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