InstructionSimplify.cpp revision 263508
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" 35using namespace llvm; 36using namespace llvm::PatternMatch; 37 38enum { RecursionLimit = 3 }; 39 40STATISTIC(NumExpand, "Number of expansions"); 41STATISTIC(NumFactor , "Number of factorizations"); 42STATISTIC(NumReassoc, "Number of reassociations"); 43 44struct 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 53static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned); 54static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &, 55 unsigned); 56static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &, 57 unsigned); 58static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned); 59static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned); 60static 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. 64static 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. 72static 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"? 79static 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? 93static 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. 128static 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. 189static 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. 259static 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. 358static 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. 430static 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. 512static 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. 555static 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. 592static 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 653Value *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. 670static Constant *stripAndComputeConstantOffsets(const DataLayout *TD, 671 Value *&V, 672 bool AllowNonInbounds = false) { 673 assert(V->getType()->getScalarType()->isPointerTy()); 674 675 // Without DataLayout, just be conservative for now. Theoretically, more could 676 // be done in this case. 677 if (!TD) 678 return ConstantInt::get(IntegerType::get(V->getContext(), 64), 0); 679 680 Type *IntPtrTy = TD->getIntPtrType(V->getType())->getScalarType(); 681 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth()); 682 683 // Even though we don't look through PHI nodes, we could be called on an 684 // instruction in an unreachable block, which may be on a cycle. 685 SmallPtrSet<Value *, 4> Visited; 686 Visited.insert(V); 687 do { 688 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 689 if ((!AllowNonInbounds && !GEP->isInBounds()) || 690 !GEP->accumulateConstantOffset(*TD, Offset)) 691 break; 692 V = GEP->getPointerOperand(); 693 } else if (Operator::getOpcode(V) == Instruction::BitCast) { 694 V = cast<Operator>(V)->getOperand(0); 695 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 696 if (GA->mayBeOverridden()) 697 break; 698 V = GA->getAliasee(); 699 } else { 700 break; 701 } 702 assert(V->getType()->getScalarType()->isPointerTy() && 703 "Unexpected operand type!"); 704 } while (Visited.insert(V)); 705 706 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset); 707 if (V->getType()->isVectorTy()) 708 return ConstantVector::getSplat(V->getType()->getVectorNumElements(), 709 OffsetIntPtr); 710 return OffsetIntPtr; 711} 712 713/// \brief Compute the constant difference between two pointer values. 714/// If the difference is not a constant, returns zero. 715static Constant *computePointerDifference(const DataLayout *TD, 716 Value *LHS, Value *RHS) { 717 Constant *LHSOffset = stripAndComputeConstantOffsets(TD, LHS); 718 Constant *RHSOffset = stripAndComputeConstantOffsets(TD, RHS); 719 720 // If LHS and RHS are not related via constant offsets to the same base 721 // value, there is nothing we can do here. 722 if (LHS != RHS) 723 return 0; 724 725 // Otherwise, the difference of LHS - RHS can be computed as: 726 // LHS - RHS 727 // = (LHSOffset + Base) - (RHSOffset + Base) 728 // = LHSOffset - RHSOffset 729 return ConstantExpr::getSub(LHSOffset, RHSOffset); 730} 731 732/// SimplifySubInst - Given operands for a Sub, see if we can 733/// fold the result. If not, this returns null. 734static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 735 const Query &Q, unsigned MaxRecurse) { 736 if (Constant *CLHS = dyn_cast<Constant>(Op0)) 737 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 738 Constant *Ops[] = { CLHS, CRHS }; 739 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(), 740 Ops, Q.TD, Q.TLI); 741 } 742 743 // X - undef -> undef 744 // undef - X -> undef 745 if (match(Op0, m_Undef()) || match(Op1, m_Undef())) 746 return UndefValue::get(Op0->getType()); 747 748 // X - 0 -> X 749 if (match(Op1, m_Zero())) 750 return Op0; 751 752 // X - X -> 0 753 if (Op0 == Op1) 754 return Constant::getNullValue(Op0->getType()); 755 756 // (X*2) - X -> X 757 // (X<<1) - X -> X 758 Value *X = 0; 759 if (match(Op0, m_Mul(m_Specific(Op1), m_ConstantInt<2>())) || 760 match(Op0, m_Shl(m_Specific(Op1), m_One()))) 761 return Op1; 762 763 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies. 764 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X 765 Value *Y = 0, *Z = Op1; 766 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z 767 // See if "V === Y - Z" simplifies. 768 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1)) 769 // It does! Now see if "X + V" simplifies. 770 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) { 771 // It does, we successfully reassociated! 772 ++NumReassoc; 773 return W; 774 } 775 // See if "V === X - Z" simplifies. 776 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 777 // It does! Now see if "Y + V" simplifies. 778 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) { 779 // It does, we successfully reassociated! 780 ++NumReassoc; 781 return W; 782 } 783 } 784 785 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies. 786 // For example, X - (X + 1) -> -1 787 X = Op0; 788 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z) 789 // See if "V === X - Y" simplifies. 790 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 791 // It does! Now see if "V - Z" simplifies. 792 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) { 793 // It does, we successfully reassociated! 794 ++NumReassoc; 795 return W; 796 } 797 // See if "V === X - Z" simplifies. 798 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1)) 799 // It does! Now see if "V - Y" simplifies. 800 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) { 801 // It does, we successfully reassociated! 802 ++NumReassoc; 803 return W; 804 } 805 } 806 807 // Z - (X - Y) -> (Z - X) + Y if everything simplifies. 808 // For example, X - (X - Y) -> Y. 809 Z = Op0; 810 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y) 811 // See if "V === Z - X" simplifies. 812 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1)) 813 // It does! Now see if "V + Y" simplifies. 814 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) { 815 // It does, we successfully reassociated! 816 ++NumReassoc; 817 return W; 818 } 819 820 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies. 821 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) && 822 match(Op1, m_Trunc(m_Value(Y)))) 823 if (X->getType() == Y->getType()) 824 // See if "V === X - Y" simplifies. 825 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1)) 826 // It does! Now see if "trunc V" simplifies. 827 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1)) 828 // It does, return the simplified "trunc V". 829 return W; 830 831 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...). 832 if (match(Op0, m_PtrToInt(m_Value(X))) && 833 match(Op1, m_PtrToInt(m_Value(Y)))) 834 if (Constant *Result = computePointerDifference(Q.TD, X, Y)) 835 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true); 836 837 // Mul distributes over Sub. Try some generic simplifications based on this. 838 if (Value *V = FactorizeBinOp(Instruction::Sub, Op0, Op1, Instruction::Mul, 839 Q, MaxRecurse)) 840 return V; 841 842 // i1 sub -> xor. 843 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 844 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1)) 845 return V; 846 847 // Threading Sub over selects and phi nodes is pointless, so don't bother. 848 // Threading over the select in "A - select(cond, B, C)" means evaluating 849 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and 850 // only if B and C are equal. If B and C are equal then (since we assume 851 // that operands have already been simplified) "select(cond, B, C)" should 852 // have been simplified to the common value of B and C already. Analysing 853 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly 854 // for threading over phi nodes. 855 856 return 0; 857} 858 859Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 860 const DataLayout *TD, const TargetLibraryInfo *TLI, 861 const DominatorTree *DT) { 862 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query (TD, TLI, DT), 863 RecursionLimit); 864} 865 866/// Given operands for an FAdd, see if we can fold the result. If not, this 867/// returns null. 868static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 869 const Query &Q, unsigned MaxRecurse) { 870 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 871 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 872 Constant *Ops[] = { CLHS, CRHS }; 873 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(), 874 Ops, Q.TD, Q.TLI); 875 } 876 877 // Canonicalize the constant to the RHS. 878 std::swap(Op0, Op1); 879 } 880 881 // fadd X, -0 ==> X 882 if (match(Op1, m_NegZero())) 883 return Op0; 884 885 // fadd X, 0 ==> X, when we know X is not -0 886 if (match(Op1, m_Zero()) && 887 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0))) 888 return Op0; 889 890 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0 891 // where nnan and ninf have to occur at least once somewhere in this 892 // expression 893 Value *SubOp = 0; 894 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0)))) 895 SubOp = Op1; 896 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1)))) 897 SubOp = Op0; 898 if (SubOp) { 899 Instruction *FSub = cast<Instruction>(SubOp); 900 if ((FMF.noNaNs() || FSub->hasNoNaNs()) && 901 (FMF.noInfs() || FSub->hasNoInfs())) 902 return Constant::getNullValue(Op0->getType()); 903 } 904 905 return 0; 906} 907 908/// Given operands for an FSub, see if we can fold the result. If not, this 909/// returns null. 910static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 911 const Query &Q, unsigned MaxRecurse) { 912 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 913 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 914 Constant *Ops[] = { CLHS, CRHS }; 915 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(), 916 Ops, Q.TD, Q.TLI); 917 } 918 } 919 920 // fsub X, 0 ==> X 921 if (match(Op1, m_Zero())) 922 return Op0; 923 924 // fsub X, -0 ==> X, when we know X is not -0 925 if (match(Op1, m_NegZero()) && 926 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0))) 927 return Op0; 928 929 // fsub 0, (fsub -0.0, X) ==> X 930 Value *X; 931 if (match(Op0, m_AnyZero())) { 932 if (match(Op1, m_FSub(m_NegZero(), m_Value(X)))) 933 return X; 934 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X)))) 935 return X; 936 } 937 938 // fsub nnan ninf x, x ==> 0.0 939 if (FMF.noNaNs() && FMF.noInfs() && Op0 == Op1) 940 return Constant::getNullValue(Op0->getType()); 941 942 return 0; 943} 944 945/// Given the operands for an FMul, see if we can fold the result 946static Value *SimplifyFMulInst(Value *Op0, Value *Op1, 947 FastMathFlags FMF, 948 const Query &Q, 949 unsigned MaxRecurse) { 950 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 951 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 952 Constant *Ops[] = { CLHS, CRHS }; 953 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(), 954 Ops, Q.TD, Q.TLI); 955 } 956 957 // Canonicalize the constant to the RHS. 958 std::swap(Op0, Op1); 959 } 960 961 // fmul X, 1.0 ==> X 962 if (match(Op1, m_FPOne())) 963 return Op0; 964 965 // fmul nnan nsz X, 0 ==> 0 966 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero())) 967 return Op1; 968 969 return 0; 970} 971 972/// SimplifyMulInst - Given operands for a Mul, see if we can 973/// fold the result. If not, this returns null. 974static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q, 975 unsigned MaxRecurse) { 976 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 977 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 978 Constant *Ops[] = { CLHS, CRHS }; 979 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(), 980 Ops, Q.TD, Q.TLI); 981 } 982 983 // Canonicalize the constant to the RHS. 984 std::swap(Op0, Op1); 985 } 986 987 // X * undef -> 0 988 if (match(Op1, m_Undef())) 989 return Constant::getNullValue(Op0->getType()); 990 991 // X * 0 -> 0 992 if (match(Op1, m_Zero())) 993 return Op1; 994 995 // X * 1 -> X 996 if (match(Op1, m_One())) 997 return Op0; 998 999 // (X / Y) * Y -> X if the division is exact. 1000 Value *X = 0; 1001 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y 1002 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y) 1003 return X; 1004 1005 // i1 mul -> and. 1006 if (MaxRecurse && Op0->getType()->isIntegerTy(1)) 1007 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1)) 1008 return V; 1009 1010 // Try some generic simplifications for associative operations. 1011 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q, 1012 MaxRecurse)) 1013 return V; 1014 1015 // Mul distributes over Add. Try some generic simplifications based on this. 1016 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add, 1017 Q, MaxRecurse)) 1018 return V; 1019 1020 // If the operation is with the result of a select instruction, check whether 1021 // operating on either branch of the select always yields the same value. 1022 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1023 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q, 1024 MaxRecurse)) 1025 return V; 1026 1027 // If the operation is with the result of a phi instruction, check whether 1028 // operating on all incoming values of the phi always yields the same value. 1029 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1030 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q, 1031 MaxRecurse)) 1032 return V; 1033 1034 return 0; 1035} 1036 1037Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1038 const DataLayout *TD, const TargetLibraryInfo *TLI, 1039 const DominatorTree *DT) { 1040 return ::SimplifyFAddInst(Op0, Op1, FMF, Query (TD, TLI, DT), RecursionLimit); 1041} 1042 1043Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, 1044 const DataLayout *TD, const TargetLibraryInfo *TLI, 1045 const DominatorTree *DT) { 1046 return ::SimplifyFSubInst(Op0, Op1, FMF, Query (TD, TLI, DT), RecursionLimit); 1047} 1048 1049Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, 1050 FastMathFlags FMF, 1051 const DataLayout *TD, 1052 const TargetLibraryInfo *TLI, 1053 const DominatorTree *DT) { 1054 return ::SimplifyFMulInst(Op0, Op1, FMF, Query (TD, TLI, DT), RecursionLimit); 1055} 1056 1057Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout *TD, 1058 const TargetLibraryInfo *TLI, 1059 const DominatorTree *DT) { 1060 return ::SimplifyMulInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1061} 1062 1063/// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can 1064/// fold the result. If not, this returns null. 1065static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1066 const Query &Q, unsigned MaxRecurse) { 1067 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1068 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1069 Constant *Ops[] = { C0, C1 }; 1070 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.TD, Q.TLI); 1071 } 1072 } 1073 1074 bool isSigned = Opcode == Instruction::SDiv; 1075 1076 // X / undef -> undef 1077 if (match(Op1, m_Undef())) 1078 return Op1; 1079 1080 // undef / X -> 0 1081 if (match(Op0, m_Undef())) 1082 return Constant::getNullValue(Op0->getType()); 1083 1084 // 0 / X -> 0, we don't need to preserve faults! 1085 if (match(Op0, m_Zero())) 1086 return Op0; 1087 1088 // X / 1 -> X 1089 if (match(Op1, m_One())) 1090 return Op0; 1091 1092 if (Op0->getType()->isIntegerTy(1)) 1093 // It can't be division by zero, hence it must be division by one. 1094 return Op0; 1095 1096 // X / X -> 1 1097 if (Op0 == Op1) 1098 return ConstantInt::get(Op0->getType(), 1); 1099 1100 // (X * Y) / Y -> X if the multiplication does not overflow. 1101 Value *X = 0, *Y = 0; 1102 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) { 1103 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1 1104 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0); 1105 // If the Mul knows it does not overflow, then we are good to go. 1106 if ((isSigned && Mul->hasNoSignedWrap()) || 1107 (!isSigned && Mul->hasNoUnsignedWrap())) 1108 return X; 1109 // If X has the form X = A / Y then X * Y cannot overflow. 1110 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X)) 1111 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y) 1112 return X; 1113 } 1114 1115 // (X rem Y) / Y -> 0 1116 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) || 1117 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1))))) 1118 return Constant::getNullValue(Op0->getType()); 1119 1120 // If the operation is with the result of a select instruction, check whether 1121 // operating on either branch of the select always yields the same value. 1122 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1123 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1124 return V; 1125 1126 // If the operation is with the result of a phi instruction, check whether 1127 // operating on all incoming values of the phi always yields the same value. 1128 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1129 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1130 return V; 1131 1132 return 0; 1133} 1134 1135/// SimplifySDivInst - Given operands for an SDiv, see if we can 1136/// fold the result. If not, this returns null. 1137static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q, 1138 unsigned MaxRecurse) { 1139 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse)) 1140 return V; 1141 1142 return 0; 1143} 1144 1145Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout *TD, 1146 const TargetLibraryInfo *TLI, 1147 const DominatorTree *DT) { 1148 return ::SimplifySDivInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1149} 1150 1151/// SimplifyUDivInst - Given operands for a UDiv, see if we can 1152/// fold the result. If not, this returns null. 1153static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q, 1154 unsigned MaxRecurse) { 1155 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse)) 1156 return V; 1157 1158 return 0; 1159} 1160 1161Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout *TD, 1162 const TargetLibraryInfo *TLI, 1163 const DominatorTree *DT) { 1164 return ::SimplifyUDivInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1165} 1166 1167static Value *SimplifyFDivInst(Value *Op0, Value *Op1, const Query &Q, 1168 unsigned) { 1169 // undef / X -> undef (the undef could be a snan). 1170 if (match(Op0, m_Undef())) 1171 return Op0; 1172 1173 // X / undef -> undef 1174 if (match(Op1, m_Undef())) 1175 return Op1; 1176 1177 return 0; 1178} 1179 1180Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, const DataLayout *TD, 1181 const TargetLibraryInfo *TLI, 1182 const DominatorTree *DT) { 1183 return ::SimplifyFDivInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1184} 1185 1186/// SimplifyRem - Given operands for an SRem or URem, see if we can 1187/// fold the result. If not, this returns null. 1188static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, 1189 const Query &Q, unsigned MaxRecurse) { 1190 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1191 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1192 Constant *Ops[] = { C0, C1 }; 1193 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.TD, Q.TLI); 1194 } 1195 } 1196 1197 // X % undef -> undef 1198 if (match(Op1, m_Undef())) 1199 return Op1; 1200 1201 // undef % X -> 0 1202 if (match(Op0, m_Undef())) 1203 return Constant::getNullValue(Op0->getType()); 1204 1205 // 0 % X -> 0, we don't need to preserve faults! 1206 if (match(Op0, m_Zero())) 1207 return Op0; 1208 1209 // X % 0 -> undef, we don't need to preserve faults! 1210 if (match(Op1, m_Zero())) 1211 return UndefValue::get(Op0->getType()); 1212 1213 // X % 1 -> 0 1214 if (match(Op1, m_One())) 1215 return Constant::getNullValue(Op0->getType()); 1216 1217 if (Op0->getType()->isIntegerTy(1)) 1218 // It can't be remainder by zero, hence it must be remainder by one. 1219 return Constant::getNullValue(Op0->getType()); 1220 1221 // X % X -> 0 1222 if (Op0 == Op1) 1223 return Constant::getNullValue(Op0->getType()); 1224 1225 // If the operation is with the result of a select instruction, check whether 1226 // operating on either branch of the select always yields the same value. 1227 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1228 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1229 return V; 1230 1231 // If the operation is with the result of a phi instruction, check whether 1232 // operating on all incoming values of the phi always yields the same value. 1233 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1234 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1235 return V; 1236 1237 return 0; 1238} 1239 1240/// SimplifySRemInst - Given operands for an SRem, see if we can 1241/// fold the result. If not, this returns null. 1242static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q, 1243 unsigned MaxRecurse) { 1244 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse)) 1245 return V; 1246 1247 return 0; 1248} 1249 1250Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout *TD, 1251 const TargetLibraryInfo *TLI, 1252 const DominatorTree *DT) { 1253 return ::SimplifySRemInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1254} 1255 1256/// SimplifyURemInst - Given operands for a URem, see if we can 1257/// fold the result. If not, this returns null. 1258static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q, 1259 unsigned MaxRecurse) { 1260 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse)) 1261 return V; 1262 1263 return 0; 1264} 1265 1266Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout *TD, 1267 const TargetLibraryInfo *TLI, 1268 const DominatorTree *DT) { 1269 return ::SimplifyURemInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1270} 1271 1272static Value *SimplifyFRemInst(Value *Op0, Value *Op1, const Query &, 1273 unsigned) { 1274 // undef % X -> undef (the undef could be a snan). 1275 if (match(Op0, m_Undef())) 1276 return Op0; 1277 1278 // X % undef -> undef 1279 if (match(Op1, m_Undef())) 1280 return Op1; 1281 1282 return 0; 1283} 1284 1285Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, const DataLayout *TD, 1286 const TargetLibraryInfo *TLI, 1287 const DominatorTree *DT) { 1288 return ::SimplifyFRemInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1289} 1290 1291/// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can 1292/// fold the result. If not, this returns null. 1293static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1, 1294 const Query &Q, unsigned MaxRecurse) { 1295 if (Constant *C0 = dyn_cast<Constant>(Op0)) { 1296 if (Constant *C1 = dyn_cast<Constant>(Op1)) { 1297 Constant *Ops[] = { C0, C1 }; 1298 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.TD, Q.TLI); 1299 } 1300 } 1301 1302 // 0 shift by X -> 0 1303 if (match(Op0, m_Zero())) 1304 return Op0; 1305 1306 // X shift by 0 -> X 1307 if (match(Op1, m_Zero())) 1308 return Op0; 1309 1310 // X shift by undef -> undef because it may shift by the bitwidth. 1311 if (match(Op1, m_Undef())) 1312 return Op1; 1313 1314 // Shifting by the bitwidth or more is undefined. 1315 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) 1316 if (CI->getValue().getLimitedValue() >= 1317 Op0->getType()->getScalarSizeInBits()) 1318 return UndefValue::get(Op0->getType()); 1319 1320 // If the operation is with the result of a select instruction, check whether 1321 // operating on either branch of the select always yields the same value. 1322 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1323 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse)) 1324 return V; 1325 1326 // If the operation is with the result of a phi instruction, check whether 1327 // operating on all incoming values of the phi always yields the same value. 1328 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1329 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse)) 1330 return V; 1331 1332 return 0; 1333} 1334 1335/// SimplifyShlInst - Given operands for an Shl, see if we can 1336/// fold the result. If not, this returns null. 1337static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1338 const Query &Q, unsigned MaxRecurse) { 1339 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse)) 1340 return V; 1341 1342 // undef << X -> 0 1343 if (match(Op0, m_Undef())) 1344 return Constant::getNullValue(Op0->getType()); 1345 1346 // (X >> A) << A -> X 1347 Value *X; 1348 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1))))) 1349 return X; 1350 return 0; 1351} 1352 1353Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW, 1354 const DataLayout *TD, const TargetLibraryInfo *TLI, 1355 const DominatorTree *DT) { 1356 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query (TD, TLI, DT), 1357 RecursionLimit); 1358} 1359 1360/// SimplifyLShrInst - Given operands for an LShr, see if we can 1361/// fold the result. If not, this returns null. 1362static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1363 const Query &Q, unsigned MaxRecurse) { 1364 if (Value *V = SimplifyShift(Instruction::LShr, Op0, Op1, Q, MaxRecurse)) 1365 return V; 1366 1367 // X >> X -> 0 1368 if (Op0 == Op1) 1369 return Constant::getNullValue(Op0->getType()); 1370 1371 // undef >>l X -> 0 1372 if (match(Op0, m_Undef())) 1373 return Constant::getNullValue(Op0->getType()); 1374 1375 // (X << A) >> A -> X 1376 Value *X; 1377 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) && 1378 cast<OverflowingBinaryOperator>(Op0)->hasNoUnsignedWrap()) 1379 return X; 1380 1381 return 0; 1382} 1383 1384Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact, 1385 const DataLayout *TD, 1386 const TargetLibraryInfo *TLI, 1387 const DominatorTree *DT) { 1388 return ::SimplifyLShrInst(Op0, Op1, isExact, Query (TD, TLI, DT), 1389 RecursionLimit); 1390} 1391 1392/// SimplifyAShrInst - Given operands for an AShr, see if we can 1393/// fold the result. If not, this returns null. 1394static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1395 const Query &Q, unsigned MaxRecurse) { 1396 if (Value *V = SimplifyShift(Instruction::AShr, Op0, Op1, Q, MaxRecurse)) 1397 return V; 1398 1399 // X >> X -> 0 1400 if (Op0 == Op1) 1401 return Constant::getNullValue(Op0->getType()); 1402 1403 // all ones >>a X -> all ones 1404 if (match(Op0, m_AllOnes())) 1405 return Op0; 1406 1407 // undef >>a X -> all ones 1408 if (match(Op0, m_Undef())) 1409 return Constant::getAllOnesValue(Op0->getType()); 1410 1411 // (X << A) >> A -> X 1412 Value *X; 1413 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) && 1414 cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap()) 1415 return X; 1416 1417 return 0; 1418} 1419 1420Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact, 1421 const DataLayout *TD, 1422 const TargetLibraryInfo *TLI, 1423 const DominatorTree *DT) { 1424 return ::SimplifyAShrInst(Op0, Op1, isExact, Query (TD, TLI, DT), 1425 RecursionLimit); 1426} 1427 1428/// SimplifyAndInst - Given operands for an And, see if we can 1429/// fold the result. If not, this returns null. 1430static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q, 1431 unsigned MaxRecurse) { 1432 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1433 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1434 Constant *Ops[] = { CLHS, CRHS }; 1435 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(), 1436 Ops, Q.TD, Q.TLI); 1437 } 1438 1439 // Canonicalize the constant to the RHS. 1440 std::swap(Op0, Op1); 1441 } 1442 1443 // X & undef -> 0 1444 if (match(Op1, m_Undef())) 1445 return Constant::getNullValue(Op0->getType()); 1446 1447 // X & X = X 1448 if (Op0 == Op1) 1449 return Op0; 1450 1451 // X & 0 = 0 1452 if (match(Op1, m_Zero())) 1453 return Op1; 1454 1455 // X & -1 = X 1456 if (match(Op1, m_AllOnes())) 1457 return Op0; 1458 1459 // A & ~A = ~A & A = 0 1460 if (match(Op0, m_Not(m_Specific(Op1))) || 1461 match(Op1, m_Not(m_Specific(Op0)))) 1462 return Constant::getNullValue(Op0->getType()); 1463 1464 // (A | ?) & A = A 1465 Value *A = 0, *B = 0; 1466 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1467 (A == Op1 || B == Op1)) 1468 return Op1; 1469 1470 // A & (A | ?) = A 1471 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1472 (A == Op0 || B == Op0)) 1473 return Op0; 1474 1475 // A & (-A) = A if A is a power of two or zero. 1476 if (match(Op0, m_Neg(m_Specific(Op1))) || 1477 match(Op1, m_Neg(m_Specific(Op0)))) { 1478 if (isKnownToBeAPowerOfTwo(Op0, /*OrZero*/true)) 1479 return Op0; 1480 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/true)) 1481 return Op1; 1482 } 1483 1484 // Try some generic simplifications for associative operations. 1485 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q, 1486 MaxRecurse)) 1487 return V; 1488 1489 // And distributes over Or. Try some generic simplifications based on this. 1490 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or, 1491 Q, MaxRecurse)) 1492 return V; 1493 1494 // And distributes over Xor. Try some generic simplifications based on this. 1495 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor, 1496 Q, MaxRecurse)) 1497 return V; 1498 1499 // Or distributes over And. Try some generic simplifications based on this. 1500 if (Value *V = FactorizeBinOp(Instruction::And, Op0, Op1, Instruction::Or, 1501 Q, MaxRecurse)) 1502 return V; 1503 1504 // If the operation is with the result of a select instruction, check whether 1505 // operating on either branch of the select always yields the same value. 1506 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1507 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q, 1508 MaxRecurse)) 1509 return V; 1510 1511 // If the operation is with the result of a phi instruction, check whether 1512 // operating on all incoming values of the phi always yields the same value. 1513 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1514 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q, 1515 MaxRecurse)) 1516 return V; 1517 1518 return 0; 1519} 1520 1521Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout *TD, 1522 const TargetLibraryInfo *TLI, 1523 const DominatorTree *DT) { 1524 return ::SimplifyAndInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1525} 1526 1527/// SimplifyOrInst - Given operands for an Or, see if we can 1528/// fold the result. If not, this returns null. 1529static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q, 1530 unsigned MaxRecurse) { 1531 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1532 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1533 Constant *Ops[] = { CLHS, CRHS }; 1534 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(), 1535 Ops, Q.TD, Q.TLI); 1536 } 1537 1538 // Canonicalize the constant to the RHS. 1539 std::swap(Op0, Op1); 1540 } 1541 1542 // X | undef -> -1 1543 if (match(Op1, m_Undef())) 1544 return Constant::getAllOnesValue(Op0->getType()); 1545 1546 // X | X = X 1547 if (Op0 == Op1) 1548 return Op0; 1549 1550 // X | 0 = X 1551 if (match(Op1, m_Zero())) 1552 return Op0; 1553 1554 // X | -1 = -1 1555 if (match(Op1, m_AllOnes())) 1556 return Op1; 1557 1558 // A | ~A = ~A | A = -1 1559 if (match(Op0, m_Not(m_Specific(Op1))) || 1560 match(Op1, m_Not(m_Specific(Op0)))) 1561 return Constant::getAllOnesValue(Op0->getType()); 1562 1563 // (A & ?) | A = A 1564 Value *A = 0, *B = 0; 1565 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 1566 (A == Op1 || B == Op1)) 1567 return Op1; 1568 1569 // A | (A & ?) = A 1570 if (match(Op1, m_And(m_Value(A), m_Value(B))) && 1571 (A == Op0 || B == Op0)) 1572 return Op0; 1573 1574 // ~(A & ?) | A = -1 1575 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) && 1576 (A == Op1 || B == Op1)) 1577 return Constant::getAllOnesValue(Op1->getType()); 1578 1579 // A | ~(A & ?) = -1 1580 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) && 1581 (A == Op0 || B == Op0)) 1582 return Constant::getAllOnesValue(Op0->getType()); 1583 1584 // Try some generic simplifications for associative operations. 1585 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q, 1586 MaxRecurse)) 1587 return V; 1588 1589 // Or distributes over And. Try some generic simplifications based on this. 1590 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q, 1591 MaxRecurse)) 1592 return V; 1593 1594 // And distributes over Or. Try some generic simplifications based on this. 1595 if (Value *V = FactorizeBinOp(Instruction::Or, Op0, Op1, Instruction::And, 1596 Q, MaxRecurse)) 1597 return V; 1598 1599 // If the operation is with the result of a select instruction, check whether 1600 // operating on either branch of the select always yields the same value. 1601 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) 1602 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q, 1603 MaxRecurse)) 1604 return V; 1605 1606 // If the operation is with the result of a phi instruction, check whether 1607 // operating on all incoming values of the phi always yields the same value. 1608 if (isa<PHINode>(Op0) || isa<PHINode>(Op1)) 1609 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse)) 1610 return V; 1611 1612 return 0; 1613} 1614 1615Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout *TD, 1616 const TargetLibraryInfo *TLI, 1617 const DominatorTree *DT) { 1618 return ::SimplifyOrInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1619} 1620 1621/// SimplifyXorInst - Given operands for a Xor, see if we can 1622/// fold the result. If not, this returns null. 1623static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q, 1624 unsigned MaxRecurse) { 1625 if (Constant *CLHS = dyn_cast<Constant>(Op0)) { 1626 if (Constant *CRHS = dyn_cast<Constant>(Op1)) { 1627 Constant *Ops[] = { CLHS, CRHS }; 1628 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(), 1629 Ops, Q.TD, Q.TLI); 1630 } 1631 1632 // Canonicalize the constant to the RHS. 1633 std::swap(Op0, Op1); 1634 } 1635 1636 // A ^ undef -> undef 1637 if (match(Op1, m_Undef())) 1638 return Op1; 1639 1640 // A ^ 0 = A 1641 if (match(Op1, m_Zero())) 1642 return Op0; 1643 1644 // A ^ A = 0 1645 if (Op0 == Op1) 1646 return Constant::getNullValue(Op0->getType()); 1647 1648 // A ^ ~A = ~A ^ A = -1 1649 if (match(Op0, m_Not(m_Specific(Op1))) || 1650 match(Op1, m_Not(m_Specific(Op0)))) 1651 return Constant::getAllOnesValue(Op0->getType()); 1652 1653 // Try some generic simplifications for associative operations. 1654 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q, 1655 MaxRecurse)) 1656 return V; 1657 1658 // And distributes over Xor. Try some generic simplifications based on this. 1659 if (Value *V = FactorizeBinOp(Instruction::Xor, Op0, Op1, Instruction::And, 1660 Q, MaxRecurse)) 1661 return V; 1662 1663 // Threading Xor over selects and phi nodes is pointless, so don't bother. 1664 // Threading over the select in "A ^ select(cond, B, C)" means evaluating 1665 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and 1666 // only if B and C are equal. If B and C are equal then (since we assume 1667 // that operands have already been simplified) "select(cond, B, C)" should 1668 // have been simplified to the common value of B and C already. Analysing 1669 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly 1670 // for threading over phi nodes. 1671 1672 return 0; 1673} 1674 1675Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout *TD, 1676 const TargetLibraryInfo *TLI, 1677 const DominatorTree *DT) { 1678 return ::SimplifyXorInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit); 1679} 1680 1681static Type *GetCompareTy(Value *Op) { 1682 return CmpInst::makeCmpResultType(Op->getType()); 1683} 1684 1685/// ExtractEquivalentCondition - Rummage around inside V looking for something 1686/// equivalent to the comparison "LHS Pred RHS". Return such a value if found, 1687/// otherwise return null. Helper function for analyzing max/min idioms. 1688static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred, 1689 Value *LHS, Value *RHS) { 1690 SelectInst *SI = dyn_cast<SelectInst>(V); 1691 if (!SI) 1692 return 0; 1693 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); 1694 if (!Cmp) 1695 return 0; 1696 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1); 1697 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS) 1698 return Cmp; 1699 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) && 1700 LHS == CmpRHS && RHS == CmpLHS) 1701 return Cmp; 1702 return 0; 1703} 1704 1705// A significant optimization not implemented here is assuming that alloca 1706// addresses are not equal to incoming argument values. They don't *alias*, 1707// as we say, but that doesn't mean they aren't equal, so we take a 1708// conservative approach. 1709// 1710// This is inspired in part by C++11 5.10p1: 1711// "Two pointers of the same type compare equal if and only if they are both 1712// null, both point to the same function, or both represent the same 1713// address." 1714// 1715// This is pretty permissive. 1716// 1717// It's also partly due to C11 6.5.9p6: 1718// "Two pointers compare equal if and only if both are null pointers, both are 1719// pointers to the same object (including a pointer to an object and a 1720// subobject at its beginning) or function, both are pointers to one past the 1721// last element of the same array object, or one is a pointer to one past the 1722// end of one array object and the other is a pointer to the start of a 1723// different array object that happens to immediately follow the first array 1724// object in the address space.) 1725// 1726// C11's version is more restrictive, however there's no reason why an argument 1727// couldn't be a one-past-the-end value for a stack object in the caller and be 1728// equal to the beginning of a stack object in the callee. 1729// 1730// If the C and C++ standards are ever made sufficiently restrictive in this 1731// area, it may be possible to update LLVM's semantics accordingly and reinstate 1732// this optimization. 1733static Constant *computePointerICmp(const DataLayout *TD, 1734 const TargetLibraryInfo *TLI, 1735 CmpInst::Predicate Pred, 1736 Value *LHS, Value *RHS) { 1737 // First, skip past any trivial no-ops. 1738 LHS = LHS->stripPointerCasts(); 1739 RHS = RHS->stripPointerCasts(); 1740 1741 // A non-null pointer is not equal to a null pointer. 1742 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) && 1743 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE)) 1744 return ConstantInt::get(GetCompareTy(LHS), 1745 !CmpInst::isTrueWhenEqual(Pred)); 1746 1747 // We can only fold certain predicates on pointer comparisons. 1748 switch (Pred) { 1749 default: 1750 return 0; 1751 1752 // Equality comaprisons are easy to fold. 1753 case CmpInst::ICMP_EQ: 1754 case CmpInst::ICMP_NE: 1755 break; 1756 1757 // We can only handle unsigned relational comparisons because 'inbounds' on 1758 // a GEP only protects against unsigned wrapping. 1759 case CmpInst::ICMP_UGT: 1760 case CmpInst::ICMP_UGE: 1761 case CmpInst::ICMP_ULT: 1762 case CmpInst::ICMP_ULE: 1763 // However, we have to switch them to their signed variants to handle 1764 // negative indices from the base pointer. 1765 Pred = ICmpInst::getSignedPredicate(Pred); 1766 break; 1767 } 1768 1769 // Strip off any constant offsets so that we can reason about them. 1770 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets 1771 // here and compare base addresses like AliasAnalysis does, however there are 1772 // numerous hazards. AliasAnalysis and its utilities rely on special rules 1773 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis 1774 // doesn't need to guarantee pointer inequality when it says NoAlias. 1775 Constant *LHSOffset = stripAndComputeConstantOffsets(TD, LHS); 1776 Constant *RHSOffset = stripAndComputeConstantOffsets(TD, RHS); 1777 1778 // If LHS and RHS are related via constant offsets to the same base 1779 // value, we can replace it with an icmp which just compares the offsets. 1780 if (LHS == RHS) 1781 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset); 1782 1783 // Various optimizations for (in)equality comparisons. 1784 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) { 1785 // Different non-empty allocations that exist at the same time have 1786 // different addresses (if the program can tell). Global variables always 1787 // exist, so they always exist during the lifetime of each other and all 1788 // allocas. Two different allocas usually have different addresses... 1789 // 1790 // However, if there's an @llvm.stackrestore dynamically in between two 1791 // allocas, they may have the same address. It's tempting to reduce the 1792 // scope of the problem by only looking at *static* allocas here. That would 1793 // cover the majority of allocas while significantly reducing the likelihood 1794 // of having an @llvm.stackrestore pop up in the middle. However, it's not 1795 // actually impossible for an @llvm.stackrestore to pop up in the middle of 1796 // an entry block. Also, if we have a block that's not attached to a 1797 // function, we can't tell if it's "static" under the current definition. 1798 // Theoretically, this problem could be fixed by creating a new kind of 1799 // instruction kind specifically for static allocas. Such a new instruction 1800 // could be required to be at the top of the entry block, thus preventing it 1801 // from being subject to a @llvm.stackrestore. Instcombine could even 1802 // convert regular allocas into these special allocas. It'd be nifty. 1803 // However, until then, this problem remains open. 1804 // 1805 // So, we'll assume that two non-empty allocas have different addresses 1806 // for now. 1807 // 1808 // With all that, if the offsets are within the bounds of their allocations 1809 // (and not one-past-the-end! so we can't use inbounds!), and their 1810 // allocations aren't the same, the pointers are not equal. 1811 // 1812 // Note that it's not necessary to check for LHS being a global variable 1813 // address, due to canonicalization and constant folding. 1814 if (isa<AllocaInst>(LHS) && 1815 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) { 1816 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset); 1817 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset); 1818 uint64_t LHSSize, RHSSize; 1819 if (LHSOffsetCI && RHSOffsetCI && 1820 getObjectSize(LHS, LHSSize, TD, TLI) && 1821 getObjectSize(RHS, RHSSize, TD, TLI)) { 1822 const APInt &LHSOffsetValue = LHSOffsetCI->getValue(); 1823 const APInt &RHSOffsetValue = RHSOffsetCI->getValue(); 1824 if (!LHSOffsetValue.isNegative() && 1825 !RHSOffsetValue.isNegative() && 1826 LHSOffsetValue.ult(LHSSize) && 1827 RHSOffsetValue.ult(RHSSize)) { 1828 return ConstantInt::get(GetCompareTy(LHS), 1829 !CmpInst::isTrueWhenEqual(Pred)); 1830 } 1831 } 1832 1833 // Repeat the above check but this time without depending on DataLayout 1834 // or being able to compute a precise size. 1835 if (!cast<PointerType>(LHS->getType())->isEmptyTy() && 1836 !cast<PointerType>(RHS->getType())->isEmptyTy() && 1837 LHSOffset->isNullValue() && 1838 RHSOffset->isNullValue()) 1839 return ConstantInt::get(GetCompareTy(LHS), 1840 !CmpInst::isTrueWhenEqual(Pred)); 1841 } 1842 1843 // Even if an non-inbounds GEP occurs along the path we can still optimize 1844 // equality comparisons concerning the result. We avoid walking the whole 1845 // chain again by starting where the last calls to 1846 // stripAndComputeConstantOffsets left off and accumulate the offsets. 1847 Constant *LHSNoBound = stripAndComputeConstantOffsets(TD, LHS, true); 1848 Constant *RHSNoBound = stripAndComputeConstantOffsets(TD, RHS, true); 1849 if (LHS == RHS) 1850 return ConstantExpr::getICmp(Pred, 1851 ConstantExpr::getAdd(LHSOffset, LHSNoBound), 1852 ConstantExpr::getAdd(RHSOffset, RHSNoBound)); 1853 } 1854 1855 // Otherwise, fail. 1856 return 0; 1857} 1858 1859/// SimplifyICmpInst - Given operands for an ICmpInst, see if we can 1860/// fold the result. If not, this returns null. 1861static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 1862 const Query &Q, unsigned MaxRecurse) { 1863 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 1864 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!"); 1865 1866 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 1867 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 1868 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.TD, Q.TLI); 1869 1870 // If we have a constant, make sure it is on the RHS. 1871 std::swap(LHS, RHS); 1872 Pred = CmpInst::getSwappedPredicate(Pred); 1873 } 1874 1875 Type *ITy = GetCompareTy(LHS); // The return type. 1876 Type *OpTy = LHS->getType(); // The operand type. 1877 1878 // icmp X, X -> true/false 1879 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false 1880 // because X could be 0. 1881 if (LHS == RHS || isa<UndefValue>(RHS)) 1882 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred)); 1883 1884 // Special case logic when the operands have i1 type. 1885 if (OpTy->getScalarType()->isIntegerTy(1)) { 1886 switch (Pred) { 1887 default: break; 1888 case ICmpInst::ICMP_EQ: 1889 // X == 1 -> X 1890 if (match(RHS, m_One())) 1891 return LHS; 1892 break; 1893 case ICmpInst::ICMP_NE: 1894 // X != 0 -> X 1895 if (match(RHS, m_Zero())) 1896 return LHS; 1897 break; 1898 case ICmpInst::ICMP_UGT: 1899 // X >u 0 -> X 1900 if (match(RHS, m_Zero())) 1901 return LHS; 1902 break; 1903 case ICmpInst::ICMP_UGE: 1904 // X >=u 1 -> X 1905 if (match(RHS, m_One())) 1906 return LHS; 1907 break; 1908 case ICmpInst::ICMP_SLT: 1909 // X <s 0 -> X 1910 if (match(RHS, m_Zero())) 1911 return LHS; 1912 break; 1913 case ICmpInst::ICMP_SLE: 1914 // X <=s -1 -> X 1915 if (match(RHS, m_One())) 1916 return LHS; 1917 break; 1918 } 1919 } 1920 1921 // If we are comparing with zero then try hard since this is a common case. 1922 if (match(RHS, m_Zero())) { 1923 bool LHSKnownNonNegative, LHSKnownNegative; 1924 switch (Pred) { 1925 default: llvm_unreachable("Unknown ICmp predicate!"); 1926 case ICmpInst::ICMP_ULT: 1927 return getFalse(ITy); 1928 case ICmpInst::ICMP_UGE: 1929 return getTrue(ITy); 1930 case ICmpInst::ICMP_EQ: 1931 case ICmpInst::ICMP_ULE: 1932 if (isKnownNonZero(LHS, Q.TD)) 1933 return getFalse(ITy); 1934 break; 1935 case ICmpInst::ICMP_NE: 1936 case ICmpInst::ICMP_UGT: 1937 if (isKnownNonZero(LHS, Q.TD)) 1938 return getTrue(ITy); 1939 break; 1940 case ICmpInst::ICMP_SLT: 1941 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.TD); 1942 if (LHSKnownNegative) 1943 return getTrue(ITy); 1944 if (LHSKnownNonNegative) 1945 return getFalse(ITy); 1946 break; 1947 case ICmpInst::ICMP_SLE: 1948 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.TD); 1949 if (LHSKnownNegative) 1950 return getTrue(ITy); 1951 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.TD)) 1952 return getFalse(ITy); 1953 break; 1954 case ICmpInst::ICMP_SGE: 1955 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.TD); 1956 if (LHSKnownNegative) 1957 return getFalse(ITy); 1958 if (LHSKnownNonNegative) 1959 return getTrue(ITy); 1960 break; 1961 case ICmpInst::ICMP_SGT: 1962 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.TD); 1963 if (LHSKnownNegative) 1964 return getFalse(ITy); 1965 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.TD)) 1966 return getTrue(ITy); 1967 break; 1968 } 1969 } 1970 1971 // See if we are doing a comparison with a constant integer. 1972 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 1973 // Rule out tautological comparisons (eg., ult 0 or uge 0). 1974 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue()); 1975 if (RHS_CR.isEmptySet()) 1976 return ConstantInt::getFalse(CI->getContext()); 1977 if (RHS_CR.isFullSet()) 1978 return ConstantInt::getTrue(CI->getContext()); 1979 1980 // Many binary operators with constant RHS have easy to compute constant 1981 // range. Use them to check whether the comparison is a tautology. 1982 uint32_t Width = CI->getBitWidth(); 1983 APInt Lower = APInt(Width, 0); 1984 APInt Upper = APInt(Width, 0); 1985 ConstantInt *CI2; 1986 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) { 1987 // 'urem x, CI2' produces [0, CI2). 1988 Upper = CI2->getValue(); 1989 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) { 1990 // 'srem x, CI2' produces (-|CI2|, |CI2|). 1991 Upper = CI2->getValue().abs(); 1992 Lower = (-Upper) + 1; 1993 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) { 1994 // 'udiv CI2, x' produces [0, CI2]. 1995 Upper = CI2->getValue() + 1; 1996 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) { 1997 // 'udiv x, CI2' produces [0, UINT_MAX / CI2]. 1998 APInt NegOne = APInt::getAllOnesValue(Width); 1999 if (!CI2->isZero()) 2000 Upper = NegOne.udiv(CI2->getValue()) + 1; 2001 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) { 2002 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2]. 2003 APInt IntMin = APInt::getSignedMinValue(Width); 2004 APInt IntMax = APInt::getSignedMaxValue(Width); 2005 APInt Val = CI2->getValue().abs(); 2006 if (!Val.isMinValue()) { 2007 Lower = IntMin.sdiv(Val); 2008 Upper = IntMax.sdiv(Val) + 1; 2009 } 2010 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) { 2011 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2]. 2012 APInt NegOne = APInt::getAllOnesValue(Width); 2013 if (CI2->getValue().ult(Width)) 2014 Upper = NegOne.lshr(CI2->getValue()) + 1; 2015 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) { 2016 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2]. 2017 APInt IntMin = APInt::getSignedMinValue(Width); 2018 APInt IntMax = APInt::getSignedMaxValue(Width); 2019 if (CI2->getValue().ult(Width)) { 2020 Lower = IntMin.ashr(CI2->getValue()); 2021 Upper = IntMax.ashr(CI2->getValue()) + 1; 2022 } 2023 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) { 2024 // 'or x, CI2' produces [CI2, UINT_MAX]. 2025 Lower = CI2->getValue(); 2026 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) { 2027 // 'and x, CI2' produces [0, CI2]. 2028 Upper = CI2->getValue() + 1; 2029 } 2030 if (Lower != Upper) { 2031 ConstantRange LHS_CR = ConstantRange(Lower, Upper); 2032 if (RHS_CR.contains(LHS_CR)) 2033 return ConstantInt::getTrue(RHS->getContext()); 2034 if (RHS_CR.inverse().contains(LHS_CR)) 2035 return ConstantInt::getFalse(RHS->getContext()); 2036 } 2037 } 2038 2039 // Compare of cast, for example (zext X) != 0 -> X != 0 2040 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) { 2041 Instruction *LI = cast<CastInst>(LHS); 2042 Value *SrcOp = LI->getOperand(0); 2043 Type *SrcTy = SrcOp->getType(); 2044 Type *DstTy = LI->getType(); 2045 2046 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input 2047 // if the integer type is the same size as the pointer type. 2048 if (MaxRecurse && Q.TD && isa<PtrToIntInst>(LI) && 2049 Q.TD->getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) { 2050 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 2051 // Transfer the cast to the constant. 2052 if (Value *V = SimplifyICmpInst(Pred, SrcOp, 2053 ConstantExpr::getIntToPtr(RHSC, SrcTy), 2054 Q, MaxRecurse-1)) 2055 return V; 2056 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) { 2057 if (RI->getOperand(0)->getType() == SrcTy) 2058 // Compare without the cast. 2059 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2060 Q, MaxRecurse-1)) 2061 return V; 2062 } 2063 } 2064 2065 if (isa<ZExtInst>(LHS)) { 2066 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the 2067 // same type. 2068 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) { 2069 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2070 // Compare X and Y. Note that signed predicates become unsigned. 2071 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2072 SrcOp, RI->getOperand(0), Q, 2073 MaxRecurse-1)) 2074 return V; 2075 } 2076 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended 2077 // too. If not, then try to deduce the result of the comparison. 2078 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2079 // Compute the constant that would happen if we truncated to SrcTy then 2080 // reextended to DstTy. 2081 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2082 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy); 2083 2084 // If the re-extended constant didn't change then this is effectively 2085 // also a case of comparing two zero-extended values. 2086 if (RExt == CI && MaxRecurse) 2087 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), 2088 SrcOp, Trunc, Q, MaxRecurse-1)) 2089 return V; 2090 2091 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit 2092 // there. Use this to work out the result of the comparison. 2093 if (RExt != CI) { 2094 switch (Pred) { 2095 default: llvm_unreachable("Unknown ICmp predicate!"); 2096 // LHS <u RHS. 2097 case ICmpInst::ICMP_EQ: 2098 case ICmpInst::ICMP_UGT: 2099 case ICmpInst::ICMP_UGE: 2100 return ConstantInt::getFalse(CI->getContext()); 2101 2102 case ICmpInst::ICMP_NE: 2103 case ICmpInst::ICMP_ULT: 2104 case ICmpInst::ICMP_ULE: 2105 return ConstantInt::getTrue(CI->getContext()); 2106 2107 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS 2108 // is non-negative then LHS <s RHS. 2109 case ICmpInst::ICMP_SGT: 2110 case ICmpInst::ICMP_SGE: 2111 return CI->getValue().isNegative() ? 2112 ConstantInt::getTrue(CI->getContext()) : 2113 ConstantInt::getFalse(CI->getContext()); 2114 2115 case ICmpInst::ICMP_SLT: 2116 case ICmpInst::ICMP_SLE: 2117 return CI->getValue().isNegative() ? 2118 ConstantInt::getFalse(CI->getContext()) : 2119 ConstantInt::getTrue(CI->getContext()); 2120 } 2121 } 2122 } 2123 } 2124 2125 if (isa<SExtInst>(LHS)) { 2126 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the 2127 // same type. 2128 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) { 2129 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) 2130 // Compare X and Y. Note that the predicate does not change. 2131 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), 2132 Q, MaxRecurse-1)) 2133 return V; 2134 } 2135 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended 2136 // too. If not, then try to deduce the result of the comparison. 2137 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { 2138 // Compute the constant that would happen if we truncated to SrcTy then 2139 // reextended to DstTy. 2140 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); 2141 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy); 2142 2143 // If the re-extended constant didn't change then this is effectively 2144 // also a case of comparing two sign-extended values. 2145 if (RExt == CI && MaxRecurse) 2146 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1)) 2147 return V; 2148 2149 // Otherwise the upper bits of LHS are all equal, while RHS has varying 2150 // bits there. Use this to work out the result of the comparison. 2151 if (RExt != CI) { 2152 switch (Pred) { 2153 default: llvm_unreachable("Unknown ICmp predicate!"); 2154 case ICmpInst::ICMP_EQ: 2155 return ConstantInt::getFalse(CI->getContext()); 2156 case ICmpInst::ICMP_NE: 2157 return ConstantInt::getTrue(CI->getContext()); 2158 2159 // If RHS is non-negative then LHS <s RHS. If RHS is negative then 2160 // LHS >s RHS. 2161 case ICmpInst::ICMP_SGT: 2162 case ICmpInst::ICMP_SGE: 2163 return CI->getValue().isNegative() ? 2164 ConstantInt::getTrue(CI->getContext()) : 2165 ConstantInt::getFalse(CI->getContext()); 2166 case ICmpInst::ICMP_SLT: 2167 case ICmpInst::ICMP_SLE: 2168 return CI->getValue().isNegative() ? 2169 ConstantInt::getFalse(CI->getContext()) : 2170 ConstantInt::getTrue(CI->getContext()); 2171 2172 // If LHS is non-negative then LHS <u RHS. If LHS is negative then 2173 // LHS >u RHS. 2174 case ICmpInst::ICMP_UGT: 2175 case ICmpInst::ICMP_UGE: 2176 // Comparison is true iff the LHS <s 0. 2177 if (MaxRecurse) 2178 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp, 2179 Constant::getNullValue(SrcTy), 2180 Q, MaxRecurse-1)) 2181 return V; 2182 break; 2183 case ICmpInst::ICMP_ULT: 2184 case ICmpInst::ICMP_ULE: 2185 // Comparison is true iff the LHS >=s 0. 2186 if (MaxRecurse) 2187 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp, 2188 Constant::getNullValue(SrcTy), 2189 Q, MaxRecurse-1)) 2190 return V; 2191 break; 2192 } 2193 } 2194 } 2195 } 2196 } 2197 2198 // Special logic for binary operators. 2199 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS); 2200 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS); 2201 if (MaxRecurse && (LBO || RBO)) { 2202 // Analyze the case when either LHS or RHS is an add instruction. 2203 Value *A = 0, *B = 0, *C = 0, *D = 0; 2204 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null). 2205 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false; 2206 if (LBO && LBO->getOpcode() == Instruction::Add) { 2207 A = LBO->getOperand(0); B = LBO->getOperand(1); 2208 NoLHSWrapProblem = ICmpInst::isEquality(Pred) || 2209 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) || 2210 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap()); 2211 } 2212 if (RBO && RBO->getOpcode() == Instruction::Add) { 2213 C = RBO->getOperand(0); D = RBO->getOperand(1); 2214 NoRHSWrapProblem = ICmpInst::isEquality(Pred) || 2215 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) || 2216 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap()); 2217 } 2218 2219 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2220 if ((A == RHS || B == RHS) && NoLHSWrapProblem) 2221 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A, 2222 Constant::getNullValue(RHS->getType()), 2223 Q, MaxRecurse-1)) 2224 return V; 2225 2226 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2227 if ((C == LHS || D == LHS) && NoRHSWrapProblem) 2228 if (Value *V = SimplifyICmpInst(Pred, 2229 Constant::getNullValue(LHS->getType()), 2230 C == LHS ? D : C, Q, MaxRecurse-1)) 2231 return V; 2232 2233 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow. 2234 if (A && C && (A == C || A == D || B == C || B == D) && 2235 NoLHSWrapProblem && NoRHSWrapProblem) { 2236 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2237 Value *Y, *Z; 2238 if (A == C) { 2239 // C + B == C + D -> B == D 2240 Y = B; 2241 Z = D; 2242 } else if (A == D) { 2243 // D + B == C + D -> B == C 2244 Y = B; 2245 Z = C; 2246 } else if (B == C) { 2247 // A + C == C + D -> A == D 2248 Y = A; 2249 Z = D; 2250 } else { 2251 assert(B == D); 2252 // A + D == C + D -> A == C 2253 Y = A; 2254 Z = C; 2255 } 2256 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1)) 2257 return V; 2258 } 2259 } 2260 2261 // icmp pred (urem X, Y), Y 2262 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) { 2263 bool KnownNonNegative, KnownNegative; 2264 switch (Pred) { 2265 default: 2266 break; 2267 case ICmpInst::ICMP_SGT: 2268 case ICmpInst::ICMP_SGE: 2269 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.TD); 2270 if (!KnownNonNegative) 2271 break; 2272 // fall-through 2273 case ICmpInst::ICMP_EQ: 2274 case ICmpInst::ICMP_UGT: 2275 case ICmpInst::ICMP_UGE: 2276 return getFalse(ITy); 2277 case ICmpInst::ICMP_SLT: 2278 case ICmpInst::ICMP_SLE: 2279 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.TD); 2280 if (!KnownNonNegative) 2281 break; 2282 // fall-through 2283 case ICmpInst::ICMP_NE: 2284 case ICmpInst::ICMP_ULT: 2285 case ICmpInst::ICMP_ULE: 2286 return getTrue(ITy); 2287 } 2288 } 2289 2290 // icmp pred X, (urem Y, X) 2291 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) { 2292 bool KnownNonNegative, KnownNegative; 2293 switch (Pred) { 2294 default: 2295 break; 2296 case ICmpInst::ICMP_SGT: 2297 case ICmpInst::ICMP_SGE: 2298 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.TD); 2299 if (!KnownNonNegative) 2300 break; 2301 // fall-through 2302 case ICmpInst::ICMP_NE: 2303 case ICmpInst::ICMP_UGT: 2304 case ICmpInst::ICMP_UGE: 2305 return getTrue(ITy); 2306 case ICmpInst::ICMP_SLT: 2307 case ICmpInst::ICMP_SLE: 2308 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.TD); 2309 if (!KnownNonNegative) 2310 break; 2311 // fall-through 2312 case ICmpInst::ICMP_EQ: 2313 case ICmpInst::ICMP_ULT: 2314 case ICmpInst::ICMP_ULE: 2315 return getFalse(ITy); 2316 } 2317 } 2318 2319 // x udiv y <=u x. 2320 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) { 2321 // icmp pred (X /u Y), X 2322 if (Pred == ICmpInst::ICMP_UGT) 2323 return getFalse(ITy); 2324 if (Pred == ICmpInst::ICMP_ULE) 2325 return getTrue(ITy); 2326 } 2327 2328 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() && 2329 LBO->getOperand(1) == RBO->getOperand(1)) { 2330 switch (LBO->getOpcode()) { 2331 default: break; 2332 case Instruction::UDiv: 2333 case Instruction::LShr: 2334 if (ICmpInst::isSigned(Pred)) 2335 break; 2336 // fall-through 2337 case Instruction::SDiv: 2338 case Instruction::AShr: 2339 if (!LBO->isExact() || !RBO->isExact()) 2340 break; 2341 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2342 RBO->getOperand(0), Q, MaxRecurse-1)) 2343 return V; 2344 break; 2345 case Instruction::Shl: { 2346 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap(); 2347 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap(); 2348 if (!NUW && !NSW) 2349 break; 2350 if (!NSW && ICmpInst::isSigned(Pred)) 2351 break; 2352 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), 2353 RBO->getOperand(0), Q, MaxRecurse-1)) 2354 return V; 2355 break; 2356 } 2357 } 2358 } 2359 2360 // Simplify comparisons involving max/min. 2361 Value *A, *B; 2362 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE; 2363 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B". 2364 2365 // Signed variants on "max(a,b)>=a -> true". 2366 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2367 if (A != RHS) std::swap(A, B); // smax(A, B) pred A. 2368 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2369 // We analyze this as smax(A, B) pred A. 2370 P = Pred; 2371 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) && 2372 (A == LHS || B == LHS)) { 2373 if (A != LHS) std::swap(A, B); // A pred smax(A, B). 2374 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". 2375 // We analyze this as smax(A, B) swapped-pred A. 2376 P = CmpInst::getSwappedPredicate(Pred); 2377 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2378 (A == RHS || B == RHS)) { 2379 if (A != RHS) std::swap(A, B); // smin(A, B) pred A. 2380 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2381 // We analyze this as smax(-A, -B) swapped-pred -A. 2382 // Note that we do not need to actually form -A or -B thanks to EqP. 2383 P = CmpInst::getSwappedPredicate(Pred); 2384 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) && 2385 (A == LHS || B == LHS)) { 2386 if (A != LHS) std::swap(A, B); // A pred smin(A, B). 2387 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". 2388 // We analyze this as smax(-A, -B) pred -A. 2389 // Note that we do not need to actually form -A or -B thanks to EqP. 2390 P = Pred; 2391 } 2392 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2393 // Cases correspond to "max(A, B) p A". 2394 switch (P) { 2395 default: 2396 break; 2397 case CmpInst::ICMP_EQ: 2398 case CmpInst::ICMP_SLE: 2399 // Equivalent to "A EqP B". This may be the same as the condition tested 2400 // in the max/min; if so, we can just return that. 2401 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2402 return V; 2403 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2404 return V; 2405 // Otherwise, see if "A EqP B" simplifies. 2406 if (MaxRecurse) 2407 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2408 return V; 2409 break; 2410 case CmpInst::ICMP_NE: 2411 case CmpInst::ICMP_SGT: { 2412 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2413 // Equivalent to "A InvEqP B". This may be the same as the condition 2414 // tested in the max/min; if so, we can just return that. 2415 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2416 return V; 2417 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2418 return V; 2419 // Otherwise, see if "A InvEqP B" simplifies. 2420 if (MaxRecurse) 2421 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2422 return V; 2423 break; 2424 } 2425 case CmpInst::ICMP_SGE: 2426 // Always true. 2427 return getTrue(ITy); 2428 case CmpInst::ICMP_SLT: 2429 // Always false. 2430 return getFalse(ITy); 2431 } 2432 } 2433 2434 // Unsigned variants on "max(a,b)>=a -> true". 2435 P = CmpInst::BAD_ICMP_PREDICATE; 2436 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { 2437 if (A != RHS) std::swap(A, B); // umax(A, B) pred A. 2438 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2439 // We analyze this as umax(A, B) pred A. 2440 P = Pred; 2441 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) && 2442 (A == LHS || B == LHS)) { 2443 if (A != LHS) std::swap(A, B); // A pred umax(A, B). 2444 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". 2445 // We analyze this as umax(A, B) swapped-pred A. 2446 P = CmpInst::getSwappedPredicate(Pred); 2447 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2448 (A == RHS || B == RHS)) { 2449 if (A != RHS) std::swap(A, B); // umin(A, B) pred A. 2450 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2451 // We analyze this as umax(-A, -B) swapped-pred -A. 2452 // Note that we do not need to actually form -A or -B thanks to EqP. 2453 P = CmpInst::getSwappedPredicate(Pred); 2454 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) && 2455 (A == LHS || B == LHS)) { 2456 if (A != LHS) std::swap(A, B); // A pred umin(A, B). 2457 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". 2458 // We analyze this as umax(-A, -B) pred -A. 2459 // Note that we do not need to actually form -A or -B thanks to EqP. 2460 P = Pred; 2461 } 2462 if (P != CmpInst::BAD_ICMP_PREDICATE) { 2463 // Cases correspond to "max(A, B) p A". 2464 switch (P) { 2465 default: 2466 break; 2467 case CmpInst::ICMP_EQ: 2468 case CmpInst::ICMP_ULE: 2469 // Equivalent to "A EqP B". This may be the same as the condition tested 2470 // in the max/min; if so, we can just return that. 2471 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) 2472 return V; 2473 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) 2474 return V; 2475 // Otherwise, see if "A EqP B" simplifies. 2476 if (MaxRecurse) 2477 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1)) 2478 return V; 2479 break; 2480 case CmpInst::ICMP_NE: 2481 case CmpInst::ICMP_UGT: { 2482 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); 2483 // Equivalent to "A InvEqP B". This may be the same as the condition 2484 // tested in the max/min; if so, we can just return that. 2485 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) 2486 return V; 2487 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) 2488 return V; 2489 // Otherwise, see if "A InvEqP B" simplifies. 2490 if (MaxRecurse) 2491 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1)) 2492 return V; 2493 break; 2494 } 2495 case CmpInst::ICMP_UGE: 2496 // Always true. 2497 return getTrue(ITy); 2498 case CmpInst::ICMP_ULT: 2499 // Always false. 2500 return getFalse(ITy); 2501 } 2502 } 2503 2504 // Variants on "max(x,y) >= min(x,z)". 2505 Value *C, *D; 2506 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && 2507 match(RHS, m_SMin(m_Value(C), m_Value(D))) && 2508 (A == C || A == D || B == C || B == D)) { 2509 // max(x, ?) pred min(x, ?). 2510 if (Pred == CmpInst::ICMP_SGE) 2511 // Always true. 2512 return getTrue(ITy); 2513 if (Pred == CmpInst::ICMP_SLT) 2514 // Always false. 2515 return getFalse(ITy); 2516 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && 2517 match(RHS, m_SMax(m_Value(C), m_Value(D))) && 2518 (A == C || A == D || B == C || B == D)) { 2519 // min(x, ?) pred max(x, ?). 2520 if (Pred == CmpInst::ICMP_SLE) 2521 // Always true. 2522 return getTrue(ITy); 2523 if (Pred == CmpInst::ICMP_SGT) 2524 // Always false. 2525 return getFalse(ITy); 2526 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && 2527 match(RHS, m_UMin(m_Value(C), m_Value(D))) && 2528 (A == C || A == D || B == C || B == D)) { 2529 // max(x, ?) pred min(x, ?). 2530 if (Pred == CmpInst::ICMP_UGE) 2531 // Always true. 2532 return getTrue(ITy); 2533 if (Pred == CmpInst::ICMP_ULT) 2534 // Always false. 2535 return getFalse(ITy); 2536 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && 2537 match(RHS, m_UMax(m_Value(C), m_Value(D))) && 2538 (A == C || A == D || B == C || B == D)) { 2539 // min(x, ?) pred max(x, ?). 2540 if (Pred == CmpInst::ICMP_ULE) 2541 // Always true. 2542 return getTrue(ITy); 2543 if (Pred == CmpInst::ICMP_UGT) 2544 // Always false. 2545 return getFalse(ITy); 2546 } 2547 2548 // Simplify comparisons of related pointers using a powerful, recursive 2549 // GEP-walk when we have target data available.. 2550 if (LHS->getType()->isPointerTy()) 2551 if (Constant *C = computePointerICmp(Q.TD, Q.TLI, Pred, LHS, RHS)) 2552 return C; 2553 2554 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) { 2555 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) { 2556 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() && 2557 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() && 2558 (ICmpInst::isEquality(Pred) || 2559 (GLHS->isInBounds() && GRHS->isInBounds() && 2560 Pred == ICmpInst::getSignedPredicate(Pred)))) { 2561 // The bases are equal and the indices are constant. Build a constant 2562 // expression GEP with the same indices and a null base pointer to see 2563 // what constant folding can make out of it. 2564 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType()); 2565 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end()); 2566 Constant *NewLHS = ConstantExpr::getGetElementPtr(Null, IndicesLHS); 2567 2568 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end()); 2569 Constant *NewRHS = ConstantExpr::getGetElementPtr(Null, IndicesRHS); 2570 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS); 2571 } 2572 } 2573 } 2574 2575 // If the comparison is with the result of a select instruction, check whether 2576 // comparing with either branch of the select always yields the same value. 2577 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 2578 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 2579 return V; 2580 2581 // If the comparison is with the result of a phi instruction, check whether 2582 // doing the compare with each incoming phi value yields a common result. 2583 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 2584 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 2585 return V; 2586 2587 return 0; 2588} 2589 2590Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2591 const DataLayout *TD, 2592 const TargetLibraryInfo *TLI, 2593 const DominatorTree *DT) { 2594 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query (TD, TLI, DT), 2595 RecursionLimit); 2596} 2597 2598/// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can 2599/// fold the result. If not, this returns null. 2600static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2601 const Query &Q, unsigned MaxRecurse) { 2602 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; 2603 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!"); 2604 2605 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 2606 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 2607 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.TD, Q.TLI); 2608 2609 // If we have a constant, make sure it is on the RHS. 2610 std::swap(LHS, RHS); 2611 Pred = CmpInst::getSwappedPredicate(Pred); 2612 } 2613 2614 // Fold trivial predicates. 2615 if (Pred == FCmpInst::FCMP_FALSE) 2616 return ConstantInt::get(GetCompareTy(LHS), 0); 2617 if (Pred == FCmpInst::FCMP_TRUE) 2618 return ConstantInt::get(GetCompareTy(LHS), 1); 2619 2620 if (isa<UndefValue>(RHS)) // fcmp pred X, undef -> undef 2621 return UndefValue::get(GetCompareTy(LHS)); 2622 2623 // fcmp x,x -> true/false. Not all compares are foldable. 2624 if (LHS == RHS) { 2625 if (CmpInst::isTrueWhenEqual(Pred)) 2626 return ConstantInt::get(GetCompareTy(LHS), 1); 2627 if (CmpInst::isFalseWhenEqual(Pred)) 2628 return ConstantInt::get(GetCompareTy(LHS), 0); 2629 } 2630 2631 // Handle fcmp with constant RHS 2632 if (Constant *RHSC = dyn_cast<Constant>(RHS)) { 2633 // If the constant is a nan, see if we can fold the comparison based on it. 2634 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) { 2635 if (CFP->getValueAPF().isNaN()) { 2636 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo" 2637 return ConstantInt::getFalse(CFP->getContext()); 2638 assert(FCmpInst::isUnordered(Pred) && 2639 "Comparison must be either ordered or unordered!"); 2640 // True if unordered. 2641 return ConstantInt::getTrue(CFP->getContext()); 2642 } 2643 // Check whether the constant is an infinity. 2644 if (CFP->getValueAPF().isInfinity()) { 2645 if (CFP->getValueAPF().isNegative()) { 2646 switch (Pred) { 2647 case FCmpInst::FCMP_OLT: 2648 // No value is ordered and less than negative infinity. 2649 return ConstantInt::getFalse(CFP->getContext()); 2650 case FCmpInst::FCMP_UGE: 2651 // All values are unordered with or at least negative infinity. 2652 return ConstantInt::getTrue(CFP->getContext()); 2653 default: 2654 break; 2655 } 2656 } else { 2657 switch (Pred) { 2658 case FCmpInst::FCMP_OGT: 2659 // No value is ordered and greater than infinity. 2660 return ConstantInt::getFalse(CFP->getContext()); 2661 case FCmpInst::FCMP_ULE: 2662 // All values are unordered with and at most infinity. 2663 return ConstantInt::getTrue(CFP->getContext()); 2664 default: 2665 break; 2666 } 2667 } 2668 } 2669 } 2670 } 2671 2672 // If the comparison is with the result of a select instruction, check whether 2673 // comparing with either branch of the select always yields the same value. 2674 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 2675 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) 2676 return V; 2677 2678 // If the comparison is with the result of a phi instruction, check whether 2679 // doing the compare with each incoming phi value yields a common result. 2680 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 2681 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) 2682 return V; 2683 2684 return 0; 2685} 2686 2687Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2688 const DataLayout *TD, 2689 const TargetLibraryInfo *TLI, 2690 const DominatorTree *DT) { 2691 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query (TD, TLI, DT), 2692 RecursionLimit); 2693} 2694 2695/// SimplifySelectInst - Given operands for a SelectInst, see if we can fold 2696/// the result. If not, this returns null. 2697static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal, 2698 Value *FalseVal, const Query &Q, 2699 unsigned MaxRecurse) { 2700 // select true, X, Y -> X 2701 // select false, X, Y -> Y 2702 if (ConstantInt *CB = dyn_cast<ConstantInt>(CondVal)) 2703 return CB->getZExtValue() ? TrueVal : FalseVal; 2704 2705 // select C, X, X -> X 2706 if (TrueVal == FalseVal) 2707 return TrueVal; 2708 2709 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y 2710 if (isa<Constant>(TrueVal)) 2711 return TrueVal; 2712 return FalseVal; 2713 } 2714 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X 2715 return FalseVal; 2716 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X 2717 return TrueVal; 2718 2719 return 0; 2720} 2721 2722Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, 2723 const DataLayout *TD, 2724 const TargetLibraryInfo *TLI, 2725 const DominatorTree *DT) { 2726 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Query (TD, TLI, DT), 2727 RecursionLimit); 2728} 2729 2730/// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can 2731/// fold the result. If not, this returns null. 2732static Value *SimplifyGEPInst(ArrayRef<Value *> Ops, const Query &Q, unsigned) { 2733 // The type of the GEP pointer operand. 2734 PointerType *PtrTy = dyn_cast<PointerType>(Ops[0]->getType()); 2735 // The GEP pointer operand is not a pointer, it's a vector of pointers. 2736 if (!PtrTy) 2737 return 0; 2738 2739 // getelementptr P -> P. 2740 if (Ops.size() == 1) 2741 return Ops[0]; 2742 2743 if (isa<UndefValue>(Ops[0])) { 2744 // Compute the (pointer) type returned by the GEP instruction. 2745 Type *LastType = GetElementPtrInst::getIndexedType(PtrTy, Ops.slice(1)); 2746 Type *GEPTy = PointerType::get(LastType, PtrTy->getAddressSpace()); 2747 return UndefValue::get(GEPTy); 2748 } 2749 2750 if (Ops.size() == 2) { 2751 // getelementptr P, 0 -> P. 2752 if (ConstantInt *C = dyn_cast<ConstantInt>(Ops[1])) 2753 if (C->isZero()) 2754 return Ops[0]; 2755 // getelementptr P, N -> P if P points to a type of zero size. 2756 if (Q.TD) { 2757 Type *Ty = PtrTy->getElementType(); 2758 if (Ty->isSized() && Q.TD->getTypeAllocSize(Ty) == 0) 2759 return Ops[0]; 2760 } 2761 } 2762 2763 // Check to see if this is constant foldable. 2764 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2765 if (!isa<Constant>(Ops[i])) 2766 return 0; 2767 2768 return ConstantExpr::getGetElementPtr(cast<Constant>(Ops[0]), Ops.slice(1)); 2769} 2770 2771Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout *TD, 2772 const TargetLibraryInfo *TLI, 2773 const DominatorTree *DT) { 2774 return ::SimplifyGEPInst(Ops, Query (TD, TLI, DT), RecursionLimit); 2775} 2776 2777/// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we 2778/// can fold the result. If not, this returns null. 2779static Value *SimplifyInsertValueInst(Value *Agg, Value *Val, 2780 ArrayRef<unsigned> Idxs, const Query &Q, 2781 unsigned) { 2782 if (Constant *CAgg = dyn_cast<Constant>(Agg)) 2783 if (Constant *CVal = dyn_cast<Constant>(Val)) 2784 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs); 2785 2786 // insertvalue x, undef, n -> x 2787 if (match(Val, m_Undef())) 2788 return Agg; 2789 2790 // insertvalue x, (extractvalue y, n), n 2791 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val)) 2792 if (EV->getAggregateOperand()->getType() == Agg->getType() && 2793 EV->getIndices() == Idxs) { 2794 // insertvalue undef, (extractvalue y, n), n -> y 2795 if (match(Agg, m_Undef())) 2796 return EV->getAggregateOperand(); 2797 2798 // insertvalue y, (extractvalue y, n), n -> y 2799 if (Agg == EV->getAggregateOperand()) 2800 return Agg; 2801 } 2802 2803 return 0; 2804} 2805 2806Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val, 2807 ArrayRef<unsigned> Idxs, 2808 const DataLayout *TD, 2809 const TargetLibraryInfo *TLI, 2810 const DominatorTree *DT) { 2811 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query (TD, TLI, DT), 2812 RecursionLimit); 2813} 2814 2815/// SimplifyPHINode - See if we can fold the given phi. If not, returns null. 2816static Value *SimplifyPHINode(PHINode *PN, const Query &Q) { 2817 // If all of the PHI's incoming values are the same then replace the PHI node 2818 // with the common value. 2819 Value *CommonValue = 0; 2820 bool HasUndefInput = false; 2821 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 2822 Value *Incoming = PN->getIncomingValue(i); 2823 // If the incoming value is the phi node itself, it can safely be skipped. 2824 if (Incoming == PN) continue; 2825 if (isa<UndefValue>(Incoming)) { 2826 // Remember that we saw an undef value, but otherwise ignore them. 2827 HasUndefInput = true; 2828 continue; 2829 } 2830 if (CommonValue && Incoming != CommonValue) 2831 return 0; // Not the same, bail out. 2832 CommonValue = Incoming; 2833 } 2834 2835 // If CommonValue is null then all of the incoming values were either undef or 2836 // equal to the phi node itself. 2837 if (!CommonValue) 2838 return UndefValue::get(PN->getType()); 2839 2840 // If we have a PHI node like phi(X, undef, X), where X is defined by some 2841 // instruction, we cannot return X as the result of the PHI node unless it 2842 // dominates the PHI block. 2843 if (HasUndefInput) 2844 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : 0; 2845 2846 return CommonValue; 2847} 2848 2849static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) { 2850 if (Constant *C = dyn_cast<Constant>(Op)) 2851 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.TD, Q.TLI); 2852 2853 return 0; 2854} 2855 2856Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout *TD, 2857 const TargetLibraryInfo *TLI, 2858 const DominatorTree *DT) { 2859 return ::SimplifyTruncInst(Op, Ty, Query (TD, TLI, DT), RecursionLimit); 2860} 2861 2862//=== Helper functions for higher up the class hierarchy. 2863 2864/// SimplifyBinOp - Given operands for a BinaryOperator, see if we can 2865/// fold the result. If not, this returns null. 2866static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 2867 const Query &Q, unsigned MaxRecurse) { 2868 switch (Opcode) { 2869 case Instruction::Add: 2870 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 2871 Q, MaxRecurse); 2872 case Instruction::FAdd: 2873 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 2874 2875 case Instruction::Sub: 2876 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 2877 Q, MaxRecurse); 2878 case Instruction::FSub: 2879 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); 2880 2881 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse); 2882 case Instruction::FMul: 2883 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse); 2884 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse); 2885 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse); 2886 case Instruction::FDiv: return SimplifyFDivInst(LHS, RHS, Q, MaxRecurse); 2887 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse); 2888 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse); 2889 case Instruction::FRem: return SimplifyFRemInst(LHS, RHS, Q, MaxRecurse); 2890 case Instruction::Shl: 2891 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false, 2892 Q, MaxRecurse); 2893 case Instruction::LShr: 2894 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 2895 case Instruction::AShr: 2896 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse); 2897 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse); 2898 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse); 2899 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse); 2900 default: 2901 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 2902 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 2903 Constant *COps[] = {CLHS, CRHS}; 2904 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.TD, 2905 Q.TLI); 2906 } 2907 2908 // If the operation is associative, try some generic simplifications. 2909 if (Instruction::isAssociative(Opcode)) 2910 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse)) 2911 return V; 2912 2913 // If the operation is with the result of a select instruction check whether 2914 // operating on either branch of the select always yields the same value. 2915 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) 2916 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse)) 2917 return V; 2918 2919 // If the operation is with the result of a phi instruction, check whether 2920 // operating on all incoming values of the phi always yields the same value. 2921 if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) 2922 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse)) 2923 return V; 2924 2925 return 0; 2926 } 2927} 2928 2929Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, 2930 const DataLayout *TD, const TargetLibraryInfo *TLI, 2931 const DominatorTree *DT) { 2932 return ::SimplifyBinOp(Opcode, LHS, RHS, Query (TD, TLI, DT), RecursionLimit); 2933} 2934 2935/// SimplifyCmpInst - Given operands for a CmpInst, see if we can 2936/// fold the result. 2937static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2938 const Query &Q, unsigned MaxRecurse) { 2939 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate)) 2940 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 2941 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse); 2942} 2943 2944Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, 2945 const DataLayout *TD, const TargetLibraryInfo *TLI, 2946 const DominatorTree *DT) { 2947 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query (TD, TLI, DT), 2948 RecursionLimit); 2949} 2950 2951static bool IsIdempotent(Intrinsic::ID ID) { 2952 switch (ID) { 2953 default: return false; 2954 2955 // Unary idempotent: f(f(x)) = f(x) 2956 case Intrinsic::fabs: 2957 case Intrinsic::floor: 2958 case Intrinsic::ceil: 2959 case Intrinsic::trunc: 2960 case Intrinsic::rint: 2961 case Intrinsic::nearbyint: 2962 case Intrinsic::round: 2963 return true; 2964 } 2965} 2966 2967template <typename IterTy> 2968static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd, 2969 const Query &Q, unsigned MaxRecurse) { 2970 // Perform idempotent optimizations 2971 if (!IsIdempotent(IID)) 2972 return 0; 2973 2974 // Unary Ops 2975 if (std::distance(ArgBegin, ArgEnd) == 1) 2976 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) 2977 if (II->getIntrinsicID() == IID) 2978 return II; 2979 2980 return 0; 2981} 2982 2983template <typename IterTy> 2984static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd, 2985 const Query &Q, unsigned MaxRecurse) { 2986 Type *Ty = V->getType(); 2987 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 2988 Ty = PTy->getElementType(); 2989 FunctionType *FTy = cast<FunctionType>(Ty); 2990 2991 // call undef -> undef 2992 if (isa<UndefValue>(V)) 2993 return UndefValue::get(FTy->getReturnType()); 2994 2995 Function *F = dyn_cast<Function>(V); 2996 if (!F) 2997 return 0; 2998 2999 if (unsigned IID = F->getIntrinsicID()) 3000 if (Value *Ret = 3001 SimplifyIntrinsic((Intrinsic::ID) IID, ArgBegin, ArgEnd, Q, MaxRecurse)) 3002 return Ret; 3003 3004 if (!canConstantFoldCallTo(F)) 3005 return 0; 3006 3007 SmallVector<Constant *, 4> ConstantArgs; 3008 ConstantArgs.reserve(ArgEnd - ArgBegin); 3009 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) { 3010 Constant *C = dyn_cast<Constant>(*I); 3011 if (!C) 3012 return 0; 3013 ConstantArgs.push_back(C); 3014 } 3015 3016 return ConstantFoldCall(F, ConstantArgs, Q.TLI); 3017} 3018 3019Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin, 3020 User::op_iterator ArgEnd, const DataLayout *TD, 3021 const TargetLibraryInfo *TLI, 3022 const DominatorTree *DT) { 3023 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(TD, TLI, DT), 3024 RecursionLimit); 3025} 3026 3027Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args, 3028 const DataLayout *TD, const TargetLibraryInfo *TLI, 3029 const DominatorTree *DT) { 3030 return ::SimplifyCall(V, Args.begin(), Args.end(), Query(TD, TLI, DT), 3031 RecursionLimit); 3032} 3033 3034/// SimplifyInstruction - See if we can compute a simplified version of this 3035/// instruction. If not, this returns null. 3036Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout *TD, 3037 const TargetLibraryInfo *TLI, 3038 const DominatorTree *DT) { 3039 Value *Result; 3040 3041 switch (I->getOpcode()) { 3042 default: 3043 Result = ConstantFoldInstruction(I, TD, TLI); 3044 break; 3045 case Instruction::FAdd: 3046 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1), 3047 I->getFastMathFlags(), TD, TLI, DT); 3048 break; 3049 case Instruction::Add: 3050 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1), 3051 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3052 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), 3053 TD, TLI, DT); 3054 break; 3055 case Instruction::FSub: 3056 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1), 3057 I->getFastMathFlags(), TD, TLI, DT); 3058 break; 3059 case Instruction::Sub: 3060 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1), 3061 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3062 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), 3063 TD, TLI, DT); 3064 break; 3065 case Instruction::FMul: 3066 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1), 3067 I->getFastMathFlags(), TD, TLI, DT); 3068 break; 3069 case Instruction::Mul: 3070 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3071 break; 3072 case Instruction::SDiv: 3073 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3074 break; 3075 case Instruction::UDiv: 3076 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3077 break; 3078 case Instruction::FDiv: 3079 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3080 break; 3081 case Instruction::SRem: 3082 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3083 break; 3084 case Instruction::URem: 3085 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3086 break; 3087 case Instruction::FRem: 3088 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3089 break; 3090 case Instruction::Shl: 3091 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1), 3092 cast<BinaryOperator>(I)->hasNoSignedWrap(), 3093 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), 3094 TD, TLI, DT); 3095 break; 3096 case Instruction::LShr: 3097 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1), 3098 cast<BinaryOperator>(I)->isExact(), 3099 TD, TLI, DT); 3100 break; 3101 case Instruction::AShr: 3102 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1), 3103 cast<BinaryOperator>(I)->isExact(), 3104 TD, TLI, DT); 3105 break; 3106 case Instruction::And: 3107 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3108 break; 3109 case Instruction::Or: 3110 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3111 break; 3112 case Instruction::Xor: 3113 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3114 break; 3115 case Instruction::ICmp: 3116 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), 3117 I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3118 break; 3119 case Instruction::FCmp: 3120 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), 3121 I->getOperand(0), I->getOperand(1), TD, TLI, DT); 3122 break; 3123 case Instruction::Select: 3124 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1), 3125 I->getOperand(2), TD, TLI, DT); 3126 break; 3127 case Instruction::GetElementPtr: { 3128 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 3129 Result = SimplifyGEPInst(Ops, TD, TLI, DT); 3130 break; 3131 } 3132 case Instruction::InsertValue: { 3133 InsertValueInst *IV = cast<InsertValueInst>(I); 3134 Result = SimplifyInsertValueInst(IV->getAggregateOperand(), 3135 IV->getInsertedValueOperand(), 3136 IV->getIndices(), TD, TLI, DT); 3137 break; 3138 } 3139 case Instruction::PHI: 3140 Result = SimplifyPHINode(cast<PHINode>(I), Query (TD, TLI, DT)); 3141 break; 3142 case Instruction::Call: { 3143 CallSite CS(cast<CallInst>(I)); 3144 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), 3145 TD, TLI, DT); 3146 break; 3147 } 3148 case Instruction::Trunc: 3149 Result = SimplifyTruncInst(I->getOperand(0), I->getType(), TD, TLI, DT); 3150 break; 3151 } 3152 3153 /// If called on unreachable code, the above logic may report that the 3154 /// instruction simplified to itself. Make life easier for users by 3155 /// detecting that case here, returning a safe value instead. 3156 return Result == I ? UndefValue::get(I->getType()) : Result; 3157} 3158 3159/// \brief Implementation of recursive simplification through an instructions 3160/// uses. 3161/// 3162/// This is the common implementation of the recursive simplification routines. 3163/// If we have a pre-simplified value in 'SimpleV', that is forcibly used to 3164/// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of 3165/// instructions to process and attempt to simplify it using 3166/// InstructionSimplify. 3167/// 3168/// This routine returns 'true' only when *it* simplifies something. The passed 3169/// in simplified value does not count toward this. 3170static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV, 3171 const DataLayout *TD, 3172 const TargetLibraryInfo *TLI, 3173 const DominatorTree *DT) { 3174 bool Simplified = false; 3175 SmallSetVector<Instruction *, 8> Worklist; 3176 3177 // If we have an explicit value to collapse to, do that round of the 3178 // simplification loop by hand initially. 3179 if (SimpleV) { 3180 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE; 3181 ++UI) 3182 if (*UI != I) 3183 Worklist.insert(cast<Instruction>(*UI)); 3184 3185 // Replace the instruction with its simplified value. 3186 I->replaceAllUsesWith(SimpleV); 3187 3188 // Gracefully handle edge cases where the instruction is not wired into any 3189 // parent block. 3190 if (I->getParent()) 3191 I->eraseFromParent(); 3192 } else { 3193 Worklist.insert(I); 3194 } 3195 3196 // Note that we must test the size on each iteration, the worklist can grow. 3197 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { 3198 I = Worklist[Idx]; 3199 3200 // See if this instruction simplifies. 3201 SimpleV = SimplifyInstruction(I, TD, TLI, DT); 3202 if (!SimpleV) 3203 continue; 3204 3205 Simplified = true; 3206 3207 // Stash away all the uses of the old instruction so we can check them for 3208 // recursive simplifications after a RAUW. This is cheaper than checking all 3209 // uses of To on the recursive step in most cases. 3210 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE; 3211 ++UI) 3212 Worklist.insert(cast<Instruction>(*UI)); 3213 3214 // Replace the instruction with its simplified value. 3215 I->replaceAllUsesWith(SimpleV); 3216 3217 // Gracefully handle edge cases where the instruction is not wired into any 3218 // parent block. 3219 if (I->getParent()) 3220 I->eraseFromParent(); 3221 } 3222 return Simplified; 3223} 3224 3225bool llvm::recursivelySimplifyInstruction(Instruction *I, 3226 const DataLayout *TD, 3227 const TargetLibraryInfo *TLI, 3228 const DominatorTree *DT) { 3229 return replaceAndRecursivelySimplifyImpl(I, 0, TD, TLI, DT); 3230} 3231 3232bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV, 3233 const DataLayout *TD, 3234 const TargetLibraryInfo *TLI, 3235 const DominatorTree *DT) { 3236 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!"); 3237 assert(SimpleV && "Must provide a simplified value."); 3238 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TD, TLI, DT); 3239} 3240