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