1 //===-- DependenceAnalysis.cpp - DA Implementation --------------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // DependenceAnalysis is an LLVM pass that analyses dependences between memory 10 // accesses. Currently, it is an (incomplete) implementation of the approach 11 // described in 12 // 13 // Practical Dependence Testing 14 // Goff, Kennedy, Tseng 15 // PLDI 1991 16 // 17 // There's a single entry point that analyzes the dependence between a pair 18 // of memory references in a function, returning either NULL, for no dependence, 19 // or a more-or-less detailed description of the dependence between them. 20 // 21 // Currently, the implementation cannot propagate constraints between 22 // coupled RDIV subscripts and lacks a multi-subscript MIV test. 23 // Both of these are conservative weaknesses; 24 // that is, not a source of correctness problems. 25 // 26 // Since Clang linearizes some array subscripts, the dependence 27 // analysis is using SCEV->delinearize to recover the representation of multiple 28 // subscripts, and thus avoid the more expensive and less precise MIV tests. The 29 // delinearization is controlled by the flag -da-delinearize. 30 // 31 // We should pay some careful attention to the possibility of integer overflow 32 // in the implementation of the various tests. This could happen with Add, 33 // Subtract, or Multiply, with both APInt's and SCEV's. 34 // 35 // Some non-linear subscript pairs can be handled by the GCD test 36 // (and perhaps other tests). 37 // Should explore how often these things occur. 38 // 39 // Finally, it seems like certain test cases expose weaknesses in the SCEV 40 // simplification, especially in the handling of sign and zero extensions. 41 // It could be useful to spend time exploring these. 42 // 43 // Please note that this is work in progress and the interface is subject to 44 // change. 45 // 46 //===----------------------------------------------------------------------===// 47 // // 48 // In memory of Ken Kennedy, 1945 - 2007 // 49 // // 50 //===----------------------------------------------------------------------===// 51 52 #include "llvm/Analysis/DependenceAnalysis.h" 53 #include "llvm/ADT/Statistic.h" 54 #include "llvm/Analysis/AliasAnalysis.h" 55 #include "llvm/Analysis/Delinearization.h" 56 #include "llvm/Analysis/LoopInfo.h" 57 #include "llvm/Analysis/ScalarEvolution.h" 58 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 59 #include "llvm/Analysis/ValueTracking.h" 60 #include "llvm/IR/InstIterator.h" 61 #include "llvm/IR/Module.h" 62 #include "llvm/InitializePasses.h" 63 #include "llvm/Support/CommandLine.h" 64 #include "llvm/Support/Debug.h" 65 #include "llvm/Support/ErrorHandling.h" 66 #include "llvm/Support/raw_ostream.h" 67 68 using namespace llvm; 69 70 #define DEBUG_TYPE "da" 71 72 //===----------------------------------------------------------------------===// 73 // statistics 74 75 STATISTIC(TotalArrayPairs, "Array pairs tested"); 76 STATISTIC(SeparableSubscriptPairs, "Separable subscript pairs"); 77 STATISTIC(CoupledSubscriptPairs, "Coupled subscript pairs"); 78 STATISTIC(NonlinearSubscriptPairs, "Nonlinear subscript pairs"); 79 STATISTIC(ZIVapplications, "ZIV applications"); 80 STATISTIC(ZIVindependence, "ZIV independence"); 81 STATISTIC(StrongSIVapplications, "Strong SIV applications"); 82 STATISTIC(StrongSIVsuccesses, "Strong SIV successes"); 83 STATISTIC(StrongSIVindependence, "Strong SIV independence"); 84 STATISTIC(WeakCrossingSIVapplications, "Weak-Crossing SIV applications"); 85 STATISTIC(WeakCrossingSIVsuccesses, "Weak-Crossing SIV successes"); 86 STATISTIC(WeakCrossingSIVindependence, "Weak-Crossing SIV independence"); 87 STATISTIC(ExactSIVapplications, "Exact SIV applications"); 88 STATISTIC(ExactSIVsuccesses, "Exact SIV successes"); 89 STATISTIC(ExactSIVindependence, "Exact SIV independence"); 90 STATISTIC(WeakZeroSIVapplications, "Weak-Zero SIV applications"); 91 STATISTIC(WeakZeroSIVsuccesses, "Weak-Zero SIV successes"); 92 STATISTIC(WeakZeroSIVindependence, "Weak-Zero SIV independence"); 93 STATISTIC(ExactRDIVapplications, "Exact RDIV applications"); 94 STATISTIC(ExactRDIVindependence, "Exact RDIV independence"); 95 STATISTIC(SymbolicRDIVapplications, "Symbolic RDIV applications"); 96 STATISTIC(SymbolicRDIVindependence, "Symbolic RDIV independence"); 97 STATISTIC(DeltaApplications, "Delta applications"); 98 STATISTIC(DeltaSuccesses, "Delta successes"); 99 STATISTIC(DeltaIndependence, "Delta independence"); 100 STATISTIC(DeltaPropagations, "Delta propagations"); 101 STATISTIC(GCDapplications, "GCD applications"); 102 STATISTIC(GCDsuccesses, "GCD successes"); 103 STATISTIC(GCDindependence, "GCD independence"); 104 STATISTIC(BanerjeeApplications, "Banerjee applications"); 105 STATISTIC(BanerjeeIndependence, "Banerjee independence"); 106 STATISTIC(BanerjeeSuccesses, "Banerjee successes"); 107 108 static cl::opt<bool> 109 Delinearize("da-delinearize", cl::init(true), cl::Hidden, 110 cl::desc("Try to delinearize array references.")); 111 static cl::opt<bool> DisableDelinearizationChecks( 112 "da-disable-delinearization-checks", cl::Hidden, 113 cl::desc( 114 "Disable checks that try to statically verify validity of " 115 "delinearized subscripts. Enabling this option may result in incorrect " 116 "dependence vectors for languages that allow the subscript of one " 117 "dimension to underflow or overflow into another dimension.")); 118 119 static cl::opt<unsigned> MIVMaxLevelThreshold( 120 "da-miv-max-level-threshold", cl::init(7), cl::Hidden, 121 cl::desc("Maximum depth allowed for the recursive algorithm used to " 122 "explore MIV direction vectors.")); 123 124 //===----------------------------------------------------------------------===// 125 // basics 126 127 DependenceAnalysis::Result 128 DependenceAnalysis::run(Function &F, FunctionAnalysisManager &FAM) { 129 auto &AA = FAM.getResult<AAManager>(F); 130 auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F); 131 auto &LI = FAM.getResult<LoopAnalysis>(F); 132 return DependenceInfo(&F, &AA, &SE, &LI); 133 } 134 135 AnalysisKey DependenceAnalysis::Key; 136 137 INITIALIZE_PASS_BEGIN(DependenceAnalysisWrapperPass, "da", 138 "Dependence Analysis", true, true) 139 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 140 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 141 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 142 INITIALIZE_PASS_END(DependenceAnalysisWrapperPass, "da", "Dependence Analysis", 143 true, true) 144 145 char DependenceAnalysisWrapperPass::ID = 0; 146 147 DependenceAnalysisWrapperPass::DependenceAnalysisWrapperPass() 148 : FunctionPass(ID) { 149 initializeDependenceAnalysisWrapperPassPass(*PassRegistry::getPassRegistry()); 150 } 151 152 FunctionPass *llvm::createDependenceAnalysisWrapperPass() { 153 return new DependenceAnalysisWrapperPass(); 154 } 155 156 bool DependenceAnalysisWrapperPass::runOnFunction(Function &F) { 157 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); 158 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 159 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 160 info.reset(new DependenceInfo(&F, &AA, &SE, &LI)); 161 return false; 162 } 163 164 DependenceInfo &DependenceAnalysisWrapperPass::getDI() const { return *info; } 165 166 void DependenceAnalysisWrapperPass::releaseMemory() { info.reset(); } 167 168 void DependenceAnalysisWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 169 AU.setPreservesAll(); 170 AU.addRequiredTransitive<AAResultsWrapperPass>(); 171 AU.addRequiredTransitive<ScalarEvolutionWrapperPass>(); 172 AU.addRequiredTransitive<LoopInfoWrapperPass>(); 173 } 174 175 // Used to test the dependence analyzer. 176 // Looks through the function, noting instructions that may access memory. 177 // Calls depends() on every possible pair and prints out the result. 178 // Ignores all other instructions. 179 static void dumpExampleDependence(raw_ostream &OS, DependenceInfo *DA) { 180 auto *F = DA->getFunction(); 181 for (inst_iterator SrcI = inst_begin(F), SrcE = inst_end(F); SrcI != SrcE; 182 ++SrcI) { 183 if (SrcI->mayReadOrWriteMemory()) { 184 for (inst_iterator DstI = SrcI, DstE = inst_end(F); 185 DstI != DstE; ++DstI) { 186 if (DstI->mayReadOrWriteMemory()) { 187 OS << "Src:" << *SrcI << " --> Dst:" << *DstI << "\n"; 188 OS << " da analyze - "; 189 if (auto D = DA->depends(&*SrcI, &*DstI, true)) { 190 D->dump(OS); 191 for (unsigned Level = 1; Level <= D->getLevels(); Level++) { 192 if (D->isSplitable(Level)) { 193 OS << " da analyze - split level = " << Level; 194 OS << ", iteration = " << *DA->getSplitIteration(*D, Level); 195 OS << "!\n"; 196 } 197 } 198 } 199 else 200 OS << "none!\n"; 201 } 202 } 203 } 204 } 205 } 206 207 void DependenceAnalysisWrapperPass::print(raw_ostream &OS, 208 const Module *) const { 209 dumpExampleDependence(OS, info.get()); 210 } 211 212 PreservedAnalyses 213 DependenceAnalysisPrinterPass::run(Function &F, FunctionAnalysisManager &FAM) { 214 OS << "'Dependence Analysis' for function '" << F.getName() << "':\n"; 215 dumpExampleDependence(OS, &FAM.getResult<DependenceAnalysis>(F)); 216 return PreservedAnalyses::all(); 217 } 218 219 //===----------------------------------------------------------------------===// 220 // Dependence methods 221 222 // Returns true if this is an input dependence. 223 bool Dependence::isInput() const { 224 return Src->mayReadFromMemory() && Dst->mayReadFromMemory(); 225 } 226 227 228 // Returns true if this is an output dependence. 229 bool Dependence::isOutput() const { 230 return Src->mayWriteToMemory() && Dst->mayWriteToMemory(); 231 } 232 233 234 // Returns true if this is an flow (aka true) dependence. 235 bool Dependence::isFlow() const { 236 return Src->mayWriteToMemory() && Dst->mayReadFromMemory(); 237 } 238 239 240 // Returns true if this is an anti dependence. 241 bool Dependence::isAnti() const { 242 return Src->mayReadFromMemory() && Dst->mayWriteToMemory(); 243 } 244 245 246 // Returns true if a particular level is scalar; that is, 247 // if no subscript in the source or destination mention the induction 248 // variable associated with the loop at this level. 249 // Leave this out of line, so it will serve as a virtual method anchor 250 bool Dependence::isScalar(unsigned level) const { 251 return false; 252 } 253 254 255 //===----------------------------------------------------------------------===// 256 // FullDependence methods 257 258 FullDependence::FullDependence(Instruction *Source, Instruction *Destination, 259 bool PossiblyLoopIndependent, 260 unsigned CommonLevels) 261 : Dependence(Source, Destination), Levels(CommonLevels), 262 LoopIndependent(PossiblyLoopIndependent) { 263 Consistent = true; 264 if (CommonLevels) 265 DV = std::make_unique<DVEntry[]>(CommonLevels); 266 } 267 268 // The rest are simple getters that hide the implementation. 269 270 // getDirection - Returns the direction associated with a particular level. 271 unsigned FullDependence::getDirection(unsigned Level) const { 272 assert(0 < Level && Level <= Levels && "Level out of range"); 273 return DV[Level - 1].Direction; 274 } 275 276 277 // Returns the distance (or NULL) associated with a particular level. 278 const SCEV *FullDependence::getDistance(unsigned Level) const { 279 assert(0 < Level && Level <= Levels && "Level out of range"); 280 return DV[Level - 1].Distance; 281 } 282 283 284 // Returns true if a particular level is scalar; that is, 285 // if no subscript in the source or destination mention the induction 286 // variable associated with the loop at this level. 287 bool FullDependence::isScalar(unsigned Level) const { 288 assert(0 < Level && Level <= Levels && "Level out of range"); 289 return DV[Level - 1].Scalar; 290 } 291 292 293 // Returns true if peeling the first iteration from this loop 294 // will break this dependence. 295 bool FullDependence::isPeelFirst(unsigned Level) const { 296 assert(0 < Level && Level <= Levels && "Level out of range"); 297 return DV[Level - 1].PeelFirst; 298 } 299 300 301 // Returns true if peeling the last iteration from this loop 302 // will break this dependence. 303 bool FullDependence::isPeelLast(unsigned Level) const { 304 assert(0 < Level && Level <= Levels && "Level out of range"); 305 return DV[Level - 1].PeelLast; 306 } 307 308 309 // Returns true if splitting this loop will break the dependence. 310 bool FullDependence::isSplitable(unsigned Level) const { 311 assert(0 < Level && Level <= Levels && "Level out of range"); 312 return DV[Level - 1].Splitable; 313 } 314 315 316 //===----------------------------------------------------------------------===// 317 // DependenceInfo::Constraint methods 318 319 // If constraint is a point <X, Y>, returns X. 320 // Otherwise assert. 321 const SCEV *DependenceInfo::Constraint::getX() const { 322 assert(Kind == Point && "Kind should be Point"); 323 return A; 324 } 325 326 327 // If constraint is a point <X, Y>, returns Y. 328 // Otherwise assert. 329 const SCEV *DependenceInfo::Constraint::getY() const { 330 assert(Kind == Point && "Kind should be Point"); 331 return B; 332 } 333 334 335 // If constraint is a line AX + BY = C, returns A. 336 // Otherwise assert. 337 const SCEV *DependenceInfo::Constraint::getA() const { 338 assert((Kind == Line || Kind == Distance) && 339 "Kind should be Line (or Distance)"); 340 return A; 341 } 342 343 344 // If constraint is a line AX + BY = C, returns B. 345 // Otherwise assert. 346 const SCEV *DependenceInfo::Constraint::getB() const { 347 assert((Kind == Line || Kind == Distance) && 348 "Kind should be Line (or Distance)"); 349 return B; 350 } 351 352 353 // If constraint is a line AX + BY = C, returns C. 354 // Otherwise assert. 355 const SCEV *DependenceInfo::Constraint::getC() const { 356 assert((Kind == Line || Kind == Distance) && 357 "Kind should be Line (or Distance)"); 358 return C; 359 } 360 361 362 // If constraint is a distance, returns D. 363 // Otherwise assert. 364 const SCEV *DependenceInfo::Constraint::getD() const { 365 assert(Kind == Distance && "Kind should be Distance"); 366 return SE->getNegativeSCEV(C); 367 } 368 369 370 // Returns the loop associated with this constraint. 371 const Loop *DependenceInfo::Constraint::getAssociatedLoop() const { 372 assert((Kind == Distance || Kind == Line || Kind == Point) && 373 "Kind should be Distance, Line, or Point"); 374 return AssociatedLoop; 375 } 376 377 void DependenceInfo::Constraint::setPoint(const SCEV *X, const SCEV *Y, 378 const Loop *CurLoop) { 379 Kind = Point; 380 A = X; 381 B = Y; 382 AssociatedLoop = CurLoop; 383 } 384 385 void DependenceInfo::Constraint::setLine(const SCEV *AA, const SCEV *BB, 386 const SCEV *CC, const Loop *CurLoop) { 387 Kind = Line; 388 A = AA; 389 B = BB; 390 C = CC; 391 AssociatedLoop = CurLoop; 392 } 393 394 void DependenceInfo::Constraint::setDistance(const SCEV *D, 395 const Loop *CurLoop) { 396 Kind = Distance; 397 A = SE->getOne(D->getType()); 398 B = SE->getNegativeSCEV(A); 399 C = SE->getNegativeSCEV(D); 400 AssociatedLoop = CurLoop; 401 } 402 403 void DependenceInfo::Constraint::setEmpty() { Kind = Empty; } 404 405 void DependenceInfo::Constraint::setAny(ScalarEvolution *NewSE) { 406 SE = NewSE; 407 Kind = Any; 408 } 409 410 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 411 // For debugging purposes. Dumps the constraint out to OS. 412 LLVM_DUMP_METHOD void DependenceInfo::Constraint::dump(raw_ostream &OS) const { 413 if (isEmpty()) 414 OS << " Empty\n"; 415 else if (isAny()) 416 OS << " Any\n"; 417 else if (isPoint()) 418 OS << " Point is <" << *getX() << ", " << *getY() << ">\n"; 419 else if (isDistance()) 420 OS << " Distance is " << *getD() << 421 " (" << *getA() << "*X + " << *getB() << "*Y = " << *getC() << ")\n"; 422 else if (isLine()) 423 OS << " Line is " << *getA() << "*X + " << 424 *getB() << "*Y = " << *getC() << "\n"; 425 else 426 llvm_unreachable("unknown constraint type in Constraint::dump"); 427 } 428 #endif 429 430 431 // Updates X with the intersection 432 // of the Constraints X and Y. Returns true if X has changed. 433 // Corresponds to Figure 4 from the paper 434 // 435 // Practical Dependence Testing 436 // Goff, Kennedy, Tseng 437 // PLDI 1991 438 bool DependenceInfo::intersectConstraints(Constraint *X, const Constraint *Y) { 439 ++DeltaApplications; 440 LLVM_DEBUG(dbgs() << "\tintersect constraints\n"); 441 LLVM_DEBUG(dbgs() << "\t X ="; X->dump(dbgs())); 442 LLVM_DEBUG(dbgs() << "\t Y ="; Y->dump(dbgs())); 443 assert(!Y->isPoint() && "Y must not be a Point"); 444 if (X->isAny()) { 445 if (Y->isAny()) 446 return false; 447 *X = *Y; 448 return true; 449 } 450 if (X->isEmpty()) 451 return false; 452 if (Y->isEmpty()) { 453 X->setEmpty(); 454 return true; 455 } 456 457 if (X->isDistance() && Y->isDistance()) { 458 LLVM_DEBUG(dbgs() << "\t intersect 2 distances\n"); 459 if (isKnownPredicate(CmpInst::ICMP_EQ, X->getD(), Y->getD())) 460 return false; 461 if (isKnownPredicate(CmpInst::ICMP_NE, X->getD(), Y->getD())) { 462 X->setEmpty(); 463 ++DeltaSuccesses; 464 return true; 465 } 466 // Hmmm, interesting situation. 467 // I guess if either is constant, keep it and ignore the other. 468 if (isa<SCEVConstant>(Y->getD())) { 469 *X = *Y; 470 return true; 471 } 472 return false; 473 } 474 475 // At this point, the pseudo-code in Figure 4 of the paper 476 // checks if (X->isPoint() && Y->isPoint()). 477 // This case can't occur in our implementation, 478 // since a Point can only arise as the result of intersecting 479 // two Line constraints, and the right-hand value, Y, is never 480 // the result of an intersection. 481 assert(!(X->isPoint() && Y->isPoint()) && 482 "We shouldn't ever see X->isPoint() && Y->isPoint()"); 483 484 if (X->isLine() && Y->isLine()) { 485 LLVM_DEBUG(dbgs() << "\t intersect 2 lines\n"); 486 const SCEV *Prod1 = SE->getMulExpr(X->getA(), Y->getB()); 487 const SCEV *Prod2 = SE->getMulExpr(X->getB(), Y->getA()); 488 if (isKnownPredicate(CmpInst::ICMP_EQ, Prod1, Prod2)) { 489 // slopes are equal, so lines are parallel 490 LLVM_DEBUG(dbgs() << "\t\tsame slope\n"); 491 Prod1 = SE->getMulExpr(X->getC(), Y->getB()); 492 Prod2 = SE->getMulExpr(X->getB(), Y->getC()); 493 if (isKnownPredicate(CmpInst::ICMP_EQ, Prod1, Prod2)) 494 return false; 495 if (isKnownPredicate(CmpInst::ICMP_NE, Prod1, Prod2)) { 496 X->setEmpty(); 497 ++DeltaSuccesses; 498 return true; 499 } 500 return false; 501 } 502 if (isKnownPredicate(CmpInst::ICMP_NE, Prod1, Prod2)) { 503 // slopes differ, so lines intersect 504 LLVM_DEBUG(dbgs() << "\t\tdifferent slopes\n"); 505 const SCEV *C1B2 = SE->getMulExpr(X->getC(), Y->getB()); 506 const SCEV *C1A2 = SE->getMulExpr(X->getC(), Y->getA()); 507 const SCEV *C2B1 = SE->getMulExpr(Y->getC(), X->getB()); 508 const SCEV *C2A1 = SE->getMulExpr(Y->getC(), X->getA()); 509 const SCEV *A1B2 = SE->getMulExpr(X->getA(), Y->getB()); 510 const SCEV *A2B1 = SE->getMulExpr(Y->getA(), X->getB()); 511 const SCEVConstant *C1A2_C2A1 = 512 dyn_cast<SCEVConstant>(SE->getMinusSCEV(C1A2, C2A1)); 513 const SCEVConstant *C1B2_C2B1 = 514 dyn_cast<SCEVConstant>(SE->getMinusSCEV(C1B2, C2B1)); 515 const SCEVConstant *A1B2_A2B1 = 516 dyn_cast<SCEVConstant>(SE->getMinusSCEV(A1B2, A2B1)); 517 const SCEVConstant *A2B1_A1B2 = 518 dyn_cast<SCEVConstant>(SE->getMinusSCEV(A2B1, A1B2)); 519 if (!C1B2_C2B1 || !C1A2_C2A1 || 520 !A1B2_A2B1 || !A2B1_A1B2) 521 return false; 522 APInt Xtop = C1B2_C2B1->getAPInt(); 523 APInt Xbot = A1B2_A2B1->getAPInt(); 524 APInt Ytop = C1A2_C2A1->getAPInt(); 525 APInt Ybot = A2B1_A1B2->getAPInt(); 526 LLVM_DEBUG(dbgs() << "\t\tXtop = " << Xtop << "\n"); 527 LLVM_DEBUG(dbgs() << "\t\tXbot = " << Xbot << "\n"); 528 LLVM_DEBUG(dbgs() << "\t\tYtop = " << Ytop << "\n"); 529 LLVM_DEBUG(dbgs() << "\t\tYbot = " << Ybot << "\n"); 530 APInt Xq = Xtop; // these need to be initialized, even 531 APInt Xr = Xtop; // though they're just going to be overwritten 532 APInt::sdivrem(Xtop, Xbot, Xq, Xr); 533 APInt Yq = Ytop; 534 APInt Yr = Ytop; 535 APInt::sdivrem(Ytop, Ybot, Yq, Yr); 536 if (Xr != 0 || Yr != 0) { 537 X->setEmpty(); 538 ++DeltaSuccesses; 539 return true; 540 } 541 LLVM_DEBUG(dbgs() << "\t\tX = " << Xq << ", Y = " << Yq << "\n"); 542 if (Xq.slt(0) || Yq.slt(0)) { 543 X->setEmpty(); 544 ++DeltaSuccesses; 545 return true; 546 } 547 if (const SCEVConstant *CUB = 548 collectConstantUpperBound(X->getAssociatedLoop(), Prod1->getType())) { 549 const APInt &UpperBound = CUB->getAPInt(); 550 LLVM_DEBUG(dbgs() << "\t\tupper bound = " << UpperBound << "\n"); 551 if (Xq.sgt(UpperBound) || Yq.sgt(UpperBound)) { 552 X->setEmpty(); 553 ++DeltaSuccesses; 554 return true; 555 } 556 } 557 X->setPoint(SE->getConstant(Xq), 558 SE->getConstant(Yq), 559 X->getAssociatedLoop()); 560 ++DeltaSuccesses; 561 return true; 562 } 563 return false; 564 } 565 566 // if (X->isLine() && Y->isPoint()) This case can't occur. 567 assert(!(X->isLine() && Y->isPoint()) && "This case should never occur"); 568 569 if (X->isPoint() && Y->isLine()) { 570 LLVM_DEBUG(dbgs() << "\t intersect Point and Line\n"); 571 const SCEV *A1X1 = SE->getMulExpr(Y->getA(), X->getX()); 572 const SCEV *B1Y1 = SE->getMulExpr(Y->getB(), X->getY()); 573 const SCEV *Sum = SE->getAddExpr(A1X1, B1Y1); 574 if (isKnownPredicate(CmpInst::ICMP_EQ, Sum, Y->getC())) 575 return false; 576 if (isKnownPredicate(CmpInst::ICMP_NE, Sum, Y->getC())) { 577 X->setEmpty(); 578 ++DeltaSuccesses; 579 return true; 580 } 581 return false; 582 } 583 584 llvm_unreachable("shouldn't reach the end of Constraint intersection"); 585 return false; 586 } 587 588 589 //===----------------------------------------------------------------------===// 590 // DependenceInfo methods 591 592 // For debugging purposes. Dumps a dependence to OS. 593 void Dependence::dump(raw_ostream &OS) const { 594 bool Splitable = false; 595 if (isConfused()) 596 OS << "confused"; 597 else { 598 if (isConsistent()) 599 OS << "consistent "; 600 if (isFlow()) 601 OS << "flow"; 602 else if (isOutput()) 603 OS << "output"; 604 else if (isAnti()) 605 OS << "anti"; 606 else if (isInput()) 607 OS << "input"; 608 unsigned Levels = getLevels(); 609 OS << " ["; 610 for (unsigned II = 1; II <= Levels; ++II) { 611 if (isSplitable(II)) 612 Splitable = true; 613 if (isPeelFirst(II)) 614 OS << 'p'; 615 const SCEV *Distance = getDistance(II); 616 if (Distance) 617 OS << *Distance; 618 else if (isScalar(II)) 619 OS << "S"; 620 else { 621 unsigned Direction = getDirection(II); 622 if (Direction == DVEntry::ALL) 623 OS << "*"; 624 else { 625 if (Direction & DVEntry::LT) 626 OS << "<"; 627 if (Direction & DVEntry::EQ) 628 OS << "="; 629 if (Direction & DVEntry::GT) 630 OS << ">"; 631 } 632 } 633 if (isPeelLast(II)) 634 OS << 'p'; 635 if (II < Levels) 636 OS << " "; 637 } 638 if (isLoopIndependent()) 639 OS << "|<"; 640 OS << "]"; 641 if (Splitable) 642 OS << " splitable"; 643 } 644 OS << "!\n"; 645 } 646 647 // Returns NoAlias/MayAliass/MustAlias for two memory locations based upon their 648 // underlaying objects. If LocA and LocB are known to not alias (for any reason: 649 // tbaa, non-overlapping regions etc), then it is known there is no dependecy. 650 // Otherwise the underlying objects are checked to see if they point to 651 // different identifiable objects. 652 static AliasResult underlyingObjectsAlias(AAResults *AA, 653 const DataLayout &DL, 654 const MemoryLocation &LocA, 655 const MemoryLocation &LocB) { 656 // Check the original locations (minus size) for noalias, which can happen for 657 // tbaa, incompatible underlying object locations, etc. 658 MemoryLocation LocAS = 659 MemoryLocation::getBeforeOrAfter(LocA.Ptr, LocA.AATags); 660 MemoryLocation LocBS = 661 MemoryLocation::getBeforeOrAfter(LocB.Ptr, LocB.AATags); 662 if (AA->isNoAlias(LocAS, LocBS)) 663 return AliasResult::NoAlias; 664 665 // Check the underlying objects are the same 666 const Value *AObj = getUnderlyingObject(LocA.Ptr); 667 const Value *BObj = getUnderlyingObject(LocB.Ptr); 668 669 // If the underlying objects are the same, they must alias 670 if (AObj == BObj) 671 return AliasResult::MustAlias; 672 673 // We may have hit the recursion limit for underlying objects, or have 674 // underlying objects where we don't know they will alias. 675 if (!isIdentifiedObject(AObj) || !isIdentifiedObject(BObj)) 676 return AliasResult::MayAlias; 677 678 // Otherwise we know the objects are different and both identified objects so 679 // must not alias. 680 return AliasResult::NoAlias; 681 } 682 683 684 // Returns true if the load or store can be analyzed. Atomic and volatile 685 // operations have properties which this analysis does not understand. 686 static 687 bool isLoadOrStore(const Instruction *I) { 688 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 689 return LI->isUnordered(); 690 else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) 691 return SI->isUnordered(); 692 return false; 693 } 694 695 696 // Examines the loop nesting of the Src and Dst 697 // instructions and establishes their shared loops. Sets the variables 698 // CommonLevels, SrcLevels, and MaxLevels. 699 // The source and destination instructions needn't be contained in the same 700 // loop. The routine establishNestingLevels finds the level of most deeply 701 // nested loop that contains them both, CommonLevels. An instruction that's 702 // not contained in a loop is at level = 0. MaxLevels is equal to the level 703 // of the source plus the level of the destination, minus CommonLevels. 704 // This lets us allocate vectors MaxLevels in length, with room for every 705 // distinct loop referenced in both the source and destination subscripts. 706 // The variable SrcLevels is the nesting depth of the source instruction. 707 // It's used to help calculate distinct loops referenced by the destination. 708 // Here's the map from loops to levels: 709 // 0 - unused 710 // 1 - outermost common loop 711 // ... - other common loops 712 // CommonLevels - innermost common loop 713 // ... - loops containing Src but not Dst 714 // SrcLevels - innermost loop containing Src but not Dst 715 // ... - loops containing Dst but not Src 716 // MaxLevels - innermost loops containing Dst but not Src 717 // Consider the follow code fragment: 718 // for (a = ...) { 719 // for (b = ...) { 720 // for (c = ...) { 721 // for (d = ...) { 722 // A[] = ...; 723 // } 724 // } 725 // for (e = ...) { 726 // for (f = ...) { 727 // for (g = ...) { 728 // ... = A[]; 729 // } 730 // } 731 // } 732 // } 733 // } 734 // If we're looking at the possibility of a dependence between the store 735 // to A (the Src) and the load from A (the Dst), we'll note that they 736 // have 2 loops in common, so CommonLevels will equal 2 and the direction 737 // vector for Result will have 2 entries. SrcLevels = 4 and MaxLevels = 7. 738 // A map from loop names to loop numbers would look like 739 // a - 1 740 // b - 2 = CommonLevels 741 // c - 3 742 // d - 4 = SrcLevels 743 // e - 5 744 // f - 6 745 // g - 7 = MaxLevels 746 void DependenceInfo::establishNestingLevels(const Instruction *Src, 747 const Instruction *Dst) { 748 const BasicBlock *SrcBlock = Src->getParent(); 749 const BasicBlock *DstBlock = Dst->getParent(); 750 unsigned SrcLevel = LI->getLoopDepth(SrcBlock); 751 unsigned DstLevel = LI->getLoopDepth(DstBlock); 752 const Loop *SrcLoop = LI->getLoopFor(SrcBlock); 753 const Loop *DstLoop = LI->getLoopFor(DstBlock); 754 SrcLevels = SrcLevel; 755 MaxLevels = SrcLevel + DstLevel; 756 while (SrcLevel > DstLevel) { 757 SrcLoop = SrcLoop->getParentLoop(); 758 SrcLevel--; 759 } 760 while (DstLevel > SrcLevel) { 761 DstLoop = DstLoop->getParentLoop(); 762 DstLevel--; 763 } 764 while (SrcLoop != DstLoop) { 765 SrcLoop = SrcLoop->getParentLoop(); 766 DstLoop = DstLoop->getParentLoop(); 767 SrcLevel--; 768 } 769 CommonLevels = SrcLevel; 770 MaxLevels -= CommonLevels; 771 } 772 773 774 // Given one of the loops containing the source, return 775 // its level index in our numbering scheme. 776 unsigned DependenceInfo::mapSrcLoop(const Loop *SrcLoop) const { 777 return SrcLoop->getLoopDepth(); 778 } 779 780 781 // Given one of the loops containing the destination, 782 // return its level index in our numbering scheme. 783 unsigned DependenceInfo::mapDstLoop(const Loop *DstLoop) const { 784 unsigned D = DstLoop->getLoopDepth(); 785 if (D > CommonLevels) 786 // This tries to make sure that we assign unique numbers to src and dst when 787 // the memory accesses reside in different loops that have the same depth. 788 return D - CommonLevels + SrcLevels; 789 else 790 return D; 791 } 792 793 794 // Returns true if Expression is loop invariant in LoopNest. 795 bool DependenceInfo::isLoopInvariant(const SCEV *Expression, 796 const Loop *LoopNest) const { 797 // Unlike ScalarEvolution::isLoopInvariant() we consider an access outside of 798 // any loop as invariant, because we only consier expression evaluation at a 799 // specific position (where the array access takes place), and not across the 800 // entire function. 801 if (!LoopNest) 802 return true; 803 804 // If the expression is invariant in the outermost loop of the loop nest, it 805 // is invariant anywhere in the loop nest. 806 return SE->isLoopInvariant(Expression, LoopNest->getOutermostLoop()); 807 } 808 809 810 811 // Finds the set of loops from the LoopNest that 812 // have a level <= CommonLevels and are referred to by the SCEV Expression. 813 void DependenceInfo::collectCommonLoops(const SCEV *Expression, 814 const Loop *LoopNest, 815 SmallBitVector &Loops) const { 816 while (LoopNest) { 817 unsigned Level = LoopNest->getLoopDepth(); 818 if (Level <= CommonLevels && !SE->isLoopInvariant(Expression, LoopNest)) 819 Loops.set(Level); 820 LoopNest = LoopNest->getParentLoop(); 821 } 822 } 823 824 void DependenceInfo::unifySubscriptType(ArrayRef<Subscript *> Pairs) { 825 826 unsigned widestWidthSeen = 0; 827 Type *widestType; 828 829 // Go through each pair and find the widest bit to which we need 830 // to extend all of them. 831 for (Subscript *Pair : Pairs) { 832 const SCEV *Src = Pair->Src; 833 const SCEV *Dst = Pair->Dst; 834 IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType()); 835 IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType()); 836 if (SrcTy == nullptr || DstTy == nullptr) { 837 assert(SrcTy == DstTy && "This function only unify integer types and " 838 "expect Src and Dst share the same type " 839 "otherwise."); 840 continue; 841 } 842 if (SrcTy->getBitWidth() > widestWidthSeen) { 843 widestWidthSeen = SrcTy->getBitWidth(); 844 widestType = SrcTy; 845 } 846 if (DstTy->getBitWidth() > widestWidthSeen) { 847 widestWidthSeen = DstTy->getBitWidth(); 848 widestType = DstTy; 849 } 850 } 851 852 853 assert(widestWidthSeen > 0); 854 855 // Now extend each pair to the widest seen. 856 for (Subscript *Pair : Pairs) { 857 const SCEV *Src = Pair->Src; 858 const SCEV *Dst = Pair->Dst; 859 IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType()); 860 IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType()); 861 if (SrcTy == nullptr || DstTy == nullptr) { 862 assert(SrcTy == DstTy && "This function only unify integer types and " 863 "expect Src and Dst share the same type " 864 "otherwise."); 865 continue; 866 } 867 if (SrcTy->getBitWidth() < widestWidthSeen) 868 // Sign-extend Src to widestType 869 Pair->Src = SE->getSignExtendExpr(Src, widestType); 870 if (DstTy->getBitWidth() < widestWidthSeen) { 871 // Sign-extend Dst to widestType 872 Pair->Dst = SE->getSignExtendExpr(Dst, widestType); 873 } 874 } 875 } 876 877 // removeMatchingExtensions - Examines a subscript pair. 878 // If the source and destination are identically sign (or zero) 879 // extended, it strips off the extension in an effect to simplify 880 // the actual analysis. 881 void DependenceInfo::removeMatchingExtensions(Subscript *Pair) { 882 const SCEV *Src = Pair->Src; 883 const SCEV *Dst = Pair->Dst; 884 if ((isa<SCEVZeroExtendExpr>(Src) && isa<SCEVZeroExtendExpr>(Dst)) || 885 (isa<SCEVSignExtendExpr>(Src) && isa<SCEVSignExtendExpr>(Dst))) { 886 const SCEVIntegralCastExpr *SrcCast = cast<SCEVIntegralCastExpr>(Src); 887 const SCEVIntegralCastExpr *DstCast = cast<SCEVIntegralCastExpr>(Dst); 888 const SCEV *SrcCastOp = SrcCast->getOperand(); 889 const SCEV *DstCastOp = DstCast->getOperand(); 890 if (SrcCastOp->getType() == DstCastOp->getType()) { 891 Pair->Src = SrcCastOp; 892 Pair->Dst = DstCastOp; 893 } 894 } 895 } 896 897 // Examine the scev and return true iff it's affine. 898 // Collect any loops mentioned in the set of "Loops". 899 bool DependenceInfo::checkSubscript(const SCEV *Expr, const Loop *LoopNest, 900 SmallBitVector &Loops, bool IsSrc) { 901 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr); 902 if (!AddRec) 903 return isLoopInvariant(Expr, LoopNest); 904 905 // The AddRec must depend on one of the containing loops. Otherwise, 906 // mapSrcLoop and mapDstLoop return indices outside the intended range. This 907 // can happen when a subscript in one loop references an IV from a sibling 908 // loop that could not be replaced with a concrete exit value by 909 // getSCEVAtScope. 910 const Loop *L = LoopNest; 911 while (L && AddRec->getLoop() != L) 912 L = L->getParentLoop(); 913 if (!L) 914 return false; 915 916 const SCEV *Start = AddRec->getStart(); 917 const SCEV *Step = AddRec->getStepRecurrence(*SE); 918 const SCEV *UB = SE->getBackedgeTakenCount(AddRec->getLoop()); 919 if (!isa<SCEVCouldNotCompute>(UB)) { 920 if (SE->getTypeSizeInBits(Start->getType()) < 921 SE->getTypeSizeInBits(UB->getType())) { 922 if (!AddRec->getNoWrapFlags()) 923 return false; 924 } 925 } 926 if (!isLoopInvariant(Step, LoopNest)) 927 return false; 928 if (IsSrc) 929 Loops.set(mapSrcLoop(AddRec->getLoop())); 930 else 931 Loops.set(mapDstLoop(AddRec->getLoop())); 932 return checkSubscript(Start, LoopNest, Loops, IsSrc); 933 } 934 935 // Examine the scev and return true iff it's linear. 936 // Collect any loops mentioned in the set of "Loops". 937 bool DependenceInfo::checkSrcSubscript(const SCEV *Src, const Loop *LoopNest, 938 SmallBitVector &Loops) { 939 return checkSubscript(Src, LoopNest, Loops, true); 940 } 941 942 // Examine the scev and return true iff it's linear. 943 // Collect any loops mentioned in the set of "Loops". 944 bool DependenceInfo::checkDstSubscript(const SCEV *Dst, const Loop *LoopNest, 945 SmallBitVector &Loops) { 946 return checkSubscript(Dst, LoopNest, Loops, false); 947 } 948 949 950 // Examines the subscript pair (the Src and Dst SCEVs) 951 // and classifies it as either ZIV, SIV, RDIV, MIV, or Nonlinear. 952 // Collects the associated loops in a set. 953 DependenceInfo::Subscript::ClassificationKind 954 DependenceInfo::classifyPair(const SCEV *Src, const Loop *SrcLoopNest, 955 const SCEV *Dst, const Loop *DstLoopNest, 956 SmallBitVector &Loops) { 957 SmallBitVector SrcLoops(MaxLevels + 1); 958 SmallBitVector DstLoops(MaxLevels + 1); 959 if (!checkSrcSubscript(Src, SrcLoopNest, SrcLoops)) 960 return Subscript::NonLinear; 961 if (!checkDstSubscript(Dst, DstLoopNest, DstLoops)) 962 return Subscript::NonLinear; 963 Loops = SrcLoops; 964 Loops |= DstLoops; 965 unsigned N = Loops.count(); 966 if (N == 0) 967 return Subscript::ZIV; 968 if (N == 1) 969 return Subscript::SIV; 970 if (N == 2 && (SrcLoops.count() == 0 || 971 DstLoops.count() == 0 || 972 (SrcLoops.count() == 1 && DstLoops.count() == 1))) 973 return Subscript::RDIV; 974 return Subscript::MIV; 975 } 976 977 978 // A wrapper around SCEV::isKnownPredicate. 979 // Looks for cases where we're interested in comparing for equality. 980 // If both X and Y have been identically sign or zero extended, 981 // it strips off the (confusing) extensions before invoking 982 // SCEV::isKnownPredicate. Perhaps, someday, the ScalarEvolution package 983 // will be similarly updated. 984 // 985 // If SCEV::isKnownPredicate can't prove the predicate, 986 // we try simple subtraction, which seems to help in some cases 987 // involving symbolics. 988 bool DependenceInfo::isKnownPredicate(ICmpInst::Predicate Pred, const SCEV *X, 989 const SCEV *Y) const { 990 if (Pred == CmpInst::ICMP_EQ || 991 Pred == CmpInst::ICMP_NE) { 992 if ((isa<SCEVSignExtendExpr>(X) && 993 isa<SCEVSignExtendExpr>(Y)) || 994 (isa<SCEVZeroExtendExpr>(X) && 995 isa<SCEVZeroExtendExpr>(Y))) { 996 const SCEVIntegralCastExpr *CX = cast<SCEVIntegralCastExpr>(X); 997 const SCEVIntegralCastExpr *CY = cast<SCEVIntegralCastExpr>(Y); 998 const SCEV *Xop = CX->getOperand(); 999 const SCEV *Yop = CY->getOperand(); 1000 if (Xop->getType() == Yop->getType()) { 1001 X = Xop; 1002 Y = Yop; 1003 } 1004 } 1005 } 1006 if (SE->isKnownPredicate(Pred, X, Y)) 1007 return true; 1008 // If SE->isKnownPredicate can't prove the condition, 1009 // we try the brute-force approach of subtracting 1010 // and testing the difference. 1011 // By testing with SE->isKnownPredicate first, we avoid 1012 // the possibility of overflow when the arguments are constants. 1013 const SCEV *Delta = SE->getMinusSCEV(X, Y); 1014 switch (Pred) { 1015 case CmpInst::ICMP_EQ: 1016 return Delta->isZero(); 1017 case CmpInst::ICMP_NE: 1018 return SE->isKnownNonZero(Delta); 1019 case CmpInst::ICMP_SGE: 1020 return SE->isKnownNonNegative(Delta); 1021 case CmpInst::ICMP_SLE: 1022 return SE->isKnownNonPositive(Delta); 1023 case CmpInst::ICMP_SGT: 1024 return SE->isKnownPositive(Delta); 1025 case CmpInst::ICMP_SLT: 1026 return SE->isKnownNegative(Delta); 1027 default: 1028 llvm_unreachable("unexpected predicate in isKnownPredicate"); 1029 } 1030 } 1031 1032 /// Compare to see if S is less than Size, using isKnownNegative(S - max(Size, 1)) 1033 /// with some extra checking if S is an AddRec and we can prove less-than using 1034 /// the loop bounds. 1035 bool DependenceInfo::isKnownLessThan(const SCEV *S, const SCEV *Size) const { 1036 // First unify to the same type 1037 auto *SType = dyn_cast<IntegerType>(S->getType()); 1038 auto *SizeType = dyn_cast<IntegerType>(Size->getType()); 1039 if (!SType || !SizeType) 1040 return false; 1041 Type *MaxType = 1042 (SType->getBitWidth() >= SizeType->getBitWidth()) ? SType : SizeType; 1043 S = SE->getTruncateOrZeroExtend(S, MaxType); 1044 Size = SE->getTruncateOrZeroExtend(Size, MaxType); 1045 1046 // Special check for addrecs using BE taken count 1047 const SCEV *Bound = SE->getMinusSCEV(S, Size); 1048 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Bound)) { 1049 if (AddRec->isAffine()) { 1050 const SCEV *BECount = SE->getBackedgeTakenCount(AddRec->getLoop()); 1051 if (!isa<SCEVCouldNotCompute>(BECount)) { 1052 const SCEV *Limit = AddRec->evaluateAtIteration(BECount, *SE); 1053 if (SE->isKnownNegative(Limit)) 1054 return true; 1055 } 1056 } 1057 } 1058 1059 // Check using normal isKnownNegative 1060 const SCEV *LimitedBound = 1061 SE->getMinusSCEV(S, SE->getSMaxExpr(Size, SE->getOne(Size->getType()))); 1062 return SE->isKnownNegative(LimitedBound); 1063 } 1064 1065 bool DependenceInfo::isKnownNonNegative(const SCEV *S, const Value *Ptr) const { 1066 bool Inbounds = false; 1067 if (auto *SrcGEP = dyn_cast<GetElementPtrInst>(Ptr)) 1068 Inbounds = SrcGEP->isInBounds(); 1069 if (Inbounds) { 1070 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) { 1071 if (AddRec->isAffine()) { 1072 // We know S is for Ptr, the operand on a load/store, so doesn't wrap. 1073 // If both parts are NonNegative, the end result will be NonNegative 1074 if (SE->isKnownNonNegative(AddRec->getStart()) && 1075 SE->isKnownNonNegative(AddRec->getOperand(1))) 1076 return true; 1077 } 1078 } 1079 } 1080 1081 return SE->isKnownNonNegative(S); 1082 } 1083 1084 // All subscripts are all the same type. 1085 // Loop bound may be smaller (e.g., a char). 1086 // Should zero extend loop bound, since it's always >= 0. 1087 // This routine collects upper bound and extends or truncates if needed. 1088 // Truncating is safe when subscripts are known not to wrap. Cases without 1089 // nowrap flags should have been rejected earlier. 1090 // Return null if no bound available. 1091 const SCEV *DependenceInfo::collectUpperBound(const Loop *L, Type *T) const { 1092 if (SE->hasLoopInvariantBackedgeTakenCount(L)) { 1093 const SCEV *UB = SE->getBackedgeTakenCount(L); 1094 return SE->getTruncateOrZeroExtend(UB, T); 1095 } 1096 return nullptr; 1097 } 1098 1099 1100 // Calls collectUpperBound(), then attempts to cast it to SCEVConstant. 1101 // If the cast fails, returns NULL. 1102 const SCEVConstant *DependenceInfo::collectConstantUpperBound(const Loop *L, 1103 Type *T) const { 1104 if (const SCEV *UB = collectUpperBound(L, T)) 1105 return dyn_cast<SCEVConstant>(UB); 1106 return nullptr; 1107 } 1108 1109 1110 // testZIV - 1111 // When we have a pair of subscripts of the form [c1] and [c2], 1112 // where c1 and c2 are both loop invariant, we attack it using 1113 // the ZIV test. Basically, we test by comparing the two values, 1114 // but there are actually three possible results: 1115 // 1) the values are equal, so there's a dependence 1116 // 2) the values are different, so there's no dependence 1117 // 3) the values might be equal, so we have to assume a dependence. 1118 // 1119 // Return true if dependence disproved. 1120 bool DependenceInfo::testZIV(const SCEV *Src, const SCEV *Dst, 1121 FullDependence &Result) const { 1122 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n"); 1123 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n"); 1124 ++ZIVapplications; 1125 if (isKnownPredicate(CmpInst::ICMP_EQ, Src, Dst)) { 1126 LLVM_DEBUG(dbgs() << " provably dependent\n"); 1127 return false; // provably dependent 1128 } 1129 if (isKnownPredicate(CmpInst::ICMP_NE, Src, Dst)) { 1130 LLVM_DEBUG(dbgs() << " provably independent\n"); 1131 ++ZIVindependence; 1132 return true; // provably independent 1133 } 1134 LLVM_DEBUG(dbgs() << " possibly dependent\n"); 1135 Result.Consistent = false; 1136 return false; // possibly dependent 1137 } 1138 1139 1140 // strongSIVtest - 1141 // From the paper, Practical Dependence Testing, Section 4.2.1 1142 // 1143 // When we have a pair of subscripts of the form [c1 + a*i] and [c2 + a*i], 1144 // where i is an induction variable, c1 and c2 are loop invariant, 1145 // and a is a constant, we can solve it exactly using the Strong SIV test. 1146 // 1147 // Can prove independence. Failing that, can compute distance (and direction). 1148 // In the presence of symbolic terms, we can sometimes make progress. 1149 // 1150 // If there's a dependence, 1151 // 1152 // c1 + a*i = c2 + a*i' 1153 // 1154 // The dependence distance is 1155 // 1156 // d = i' - i = (c1 - c2)/a 1157 // 1158 // A dependence only exists if d is an integer and abs(d) <= U, where U is the 1159 // loop's upper bound. If a dependence exists, the dependence direction is 1160 // defined as 1161 // 1162 // { < if d > 0 1163 // direction = { = if d = 0 1164 // { > if d < 0 1165 // 1166 // Return true if dependence disproved. 1167 bool DependenceInfo::strongSIVtest(const SCEV *Coeff, const SCEV *SrcConst, 1168 const SCEV *DstConst, const Loop *CurLoop, 1169 unsigned Level, FullDependence &Result, 1170 Constraint &NewConstraint) const { 1171 LLVM_DEBUG(dbgs() << "\tStrong SIV test\n"); 1172 LLVM_DEBUG(dbgs() << "\t Coeff = " << *Coeff); 1173 LLVM_DEBUG(dbgs() << ", " << *Coeff->getType() << "\n"); 1174 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst); 1175 LLVM_DEBUG(dbgs() << ", " << *SrcConst->getType() << "\n"); 1176 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst); 1177 LLVM_DEBUG(dbgs() << ", " << *DstConst->getType() << "\n"); 1178 ++StrongSIVapplications; 1179 assert(0 < Level && Level <= CommonLevels && "level out of range"); 1180 Level--; 1181 1182 const SCEV *Delta = SE->getMinusSCEV(SrcConst, DstConst); 1183 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta); 1184 LLVM_DEBUG(dbgs() << ", " << *Delta->getType() << "\n"); 1185 1186 // check that |Delta| < iteration count 1187 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) { 1188 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound); 1189 LLVM_DEBUG(dbgs() << ", " << *UpperBound->getType() << "\n"); 1190 const SCEV *AbsDelta = 1191 SE->isKnownNonNegative(Delta) ? Delta : SE->getNegativeSCEV(Delta); 1192 const SCEV *AbsCoeff = 1193 SE->isKnownNonNegative(Coeff) ? Coeff : SE->getNegativeSCEV(Coeff); 1194 const SCEV *Product = SE->getMulExpr(UpperBound, AbsCoeff); 1195 if (isKnownPredicate(CmpInst::ICMP_SGT, AbsDelta, Product)) { 1196 // Distance greater than trip count - no dependence 1197 ++StrongSIVindependence; 1198 ++StrongSIVsuccesses; 1199 return true; 1200 } 1201 } 1202 1203 // Can we compute distance? 1204 if (isa<SCEVConstant>(Delta) && isa<SCEVConstant>(Coeff)) { 1205 APInt ConstDelta = cast<SCEVConstant>(Delta)->getAPInt(); 1206 APInt ConstCoeff = cast<SCEVConstant>(Coeff)->getAPInt(); 1207 APInt Distance = ConstDelta; // these need to be initialized 1208 APInt Remainder = ConstDelta; 1209 APInt::sdivrem(ConstDelta, ConstCoeff, Distance, Remainder); 1210 LLVM_DEBUG(dbgs() << "\t Distance = " << Distance << "\n"); 1211 LLVM_DEBUG(dbgs() << "\t Remainder = " << Remainder << "\n"); 1212 // Make sure Coeff divides Delta exactly 1213 if (Remainder != 0) { 1214 // Coeff doesn't divide Distance, no dependence 1215 ++StrongSIVindependence; 1216 ++StrongSIVsuccesses; 1217 return true; 1218 } 1219 Result.DV[Level].Distance = SE->getConstant(Distance); 1220 NewConstraint.setDistance(SE->getConstant(Distance), CurLoop); 1221 if (Distance.sgt(0)) 1222 Result.DV[Level].Direction &= Dependence::DVEntry::LT; 1223 else if (Distance.slt(0)) 1224 Result.DV[Level].Direction &= Dependence::DVEntry::GT; 1225 else 1226 Result.DV[Level].Direction &= Dependence::DVEntry::EQ; 1227 ++StrongSIVsuccesses; 1228 } 1229 else if (Delta->isZero()) { 1230 // since 0/X == 0 1231 Result.DV[Level].Distance = Delta; 1232 NewConstraint.setDistance(Delta, CurLoop); 1233 Result.DV[Level].Direction &= Dependence::DVEntry::EQ; 1234 ++StrongSIVsuccesses; 1235 } 1236 else { 1237 if (Coeff->isOne()) { 1238 LLVM_DEBUG(dbgs() << "\t Distance = " << *Delta << "\n"); 1239 Result.DV[Level].Distance = Delta; // since X/1 == X 1240 NewConstraint.setDistance(Delta, CurLoop); 1241 } 1242 else { 1243 Result.Consistent = false; 1244 NewConstraint.setLine(Coeff, 1245 SE->getNegativeSCEV(Coeff), 1246 SE->getNegativeSCEV(Delta), CurLoop); 1247 } 1248 1249 // maybe we can get a useful direction 1250 bool DeltaMaybeZero = !SE->isKnownNonZero(Delta); 1251 bool DeltaMaybePositive = !SE->isKnownNonPositive(Delta); 1252 bool DeltaMaybeNegative = !SE->isKnownNonNegative(Delta); 1253 bool CoeffMaybePositive = !SE->isKnownNonPositive(Coeff); 1254 bool CoeffMaybeNegative = !SE->isKnownNonNegative(Coeff); 1255 // The double negatives above are confusing. 1256 // It helps to read !SE->isKnownNonZero(Delta) 1257 // as "Delta might be Zero" 1258 unsigned NewDirection = Dependence::DVEntry::NONE; 1259 if ((DeltaMaybePositive && CoeffMaybePositive) || 1260 (DeltaMaybeNegative && CoeffMaybeNegative)) 1261 NewDirection = Dependence::DVEntry::LT; 1262 if (DeltaMaybeZero) 1263 NewDirection |= Dependence::DVEntry::EQ; 1264 if ((DeltaMaybeNegative && CoeffMaybePositive) || 1265 (DeltaMaybePositive && CoeffMaybeNegative)) 1266 NewDirection |= Dependence::DVEntry::GT; 1267 if (NewDirection < Result.DV[Level].Direction) 1268 ++StrongSIVsuccesses; 1269 Result.DV[Level].Direction &= NewDirection; 1270 } 1271 return false; 1272 } 1273 1274 1275 // weakCrossingSIVtest - 1276 // From the paper, Practical Dependence Testing, Section 4.2.2 1277 // 1278 // When we have a pair of subscripts of the form [c1 + a*i] and [c2 - a*i], 1279 // where i is an induction variable, c1 and c2 are loop invariant, 1280 // and a is a constant, we can solve it exactly using the 1281 // Weak-Crossing SIV test. 1282 // 1283 // Given c1 + a*i = c2 - a*i', we can look for the intersection of 1284 // the two lines, where i = i', yielding 1285 // 1286 // c1 + a*i = c2 - a*i 1287 // 2a*i = c2 - c1 1288 // i = (c2 - c1)/2a 1289 // 1290 // If i < 0, there is no dependence. 1291 // If i > upperbound, there is no dependence. 1292 // If i = 0 (i.e., if c1 = c2), there's a dependence with distance = 0. 1293 // If i = upperbound, there's a dependence with distance = 0. 1294 // If i is integral, there's a dependence (all directions). 1295 // If the non-integer part = 1/2, there's a dependence (<> directions). 1296 // Otherwise, there's no dependence. 1297 // 1298 // Can prove independence. Failing that, 1299 // can sometimes refine the directions. 1300 // Can determine iteration for splitting. 1301 // 1302 // Return true if dependence disproved. 1303 bool DependenceInfo::weakCrossingSIVtest( 1304 const SCEV *Coeff, const SCEV *SrcConst, const SCEV *DstConst, 1305 const Loop *CurLoop, unsigned Level, FullDependence &Result, 1306 Constraint &NewConstraint, const SCEV *&SplitIter) const { 1307 LLVM_DEBUG(dbgs() << "\tWeak-Crossing SIV test\n"); 1308 LLVM_DEBUG(dbgs() << "\t Coeff = " << *Coeff << "\n"); 1309 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n"); 1310 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n"); 1311 ++WeakCrossingSIVapplications; 1312 assert(0 < Level && Level <= CommonLevels && "Level out of range"); 1313 Level--; 1314 Result.Consistent = false; 1315 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst); 1316 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n"); 1317 NewConstraint.setLine(Coeff, Coeff, Delta, CurLoop); 1318 if (Delta->isZero()) { 1319 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::LT); 1320 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::GT); 1321 ++WeakCrossingSIVsuccesses; 1322 if (!Result.DV[Level].Direction) { 1323 ++WeakCrossingSIVindependence; 1324 return true; 1325 } 1326 Result.DV[Level].Distance = Delta; // = 0 1327 return false; 1328 } 1329 const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(Coeff); 1330 if (!ConstCoeff) 1331 return false; 1332 1333 Result.DV[Level].Splitable = true; 1334 if (SE->isKnownNegative(ConstCoeff)) { 1335 ConstCoeff = dyn_cast<SCEVConstant>(SE->getNegativeSCEV(ConstCoeff)); 1336 assert(ConstCoeff && 1337 "dynamic cast of negative of ConstCoeff should yield constant"); 1338 Delta = SE->getNegativeSCEV(Delta); 1339 } 1340 assert(SE->isKnownPositive(ConstCoeff) && "ConstCoeff should be positive"); 1341 1342 // compute SplitIter for use by DependenceInfo::getSplitIteration() 1343 SplitIter = SE->getUDivExpr( 1344 SE->getSMaxExpr(SE->getZero(Delta->getType()), Delta), 1345 SE->getMulExpr(SE->getConstant(Delta->getType(), 2), ConstCoeff)); 1346 LLVM_DEBUG(dbgs() << "\t Split iter = " << *SplitIter << "\n"); 1347 1348 const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta); 1349 if (!ConstDelta) 1350 return false; 1351 1352 // We're certain that ConstCoeff > 0; therefore, 1353 // if Delta < 0, then no dependence. 1354 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n"); 1355 LLVM_DEBUG(dbgs() << "\t ConstCoeff = " << *ConstCoeff << "\n"); 1356 if (SE->isKnownNegative(Delta)) { 1357 // No dependence, Delta < 0 1358 ++WeakCrossingSIVindependence; 1359 ++WeakCrossingSIVsuccesses; 1360 return true; 1361 } 1362 1363 // We're certain that Delta > 0 and ConstCoeff > 0. 1364 // Check Delta/(2*ConstCoeff) against upper loop bound 1365 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) { 1366 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound << "\n"); 1367 const SCEV *ConstantTwo = SE->getConstant(UpperBound->getType(), 2); 1368 const SCEV *ML = SE->getMulExpr(SE->getMulExpr(ConstCoeff, UpperBound), 1369 ConstantTwo); 1370 LLVM_DEBUG(dbgs() << "\t ML = " << *ML << "\n"); 1371 if (isKnownPredicate(CmpInst::ICMP_SGT, Delta, ML)) { 1372 // Delta too big, no dependence 1373 ++WeakCrossingSIVindependence; 1374 ++WeakCrossingSIVsuccesses; 1375 return true; 1376 } 1377 if (isKnownPredicate(CmpInst::ICMP_EQ, Delta, ML)) { 1378 // i = i' = UB 1379 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::LT); 1380 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::GT); 1381 ++WeakCrossingSIVsuccesses; 1382 if (!Result.DV[Level].Direction) { 1383 ++WeakCrossingSIVindependence; 1384 return true; 1385 } 1386 Result.DV[Level].Splitable = false; 1387 Result.DV[Level].Distance = SE->getZero(Delta->getType()); 1388 return false; 1389 } 1390 } 1391 1392 // check that Coeff divides Delta 1393 APInt APDelta = ConstDelta->getAPInt(); 1394 APInt APCoeff = ConstCoeff->getAPInt(); 1395 APInt Distance = APDelta; // these need to be initialzed 1396 APInt Remainder = APDelta; 1397 APInt::sdivrem(APDelta, APCoeff, Distance, Remainder); 1398 LLVM_DEBUG(dbgs() << "\t Remainder = " << Remainder << "\n"); 1399 if (Remainder != 0) { 1400 // Coeff doesn't divide Delta, no dependence 1401 ++WeakCrossingSIVindependence; 1402 ++WeakCrossingSIVsuccesses; 1403 return true; 1404 } 1405 LLVM_DEBUG(dbgs() << "\t Distance = " << Distance << "\n"); 1406 1407 // if 2*Coeff doesn't divide Delta, then the equal direction isn't possible 1408 APInt Two = APInt(Distance.getBitWidth(), 2, true); 1409 Remainder = Distance.srem(Two); 1410 LLVM_DEBUG(dbgs() << "\t Remainder = " << Remainder << "\n"); 1411 if (Remainder != 0) { 1412 // Equal direction isn't possible 1413 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::EQ); 1414 ++WeakCrossingSIVsuccesses; 1415 } 1416 return false; 1417 } 1418 1419 1420 // Kirch's algorithm, from 1421 // 1422 // Optimizing Supercompilers for Supercomputers 1423 // Michael Wolfe 1424 // MIT Press, 1989 1425 // 1426 // Program 2.1, page 29. 1427 // Computes the GCD of AM and BM. 1428 // Also finds a solution to the equation ax - by = gcd(a, b). 1429 // Returns true if dependence disproved; i.e., gcd does not divide Delta. 1430 static bool findGCD(unsigned Bits, const APInt &AM, const APInt &BM, 1431 const APInt &Delta, APInt &G, APInt &X, APInt &Y) { 1432 APInt A0(Bits, 1, true), A1(Bits, 0, true); 1433 APInt B0(Bits, 0, true), B1(Bits, 1, true); 1434 APInt G0 = AM.abs(); 1435 APInt G1 = BM.abs(); 1436 APInt Q = G0; // these need to be initialized 1437 APInt R = G0; 1438 APInt::sdivrem(G0, G1, Q, R); 1439 while (R != 0) { 1440 APInt A2 = A0 - Q*A1; A0 = A1; A1 = A2; 1441 APInt B2 = B0 - Q*B1; B0 = B1; B1 = B2; 1442 G0 = G1; G1 = R; 1443 APInt::sdivrem(G0, G1, Q, R); 1444 } 1445 G = G1; 1446 LLVM_DEBUG(dbgs() << "\t GCD = " << G << "\n"); 1447 X = AM.slt(0) ? -A1 : A1; 1448 Y = BM.slt(0) ? B1 : -B1; 1449 1450 // make sure gcd divides Delta 1451 R = Delta.srem(G); 1452 if (R != 0) 1453 return true; // gcd doesn't divide Delta, no dependence 1454 Q = Delta.sdiv(G); 1455 return false; 1456 } 1457 1458 static APInt floorOfQuotient(const APInt &A, const APInt &B) { 1459 APInt Q = A; // these need to be initialized 1460 APInt R = A; 1461 APInt::sdivrem(A, B, Q, R); 1462 if (R == 0) 1463 return Q; 1464 if ((A.sgt(0) && B.sgt(0)) || 1465 (A.slt(0) && B.slt(0))) 1466 return Q; 1467 else 1468 return Q - 1; 1469 } 1470 1471 static APInt ceilingOfQuotient(const APInt &A, const APInt &B) { 1472 APInt Q = A; // these need to be initialized 1473 APInt R = A; 1474 APInt::sdivrem(A, B, Q, R); 1475 if (R == 0) 1476 return Q; 1477 if ((A.sgt(0) && B.sgt(0)) || 1478 (A.slt(0) && B.slt(0))) 1479 return Q + 1; 1480 else 1481 return Q; 1482 } 1483 1484 // exactSIVtest - 1485 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*i], 1486 // where i is an induction variable, c1 and c2 are loop invariant, and a1 1487 // and a2 are constant, we can solve it exactly using an algorithm developed 1488 // by Banerjee and Wolfe. See Algorithm 6.2.1 (case 2.5) in: 1489 // 1490 // Dependence Analysis for Supercomputing 1491 // Utpal Banerjee 1492 // Kluwer Academic Publishers, 1988 1493 // 1494 // It's slower than the specialized tests (strong SIV, weak-zero SIV, etc), 1495 // so use them if possible. They're also a bit better with symbolics and, 1496 // in the case of the strong SIV test, can compute Distances. 1497 // 1498 // Return true if dependence disproved. 1499 // 1500 // This is a modified version of the original Banerjee algorithm. The original 1501 // only tested whether Dst depends on Src. This algorithm extends that and 1502 // returns all the dependencies that exist between Dst and Src. 1503 bool DependenceInfo::exactSIVtest(const SCEV *SrcCoeff, const SCEV *DstCoeff, 1504 const SCEV *SrcConst, const SCEV *DstConst, 1505 const Loop *CurLoop, unsigned Level, 1506 FullDependence &Result, 1507 Constraint &NewConstraint) const { 1508 LLVM_DEBUG(dbgs() << "\tExact SIV test\n"); 1509 LLVM_DEBUG(dbgs() << "\t SrcCoeff = " << *SrcCoeff << " = AM\n"); 1510 LLVM_DEBUG(dbgs() << "\t DstCoeff = " << *DstCoeff << " = BM\n"); 1511 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n"); 1512 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n"); 1513 ++ExactSIVapplications; 1514 assert(0 < Level && Level <= CommonLevels && "Level out of range"); 1515 Level--; 1516 Result.Consistent = false; 1517 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst); 1518 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n"); 1519 NewConstraint.setLine(SrcCoeff, SE->getNegativeSCEV(DstCoeff), Delta, 1520 CurLoop); 1521 const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta); 1522 const SCEVConstant *ConstSrcCoeff = dyn_cast<SCEVConstant>(SrcCoeff); 1523 const SCEVConstant *ConstDstCoeff = dyn_cast<SCEVConstant>(DstCoeff); 1524 if (!ConstDelta || !ConstSrcCoeff || !ConstDstCoeff) 1525 return false; 1526 1527 // find gcd 1528 APInt G, X, Y; 1529 APInt AM = ConstSrcCoeff->getAPInt(); 1530 APInt BM = ConstDstCoeff->getAPInt(); 1531 APInt CM = ConstDelta->getAPInt(); 1532 unsigned Bits = AM.getBitWidth(); 1533 if (findGCD(Bits, AM, BM, CM, G, X, Y)) { 1534 // gcd doesn't divide Delta, no dependence 1535 ++ExactSIVindependence; 1536 ++ExactSIVsuccesses; 1537 return true; 1538 } 1539 1540 LLVM_DEBUG(dbgs() << "\t X = " << X << ", Y = " << Y << "\n"); 1541 1542 // since SCEV construction normalizes, LM = 0 1543 APInt UM(Bits, 1, true); 1544 bool UMValid = false; 1545 // UM is perhaps unavailable, let's check 1546 if (const SCEVConstant *CUB = 1547 collectConstantUpperBound(CurLoop, Delta->getType())) { 1548 UM = CUB->getAPInt(); 1549 LLVM_DEBUG(dbgs() << "\t UM = " << UM << "\n"); 1550 UMValid = true; 1551 } 1552 1553 APInt TU(APInt::getSignedMaxValue(Bits)); 1554 APInt TL(APInt::getSignedMinValue(Bits)); 1555 APInt TC = CM.sdiv(G); 1556 APInt TX = X * TC; 1557 APInt TY = Y * TC; 1558 LLVM_DEBUG(dbgs() << "\t TC = " << TC << "\n"); 1559 LLVM_DEBUG(dbgs() << "\t TX = " << TX << "\n"); 1560 LLVM_DEBUG(dbgs() << "\t TY = " << TY << "\n"); 1561 1562 SmallVector<APInt, 2> TLVec, TUVec; 1563 APInt TB = BM.sdiv(G); 1564 if (TB.sgt(0)) { 1565 TLVec.push_back(ceilingOfQuotient(-TX, TB)); 1566 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n"); 1567 // New bound check - modification to Banerjee's e3 check 1568 if (UMValid) { 1569 TUVec.push_back(floorOfQuotient(UM - TX, TB)); 1570 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n"); 1571 } 1572 } else { 1573 TUVec.push_back(floorOfQuotient(-TX, TB)); 1574 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n"); 1575 // New bound check - modification to Banerjee's e3 check 1576 if (UMValid) { 1577 TLVec.push_back(ceilingOfQuotient(UM - TX, TB)); 1578 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n"); 1579 } 1580 } 1581 1582 APInt TA = AM.sdiv(G); 1583 if (TA.sgt(0)) { 1584 if (UMValid) { 1585 TUVec.push_back(floorOfQuotient(UM - TY, TA)); 1586 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n"); 1587 } 1588 // New bound check - modification to Banerjee's e3 check 1589 TLVec.push_back(ceilingOfQuotient(-TY, TA)); 1590 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n"); 1591 } else { 1592 if (UMValid) { 1593 TLVec.push_back(ceilingOfQuotient(UM - TY, TA)); 1594 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n"); 1595 } 1596 // New bound check - modification to Banerjee's e3 check 1597 TUVec.push_back(floorOfQuotient(-TY, TA)); 1598 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n"); 1599 } 1600 1601 LLVM_DEBUG(dbgs() << "\t TA = " << TA << "\n"); 1602 LLVM_DEBUG(dbgs() << "\t TB = " << TB << "\n"); 1603 1604 if (TLVec.empty() || TUVec.empty()) 1605 return false; 1606 TL = APIntOps::smax(TLVec.front(), TLVec.back()); 1607 TU = APIntOps::smin(TUVec.front(), TUVec.back()); 1608 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n"); 1609 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n"); 1610 1611 if (TL.sgt(TU)) { 1612 ++ExactSIVindependence; 1613 ++ExactSIVsuccesses; 1614 return true; 1615 } 1616 1617 // explore directions 1618 unsigned NewDirection = Dependence::DVEntry::NONE; 1619 APInt LowerDistance, UpperDistance; 1620 if (TA.sgt(TB)) { 1621 LowerDistance = (TY - TX) + (TA - TB) * TL; 1622 UpperDistance = (TY - TX) + (TA - TB) * TU; 1623 } else { 1624 LowerDistance = (TY - TX) + (TA - TB) * TU; 1625 UpperDistance = (TY - TX) + (TA - TB) * TL; 1626 } 1627 1628 LLVM_DEBUG(dbgs() << "\t LowerDistance = " << LowerDistance << "\n"); 1629 LLVM_DEBUG(dbgs() << "\t UpperDistance = " << UpperDistance << "\n"); 1630 1631 APInt Zero(Bits, 0, true); 1632 if (LowerDistance.sle(Zero) && UpperDistance.sge(Zero)) { 1633 NewDirection |= Dependence::DVEntry::EQ; 1634 ++ExactSIVsuccesses; 1635 } 1636 if (LowerDistance.slt(0)) { 1637 NewDirection |= Dependence::DVEntry::GT; 1638 ++ExactSIVsuccesses; 1639 } 1640 if (UpperDistance.sgt(0)) { 1641 NewDirection |= Dependence::DVEntry::LT; 1642 ++ExactSIVsuccesses; 1643 } 1644 1645 // finished 1646 Result.DV[Level].Direction &= NewDirection; 1647 if (Result.DV[Level].Direction == Dependence::DVEntry::NONE) 1648 ++ExactSIVindependence; 1649 LLVM_DEBUG(dbgs() << "\t Result = "); 1650 LLVM_DEBUG(Result.dump(dbgs())); 1651 return Result.DV[Level].Direction == Dependence::DVEntry::NONE; 1652 } 1653 1654 1655 // Return true if the divisor evenly divides the dividend. 1656 static 1657 bool isRemainderZero(const SCEVConstant *Dividend, 1658 const SCEVConstant *Divisor) { 1659 const APInt &ConstDividend = Dividend->getAPInt(); 1660 const APInt &ConstDivisor = Divisor->getAPInt(); 1661 return ConstDividend.srem(ConstDivisor) == 0; 1662 } 1663 1664 1665 // weakZeroSrcSIVtest - 1666 // From the paper, Practical Dependence Testing, Section 4.2.2 1667 // 1668 // When we have a pair of subscripts of the form [c1] and [c2 + a*i], 1669 // where i is an induction variable, c1 and c2 are loop invariant, 1670 // and a is a constant, we can solve it exactly using the 1671 // Weak-Zero SIV test. 1672 // 1673 // Given 1674 // 1675 // c1 = c2 + a*i 1676 // 1677 // we get 1678 // 1679 // (c1 - c2)/a = i 1680 // 1681 // If i is not an integer, there's no dependence. 1682 // If i < 0 or > UB, there's no dependence. 1683 // If i = 0, the direction is >= and peeling the 1684 // 1st iteration will break the dependence. 1685 // If i = UB, the direction is <= and peeling the 1686 // last iteration will break the dependence. 1687 // Otherwise, the direction is *. 1688 // 1689 // Can prove independence. Failing that, we can sometimes refine 1690 // the directions. Can sometimes show that first or last 1691 // iteration carries all the dependences (so worth peeling). 1692 // 1693 // (see also weakZeroDstSIVtest) 1694 // 1695 // Return true if dependence disproved. 1696 bool DependenceInfo::weakZeroSrcSIVtest(const SCEV *DstCoeff, 1697 const SCEV *SrcConst, 1698 const SCEV *DstConst, 1699 const Loop *CurLoop, unsigned Level, 1700 FullDependence &Result, 1701 Constraint &NewConstraint) const { 1702 // For the WeakSIV test, it's possible the loop isn't common to 1703 // the Src and Dst loops. If it isn't, then there's no need to 1704 // record a direction. 1705 LLVM_DEBUG(dbgs() << "\tWeak-Zero (src) SIV test\n"); 1706 LLVM_DEBUG(dbgs() << "\t DstCoeff = " << *DstCoeff << "\n"); 1707 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n"); 1708 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n"); 1709 ++WeakZeroSIVapplications; 1710 assert(0 < Level && Level <= MaxLevels && "Level out of range"); 1711 Level--; 1712 Result.Consistent = false; 1713 const SCEV *Delta = SE->getMinusSCEV(SrcConst, DstConst); 1714 NewConstraint.setLine(SE->getZero(Delta->getType()), DstCoeff, Delta, 1715 CurLoop); 1716 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n"); 1717 if (isKnownPredicate(CmpInst::ICMP_EQ, SrcConst, DstConst)) { 1718 if (Level < CommonLevels) { 1719 Result.DV[Level].Direction &= Dependence::DVEntry::GE; 1720 Result.DV[Level].PeelFirst = true; 1721 ++WeakZeroSIVsuccesses; 1722 } 1723 return false; // dependences caused by first iteration 1724 } 1725 const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(DstCoeff); 1726 if (!ConstCoeff) 1727 return false; 1728 const SCEV *AbsCoeff = 1729 SE->isKnownNegative(ConstCoeff) ? 1730 SE->getNegativeSCEV(ConstCoeff) : ConstCoeff; 1731 const SCEV *NewDelta = 1732 SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(Delta) : Delta; 1733 1734 // check that Delta/SrcCoeff < iteration count 1735 // really check NewDelta < count*AbsCoeff 1736 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) { 1737 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound << "\n"); 1738 const SCEV *Product = SE->getMulExpr(AbsCoeff, UpperBound); 1739 if (isKnownPredicate(CmpInst::ICMP_SGT, NewDelta, Product)) { 1740 ++WeakZeroSIVindependence; 1741 ++WeakZeroSIVsuccesses; 1742 return true; 1743 } 1744 if (isKnownPredicate(CmpInst::ICMP_EQ, NewDelta, Product)) { 1745 // dependences caused by last iteration 1746 if (Level < CommonLevels) { 1747 Result.DV[Level].Direction &= Dependence::DVEntry::LE; 1748 Result.DV[Level].PeelLast = true; 1749 ++WeakZeroSIVsuccesses; 1750 } 1751 return false; 1752 } 1753 } 1754 1755 // check that Delta/SrcCoeff >= 0 1756 // really check that NewDelta >= 0 1757 if (SE->isKnownNegative(NewDelta)) { 1758 // No dependence, newDelta < 0 1759 ++WeakZeroSIVindependence; 1760 ++WeakZeroSIVsuccesses; 1761 return true; 1762 } 1763 1764 // if SrcCoeff doesn't divide Delta, then no dependence 1765 if (isa<SCEVConstant>(Delta) && 1766 !isRemainderZero(cast<SCEVConstant>(Delta), ConstCoeff)) { 1767 ++WeakZeroSIVindependence; 1768 ++WeakZeroSIVsuccesses; 1769 return true; 1770 } 1771 return false; 1772 } 1773 1774 1775 // weakZeroDstSIVtest - 1776 // From the paper, Practical Dependence Testing, Section 4.2.2 1777 // 1778 // When we have a pair of subscripts of the form [c1 + a*i] and [c2], 1779 // where i is an induction variable, c1 and c2 are loop invariant, 1780 // and a is a constant, we can solve it exactly using the 1781 // Weak-Zero SIV test. 1782 // 1783 // Given 1784 // 1785 // c1 + a*i = c2 1786 // 1787 // we get 1788 // 1789 // i = (c2 - c1)/a 1790 // 1791 // If i is not an integer, there's no dependence. 1792 // If i < 0 or > UB, there's no dependence. 1793 // If i = 0, the direction is <= and peeling the 1794 // 1st iteration will break the dependence. 1795 // If i = UB, the direction is >= and peeling the 1796 // last iteration will break the dependence. 1797 // Otherwise, the direction is *. 1798 // 1799 // Can prove independence. Failing that, we can sometimes refine 1800 // the directions. Can sometimes show that first or last 1801 // iteration carries all the dependences (so worth peeling). 1802 // 1803 // (see also weakZeroSrcSIVtest) 1804 // 1805 // Return true if dependence disproved. 1806 bool DependenceInfo::weakZeroDstSIVtest(const SCEV *SrcCoeff, 1807 const SCEV *SrcConst, 1808 const SCEV *DstConst, 1809 const Loop *CurLoop, unsigned Level, 1810 FullDependence &Result, 1811 Constraint &NewConstraint) const { 1812 // For the WeakSIV test, it's possible the loop isn't common to the 1813 // Src and Dst loops. If it isn't, then there's no need to record a direction. 1814 LLVM_DEBUG(dbgs() << "\tWeak-Zero (dst) SIV test\n"); 1815 LLVM_DEBUG(dbgs() << "\t SrcCoeff = " << *SrcCoeff << "\n"); 1816 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n"); 1817 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n"); 1818 ++WeakZeroSIVapplications; 1819 assert(0 < Level && Level <= SrcLevels && "Level out of range"); 1820 Level--; 1821 Result.Consistent = false; 1822 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst); 1823 NewConstraint.setLine(SrcCoeff, SE->getZero(Delta->getType()), Delta, 1824 CurLoop); 1825 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n"); 1826 if (isKnownPredicate(CmpInst::ICMP_EQ, DstConst, SrcConst)) { 1827 if (Level < CommonLevels) { 1828 Result.DV[Level].Direction &= Dependence::DVEntry::LE; 1829 Result.DV[Level].PeelFirst = true; 1830 ++WeakZeroSIVsuccesses; 1831 } 1832 return false; // dependences caused by first iteration 1833 } 1834 const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(SrcCoeff); 1835 if (!ConstCoeff) 1836 return false; 1837 const SCEV *AbsCoeff = 1838 SE->isKnownNegative(ConstCoeff) ? 1839 SE->getNegativeSCEV(ConstCoeff) : ConstCoeff; 1840 const SCEV *NewDelta = 1841 SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(Delta) : Delta; 1842 1843 // check that Delta/SrcCoeff < iteration count 1844 // really check NewDelta < count*AbsCoeff 1845 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) { 1846 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound << "\n"); 1847 const SCEV *Product = SE->getMulExpr(AbsCoeff, UpperBound); 1848 if (isKnownPredicate(CmpInst::ICMP_SGT, NewDelta, Product)) { 1849 ++WeakZeroSIVindependence; 1850 ++WeakZeroSIVsuccesses; 1851 return true; 1852 } 1853 if (isKnownPredicate(CmpInst::ICMP_EQ, NewDelta, Product)) { 1854 // dependences caused by last iteration 1855 if (Level < CommonLevels) { 1856 Result.DV[Level].Direction &= Dependence::DVEntry::GE; 1857 Result.DV[Level].PeelLast = true; 1858 ++WeakZeroSIVsuccesses; 1859 } 1860 return false; 1861 } 1862 } 1863 1864 // check that Delta/SrcCoeff >= 0 1865 // really check that NewDelta >= 0 1866 if (SE->isKnownNegative(NewDelta)) { 1867 // No dependence, newDelta < 0 1868 ++WeakZeroSIVindependence; 1869 ++WeakZeroSIVsuccesses; 1870 return true; 1871 } 1872 1873 // if SrcCoeff doesn't divide Delta, then no dependence 1874 if (isa<SCEVConstant>(Delta) && 1875 !isRemainderZero(cast<SCEVConstant>(Delta), ConstCoeff)) { 1876 ++WeakZeroSIVindependence; 1877 ++WeakZeroSIVsuccesses; 1878 return true; 1879 } 1880 return false; 1881 } 1882 1883 1884 // exactRDIVtest - Tests the RDIV subscript pair for dependence. 1885 // Things of the form [c1 + a*i] and [c2 + b*j], 1886 // where i and j are induction variable, c1 and c2 are loop invariant, 1887 // and a and b are constants. 1888 // Returns true if any possible dependence is disproved. 1889 // Marks the result as inconsistent. 1890 // Works in some cases that symbolicRDIVtest doesn't, and vice versa. 1891 bool DependenceInfo::exactRDIVtest(const SCEV *SrcCoeff, const SCEV *DstCoeff, 1892 const SCEV *SrcConst, const SCEV *DstConst, 1893 const Loop *SrcLoop, const Loop *DstLoop, 1894 FullDependence &Result) const { 1895 LLVM_DEBUG(dbgs() << "\tExact RDIV test\n"); 1896 LLVM_DEBUG(dbgs() << "\t SrcCoeff = " << *SrcCoeff << " = AM\n"); 1897 LLVM_DEBUG(dbgs() << "\t DstCoeff = " << *DstCoeff << " = BM\n"); 1898 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n"); 1899 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n"); 1900 ++ExactRDIVapplications; 1901 Result.Consistent = false; 1902 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst); 1903 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n"); 1904 const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta); 1905 const SCEVConstant *ConstSrcCoeff = dyn_cast<SCEVConstant>(SrcCoeff); 1906 const SCEVConstant *ConstDstCoeff = dyn_cast<SCEVConstant>(DstCoeff); 1907 if (!ConstDelta || !ConstSrcCoeff || !ConstDstCoeff) 1908 return false; 1909 1910 // find gcd 1911 APInt G, X, Y; 1912 APInt AM = ConstSrcCoeff->getAPInt(); 1913 APInt BM = ConstDstCoeff->getAPInt(); 1914 APInt CM = ConstDelta->getAPInt(); 1915 unsigned Bits = AM.getBitWidth(); 1916 if (findGCD(Bits, AM, BM, CM, G, X, Y)) { 1917 // gcd doesn't divide Delta, no dependence 1918 ++ExactRDIVindependence; 1919 return true; 1920 } 1921 1922 LLVM_DEBUG(dbgs() << "\t X = " << X << ", Y = " << Y << "\n"); 1923 1924 // since SCEV construction seems to normalize, LM = 0 1925 APInt SrcUM(Bits, 1, true); 1926 bool SrcUMvalid = false; 1927 // SrcUM is perhaps unavailable, let's check 1928 if (const SCEVConstant *UpperBound = 1929 collectConstantUpperBound(SrcLoop, Delta->getType())) { 1930 SrcUM = UpperBound->getAPInt(); 1931 LLVM_DEBUG(dbgs() << "\t SrcUM = " << SrcUM << "\n"); 1932 SrcUMvalid = true; 1933 } 1934 1935 APInt DstUM(Bits, 1, true); 1936 bool DstUMvalid = false; 1937 // UM is perhaps unavailable, let's check 1938 if (const SCEVConstant *UpperBound = 1939 collectConstantUpperBound(DstLoop, Delta->getType())) { 1940 DstUM = UpperBound->getAPInt(); 1941 LLVM_DEBUG(dbgs() << "\t DstUM = " << DstUM << "\n"); 1942 DstUMvalid = true; 1943 } 1944 1945 APInt TU(APInt::getSignedMaxValue(Bits)); 1946 APInt TL(APInt::getSignedMinValue(Bits)); 1947 APInt TC = CM.sdiv(G); 1948 APInt TX = X * TC; 1949 APInt TY = Y * TC; 1950 LLVM_DEBUG(dbgs() << "\t TC = " << TC << "\n"); 1951 LLVM_DEBUG(dbgs() << "\t TX = " << TX << "\n"); 1952 LLVM_DEBUG(dbgs() << "\t TY = " << TY << "\n"); 1953 1954 SmallVector<APInt, 2> TLVec, TUVec; 1955 APInt TB = BM.sdiv(G); 1956 if (TB.sgt(0)) { 1957 TLVec.push_back(ceilingOfQuotient(-TX, TB)); 1958 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n"); 1959 if (SrcUMvalid) { 1960 TUVec.push_back(floorOfQuotient(SrcUM - TX, TB)); 1961 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n"); 1962 } 1963 } else { 1964 TUVec.push_back(floorOfQuotient(-TX, TB)); 1965 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n"); 1966 if (SrcUMvalid) { 1967 TLVec.push_back(ceilingOfQuotient(SrcUM - TX, TB)); 1968 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n"); 1969 } 1970 } 1971 1972 APInt TA = AM.sdiv(G); 1973 if (TA.sgt(0)) { 1974 TLVec.push_back(ceilingOfQuotient(-TY, TA)); 1975 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n"); 1976 if (DstUMvalid) { 1977 TUVec.push_back(floorOfQuotient(DstUM - TY, TA)); 1978 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n"); 1979 } 1980 } else { 1981 TUVec.push_back(floorOfQuotient(-TY, TA)); 1982 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n"); 1983 if (DstUMvalid) { 1984 TLVec.push_back(ceilingOfQuotient(DstUM - TY, TA)); 1985 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n"); 1986 } 1987 } 1988 1989 if (TLVec.empty() || TUVec.empty()) 1990 return false; 1991 1992 LLVM_DEBUG(dbgs() << "\t TA = " << TA << "\n"); 1993 LLVM_DEBUG(dbgs() << "\t TB = " << TB << "\n"); 1994 1995 TL = APIntOps::smax(TLVec.front(), TLVec.back()); 1996 TU = APIntOps::smin(TUVec.front(), TUVec.back()); 1997 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n"); 1998 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n"); 1999 2000 if (TL.sgt(TU)) 2001 ++ExactRDIVindependence; 2002 return TL.sgt(TU); 2003 } 2004 2005 2006 // symbolicRDIVtest - 2007 // In Section 4.5 of the Practical Dependence Testing paper,the authors 2008 // introduce a special case of Banerjee's Inequalities (also called the 2009 // Extreme-Value Test) that can handle some of the SIV and RDIV cases, 2010 // particularly cases with symbolics. Since it's only able to disprove 2011 // dependence (not compute distances or directions), we'll use it as a 2012 // fall back for the other tests. 2013 // 2014 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*j] 2015 // where i and j are induction variables and c1 and c2 are loop invariants, 2016 // we can use the symbolic tests to disprove some dependences, serving as a 2017 // backup for the RDIV test. Note that i and j can be the same variable, 2018 // letting this test serve as a backup for the various SIV tests. 2019 // 2020 // For a dependence to exist, c1 + a1*i must equal c2 + a2*j for some 2021 // 0 <= i <= N1 and some 0 <= j <= N2, where N1 and N2 are the (normalized) 2022 // loop bounds for the i and j loops, respectively. So, ... 2023 // 2024 // c1 + a1*i = c2 + a2*j 2025 // a1*i - a2*j = c2 - c1 2026 // 2027 // To test for a dependence, we compute c2 - c1 and make sure it's in the 2028 // range of the maximum and minimum possible values of a1*i - a2*j. 2029 // Considering the signs of a1 and a2, we have 4 possible cases: 2030 // 2031 // 1) If a1 >= 0 and a2 >= 0, then 2032 // a1*0 - a2*N2 <= c2 - c1 <= a1*N1 - a2*0 2033 // -a2*N2 <= c2 - c1 <= a1*N1 2034 // 2035 // 2) If a1 >= 0 and a2 <= 0, then 2036 // a1*0 - a2*0 <= c2 - c1 <= a1*N1 - a2*N2 2037 // 0 <= c2 - c1 <= a1*N1 - a2*N2 2038 // 2039 // 3) If a1 <= 0 and a2 >= 0, then 2040 // a1*N1 - a2*N2 <= c2 - c1 <= a1*0 - a2*0 2041 // a1*N1 - a2*N2 <= c2 - c1 <= 0 2042 // 2043 // 4) If a1 <= 0 and a2 <= 0, then 2044 // a1*N1 - a2*0 <= c2 - c1 <= a1*0 - a2*N2 2045 // a1*N1 <= c2 - c1 <= -a2*N2 2046 // 2047 // return true if dependence disproved 2048 bool DependenceInfo::symbolicRDIVtest(const SCEV *A1, const SCEV *A2, 2049 const SCEV *C1, const SCEV *C2, 2050 const Loop *Loop1, 2051 const Loop *Loop2) const { 2052 ++SymbolicRDIVapplications; 2053 LLVM_DEBUG(dbgs() << "\ttry symbolic RDIV test\n"); 2054 LLVM_DEBUG(dbgs() << "\t A1 = " << *A1); 2055 LLVM_DEBUG(dbgs() << ", type = " << *A1->getType() << "\n"); 2056 LLVM_DEBUG(dbgs() << "\t A2 = " << *A2 << "\n"); 2057 LLVM_DEBUG(dbgs() << "\t C1 = " << *C1 << "\n"); 2058 LLVM_DEBUG(dbgs() << "\t C2 = " << *C2 << "\n"); 2059 const SCEV *N1 = collectUpperBound(Loop1, A1->getType()); 2060 const SCEV *N2 = collectUpperBound(Loop2, A1->getType()); 2061 LLVM_DEBUG(if (N1) dbgs() << "\t N1 = " << *N1 << "\n"); 2062 LLVM_DEBUG(if (N2) dbgs() << "\t N2 = " << *N2 << "\n"); 2063 const SCEV *C2_C1 = SE->getMinusSCEV(C2, C1); 2064 const SCEV *C1_C2 = SE->getMinusSCEV(C1, C2); 2065 LLVM_DEBUG(dbgs() << "\t C2 - C1 = " << *C2_C1 << "\n"); 2066 LLVM_DEBUG(dbgs() << "\t C1 - C2 = " << *C1_C2 << "\n"); 2067 if (SE->isKnownNonNegative(A1)) { 2068 if (SE->isKnownNonNegative(A2)) { 2069 // A1 >= 0 && A2 >= 0 2070 if (N1) { 2071 // make sure that c2 - c1 <= a1*N1 2072 const SCEV *A1N1 = SE->getMulExpr(A1, N1); 2073 LLVM_DEBUG(dbgs() << "\t A1*N1 = " << *A1N1 << "\n"); 2074 if (isKnownPredicate(CmpInst::ICMP_SGT, C2_C1, A1N1)) { 2075 ++SymbolicRDIVindependence; 2076 return true; 2077 } 2078 } 2079 if (N2) { 2080 // make sure that -a2*N2 <= c2 - c1, or a2*N2 >= c1 - c2 2081 const SCEV *A2N2 = SE->getMulExpr(A2, N2); 2082 LLVM_DEBUG(dbgs() << "\t A2*N2 = " << *A2N2 << "\n"); 2083 if (isKnownPredicate(CmpInst::ICMP_SLT, A2N2, C1_C2)) { 2084 ++SymbolicRDIVindependence; 2085 return true; 2086 } 2087 } 2088 } 2089 else if (SE->isKnownNonPositive(A2)) { 2090 // a1 >= 0 && a2 <= 0 2091 if (N1 && N2) { 2092 // make sure that c2 - c1 <= a1*N1 - a2*N2 2093 const SCEV *A1N1 = SE->getMulExpr(A1, N1); 2094 const SCEV *A2N2 = SE->getMulExpr(A2, N2); 2095 const SCEV *A1N1_A2N2 = SE->getMinusSCEV(A1N1, A2N2); 2096 LLVM_DEBUG(dbgs() << "\t A1*N1 - A2*N2 = " << *A1N1_A2N2 << "\n"); 2097 if (isKnownPredicate(CmpInst::ICMP_SGT, C2_C1, A1N1_A2N2)) { 2098 ++SymbolicRDIVindependence; 2099 return true; 2100 } 2101 } 2102 // make sure that 0 <= c2 - c1 2103 if (SE->isKnownNegative(C2_C1)) { 2104 ++SymbolicRDIVindependence; 2105 return true; 2106 } 2107 } 2108 } 2109 else if (SE->isKnownNonPositive(A1)) { 2110 if (SE->isKnownNonNegative(A2)) { 2111 // a1 <= 0 && a2 >= 0 2112 if (N1 && N2) { 2113 // make sure that a1*N1 - a2*N2 <= c2 - c1 2114 const SCEV *A1N1 = SE->getMulExpr(A1, N1); 2115 const SCEV *A2N2 = SE->getMulExpr(A2, N2); 2116 const SCEV *A1N1_A2N2 = SE->getMinusSCEV(A1N1, A2N2); 2117 LLVM_DEBUG(dbgs() << "\t A1*N1 - A2*N2 = " << *A1N1_A2N2 << "\n"); 2118 if (isKnownPredicate(CmpInst::ICMP_SGT, A1N1_A2N2, C2_C1)) { 2119 ++SymbolicRDIVindependence; 2120 return true; 2121 } 2122 } 2123 // make sure that c2 - c1 <= 0 2124 if (SE->isKnownPositive(C2_C1)) { 2125 ++SymbolicRDIVindependence; 2126 return true; 2127 } 2128 } 2129 else if (SE->isKnownNonPositive(A2)) { 2130 // a1 <= 0 && a2 <= 0 2131 if (N1) { 2132 // make sure that a1*N1 <= c2 - c1 2133 const SCEV *A1N1 = SE->getMulExpr(A1, N1); 2134 LLVM_DEBUG(dbgs() << "\t A1*N1 = " << *A1N1 << "\n"); 2135 if (isKnownPredicate(CmpInst::ICMP_SGT, A1N1, C2_C1)) { 2136 ++SymbolicRDIVindependence; 2137 return true; 2138 } 2139 } 2140 if (N2) { 2141 // make sure that c2 - c1 <= -a2*N2, or c1 - c2 >= a2*N2 2142 const SCEV *A2N2 = SE->getMulExpr(A2, N2); 2143 LLVM_DEBUG(dbgs() << "\t A2*N2 = " << *A2N2 << "\n"); 2144 if (isKnownPredicate(CmpInst::ICMP_SLT, C1_C2, A2N2)) { 2145 ++SymbolicRDIVindependence; 2146 return true; 2147 } 2148 } 2149 } 2150 } 2151 return false; 2152 } 2153 2154 2155 // testSIV - 2156 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 - a2*i] 2157 // where i is an induction variable, c1 and c2 are loop invariant, and a1 and 2158 // a2 are constant, we attack it with an SIV test. While they can all be 2159 // solved with the Exact SIV test, it's worthwhile to use simpler tests when 2160 // they apply; they're cheaper and sometimes more precise. 2161 // 2162 // Return true if dependence disproved. 2163 bool DependenceInfo::testSIV(const SCEV *Src, const SCEV *Dst, unsigned &Level, 2164 FullDependence &Result, Constraint &NewConstraint, 2165 const SCEV *&SplitIter) const { 2166 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n"); 2167 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n"); 2168 const SCEVAddRecExpr *SrcAddRec = dyn_cast<SCEVAddRecExpr>(Src); 2169 const SCEVAddRecExpr *DstAddRec = dyn_cast<SCEVAddRecExpr>(Dst); 2170 if (SrcAddRec && DstAddRec) { 2171 const SCEV *SrcConst = SrcAddRec->getStart(); 2172 const SCEV *DstConst = DstAddRec->getStart(); 2173 const SCEV *SrcCoeff = SrcAddRec->getStepRecurrence(*SE); 2174 const SCEV *DstCoeff = DstAddRec->getStepRecurrence(*SE); 2175 const Loop *CurLoop = SrcAddRec->getLoop(); 2176 assert(CurLoop == DstAddRec->getLoop() && 2177 "both loops in SIV should be same"); 2178 Level = mapSrcLoop(CurLoop); 2179 bool disproven; 2180 if (SrcCoeff == DstCoeff) 2181 disproven = strongSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop, 2182 Level, Result, NewConstraint); 2183 else if (SrcCoeff == SE->getNegativeSCEV(DstCoeff)) 2184 disproven = weakCrossingSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop, 2185 Level, Result, NewConstraint, SplitIter); 2186 else 2187 disproven = exactSIVtest(SrcCoeff, DstCoeff, SrcConst, DstConst, CurLoop, 2188 Level, Result, NewConstraint); 2189 return disproven || 2190 gcdMIVtest(Src, Dst, Result) || 2191 symbolicRDIVtest(SrcCoeff, DstCoeff, SrcConst, DstConst, CurLoop, CurLoop); 2192 } 2193 if (SrcAddRec) { 2194 const SCEV *SrcConst = SrcAddRec->getStart(); 2195 const SCEV *SrcCoeff = SrcAddRec->getStepRecurrence(*SE); 2196 const SCEV *DstConst = Dst; 2197 const Loop *CurLoop = SrcAddRec->getLoop(); 2198 Level = mapSrcLoop(CurLoop); 2199 return weakZeroDstSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop, 2200 Level, Result, NewConstraint) || 2201 gcdMIVtest(Src, Dst, Result); 2202 } 2203 if (DstAddRec) { 2204 const SCEV *DstConst = DstAddRec->getStart(); 2205 const SCEV *DstCoeff = DstAddRec->getStepRecurrence(*SE); 2206 const SCEV *SrcConst = Src; 2207 const Loop *CurLoop = DstAddRec->getLoop(); 2208 Level = mapDstLoop(CurLoop); 2209 return weakZeroSrcSIVtest(DstCoeff, SrcConst, DstConst, 2210 CurLoop, Level, Result, NewConstraint) || 2211 gcdMIVtest(Src, Dst, Result); 2212 } 2213 llvm_unreachable("SIV test expected at least one AddRec"); 2214 return false; 2215 } 2216 2217 2218 // testRDIV - 2219 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*j] 2220 // where i and j are induction variables, c1 and c2 are loop invariant, 2221 // and a1 and a2 are constant, we can solve it exactly with an easy adaptation 2222 // of the Exact SIV test, the Restricted Double Index Variable (RDIV) test. 2223 // It doesn't make sense to talk about distance or direction in this case, 2224 // so there's no point in making special versions of the Strong SIV test or 2225 // the Weak-crossing SIV test. 2226 // 2227 // With minor algebra, this test can also be used for things like 2228 // [c1 + a1*i + a2*j][c2]. 2229 // 2230 // Return true if dependence disproved. 2231 bool DependenceInfo::testRDIV(const SCEV *Src, const SCEV *Dst, 2232 FullDependence &Result) const { 2233 // we have 3 possible situations here: 2234 // 1) [a*i + b] and [c*j + d] 2235 // 2) [a*i + c*j + b] and [d] 2236 // 3) [b] and [a*i + c*j + d] 2237 // We need to find what we've got and get organized 2238 2239 const SCEV *SrcConst, *DstConst; 2240 const SCEV *SrcCoeff, *DstCoeff; 2241 const Loop *SrcLoop, *DstLoop; 2242 2243 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n"); 2244 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n"); 2245 const SCEVAddRecExpr *SrcAddRec = dyn_cast<SCEVAddRecExpr>(Src); 2246 const SCEVAddRecExpr *DstAddRec = dyn_cast<SCEVAddRecExpr>(Dst); 2247 if (SrcAddRec && DstAddRec) { 2248 SrcConst = SrcAddRec->getStart(); 2249 SrcCoeff = SrcAddRec->getStepRecurrence(*SE); 2250 SrcLoop = SrcAddRec->getLoop(); 2251 DstConst = DstAddRec->getStart(); 2252 DstCoeff = DstAddRec->getStepRecurrence(*SE); 2253 DstLoop = DstAddRec->getLoop(); 2254 } 2255 else if (SrcAddRec) { 2256 if (const SCEVAddRecExpr *tmpAddRec = 2257 dyn_cast<SCEVAddRecExpr>(SrcAddRec->getStart())) { 2258 SrcConst = tmpAddRec->getStart(); 2259 SrcCoeff = tmpAddRec->getStepRecurrence(*SE); 2260 SrcLoop = tmpAddRec->getLoop(); 2261 DstConst = Dst; 2262 DstCoeff = SE->getNegativeSCEV(SrcAddRec->getStepRecurrence(*SE)); 2263 DstLoop = SrcAddRec->getLoop(); 2264 } 2265 else 2266 llvm_unreachable("RDIV reached by surprising SCEVs"); 2267 } 2268 else if (DstAddRec) { 2269 if (const SCEVAddRecExpr *tmpAddRec = 2270 dyn_cast<SCEVAddRecExpr>(DstAddRec->getStart())) { 2271 DstConst = tmpAddRec->getStart(); 2272 DstCoeff = tmpAddRec->getStepRecurrence(*SE); 2273 DstLoop = tmpAddRec->getLoop(); 2274 SrcConst = Src; 2275 SrcCoeff = SE->getNegativeSCEV(DstAddRec->getStepRecurrence(*SE)); 2276 SrcLoop = DstAddRec->getLoop(); 2277 } 2278 else 2279 llvm_unreachable("RDIV reached by surprising SCEVs"); 2280 } 2281 else 2282 llvm_unreachable("RDIV expected at least one AddRec"); 2283 return exactRDIVtest(SrcCoeff, DstCoeff, 2284 SrcConst, DstConst, 2285 SrcLoop, DstLoop, 2286 Result) || 2287 gcdMIVtest(Src, Dst, Result) || 2288 symbolicRDIVtest(SrcCoeff, DstCoeff, 2289 SrcConst, DstConst, 2290 SrcLoop, DstLoop); 2291 } 2292 2293 2294 // Tests the single-subscript MIV pair (Src and Dst) for dependence. 2295 // Return true if dependence disproved. 2296 // Can sometimes refine direction vectors. 2297 bool DependenceInfo::testMIV(const SCEV *Src, const SCEV *Dst, 2298 const SmallBitVector &Loops, 2299 FullDependence &Result) const { 2300 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n"); 2301 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n"); 2302 Result.Consistent = false; 2303 return gcdMIVtest(Src, Dst, Result) || 2304 banerjeeMIVtest(Src, Dst, Loops, Result); 2305 } 2306 2307 2308 // Given a product, e.g., 10*X*Y, returns the first constant operand, 2309 // in this case 10. If there is no constant part, returns NULL. 2310 static 2311 const SCEVConstant *getConstantPart(const SCEV *Expr) { 2312 if (const auto *Constant = dyn_cast<SCEVConstant>(Expr)) 2313 return Constant; 2314 else if (const auto *Product = dyn_cast<SCEVMulExpr>(Expr)) 2315 if (const auto *Constant = dyn_cast<SCEVConstant>(Product->getOperand(0))) 2316 return Constant; 2317 return nullptr; 2318 } 2319 2320 2321 //===----------------------------------------------------------------------===// 2322 // gcdMIVtest - 2323 // Tests an MIV subscript pair for dependence. 2324 // Returns true if any possible dependence is disproved. 2325 // Marks the result as inconsistent. 2326 // Can sometimes disprove the equal direction for 1 or more loops, 2327 // as discussed in Michael Wolfe's book, 2328 // High Performance Compilers for Parallel Computing, page 235. 2329 // 2330 // We spend some effort (code!) to handle cases like 2331 // [10*i + 5*N*j + 15*M + 6], where i and j are induction variables, 2332 // but M and N are just loop-invariant variables. 2333 // This should help us handle linearized subscripts; 2334 // also makes this test a useful backup to the various SIV tests. 2335 // 2336 // It occurs to me that the presence of loop-invariant variables 2337 // changes the nature of the test from "greatest common divisor" 2338 // to "a common divisor". 2339 bool DependenceInfo::gcdMIVtest(const SCEV *Src, const SCEV *Dst, 2340 FullDependence &Result) const { 2341 LLVM_DEBUG(dbgs() << "starting gcd\n"); 2342 ++GCDapplications; 2343 unsigned BitWidth = SE->getTypeSizeInBits(Src->getType()); 2344 APInt RunningGCD = APInt::getZero(BitWidth); 2345 2346 // Examine Src coefficients. 2347 // Compute running GCD and record source constant. 2348 // Because we're looking for the constant at the end of the chain, 2349 // we can't quit the loop just because the GCD == 1. 2350 const SCEV *Coefficients = Src; 2351 while (const SCEVAddRecExpr *AddRec = 2352 dyn_cast<SCEVAddRecExpr>(Coefficients)) { 2353 const SCEV *Coeff = AddRec->getStepRecurrence(*SE); 2354 // If the coefficient is the product of a constant and other stuff, 2355 // we can use the constant in the GCD computation. 2356 const auto *Constant = getConstantPart(Coeff); 2357 if (!Constant) 2358 return false; 2359 APInt ConstCoeff = Constant->getAPInt(); 2360 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs()); 2361 Coefficients = AddRec->getStart(); 2362 } 2363 const SCEV *SrcConst = Coefficients; 2364 2365 // Examine Dst coefficients. 2366 // Compute running GCD and record destination constant. 2367 // Because we're looking for the constant at the end of the chain, 2368 // we can't quit the loop just because the GCD == 1. 2369 Coefficients = Dst; 2370 while (const SCEVAddRecExpr *AddRec = 2371 dyn_cast<SCEVAddRecExpr>(Coefficients)) { 2372 const SCEV *Coeff = AddRec->getStepRecurrence(*SE); 2373 // If the coefficient is the product of a constant and other stuff, 2374 // we can use the constant in the GCD computation. 2375 const auto *Constant = getConstantPart(Coeff); 2376 if (!Constant) 2377 return false; 2378 APInt ConstCoeff = Constant->getAPInt(); 2379 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs()); 2380 Coefficients = AddRec->getStart(); 2381 } 2382 const SCEV *DstConst = Coefficients; 2383 2384 APInt ExtraGCD = APInt::getZero(BitWidth); 2385 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst); 2386 LLVM_DEBUG(dbgs() << " Delta = " << *Delta << "\n"); 2387 const SCEVConstant *Constant = dyn_cast<SCEVConstant>(Delta); 2388 if (const SCEVAddExpr *Sum = dyn_cast<SCEVAddExpr>(Delta)) { 2389 // If Delta is a sum of products, we may be able to make further progress. 2390 for (unsigned Op = 0, Ops = Sum->getNumOperands(); Op < Ops; Op++) { 2391 const SCEV *Operand = Sum->getOperand(Op); 2392 if (isa<SCEVConstant>(Operand)) { 2393 assert(!Constant && "Surprised to find multiple constants"); 2394 Constant = cast<SCEVConstant>(Operand); 2395 } 2396 else if (const SCEVMulExpr *Product = dyn_cast<SCEVMulExpr>(Operand)) { 2397 // Search for constant operand to participate in GCD; 2398 // If none found; return false. 2399 const SCEVConstant *ConstOp = getConstantPart(Product); 2400 if (!ConstOp) 2401 return false; 2402 APInt ConstOpValue = ConstOp->getAPInt(); 2403 ExtraGCD = APIntOps::GreatestCommonDivisor(ExtraGCD, 2404 ConstOpValue.abs()); 2405 } 2406 else 2407 return false; 2408 } 2409 } 2410 if (!Constant) 2411 return false; 2412 APInt ConstDelta = cast<SCEVConstant>(Constant)->getAPInt(); 2413 LLVM_DEBUG(dbgs() << " ConstDelta = " << ConstDelta << "\n"); 2414 if (ConstDelta == 0) 2415 return false; 2416 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ExtraGCD); 2417 LLVM_DEBUG(dbgs() << " RunningGCD = " << RunningGCD << "\n"); 2418 APInt Remainder = ConstDelta.srem(RunningGCD); 2419 if (Remainder != 0) { 2420 ++GCDindependence; 2421 return true; 2422 } 2423 2424 // Try to disprove equal directions. 2425 // For example, given a subscript pair [3*i + 2*j] and [i' + 2*j' - 1], 2426 // the code above can't disprove the dependence because the GCD = 1. 2427 // So we consider what happen if i = i' and what happens if j = j'. 2428 // If i = i', we can simplify the subscript to [2*i + 2*j] and [2*j' - 1], 2429 // which is infeasible, so we can disallow the = direction for the i level. 2430 // Setting j = j' doesn't help matters, so we end up with a direction vector 2431 // of [<>, *] 2432 // 2433 // Given A[5*i + 10*j*M + 9*M*N] and A[15*i + 20*j*M - 21*N*M + 5], 2434 // we need to remember that the constant part is 5 and the RunningGCD should 2435 // be initialized to ExtraGCD = 30. 2436 LLVM_DEBUG(dbgs() << " ExtraGCD = " << ExtraGCD << '\n'); 2437 2438 bool Improved = false; 2439 Coefficients = Src; 2440 while (const SCEVAddRecExpr *AddRec = 2441 dyn_cast<SCEVAddRecExpr>(Coefficients)) { 2442 Coefficients = AddRec->getStart(); 2443 const Loop *CurLoop = AddRec->getLoop(); 2444 RunningGCD = ExtraGCD; 2445 const SCEV *SrcCoeff = AddRec->getStepRecurrence(*SE); 2446 const SCEV *DstCoeff = SE->getMinusSCEV(SrcCoeff, SrcCoeff); 2447 const SCEV *Inner = Src; 2448 while (RunningGCD != 1 && isa<SCEVAddRecExpr>(Inner)) { 2449 AddRec = cast<SCEVAddRecExpr>(Inner); 2450 const SCEV *Coeff = AddRec->getStepRecurrence(*SE); 2451 if (CurLoop == AddRec->getLoop()) 2452 ; // SrcCoeff == Coeff 2453 else { 2454 // If the coefficient is the product of a constant and other stuff, 2455 // we can use the constant in the GCD computation. 2456 Constant = getConstantPart(Coeff); 2457 if (!Constant) 2458 return false; 2459 APInt ConstCoeff = Constant->getAPInt(); 2460 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs()); 2461 } 2462 Inner = AddRec->getStart(); 2463 } 2464 Inner = Dst; 2465 while (RunningGCD != 1 && isa<SCEVAddRecExpr>(Inner)) { 2466 AddRec = cast<SCEVAddRecExpr>(Inner); 2467 const SCEV *Coeff = AddRec->getStepRecurrence(*SE); 2468 if (CurLoop == AddRec->getLoop()) 2469 DstCoeff = Coeff; 2470 else { 2471 // If the coefficient is the product of a constant and other stuff, 2472 // we can use the constant in the GCD computation. 2473 Constant = getConstantPart(Coeff); 2474 if (!Constant) 2475 return false; 2476 APInt ConstCoeff = Constant->getAPInt(); 2477 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs()); 2478 } 2479 Inner = AddRec->getStart(); 2480 } 2481 Delta = SE->getMinusSCEV(SrcCoeff, DstCoeff); 2482 // If the coefficient is the product of a constant and other stuff, 2483 // we can use the constant in the GCD computation. 2484 Constant = getConstantPart(Delta); 2485 if (!Constant) 2486 // The difference of the two coefficients might not be a product 2487 // or constant, in which case we give up on this direction. 2488 continue; 2489 APInt ConstCoeff = Constant->getAPInt(); 2490 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs()); 2491 LLVM_DEBUG(dbgs() << "\tRunningGCD = " << RunningGCD << "\n"); 2492 if (RunningGCD != 0) { 2493 Remainder = ConstDelta.srem(RunningGCD); 2494 LLVM_DEBUG(dbgs() << "\tRemainder = " << Remainder << "\n"); 2495 if (Remainder != 0) { 2496 unsigned Level = mapSrcLoop(CurLoop); 2497 Result.DV[Level - 1].Direction &= unsigned(~Dependence::DVEntry::EQ); 2498 Improved = true; 2499 } 2500 } 2501 } 2502 if (Improved) 2503 ++GCDsuccesses; 2504 LLVM_DEBUG(dbgs() << "all done\n"); 2505 return false; 2506 } 2507 2508 2509 //===----------------------------------------------------------------------===// 2510 // banerjeeMIVtest - 2511 // Use Banerjee's Inequalities to test an MIV subscript pair. 2512 // (Wolfe, in the race-car book, calls this the Extreme Value Test.) 2513 // Generally follows the discussion in Section 2.5.2 of 2514 // 2515 // Optimizing Supercompilers for Supercomputers 2516 // Michael Wolfe 2517 // 2518 // The inequalities given on page 25 are simplified in that loops are 2519 // normalized so that the lower bound is always 0 and the stride is always 1. 2520 // For example, Wolfe gives 2521 // 2522 // LB^<_k = (A^-_k - B_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k 2523 // 2524 // where A_k is the coefficient of the kth index in the source subscript, 2525 // B_k is the coefficient of the kth index in the destination subscript, 2526 // U_k is the upper bound of the kth index, L_k is the lower bound of the Kth 2527 // index, and N_k is the stride of the kth index. Since all loops are normalized 2528 // by the SCEV package, N_k = 1 and L_k = 0, allowing us to simplify the 2529 // equation to 2530 // 2531 // LB^<_k = (A^-_k - B_k)^- (U_k - 0 - 1) + (A_k - B_k)0 - B_k 1 2532 // = (A^-_k - B_k)^- (U_k - 1) - B_k 2533 // 2534 // Similar simplifications are possible for the other equations. 2535 // 2536 // When we can't determine the number of iterations for a loop, 2537 // we use NULL as an indicator for the worst case, infinity. 2538 // When computing the upper bound, NULL denotes +inf; 2539 // for the lower bound, NULL denotes -inf. 2540 // 2541 // Return true if dependence disproved. 2542 bool DependenceInfo::banerjeeMIVtest(const SCEV *Src, const SCEV *Dst, 2543 const SmallBitVector &Loops, 2544 FullDependence &Result) const { 2545 LLVM_DEBUG(dbgs() << "starting Banerjee\n"); 2546 ++BanerjeeApplications; 2547 LLVM_DEBUG(dbgs() << " Src = " << *Src << '\n'); 2548 const SCEV *A0; 2549 CoefficientInfo *A = collectCoeffInfo(Src, true, A0); 2550 LLVM_DEBUG(dbgs() << " Dst = " << *Dst << '\n'); 2551 const SCEV *B0; 2552 CoefficientInfo *B = collectCoeffInfo(Dst, false, B0); 2553 BoundInfo *Bound = new BoundInfo[MaxLevels + 1]; 2554 const SCEV *Delta = SE->getMinusSCEV(B0, A0); 2555 LLVM_DEBUG(dbgs() << "\tDelta = " << *Delta << '\n'); 2556 2557 // Compute bounds for all the * directions. 2558 LLVM_DEBUG(dbgs() << "\tBounds[*]\n"); 2559 for (unsigned K = 1; K <= MaxLevels; ++K) { 2560 Bound[K].Iterations = A[K].Iterations ? A[K].Iterations : B[K].Iterations; 2561 Bound[K].Direction = Dependence::DVEntry::ALL; 2562 Bound[K].DirSet = Dependence::DVEntry::NONE; 2563 findBoundsALL(A, B, Bound, K); 2564 #ifndef NDEBUG 2565 LLVM_DEBUG(dbgs() << "\t " << K << '\t'); 2566 if (Bound[K].Lower[Dependence::DVEntry::ALL]) 2567 LLVM_DEBUG(dbgs() << *Bound[K].Lower[Dependence::DVEntry::ALL] << '\t'); 2568 else 2569 LLVM_DEBUG(dbgs() << "-inf\t"); 2570 if (Bound[K].Upper[Dependence::DVEntry::ALL]) 2571 LLVM_DEBUG(dbgs() << *Bound[K].Upper[Dependence::DVEntry::ALL] << '\n'); 2572 else 2573 LLVM_DEBUG(dbgs() << "+inf\n"); 2574 #endif 2575 } 2576 2577 // Test the *, *, *, ... case. 2578 bool Disproved = false; 2579 if (testBounds(Dependence::DVEntry::ALL, 0, Bound, Delta)) { 2580 // Explore the direction vector hierarchy. 2581 unsigned DepthExpanded = 0; 2582 unsigned NewDeps = exploreDirections(1, A, B, Bound, 2583 Loops, DepthExpanded, Delta); 2584 if (NewDeps > 0) { 2585 bool Improved = false; 2586 for (unsigned K = 1; K <= CommonLevels; ++K) { 2587 if (Loops[K]) { 2588 unsigned Old = Result.DV[K - 1].Direction; 2589 Result.DV[K - 1].Direction = Old & Bound[K].DirSet; 2590 Improved |= Old != Result.DV[K - 1].Direction; 2591 if (!Result.DV[K - 1].Direction) { 2592 Improved = false; 2593 Disproved = true; 2594 break; 2595 } 2596 } 2597 } 2598 if (Improved) 2599 ++BanerjeeSuccesses; 2600 } 2601 else { 2602 ++BanerjeeIndependence; 2603 Disproved = true; 2604 } 2605 } 2606 else { 2607 ++BanerjeeIndependence; 2608 Disproved = true; 2609 } 2610 delete [] Bound; 2611 delete [] A; 2612 delete [] B; 2613 return Disproved; 2614 } 2615 2616 2617 // Hierarchically expands the direction vector 2618 // search space, combining the directions of discovered dependences 2619 // in the DirSet field of Bound. Returns the number of distinct 2620 // dependences discovered. If the dependence is disproved, 2621 // it will return 0. 2622 unsigned DependenceInfo::exploreDirections(unsigned Level, CoefficientInfo *A, 2623 CoefficientInfo *B, BoundInfo *Bound, 2624 const SmallBitVector &Loops, 2625 unsigned &DepthExpanded, 2626 const SCEV *Delta) const { 2627 // This algorithm has worst case complexity of O(3^n), where 'n' is the number 2628 // of common loop levels. To avoid excessive compile-time, pessimize all the 2629 // results and immediately return when the number of common levels is beyond 2630 // the given threshold. 2631 if (CommonLevels > MIVMaxLevelThreshold) { 2632 LLVM_DEBUG(dbgs() << "Number of common levels exceeded the threshold. MIV " 2633 "direction exploration is terminated.\n"); 2634 for (unsigned K = 1; K <= CommonLevels; ++K) 2635 if (Loops[K]) 2636 Bound[K].DirSet = Dependence::DVEntry::ALL; 2637 return 1; 2638 } 2639 2640 if (Level > CommonLevels) { 2641 // record result 2642 LLVM_DEBUG(dbgs() << "\t["); 2643 for (unsigned K = 1; K <= CommonLevels; ++K) { 2644 if (Loops[K]) { 2645 Bound[K].DirSet |= Bound[K].Direction; 2646 #ifndef NDEBUG 2647 switch (Bound[K].Direction) { 2648 case Dependence::DVEntry::LT: 2649 LLVM_DEBUG(dbgs() << " <"); 2650 break; 2651 case Dependence::DVEntry::EQ: 2652 LLVM_DEBUG(dbgs() << " ="); 2653 break; 2654 case Dependence::DVEntry::GT: 2655 LLVM_DEBUG(dbgs() << " >"); 2656 break; 2657 case Dependence::DVEntry::ALL: 2658 LLVM_DEBUG(dbgs() << " *"); 2659 break; 2660 default: 2661 llvm_unreachable("unexpected Bound[K].Direction"); 2662 } 2663 #endif 2664 } 2665 } 2666 LLVM_DEBUG(dbgs() << " ]\n"); 2667 return 1; 2668 } 2669 if (Loops[Level]) { 2670 if (Level > DepthExpanded) { 2671 DepthExpanded = Level; 2672 // compute bounds for <, =, > at current level 2673 findBoundsLT(A, B, Bound, Level); 2674 findBoundsGT(A, B, Bound, Level); 2675 findBoundsEQ(A, B, Bound, Level); 2676 #ifndef NDEBUG 2677 LLVM_DEBUG(dbgs() << "\tBound for level = " << Level << '\n'); 2678 LLVM_DEBUG(dbgs() << "\t <\t"); 2679 if (Bound[Level].Lower[Dependence::DVEntry::LT]) 2680 LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::LT] 2681 << '\t'); 2682 else 2683 LLVM_DEBUG(dbgs() << "-inf\t"); 2684 if (Bound[Level].Upper[Dependence::DVEntry::LT]) 2685 LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::LT] 2686 << '\n'); 2687 else 2688 LLVM_DEBUG(dbgs() << "+inf\n"); 2689 LLVM_DEBUG(dbgs() << "\t =\t"); 2690 if (Bound[Level].Lower[Dependence::DVEntry::EQ]) 2691 LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::EQ] 2692 << '\t'); 2693 else 2694 LLVM_DEBUG(dbgs() << "-inf\t"); 2695 if (Bound[Level].Upper[Dependence::DVEntry::EQ]) 2696 LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::EQ] 2697 << '\n'); 2698 else 2699 LLVM_DEBUG(dbgs() << "+inf\n"); 2700 LLVM_DEBUG(dbgs() << "\t >\t"); 2701 if (Bound[Level].Lower[Dependence::DVEntry::GT]) 2702 LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::GT] 2703 << '\t'); 2704 else 2705 LLVM_DEBUG(dbgs() << "-inf\t"); 2706 if (Bound[Level].Upper[Dependence::DVEntry::GT]) 2707 LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::GT] 2708 << '\n'); 2709 else 2710 LLVM_DEBUG(dbgs() << "+inf\n"); 2711 #endif 2712 } 2713 2714 unsigned NewDeps = 0; 2715 2716 // test bounds for <, *, *, ... 2717 if (testBounds(Dependence::DVEntry::LT, Level, Bound, Delta)) 2718 NewDeps += exploreDirections(Level + 1, A, B, Bound, 2719 Loops, DepthExpanded, Delta); 2720 2721 // Test bounds for =, *, *, ... 2722 if (testBounds(Dependence::DVEntry::EQ, Level, Bound, Delta)) 2723 NewDeps += exploreDirections(Level + 1, A, B, Bound, 2724 Loops, DepthExpanded, Delta); 2725 2726 // test bounds for >, *, *, ... 2727 if (testBounds(Dependence::DVEntry::GT, Level, Bound, Delta)) 2728 NewDeps += exploreDirections(Level + 1, A, B, Bound, 2729 Loops, DepthExpanded, Delta); 2730 2731 Bound[Level].Direction = Dependence::DVEntry::ALL; 2732 return NewDeps; 2733 } 2734 else 2735 return exploreDirections(Level + 1, A, B, Bound, Loops, DepthExpanded, Delta); 2736 } 2737 2738 2739 // Returns true iff the current bounds are plausible. 2740 bool DependenceInfo::testBounds(unsigned char DirKind, unsigned Level, 2741 BoundInfo *Bound, const SCEV *Delta) const { 2742 Bound[Level].Direction = DirKind; 2743 if (const SCEV *LowerBound = getLowerBound(Bound)) 2744 if (isKnownPredicate(CmpInst::ICMP_SGT, LowerBound, Delta)) 2745 return false; 2746 if (const SCEV *UpperBound = getUpperBound(Bound)) 2747 if (isKnownPredicate(CmpInst::ICMP_SGT, Delta, UpperBound)) 2748 return false; 2749 return true; 2750 } 2751 2752 2753 // Computes the upper and lower bounds for level K 2754 // using the * direction. Records them in Bound. 2755 // Wolfe gives the equations 2756 // 2757 // LB^*_k = (A^-_k - B^+_k)(U_k - L_k) + (A_k - B_k)L_k 2758 // UB^*_k = (A^+_k - B^-_k)(U_k - L_k) + (A_k - B_k)L_k 2759 // 2760 // Since we normalize loops, we can simplify these equations to 2761 // 2762 // LB^*_k = (A^-_k - B^+_k)U_k 2763 // UB^*_k = (A^+_k - B^-_k)U_k 2764 // 2765 // We must be careful to handle the case where the upper bound is unknown. 2766 // Note that the lower bound is always <= 0 2767 // and the upper bound is always >= 0. 2768 void DependenceInfo::findBoundsALL(CoefficientInfo *A, CoefficientInfo *B, 2769 BoundInfo *Bound, unsigned K) const { 2770 Bound[K].Lower[Dependence::DVEntry::ALL] = nullptr; // Default value = -infinity. 2771 Bound[K].Upper[Dependence::DVEntry::ALL] = nullptr; // Default value = +infinity. 2772 if (Bound[K].Iterations) { 2773 Bound[K].Lower[Dependence::DVEntry::ALL] = 2774 SE->getMulExpr(SE->getMinusSCEV(A[K].NegPart, B[K].PosPart), 2775 Bound[K].Iterations); 2776 Bound[K].Upper[Dependence::DVEntry::ALL] = 2777 SE->getMulExpr(SE->getMinusSCEV(A[K].PosPart, B[K].NegPart), 2778 Bound[K].Iterations); 2779 } 2780 else { 2781 // If the difference is 0, we won't need to know the number of iterations. 2782 if (isKnownPredicate(CmpInst::ICMP_EQ, A[K].NegPart, B[K].PosPart)) 2783 Bound[K].Lower[Dependence::DVEntry::ALL] = 2784 SE->getZero(A[K].Coeff->getType()); 2785 if (isKnownPredicate(CmpInst::ICMP_EQ, A[K].PosPart, B[K].NegPart)) 2786 Bound[K].Upper[Dependence::DVEntry::ALL] = 2787 SE->getZero(A[K].Coeff->getType()); 2788 } 2789 } 2790 2791 2792 // Computes the upper and lower bounds for level K 2793 // using the = direction. Records them in Bound. 2794 // Wolfe gives the equations 2795 // 2796 // LB^=_k = (A_k - B_k)^- (U_k - L_k) + (A_k - B_k)L_k 2797 // UB^=_k = (A_k - B_k)^+ (U_k - L_k) + (A_k - B_k)L_k 2798 // 2799 // Since we normalize loops, we can simplify these equations to 2800 // 2801 // LB^=_k = (A_k - B_k)^- U_k 2802 // UB^=_k = (A_k - B_k)^+ U_k 2803 // 2804 // We must be careful to handle the case where the upper bound is unknown. 2805 // Note that the lower bound is always <= 0 2806 // and the upper bound is always >= 0. 2807 void DependenceInfo::findBoundsEQ(CoefficientInfo *A, CoefficientInfo *B, 2808 BoundInfo *Bound, unsigned K) const { 2809 Bound[K].Lower[Dependence::DVEntry::EQ] = nullptr; // Default value = -infinity. 2810 Bound[K].Upper[Dependence::DVEntry::EQ] = nullptr; // Default value = +infinity. 2811 if (Bound[K].Iterations) { 2812 const SCEV *Delta = SE->getMinusSCEV(A[K].Coeff, B[K].Coeff); 2813 const SCEV *NegativePart = getNegativePart(Delta); 2814 Bound[K].Lower[Dependence::DVEntry::EQ] = 2815 SE->getMulExpr(NegativePart, Bound[K].Iterations); 2816 const SCEV *PositivePart = getPositivePart(Delta); 2817 Bound[K].Upper[Dependence::DVEntry::EQ] = 2818 SE->getMulExpr(PositivePart, Bound[K].Iterations); 2819 } 2820 else { 2821 // If the positive/negative part of the difference is 0, 2822 // we won't need to know the number of iterations. 2823 const SCEV *Delta = SE->getMinusSCEV(A[K].Coeff, B[K].Coeff); 2824 const SCEV *NegativePart = getNegativePart(Delta); 2825 if (NegativePart->isZero()) 2826 Bound[K].Lower[Dependence::DVEntry::EQ] = NegativePart; // Zero 2827 const SCEV *PositivePart = getPositivePart(Delta); 2828 if (PositivePart->isZero()) 2829 Bound[K].Upper[Dependence::DVEntry::EQ] = PositivePart; // Zero 2830 } 2831 } 2832 2833 2834 // Computes the upper and lower bounds for level K 2835 // using the < direction. Records them in Bound. 2836 // Wolfe gives the equations 2837 // 2838 // LB^<_k = (A^-_k - B_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k 2839 // UB^<_k = (A^+_k - B_k)^+ (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k 2840 // 2841 // Since we normalize loops, we can simplify these equations to 2842 // 2843 // LB^<_k = (A^-_k - B_k)^- (U_k - 1) - B_k 2844 // UB^<_k = (A^+_k - B_k)^+ (U_k - 1) - B_k 2845 // 2846 // We must be careful to handle the case where the upper bound is unknown. 2847 void DependenceInfo::findBoundsLT(CoefficientInfo *A, CoefficientInfo *B, 2848 BoundInfo *Bound, unsigned K) const { 2849 Bound[K].Lower[Dependence::DVEntry::LT] = nullptr; // Default value = -infinity. 2850 Bound[K].Upper[Dependence::DVEntry::LT] = nullptr; // Default value = +infinity. 2851 if (Bound[K].Iterations) { 2852 const SCEV *Iter_1 = SE->getMinusSCEV( 2853 Bound[K].Iterations, SE->getOne(Bound[K].Iterations->getType())); 2854 const SCEV *NegPart = 2855 getNegativePart(SE->getMinusSCEV(A[K].NegPart, B[K].Coeff)); 2856 Bound[K].Lower[Dependence::DVEntry::LT] = 2857 SE->getMinusSCEV(SE->getMulExpr(NegPart, Iter_1), B[K].Coeff); 2858 const SCEV *PosPart = 2859 getPositivePart(SE->getMinusSCEV(A[K].PosPart, B[K].Coeff)); 2860 Bound[K].Upper[Dependence::DVEntry::LT] = 2861 SE->getMinusSCEV(SE->getMulExpr(PosPart, Iter_1), B[K].Coeff); 2862 } 2863 else { 2864 // If the positive/negative part of the difference is 0, 2865 // we won't need to know the number of iterations. 2866 const SCEV *NegPart = 2867 getNegativePart(SE->getMinusSCEV(A[K].NegPart, B[K].Coeff)); 2868 if (NegPart->isZero()) 2869 Bound[K].Lower[Dependence::DVEntry::LT] = SE->getNegativeSCEV(B[K].Coeff); 2870 const SCEV *PosPart = 2871 getPositivePart(SE->getMinusSCEV(A[K].PosPart, B[K].Coeff)); 2872 if (PosPart->isZero()) 2873 Bound[K].Upper[Dependence::DVEntry::LT] = SE->getNegativeSCEV(B[K].Coeff); 2874 } 2875 } 2876 2877 2878 // Computes the upper and lower bounds for level K 2879 // using the > direction. Records them in Bound. 2880 // Wolfe gives the equations 2881 // 2882 // LB^>_k = (A_k - B^+_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k + A_k N_k 2883 // UB^>_k = (A_k - B^-_k)^+ (U_k - L_k - N_k) + (A_k - B_k)L_k + A_k N_k 2884 // 2885 // Since we normalize loops, we can simplify these equations to 2886 // 2887 // LB^>_k = (A_k - B^+_k)^- (U_k - 1) + A_k 2888 // UB^>_k = (A_k - B^-_k)^+ (U_k - 1) + A_k 2889 // 2890 // We must be careful to handle the case where the upper bound is unknown. 2891 void DependenceInfo::findBoundsGT(CoefficientInfo *A, CoefficientInfo *B, 2892 BoundInfo *Bound, unsigned K) const { 2893 Bound[K].Lower[Dependence::DVEntry::GT] = nullptr; // Default value = -infinity. 2894 Bound[K].Upper[Dependence::DVEntry::GT] = nullptr; // Default value = +infinity. 2895 if (Bound[K].Iterations) { 2896 const SCEV *Iter_1 = SE->getMinusSCEV( 2897 Bound[K].Iterations, SE->getOne(Bound[K].Iterations->getType())); 2898 const SCEV *NegPart = 2899 getNegativePart(SE->getMinusSCEV(A[K].Coeff, B[K].PosPart)); 2900 Bound[K].Lower[Dependence::DVEntry::GT] = 2901 SE->getAddExpr(SE->getMulExpr(NegPart, Iter_1), A[K].Coeff); 2902 const SCEV *PosPart = 2903 getPositivePart(SE->getMinusSCEV(A[K].Coeff, B[K].NegPart)); 2904 Bound[K].Upper[Dependence::DVEntry::GT] = 2905 SE->getAddExpr(SE->getMulExpr(PosPart, Iter_1), A[K].Coeff); 2906 } 2907 else { 2908 // If the positive/negative part of the difference is 0, 2909 // we won't need to know the number of iterations. 2910 const SCEV *NegPart = getNegativePart(SE->getMinusSCEV(A[K].Coeff, B[K].PosPart)); 2911 if (NegPart->isZero()) 2912 Bound[K].Lower[Dependence::DVEntry::GT] = A[K].Coeff; 2913 const SCEV *PosPart = getPositivePart(SE->getMinusSCEV(A[K].Coeff, B[K].NegPart)); 2914 if (PosPart->isZero()) 2915 Bound[K].Upper[Dependence::DVEntry::GT] = A[K].Coeff; 2916 } 2917 } 2918 2919 2920 // X^+ = max(X, 0) 2921 const SCEV *DependenceInfo::getPositivePart(const SCEV *X) const { 2922 return SE->getSMaxExpr(X, SE->getZero(X->getType())); 2923 } 2924 2925 2926 // X^- = min(X, 0) 2927 const SCEV *DependenceInfo::getNegativePart(const SCEV *X) const { 2928 return SE->getSMinExpr(X, SE->getZero(X->getType())); 2929 } 2930 2931 2932 // Walks through the subscript, 2933 // collecting each coefficient, the associated loop bounds, 2934 // and recording its positive and negative parts for later use. 2935 DependenceInfo::CoefficientInfo * 2936 DependenceInfo::collectCoeffInfo(const SCEV *Subscript, bool SrcFlag, 2937 const SCEV *&Constant) const { 2938 const SCEV *Zero = SE->getZero(Subscript->getType()); 2939 CoefficientInfo *CI = new CoefficientInfo[MaxLevels + 1]; 2940 for (unsigned K = 1; K <= MaxLevels; ++K) { 2941 CI[K].Coeff = Zero; 2942 CI[K].PosPart = Zero; 2943 CI[K].NegPart = Zero; 2944 CI[K].Iterations = nullptr; 2945 } 2946 while (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Subscript)) { 2947 const Loop *L = AddRec->getLoop(); 2948 unsigned K = SrcFlag ? mapSrcLoop(L) : mapDstLoop(L); 2949 CI[K].Coeff = AddRec->getStepRecurrence(*SE); 2950 CI[K].PosPart = getPositivePart(CI[K].Coeff); 2951 CI[K].NegPart = getNegativePart(CI[K].Coeff); 2952 CI[K].Iterations = collectUpperBound(L, Subscript->getType()); 2953 Subscript = AddRec->getStart(); 2954 } 2955 Constant = Subscript; 2956 #ifndef NDEBUG 2957 LLVM_DEBUG(dbgs() << "\tCoefficient Info\n"); 2958 for (unsigned K = 1; K <= MaxLevels; ++K) { 2959 LLVM_DEBUG(dbgs() << "\t " << K << "\t" << *CI[K].Coeff); 2960 LLVM_DEBUG(dbgs() << "\tPos Part = "); 2961 LLVM_DEBUG(dbgs() << *CI[K].PosPart); 2962 LLVM_DEBUG(dbgs() << "\tNeg Part = "); 2963 LLVM_DEBUG(dbgs() << *CI[K].NegPart); 2964 LLVM_DEBUG(dbgs() << "\tUpper Bound = "); 2965 if (CI[K].Iterations) 2966 LLVM_DEBUG(dbgs() << *CI[K].Iterations); 2967 else 2968 LLVM_DEBUG(dbgs() << "+inf"); 2969 LLVM_DEBUG(dbgs() << '\n'); 2970 } 2971 LLVM_DEBUG(dbgs() << "\t Constant = " << *Subscript << '\n'); 2972 #endif 2973 return CI; 2974 } 2975 2976 2977 // Looks through all the bounds info and 2978 // computes the lower bound given the current direction settings 2979 // at each level. If the lower bound for any level is -inf, 2980 // the result is -inf. 2981 const SCEV *DependenceInfo::getLowerBound(BoundInfo *Bound) const { 2982 const SCEV *Sum = Bound[1].Lower[Bound[1].Direction]; 2983 for (unsigned K = 2; Sum && K <= MaxLevels; ++K) { 2984 if (Bound[K].Lower[Bound[K].Direction]) 2985 Sum = SE->getAddExpr(Sum, Bound[K].Lower[Bound[K].Direction]); 2986 else 2987 Sum = nullptr; 2988 } 2989 return Sum; 2990 } 2991 2992 2993 // Looks through all the bounds info and 2994 // computes the upper bound given the current direction settings 2995 // at each level. If the upper bound at any level is +inf, 2996 // the result is +inf. 2997 const SCEV *DependenceInfo::getUpperBound(BoundInfo *Bound) const { 2998 const SCEV *Sum = Bound[1].Upper[Bound[1].Direction]; 2999 for (unsigned K = 2; Sum && K <= MaxLevels; ++K) { 3000 if (Bound[K].Upper[Bound[K].Direction]) 3001 Sum = SE->getAddExpr(Sum, Bound[K].Upper[Bound[K].Direction]); 3002 else 3003 Sum = nullptr; 3004 } 3005 return Sum; 3006 } 3007 3008 3009 //===----------------------------------------------------------------------===// 3010 // Constraint manipulation for Delta test. 3011 3012 // Given a linear SCEV, 3013 // return the coefficient (the step) 3014 // corresponding to the specified loop. 3015 // If there isn't one, return 0. 3016 // For example, given a*i + b*j + c*k, finding the coefficient 3017 // corresponding to the j loop would yield b. 3018 const SCEV *DependenceInfo::findCoefficient(const SCEV *Expr, 3019 const Loop *TargetLoop) const { 3020 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr); 3021 if (!AddRec) 3022 return SE->getZero(Expr->getType()); 3023 if (AddRec->getLoop() == TargetLoop) 3024 return AddRec->getStepRecurrence(*SE); 3025 return findCoefficient(AddRec->getStart(), TargetLoop); 3026 } 3027 3028 3029 // Given a linear SCEV, 3030 // return the SCEV given by zeroing out the coefficient 3031 // corresponding to the specified loop. 3032 // For example, given a*i + b*j + c*k, zeroing the coefficient 3033 // corresponding to the j loop would yield a*i + c*k. 3034 const SCEV *DependenceInfo::zeroCoefficient(const SCEV *Expr, 3035 const Loop *TargetLoop) const { 3036 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr); 3037 if (!AddRec) 3038 return Expr; // ignore 3039 if (AddRec->getLoop() == TargetLoop) 3040 return AddRec->getStart(); 3041 return SE->getAddRecExpr(zeroCoefficient(AddRec->getStart(), TargetLoop), 3042 AddRec->getStepRecurrence(*SE), 3043 AddRec->getLoop(), 3044 AddRec->getNoWrapFlags()); 3045 } 3046 3047 3048 // Given a linear SCEV Expr, 3049 // return the SCEV given by adding some Value to the 3050 // coefficient corresponding to the specified TargetLoop. 3051 // For example, given a*i + b*j + c*k, adding 1 to the coefficient 3052 // corresponding to the j loop would yield a*i + (b+1)*j + c*k. 3053 const SCEV *DependenceInfo::addToCoefficient(const SCEV *Expr, 3054 const Loop *TargetLoop, 3055 const SCEV *Value) const { 3056 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr); 3057 if (!AddRec) // create a new addRec 3058 return SE->getAddRecExpr(Expr, 3059 Value, 3060 TargetLoop, 3061 SCEV::FlagAnyWrap); // Worst case, with no info. 3062 if (AddRec->getLoop() == TargetLoop) { 3063 const SCEV *Sum = SE->getAddExpr(AddRec->getStepRecurrence(*SE), Value); 3064 if (Sum->isZero()) 3065 return AddRec->getStart(); 3066 return SE->getAddRecExpr(AddRec->getStart(), 3067 Sum, 3068 AddRec->getLoop(), 3069 AddRec->getNoWrapFlags()); 3070 } 3071 if (SE->isLoopInvariant(AddRec, TargetLoop)) 3072 return SE->getAddRecExpr(AddRec, Value, TargetLoop, SCEV::FlagAnyWrap); 3073 return SE->getAddRecExpr( 3074 addToCoefficient(AddRec->getStart(), TargetLoop, Value), 3075 AddRec->getStepRecurrence(*SE), AddRec->getLoop(), 3076 AddRec->getNoWrapFlags()); 3077 } 3078 3079 3080 // Review the constraints, looking for opportunities 3081 // to simplify a subscript pair (Src and Dst). 3082 // Return true if some simplification occurs. 3083 // If the simplification isn't exact (that is, if it is conservative 3084 // in terms of dependence), set consistent to false. 3085 // Corresponds to Figure 5 from the paper 3086 // 3087 // Practical Dependence Testing 3088 // Goff, Kennedy, Tseng 3089 // PLDI 1991 3090 bool DependenceInfo::propagate(const SCEV *&Src, const SCEV *&Dst, 3091 SmallBitVector &Loops, 3092 SmallVectorImpl<Constraint> &Constraints, 3093 bool &Consistent) { 3094 bool Result = false; 3095 for (unsigned LI : Loops.set_bits()) { 3096 LLVM_DEBUG(dbgs() << "\t Constraint[" << LI << "] is"); 3097 LLVM_DEBUG(Constraints[LI].dump(dbgs())); 3098 if (Constraints[LI].isDistance()) 3099 Result |= propagateDistance(Src, Dst, Constraints[LI], Consistent); 3100 else if (Constraints[LI].isLine()) 3101 Result |= propagateLine(Src, Dst, Constraints[LI], Consistent); 3102 else if (Constraints[LI].isPoint()) 3103 Result |= propagatePoint(Src, Dst, Constraints[LI]); 3104 } 3105 return Result; 3106 } 3107 3108 3109 // Attempt to propagate a distance 3110 // constraint into a subscript pair (Src and Dst). 3111 // Return true if some simplification occurs. 3112 // If the simplification isn't exact (that is, if it is conservative 3113 // in terms of dependence), set consistent to false. 3114 bool DependenceInfo::propagateDistance(const SCEV *&Src, const SCEV *&Dst, 3115 Constraint &CurConstraint, 3116 bool &Consistent) { 3117 const Loop *CurLoop = CurConstraint.getAssociatedLoop(); 3118 LLVM_DEBUG(dbgs() << "\t\tSrc is " << *Src << "\n"); 3119 const SCEV *A_K = findCoefficient(Src, CurLoop); 3120 if (A_K->isZero()) 3121 return false; 3122 const SCEV *DA_K = SE->getMulExpr(A_K, CurConstraint.getD()); 3123 Src = SE->getMinusSCEV(Src, DA_K); 3124 Src = zeroCoefficient(Src, CurLoop); 3125 LLVM_DEBUG(dbgs() << "\t\tnew Src is " << *Src << "\n"); 3126 LLVM_DEBUG(dbgs() << "\t\tDst is " << *Dst << "\n"); 3127 Dst = addToCoefficient(Dst, CurLoop, SE->getNegativeSCEV(A_K)); 3128 LLVM_DEBUG(dbgs() << "\t\tnew Dst is " << *Dst << "\n"); 3129 if (!findCoefficient(Dst, CurLoop)->isZero()) 3130 Consistent = false; 3131 return true; 3132 } 3133 3134 3135 // Attempt to propagate a line 3136 // constraint into a subscript pair (Src and Dst). 3137 // Return true if some simplification occurs. 3138 // If the simplification isn't exact (that is, if it is conservative 3139 // in terms of dependence), set consistent to false. 3140 bool DependenceInfo::propagateLine(const SCEV *&Src, const SCEV *&Dst, 3141 Constraint &CurConstraint, 3142 bool &Consistent) { 3143 const Loop *CurLoop = CurConstraint.getAssociatedLoop(); 3144 const SCEV *A = CurConstraint.getA(); 3145 const SCEV *B = CurConstraint.getB(); 3146 const SCEV *C = CurConstraint.getC(); 3147 LLVM_DEBUG(dbgs() << "\t\tA = " << *A << ", B = " << *B << ", C = " << *C 3148 << "\n"); 3149 LLVM_DEBUG(dbgs() << "\t\tSrc = " << *Src << "\n"); 3150 LLVM_DEBUG(dbgs() << "\t\tDst = " << *Dst << "\n"); 3151 if (A->isZero()) { 3152 const SCEVConstant *Bconst = dyn_cast<SCEVConstant>(B); 3153 const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C); 3154 if (!Bconst || !Cconst) return false; 3155 APInt Beta = Bconst->getAPInt(); 3156 APInt Charlie = Cconst->getAPInt(); 3157 APInt CdivB = Charlie.sdiv(Beta); 3158 assert(Charlie.srem(Beta) == 0 && "C should be evenly divisible by B"); 3159 const SCEV *AP_K = findCoefficient(Dst, CurLoop); 3160 // Src = SE->getAddExpr(Src, SE->getMulExpr(AP_K, SE->getConstant(CdivB))); 3161 Src = SE->getMinusSCEV(Src, SE->getMulExpr(AP_K, SE->getConstant(CdivB))); 3162 Dst = zeroCoefficient(Dst, CurLoop); 3163 if (!findCoefficient(Src, CurLoop)->isZero()) 3164 Consistent = false; 3165 } 3166 else if (B->isZero()) { 3167 const SCEVConstant *Aconst = dyn_cast<SCEVConstant>(A); 3168 const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C); 3169 if (!Aconst || !Cconst) return false; 3170 APInt Alpha = Aconst->getAPInt(); 3171 APInt Charlie = Cconst->getAPInt(); 3172 APInt CdivA = Charlie.sdiv(Alpha); 3173 assert(Charlie.srem(Alpha) == 0 && "C should be evenly divisible by A"); 3174 const SCEV *A_K = findCoefficient(Src, CurLoop); 3175 Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, SE->getConstant(CdivA))); 3176 Src = zeroCoefficient(Src, CurLoop); 3177 if (!findCoefficient(Dst, CurLoop)->isZero()) 3178 Consistent = false; 3179 } 3180 else if (isKnownPredicate(CmpInst::ICMP_EQ, A, B)) { 3181 const SCEVConstant *Aconst = dyn_cast<SCEVConstant>(A); 3182 const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C); 3183 if (!Aconst || !Cconst) return false; 3184 APInt Alpha = Aconst->getAPInt(); 3185 APInt Charlie = Cconst->getAPInt(); 3186 APInt CdivA = Charlie.sdiv(Alpha); 3187 assert(Charlie.srem(Alpha) == 0 && "C should be evenly divisible by A"); 3188 const SCEV *A_K = findCoefficient(Src, CurLoop); 3189 Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, SE->getConstant(CdivA))); 3190 Src = zeroCoefficient(Src, CurLoop); 3191 Dst = addToCoefficient(Dst, CurLoop, A_K); 3192 if (!findCoefficient(Dst, CurLoop)->isZero()) 3193 Consistent = false; 3194 } 3195 else { 3196 // paper is incorrect here, or perhaps just misleading 3197 const SCEV *A_K = findCoefficient(Src, CurLoop); 3198 Src = SE->getMulExpr(Src, A); 3199 Dst = SE->getMulExpr(Dst, A); 3200 Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, C)); 3201 Src = zeroCoefficient(Src, CurLoop); 3202 Dst = addToCoefficient(Dst, CurLoop, SE->getMulExpr(A_K, B)); 3203 if (!findCoefficient(Dst, CurLoop)->isZero()) 3204 Consistent = false; 3205 } 3206 LLVM_DEBUG(dbgs() << "\t\tnew Src = " << *Src << "\n"); 3207 LLVM_DEBUG(dbgs() << "\t\tnew Dst = " << *Dst << "\n"); 3208 return true; 3209 } 3210 3211 3212 // Attempt to propagate a point 3213 // constraint into a subscript pair (Src and Dst). 3214 // Return true if some simplification occurs. 3215 bool DependenceInfo::propagatePoint(const SCEV *&Src, const SCEV *&Dst, 3216 Constraint &CurConstraint) { 3217 const Loop *CurLoop = CurConstraint.getAssociatedLoop(); 3218 const SCEV *A_K = findCoefficient(Src, CurLoop); 3219 const SCEV *AP_K = findCoefficient(Dst, CurLoop); 3220 const SCEV *XA_K = SE->getMulExpr(A_K, CurConstraint.getX()); 3221 const SCEV *YAP_K = SE->getMulExpr(AP_K, CurConstraint.getY()); 3222 LLVM_DEBUG(dbgs() << "\t\tSrc is " << *Src << "\n"); 3223 Src = SE->getAddExpr(Src, SE->getMinusSCEV(XA_K, YAP_K)); 3224 Src = zeroCoefficient(Src, CurLoop); 3225 LLVM_DEBUG(dbgs() << "\t\tnew Src is " << *Src << "\n"); 3226 LLVM_DEBUG(dbgs() << "\t\tDst is " << *Dst << "\n"); 3227 Dst = zeroCoefficient(Dst, CurLoop); 3228 LLVM_DEBUG(dbgs() << "\t\tnew Dst is " << *Dst << "\n"); 3229 return true; 3230 } 3231 3232 3233 // Update direction vector entry based on the current constraint. 3234 void DependenceInfo::updateDirection(Dependence::DVEntry &Level, 3235 const Constraint &CurConstraint) const { 3236 LLVM_DEBUG(dbgs() << "\tUpdate direction, constraint ="); 3237 LLVM_DEBUG(CurConstraint.dump(dbgs())); 3238 if (CurConstraint.isAny()) 3239 ; // use defaults 3240 else if (CurConstraint.isDistance()) { 3241 // this one is consistent, the others aren't 3242 Level.Scalar = false; 3243 Level.Distance = CurConstraint.getD(); 3244 unsigned NewDirection = Dependence::DVEntry::NONE; 3245 if (!SE->isKnownNonZero(Level.Distance)) // if may be zero 3246 NewDirection = Dependence::DVEntry::EQ; 3247 if (!SE->isKnownNonPositive(Level.Distance)) // if may be positive 3248 NewDirection |= Dependence::DVEntry::LT; 3249 if (!SE->isKnownNonNegative(Level.Distance)) // if may be negative 3250 NewDirection |= Dependence::DVEntry::GT; 3251 Level.Direction &= NewDirection; 3252 } 3253 else if (CurConstraint.isLine()) { 3254 Level.Scalar = false; 3255 Level.Distance = nullptr; 3256 // direction should be accurate 3257 } 3258 else if (CurConstraint.isPoint()) { 3259 Level.Scalar = false; 3260 Level.Distance = nullptr; 3261 unsigned NewDirection = Dependence::DVEntry::NONE; 3262 if (!isKnownPredicate(CmpInst::ICMP_NE, 3263 CurConstraint.getY(), 3264 CurConstraint.getX())) 3265 // if X may be = Y 3266 NewDirection |= Dependence::DVEntry::EQ; 3267 if (!isKnownPredicate(CmpInst::ICMP_SLE, 3268 CurConstraint.getY(), 3269 CurConstraint.getX())) 3270 // if Y may be > X 3271 NewDirection |= Dependence::DVEntry::LT; 3272 if (!isKnownPredicate(CmpInst::ICMP_SGE, 3273 CurConstraint.getY(), 3274 CurConstraint.getX())) 3275 // if Y may be < X 3276 NewDirection |= Dependence::DVEntry::GT; 3277 Level.Direction &= NewDirection; 3278 } 3279 else 3280 llvm_unreachable("constraint has unexpected kind"); 3281 } 3282 3283 /// Check if we can delinearize the subscripts. If the SCEVs representing the 3284 /// source and destination array references are recurrences on a nested loop, 3285 /// this function flattens the nested recurrences into separate recurrences 3286 /// for each loop level. 3287 bool DependenceInfo::tryDelinearize(Instruction *Src, Instruction *Dst, 3288 SmallVectorImpl<Subscript> &Pair) { 3289 assert(isLoadOrStore(Src) && "instruction is not load or store"); 3290 assert(isLoadOrStore(Dst) && "instruction is not load or store"); 3291 Value *SrcPtr = getLoadStorePointerOperand(Src); 3292 Value *DstPtr = getLoadStorePointerOperand(Dst); 3293 Loop *SrcLoop = LI->getLoopFor(Src->getParent()); 3294 Loop *DstLoop = LI->getLoopFor(Dst->getParent()); 3295 const SCEV *SrcAccessFn = SE->getSCEVAtScope(SrcPtr, SrcLoop); 3296 const SCEV *DstAccessFn = SE->getSCEVAtScope(DstPtr, DstLoop); 3297 const SCEVUnknown *SrcBase = 3298 dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn)); 3299 const SCEVUnknown *DstBase = 3300 dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn)); 3301 3302 if (!SrcBase || !DstBase || SrcBase != DstBase) 3303 return false; 3304 3305 SmallVector<const SCEV *, 4> SrcSubscripts, DstSubscripts; 3306 3307 if (!tryDelinearizeFixedSize(Src, Dst, SrcAccessFn, DstAccessFn, 3308 SrcSubscripts, DstSubscripts) && 3309 !tryDelinearizeParametricSize(Src, Dst, SrcAccessFn, DstAccessFn, 3310 SrcSubscripts, DstSubscripts)) 3311 return false; 3312 3313 int Size = SrcSubscripts.size(); 3314 LLVM_DEBUG({ 3315 dbgs() << "\nSrcSubscripts: "; 3316 for (int I = 0; I < Size; I++) 3317 dbgs() << *SrcSubscripts[I]; 3318 dbgs() << "\nDstSubscripts: "; 3319 for (int I = 0; I < Size; I++) 3320 dbgs() << *DstSubscripts[I]; 3321 }); 3322 3323 // The delinearization transforms a single-subscript MIV dependence test into 3324 // a multi-subscript SIV dependence test that is easier to compute. So we 3325 // resize Pair to contain as many pairs of subscripts as the delinearization 3326 // has found, and then initialize the pairs following the delinearization. 3327 Pair.resize(Size); 3328 for (int I = 0; I < Size; ++I) { 3329 Pair[I].Src = SrcSubscripts[I]; 3330 Pair[I].Dst = DstSubscripts[I]; 3331 unifySubscriptType(&Pair[I]); 3332 } 3333 3334 return true; 3335 } 3336 3337 /// Try to delinearize \p SrcAccessFn and \p DstAccessFn if the underlying 3338 /// arrays accessed are fixed-size arrays. Return true if delinearization was 3339 /// successful. 3340 bool DependenceInfo::tryDelinearizeFixedSize( 3341 Instruction *Src, Instruction *Dst, const SCEV *SrcAccessFn, 3342 const SCEV *DstAccessFn, SmallVectorImpl<const SCEV *> &SrcSubscripts, 3343 SmallVectorImpl<const SCEV *> &DstSubscripts) { 3344 LLVM_DEBUG({ 3345 const SCEVUnknown *SrcBase = 3346 dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn)); 3347 const SCEVUnknown *DstBase = 3348 dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn)); 3349 assert(SrcBase && DstBase && SrcBase == DstBase && 3350 "expected src and dst scev unknowns to be equal"); 3351 }); 3352 3353 SmallVector<int, 4> SrcSizes; 3354 SmallVector<int, 4> DstSizes; 3355 if (!tryDelinearizeFixedSizeImpl(SE, Src, SrcAccessFn, SrcSubscripts, 3356 SrcSizes) || 3357 !tryDelinearizeFixedSizeImpl(SE, Dst, DstAccessFn, DstSubscripts, 3358 DstSizes)) 3359 return false; 3360 3361 // Check that the two size arrays are non-empty and equal in length and 3362 // value. 3363 if (SrcSizes.size() != DstSizes.size() || 3364 !std::equal(SrcSizes.begin(), SrcSizes.end(), DstSizes.begin())) { 3365 SrcSubscripts.clear(); 3366 DstSubscripts.clear(); 3367 return false; 3368 } 3369 3370 assert(SrcSubscripts.size() == DstSubscripts.size() && 3371 "Expected equal number of entries in the list of SrcSubscripts and " 3372 "DstSubscripts."); 3373 3374 Value *SrcPtr = getLoadStorePointerOperand(Src); 3375 Value *DstPtr = getLoadStorePointerOperand(Dst); 3376 3377 // In general we cannot safely assume that the subscripts recovered from GEPs 3378 // are in the range of values defined for their corresponding array 3379 // dimensions. For example some C language usage/interpretation make it 3380 // impossible to verify this at compile-time. As such we can only delinearize 3381 // iff the subscripts are positive and are less than the range of the 3382 // dimension. 3383 if (!DisableDelinearizationChecks) { 3384 auto AllIndiciesInRange = [&](SmallVector<int, 4> &DimensionSizes, 3385 SmallVectorImpl<const SCEV *> &Subscripts, 3386 Value *Ptr) { 3387 size_t SSize = Subscripts.size(); 3388 for (size_t I = 1; I < SSize; ++I) { 3389 const SCEV *S = Subscripts[I]; 3390 if (!isKnownNonNegative(S, Ptr)) 3391 return false; 3392 if (auto *SType = dyn_cast<IntegerType>(S->getType())) { 3393 const SCEV *Range = SE->getConstant( 3394 ConstantInt::get(SType, DimensionSizes[I - 1], false)); 3395 if (!isKnownLessThan(S, Range)) 3396 return false; 3397 } 3398 } 3399 return true; 3400 }; 3401 3402 if (!AllIndiciesInRange(SrcSizes, SrcSubscripts, SrcPtr) || 3403 !AllIndiciesInRange(DstSizes, DstSubscripts, DstPtr)) { 3404 SrcSubscripts.clear(); 3405 DstSubscripts.clear(); 3406 return false; 3407 } 3408 } 3409 LLVM_DEBUG({ 3410 dbgs() << "Delinearized subscripts of fixed-size array\n" 3411 << "SrcGEP:" << *SrcPtr << "\n" 3412 << "DstGEP:" << *DstPtr << "\n"; 3413 }); 3414 return true; 3415 } 3416 3417 bool DependenceInfo::tryDelinearizeParametricSize( 3418 Instruction *Src, Instruction *Dst, const SCEV *SrcAccessFn, 3419 const SCEV *DstAccessFn, SmallVectorImpl<const SCEV *> &SrcSubscripts, 3420 SmallVectorImpl<const SCEV *> &DstSubscripts) { 3421 3422 Value *SrcPtr = getLoadStorePointerOperand(Src); 3423 Value *DstPtr = getLoadStorePointerOperand(Dst); 3424 const SCEVUnknown *SrcBase = 3425 dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn)); 3426 const SCEVUnknown *DstBase = 3427 dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn)); 3428 assert(SrcBase && DstBase && SrcBase == DstBase && 3429 "expected src and dst scev unknowns to be equal"); 3430 3431 const SCEV *ElementSize = SE->getElementSize(Src); 3432 if (ElementSize != SE->getElementSize(Dst)) 3433 return false; 3434 3435 const SCEV *SrcSCEV = SE->getMinusSCEV(SrcAccessFn, SrcBase); 3436 const SCEV *DstSCEV = SE->getMinusSCEV(DstAccessFn, DstBase); 3437 3438 const SCEVAddRecExpr *SrcAR = dyn_cast<SCEVAddRecExpr>(SrcSCEV); 3439 const SCEVAddRecExpr *DstAR = dyn_cast<SCEVAddRecExpr>(DstSCEV); 3440 if (!SrcAR || !DstAR || !SrcAR->isAffine() || !DstAR->isAffine()) 3441 return false; 3442 3443 // First step: collect parametric terms in both array references. 3444 SmallVector<const SCEV *, 4> Terms; 3445 collectParametricTerms(*SE, SrcAR, Terms); 3446 collectParametricTerms(*SE, DstAR, Terms); 3447 3448 // Second step: find subscript sizes. 3449 SmallVector<const SCEV *, 4> Sizes; 3450 findArrayDimensions(*SE, Terms, Sizes, ElementSize); 3451 3452 // Third step: compute the access functions for each subscript. 3453 computeAccessFunctions(*SE, SrcAR, SrcSubscripts, Sizes); 3454 computeAccessFunctions(*SE, DstAR, DstSubscripts, Sizes); 3455 3456 // Fail when there is only a subscript: that's a linearized access function. 3457 if (SrcSubscripts.size() < 2 || DstSubscripts.size() < 2 || 3458 SrcSubscripts.size() != DstSubscripts.size()) 3459 return false; 3460 3461 size_t Size = SrcSubscripts.size(); 3462 3463 // Statically check that the array bounds are in-range. The first subscript we 3464 // don't have a size for and it cannot overflow into another subscript, so is 3465 // always safe. The others need to be 0 <= subscript[i] < bound, for both src 3466 // and dst. 3467 // FIXME: It may be better to record these sizes and add them as constraints 3468 // to the dependency checks. 3469 if (!DisableDelinearizationChecks) 3470 for (size_t I = 1; I < Size; ++I) { 3471 if (!isKnownNonNegative(SrcSubscripts[I], SrcPtr)) 3472 return false; 3473 3474 if (!isKnownLessThan(SrcSubscripts[I], Sizes[I - 1])) 3475 return false; 3476 3477 if (!isKnownNonNegative(DstSubscripts[I], DstPtr)) 3478 return false; 3479 3480 if (!isKnownLessThan(DstSubscripts[I], Sizes[I - 1])) 3481 return false; 3482 } 3483 3484 return true; 3485 } 3486 3487 //===----------------------------------------------------------------------===// 3488 3489 #ifndef NDEBUG 3490 // For debugging purposes, dump a small bit vector to dbgs(). 3491 static void dumpSmallBitVector(SmallBitVector &BV) { 3492 dbgs() << "{"; 3493 for (unsigned VI : BV.set_bits()) { 3494 dbgs() << VI; 3495 if (BV.find_next(VI) >= 0) 3496 dbgs() << ' '; 3497 } 3498 dbgs() << "}\n"; 3499 } 3500 #endif 3501 3502 bool DependenceInfo::invalidate(Function &F, const PreservedAnalyses &PA, 3503 FunctionAnalysisManager::Invalidator &Inv) { 3504 // Check if the analysis itself has been invalidated. 3505 auto PAC = PA.getChecker<DependenceAnalysis>(); 3506 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>()) 3507 return true; 3508 3509 // Check transitive dependencies. 3510 return Inv.invalidate<AAManager>(F, PA) || 3511 Inv.invalidate<ScalarEvolutionAnalysis>(F, PA) || 3512 Inv.invalidate<LoopAnalysis>(F, PA); 3513 } 3514 3515 // depends - 3516 // Returns NULL if there is no dependence. 3517 // Otherwise, return a Dependence with as many details as possible. 3518 // Corresponds to Section 3.1 in the paper 3519 // 3520 // Practical Dependence Testing 3521 // Goff, Kennedy, Tseng 3522 // PLDI 1991 3523 // 3524 // Care is required to keep the routine below, getSplitIteration(), 3525 // up to date with respect to this routine. 3526 std::unique_ptr<Dependence> 3527 DependenceInfo::depends(Instruction *Src, Instruction *Dst, 3528 bool PossiblyLoopIndependent) { 3529 if (Src == Dst) 3530 PossiblyLoopIndependent = false; 3531 3532 if (!(Src->mayReadOrWriteMemory() && Dst->mayReadOrWriteMemory())) 3533 // if both instructions don't reference memory, there's no dependence 3534 return nullptr; 3535 3536 if (!isLoadOrStore(Src) || !isLoadOrStore(Dst)) { 3537 // can only analyze simple loads and stores, i.e., no calls, invokes, etc. 3538 LLVM_DEBUG(dbgs() << "can only handle simple loads and stores\n"); 3539 return std::make_unique<Dependence>(Src, Dst); 3540 } 3541 3542 assert(isLoadOrStore(Src) && "instruction is not load or store"); 3543 assert(isLoadOrStore(Dst) && "instruction is not load or store"); 3544 Value *SrcPtr = getLoadStorePointerOperand(Src); 3545 Value *DstPtr = getLoadStorePointerOperand(Dst); 3546 3547 switch (underlyingObjectsAlias(AA, F->getParent()->getDataLayout(), 3548 MemoryLocation::get(Dst), 3549 MemoryLocation::get(Src))) { 3550 case AliasResult::MayAlias: 3551 case AliasResult::PartialAlias: 3552 // cannot analyse objects if we don't understand their aliasing. 3553 LLVM_DEBUG(dbgs() << "can't analyze may or partial alias\n"); 3554 return std::make_unique<Dependence>(Src, Dst); 3555 case AliasResult::NoAlias: 3556 // If the objects noalias, they are distinct, accesses are independent. 3557 LLVM_DEBUG(dbgs() << "no alias\n"); 3558 return nullptr; 3559 case AliasResult::MustAlias: 3560 break; // The underlying objects alias; test accesses for dependence. 3561 } 3562 3563 // establish loop nesting levels 3564 establishNestingLevels(Src, Dst); 3565 LLVM_DEBUG(dbgs() << " common nesting levels = " << CommonLevels << "\n"); 3566 LLVM_DEBUG(dbgs() << " maximum nesting levels = " << MaxLevels << "\n"); 3567 3568 FullDependence Result(Src, Dst, PossiblyLoopIndependent, CommonLevels); 3569 ++TotalArrayPairs; 3570 3571 unsigned Pairs = 1; 3572 SmallVector<Subscript, 2> Pair(Pairs); 3573 const SCEV *SrcSCEV = SE->getSCEV(SrcPtr); 3574 const SCEV *DstSCEV = SE->getSCEV(DstPtr); 3575 LLVM_DEBUG(dbgs() << " SrcSCEV = " << *SrcSCEV << "\n"); 3576 LLVM_DEBUG(dbgs() << " DstSCEV = " << *DstSCEV << "\n"); 3577 if (SE->getPointerBase(SrcSCEV) != SE->getPointerBase(DstSCEV)) { 3578 // If two pointers have different bases, trying to analyze indexes won't 3579 // work; we can't compare them to each other. This can happen, for example, 3580 // if one is produced by an LCSSA PHI node. 3581 // 3582 // We check this upfront so we don't crash in cases where getMinusSCEV() 3583 // returns a SCEVCouldNotCompute. 3584 LLVM_DEBUG(dbgs() << "can't analyze SCEV with different pointer base\n"); 3585 return std::make_unique<Dependence>(Src, Dst); 3586 } 3587 Pair[0].Src = SrcSCEV; 3588 Pair[0].Dst = DstSCEV; 3589 3590 if (Delinearize) { 3591 if (tryDelinearize(Src, Dst, Pair)) { 3592 LLVM_DEBUG(dbgs() << " delinearized\n"); 3593 Pairs = Pair.size(); 3594 } 3595 } 3596 3597 for (unsigned P = 0; P < Pairs; ++P) { 3598 Pair[P].Loops.resize(MaxLevels + 1); 3599 Pair[P].GroupLoops.resize(MaxLevels + 1); 3600 Pair[P].Group.resize(Pairs); 3601 removeMatchingExtensions(&Pair[P]); 3602 Pair[P].Classification = 3603 classifyPair(Pair[P].Src, LI->getLoopFor(Src->getParent()), 3604 Pair[P].Dst, LI->getLoopFor(Dst->getParent()), 3605 Pair[P].Loops); 3606 Pair[P].GroupLoops = Pair[P].Loops; 3607 Pair[P].Group.set(P); 3608 LLVM_DEBUG(dbgs() << " subscript " << P << "\n"); 3609 LLVM_DEBUG(dbgs() << "\tsrc = " << *Pair[P].Src << "\n"); 3610 LLVM_DEBUG(dbgs() << "\tdst = " << *Pair[P].Dst << "\n"); 3611 LLVM_DEBUG(dbgs() << "\tclass = " << Pair[P].Classification << "\n"); 3612 LLVM_DEBUG(dbgs() << "\tloops = "); 3613 LLVM_DEBUG(dumpSmallBitVector(Pair[P].Loops)); 3614 } 3615 3616 SmallBitVector Separable(Pairs); 3617 SmallBitVector Coupled(Pairs); 3618 3619 // Partition subscripts into separable and minimally-coupled groups 3620 // Algorithm in paper is algorithmically better; 3621 // this may be faster in practice. Check someday. 3622 // 3623 // Here's an example of how it works. Consider this code: 3624 // 3625 // for (i = ...) { 3626 // for (j = ...) { 3627 // for (k = ...) { 3628 // for (l = ...) { 3629 // for (m = ...) { 3630 // A[i][j][k][m] = ...; 3631 // ... = A[0][j][l][i + j]; 3632 // } 3633 // } 3634 // } 3635 // } 3636 // } 3637 // 3638 // There are 4 subscripts here: 3639 // 0 [i] and [0] 3640 // 1 [j] and [j] 3641 // 2 [k] and [l] 3642 // 3 [m] and [i + j] 3643 // 3644 // We've already classified each subscript pair as ZIV, SIV, etc., 3645 // and collected all the loops mentioned by pair P in Pair[P].Loops. 3646 // In addition, we've initialized Pair[P].GroupLoops to Pair[P].Loops 3647 // and set Pair[P].Group = {P}. 3648 // 3649 // Src Dst Classification Loops GroupLoops Group 3650 // 0 [i] [0] SIV {1} {1} {0} 3651 // 1 [j] [j] SIV {2} {2} {1} 3652 // 2 [k] [l] RDIV {3,4} {3,4} {2} 3653 // 3 [m] [i + j] MIV {1,2,5} {1,2,5} {3} 3654 // 3655 // For each subscript SI 0 .. 3, we consider each remaining subscript, SJ. 3656 // So, 0 is compared against 1, 2, and 3; 1 is compared against 2 and 3, etc. 3657 // 3658 // We begin by comparing 0 and 1. The intersection of the GroupLoops is empty. 3659 // Next, 0 and 2. Again, the intersection of their GroupLoops is empty. 3660 // Next 0 and 3. The intersection of their GroupLoop = {1}, not empty, 3661 // so Pair[3].Group = {0,3} and Done = false (that is, 0 will not be added 3662 // to either Separable or Coupled). 3663 // 3664 // Next, we consider 1 and 2. The intersection of the GroupLoops is empty. 3665 // Next, 1 and 3. The intersection of their GroupLoops = {2}, not empty, 3666 // so Pair[3].Group = {0, 1, 3} and Done = false. 3667 // 3668 // Next, we compare 2 against 3. The intersection of the GroupLoops is empty. 3669 // Since Done remains true, we add 2 to the set of Separable pairs. 3670 // 3671 // Finally, we consider 3. There's nothing to compare it with, 3672 // so Done remains true and we add it to the Coupled set. 3673 // Pair[3].Group = {0, 1, 3} and GroupLoops = {1, 2, 5}. 3674 // 3675 // In the end, we've got 1 separable subscript and 1 coupled group. 3676 for (unsigned SI = 0; SI < Pairs; ++SI) { 3677 if (Pair[SI].Classification == Subscript::NonLinear) { 3678 // ignore these, but collect loops for later 3679 ++NonlinearSubscriptPairs; 3680 collectCommonLoops(Pair[SI].Src, 3681 LI->getLoopFor(Src->getParent()), 3682 Pair[SI].Loops); 3683 collectCommonLoops(Pair[SI].Dst, 3684 LI->getLoopFor(Dst->getParent()), 3685 Pair[SI].Loops); 3686 Result.Consistent = false; 3687 } else if (Pair[SI].Classification == Subscript::ZIV) { 3688 // always separable 3689 Separable.set(SI); 3690 } 3691 else { 3692 // SIV, RDIV, or MIV, so check for coupled group 3693 bool Done = true; 3694 for (unsigned SJ = SI + 1; SJ < Pairs; ++SJ) { 3695 SmallBitVector Intersection = Pair[SI].GroupLoops; 3696 Intersection &= Pair[SJ].GroupLoops; 3697 if (Intersection.any()) { 3698 // accumulate set of all the loops in group 3699 Pair[SJ].GroupLoops |= Pair[SI].GroupLoops; 3700 // accumulate set of all subscripts in group 3701 Pair[SJ].Group |= Pair[SI].Group; 3702 Done = false; 3703 } 3704 } 3705 if (Done) { 3706 if (Pair[SI].Group.count() == 1) { 3707 Separable.set(SI); 3708 ++SeparableSubscriptPairs; 3709 } 3710 else { 3711 Coupled.set(SI); 3712 ++CoupledSubscriptPairs; 3713 } 3714 } 3715 } 3716 } 3717 3718 LLVM_DEBUG(dbgs() << " Separable = "); 3719 LLVM_DEBUG(dumpSmallBitVector(Separable)); 3720 LLVM_DEBUG(dbgs() << " Coupled = "); 3721 LLVM_DEBUG(dumpSmallBitVector(Coupled)); 3722 3723 Constraint NewConstraint; 3724 NewConstraint.setAny(SE); 3725 3726 // test separable subscripts 3727 for (unsigned SI : Separable.set_bits()) { 3728 LLVM_DEBUG(dbgs() << "testing subscript " << SI); 3729 switch (Pair[SI].Classification) { 3730 case Subscript::ZIV: 3731 LLVM_DEBUG(dbgs() << ", ZIV\n"); 3732 if (testZIV(Pair[SI].Src, Pair[SI].Dst, Result)) 3733 return nullptr; 3734 break; 3735 case Subscript::SIV: { 3736 LLVM_DEBUG(dbgs() << ", SIV\n"); 3737 unsigned Level; 3738 const SCEV *SplitIter = nullptr; 3739 if (testSIV(Pair[SI].Src, Pair[SI].Dst, Level, Result, NewConstraint, 3740 SplitIter)) 3741 return nullptr; 3742 break; 3743 } 3744 case Subscript::RDIV: 3745 LLVM_DEBUG(dbgs() << ", RDIV\n"); 3746 if (testRDIV(Pair[SI].Src, Pair[SI].Dst, Result)) 3747 return nullptr; 3748 break; 3749 case Subscript::MIV: 3750 LLVM_DEBUG(dbgs() << ", MIV\n"); 3751 if (testMIV(Pair[SI].Src, Pair[SI].Dst, Pair[SI].Loops, Result)) 3752 return nullptr; 3753 break; 3754 default: 3755 llvm_unreachable("subscript has unexpected classification"); 3756 } 3757 } 3758 3759 if (Coupled.count()) { 3760 // test coupled subscript groups 3761 LLVM_DEBUG(dbgs() << "starting on coupled subscripts\n"); 3762 LLVM_DEBUG(dbgs() << "MaxLevels + 1 = " << MaxLevels + 1 << "\n"); 3763 SmallVector<Constraint, 4> Constraints(MaxLevels + 1); 3764 for (unsigned II = 0; II <= MaxLevels; ++II) 3765 Constraints[II].setAny(SE); 3766 for (unsigned SI : Coupled.set_bits()) { 3767 LLVM_DEBUG(dbgs() << "testing subscript group " << SI << " { "); 3768 SmallBitVector Group(Pair[SI].Group); 3769 SmallBitVector Sivs(Pairs); 3770 SmallBitVector Mivs(Pairs); 3771 SmallBitVector ConstrainedLevels(MaxLevels + 1); 3772 SmallVector<Subscript *, 4> PairsInGroup; 3773 for (unsigned SJ : Group.set_bits()) { 3774 LLVM_DEBUG(dbgs() << SJ << " "); 3775 if (Pair[SJ].Classification == Subscript::SIV) 3776 Sivs.set(SJ); 3777 else 3778 Mivs.set(SJ); 3779 PairsInGroup.push_back(&Pair[SJ]); 3780 } 3781 unifySubscriptType(PairsInGroup); 3782 LLVM_DEBUG(dbgs() << "}\n"); 3783 while (Sivs.any()) { 3784 bool Changed = false; 3785 for (unsigned SJ : Sivs.set_bits()) { 3786 LLVM_DEBUG(dbgs() << "testing subscript " << SJ << ", SIV\n"); 3787 // SJ is an SIV subscript that's part of the current coupled group 3788 unsigned Level; 3789 const SCEV *SplitIter = nullptr; 3790 LLVM_DEBUG(dbgs() << "SIV\n"); 3791 if (testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level, Result, NewConstraint, 3792 SplitIter)) 3793 return nullptr; 3794 ConstrainedLevels.set(Level); 3795 if (intersectConstraints(&Constraints[Level], &NewConstraint)) { 3796 if (Constraints[Level].isEmpty()) { 3797 ++DeltaIndependence; 3798 return nullptr; 3799 } 3800 Changed = true; 3801 } 3802 Sivs.reset(SJ); 3803 } 3804 if (Changed) { 3805 // propagate, possibly creating new SIVs and ZIVs 3806 LLVM_DEBUG(dbgs() << " propagating\n"); 3807 LLVM_DEBUG(dbgs() << "\tMivs = "); 3808 LLVM_DEBUG(dumpSmallBitVector(Mivs)); 3809 for (unsigned SJ : Mivs.set_bits()) { 3810 // SJ is an MIV subscript that's part of the current coupled group 3811 LLVM_DEBUG(dbgs() << "\tSJ = " << SJ << "\n"); 3812 if (propagate(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops, 3813 Constraints, Result.Consistent)) { 3814 LLVM_DEBUG(dbgs() << "\t Changed\n"); 3815 ++DeltaPropagations; 3816 Pair[SJ].Classification = 3817 classifyPair(Pair[SJ].Src, LI->getLoopFor(Src->getParent()), 3818 Pair[SJ].Dst, LI->getLoopFor(Dst->getParent()), 3819 Pair[SJ].Loops); 3820 switch (Pair[SJ].Classification) { 3821 case Subscript::ZIV: 3822 LLVM_DEBUG(dbgs() << "ZIV\n"); 3823 if (testZIV(Pair[SJ].Src, Pair[SJ].Dst, Result)) 3824 return nullptr; 3825 Mivs.reset(SJ); 3826 break; 3827 case Subscript::SIV: 3828 Sivs.set(SJ); 3829 Mivs.reset(SJ); 3830 break; 3831 case Subscript::RDIV: 3832 case Subscript::MIV: 3833 break; 3834 default: 3835 llvm_unreachable("bad subscript classification"); 3836 } 3837 } 3838 } 3839 } 3840 } 3841 3842 // test & propagate remaining RDIVs 3843 for (unsigned SJ : Mivs.set_bits()) { 3844 if (Pair[SJ].Classification == Subscript::RDIV) { 3845 LLVM_DEBUG(dbgs() << "RDIV test\n"); 3846 if (testRDIV(Pair[SJ].Src, Pair[SJ].Dst, Result)) 3847 return nullptr; 3848 // I don't yet understand how to propagate RDIV results 3849 Mivs.reset(SJ); 3850 } 3851 } 3852 3853 // test remaining MIVs 3854 // This code is temporary. 3855 // Better to somehow test all remaining subscripts simultaneously. 3856 for (unsigned SJ : Mivs.set_bits()) { 3857 if (Pair[SJ].Classification == Subscript::MIV) { 3858 LLVM_DEBUG(dbgs() << "MIV test\n"); 3859 if (testMIV(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops, Result)) 3860 return nullptr; 3861 } 3862 else 3863 llvm_unreachable("expected only MIV subscripts at this point"); 3864 } 3865 3866 // update Result.DV from constraint vector 3867 LLVM_DEBUG(dbgs() << " updating\n"); 3868 for (unsigned SJ : ConstrainedLevels.set_bits()) { 3869 if (SJ > CommonLevels) 3870 break; 3871 updateDirection(Result.DV[SJ - 1], Constraints[SJ]); 3872 if (Result.DV[SJ - 1].Direction == Dependence::DVEntry::NONE) 3873 return nullptr; 3874 } 3875 } 3876 } 3877 3878 // Make sure the Scalar flags are set correctly. 3879 SmallBitVector CompleteLoops(MaxLevels + 1); 3880 for (unsigned SI = 0; SI < Pairs; ++SI) 3881 CompleteLoops |= Pair[SI].Loops; 3882 for (unsigned II = 1; II <= CommonLevels; ++II) 3883 if (CompleteLoops[II]) 3884 Result.DV[II - 1].Scalar = false; 3885 3886 if (PossiblyLoopIndependent) { 3887 // Make sure the LoopIndependent flag is set correctly. 3888 // All directions must include equal, otherwise no 3889 // loop-independent dependence is possible. 3890 for (unsigned II = 1; II <= CommonLevels; ++II) { 3891 if (!(Result.getDirection(II) & Dependence::DVEntry::EQ)) { 3892 Result.LoopIndependent = false; 3893 break; 3894 } 3895 } 3896 } 3897 else { 3898 // On the other hand, if all directions are equal and there's no 3899 // loop-independent dependence possible, then no dependence exists. 3900 bool AllEqual = true; 3901 for (unsigned II = 1; II <= CommonLevels; ++II) { 3902 if (Result.getDirection(II) != Dependence::DVEntry::EQ) { 3903 AllEqual = false; 3904 break; 3905 } 3906 } 3907 if (AllEqual) 3908 return nullptr; 3909 } 3910 3911 return std::make_unique<FullDependence>(std::move(Result)); 3912 } 3913 3914 //===----------------------------------------------------------------------===// 3915 // getSplitIteration - 3916 // Rather than spend rarely-used space recording the splitting iteration 3917 // during the Weak-Crossing SIV test, we re-compute it on demand. 3918 // The re-computation is basically a repeat of the entire dependence test, 3919 // though simplified since we know that the dependence exists. 3920 // It's tedious, since we must go through all propagations, etc. 3921 // 3922 // Care is required to keep this code up to date with respect to the routine 3923 // above, depends(). 3924 // 3925 // Generally, the dependence analyzer will be used to build 3926 // a dependence graph for a function (basically a map from instructions 3927 // to dependences). Looking for cycles in the graph shows us loops 3928 // that cannot be trivially vectorized/parallelized. 3929 // 3930 // We can try to improve the situation by examining all the dependences 3931 // that make up the cycle, looking for ones we can break. 3932 // Sometimes, peeling the first or last iteration of a loop will break 3933 // dependences, and we've got flags for those possibilities. 3934 // Sometimes, splitting a loop at some other iteration will do the trick, 3935 // and we've got a flag for that case. Rather than waste the space to 3936 // record the exact iteration (since we rarely know), we provide 3937 // a method that calculates the iteration. It's a drag that it must work 3938 // from scratch, but wonderful in that it's possible. 3939 // 3940 // Here's an example: 3941 // 3942 // for (i = 0; i < 10; i++) 3943 // A[i] = ... 3944 // ... = A[11 - i] 3945 // 3946 // There's a loop-carried flow dependence from the store to the load, 3947 // found by the weak-crossing SIV test. The dependence will have a flag, 3948 // indicating that the dependence can be broken by splitting the loop. 3949 // Calling getSplitIteration will return 5. 3950 // Splitting the loop breaks the dependence, like so: 3951 // 3952 // for (i = 0; i <= 5; i++) 3953 // A[i] = ... 3954 // ... = A[11 - i] 3955 // for (i = 6; i < 10; i++) 3956 // A[i] = ... 3957 // ... = A[11 - i] 3958 // 3959 // breaks the dependence and allows us to vectorize/parallelize 3960 // both loops. 3961 const SCEV *DependenceInfo::getSplitIteration(const Dependence &Dep, 3962 unsigned SplitLevel) { 3963 assert(Dep.isSplitable(SplitLevel) && 3964 "Dep should be splitable at SplitLevel"); 3965 Instruction *Src = Dep.getSrc(); 3966 Instruction *Dst = Dep.getDst(); 3967 assert(Src->mayReadFromMemory() || Src->mayWriteToMemory()); 3968 assert(Dst->mayReadFromMemory() || Dst->mayWriteToMemory()); 3969 assert(isLoadOrStore(Src)); 3970 assert(isLoadOrStore(Dst)); 3971 Value *SrcPtr = getLoadStorePointerOperand(Src); 3972 Value *DstPtr = getLoadStorePointerOperand(Dst); 3973 assert(underlyingObjectsAlias( 3974 AA, F->getParent()->getDataLayout(), MemoryLocation::get(Dst), 3975 MemoryLocation::get(Src)) == AliasResult::MustAlias); 3976 3977 // establish loop nesting levels 3978 establishNestingLevels(Src, Dst); 3979 3980 FullDependence Result(Src, Dst, false, CommonLevels); 3981 3982 unsigned Pairs = 1; 3983 SmallVector<Subscript, 2> Pair(Pairs); 3984 const SCEV *SrcSCEV = SE->getSCEV(SrcPtr); 3985 const SCEV *DstSCEV = SE->getSCEV(DstPtr); 3986 Pair[0].Src = SrcSCEV; 3987 Pair[0].Dst = DstSCEV; 3988 3989 if (Delinearize) { 3990 if (tryDelinearize(Src, Dst, Pair)) { 3991 LLVM_DEBUG(dbgs() << " delinearized\n"); 3992 Pairs = Pair.size(); 3993 } 3994 } 3995 3996 for (unsigned P = 0; P < Pairs; ++P) { 3997 Pair[P].Loops.resize(MaxLevels + 1); 3998 Pair[P].GroupLoops.resize(MaxLevels + 1); 3999 Pair[P].Group.resize(Pairs); 4000 removeMatchingExtensions(&Pair[P]); 4001 Pair[P].Classification = 4002 classifyPair(Pair[P].Src, LI->getLoopFor(Src->getParent()), 4003 Pair[P].Dst, LI->getLoopFor(Dst->getParent()), 4004 Pair[P].Loops); 4005 Pair[P].GroupLoops = Pair[P].Loops; 4006 Pair[P].Group.set(P); 4007 } 4008 4009 SmallBitVector Separable(Pairs); 4010 SmallBitVector Coupled(Pairs); 4011 4012 // partition subscripts into separable and minimally-coupled groups 4013 for (unsigned SI = 0; SI < Pairs; ++SI) { 4014 if (Pair[SI].Classification == Subscript::NonLinear) { 4015 // ignore these, but collect loops for later 4016 collectCommonLoops(Pair[SI].Src, 4017 LI->getLoopFor(Src->getParent()), 4018 Pair[SI].Loops); 4019 collectCommonLoops(Pair[SI].Dst, 4020 LI->getLoopFor(Dst->getParent()), 4021 Pair[SI].Loops); 4022 Result.Consistent = false; 4023 } 4024 else if (Pair[SI].Classification == Subscript::ZIV) 4025 Separable.set(SI); 4026 else { 4027 // SIV, RDIV, or MIV, so check for coupled group 4028 bool Done = true; 4029 for (unsigned SJ = SI + 1; SJ < Pairs; ++SJ) { 4030 SmallBitVector Intersection = Pair[SI].GroupLoops; 4031 Intersection &= Pair[SJ].GroupLoops; 4032 if (Intersection.any()) { 4033 // accumulate set of all the loops in group 4034 Pair[SJ].GroupLoops |= Pair[SI].GroupLoops; 4035 // accumulate set of all subscripts in group 4036 Pair[SJ].Group |= Pair[SI].Group; 4037 Done = false; 4038 } 4039 } 4040 if (Done) { 4041 if (Pair[SI].Group.count() == 1) 4042 Separable.set(SI); 4043 else 4044 Coupled.set(SI); 4045 } 4046 } 4047 } 4048 4049 Constraint NewConstraint; 4050 NewConstraint.setAny(SE); 4051 4052 // test separable subscripts 4053 for (unsigned SI : Separable.set_bits()) { 4054 switch (Pair[SI].Classification) { 4055 case Subscript::SIV: { 4056 unsigned Level; 4057 const SCEV *SplitIter = nullptr; 4058 (void) testSIV(Pair[SI].Src, Pair[SI].Dst, Level, 4059 Result, NewConstraint, SplitIter); 4060 if (Level == SplitLevel) { 4061 assert(SplitIter != nullptr); 4062 return SplitIter; 4063 } 4064 break; 4065 } 4066 case Subscript::ZIV: 4067 case Subscript::RDIV: 4068 case Subscript::MIV: 4069 break; 4070 default: 4071 llvm_unreachable("subscript has unexpected classification"); 4072 } 4073 } 4074 4075 if (Coupled.count()) { 4076 // test coupled subscript groups 4077 SmallVector<Constraint, 4> Constraints(MaxLevels + 1); 4078 for (unsigned II = 0; II <= MaxLevels; ++II) 4079 Constraints[II].setAny(SE); 4080 for (unsigned SI : Coupled.set_bits()) { 4081 SmallBitVector Group(Pair[SI].Group); 4082 SmallBitVector Sivs(Pairs); 4083 SmallBitVector Mivs(Pairs); 4084 SmallBitVector ConstrainedLevels(MaxLevels + 1); 4085 for (unsigned SJ : Group.set_bits()) { 4086 if (Pair[SJ].Classification == Subscript::SIV) 4087 Sivs.set(SJ); 4088 else 4089 Mivs.set(SJ); 4090 } 4091 while (Sivs.any()) { 4092 bool Changed = false; 4093 for (unsigned SJ : Sivs.set_bits()) { 4094 // SJ is an SIV subscript that's part of the current coupled group 4095 unsigned Level; 4096 const SCEV *SplitIter = nullptr; 4097 (void) testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level, 4098 Result, NewConstraint, SplitIter); 4099 if (Level == SplitLevel && SplitIter) 4100 return SplitIter; 4101 ConstrainedLevels.set(Level); 4102 if (intersectConstraints(&Constraints[Level], &NewConstraint)) 4103 Changed = true; 4104 Sivs.reset(SJ); 4105 } 4106 if (Changed) { 4107 // propagate, possibly creating new SIVs and ZIVs 4108 for (unsigned SJ : Mivs.set_bits()) { 4109 // SJ is an MIV subscript that's part of the current coupled group 4110 if (propagate(Pair[SJ].Src, Pair[SJ].Dst, 4111 Pair[SJ].Loops, Constraints, Result.Consistent)) { 4112 Pair[SJ].Classification = 4113 classifyPair(Pair[SJ].Src, LI->getLoopFor(Src->getParent()), 4114 Pair[SJ].Dst, LI->getLoopFor(Dst->getParent()), 4115 Pair[SJ].Loops); 4116 switch (Pair[SJ].Classification) { 4117 case Subscript::ZIV: 4118 Mivs.reset(SJ); 4119 break; 4120 case Subscript::SIV: 4121 Sivs.set(SJ); 4122 Mivs.reset(SJ); 4123 break; 4124 case Subscript::RDIV: 4125 case Subscript::MIV: 4126 break; 4127 default: 4128 llvm_unreachable("bad subscript classification"); 4129 } 4130 } 4131 } 4132 } 4133 } 4134 } 4135 } 4136 llvm_unreachable("somehow reached end of routine"); 4137 return nullptr; 4138 } 4139