1 //===- LoopVectorizationLegality.cpp --------------------------------------===// 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 // This file provides loop vectorization legality analysis. Original code 10 // resided in LoopVectorize.cpp for a long time. 11 // 12 // At this point, it is implemented as a utility class, not as an analysis 13 // pass. It should be easy to create an analysis pass around it if there 14 // is a need (but D45420 needs to happen first). 15 // 16 17 #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" 18 #include "llvm/Analysis/Loads.h" 19 #include "llvm/Analysis/LoopInfo.h" 20 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 21 #include "llvm/Analysis/TargetLibraryInfo.h" 22 #include "llvm/Analysis/TargetTransformInfo.h" 23 #include "llvm/Analysis/ValueTracking.h" 24 #include "llvm/Analysis/VectorUtils.h" 25 #include "llvm/IR/IntrinsicInst.h" 26 #include "llvm/IR/PatternMatch.h" 27 #include "llvm/Transforms/Utils/SizeOpts.h" 28 #include "llvm/Transforms/Vectorize/LoopVectorize.h" 29 30 using namespace llvm; 31 using namespace PatternMatch; 32 33 #define LV_NAME "loop-vectorize" 34 #define DEBUG_TYPE LV_NAME 35 36 static cl::opt<bool> 37 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden, 38 cl::desc("Enable if-conversion during vectorization.")); 39 40 namespace llvm { 41 cl::opt<bool> 42 HintsAllowReordering("hints-allow-reordering", cl::init(true), cl::Hidden, 43 cl::desc("Allow enabling loop hints to reorder " 44 "FP operations during vectorization.")); 45 } 46 47 // TODO: Move size-based thresholds out of legality checking, make cost based 48 // decisions instead of hard thresholds. 49 static cl::opt<unsigned> VectorizeSCEVCheckThreshold( 50 "vectorize-scev-check-threshold", cl::init(16), cl::Hidden, 51 cl::desc("The maximum number of SCEV checks allowed.")); 52 53 static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold( 54 "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden, 55 cl::desc("The maximum number of SCEV checks allowed with a " 56 "vectorize(enable) pragma")); 57 58 static cl::opt<LoopVectorizeHints::ScalableForceKind> 59 ForceScalableVectorization( 60 "scalable-vectorization", cl::init(LoopVectorizeHints::SK_Unspecified), 61 cl::Hidden, 62 cl::desc("Control whether the compiler can use scalable vectors to " 63 "vectorize a loop"), 64 cl::values( 65 clEnumValN(LoopVectorizeHints::SK_FixedWidthOnly, "off", 66 "Scalable vectorization is disabled."), 67 clEnumValN( 68 LoopVectorizeHints::SK_PreferScalable, "preferred", 69 "Scalable vectorization is available and favored when the " 70 "cost is inconclusive."), 71 clEnumValN( 72 LoopVectorizeHints::SK_PreferScalable, "on", 73 "Scalable vectorization is available and favored when the " 74 "cost is inconclusive."))); 75 76 /// Maximum vectorization interleave count. 77 static const unsigned MaxInterleaveFactor = 16; 78 79 namespace llvm { 80 81 bool LoopVectorizeHints::Hint::validate(unsigned Val) { 82 switch (Kind) { 83 case HK_WIDTH: 84 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth; 85 case HK_INTERLEAVE: 86 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor; 87 case HK_FORCE: 88 return (Val <= 1); 89 case HK_ISVECTORIZED: 90 case HK_PREDICATE: 91 case HK_SCALABLE: 92 return (Val == 0 || Val == 1); 93 } 94 return false; 95 } 96 97 LoopVectorizeHints::LoopVectorizeHints(const Loop *L, 98 bool InterleaveOnlyWhenForced, 99 OptimizationRemarkEmitter &ORE, 100 const TargetTransformInfo *TTI) 101 : Width("vectorize.width", VectorizerParams::VectorizationFactor, HK_WIDTH), 102 Interleave("interleave.count", InterleaveOnlyWhenForced, HK_INTERLEAVE), 103 Force("vectorize.enable", FK_Undefined, HK_FORCE), 104 IsVectorized("isvectorized", 0, HK_ISVECTORIZED), 105 Predicate("vectorize.predicate.enable", FK_Undefined, HK_PREDICATE), 106 Scalable("vectorize.scalable.enable", SK_Unspecified, HK_SCALABLE), 107 TheLoop(L), ORE(ORE) { 108 // Populate values with existing loop metadata. 109 getHintsFromMetadata(); 110 111 // force-vector-interleave overrides DisableInterleaving. 112 if (VectorizerParams::isInterleaveForced()) 113 Interleave.Value = VectorizerParams::VectorizationInterleave; 114 115 // If the metadata doesn't explicitly specify whether to enable scalable 116 // vectorization, then decide based on the following criteria (increasing 117 // level of priority): 118 // - Target default 119 // - Metadata width 120 // - Force option (always overrides) 121 if ((LoopVectorizeHints::ScalableForceKind)Scalable.Value == SK_Unspecified) { 122 if (TTI) 123 Scalable.Value = TTI->enableScalableVectorization() ? SK_PreferScalable 124 : SK_FixedWidthOnly; 125 126 if (Width.Value) 127 // If the width is set, but the metadata says nothing about the scalable 128 // property, then assume it concerns only a fixed-width UserVF. 129 // If width is not set, the flag takes precedence. 130 Scalable.Value = SK_FixedWidthOnly; 131 } 132 133 // If the flag is set to force any use of scalable vectors, override the loop 134 // hints. 135 if (ForceScalableVectorization.getValue() != 136 LoopVectorizeHints::SK_Unspecified) 137 Scalable.Value = ForceScalableVectorization.getValue(); 138 139 // Scalable vectorization is disabled if no preference is specified. 140 if ((LoopVectorizeHints::ScalableForceKind)Scalable.Value == SK_Unspecified) 141 Scalable.Value = SK_FixedWidthOnly; 142 143 if (IsVectorized.Value != 1) 144 // If the vectorization width and interleaving count are both 1 then 145 // consider the loop to have been already vectorized because there's 146 // nothing more that we can do. 147 IsVectorized.Value = 148 getWidth() == ElementCount::getFixed(1) && getInterleave() == 1; 149 LLVM_DEBUG(if (InterleaveOnlyWhenForced && getInterleave() == 1) dbgs() 150 << "LV: Interleaving disabled by the pass manager\n"); 151 } 152 153 void LoopVectorizeHints::setAlreadyVectorized() { 154 LLVMContext &Context = TheLoop->getHeader()->getContext(); 155 156 MDNode *IsVectorizedMD = MDNode::get( 157 Context, 158 {MDString::get(Context, "llvm.loop.isvectorized"), 159 ConstantAsMetadata::get(ConstantInt::get(Context, APInt(32, 1)))}); 160 MDNode *LoopID = TheLoop->getLoopID(); 161 MDNode *NewLoopID = 162 makePostTransformationMetadata(Context, LoopID, 163 {Twine(Prefix(), "vectorize.").str(), 164 Twine(Prefix(), "interleave.").str()}, 165 {IsVectorizedMD}); 166 TheLoop->setLoopID(NewLoopID); 167 168 // Update internal cache. 169 IsVectorized.Value = 1; 170 } 171 172 bool LoopVectorizeHints::allowVectorization( 173 Function *F, Loop *L, bool VectorizeOnlyWhenForced) const { 174 if (getForce() == LoopVectorizeHints::FK_Disabled) { 175 LLVM_DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n"); 176 emitRemarkWithHints(); 177 return false; 178 } 179 180 if (VectorizeOnlyWhenForced && getForce() != LoopVectorizeHints::FK_Enabled) { 181 LLVM_DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n"); 182 emitRemarkWithHints(); 183 return false; 184 } 185 186 if (getIsVectorized() == 1) { 187 LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n"); 188 // FIXME: Add interleave.disable metadata. This will allow 189 // vectorize.disable to be used without disabling the pass and errors 190 // to differentiate between disabled vectorization and a width of 1. 191 ORE.emit([&]() { 192 return OptimizationRemarkAnalysis(vectorizeAnalysisPassName(), 193 "AllDisabled", L->getStartLoc(), 194 L->getHeader()) 195 << "loop not vectorized: vectorization and interleaving are " 196 "explicitly disabled, or the loop has already been " 197 "vectorized"; 198 }); 199 return false; 200 } 201 202 return true; 203 } 204 205 void LoopVectorizeHints::emitRemarkWithHints() const { 206 using namespace ore; 207 208 ORE.emit([&]() { 209 if (Force.Value == LoopVectorizeHints::FK_Disabled) 210 return OptimizationRemarkMissed(LV_NAME, "MissedExplicitlyDisabled", 211 TheLoop->getStartLoc(), 212 TheLoop->getHeader()) 213 << "loop not vectorized: vectorization is explicitly disabled"; 214 else { 215 OptimizationRemarkMissed R(LV_NAME, "MissedDetails", 216 TheLoop->getStartLoc(), TheLoop->getHeader()); 217 R << "loop not vectorized"; 218 if (Force.Value == LoopVectorizeHints::FK_Enabled) { 219 R << " (Force=" << NV("Force", true); 220 if (Width.Value != 0) 221 R << ", Vector Width=" << NV("VectorWidth", getWidth()); 222 if (getInterleave() != 0) 223 R << ", Interleave Count=" << NV("InterleaveCount", getInterleave()); 224 R << ")"; 225 } 226 return R; 227 } 228 }); 229 } 230 231 const char *LoopVectorizeHints::vectorizeAnalysisPassName() const { 232 if (getWidth() == ElementCount::getFixed(1)) 233 return LV_NAME; 234 if (getForce() == LoopVectorizeHints::FK_Disabled) 235 return LV_NAME; 236 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth().isZero()) 237 return LV_NAME; 238 return OptimizationRemarkAnalysis::AlwaysPrint; 239 } 240 241 bool LoopVectorizeHints::allowReordering() const { 242 // Allow the vectorizer to change the order of operations if enabling 243 // loop hints are provided 244 ElementCount EC = getWidth(); 245 return HintsAllowReordering && 246 (getForce() == LoopVectorizeHints::FK_Enabled || 247 EC.getKnownMinValue() > 1); 248 } 249 250 void LoopVectorizeHints::getHintsFromMetadata() { 251 MDNode *LoopID = TheLoop->getLoopID(); 252 if (!LoopID) 253 return; 254 255 // First operand should refer to the loop id itself. 256 assert(LoopID->getNumOperands() > 0 && "requires at least one operand"); 257 assert(LoopID->getOperand(0) == LoopID && "invalid loop id"); 258 259 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 260 const MDString *S = nullptr; 261 SmallVector<Metadata *, 4> Args; 262 263 // The expected hint is either a MDString or a MDNode with the first 264 // operand a MDString. 265 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) { 266 if (!MD || MD->getNumOperands() == 0) 267 continue; 268 S = dyn_cast<MDString>(MD->getOperand(0)); 269 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i) 270 Args.push_back(MD->getOperand(i)); 271 } else { 272 S = dyn_cast<MDString>(LoopID->getOperand(i)); 273 assert(Args.size() == 0 && "too many arguments for MDString"); 274 } 275 276 if (!S) 277 continue; 278 279 // Check if the hint starts with the loop metadata prefix. 280 StringRef Name = S->getString(); 281 if (Args.size() == 1) 282 setHint(Name, Args[0]); 283 } 284 } 285 286 void LoopVectorizeHints::setHint(StringRef Name, Metadata *Arg) { 287 if (!Name.startswith(Prefix())) 288 return; 289 Name = Name.substr(Prefix().size(), StringRef::npos); 290 291 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg); 292 if (!C) 293 return; 294 unsigned Val = C->getZExtValue(); 295 296 Hint *Hints[] = {&Width, &Interleave, &Force, 297 &IsVectorized, &Predicate, &Scalable}; 298 for (auto H : Hints) { 299 if (Name == H->Name) { 300 if (H->validate(Val)) 301 H->Value = Val; 302 else 303 LLVM_DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n"); 304 break; 305 } 306 } 307 } 308 309 // Return true if the inner loop \p Lp is uniform with regard to the outer loop 310 // \p OuterLp (i.e., if the outer loop is vectorized, all the vector lanes 311 // executing the inner loop will execute the same iterations). This check is 312 // very constrained for now but it will be relaxed in the future. \p Lp is 313 // considered uniform if it meets all the following conditions: 314 // 1) it has a canonical IV (starting from 0 and with stride 1), 315 // 2) its latch terminator is a conditional branch and, 316 // 3) its latch condition is a compare instruction whose operands are the 317 // canonical IV and an OuterLp invariant. 318 // This check doesn't take into account the uniformity of other conditions not 319 // related to the loop latch because they don't affect the loop uniformity. 320 // 321 // NOTE: We decided to keep all these checks and its associated documentation 322 // together so that we can easily have a picture of the current supported loop 323 // nests. However, some of the current checks don't depend on \p OuterLp and 324 // would be redundantly executed for each \p Lp if we invoked this function for 325 // different candidate outer loops. This is not the case for now because we 326 // don't currently have the infrastructure to evaluate multiple candidate outer 327 // loops and \p OuterLp will be a fixed parameter while we only support explicit 328 // outer loop vectorization. It's also very likely that these checks go away 329 // before introducing the aforementioned infrastructure. However, if this is not 330 // the case, we should move the \p OuterLp independent checks to a separate 331 // function that is only executed once for each \p Lp. 332 static bool isUniformLoop(Loop *Lp, Loop *OuterLp) { 333 assert(Lp->getLoopLatch() && "Expected loop with a single latch."); 334 335 // If Lp is the outer loop, it's uniform by definition. 336 if (Lp == OuterLp) 337 return true; 338 assert(OuterLp->contains(Lp) && "OuterLp must contain Lp."); 339 340 // 1. 341 PHINode *IV = Lp->getCanonicalInductionVariable(); 342 if (!IV) { 343 LLVM_DEBUG(dbgs() << "LV: Canonical IV not found.\n"); 344 return false; 345 } 346 347 // 2. 348 BasicBlock *Latch = Lp->getLoopLatch(); 349 auto *LatchBr = dyn_cast<BranchInst>(Latch->getTerminator()); 350 if (!LatchBr || LatchBr->isUnconditional()) { 351 LLVM_DEBUG(dbgs() << "LV: Unsupported loop latch branch.\n"); 352 return false; 353 } 354 355 // 3. 356 auto *LatchCmp = dyn_cast<CmpInst>(LatchBr->getCondition()); 357 if (!LatchCmp) { 358 LLVM_DEBUG( 359 dbgs() << "LV: Loop latch condition is not a compare instruction.\n"); 360 return false; 361 } 362 363 Value *CondOp0 = LatchCmp->getOperand(0); 364 Value *CondOp1 = LatchCmp->getOperand(1); 365 Value *IVUpdate = IV->getIncomingValueForBlock(Latch); 366 if (!(CondOp0 == IVUpdate && OuterLp->isLoopInvariant(CondOp1)) && 367 !(CondOp1 == IVUpdate && OuterLp->isLoopInvariant(CondOp0))) { 368 LLVM_DEBUG(dbgs() << "LV: Loop latch condition is not uniform.\n"); 369 return false; 370 } 371 372 return true; 373 } 374 375 // Return true if \p Lp and all its nested loops are uniform with regard to \p 376 // OuterLp. 377 static bool isUniformLoopNest(Loop *Lp, Loop *OuterLp) { 378 if (!isUniformLoop(Lp, OuterLp)) 379 return false; 380 381 // Check if nested loops are uniform. 382 for (Loop *SubLp : *Lp) 383 if (!isUniformLoopNest(SubLp, OuterLp)) 384 return false; 385 386 return true; 387 } 388 389 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) { 390 if (Ty->isPointerTy()) 391 return DL.getIntPtrType(Ty); 392 393 // It is possible that char's or short's overflow when we ask for the loop's 394 // trip count, work around this by changing the type size. 395 if (Ty->getScalarSizeInBits() < 32) 396 return Type::getInt32Ty(Ty->getContext()); 397 398 return Ty; 399 } 400 401 static Type *getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) { 402 Ty0 = convertPointerToIntegerType(DL, Ty0); 403 Ty1 = convertPointerToIntegerType(DL, Ty1); 404 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits()) 405 return Ty0; 406 return Ty1; 407 } 408 409 /// Check that the instruction has outside loop users and is not an 410 /// identified reduction variable. 411 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst, 412 SmallPtrSetImpl<Value *> &AllowedExit) { 413 // Reductions, Inductions and non-header phis are allowed to have exit users. All 414 // other instructions must not have external users. 415 if (!AllowedExit.count(Inst)) 416 // Check that all of the users of the loop are inside the BB. 417 for (User *U : Inst->users()) { 418 Instruction *UI = cast<Instruction>(U); 419 // This user may be a reduction exit value. 420 if (!TheLoop->contains(UI)) { 421 LLVM_DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n'); 422 return true; 423 } 424 } 425 return false; 426 } 427 428 /// Returns true if A and B have same pointer operands or same SCEVs addresses 429 static bool storeToSameAddress(ScalarEvolution *SE, StoreInst *A, 430 StoreInst *B) { 431 // Compare store 432 if (A == B) 433 return true; 434 435 // Otherwise Compare pointers 436 Value *APtr = A->getPointerOperand(); 437 Value *BPtr = B->getPointerOperand(); 438 if (APtr == BPtr) 439 return true; 440 441 // Otherwise compare address SCEVs 442 if (SE->getSCEV(APtr) == SE->getSCEV(BPtr)) 443 return true; 444 445 return false; 446 } 447 448 int LoopVectorizationLegality::isConsecutivePtr(Type *AccessTy, 449 Value *Ptr) const { 450 const ValueToValueMap &Strides = 451 getSymbolicStrides() ? *getSymbolicStrides() : ValueToValueMap(); 452 453 Function *F = TheLoop->getHeader()->getParent(); 454 bool OptForSize = F->hasOptSize() || 455 llvm::shouldOptimizeForSize(TheLoop->getHeader(), PSI, BFI, 456 PGSOQueryType::IRPass); 457 bool CanAddPredicate = !OptForSize; 458 int Stride = getPtrStride(PSE, AccessTy, Ptr, TheLoop, Strides, 459 CanAddPredicate, false); 460 if (Stride == 1 || Stride == -1) 461 return Stride; 462 return 0; 463 } 464 465 bool LoopVectorizationLegality::isUniform(Value *V) { 466 return LAI->isUniform(V); 467 } 468 469 bool LoopVectorizationLegality::canVectorizeOuterLoop() { 470 assert(!TheLoop->isInnermost() && "We are not vectorizing an outer loop."); 471 // Store the result and return it at the end instead of exiting early, in case 472 // allowExtraAnalysis is used to report multiple reasons for not vectorizing. 473 bool Result = true; 474 bool DoExtraAnalysis = ORE->allowExtraAnalysis(DEBUG_TYPE); 475 476 for (BasicBlock *BB : TheLoop->blocks()) { 477 // Check whether the BB terminator is a BranchInst. Any other terminator is 478 // not supported yet. 479 auto *Br = dyn_cast<BranchInst>(BB->getTerminator()); 480 if (!Br) { 481 reportVectorizationFailure("Unsupported basic block terminator", 482 "loop control flow is not understood by vectorizer", 483 "CFGNotUnderstood", ORE, TheLoop); 484 if (DoExtraAnalysis) 485 Result = false; 486 else 487 return false; 488 } 489 490 // Check whether the BranchInst is a supported one. Only unconditional 491 // branches, conditional branches with an outer loop invariant condition or 492 // backedges are supported. 493 // FIXME: We skip these checks when VPlan predication is enabled as we 494 // want to allow divergent branches. This whole check will be removed 495 // once VPlan predication is on by default. 496 if (Br && Br->isConditional() && 497 !TheLoop->isLoopInvariant(Br->getCondition()) && 498 !LI->isLoopHeader(Br->getSuccessor(0)) && 499 !LI->isLoopHeader(Br->getSuccessor(1))) { 500 reportVectorizationFailure("Unsupported conditional branch", 501 "loop control flow is not understood by vectorizer", 502 "CFGNotUnderstood", ORE, TheLoop); 503 if (DoExtraAnalysis) 504 Result = false; 505 else 506 return false; 507 } 508 } 509 510 // Check whether inner loops are uniform. At this point, we only support 511 // simple outer loops scenarios with uniform nested loops. 512 if (!isUniformLoopNest(TheLoop /*loop nest*/, 513 TheLoop /*context outer loop*/)) { 514 reportVectorizationFailure("Outer loop contains divergent loops", 515 "loop control flow is not understood by vectorizer", 516 "CFGNotUnderstood", ORE, TheLoop); 517 if (DoExtraAnalysis) 518 Result = false; 519 else 520 return false; 521 } 522 523 // Check whether we are able to set up outer loop induction. 524 if (!setupOuterLoopInductions()) { 525 reportVectorizationFailure("Unsupported outer loop Phi(s)", 526 "Unsupported outer loop Phi(s)", 527 "UnsupportedPhi", ORE, TheLoop); 528 if (DoExtraAnalysis) 529 Result = false; 530 else 531 return false; 532 } 533 534 return Result; 535 } 536 537 void LoopVectorizationLegality::addInductionPhi( 538 PHINode *Phi, const InductionDescriptor &ID, 539 SmallPtrSetImpl<Value *> &AllowedExit) { 540 Inductions[Phi] = ID; 541 542 // In case this induction also comes with casts that we know we can ignore 543 // in the vectorized loop body, record them here. All casts could be recorded 544 // here for ignoring, but suffices to record only the first (as it is the 545 // only one that may bw used outside the cast sequence). 546 const SmallVectorImpl<Instruction *> &Casts = ID.getCastInsts(); 547 if (!Casts.empty()) 548 InductionCastsToIgnore.insert(*Casts.begin()); 549 550 Type *PhiTy = Phi->getType(); 551 const DataLayout &DL = Phi->getModule()->getDataLayout(); 552 553 // Get the widest type. 554 if (!PhiTy->isFloatingPointTy()) { 555 if (!WidestIndTy) 556 WidestIndTy = convertPointerToIntegerType(DL, PhiTy); 557 else 558 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy); 559 } 560 561 // Int inductions are special because we only allow one IV. 562 if (ID.getKind() == InductionDescriptor::IK_IntInduction && 563 ID.getConstIntStepValue() && ID.getConstIntStepValue()->isOne() && 564 isa<Constant>(ID.getStartValue()) && 565 cast<Constant>(ID.getStartValue())->isNullValue()) { 566 567 // Use the phi node with the widest type as induction. Use the last 568 // one if there are multiple (no good reason for doing this other 569 // than it is expedient). We've checked that it begins at zero and 570 // steps by one, so this is a canonical induction variable. 571 if (!PrimaryInduction || PhiTy == WidestIndTy) 572 PrimaryInduction = Phi; 573 } 574 575 // Both the PHI node itself, and the "post-increment" value feeding 576 // back into the PHI node may have external users. 577 // We can allow those uses, except if the SCEVs we have for them rely 578 // on predicates that only hold within the loop, since allowing the exit 579 // currently means re-using this SCEV outside the loop (see PR33706 for more 580 // details). 581 if (PSE.getPredicate().isAlwaysTrue()) { 582 AllowedExit.insert(Phi); 583 AllowedExit.insert(Phi->getIncomingValueForBlock(TheLoop->getLoopLatch())); 584 } 585 586 LLVM_DEBUG(dbgs() << "LV: Found an induction variable.\n"); 587 } 588 589 bool LoopVectorizationLegality::setupOuterLoopInductions() { 590 BasicBlock *Header = TheLoop->getHeader(); 591 592 // Returns true if a given Phi is a supported induction. 593 auto isSupportedPhi = [&](PHINode &Phi) -> bool { 594 InductionDescriptor ID; 595 if (InductionDescriptor::isInductionPHI(&Phi, TheLoop, PSE, ID) && 596 ID.getKind() == InductionDescriptor::IK_IntInduction) { 597 addInductionPhi(&Phi, ID, AllowedExit); 598 return true; 599 } else { 600 // Bail out for any Phi in the outer loop header that is not a supported 601 // induction. 602 LLVM_DEBUG( 603 dbgs() 604 << "LV: Found unsupported PHI for outer loop vectorization.\n"); 605 return false; 606 } 607 }; 608 609 if (llvm::all_of(Header->phis(), isSupportedPhi)) 610 return true; 611 else 612 return false; 613 } 614 615 /// Checks if a function is scalarizable according to the TLI, in 616 /// the sense that it should be vectorized and then expanded in 617 /// multiple scalar calls. This is represented in the 618 /// TLI via mappings that do not specify a vector name, as in the 619 /// following example: 620 /// 621 /// const VecDesc VecIntrinsics[] = { 622 /// {"llvm.phx.abs.i32", "", 4} 623 /// }; 624 static bool isTLIScalarize(const TargetLibraryInfo &TLI, const CallInst &CI) { 625 const StringRef ScalarName = CI.getCalledFunction()->getName(); 626 bool Scalarize = TLI.isFunctionVectorizable(ScalarName); 627 // Check that all known VFs are not associated to a vector 628 // function, i.e. the vector name is emty. 629 if (Scalarize) { 630 ElementCount WidestFixedVF, WidestScalableVF; 631 TLI.getWidestVF(ScalarName, WidestFixedVF, WidestScalableVF); 632 for (ElementCount VF = ElementCount::getFixed(2); 633 ElementCount::isKnownLE(VF, WidestFixedVF); VF *= 2) 634 Scalarize &= !TLI.isFunctionVectorizable(ScalarName, VF); 635 for (ElementCount VF = ElementCount::getScalable(1); 636 ElementCount::isKnownLE(VF, WidestScalableVF); VF *= 2) 637 Scalarize &= !TLI.isFunctionVectorizable(ScalarName, VF); 638 assert((WidestScalableVF.isZero() || !Scalarize) && 639 "Caller may decide to scalarize a variant using a scalable VF"); 640 } 641 return Scalarize; 642 } 643 644 bool LoopVectorizationLegality::canVectorizeInstrs() { 645 BasicBlock *Header = TheLoop->getHeader(); 646 647 // For each block in the loop. 648 for (BasicBlock *BB : TheLoop->blocks()) { 649 // Scan the instructions in the block and look for hazards. 650 for (Instruction &I : *BB) { 651 if (auto *Phi = dyn_cast<PHINode>(&I)) { 652 Type *PhiTy = Phi->getType(); 653 // Check that this PHI type is allowed. 654 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() && 655 !PhiTy->isPointerTy()) { 656 reportVectorizationFailure("Found a non-int non-pointer PHI", 657 "loop control flow is not understood by vectorizer", 658 "CFGNotUnderstood", ORE, TheLoop); 659 return false; 660 } 661 662 // If this PHINode is not in the header block, then we know that we 663 // can convert it to select during if-conversion. No need to check if 664 // the PHIs in this block are induction or reduction variables. 665 if (BB != Header) { 666 // Non-header phi nodes that have outside uses can be vectorized. Add 667 // them to the list of allowed exits. 668 // Unsafe cyclic dependencies with header phis are identified during 669 // legalization for reduction, induction and first order 670 // recurrences. 671 AllowedExit.insert(&I); 672 continue; 673 } 674 675 // We only allow if-converted PHIs with exactly two incoming values. 676 if (Phi->getNumIncomingValues() != 2) { 677 reportVectorizationFailure("Found an invalid PHI", 678 "loop control flow is not understood by vectorizer", 679 "CFGNotUnderstood", ORE, TheLoop, Phi); 680 return false; 681 } 682 683 RecurrenceDescriptor RedDes; 684 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, RedDes, DB, AC, 685 DT, PSE.getSE())) { 686 Requirements->addExactFPMathInst(RedDes.getExactFPMathInst()); 687 AllowedExit.insert(RedDes.getLoopExitInstr()); 688 Reductions[Phi] = RedDes; 689 continue; 690 } 691 692 // TODO: Instead of recording the AllowedExit, it would be good to record the 693 // complementary set: NotAllowedExit. These include (but may not be 694 // limited to): 695 // 1. Reduction phis as they represent the one-before-last value, which 696 // is not available when vectorized 697 // 2. Induction phis and increment when SCEV predicates cannot be used 698 // outside the loop - see addInductionPhi 699 // 3. Non-Phis with outside uses when SCEV predicates cannot be used 700 // outside the loop - see call to hasOutsideLoopUser in the non-phi 701 // handling below 702 // 4. FirstOrderRecurrence phis that can possibly be handled by 703 // extraction. 704 // By recording these, we can then reason about ways to vectorize each 705 // of these NotAllowedExit. 706 InductionDescriptor ID; 707 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID)) { 708 addInductionPhi(Phi, ID, AllowedExit); 709 Requirements->addExactFPMathInst(ID.getExactFPMathInst()); 710 continue; 711 } 712 713 if (RecurrenceDescriptor::isFirstOrderRecurrence(Phi, TheLoop, 714 SinkAfter, DT)) { 715 AllowedExit.insert(Phi); 716 FirstOrderRecurrences.insert(Phi); 717 continue; 718 } 719 720 // As a last resort, coerce the PHI to a AddRec expression 721 // and re-try classifying it a an induction PHI. 722 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID, true)) { 723 addInductionPhi(Phi, ID, AllowedExit); 724 continue; 725 } 726 727 reportVectorizationFailure("Found an unidentified PHI", 728 "value that could not be identified as " 729 "reduction is used outside the loop", 730 "NonReductionValueUsedOutsideLoop", ORE, TheLoop, Phi); 731 return false; 732 } // end of PHI handling 733 734 // We handle calls that: 735 // * Are debug info intrinsics. 736 // * Have a mapping to an IR intrinsic. 737 // * Have a vector version available. 738 auto *CI = dyn_cast<CallInst>(&I); 739 740 if (CI && !getVectorIntrinsicIDForCall(CI, TLI) && 741 !isa<DbgInfoIntrinsic>(CI) && 742 !(CI->getCalledFunction() && TLI && 743 (!VFDatabase::getMappings(*CI).empty() || 744 isTLIScalarize(*TLI, *CI)))) { 745 // If the call is a recognized math libary call, it is likely that 746 // we can vectorize it given loosened floating-point constraints. 747 LibFunc Func; 748 bool IsMathLibCall = 749 TLI && CI->getCalledFunction() && 750 CI->getType()->isFloatingPointTy() && 751 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) && 752 TLI->hasOptimizedCodeGen(Func); 753 754 if (IsMathLibCall) { 755 // TODO: Ideally, we should not use clang-specific language here, 756 // but it's hard to provide meaningful yet generic advice. 757 // Also, should this be guarded by allowExtraAnalysis() and/or be part 758 // of the returned info from isFunctionVectorizable()? 759 reportVectorizationFailure( 760 "Found a non-intrinsic callsite", 761 "library call cannot be vectorized. " 762 "Try compiling with -fno-math-errno, -ffast-math, " 763 "or similar flags", 764 "CantVectorizeLibcall", ORE, TheLoop, CI); 765 } else { 766 reportVectorizationFailure("Found a non-intrinsic callsite", 767 "call instruction cannot be vectorized", 768 "CantVectorizeLibcall", ORE, TheLoop, CI); 769 } 770 return false; 771 } 772 773 // Some intrinsics have scalar arguments and should be same in order for 774 // them to be vectorized (i.e. loop invariant). 775 if (CI) { 776 auto *SE = PSE.getSE(); 777 Intrinsic::ID IntrinID = getVectorIntrinsicIDForCall(CI, TLI); 778 for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) 779 if (isVectorIntrinsicWithScalarOpAtArg(IntrinID, i)) { 780 if (!SE->isLoopInvariant(PSE.getSCEV(CI->getOperand(i)), TheLoop)) { 781 reportVectorizationFailure("Found unvectorizable intrinsic", 782 "intrinsic instruction cannot be vectorized", 783 "CantVectorizeIntrinsic", ORE, TheLoop, CI); 784 return false; 785 } 786 } 787 } 788 789 // Check that the instruction return type is vectorizable. 790 // Also, we can't vectorize extractelement instructions. 791 if ((!VectorType::isValidElementType(I.getType()) && 792 !I.getType()->isVoidTy()) || 793 isa<ExtractElementInst>(I)) { 794 reportVectorizationFailure("Found unvectorizable type", 795 "instruction return type cannot be vectorized", 796 "CantVectorizeInstructionReturnType", ORE, TheLoop, &I); 797 return false; 798 } 799 800 // Check that the stored type is vectorizable. 801 if (auto *ST = dyn_cast<StoreInst>(&I)) { 802 Type *T = ST->getValueOperand()->getType(); 803 if (!VectorType::isValidElementType(T)) { 804 reportVectorizationFailure("Store instruction cannot be vectorized", 805 "store instruction cannot be vectorized", 806 "CantVectorizeStore", ORE, TheLoop, ST); 807 return false; 808 } 809 810 // For nontemporal stores, check that a nontemporal vector version is 811 // supported on the target. 812 if (ST->getMetadata(LLVMContext::MD_nontemporal)) { 813 // Arbitrarily try a vector of 2 elements. 814 auto *VecTy = FixedVectorType::get(T, /*NumElts=*/2); 815 assert(VecTy && "did not find vectorized version of stored type"); 816 if (!TTI->isLegalNTStore(VecTy, ST->getAlign())) { 817 reportVectorizationFailure( 818 "nontemporal store instruction cannot be vectorized", 819 "nontemporal store instruction cannot be vectorized", 820 "CantVectorizeNontemporalStore", ORE, TheLoop, ST); 821 return false; 822 } 823 } 824 825 } else if (auto *LD = dyn_cast<LoadInst>(&I)) { 826 if (LD->getMetadata(LLVMContext::MD_nontemporal)) { 827 // For nontemporal loads, check that a nontemporal vector version is 828 // supported on the target (arbitrarily try a vector of 2 elements). 829 auto *VecTy = FixedVectorType::get(I.getType(), /*NumElts=*/2); 830 assert(VecTy && "did not find vectorized version of load type"); 831 if (!TTI->isLegalNTLoad(VecTy, LD->getAlign())) { 832 reportVectorizationFailure( 833 "nontemporal load instruction cannot be vectorized", 834 "nontemporal load instruction cannot be vectorized", 835 "CantVectorizeNontemporalLoad", ORE, TheLoop, LD); 836 return false; 837 } 838 } 839 840 // FP instructions can allow unsafe algebra, thus vectorizable by 841 // non-IEEE-754 compliant SIMD units. 842 // This applies to floating-point math operations and calls, not memory 843 // operations, shuffles, or casts, as they don't change precision or 844 // semantics. 845 } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) && 846 !I.isFast()) { 847 LLVM_DEBUG(dbgs() << "LV: Found FP op with unsafe algebra.\n"); 848 Hints->setPotentiallyUnsafe(); 849 } 850 851 // Reduction instructions are allowed to have exit users. 852 // All other instructions must not have external users. 853 if (hasOutsideLoopUser(TheLoop, &I, AllowedExit)) { 854 // We can safely vectorize loops where instructions within the loop are 855 // used outside the loop only if the SCEV predicates within the loop is 856 // same as outside the loop. Allowing the exit means reusing the SCEV 857 // outside the loop. 858 if (PSE.getPredicate().isAlwaysTrue()) { 859 AllowedExit.insert(&I); 860 continue; 861 } 862 reportVectorizationFailure("Value cannot be used outside the loop", 863 "value cannot be used outside the loop", 864 "ValueUsedOutsideLoop", ORE, TheLoop, &I); 865 return false; 866 } 867 } // next instr. 868 } 869 870 if (!PrimaryInduction) { 871 if (Inductions.empty()) { 872 reportVectorizationFailure("Did not find one integer induction var", 873 "loop induction variable could not be identified", 874 "NoInductionVariable", ORE, TheLoop); 875 return false; 876 } else if (!WidestIndTy) { 877 reportVectorizationFailure("Did not find one integer induction var", 878 "integer loop induction variable could not be identified", 879 "NoIntegerInductionVariable", ORE, TheLoop); 880 return false; 881 } else { 882 LLVM_DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); 883 } 884 } 885 886 // For first order recurrences, we use the previous value (incoming value from 887 // the latch) to check if it dominates all users of the recurrence. Bail out 888 // if we have to sink such an instruction for another recurrence, as the 889 // dominance requirement may not hold after sinking. 890 BasicBlock *LoopLatch = TheLoop->getLoopLatch(); 891 if (any_of(FirstOrderRecurrences, [LoopLatch, this](const PHINode *Phi) { 892 Instruction *V = 893 cast<Instruction>(Phi->getIncomingValueForBlock(LoopLatch)); 894 return SinkAfter.find(V) != SinkAfter.end(); 895 })) 896 return false; 897 898 // Now we know the widest induction type, check if our found induction 899 // is the same size. If it's not, unset it here and InnerLoopVectorizer 900 // will create another. 901 if (PrimaryInduction && WidestIndTy != PrimaryInduction->getType()) 902 PrimaryInduction = nullptr; 903 904 return true; 905 } 906 907 bool LoopVectorizationLegality::canVectorizeMemory() { 908 LAI = &(*GetLAA)(*TheLoop); 909 const OptimizationRemarkAnalysis *LAR = LAI->getReport(); 910 if (LAR) { 911 ORE->emit([&]() { 912 return OptimizationRemarkAnalysis(Hints->vectorizeAnalysisPassName(), 913 "loop not vectorized: ", *LAR); 914 }); 915 } 916 917 if (!LAI->canVectorizeMemory()) 918 return false; 919 920 // We can vectorize stores to invariant address when final reduction value is 921 // guaranteed to be stored at the end of the loop. Also, if decision to 922 // vectorize loop is made, runtime checks are added so as to make sure that 923 // invariant address won't alias with any other objects. 924 if (!LAI->getStoresToInvariantAddresses().empty()) { 925 // For each invariant address, check its last stored value is unconditional. 926 for (StoreInst *SI : LAI->getStoresToInvariantAddresses()) { 927 if (isInvariantStoreOfReduction(SI) && 928 blockNeedsPredication(SI->getParent())) { 929 reportVectorizationFailure( 930 "We don't allow storing to uniform addresses", 931 "write of conditional recurring variant value to a loop " 932 "invariant address could not be vectorized", 933 "CantVectorizeStoreToLoopInvariantAddress", ORE, TheLoop); 934 return false; 935 } 936 } 937 938 if (LAI->hasDependenceInvolvingLoopInvariantAddress()) { 939 // For each invariant address, check its last stored value is the result 940 // of one of our reductions. 941 // 942 // We do not check if dependence with loads exists because they are 943 // currently rejected earlier in LoopAccessInfo::analyzeLoop. In case this 944 // behaviour changes we have to modify this code. 945 ScalarEvolution *SE = PSE.getSE(); 946 SmallVector<StoreInst *, 4> UnhandledStores; 947 for (StoreInst *SI : LAI->getStoresToInvariantAddresses()) { 948 if (isInvariantStoreOfReduction(SI)) { 949 // Earlier stores to this address are effectively deadcode. 950 // With opaque pointers it is possible for one pointer to be used with 951 // different sizes of stored values: 952 // store i32 0, ptr %x 953 // store i8 0, ptr %x 954 // The latest store doesn't complitely overwrite the first one in the 955 // example. That is why we have to make sure that types of stored 956 // values are same. 957 // TODO: Check that bitwidth of unhandled store is smaller then the 958 // one that overwrites it and add a test. 959 erase_if(UnhandledStores, [SE, SI](StoreInst *I) { 960 return storeToSameAddress(SE, SI, I) && 961 I->getValueOperand()->getType() == 962 SI->getValueOperand()->getType(); 963 }); 964 continue; 965 } 966 UnhandledStores.push_back(SI); 967 } 968 969 bool IsOK = UnhandledStores.empty(); 970 // TODO: we should also validate against InvariantMemSets. 971 if (!IsOK) { 972 reportVectorizationFailure( 973 "We don't allow storing to uniform addresses", 974 "write to a loop invariant address could not " 975 "be vectorized", 976 "CantVectorizeStoreToLoopInvariantAddress", ORE, TheLoop); 977 return false; 978 } 979 } 980 } 981 982 PSE.addPredicate(LAI->getPSE().getPredicate()); 983 return true; 984 } 985 986 bool LoopVectorizationLegality::canVectorizeFPMath( 987 bool EnableStrictReductions) { 988 989 // First check if there is any ExactFP math or if we allow reassociations 990 if (!Requirements->getExactFPInst() || Hints->allowReordering()) 991 return true; 992 993 // If the above is false, we have ExactFPMath & do not allow reordering. 994 // If the EnableStrictReductions flag is set, first check if we have any 995 // Exact FP induction vars, which we cannot vectorize. 996 if (!EnableStrictReductions || 997 any_of(getInductionVars(), [&](auto &Induction) -> bool { 998 InductionDescriptor IndDesc = Induction.second; 999 return IndDesc.getExactFPMathInst(); 1000 })) 1001 return false; 1002 1003 // We can now only vectorize if all reductions with Exact FP math also 1004 // have the isOrdered flag set, which indicates that we can move the 1005 // reduction operations in-loop. 1006 return (all_of(getReductionVars(), [&](auto &Reduction) -> bool { 1007 const RecurrenceDescriptor &RdxDesc = Reduction.second; 1008 return !RdxDesc.hasExactFPMath() || RdxDesc.isOrdered(); 1009 })); 1010 } 1011 1012 bool LoopVectorizationLegality::isInvariantStoreOfReduction(StoreInst *SI) { 1013 return any_of(getReductionVars(), [&](auto &Reduction) -> bool { 1014 const RecurrenceDescriptor &RdxDesc = Reduction.second; 1015 return RdxDesc.IntermediateStore == SI; 1016 }); 1017 } 1018 1019 bool LoopVectorizationLegality::isInvariantAddressOfReduction(Value *V) { 1020 return any_of(getReductionVars(), [&](auto &Reduction) -> bool { 1021 const RecurrenceDescriptor &RdxDesc = Reduction.second; 1022 if (!RdxDesc.IntermediateStore) 1023 return false; 1024 1025 ScalarEvolution *SE = PSE.getSE(); 1026 Value *InvariantAddress = RdxDesc.IntermediateStore->getPointerOperand(); 1027 return V == InvariantAddress || 1028 SE->getSCEV(V) == SE->getSCEV(InvariantAddress); 1029 }); 1030 } 1031 1032 bool LoopVectorizationLegality::isInductionPhi(const Value *V) const { 1033 Value *In0 = const_cast<Value *>(V); 1034 PHINode *PN = dyn_cast_or_null<PHINode>(In0); 1035 if (!PN) 1036 return false; 1037 1038 return Inductions.count(PN); 1039 } 1040 1041 const InductionDescriptor * 1042 LoopVectorizationLegality::getIntOrFpInductionDescriptor(PHINode *Phi) const { 1043 if (!isInductionPhi(Phi)) 1044 return nullptr; 1045 auto &ID = getInductionVars().find(Phi)->second; 1046 if (ID.getKind() == InductionDescriptor::IK_IntInduction || 1047 ID.getKind() == InductionDescriptor::IK_FpInduction) 1048 return &ID; 1049 return nullptr; 1050 } 1051 1052 const InductionDescriptor * 1053 LoopVectorizationLegality::getPointerInductionDescriptor(PHINode *Phi) const { 1054 if (!isInductionPhi(Phi)) 1055 return nullptr; 1056 auto &ID = getInductionVars().find(Phi)->second; 1057 if (ID.getKind() == InductionDescriptor::IK_PtrInduction) 1058 return &ID; 1059 return nullptr; 1060 } 1061 1062 bool LoopVectorizationLegality::isCastedInductionVariable( 1063 const Value *V) const { 1064 auto *Inst = dyn_cast<Instruction>(V); 1065 return (Inst && InductionCastsToIgnore.count(Inst)); 1066 } 1067 1068 bool LoopVectorizationLegality::isInductionVariable(const Value *V) const { 1069 return isInductionPhi(V) || isCastedInductionVariable(V); 1070 } 1071 1072 bool LoopVectorizationLegality::isFirstOrderRecurrence( 1073 const PHINode *Phi) const { 1074 return FirstOrderRecurrences.count(Phi); 1075 } 1076 1077 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) const { 1078 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT); 1079 } 1080 1081 bool LoopVectorizationLegality::blockCanBePredicated( 1082 BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs, 1083 SmallPtrSetImpl<const Instruction *> &MaskedOp, 1084 SmallPtrSetImpl<Instruction *> &ConditionalAssumes) const { 1085 for (Instruction &I : *BB) { 1086 // We can predicate blocks with calls to assume, as long as we drop them in 1087 // case we flatten the CFG via predication. 1088 if (match(&I, m_Intrinsic<Intrinsic::assume>())) { 1089 ConditionalAssumes.insert(&I); 1090 continue; 1091 } 1092 1093 // Do not let llvm.experimental.noalias.scope.decl block the vectorization. 1094 // TODO: there might be cases that it should block the vectorization. Let's 1095 // ignore those for now. 1096 if (isa<NoAliasScopeDeclInst>(&I)) 1097 continue; 1098 1099 // We might be able to hoist the load. 1100 if (I.mayReadFromMemory()) { 1101 auto *LI = dyn_cast<LoadInst>(&I); 1102 if (!LI) 1103 return false; 1104 if (!SafePtrs.count(LI->getPointerOperand())) { 1105 MaskedOp.insert(LI); 1106 continue; 1107 } 1108 } 1109 1110 if (I.mayWriteToMemory()) { 1111 auto *SI = dyn_cast<StoreInst>(&I); 1112 if (!SI) 1113 return false; 1114 // Predicated store requires some form of masking: 1115 // 1) masked store HW instruction, 1116 // 2) emulation via load-blend-store (only if safe and legal to do so, 1117 // be aware on the race conditions), or 1118 // 3) element-by-element predicate check and scalar store. 1119 MaskedOp.insert(SI); 1120 continue; 1121 } 1122 if (I.mayThrow()) 1123 return false; 1124 } 1125 1126 return true; 1127 } 1128 1129 bool LoopVectorizationLegality::canVectorizeWithIfConvert() { 1130 if (!EnableIfConversion) { 1131 reportVectorizationFailure("If-conversion is disabled", 1132 "if-conversion is disabled", 1133 "IfConversionDisabled", 1134 ORE, TheLoop); 1135 return false; 1136 } 1137 1138 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"); 1139 1140 // A list of pointers which are known to be dereferenceable within scope of 1141 // the loop body for each iteration of the loop which executes. That is, 1142 // the memory pointed to can be dereferenced (with the access size implied by 1143 // the value's type) unconditionally within the loop header without 1144 // introducing a new fault. 1145 SmallPtrSet<Value *, 8> SafePointers; 1146 1147 // Collect safe addresses. 1148 for (BasicBlock *BB : TheLoop->blocks()) { 1149 if (!blockNeedsPredication(BB)) { 1150 for (Instruction &I : *BB) 1151 if (auto *Ptr = getLoadStorePointerOperand(&I)) 1152 SafePointers.insert(Ptr); 1153 continue; 1154 } 1155 1156 // For a block which requires predication, a address may be safe to access 1157 // in the loop w/o predication if we can prove dereferenceability facts 1158 // sufficient to ensure it'll never fault within the loop. For the moment, 1159 // we restrict this to loads; stores are more complicated due to 1160 // concurrency restrictions. 1161 ScalarEvolution &SE = *PSE.getSE(); 1162 for (Instruction &I : *BB) { 1163 LoadInst *LI = dyn_cast<LoadInst>(&I); 1164 if (LI && !LI->getType()->isVectorTy() && !mustSuppressSpeculation(*LI) && 1165 isDereferenceableAndAlignedInLoop(LI, TheLoop, SE, *DT)) 1166 SafePointers.insert(LI->getPointerOperand()); 1167 } 1168 } 1169 1170 // Collect the blocks that need predication. 1171 for (BasicBlock *BB : TheLoop->blocks()) { 1172 // We don't support switch statements inside loops. 1173 if (!isa<BranchInst>(BB->getTerminator())) { 1174 reportVectorizationFailure("Loop contains a switch statement", 1175 "loop contains a switch statement", 1176 "LoopContainsSwitch", ORE, TheLoop, 1177 BB->getTerminator()); 1178 return false; 1179 } 1180 1181 // We must be able to predicate all blocks that need to be predicated. 1182 if (blockNeedsPredication(BB)) { 1183 if (!blockCanBePredicated(BB, SafePointers, MaskedOp, 1184 ConditionalAssumes)) { 1185 reportVectorizationFailure( 1186 "Control flow cannot be substituted for a select", 1187 "control flow cannot be substituted for a select", 1188 "NoCFGForSelect", ORE, TheLoop, 1189 BB->getTerminator()); 1190 return false; 1191 } 1192 } 1193 } 1194 1195 // We can if-convert this loop. 1196 return true; 1197 } 1198 1199 // Helper function to canVectorizeLoopNestCFG. 1200 bool LoopVectorizationLegality::canVectorizeLoopCFG(Loop *Lp, 1201 bool UseVPlanNativePath) { 1202 assert((UseVPlanNativePath || Lp->isInnermost()) && 1203 "VPlan-native path is not enabled."); 1204 1205 // TODO: ORE should be improved to show more accurate information when an 1206 // outer loop can't be vectorized because a nested loop is not understood or 1207 // legal. Something like: "outer_loop_location: loop not vectorized: 1208 // (inner_loop_location) loop control flow is not understood by vectorizer". 1209 1210 // Store the result and return it at the end instead of exiting early, in case 1211 // allowExtraAnalysis is used to report multiple reasons for not vectorizing. 1212 bool Result = true; 1213 bool DoExtraAnalysis = ORE->allowExtraAnalysis(DEBUG_TYPE); 1214 1215 // We must have a loop in canonical form. Loops with indirectbr in them cannot 1216 // be canonicalized. 1217 if (!Lp->getLoopPreheader()) { 1218 reportVectorizationFailure("Loop doesn't have a legal pre-header", 1219 "loop control flow is not understood by vectorizer", 1220 "CFGNotUnderstood", ORE, TheLoop); 1221 if (DoExtraAnalysis) 1222 Result = false; 1223 else 1224 return false; 1225 } 1226 1227 // We must have a single backedge. 1228 if (Lp->getNumBackEdges() != 1) { 1229 reportVectorizationFailure("The loop must have a single backedge", 1230 "loop control flow is not understood by vectorizer", 1231 "CFGNotUnderstood", ORE, TheLoop); 1232 if (DoExtraAnalysis) 1233 Result = false; 1234 else 1235 return false; 1236 } 1237 1238 return Result; 1239 } 1240 1241 bool LoopVectorizationLegality::canVectorizeLoopNestCFG( 1242 Loop *Lp, bool UseVPlanNativePath) { 1243 // Store the result and return it at the end instead of exiting early, in case 1244 // allowExtraAnalysis is used to report multiple reasons for not vectorizing. 1245 bool Result = true; 1246 bool DoExtraAnalysis = ORE->allowExtraAnalysis(DEBUG_TYPE); 1247 if (!canVectorizeLoopCFG(Lp, UseVPlanNativePath)) { 1248 if (DoExtraAnalysis) 1249 Result = false; 1250 else 1251 return false; 1252 } 1253 1254 // Recursively check whether the loop control flow of nested loops is 1255 // understood. 1256 for (Loop *SubLp : *Lp) 1257 if (!canVectorizeLoopNestCFG(SubLp, UseVPlanNativePath)) { 1258 if (DoExtraAnalysis) 1259 Result = false; 1260 else 1261 return false; 1262 } 1263 1264 return Result; 1265 } 1266 1267 bool LoopVectorizationLegality::canVectorize(bool UseVPlanNativePath) { 1268 // Store the result and return it at the end instead of exiting early, in case 1269 // allowExtraAnalysis is used to report multiple reasons for not vectorizing. 1270 bool Result = true; 1271 1272 bool DoExtraAnalysis = ORE->allowExtraAnalysis(DEBUG_TYPE); 1273 // Check whether the loop-related control flow in the loop nest is expected by 1274 // vectorizer. 1275 if (!canVectorizeLoopNestCFG(TheLoop, UseVPlanNativePath)) { 1276 if (DoExtraAnalysis) 1277 Result = false; 1278 else 1279 return false; 1280 } 1281 1282 // We need to have a loop header. 1283 LLVM_DEBUG(dbgs() << "LV: Found a loop: " << TheLoop->getHeader()->getName() 1284 << '\n'); 1285 1286 // Specific checks for outer loops. We skip the remaining legal checks at this 1287 // point because they don't support outer loops. 1288 if (!TheLoop->isInnermost()) { 1289 assert(UseVPlanNativePath && "VPlan-native path is not enabled."); 1290 1291 if (!canVectorizeOuterLoop()) { 1292 reportVectorizationFailure("Unsupported outer loop", 1293 "unsupported outer loop", 1294 "UnsupportedOuterLoop", 1295 ORE, TheLoop); 1296 // TODO: Implement DoExtraAnalysis when subsequent legal checks support 1297 // outer loops. 1298 return false; 1299 } 1300 1301 LLVM_DEBUG(dbgs() << "LV: We can vectorize this outer loop!\n"); 1302 return Result; 1303 } 1304 1305 assert(TheLoop->isInnermost() && "Inner loop expected."); 1306 // Check if we can if-convert non-single-bb loops. 1307 unsigned NumBlocks = TheLoop->getNumBlocks(); 1308 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { 1309 LLVM_DEBUG(dbgs() << "LV: Can't if-convert the loop.\n"); 1310 if (DoExtraAnalysis) 1311 Result = false; 1312 else 1313 return false; 1314 } 1315 1316 // Check if we can vectorize the instructions and CFG in this loop. 1317 if (!canVectorizeInstrs()) { 1318 LLVM_DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n"); 1319 if (DoExtraAnalysis) 1320 Result = false; 1321 else 1322 return false; 1323 } 1324 1325 // Go over each instruction and look at memory deps. 1326 if (!canVectorizeMemory()) { 1327 LLVM_DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n"); 1328 if (DoExtraAnalysis) 1329 Result = false; 1330 else 1331 return false; 1332 } 1333 1334 LLVM_DEBUG(dbgs() << "LV: We can vectorize this loop" 1335 << (LAI->getRuntimePointerChecking()->Need 1336 ? " (with a runtime bound check)" 1337 : "") 1338 << "!\n"); 1339 1340 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold; 1341 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled) 1342 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold; 1343 1344 if (PSE.getPredicate().getComplexity() > SCEVThreshold) { 1345 reportVectorizationFailure("Too many SCEV checks needed", 1346 "Too many SCEV assumptions need to be made and checked at runtime", 1347 "TooManySCEVRunTimeChecks", ORE, TheLoop); 1348 if (DoExtraAnalysis) 1349 Result = false; 1350 else 1351 return false; 1352 } 1353 1354 // Okay! We've done all the tests. If any have failed, return false. Otherwise 1355 // we can vectorize, and at this point we don't have any other mem analysis 1356 // which may limit our maximum vectorization factor, so just return true with 1357 // no restrictions. 1358 return Result; 1359 } 1360 1361 bool LoopVectorizationLegality::prepareToFoldTailByMasking() { 1362 1363 LLVM_DEBUG(dbgs() << "LV: checking if tail can be folded by masking.\n"); 1364 1365 SmallPtrSet<const Value *, 8> ReductionLiveOuts; 1366 1367 for (auto &Reduction : getReductionVars()) 1368 ReductionLiveOuts.insert(Reduction.second.getLoopExitInstr()); 1369 1370 // TODO: handle non-reduction outside users when tail is folded by masking. 1371 for (auto *AE : AllowedExit) { 1372 // Check that all users of allowed exit values are inside the loop or 1373 // are the live-out of a reduction. 1374 if (ReductionLiveOuts.count(AE)) 1375 continue; 1376 for (User *U : AE->users()) { 1377 Instruction *UI = cast<Instruction>(U); 1378 if (TheLoop->contains(UI)) 1379 continue; 1380 LLVM_DEBUG( 1381 dbgs() 1382 << "LV: Cannot fold tail by masking, loop has an outside user for " 1383 << *UI << "\n"); 1384 return false; 1385 } 1386 } 1387 1388 // The list of pointers that we can safely read and write to remains empty. 1389 SmallPtrSet<Value *, 8> SafePointers; 1390 1391 SmallPtrSet<const Instruction *, 8> TmpMaskedOp; 1392 SmallPtrSet<Instruction *, 8> TmpConditionalAssumes; 1393 1394 // Check and mark all blocks for predication, including those that ordinarily 1395 // do not need predication such as the header block. 1396 for (BasicBlock *BB : TheLoop->blocks()) { 1397 if (!blockCanBePredicated(BB, SafePointers, TmpMaskedOp, 1398 TmpConditionalAssumes)) { 1399 LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking as requested.\n"); 1400 return false; 1401 } 1402 } 1403 1404 LLVM_DEBUG(dbgs() << "LV: can fold tail by masking.\n"); 1405 1406 MaskedOp.insert(TmpMaskedOp.begin(), TmpMaskedOp.end()); 1407 ConditionalAssumes.insert(TmpConditionalAssumes.begin(), 1408 TmpConditionalAssumes.end()); 1409 1410 return true; 1411 } 1412 1413 } // namespace llvm 1414