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