819 if (MaxRecurse) 2820 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2821 return V; 2822 break; 2823 case CmpInst::ICMP_NE: 2824 case CmpInst::ICMP_SGT: { 2825 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2826 // Equivalent to "A InvEqP B". This may be the same as the condition 2827 // tested in the max/min; if so, we can just return that. 2828 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2829 return V; 2830 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2831 return V; 2832 // Otherwise, see if "A InvEqP B" simplifies. 2833 if (MaxRecurse) 2834 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2835 return V; 2836 break; 2837 } 2838 case CmpInst::ICMP_SGE: 2839 // Always true. 2840 return getTrue(ITy); 2841 case CmpInst::ICMP_SLT: 2842 // Always false. 2843 return getFalse(ITy); 2844 } 2845 } 2846 2847 // Unsigned variants on "max(a,b)>=a -> true". 2848 P = CmpInst::BAD_ICMP_PREDICATE; 2849 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2850 if (A != RHS) std::swap(A, B); // umax(A, B) pred A. 2851 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2852 // We analyze this as umax(A, B) pred A. 2853 P = Pred; 2854 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) && 2855 (A == LHS || B == LHS)) { 2856 if (A != LHS) std::swap(A, B); // A pred umax(A, B). 2857 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2858 // We analyze this as umax(A, B) swapped-pred A. 2859 P = CmpInst::getSwappedPredicate(Pred); 2860 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2861 (A == RHS || B == RHS)) { 2862 if (A != RHS) std::swap(A, B); // umin(A, B) pred A. 2863 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2864 // We analyze this as umax(-A, -B) swapped-pred -A. 2865 // Note that we do not need to actually form -A or -B thanks to EqP. 2866 P = CmpInst::getSwappedPredicate(Pred); 2867 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) && 2868 (A == LHS || B == LHS)) { 2869 if (A != LHS) std::swap(A, B); // A pred umin(A, B). 2870 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2871 // We analyze this as umax(-A, -B) pred -A. 2872 // Note that we do not need to actually form -A or -B thanks to EqP. 2873 P = Pred; 2874 } 2875 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2876 // Cases correspond to "max(A, B) p A". 2877 switch (P) { 2878 default: 2879 break; 2880 case CmpInst::ICMP_EQ: 2881 case CmpInst::ICMP_ULE: 2882 // Equivalent to "A EqP B". This may be the same as the condition tested 2883 // in the max/min; if so, we can just return that. 2884 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2885 return V; 2886 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2887 return V; 2888 // Otherwise, see if "A EqP B" simplifies. 2889 if (MaxRecurse) 2890 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2891 return V; 2892 break; 2893 case CmpInst::ICMP_NE: 2894 case CmpInst::ICMP_UGT: { 2895 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2896 // Equivalent to "A InvEqP B". This may be the same as the condition 2897 // tested in the max/min; if so, we can just return that. 2898 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2899 return V; 2900 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2901 return V; 2902 // Otherwise, see if "A InvEqP B" simplifies. 2903 if (MaxRecurse) 2904 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2905 return V; 2906 break; 2907 } 2908 case CmpInst::ICMP_UGE: 2909 // Always true. 2910 return getTrue(ITy); 2911 case CmpInst::ICMP_ULT: 2912 // Always false. 2913 return getFalse(ITy); 2914 } 2915 } 2916 2917 // Variants on "max(x,y) >= min(x,z)". 2918 Value *C, *D; 2919 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && 2920 match(RHS, m_SMin(m_Value(C), m_Value(D))) && 2921 (A == C || A == D || B == C || B == D)) { 2922 // max(x, ?) pred min(x, ?). 2923 if (Pred == CmpInst::ICMP_SGE) 2924 // Always true. 2925 return getTrue(ITy); 2926 if (Pred == CmpInst::ICMP_SLT) 2927 // Always false. 2928 return getFalse(ITy); 2929 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2930 match(RHS, m_SMax(m_Value(C), m_Value(D))) && 2931 (A == C || A == D || B == C || B == D)) { 2932 // min(x, ?) pred max(x, ?). 2933 if (Pred == CmpInst::ICMP_SLE) 2934 // Always true. 2935 return getTrue(ITy); 2936 if (Pred == CmpInst::ICMP_SGT) 2937 // Always false. 2938 return getFalse(ITy); 2939 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && 2940 match(RHS, m_UMin(m_Value(C), m_Value(D))) && 2941 (A == C || A == D || B == C || B == D)) { 2942 // max(x, ?) pred min(x, ?). 2943 if (Pred == CmpInst::ICMP_UGE) 2944 // Always true. 2945 return getTrue(ITy); 2946 if (Pred == CmpInst::ICMP_ULT) 2947 // Always false. 2948 return getFalse(ITy); 2949 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2950 match(RHS, m_UMax(m_Value(C), m_Value(D))) && 2951 (A == C || A == D || B == C || B == D)) { 2952 // min(x, ?) pred max(x, ?). 2953 if (Pred == CmpInst::ICMP_ULE) 2954 // Always true. 2955 return getTrue(ITy); 2956 if (Pred == CmpInst::ICMP_UGT) 2957 // Always false. 2958 return getFalse(ITy); 2959 } 2960 2961 // Simplify comparisons of related pointers using a powerful, recursive 2962 // GEP-walk when we have target data available.. 2963 if (LHS->getType()->isPointerTy()) 2964 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS)) 2965 return C; 2966 2967 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) { 2968 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) { 2969 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() && 2970 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() && 2971 (ICmpInst::isEquality(Pred) || 2972 (GLHS->isInBounds() && GRHS->isInBounds() && 2973 Pred == ICmpInst::getSignedPredicate(Pred)))) { 2974 // The bases are equal and the indices are constant. Build a constant 2975 // expression GEP with the same indices and a null base pointer to see 2976 // what constant folding can make out of it. 2977 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType()); 2978 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end()); 2979 Constant *NewLHS = ConstantExpr::getGetElementPtr( 2980 GLHS->getSourceElementType(), Null, IndicesLHS); 2981 2982 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end()); 2983 Constant *NewRHS = ConstantExpr::getGetElementPtr( 2984 GLHS->getSourceElementType(), Null, IndicesRHS); 2985 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS); 2986 } 2987 } 2988 } 2989 2990 // If a bit is known to be zero for A and known to be one for B, 2991 // then A and B cannot be equal. 2992 if (ICmpInst::isEquality(Pred)) { 2993 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2994 uint32_t BitWidth = CI->getBitWidth(); 2995 APInt LHSKnownZero(BitWidth, 0); 2996 APInt LHSKnownOne(BitWidth, 0); 2997 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC, 2998 Q.CxtI, Q.DT); 2999 const APInt &RHSVal = CI->getValue(); 3000 if (((LHSKnownZero & RHSVal) != 0) || ((LHSKnownOne & ~RHSVal) != 0)) 3001 return Pred == ICmpInst::ICMP_EQ 3002 ? ConstantInt::getFalse(CI->getContext()) 3003 : ConstantInt::getTrue(CI->getContext()); 3004 } 3005 } 3006 3007 // If the comparison is with the result of a select instruction, check whether 3008 // comparing with either branch of the select always yields the same value. 3009 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3010 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3011 return V; 3012 3013 // If the comparison is with the result of a phi instruction, check whether 3014 // doing the compare with each incoming phi value yields a common result. 3015 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3016 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3017 return V; 3018 3019 return nullptr; 3020 } 3021 3022 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3023 const DataLayout &DL, 3024 const TargetLibraryInfo *TLI, 3025 const DominatorTree *DT, AssumptionCache *AC, 3026 Instruction *CxtI) { 3027 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3028 RecursionLimit); 3029 } 3030 3031 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can 3032 /// fold the result. If not, this returns null. 3033 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3034 const Query &Q, unsigned MaxRecurse) { 3035 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 3036 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!"); 3037 3038 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 3039 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 3040 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); 3041 3042 // If we have a constant, make sure it is on the RHS. 3043 std::swap(LHS, RHS); 3044 Pred = CmpInst::getSwappedPredicate(Pred); 3045 } 3046 3047 // Fold trivial predicates. 3048 if (Pred == FCmpInst::FCMP_FALSE) 3049 return ConstantInt::get(GetCompareTy(LHS), 0); 3050 if (Pred == FCmpInst::FCMP_TRUE) 3051 return ConstantInt::get(GetCompareTy(LHS), 1); 3052 3053 // fcmp pred x, undef and fcmp pred undef, x 3054 // fold to true if unordered, false if ordered 3055 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) { 3056 // Choosing NaN for the undef will always make unordered comparison succeed 3057 // and ordered comparison fail. 3058 return ConstantInt::get(GetCompareTy(LHS), CmpInst::isUnordered(Pred)); 3059 } 3060 3061 // fcmp x,x -> true/false. Not all compares are foldable. 3062 if (LHS == RHS) { 3063 if (CmpInst::isTrueWhenEqual(Pred)) 3064 return ConstantInt::get(GetCompareTy(LHS), 1); 3065 if (CmpInst::isFalseWhenEqual(Pred)) 3066 return ConstantInt::get(GetCompareTy(LHS), 0); 3067 } 3068 3069 // Handle fcmp with constant RHS 3070 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) { 3071 // If the constant is a nan, see if we can fold the comparison based on it. 3072 if (CFP->getValueAPF().isNaN()) { 3073 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo" 3074 return ConstantInt::getFalse(CFP->getContext()); 3075 assert(FCmpInst::isUnordered(Pred) && 3076 "Comparison must be either ordered or unordered!"); 3077 // True if unordered. 3078 return ConstantInt::getTrue(CFP->getContext()); 3079 } 3080 // Check whether the constant is an infinity. 3081 if (CFP->getValueAPF().isInfinity()) { 3082 if (CFP->getValueAPF().isNegative()) { 3083 switch (Pred) { 3084 case FCmpInst::FCMP_OLT: 3085 // No value is ordered and less than negative infinity. 3086 return ConstantInt::getFalse(CFP->getContext()); 3087 case FCmpInst::FCMP_UGE: 3088 // All values are unordered with or at least negative infinity. 3089 return ConstantInt::getTrue(CFP->getContext()); 3090 default: 3091 break; 3092 } 3093 } else { 3094 switch (Pred) { 3095 case FCmpInst::FCMP_OGT: 3096 // No value is ordered and greater than infinity. 3097 return ConstantInt::getFalse(CFP->getContext()); 3098 case FCmpInst::FCMP_ULE: 3099 // All values are unordered with and at most infinity. 3100 return ConstantInt::getTrue(CFP->getContext()); 3101 default: 3102 break; 3103 } 3104 } 3105 } 3106 if (CFP->getValueAPF().isZero()) { 3107 switch (Pred) { 3108 case FCmpInst::FCMP_UGE: 3109 if (CannotBeOrderedLessThanZero(LHS)) 3110 return ConstantInt::getTrue(CFP->getContext()); 3111 break; 3112 case FCmpInst::FCMP_OLT: 3113 // X < 0 3114 if (CannotBeOrderedLessThanZero(LHS)) 3115 return ConstantInt::getFalse(CFP->getContext()); 3116 break; 3117 default: 3118 break; 3119 } 3120 } 3121 } 3122 3123 // If the comparison is with the result of a select instruction, check whether 3124 // comparing with either branch of the select always yields the same value. 3125 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3126 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 3127 return V; 3128 3129 // If the comparison is with the result of a phi instruction, check whether 3130 // doing the compare with each incoming phi value yields a common result. 3131 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3132 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 3133 return V; 3134 3135 return nullptr; 3136 } 3137 3138 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3139 const DataLayout &DL, 3140 const TargetLibraryInfo *TLI, 3141 const DominatorTree *DT, AssumptionCache *AC, 3142 const Instruction *CxtI) { 3143 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3144 RecursionLimit); 3145 } 3146 3147 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold 3148 /// the result. If not, this returns null. 3149 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal, 3150 Value *FalseVal, const Query &Q, 3151 unsigned MaxRecurse) { 3152 // select true, X, Y -> X 3153 // select false, X, Y -> Y 3154 if (Constant *CB = dyn_cast<Constant>(CondVal)) { 3155 if (CB->isAllOnesValue()) 3156 return TrueVal; 3157 if (CB->isNullValue()) 3158 return FalseVal; 3159 } 3160 3161 // select C, X, X -> X 3162 if (TrueVal == FalseVal) 3163 return TrueVal; 3164 3165 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y 3166 if (isa<Constant>(TrueVal)) 3167 return TrueVal; 3168 return FalseVal; 3169 } 3170 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X 3171 return FalseVal; 3172 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X 3173 return TrueVal; 3174 3175 const auto *ICI = dyn_cast<ICmpInst>(CondVal); 3176 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); 3177 if (ICI && BitWidth) { 3178 ICmpInst::Predicate Pred = ICI->getPredicate(); 3179 APInt MinSignedValue = APInt::getSignBit(BitWidth); 3180 Value *X; 3181 const APInt *Y; 3182 bool TrueWhenUnset; 3183 bool IsBitTest = false; 3184 if (ICmpInst::isEquality(Pred) && 3185 match(ICI->getOperand(0), m_And(m_Value(X), m_APInt(Y))) && 3186 match(ICI->getOperand(1), m_Zero())) { 3187 IsBitTest = true; 3188 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 3189 } else if (Pred == ICmpInst::ICMP_SLT && 3190 match(ICI->getOperand(1), m_Zero())) { 3191 X = ICI->getOperand(0); 3192 Y = &MinSignedValue; 3193 IsBitTest = true; 3194 TrueWhenUnset = false; 3195 } else if (Pred == ICmpInst::ICMP_SGT && 3196 match(ICI->getOperand(1), m_AllOnes())) { 3197 X = ICI->getOperand(0); 3198 Y = &MinSignedValue; 3199 IsBitTest = true; 3200 TrueWhenUnset = true; 3201 } 3202 if (IsBitTest) { 3203 const APInt *C; 3204 // (X & Y) == 0 ? X & ~Y : X --> X 3205 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y 3206 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) && 3207 *Y == ~*C) 3208 return TrueWhenUnset ? FalseVal : TrueVal; 3209 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y 3210 // (X & Y) != 0 ? X : X & ~Y --> X 3211 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) && 3212 *Y == ~*C) 3213 return TrueWhenUnset ? FalseVal : TrueVal; 3214 3215 if (Y->isPowerOf2()) { 3216 // (X & Y) == 0 ? X | Y : X --> X | Y 3217 // (X & Y) != 0 ? X | Y : X --> X 3218 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) && 3219 *Y == *C) 3220 return TrueWhenUnset ? TrueVal : FalseVal; 3221 // (X & Y) == 0 ? X : X | Y --> X 3222 // (X & Y) != 0 ? X : X | Y --> X | Y 3223 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) && 3224 *Y == *C) 3225 return TrueWhenUnset ? TrueVal : FalseVal; 3226 } 3227 } 3228 } 3229 3230 return nullptr; 3231 } 3232 3233 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, 3234 const DataLayout &DL, 3235 const TargetLibraryInfo *TLI, 3236 const DominatorTree *DT, AssumptionCache *AC, 3237 const Instruction *CxtI) { 3238 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, 3239 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3240 } 3241 3242 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can 3243 /// fold the result. If not, this returns null. 3244 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, 3245 const Query &Q, unsigned) { 3246 // The type of the GEP pointer operand. 3247 unsigned AS = 3248 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace(); 3249 3250 // getelementptr P -> P. 3251 if (Ops.size() == 1) 3252 return Ops[0]; 3253 3254 // Compute the (pointer) type returned by the GEP instruction. 3255 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1)); 3256 Type *GEPTy = PointerType::get(LastType, AS); 3257 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType())) 3258 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 3259 3260 if (isa<UndefValue>(Ops[0])) 3261 return UndefValue::get(GEPTy); 3262 3263 if (Ops.size() == 2) { 3264 // getelementptr P, 0 -> P. 3265 if (match(Ops[1], m_Zero())) 3266 return Ops[0]; 3267 3268 Type *Ty = SrcTy; 3269 if (Ty->isSized()) { 3270 Value *P; 3271 uint64_t C; 3272 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty); 3273 // getelementptr P, N -> P if P points to a type of zero size. 3274 if (TyAllocSize == 0) 3275 return Ops[0]; 3276 3277 // The following transforms are only safe if the ptrtoint cast 3278 // doesn't truncate the pointers. 3279 if (Ops[1]->getType()->getScalarSizeInBits() == 3280 Q.DL.getPointerSizeInBits(AS)) { 3281 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * { 3282 if (match(P, m_Zero())) 3283 return Constant::getNullValue(GEPTy); 3284 Value *Temp; 3285 if (match(P, m_PtrToInt(m_Value(Temp)))) 3286 if (Temp->getType() == GEPTy) 3287 return Temp; 3288 return nullptr; 3289 }; 3290 3291 // getelementptr V, (sub P, V) -> P if P points to a type of size 1. 3292 if (TyAllocSize == 1 && 3293 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))))) 3294 if (Value *R = PtrToIntOrZero(P)) 3295 return R; 3296 3297 // getelementptr V, (ashr (sub P, V), C) -> Q 3298 // if P points to a type of size 1 << C. 3299 if (match(Ops[1], 3300 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3301 m_ConstantInt(C))) && 3302 TyAllocSize == 1ULL << C) 3303 if (Value *R = PtrToIntOrZero(P)) 3304 return R; 3305 3306 // getelementptr V, (sdiv (sub P, V), C) -> Q 3307 // if P points to a type of size C. 3308 if (match(Ops[1], 3309 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), 3310 m_SpecificInt(TyAllocSize)))) 3311 if (Value *R = PtrToIntOrZero(P)) 3312 return R; 3313 } 3314 } 3315 } 3316 3317 // Check to see if this is constant foldable. 3318 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 3319 if (!isa<Constant>(Ops[i])) 3320 return nullptr; 3321 3322 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]), 3323 Ops.slice(1)); 3324 } 3325 3326 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout &DL, 3327 const TargetLibraryInfo *TLI, 3328 const DominatorTree *DT, AssumptionCache *AC, 3329 const Instruction *CxtI) { 3330 return ::SimplifyGEPInst( 3331 cast<PointerType>(Ops[0]->getType()->getScalarType())->getElementType(), 3332 Ops, Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3333 } 3334 3335 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we 3336 /// can fold the result. If not, this returns null. 3337 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val, 3338 ArrayRef<unsigned> Idxs, const Query &Q, 3339 unsigned) { 3340 if (Constant *CAgg = dyn_cast<Constant>(Agg)) 3341 if (Constant *CVal = dyn_cast<Constant>(Val)) 3342 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs); 3343 3344 // insertvalue x, undef, n -> x 3345 if (match(Val, m_Undef())) 3346 return Agg; 3347 3348 // insertvalue x, (extractvalue y, n), n 3349 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val)) 3350 if (EV->getAggregateOperand()->getType() == Agg->getType() && 3351 EV->getIndices() == Idxs) { 3352 // insertvalue undef, (extractvalue y, n), n -> y 3353 if (match(Agg, m_Undef())) 3354 return EV->getAggregateOperand(); 3355 3356 // insertvalue y, (extractvalue y, n), n -> y 3357 if (Agg == EV->getAggregateOperand()) 3358 return Agg; 3359 } 3360 3361 return nullptr; 3362 } 3363 3364 Value *llvm::SimplifyInsertValueInst( 3365 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout &DL, 3366 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC, 3367 const Instruction *CxtI) { 3368 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI), 3369 RecursionLimit); 3370 } 3371 3372 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null. 3373 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) { 3374 // If all of the PHI's incoming values are the same then replace the PHI node 3375 // with the common value. 3376 Value *CommonValue = nullptr; 3377 bool HasUndefInput = false; 3378 for (Value *Incoming : PN->incoming_values()) { 3379 // If the incoming value is the phi node itself, it can safely be skipped. 3380 if (Incoming == PN) continue; 3381 if (isa<UndefValue>(Incoming)) { 3382 // Remember that we saw an undef value, but otherwise ignore them. 3383 HasUndefInput = true; 3384 continue; 3385 } 3386 if (CommonValue && Incoming != CommonValue) 3387 return nullptr; // Not the same, bail out. 3388 CommonValue = Incoming; 3389 } 3390 3391 // If CommonValue is null then all of the incoming values were either undef or 3392 // equal to the phi node itself. 3393 if (!CommonValue) 3394 return UndefValue::get(PN->getType()); 3395 3396 // If we have a PHI node like phi(X, undef, X), where X is defined by some 3397 // instruction, we cannot return X as the result of the PHI node unless it 3398 // dominates the PHI block. 3399 if (HasUndefInput) 3400 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr; 3401 3402 return CommonValue; 3403 } 3404 3405 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) { 3406 if (Constant *C = dyn_cast<Constant>(Op)) 3407 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI); 3408 3409 return nullptr; 3410 } 3411 3412 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout &DL, 3413 const TargetLibraryInfo *TLI, 3414 const DominatorTree *DT, AssumptionCache *AC, 3415 const Instruction *CxtI) { 3416 return ::SimplifyTruncInst(Op, Ty, Query(DL, TLI, DT, AC, CxtI), 3417 RecursionLimit); 3418 } 3419 3420 //=== Helper functions for higher up the class hierarchy. 3421 3422 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can 3423 /// fold the result. If not, this returns null. 3424 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3425 const Query &Q, unsigned MaxRecurse) { 3426 switch (Opcode) { 3427 case Instruction::Add: 3428 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3429 Q, MaxRecurse); 3430 case Instruction::FAdd: 3431 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3432 3433 case Instruction::Sub: 3434 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3435 Q, MaxRecurse); 3436 case Instruction::FSub: 3437 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3438 3439 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse); 3440 case Instruction::FMul: 3441 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3442 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse); 3443 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse); 3444 case Instruction::FDiv: 3445 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3446 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse); 3447 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse); 3448 case Instruction::FRem: 3449 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 3450 case Instruction::Shl: 3451 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 3452 Q, MaxRecurse); 3453 case Instruction::LShr: 3454 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3455 case Instruction::AShr: 3456 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 3457 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse); 3458 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse); 3459 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse); 3460 default: 3461 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 3462 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 3463 Constant *COps[] = {CLHS, CRHS}; 3464 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL, 3465 Q.TLI); 3466 } 3467 3468 // If the operation is associative, try some generic simplifications. 3469 if (Instruction::isAssociative(Opcode)) 3470 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse)) 3471 return V; 3472 3473 // If the operation is with the result of a select instruction check whether 3474 // operating on either branch of the select always yields the same value. 3475 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 3476 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse)) 3477 return V; 3478 3479 // If the operation is with the result of a phi instruction, check whether 3480 // operating on all incoming values of the phi always yields the same value. 3481 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 3482 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse)) 3483 return V; 3484 3485 return nullptr; 3486 } 3487 } 3488 3489 /// SimplifyFPBinOp - Given operands for a BinaryOperator, see if we can 3490 /// fold the result. If not, this returns null. 3491 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the 3492 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp. 3493 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3494 const FastMathFlags &FMF, const Query &Q, 3495 unsigned MaxRecurse) { 3496 switch (Opcode) { 3497 case Instruction::FAdd: 3498 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse); 3499 case Instruction::FSub: 3500 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse); 3501 case Instruction::FMul: 3502 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse); 3503 default: 3504 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse); 3505 } 3506 } 3507 3508 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3509 const DataLayout &DL, const TargetLibraryInfo *TLI, 3510 const DominatorTree *DT, AssumptionCache *AC, 3511 const Instruction *CxtI) { 3512 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3513 RecursionLimit); 3514 } 3515 3516 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, 3517 const FastMathFlags &FMF, const DataLayout &DL, 3518 const TargetLibraryInfo *TLI, 3519 const DominatorTree *DT, AssumptionCache *AC, 3520 const Instruction *CxtI) { 3521 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI), 3522 RecursionLimit); 3523 } 3524 3525 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can 3526 /// fold the result. 3527 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3528 const Query &Q, unsigned MaxRecurse) { 3529 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate)) 3530 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 3531 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 3532 } 3533 3534 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 3535 const DataLayout &DL, const TargetLibraryInfo *TLI, 3536 const DominatorTree *DT, AssumptionCache *AC, 3537 const Instruction *CxtI) { 3538 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI), 3539 RecursionLimit); 3540 } 3541 3542 static bool IsIdempotent(Intrinsic::ID ID) { 3543 switch (ID) { 3544 default: return false; 3545 3546 // Unary idempotent: f(f(x)) = f(x) 3547 case Intrinsic::fabs: 3548 case Intrinsic::floor: 3549 case Intrinsic::ceil: 3550 case Intrinsic::trunc: 3551 case Intrinsic::rint: 3552 case Intrinsic::nearbyint: 3553 case Intrinsic::round: 3554 return true; 3555 } 3556 } 3557 3558 template <typename IterTy> 3559 static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd, 3560 const Query &Q, unsigned MaxRecurse) { 3561 // Perform idempotent optimizations 3562 if (!IsIdempotent(IID)) 3563 return nullptr; 3564 3565 // Unary Ops 3566 if (std::distance(ArgBegin, ArgEnd) == 1) 3567 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) 3568 if (II->getIntrinsicID() == IID) 3569 return II; 3570 3571 return nullptr; 3572 } 3573 3574 template <typename IterTy> 3575 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd, 3576 const Query &Q, unsigned MaxRecurse) { 3577 Type *Ty = V->getType(); 3578 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 3579 Ty = PTy->getElementType(); 3580 FunctionType *FTy = cast<FunctionType>(Ty); 3581 3582 // call undef -> undef 3583 if (isa<UndefValue>(V)) 3584 return UndefValue::get(FTy->getReturnType()); 3585 3586 Function *F = dyn_cast<Function>(V); 3587 if (!F) 3588 return nullptr; 3589 3590 if (unsigned IID = F->getIntrinsicID()) 3591 if (Value *Ret = 3592 SimplifyIntrinsic((Intrinsic::ID) IID, ArgBegin, ArgEnd, Q, MaxRecurse)) 3593 return Ret; 3594 3595 if (!canConstantFoldCallTo(F)) 3596 return nullptr; 3597 3598 SmallVector<Constant *, 4> ConstantArgs; 3599 ConstantArgs.reserve(ArgEnd - ArgBegin); 3600 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) { 3601 Constant *C = dyn_cast<Constant>(*I); 3602 if (!C) 3603 return nullptr; 3604 ConstantArgs.push_back(C); 3605 } 3606 3607 return ConstantFoldCall(F, ConstantArgs, Q.TLI); 3608 } 3609 3610 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin, 3611 User::op_iterator ArgEnd, const DataLayout &DL, 3612 const TargetLibraryInfo *TLI, const DominatorTree *DT, 3613 AssumptionCache *AC, const Instruction *CxtI) { 3614 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI), 3615 RecursionLimit); 3616 } 3617 3618 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args, 3619 const DataLayout &DL, const TargetLibraryInfo *TLI, 3620 const DominatorTree *DT, AssumptionCache *AC, 3621 const Instruction *CxtI) { 3622 return ::SimplifyCall(V, Args.begin(), Args.end(), 3623 Query(DL, TLI, DT, AC, CxtI), RecursionLimit); 3624 } 3625 3626 /// SimplifyInstruction - See if we can compute a simplified version of this 3627 /// instruction. If not, this returns null. 3628 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout &DL, 3629 const TargetLibraryInfo *TLI, 3630 const DominatorTree *DT, AssumptionCache *AC) { 3631 Value *Result; 3632 3633 switch (I->getOpcode()) { 3634 default: 3635 Result = ConstantFoldInstruction(I, DL, TLI); 3636 break; 3637 case Instruction::FAdd: 3638 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1), 3639 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3640 break; 3641 case Instruction::Add: 3642 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1), 3643 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3644 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3645 TLI, DT, AC, I); 3646 break; 3647 case Instruction::FSub: 3648 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1), 3649 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3650 break; 3651 case Instruction::Sub: 3652 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1), 3653 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3654 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3655 TLI, DT, AC, I); 3656 break; 3657 case Instruction::FMul: 3658 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1), 3659 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3660 break; 3661 case Instruction::Mul: 3662 Result = 3663 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3664 break; 3665 case Instruction::SDiv: 3666 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3667 AC, I); 3668 break; 3669 case Instruction::UDiv: 3670 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3671 AC, I); 3672 break; 3673 case Instruction::FDiv: 3674 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), 3675 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3676 break; 3677 case Instruction::SRem: 3678 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3679 AC, I); 3680 break; 3681 case Instruction::URem: 3682 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, 3683 AC, I); 3684 break; 3685 case Instruction::FRem: 3686 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), 3687 I->getFastMathFlags(), DL, TLI, DT, AC, I); 3688 break; 3689 case Instruction::Shl: 3690 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1), 3691 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3692 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL, 3693 TLI, DT, AC, I); 3694 break; 3695 case Instruction::LShr: 3696 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1), 3697 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 3698 AC, I); 3699 break; 3700 case Instruction::AShr: 3701 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1), 3702 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT, 3703 AC, I); 3704 break; 3705 case Instruction::And: 3706 Result = 3707 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3708 break; 3709 case Instruction::Or: 3710 Result = 3711 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3712 break; 3713 case Instruction::Xor: 3714 Result = 3715 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I); 3716 break; 3717 case Instruction::ICmp: 3718 Result = 3719 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0), 3720 I->getOperand(1), DL, TLI, DT, AC, I); 3721 break; 3722 case Instruction::FCmp: 3723 Result = 3724 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0), 3725 I->getOperand(1), DL, TLI, DT, AC, I); 3726 break; 3727 case Instruction::Select: 3728 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1), 3729 I->getOperand(2), DL, TLI, DT, AC, I); 3730 break; 3731 case Instruction::GetElementPtr: { 3732 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 3733 Result = SimplifyGEPInst(Ops, DL, TLI, DT, AC, I); 3734 break; 3735 } 3736 case Instruction::InsertValue: { 3737 InsertValueInst *IV = cast<InsertValueInst>(I); 3738 Result = SimplifyInsertValueInst(IV->getAggregateOperand(), 3739 IV->getInsertedValueOperand(), 3740 IV->getIndices(), DL, TLI, DT, AC, I); 3741 break; 3742 } 3743 case Instruction::PHI: 3744 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I)); 3745 break; 3746 case Instruction::Call: { 3747 CallSite CS(cast<CallInst>(I)); 3748 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL, 3749 TLI, DT, AC, I); 3750 break; 3751 } 3752 case Instruction::Trunc: 3753 Result = 3754 SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT, AC, I); 3755 break; 3756 } 3757 3758 /// If called on unreachable code, the above logic may report that the 3759 /// instruction simplified to itself. Make life easier for users by 3760 /// detecting that case here, returning a safe value instead. 3761 return Result == I ? UndefValue::get(I->getType()) : Result; 3762 } 3763 3764 /// \brief Implementation of recursive simplification through an instructions 3765 /// uses. 3766 /// 3767 /// This is the common implementation of the recursive simplification routines. 3768 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to 3769 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of 3770 /// instructions to process and attempt to simplify it using 3771 /// InstructionSimplify. 3772 /// 3773 /// This routine returns 'true' only when *it* simplifies something. The passed 3774 /// in simplified value does not count toward this. 3775 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV, 3776 const TargetLibraryInfo *TLI, 3777 const DominatorTree *DT, 3778 AssumptionCache *AC) { 3779 bool Simplified = false; 3780 SmallSetVector<Instruction *, 8> Worklist; 3781 const DataLayout &DL = I->getModule()->getDataLayout(); 3782 3783 // If we have an explicit value to collapse to, do that round of the 3784 // simplification loop by hand initially. 3785 if (SimpleV) { 3786 for (User *U : I->users()) 3787 if (U != I) 3788 Worklist.insert(cast<Instruction>(U)); 3789 3790 // Replace the instruction with its simplified value. 3791 I->replaceAllUsesWith(SimpleV); 3792 3793 // Gracefully handle edge cases where the instruction is not wired into any 3794 // parent block. 3795 if (I->getParent()) 3796 I->eraseFromParent(); 3797 } else { 3798 Worklist.insert(I); 3799 } 3800 3801 // Note that we must test the size on each iteration, the worklist can grow. 3802 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { 3803 I = Worklist[Idx]; 3804 3805 // See if this instruction simplifies. 3806 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC); 3807 if (!SimpleV) 3808 continue; 3809 3810 Simplified = true; 3811 3812 // Stash away all the uses of the old instruction so we can check them for 3813 // recursive simplifications after a RAUW. This is cheaper than checking all 3814 // uses of To on the recursive step in most cases. 3815 for (User *U : I->users()) 3816 Worklist.insert(cast<Instruction>(U)); 3817 3818 // Replace the instruction with its simplified value. 3819 I->replaceAllUsesWith(SimpleV); 3820 3821 // Gracefully handle edge cases where the instruction is not wired into any 3822 // parent block. 3823 if (I->getParent()) 3824 I->eraseFromParent(); 3825 } 3826 return Simplified; 3827 } 3828 3829 bool llvm::recursivelySimplifyInstruction(Instruction *I, 3830 const TargetLibraryInfo *TLI, 3831 const DominatorTree *DT, 3832 AssumptionCache *AC) { 3833 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC); 3834 } 3835 3836 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV, 3837 const TargetLibraryInfo *TLI, 3838 const DominatorTree *DT, 3839 AssumptionCache *AC) { 3840 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!"); 3841 assert(SimpleV && "Must provide a simplified value."); 3842 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC); 3843 } 3844