1 //===-- LoopUtils.cpp - Loop Utility functions -------------------------===// 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 defines common loop utility functions. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/Transforms/Utils/LoopUtils.h" 14 #include "llvm/ADT/DenseSet.h" 15 #include "llvm/ADT/PriorityWorklist.h" 16 #include "llvm/ADT/ScopeExit.h" 17 #include "llvm/ADT/SetVector.h" 18 #include "llvm/ADT/SmallPtrSet.h" 19 #include "llvm/ADT/SmallVector.h" 20 #include "llvm/Analysis/AliasAnalysis.h" 21 #include "llvm/Analysis/BasicAliasAnalysis.h" 22 #include "llvm/Analysis/DomTreeUpdater.h" 23 #include "llvm/Analysis/GlobalsModRef.h" 24 #include "llvm/Analysis/InstSimplifyFolder.h" 25 #include "llvm/Analysis/LoopAccessAnalysis.h" 26 #include "llvm/Analysis/LoopInfo.h" 27 #include "llvm/Analysis/LoopPass.h" 28 #include "llvm/Analysis/MemorySSA.h" 29 #include "llvm/Analysis/MemorySSAUpdater.h" 30 #include "llvm/Analysis/ScalarEvolution.h" 31 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" 32 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 33 #include "llvm/IR/DIBuilder.h" 34 #include "llvm/IR/Dominators.h" 35 #include "llvm/IR/Instructions.h" 36 #include "llvm/IR/IntrinsicInst.h" 37 #include "llvm/IR/MDBuilder.h" 38 #include "llvm/IR/Module.h" 39 #include "llvm/IR/PatternMatch.h" 40 #include "llvm/IR/ProfDataUtils.h" 41 #include "llvm/IR/ValueHandle.h" 42 #include "llvm/InitializePasses.h" 43 #include "llvm/Pass.h" 44 #include "llvm/Support/Debug.h" 45 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 46 #include "llvm/Transforms/Utils/Local.h" 47 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 48 49 using namespace llvm; 50 using namespace llvm::PatternMatch; 51 52 #define DEBUG_TYPE "loop-utils" 53 54 static const char *LLVMLoopDisableNonforced = "llvm.loop.disable_nonforced"; 55 static const char *LLVMLoopDisableLICM = "llvm.licm.disable"; 56 57 bool llvm::formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI, 58 MemorySSAUpdater *MSSAU, 59 bool PreserveLCSSA) { 60 bool Changed = false; 61 62 // We re-use a vector for the in-loop predecesosrs. 63 SmallVector<BasicBlock *, 4> InLoopPredecessors; 64 65 auto RewriteExit = [&](BasicBlock *BB) { 66 assert(InLoopPredecessors.empty() && 67 "Must start with an empty predecessors list!"); 68 auto Cleanup = make_scope_exit([&] { InLoopPredecessors.clear(); }); 69 70 // See if there are any non-loop predecessors of this exit block and 71 // keep track of the in-loop predecessors. 72 bool IsDedicatedExit = true; 73 for (auto *PredBB : predecessors(BB)) 74 if (L->contains(PredBB)) { 75 if (isa<IndirectBrInst>(PredBB->getTerminator())) 76 // We cannot rewrite exiting edges from an indirectbr. 77 return false; 78 79 InLoopPredecessors.push_back(PredBB); 80 } else { 81 IsDedicatedExit = false; 82 } 83 84 assert(!InLoopPredecessors.empty() && "Must have *some* loop predecessor!"); 85 86 // Nothing to do if this is already a dedicated exit. 87 if (IsDedicatedExit) 88 return false; 89 90 auto *NewExitBB = SplitBlockPredecessors( 91 BB, InLoopPredecessors, ".loopexit", DT, LI, MSSAU, PreserveLCSSA); 92 93 if (!NewExitBB) 94 LLVM_DEBUG( 95 dbgs() << "WARNING: Can't create a dedicated exit block for loop: " 96 << *L << "\n"); 97 else 98 LLVM_DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block " 99 << NewExitBB->getName() << "\n"); 100 return true; 101 }; 102 103 // Walk the exit blocks directly rather than building up a data structure for 104 // them, but only visit each one once. 105 SmallPtrSet<BasicBlock *, 4> Visited; 106 for (auto *BB : L->blocks()) 107 for (auto *SuccBB : successors(BB)) { 108 // We're looking for exit blocks so skip in-loop successors. 109 if (L->contains(SuccBB)) 110 continue; 111 112 // Visit each exit block exactly once. 113 if (!Visited.insert(SuccBB).second) 114 continue; 115 116 Changed |= RewriteExit(SuccBB); 117 } 118 119 return Changed; 120 } 121 122 /// Returns the instructions that use values defined in the loop. 123 SmallVector<Instruction *, 8> llvm::findDefsUsedOutsideOfLoop(Loop *L) { 124 SmallVector<Instruction *, 8> UsedOutside; 125 126 for (auto *Block : L->getBlocks()) 127 // FIXME: I believe that this could use copy_if if the Inst reference could 128 // be adapted into a pointer. 129 for (auto &Inst : *Block) { 130 auto Users = Inst.users(); 131 if (any_of(Users, [&](User *U) { 132 auto *Use = cast<Instruction>(U); 133 return !L->contains(Use->getParent()); 134 })) 135 UsedOutside.push_back(&Inst); 136 } 137 138 return UsedOutside; 139 } 140 141 void llvm::getLoopAnalysisUsage(AnalysisUsage &AU) { 142 // By definition, all loop passes need the LoopInfo analysis and the 143 // Dominator tree it depends on. Because they all participate in the loop 144 // pass manager, they must also preserve these. 145 AU.addRequired<DominatorTreeWrapperPass>(); 146 AU.addPreserved<DominatorTreeWrapperPass>(); 147 AU.addRequired<LoopInfoWrapperPass>(); 148 AU.addPreserved<LoopInfoWrapperPass>(); 149 150 // We must also preserve LoopSimplify and LCSSA. We locally access their IDs 151 // here because users shouldn't directly get them from this header. 152 extern char &LoopSimplifyID; 153 extern char &LCSSAID; 154 AU.addRequiredID(LoopSimplifyID); 155 AU.addPreservedID(LoopSimplifyID); 156 AU.addRequiredID(LCSSAID); 157 AU.addPreservedID(LCSSAID); 158 // This is used in the LPPassManager to perform LCSSA verification on passes 159 // which preserve lcssa form 160 AU.addRequired<LCSSAVerificationPass>(); 161 AU.addPreserved<LCSSAVerificationPass>(); 162 163 // Loop passes are designed to run inside of a loop pass manager which means 164 // that any function analyses they require must be required by the first loop 165 // pass in the manager (so that it is computed before the loop pass manager 166 // runs) and preserved by all loop pasess in the manager. To make this 167 // reasonably robust, the set needed for most loop passes is maintained here. 168 // If your loop pass requires an analysis not listed here, you will need to 169 // carefully audit the loop pass manager nesting structure that results. 170 AU.addRequired<AAResultsWrapperPass>(); 171 AU.addPreserved<AAResultsWrapperPass>(); 172 AU.addPreserved<BasicAAWrapperPass>(); 173 AU.addPreserved<GlobalsAAWrapperPass>(); 174 AU.addPreserved<SCEVAAWrapperPass>(); 175 AU.addRequired<ScalarEvolutionWrapperPass>(); 176 AU.addPreserved<ScalarEvolutionWrapperPass>(); 177 // FIXME: When all loop passes preserve MemorySSA, it can be required and 178 // preserved here instead of the individual handling in each pass. 179 } 180 181 /// Manually defined generic "LoopPass" dependency initialization. This is used 182 /// to initialize the exact set of passes from above in \c 183 /// getLoopAnalysisUsage. It can be used within a loop pass's initialization 184 /// with: 185 /// 186 /// INITIALIZE_PASS_DEPENDENCY(LoopPass) 187 /// 188 /// As-if "LoopPass" were a pass. 189 void llvm::initializeLoopPassPass(PassRegistry &Registry) { 190 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 191 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 192 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 193 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass) 194 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 195 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass) 196 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) 197 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass) 198 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 199 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) 200 } 201 202 /// Create MDNode for input string. 203 static MDNode *createStringMetadata(Loop *TheLoop, StringRef Name, unsigned V) { 204 LLVMContext &Context = TheLoop->getHeader()->getContext(); 205 Metadata *MDs[] = { 206 MDString::get(Context, Name), 207 ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(Context), V))}; 208 return MDNode::get(Context, MDs); 209 } 210 211 /// Set input string into loop metadata by keeping other values intact. 212 /// If the string is already in loop metadata update value if it is 213 /// different. 214 void llvm::addStringMetadataToLoop(Loop *TheLoop, const char *StringMD, 215 unsigned V) { 216 SmallVector<Metadata *, 4> MDs(1); 217 // If the loop already has metadata, retain it. 218 MDNode *LoopID = TheLoop->getLoopID(); 219 if (LoopID) { 220 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 221 MDNode *Node = cast<MDNode>(LoopID->getOperand(i)); 222 // If it is of form key = value, try to parse it. 223 if (Node->getNumOperands() == 2) { 224 MDString *S = dyn_cast<MDString>(Node->getOperand(0)); 225 if (S && S->getString().equals(StringMD)) { 226 ConstantInt *IntMD = 227 mdconst::extract_or_null<ConstantInt>(Node->getOperand(1)); 228 if (IntMD && IntMD->getSExtValue() == V) 229 // It is already in place. Do nothing. 230 return; 231 // We need to update the value, so just skip it here and it will 232 // be added after copying other existed nodes. 233 continue; 234 } 235 } 236 MDs.push_back(Node); 237 } 238 } 239 // Add new metadata. 240 MDs.push_back(createStringMetadata(TheLoop, StringMD, V)); 241 // Replace current metadata node with new one. 242 LLVMContext &Context = TheLoop->getHeader()->getContext(); 243 MDNode *NewLoopID = MDNode::get(Context, MDs); 244 // Set operand 0 to refer to the loop id itself. 245 NewLoopID->replaceOperandWith(0, NewLoopID); 246 TheLoop->setLoopID(NewLoopID); 247 } 248 249 std::optional<ElementCount> 250 llvm::getOptionalElementCountLoopAttribute(const Loop *TheLoop) { 251 std::optional<int> Width = 252 getOptionalIntLoopAttribute(TheLoop, "llvm.loop.vectorize.width"); 253 254 if (Width) { 255 std::optional<int> IsScalable = getOptionalIntLoopAttribute( 256 TheLoop, "llvm.loop.vectorize.scalable.enable"); 257 return ElementCount::get(*Width, IsScalable.value_or(false)); 258 } 259 260 return std::nullopt; 261 } 262 263 std::optional<MDNode *> llvm::makeFollowupLoopID( 264 MDNode *OrigLoopID, ArrayRef<StringRef> FollowupOptions, 265 const char *InheritOptionsExceptPrefix, bool AlwaysNew) { 266 if (!OrigLoopID) { 267 if (AlwaysNew) 268 return nullptr; 269 return std::nullopt; 270 } 271 272 assert(OrigLoopID->getOperand(0) == OrigLoopID); 273 274 bool InheritAllAttrs = !InheritOptionsExceptPrefix; 275 bool InheritSomeAttrs = 276 InheritOptionsExceptPrefix && InheritOptionsExceptPrefix[0] != '\0'; 277 SmallVector<Metadata *, 8> MDs; 278 MDs.push_back(nullptr); 279 280 bool Changed = false; 281 if (InheritAllAttrs || InheritSomeAttrs) { 282 for (const MDOperand &Existing : drop_begin(OrigLoopID->operands())) { 283 MDNode *Op = cast<MDNode>(Existing.get()); 284 285 auto InheritThisAttribute = [InheritSomeAttrs, 286 InheritOptionsExceptPrefix](MDNode *Op) { 287 if (!InheritSomeAttrs) 288 return false; 289 290 // Skip malformatted attribute metadata nodes. 291 if (Op->getNumOperands() == 0) 292 return true; 293 Metadata *NameMD = Op->getOperand(0).get(); 294 if (!isa<MDString>(NameMD)) 295 return true; 296 StringRef AttrName = cast<MDString>(NameMD)->getString(); 297 298 // Do not inherit excluded attributes. 299 return !AttrName.startswith(InheritOptionsExceptPrefix); 300 }; 301 302 if (InheritThisAttribute(Op)) 303 MDs.push_back(Op); 304 else 305 Changed = true; 306 } 307 } else { 308 // Modified if we dropped at least one attribute. 309 Changed = OrigLoopID->getNumOperands() > 1; 310 } 311 312 bool HasAnyFollowup = false; 313 for (StringRef OptionName : FollowupOptions) { 314 MDNode *FollowupNode = findOptionMDForLoopID(OrigLoopID, OptionName); 315 if (!FollowupNode) 316 continue; 317 318 HasAnyFollowup = true; 319 for (const MDOperand &Option : drop_begin(FollowupNode->operands())) { 320 MDs.push_back(Option.get()); 321 Changed = true; 322 } 323 } 324 325 // Attributes of the followup loop not specified explicity, so signal to the 326 // transformation pass to add suitable attributes. 327 if (!AlwaysNew && !HasAnyFollowup) 328 return std::nullopt; 329 330 // If no attributes were added or remove, the previous loop Id can be reused. 331 if (!AlwaysNew && !Changed) 332 return OrigLoopID; 333 334 // No attributes is equivalent to having no !llvm.loop metadata at all. 335 if (MDs.size() == 1) 336 return nullptr; 337 338 // Build the new loop ID. 339 MDTuple *FollowupLoopID = MDNode::get(OrigLoopID->getContext(), MDs); 340 FollowupLoopID->replaceOperandWith(0, FollowupLoopID); 341 return FollowupLoopID; 342 } 343 344 bool llvm::hasDisableAllTransformsHint(const Loop *L) { 345 return getBooleanLoopAttribute(L, LLVMLoopDisableNonforced); 346 } 347 348 bool llvm::hasDisableLICMTransformsHint(const Loop *L) { 349 return getBooleanLoopAttribute(L, LLVMLoopDisableLICM); 350 } 351 352 TransformationMode llvm::hasUnrollTransformation(const Loop *L) { 353 if (getBooleanLoopAttribute(L, "llvm.loop.unroll.disable")) 354 return TM_SuppressedByUser; 355 356 std::optional<int> Count = 357 getOptionalIntLoopAttribute(L, "llvm.loop.unroll.count"); 358 if (Count) 359 return *Count == 1 ? TM_SuppressedByUser : TM_ForcedByUser; 360 361 if (getBooleanLoopAttribute(L, "llvm.loop.unroll.enable")) 362 return TM_ForcedByUser; 363 364 if (getBooleanLoopAttribute(L, "llvm.loop.unroll.full")) 365 return TM_ForcedByUser; 366 367 if (hasDisableAllTransformsHint(L)) 368 return TM_Disable; 369 370 return TM_Unspecified; 371 } 372 373 TransformationMode llvm::hasUnrollAndJamTransformation(const Loop *L) { 374 if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.disable")) 375 return TM_SuppressedByUser; 376 377 std::optional<int> Count = 378 getOptionalIntLoopAttribute(L, "llvm.loop.unroll_and_jam.count"); 379 if (Count) 380 return *Count == 1 ? TM_SuppressedByUser : TM_ForcedByUser; 381 382 if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.enable")) 383 return TM_ForcedByUser; 384 385 if (hasDisableAllTransformsHint(L)) 386 return TM_Disable; 387 388 return TM_Unspecified; 389 } 390 391 TransformationMode llvm::hasVectorizeTransformation(const Loop *L) { 392 std::optional<bool> Enable = 393 getOptionalBoolLoopAttribute(L, "llvm.loop.vectorize.enable"); 394 395 if (Enable == false) 396 return TM_SuppressedByUser; 397 398 std::optional<ElementCount> VectorizeWidth = 399 getOptionalElementCountLoopAttribute(L); 400 std::optional<int> InterleaveCount = 401 getOptionalIntLoopAttribute(L, "llvm.loop.interleave.count"); 402 403 // 'Forcing' vector width and interleave count to one effectively disables 404 // this tranformation. 405 if (Enable == true && VectorizeWidth && VectorizeWidth->isScalar() && 406 InterleaveCount == 1) 407 return TM_SuppressedByUser; 408 409 if (getBooleanLoopAttribute(L, "llvm.loop.isvectorized")) 410 return TM_Disable; 411 412 if (Enable == true) 413 return TM_ForcedByUser; 414 415 if ((VectorizeWidth && VectorizeWidth->isScalar()) && InterleaveCount == 1) 416 return TM_Disable; 417 418 if ((VectorizeWidth && VectorizeWidth->isVector()) || InterleaveCount > 1) 419 return TM_Enable; 420 421 if (hasDisableAllTransformsHint(L)) 422 return TM_Disable; 423 424 return TM_Unspecified; 425 } 426 427 TransformationMode llvm::hasDistributeTransformation(const Loop *L) { 428 if (getBooleanLoopAttribute(L, "llvm.loop.distribute.enable")) 429 return TM_ForcedByUser; 430 431 if (hasDisableAllTransformsHint(L)) 432 return TM_Disable; 433 434 return TM_Unspecified; 435 } 436 437 TransformationMode llvm::hasLICMVersioningTransformation(const Loop *L) { 438 if (getBooleanLoopAttribute(L, "llvm.loop.licm_versioning.disable")) 439 return TM_SuppressedByUser; 440 441 if (hasDisableAllTransformsHint(L)) 442 return TM_Disable; 443 444 return TM_Unspecified; 445 } 446 447 /// Does a BFS from a given node to all of its children inside a given loop. 448 /// The returned vector of nodes includes the starting point. 449 SmallVector<DomTreeNode *, 16> 450 llvm::collectChildrenInLoop(DomTreeNode *N, const Loop *CurLoop) { 451 SmallVector<DomTreeNode *, 16> Worklist; 452 auto AddRegionToWorklist = [&](DomTreeNode *DTN) { 453 // Only include subregions in the top level loop. 454 BasicBlock *BB = DTN->getBlock(); 455 if (CurLoop->contains(BB)) 456 Worklist.push_back(DTN); 457 }; 458 459 AddRegionToWorklist(N); 460 461 for (size_t I = 0; I < Worklist.size(); I++) { 462 for (DomTreeNode *Child : Worklist[I]->children()) 463 AddRegionToWorklist(Child); 464 } 465 466 return Worklist; 467 } 468 469 void llvm::deleteDeadLoop(Loop *L, DominatorTree *DT, ScalarEvolution *SE, 470 LoopInfo *LI, MemorySSA *MSSA) { 471 assert((!DT || L->isLCSSAForm(*DT)) && "Expected LCSSA!"); 472 auto *Preheader = L->getLoopPreheader(); 473 assert(Preheader && "Preheader should exist!"); 474 475 std::unique_ptr<MemorySSAUpdater> MSSAU; 476 if (MSSA) 477 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); 478 479 // Now that we know the removal is safe, remove the loop by changing the 480 // branch from the preheader to go to the single exit block. 481 // 482 // Because we're deleting a large chunk of code at once, the sequence in which 483 // we remove things is very important to avoid invalidation issues. 484 485 // Tell ScalarEvolution that the loop is deleted. Do this before 486 // deleting the loop so that ScalarEvolution can look at the loop 487 // to determine what it needs to clean up. 488 if (SE) { 489 SE->forgetLoop(L); 490 SE->forgetBlockAndLoopDispositions(); 491 } 492 493 Instruction *OldTerm = Preheader->getTerminator(); 494 assert(!OldTerm->mayHaveSideEffects() && 495 "Preheader must end with a side-effect-free terminator"); 496 assert(OldTerm->getNumSuccessors() == 1 && 497 "Preheader must have a single successor"); 498 // Connect the preheader to the exit block. Keep the old edge to the header 499 // around to perform the dominator tree update in two separate steps 500 // -- #1 insertion of the edge preheader -> exit and #2 deletion of the edge 501 // preheader -> header. 502 // 503 // 504 // 0. Preheader 1. Preheader 2. Preheader 505 // | | | | 506 // V | V | 507 // Header <--\ | Header <--\ | Header <--\ 508 // | | | | | | | | | | | 509 // | V | | | V | | | V | 510 // | Body --/ | | Body --/ | | Body --/ 511 // V V V V V 512 // Exit Exit Exit 513 // 514 // By doing this is two separate steps we can perform the dominator tree 515 // update without using the batch update API. 516 // 517 // Even when the loop is never executed, we cannot remove the edge from the 518 // source block to the exit block. Consider the case where the unexecuted loop 519 // branches back to an outer loop. If we deleted the loop and removed the edge 520 // coming to this inner loop, this will break the outer loop structure (by 521 // deleting the backedge of the outer loop). If the outer loop is indeed a 522 // non-loop, it will be deleted in a future iteration of loop deletion pass. 523 IRBuilder<> Builder(OldTerm); 524 525 auto *ExitBlock = L->getUniqueExitBlock(); 526 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 527 if (ExitBlock) { 528 assert(ExitBlock && "Should have a unique exit block!"); 529 assert(L->hasDedicatedExits() && "Loop should have dedicated exits!"); 530 531 Builder.CreateCondBr(Builder.getFalse(), L->getHeader(), ExitBlock); 532 // Remove the old branch. The conditional branch becomes a new terminator. 533 OldTerm->eraseFromParent(); 534 535 // Rewrite phis in the exit block to get their inputs from the Preheader 536 // instead of the exiting block. 537 for (PHINode &P : ExitBlock->phis()) { 538 // Set the zero'th element of Phi to be from the preheader and remove all 539 // other incoming values. Given the loop has dedicated exits, all other 540 // incoming values must be from the exiting blocks. 541 int PredIndex = 0; 542 P.setIncomingBlock(PredIndex, Preheader); 543 // Removes all incoming values from all other exiting blocks (including 544 // duplicate values from an exiting block). 545 // Nuke all entries except the zero'th entry which is the preheader entry. 546 // NOTE! We need to remove Incoming Values in the reverse order as done 547 // below, to keep the indices valid for deletion (removeIncomingValues 548 // updates getNumIncomingValues and shifts all values down into the 549 // operand being deleted). 550 for (unsigned i = 0, e = P.getNumIncomingValues() - 1; i != e; ++i) 551 P.removeIncomingValue(e - i, false); 552 553 assert((P.getNumIncomingValues() == 1 && 554 P.getIncomingBlock(PredIndex) == Preheader) && 555 "Should have exactly one value and that's from the preheader!"); 556 } 557 558 if (DT) { 559 DTU.applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}}); 560 if (MSSA) { 561 MSSAU->applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}}, 562 *DT); 563 if (VerifyMemorySSA) 564 MSSA->verifyMemorySSA(); 565 } 566 } 567 568 // Disconnect the loop body by branching directly to its exit. 569 Builder.SetInsertPoint(Preheader->getTerminator()); 570 Builder.CreateBr(ExitBlock); 571 // Remove the old branch. 572 Preheader->getTerminator()->eraseFromParent(); 573 } else { 574 assert(L->hasNoExitBlocks() && 575 "Loop should have either zero or one exit blocks."); 576 577 Builder.SetInsertPoint(OldTerm); 578 Builder.CreateUnreachable(); 579 Preheader->getTerminator()->eraseFromParent(); 580 } 581 582 if (DT) { 583 DTU.applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}}); 584 if (MSSA) { 585 MSSAU->applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}}, 586 *DT); 587 SmallSetVector<BasicBlock *, 8> DeadBlockSet(L->block_begin(), 588 L->block_end()); 589 MSSAU->removeBlocks(DeadBlockSet); 590 if (VerifyMemorySSA) 591 MSSA->verifyMemorySSA(); 592 } 593 } 594 595 // Use a map to unique and a vector to guarantee deterministic ordering. 596 llvm::SmallDenseSet<DebugVariable, 4> DeadDebugSet; 597 llvm::SmallVector<DbgVariableIntrinsic *, 4> DeadDebugInst; 598 599 if (ExitBlock) { 600 // Given LCSSA form is satisfied, we should not have users of instructions 601 // within the dead loop outside of the loop. However, LCSSA doesn't take 602 // unreachable uses into account. We handle them here. 603 // We could do it after drop all references (in this case all users in the 604 // loop will be already eliminated and we have less work to do but according 605 // to API doc of User::dropAllReferences only valid operation after dropping 606 // references, is deletion. So let's substitute all usages of 607 // instruction from the loop with poison value of corresponding type first. 608 for (auto *Block : L->blocks()) 609 for (Instruction &I : *Block) { 610 auto *Poison = PoisonValue::get(I.getType()); 611 for (Use &U : llvm::make_early_inc_range(I.uses())) { 612 if (auto *Usr = dyn_cast<Instruction>(U.getUser())) 613 if (L->contains(Usr->getParent())) 614 continue; 615 // If we have a DT then we can check that uses outside a loop only in 616 // unreachable block. 617 if (DT) 618 assert(!DT->isReachableFromEntry(U) && 619 "Unexpected user in reachable block"); 620 U.set(Poison); 621 } 622 auto *DVI = dyn_cast<DbgVariableIntrinsic>(&I); 623 if (!DVI) 624 continue; 625 if (!DeadDebugSet.insert(DebugVariable(DVI)).second) 626 continue; 627 DeadDebugInst.push_back(DVI); 628 } 629 630 // After the loop has been deleted all the values defined and modified 631 // inside the loop are going to be unavailable. 632 // Since debug values in the loop have been deleted, inserting an undef 633 // dbg.value truncates the range of any dbg.value before the loop where the 634 // loop used to be. This is particularly important for constant values. 635 Instruction *InsertDbgValueBefore = ExitBlock->getFirstNonPHI(); 636 assert(InsertDbgValueBefore && 637 "There should be a non-PHI instruction in exit block, else these " 638 "instructions will have no parent."); 639 for (auto *DVI : DeadDebugInst) { 640 DVI->setKillLocation(); 641 DVI->moveBefore(InsertDbgValueBefore); 642 } 643 } 644 645 // Remove the block from the reference counting scheme, so that we can 646 // delete it freely later. 647 for (auto *Block : L->blocks()) 648 Block->dropAllReferences(); 649 650 if (MSSA && VerifyMemorySSA) 651 MSSA->verifyMemorySSA(); 652 653 if (LI) { 654 // Erase the instructions and the blocks without having to worry 655 // about ordering because we already dropped the references. 656 // NOTE: This iteration is safe because erasing the block does not remove 657 // its entry from the loop's block list. We do that in the next section. 658 for (BasicBlock *BB : L->blocks()) 659 BB->eraseFromParent(); 660 661 // Finally, the blocks from loopinfo. This has to happen late because 662 // otherwise our loop iterators won't work. 663 664 SmallPtrSet<BasicBlock *, 8> blocks; 665 blocks.insert(L->block_begin(), L->block_end()); 666 for (BasicBlock *BB : blocks) 667 LI->removeBlock(BB); 668 669 // The last step is to update LoopInfo now that we've eliminated this loop. 670 // Note: LoopInfo::erase remove the given loop and relink its subloops with 671 // its parent. While removeLoop/removeChildLoop remove the given loop but 672 // not relink its subloops, which is what we want. 673 if (Loop *ParentLoop = L->getParentLoop()) { 674 Loop::iterator I = find(*ParentLoop, L); 675 assert(I != ParentLoop->end() && "Couldn't find loop"); 676 ParentLoop->removeChildLoop(I); 677 } else { 678 Loop::iterator I = find(*LI, L); 679 assert(I != LI->end() && "Couldn't find loop"); 680 LI->removeLoop(I); 681 } 682 LI->destroy(L); 683 } 684 } 685 686 void llvm::breakLoopBackedge(Loop *L, DominatorTree &DT, ScalarEvolution &SE, 687 LoopInfo &LI, MemorySSA *MSSA) { 688 auto *Latch = L->getLoopLatch(); 689 assert(Latch && "multiple latches not yet supported"); 690 auto *Header = L->getHeader(); 691 Loop *OutermostLoop = L->getOutermostLoop(); 692 693 SE.forgetLoop(L); 694 SE.forgetBlockAndLoopDispositions(); 695 696 std::unique_ptr<MemorySSAUpdater> MSSAU; 697 if (MSSA) 698 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); 699 700 // Update the CFG and domtree. We chose to special case a couple of 701 // of common cases for code quality and test readability reasons. 702 [&]() -> void { 703 if (auto *BI = dyn_cast<BranchInst>(Latch->getTerminator())) { 704 if (!BI->isConditional()) { 705 DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager); 706 (void)changeToUnreachable(BI, /*PreserveLCSSA*/ true, &DTU, 707 MSSAU.get()); 708 return; 709 } 710 711 // Conditional latch/exit - note that latch can be shared by inner 712 // and outer loop so the other target doesn't need to an exit 713 if (L->isLoopExiting(Latch)) { 714 // TODO: Generalize ConstantFoldTerminator so that it can be used 715 // here without invalidating LCSSA or MemorySSA. (Tricky case for 716 // LCSSA: header is an exit block of a preceeding sibling loop w/o 717 // dedicated exits.) 718 const unsigned ExitIdx = L->contains(BI->getSuccessor(0)) ? 1 : 0; 719 BasicBlock *ExitBB = BI->getSuccessor(ExitIdx); 720 721 DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager); 722 Header->removePredecessor(Latch, true); 723 724 IRBuilder<> Builder(BI); 725 auto *NewBI = Builder.CreateBr(ExitBB); 726 // Transfer the metadata to the new branch instruction (minus the 727 // loop info since this is no longer a loop) 728 NewBI->copyMetadata(*BI, {LLVMContext::MD_dbg, 729 LLVMContext::MD_annotation}); 730 731 BI->eraseFromParent(); 732 DTU.applyUpdates({{DominatorTree::Delete, Latch, Header}}); 733 if (MSSA) 734 MSSAU->applyUpdates({{DominatorTree::Delete, Latch, Header}}, DT); 735 return; 736 } 737 } 738 739 // General case. By splitting the backedge, and then explicitly making it 740 // unreachable we gracefully handle corner cases such as switch and invoke 741 // termiantors. 742 auto *BackedgeBB = SplitEdge(Latch, Header, &DT, &LI, MSSAU.get()); 743 744 DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager); 745 (void)changeToUnreachable(BackedgeBB->getTerminator(), 746 /*PreserveLCSSA*/ true, &DTU, MSSAU.get()); 747 }(); 748 749 // Erase (and destroy) this loop instance. Handles relinking sub-loops 750 // and blocks within the loop as needed. 751 LI.erase(L); 752 753 // If the loop we broke had a parent, then changeToUnreachable might have 754 // caused a block to be removed from the parent loop (see loop_nest_lcssa 755 // test case in zero-btc.ll for an example), thus changing the parent's 756 // exit blocks. If that happened, we need to rebuild LCSSA on the outermost 757 // loop which might have a had a block removed. 758 if (OutermostLoop != L) 759 formLCSSARecursively(*OutermostLoop, DT, &LI, &SE); 760 } 761 762 763 /// Checks if \p L has an exiting latch branch. There may also be other 764 /// exiting blocks. Returns branch instruction terminating the loop 765 /// latch if above check is successful, nullptr otherwise. 766 static BranchInst *getExpectedExitLoopLatchBranch(Loop *L) { 767 BasicBlock *Latch = L->getLoopLatch(); 768 if (!Latch) 769 return nullptr; 770 771 BranchInst *LatchBR = dyn_cast<BranchInst>(Latch->getTerminator()); 772 if (!LatchBR || LatchBR->getNumSuccessors() != 2 || !L->isLoopExiting(Latch)) 773 return nullptr; 774 775 assert((LatchBR->getSuccessor(0) == L->getHeader() || 776 LatchBR->getSuccessor(1) == L->getHeader()) && 777 "At least one edge out of the latch must go to the header"); 778 779 return LatchBR; 780 } 781 782 /// Return the estimated trip count for any exiting branch which dominates 783 /// the loop latch. 784 static std::optional<uint64_t> getEstimatedTripCount(BranchInst *ExitingBranch, 785 Loop *L, 786 uint64_t &OrigExitWeight) { 787 // To estimate the number of times the loop body was executed, we want to 788 // know the number of times the backedge was taken, vs. the number of times 789 // we exited the loop. 790 uint64_t LoopWeight, ExitWeight; 791 if (!extractBranchWeights(*ExitingBranch, LoopWeight, ExitWeight)) 792 return std::nullopt; 793 794 if (L->contains(ExitingBranch->getSuccessor(1))) 795 std::swap(LoopWeight, ExitWeight); 796 797 if (!ExitWeight) 798 // Don't have a way to return predicated infinite 799 return std::nullopt; 800 801 OrigExitWeight = ExitWeight; 802 803 // Estimated exit count is a ratio of the loop weight by the weight of the 804 // edge exiting the loop, rounded to nearest. 805 uint64_t ExitCount = llvm::divideNearest(LoopWeight, ExitWeight); 806 // Estimated trip count is one plus estimated exit count. 807 return ExitCount + 1; 808 } 809 810 std::optional<unsigned> 811 llvm::getLoopEstimatedTripCount(Loop *L, 812 unsigned *EstimatedLoopInvocationWeight) { 813 // Currently we take the estimate exit count only from the loop latch, 814 // ignoring other exiting blocks. This can overestimate the trip count 815 // if we exit through another exit, but can never underestimate it. 816 // TODO: incorporate information from other exits 817 if (BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L)) { 818 uint64_t ExitWeight; 819 if (std::optional<uint64_t> EstTripCount = 820 getEstimatedTripCount(LatchBranch, L, ExitWeight)) { 821 if (EstimatedLoopInvocationWeight) 822 *EstimatedLoopInvocationWeight = ExitWeight; 823 return *EstTripCount; 824 } 825 } 826 return std::nullopt; 827 } 828 829 bool llvm::setLoopEstimatedTripCount(Loop *L, unsigned EstimatedTripCount, 830 unsigned EstimatedloopInvocationWeight) { 831 // At the moment, we currently support changing the estimate trip count of 832 // the latch branch only. We could extend this API to manipulate estimated 833 // trip counts for any exit. 834 BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L); 835 if (!LatchBranch) 836 return false; 837 838 // Calculate taken and exit weights. 839 unsigned LatchExitWeight = 0; 840 unsigned BackedgeTakenWeight = 0; 841 842 if (EstimatedTripCount > 0) { 843 LatchExitWeight = EstimatedloopInvocationWeight; 844 BackedgeTakenWeight = (EstimatedTripCount - 1) * LatchExitWeight; 845 } 846 847 // Make a swap if back edge is taken when condition is "false". 848 if (LatchBranch->getSuccessor(0) != L->getHeader()) 849 std::swap(BackedgeTakenWeight, LatchExitWeight); 850 851 MDBuilder MDB(LatchBranch->getContext()); 852 853 // Set/Update profile metadata. 854 LatchBranch->setMetadata( 855 LLVMContext::MD_prof, 856 MDB.createBranchWeights(BackedgeTakenWeight, LatchExitWeight)); 857 858 return true; 859 } 860 861 bool llvm::hasIterationCountInvariantInParent(Loop *InnerLoop, 862 ScalarEvolution &SE) { 863 Loop *OuterL = InnerLoop->getParentLoop(); 864 if (!OuterL) 865 return true; 866 867 // Get the backedge taken count for the inner loop 868 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); 869 const SCEV *InnerLoopBECountSC = SE.getExitCount(InnerLoop, InnerLoopLatch); 870 if (isa<SCEVCouldNotCompute>(InnerLoopBECountSC) || 871 !InnerLoopBECountSC->getType()->isIntegerTy()) 872 return false; 873 874 // Get whether count is invariant to the outer loop 875 ScalarEvolution::LoopDisposition LD = 876 SE.getLoopDisposition(InnerLoopBECountSC, OuterL); 877 if (LD != ScalarEvolution::LoopInvariant) 878 return false; 879 880 return true; 881 } 882 883 CmpInst::Predicate llvm::getMinMaxReductionPredicate(RecurKind RK) { 884 switch (RK) { 885 default: 886 llvm_unreachable("Unknown min/max recurrence kind"); 887 case RecurKind::UMin: 888 return CmpInst::ICMP_ULT; 889 case RecurKind::UMax: 890 return CmpInst::ICMP_UGT; 891 case RecurKind::SMin: 892 return CmpInst::ICMP_SLT; 893 case RecurKind::SMax: 894 return CmpInst::ICMP_SGT; 895 case RecurKind::FMin: 896 return CmpInst::FCMP_OLT; 897 case RecurKind::FMax: 898 return CmpInst::FCMP_OGT; 899 } 900 } 901 902 Value *llvm::createSelectCmpOp(IRBuilderBase &Builder, Value *StartVal, 903 RecurKind RK, Value *Left, Value *Right) { 904 if (auto VTy = dyn_cast<VectorType>(Left->getType())) 905 StartVal = Builder.CreateVectorSplat(VTy->getElementCount(), StartVal); 906 Value *Cmp = 907 Builder.CreateCmp(CmpInst::ICMP_NE, Left, StartVal, "rdx.select.cmp"); 908 return Builder.CreateSelect(Cmp, Left, Right, "rdx.select"); 909 } 910 911 Value *llvm::createMinMaxOp(IRBuilderBase &Builder, RecurKind RK, Value *Left, 912 Value *Right) { 913 CmpInst::Predicate Pred = getMinMaxReductionPredicate(RK); 914 Value *Cmp = Builder.CreateCmp(Pred, Left, Right, "rdx.minmax.cmp"); 915 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select"); 916 return Select; 917 } 918 919 // Helper to generate an ordered reduction. 920 Value *llvm::getOrderedReduction(IRBuilderBase &Builder, Value *Acc, Value *Src, 921 unsigned Op, RecurKind RdxKind) { 922 unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements(); 923 924 // Extract and apply reduction ops in ascending order: 925 // e.g. ((((Acc + Scl[0]) + Scl[1]) + Scl[2]) + ) ... + Scl[VF-1] 926 Value *Result = Acc; 927 for (unsigned ExtractIdx = 0; ExtractIdx != VF; ++ExtractIdx) { 928 Value *Ext = 929 Builder.CreateExtractElement(Src, Builder.getInt32(ExtractIdx)); 930 931 if (Op != Instruction::ICmp && Op != Instruction::FCmp) { 932 Result = Builder.CreateBinOp((Instruction::BinaryOps)Op, Result, Ext, 933 "bin.rdx"); 934 } else { 935 assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) && 936 "Invalid min/max"); 937 Result = createMinMaxOp(Builder, RdxKind, Result, Ext); 938 } 939 } 940 941 return Result; 942 } 943 944 // Helper to generate a log2 shuffle reduction. 945 Value *llvm::getShuffleReduction(IRBuilderBase &Builder, Value *Src, 946 unsigned Op, RecurKind RdxKind) { 947 unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements(); 948 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 949 // and vector ops, reducing the set of values being computed by half each 950 // round. 951 assert(isPowerOf2_32(VF) && 952 "Reduction emission only supported for pow2 vectors!"); 953 // Note: fast-math-flags flags are controlled by the builder configuration 954 // and are assumed to apply to all generated arithmetic instructions. Other 955 // poison generating flags (nsw/nuw/inbounds/inrange/exact) are not part 956 // of the builder configuration, and since they're not passed explicitly, 957 // will never be relevant here. Note that it would be generally unsound to 958 // propagate these from an intrinsic call to the expansion anyways as we/ 959 // change the order of operations. 960 Value *TmpVec = Src; 961 SmallVector<int, 32> ShuffleMask(VF); 962 for (unsigned i = VF; i != 1; i >>= 1) { 963 // Move the upper half of the vector to the lower half. 964 for (unsigned j = 0; j != i / 2; ++j) 965 ShuffleMask[j] = i / 2 + j; 966 967 // Fill the rest of the mask with undef. 968 std::fill(&ShuffleMask[i / 2], ShuffleMask.end(), -1); 969 970 Value *Shuf = Builder.CreateShuffleVector(TmpVec, ShuffleMask, "rdx.shuf"); 971 972 if (Op != Instruction::ICmp && Op != Instruction::FCmp) { 973 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf, 974 "bin.rdx"); 975 } else { 976 assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) && 977 "Invalid min/max"); 978 TmpVec = createMinMaxOp(Builder, RdxKind, TmpVec, Shuf); 979 } 980 } 981 // The result is in the first element of the vector. 982 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); 983 } 984 985 Value *llvm::createSelectCmpTargetReduction(IRBuilderBase &Builder, 986 const TargetTransformInfo *TTI, 987 Value *Src, 988 const RecurrenceDescriptor &Desc, 989 PHINode *OrigPhi) { 990 assert(RecurrenceDescriptor::isSelectCmpRecurrenceKind( 991 Desc.getRecurrenceKind()) && 992 "Unexpected reduction kind"); 993 Value *InitVal = Desc.getRecurrenceStartValue(); 994 Value *NewVal = nullptr; 995 996 // First use the original phi to determine the new value we're trying to 997 // select from in the loop. 998 SelectInst *SI = nullptr; 999 for (auto *U : OrigPhi->users()) { 1000 if ((SI = dyn_cast<SelectInst>(U))) 1001 break; 1002 } 1003 assert(SI && "One user of the original phi should be a select"); 1004 1005 if (SI->getTrueValue() == OrigPhi) 1006 NewVal = SI->getFalseValue(); 1007 else { 1008 assert(SI->getFalseValue() == OrigPhi && 1009 "At least one input to the select should be the original Phi"); 1010 NewVal = SI->getTrueValue(); 1011 } 1012 1013 // Create a splat vector with the new value and compare this to the vector 1014 // we want to reduce. 1015 ElementCount EC = cast<VectorType>(Src->getType())->getElementCount(); 1016 Value *Right = Builder.CreateVectorSplat(EC, InitVal); 1017 Value *Cmp = 1018 Builder.CreateCmp(CmpInst::ICMP_NE, Src, Right, "rdx.select.cmp"); 1019 1020 // If any predicate is true it means that we want to select the new value. 1021 Cmp = Builder.CreateOrReduce(Cmp); 1022 return Builder.CreateSelect(Cmp, NewVal, InitVal, "rdx.select"); 1023 } 1024 1025 Value *llvm::createSimpleTargetReduction(IRBuilderBase &Builder, 1026 const TargetTransformInfo *TTI, 1027 Value *Src, RecurKind RdxKind) { 1028 auto *SrcVecEltTy = cast<VectorType>(Src->getType())->getElementType(); 1029 switch (RdxKind) { 1030 case RecurKind::Add: 1031 return Builder.CreateAddReduce(Src); 1032 case RecurKind::Mul: 1033 return Builder.CreateMulReduce(Src); 1034 case RecurKind::And: 1035 return Builder.CreateAndReduce(Src); 1036 case RecurKind::Or: 1037 return Builder.CreateOrReduce(Src); 1038 case RecurKind::Xor: 1039 return Builder.CreateXorReduce(Src); 1040 case RecurKind::FMulAdd: 1041 case RecurKind::FAdd: 1042 return Builder.CreateFAddReduce(ConstantFP::getNegativeZero(SrcVecEltTy), 1043 Src); 1044 case RecurKind::FMul: 1045 return Builder.CreateFMulReduce(ConstantFP::get(SrcVecEltTy, 1.0), Src); 1046 case RecurKind::SMax: 1047 return Builder.CreateIntMaxReduce(Src, true); 1048 case RecurKind::SMin: 1049 return Builder.CreateIntMinReduce(Src, true); 1050 case RecurKind::UMax: 1051 return Builder.CreateIntMaxReduce(Src, false); 1052 case RecurKind::UMin: 1053 return Builder.CreateIntMinReduce(Src, false); 1054 case RecurKind::FMax: 1055 return Builder.CreateFPMaxReduce(Src); 1056 case RecurKind::FMin: 1057 return Builder.CreateFPMinReduce(Src); 1058 default: 1059 llvm_unreachable("Unhandled opcode"); 1060 } 1061 } 1062 1063 Value *llvm::createTargetReduction(IRBuilderBase &B, 1064 const TargetTransformInfo *TTI, 1065 const RecurrenceDescriptor &Desc, Value *Src, 1066 PHINode *OrigPhi) { 1067 // TODO: Support in-order reductions based on the recurrence descriptor. 1068 // All ops in the reduction inherit fast-math-flags from the recurrence 1069 // descriptor. 1070 IRBuilderBase::FastMathFlagGuard FMFGuard(B); 1071 B.setFastMathFlags(Desc.getFastMathFlags()); 1072 1073 RecurKind RK = Desc.getRecurrenceKind(); 1074 if (RecurrenceDescriptor::isSelectCmpRecurrenceKind(RK)) 1075 return createSelectCmpTargetReduction(B, TTI, Src, Desc, OrigPhi); 1076 1077 return createSimpleTargetReduction(B, TTI, Src, RK); 1078 } 1079 1080 Value *llvm::createOrderedReduction(IRBuilderBase &B, 1081 const RecurrenceDescriptor &Desc, 1082 Value *Src, Value *Start) { 1083 assert((Desc.getRecurrenceKind() == RecurKind::FAdd || 1084 Desc.getRecurrenceKind() == RecurKind::FMulAdd) && 1085 "Unexpected reduction kind"); 1086 assert(Src->getType()->isVectorTy() && "Expected a vector type"); 1087 assert(!Start->getType()->isVectorTy() && "Expected a scalar type"); 1088 1089 return B.CreateFAddReduce(Start, Src); 1090 } 1091 1092 void llvm::propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue, 1093 bool IncludeWrapFlags) { 1094 auto *VecOp = dyn_cast<Instruction>(I); 1095 if (!VecOp) 1096 return; 1097 auto *Intersection = (OpValue == nullptr) ? dyn_cast<Instruction>(VL[0]) 1098 : dyn_cast<Instruction>(OpValue); 1099 if (!Intersection) 1100 return; 1101 const unsigned Opcode = Intersection->getOpcode(); 1102 VecOp->copyIRFlags(Intersection, IncludeWrapFlags); 1103 for (auto *V : VL) { 1104 auto *Instr = dyn_cast<Instruction>(V); 1105 if (!Instr) 1106 continue; 1107 if (OpValue == nullptr || Opcode == Instr->getOpcode()) 1108 VecOp->andIRFlags(V); 1109 } 1110 } 1111 1112 bool llvm::isKnownNegativeInLoop(const SCEV *S, const Loop *L, 1113 ScalarEvolution &SE) { 1114 const SCEV *Zero = SE.getZero(S->getType()); 1115 return SE.isAvailableAtLoopEntry(S, L) && 1116 SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, S, Zero); 1117 } 1118 1119 bool llvm::isKnownNonNegativeInLoop(const SCEV *S, const Loop *L, 1120 ScalarEvolution &SE) { 1121 const SCEV *Zero = SE.getZero(S->getType()); 1122 return SE.isAvailableAtLoopEntry(S, L) && 1123 SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGE, S, Zero); 1124 } 1125 1126 bool llvm::cannotBeMinInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE, 1127 bool Signed) { 1128 unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth(); 1129 APInt Min = Signed ? APInt::getSignedMinValue(BitWidth) : 1130 APInt::getMinValue(BitWidth); 1131 auto Predicate = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1132 return SE.isAvailableAtLoopEntry(S, L) && 1133 SE.isLoopEntryGuardedByCond(L, Predicate, S, 1134 SE.getConstant(Min)); 1135 } 1136 1137 bool llvm::cannotBeMaxInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE, 1138 bool Signed) { 1139 unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth(); 1140 APInt Max = Signed ? APInt::getSignedMaxValue(BitWidth) : 1141 APInt::getMaxValue(BitWidth); 1142 auto Predicate = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1143 return SE.isAvailableAtLoopEntry(S, L) && 1144 SE.isLoopEntryGuardedByCond(L, Predicate, S, 1145 SE.getConstant(Max)); 1146 } 1147 1148 //===----------------------------------------------------------------------===// 1149 // rewriteLoopExitValues - Optimize IV users outside the loop. 1150 // As a side effect, reduces the amount of IV processing within the loop. 1151 //===----------------------------------------------------------------------===// 1152 1153 static bool hasHardUserWithinLoop(const Loop *L, const Instruction *I) { 1154 SmallPtrSet<const Instruction *, 8> Visited; 1155 SmallVector<const Instruction *, 8> WorkList; 1156 Visited.insert(I); 1157 WorkList.push_back(I); 1158 while (!WorkList.empty()) { 1159 const Instruction *Curr = WorkList.pop_back_val(); 1160 // This use is outside the loop, nothing to do. 1161 if (!L->contains(Curr)) 1162 continue; 1163 // Do we assume it is a "hard" use which will not be eliminated easily? 1164 if (Curr->mayHaveSideEffects()) 1165 return true; 1166 // Otherwise, add all its users to worklist. 1167 for (const auto *U : Curr->users()) { 1168 auto *UI = cast<Instruction>(U); 1169 if (Visited.insert(UI).second) 1170 WorkList.push_back(UI); 1171 } 1172 } 1173 return false; 1174 } 1175 1176 // Collect information about PHI nodes which can be transformed in 1177 // rewriteLoopExitValues. 1178 struct RewritePhi { 1179 PHINode *PN; // For which PHI node is this replacement? 1180 unsigned Ith; // For which incoming value? 1181 const SCEV *ExpansionSCEV; // The SCEV of the incoming value we are rewriting. 1182 Instruction *ExpansionPoint; // Where we'd like to expand that SCEV? 1183 bool HighCost; // Is this expansion a high-cost? 1184 1185 RewritePhi(PHINode *P, unsigned I, const SCEV *Val, Instruction *ExpansionPt, 1186 bool H) 1187 : PN(P), Ith(I), ExpansionSCEV(Val), ExpansionPoint(ExpansionPt), 1188 HighCost(H) {} 1189 }; 1190 1191 // Check whether it is possible to delete the loop after rewriting exit 1192 // value. If it is possible, ignore ReplaceExitValue and do rewriting 1193 // aggressively. 1194 static bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) { 1195 BasicBlock *Preheader = L->getLoopPreheader(); 1196 // If there is no preheader, the loop will not be deleted. 1197 if (!Preheader) 1198 return false; 1199 1200 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1. 1201 // We obviate multiple ExitingBlocks case for simplicity. 1202 // TODO: If we see testcase with multiple ExitingBlocks can be deleted 1203 // after exit value rewriting, we can enhance the logic here. 1204 SmallVector<BasicBlock *, 4> ExitingBlocks; 1205 L->getExitingBlocks(ExitingBlocks); 1206 SmallVector<BasicBlock *, 8> ExitBlocks; 1207 L->getUniqueExitBlocks(ExitBlocks); 1208 if (ExitBlocks.size() != 1 || ExitingBlocks.size() != 1) 1209 return false; 1210 1211 BasicBlock *ExitBlock = ExitBlocks[0]; 1212 BasicBlock::iterator BI = ExitBlock->begin(); 1213 while (PHINode *P = dyn_cast<PHINode>(BI)) { 1214 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]); 1215 1216 // If the Incoming value of P is found in RewritePhiSet, we know it 1217 // could be rewritten to use a loop invariant value in transformation 1218 // phase later. Skip it in the loop invariant check below. 1219 bool found = false; 1220 for (const RewritePhi &Phi : RewritePhiSet) { 1221 unsigned i = Phi.Ith; 1222 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) { 1223 found = true; 1224 break; 1225 } 1226 } 1227 1228 Instruction *I; 1229 if (!found && (I = dyn_cast<Instruction>(Incoming))) 1230 if (!L->hasLoopInvariantOperands(I)) 1231 return false; 1232 1233 ++BI; 1234 } 1235 1236 for (auto *BB : L->blocks()) 1237 if (llvm::any_of(*BB, [](Instruction &I) { 1238 return I.mayHaveSideEffects(); 1239 })) 1240 return false; 1241 1242 return true; 1243 } 1244 1245 /// Checks if it is safe to call InductionDescriptor::isInductionPHI for \p Phi, 1246 /// and returns true if this Phi is an induction phi in the loop. When 1247 /// isInductionPHI returns true, \p ID will be also be set by isInductionPHI. 1248 static bool checkIsIndPhi(PHINode *Phi, Loop *L, ScalarEvolution *SE, 1249 InductionDescriptor &ID) { 1250 if (!Phi) 1251 return false; 1252 if (!L->getLoopPreheader()) 1253 return false; 1254 if (Phi->getParent() != L->getHeader()) 1255 return false; 1256 return InductionDescriptor::isInductionPHI(Phi, L, SE, ID); 1257 } 1258 1259 int llvm::rewriteLoopExitValues(Loop *L, LoopInfo *LI, TargetLibraryInfo *TLI, 1260 ScalarEvolution *SE, 1261 const TargetTransformInfo *TTI, 1262 SCEVExpander &Rewriter, DominatorTree *DT, 1263 ReplaceExitVal ReplaceExitValue, 1264 SmallVector<WeakTrackingVH, 16> &DeadInsts) { 1265 // Check a pre-condition. 1266 assert(L->isRecursivelyLCSSAForm(*DT, *LI) && 1267 "Indvars did not preserve LCSSA!"); 1268 1269 SmallVector<BasicBlock*, 8> ExitBlocks; 1270 L->getUniqueExitBlocks(ExitBlocks); 1271 1272 SmallVector<RewritePhi, 8> RewritePhiSet; 1273 // Find all values that are computed inside the loop, but used outside of it. 1274 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan 1275 // the exit blocks of the loop to find them. 1276 for (BasicBlock *ExitBB : ExitBlocks) { 1277 // If there are no PHI nodes in this exit block, then no values defined 1278 // inside the loop are used on this path, skip it. 1279 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin()); 1280 if (!PN) continue; 1281 1282 unsigned NumPreds = PN->getNumIncomingValues(); 1283 1284 // Iterate over all of the PHI nodes. 1285 BasicBlock::iterator BBI = ExitBB->begin(); 1286 while ((PN = dyn_cast<PHINode>(BBI++))) { 1287 if (PN->use_empty()) 1288 continue; // dead use, don't replace it 1289 1290 if (!SE->isSCEVable(PN->getType())) 1291 continue; 1292 1293 // Iterate over all of the values in all the PHI nodes. 1294 for (unsigned i = 0; i != NumPreds; ++i) { 1295 // If the value being merged in is not integer or is not defined 1296 // in the loop, skip it. 1297 Value *InVal = PN->getIncomingValue(i); 1298 if (!isa<Instruction>(InVal)) 1299 continue; 1300 1301 // If this pred is for a subloop, not L itself, skip it. 1302 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L) 1303 continue; // The Block is in a subloop, skip it. 1304 1305 // Check that InVal is defined in the loop. 1306 Instruction *Inst = cast<Instruction>(InVal); 1307 if (!L->contains(Inst)) 1308 continue; 1309 1310 // Find exit values which are induction variables in the loop, and are 1311 // unused in the loop, with the only use being the exit block PhiNode, 1312 // and the induction variable update binary operator. 1313 // The exit value can be replaced with the final value when it is cheap 1314 // to do so. 1315 if (ReplaceExitValue == UnusedIndVarInLoop) { 1316 InductionDescriptor ID; 1317 PHINode *IndPhi = dyn_cast<PHINode>(Inst); 1318 if (IndPhi) { 1319 if (!checkIsIndPhi(IndPhi, L, SE, ID)) 1320 continue; 1321 // This is an induction PHI. Check that the only users are PHI 1322 // nodes, and induction variable update binary operators. 1323 if (llvm::any_of(Inst->users(), [&](User *U) { 1324 if (!isa<PHINode>(U) && !isa<BinaryOperator>(U)) 1325 return true; 1326 BinaryOperator *B = dyn_cast<BinaryOperator>(U); 1327 if (B && B != ID.getInductionBinOp()) 1328 return true; 1329 return false; 1330 })) 1331 continue; 1332 } else { 1333 // If it is not an induction phi, it must be an induction update 1334 // binary operator with an induction phi user. 1335 BinaryOperator *B = dyn_cast<BinaryOperator>(Inst); 1336 if (!B) 1337 continue; 1338 if (llvm::any_of(Inst->users(), [&](User *U) { 1339 PHINode *Phi = dyn_cast<PHINode>(U); 1340 if (Phi != PN && !checkIsIndPhi(Phi, L, SE, ID)) 1341 return true; 1342 return false; 1343 })) 1344 continue; 1345 if (B != ID.getInductionBinOp()) 1346 continue; 1347 } 1348 } 1349 1350 // Okay, this instruction has a user outside of the current loop 1351 // and varies predictably *inside* the loop. Evaluate the value it 1352 // contains when the loop exits, if possible. We prefer to start with 1353 // expressions which are true for all exits (so as to maximize 1354 // expression reuse by the SCEVExpander), but resort to per-exit 1355 // evaluation if that fails. 1356 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop()); 1357 if (isa<SCEVCouldNotCompute>(ExitValue) || 1358 !SE->isLoopInvariant(ExitValue, L) || 1359 !Rewriter.isSafeToExpand(ExitValue)) { 1360 // TODO: This should probably be sunk into SCEV in some way; maybe a 1361 // getSCEVForExit(SCEV*, L, ExitingBB)? It can be generalized for 1362 // most SCEV expressions and other recurrence types (e.g. shift 1363 // recurrences). Is there existing code we can reuse? 1364 const SCEV *ExitCount = SE->getExitCount(L, PN->getIncomingBlock(i)); 1365 if (isa<SCEVCouldNotCompute>(ExitCount)) 1366 continue; 1367 if (auto *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Inst))) 1368 if (AddRec->getLoop() == L) 1369 ExitValue = AddRec->evaluateAtIteration(ExitCount, *SE); 1370 if (isa<SCEVCouldNotCompute>(ExitValue) || 1371 !SE->isLoopInvariant(ExitValue, L) || 1372 !Rewriter.isSafeToExpand(ExitValue)) 1373 continue; 1374 } 1375 1376 // Computing the value outside of the loop brings no benefit if it is 1377 // definitely used inside the loop in a way which can not be optimized 1378 // away. Avoid doing so unless we know we have a value which computes 1379 // the ExitValue already. TODO: This should be merged into SCEV 1380 // expander to leverage its knowledge of existing expressions. 1381 if (ReplaceExitValue != AlwaysRepl && !isa<SCEVConstant>(ExitValue) && 1382 !isa<SCEVUnknown>(ExitValue) && hasHardUserWithinLoop(L, Inst)) 1383 continue; 1384 1385 // Check if expansions of this SCEV would count as being high cost. 1386 bool HighCost = Rewriter.isHighCostExpansion( 1387 ExitValue, L, SCEVCheapExpansionBudget, TTI, Inst); 1388 1389 // Note that we must not perform expansions until after 1390 // we query *all* the costs, because if we perform temporary expansion 1391 // inbetween, one that we might not intend to keep, said expansion 1392 // *may* affect cost calculation of the the next SCEV's we'll query, 1393 // and next SCEV may errneously get smaller cost. 1394 1395 // Collect all the candidate PHINodes to be rewritten. 1396 Instruction *InsertPt = 1397 (isa<PHINode>(Inst) || isa<LandingPadInst>(Inst)) ? 1398 &*Inst->getParent()->getFirstInsertionPt() : Inst; 1399 RewritePhiSet.emplace_back(PN, i, ExitValue, InsertPt, HighCost); 1400 } 1401 } 1402 } 1403 1404 // TODO: evaluate whether it is beneficial to change how we calculate 1405 // high-cost: if we have SCEV 'A' which we know we will expand, should we 1406 // calculate the cost of other SCEV's after expanding SCEV 'A', thus 1407 // potentially giving cost bonus to those other SCEV's? 1408 1409 bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet); 1410 int NumReplaced = 0; 1411 1412 // Transformation. 1413 for (const RewritePhi &Phi : RewritePhiSet) { 1414 PHINode *PN = Phi.PN; 1415 1416 // Only do the rewrite when the ExitValue can be expanded cheaply. 1417 // If LoopCanBeDel is true, rewrite exit value aggressively. 1418 if ((ReplaceExitValue == OnlyCheapRepl || 1419 ReplaceExitValue == UnusedIndVarInLoop) && 1420 !LoopCanBeDel && Phi.HighCost) 1421 continue; 1422 1423 Value *ExitVal = Rewriter.expandCodeFor( 1424 Phi.ExpansionSCEV, Phi.PN->getType(), Phi.ExpansionPoint); 1425 1426 LLVM_DEBUG(dbgs() << "rewriteLoopExitValues: AfterLoopVal = " << *ExitVal 1427 << '\n' 1428 << " LoopVal = " << *(Phi.ExpansionPoint) << "\n"); 1429 1430 #ifndef NDEBUG 1431 // If we reuse an instruction from a loop which is neither L nor one of 1432 // its containing loops, we end up breaking LCSSA form for this loop by 1433 // creating a new use of its instruction. 1434 if (auto *ExitInsn = dyn_cast<Instruction>(ExitVal)) 1435 if (auto *EVL = LI->getLoopFor(ExitInsn->getParent())) 1436 if (EVL != L) 1437 assert(EVL->contains(L) && "LCSSA breach detected!"); 1438 #endif 1439 1440 NumReplaced++; 1441 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith)); 1442 PN->setIncomingValue(Phi.Ith, ExitVal); 1443 // It's necessary to tell ScalarEvolution about this explicitly so that 1444 // it can walk the def-use list and forget all SCEVs, as it may not be 1445 // watching the PHI itself. Once the new exit value is in place, there 1446 // may not be a def-use connection between the loop and every instruction 1447 // which got a SCEVAddRecExpr for that loop. 1448 SE->forgetValue(PN); 1449 1450 // If this instruction is dead now, delete it. Don't do it now to avoid 1451 // invalidating iterators. 1452 if (isInstructionTriviallyDead(Inst, TLI)) 1453 DeadInsts.push_back(Inst); 1454 1455 // Replace PN with ExitVal if that is legal and does not break LCSSA. 1456 if (PN->getNumIncomingValues() == 1 && 1457 LI->replacementPreservesLCSSAForm(PN, ExitVal)) { 1458 PN->replaceAllUsesWith(ExitVal); 1459 PN->eraseFromParent(); 1460 } 1461 } 1462 1463 // The insertion point instruction may have been deleted; clear it out 1464 // so that the rewriter doesn't trip over it later. 1465 Rewriter.clearInsertPoint(); 1466 return NumReplaced; 1467 } 1468 1469 /// Set weights for \p UnrolledLoop and \p RemainderLoop based on weights for 1470 /// \p OrigLoop. 1471 void llvm::setProfileInfoAfterUnrolling(Loop *OrigLoop, Loop *UnrolledLoop, 1472 Loop *RemainderLoop, uint64_t UF) { 1473 assert(UF > 0 && "Zero unrolled factor is not supported"); 1474 assert(UnrolledLoop != RemainderLoop && 1475 "Unrolled and Remainder loops are expected to distinct"); 1476 1477 // Get number of iterations in the original scalar loop. 1478 unsigned OrigLoopInvocationWeight = 0; 1479 std::optional<unsigned> OrigAverageTripCount = 1480 getLoopEstimatedTripCount(OrigLoop, &OrigLoopInvocationWeight); 1481 if (!OrigAverageTripCount) 1482 return; 1483 1484 // Calculate number of iterations in unrolled loop. 1485 unsigned UnrolledAverageTripCount = *OrigAverageTripCount / UF; 1486 // Calculate number of iterations for remainder loop. 1487 unsigned RemainderAverageTripCount = *OrigAverageTripCount % UF; 1488 1489 setLoopEstimatedTripCount(UnrolledLoop, UnrolledAverageTripCount, 1490 OrigLoopInvocationWeight); 1491 setLoopEstimatedTripCount(RemainderLoop, RemainderAverageTripCount, 1492 OrigLoopInvocationWeight); 1493 } 1494 1495 /// Utility that implements appending of loops onto a worklist. 1496 /// Loops are added in preorder (analogous for reverse postorder for trees), 1497 /// and the worklist is processed LIFO. 1498 template <typename RangeT> 1499 void llvm::appendReversedLoopsToWorklist( 1500 RangeT &&Loops, SmallPriorityWorklist<Loop *, 4> &Worklist) { 1501 // We use an internal worklist to build up the preorder traversal without 1502 // recursion. 1503 SmallVector<Loop *, 4> PreOrderLoops, PreOrderWorklist; 1504 1505 // We walk the initial sequence of loops in reverse because we generally want 1506 // to visit defs before uses and the worklist is LIFO. 1507 for (Loop *RootL : Loops) { 1508 assert(PreOrderLoops.empty() && "Must start with an empty preorder walk."); 1509 assert(PreOrderWorklist.empty() && 1510 "Must start with an empty preorder walk worklist."); 1511 PreOrderWorklist.push_back(RootL); 1512 do { 1513 Loop *L = PreOrderWorklist.pop_back_val(); 1514 PreOrderWorklist.append(L->begin(), L->end()); 1515 PreOrderLoops.push_back(L); 1516 } while (!PreOrderWorklist.empty()); 1517 1518 Worklist.insert(std::move(PreOrderLoops)); 1519 PreOrderLoops.clear(); 1520 } 1521 } 1522 1523 template <typename RangeT> 1524 void llvm::appendLoopsToWorklist(RangeT &&Loops, 1525 SmallPriorityWorklist<Loop *, 4> &Worklist) { 1526 appendReversedLoopsToWorklist(reverse(Loops), Worklist); 1527 } 1528 1529 template void llvm::appendLoopsToWorklist<ArrayRef<Loop *> &>( 1530 ArrayRef<Loop *> &Loops, SmallPriorityWorklist<Loop *, 4> &Worklist); 1531 1532 template void 1533 llvm::appendLoopsToWorklist<Loop &>(Loop &L, 1534 SmallPriorityWorklist<Loop *, 4> &Worklist); 1535 1536 void llvm::appendLoopsToWorklist(LoopInfo &LI, 1537 SmallPriorityWorklist<Loop *, 4> &Worklist) { 1538 appendReversedLoopsToWorklist(LI, Worklist); 1539 } 1540 1541 Loop *llvm::cloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, 1542 LoopInfo *LI, LPPassManager *LPM) { 1543 Loop &New = *LI->AllocateLoop(); 1544 if (PL) 1545 PL->addChildLoop(&New); 1546 else 1547 LI->addTopLevelLoop(&New); 1548 1549 if (LPM) 1550 LPM->addLoop(New); 1551 1552 // Add all of the blocks in L to the new loop. 1553 for (BasicBlock *BB : L->blocks()) 1554 if (LI->getLoopFor(BB) == L) 1555 New.addBasicBlockToLoop(cast<BasicBlock>(VM[BB]), *LI); 1556 1557 // Add all of the subloops to the new loop. 1558 for (Loop *I : *L) 1559 cloneLoop(I, &New, VM, LI, LPM); 1560 1561 return &New; 1562 } 1563 1564 /// IR Values for the lower and upper bounds of a pointer evolution. We 1565 /// need to use value-handles because SCEV expansion can invalidate previously 1566 /// expanded values. Thus expansion of a pointer can invalidate the bounds for 1567 /// a previous one. 1568 struct PointerBounds { 1569 TrackingVH<Value> Start; 1570 TrackingVH<Value> End; 1571 }; 1572 1573 /// Expand code for the lower and upper bound of the pointer group \p CG 1574 /// in \p TheLoop. \return the values for the bounds. 1575 static PointerBounds expandBounds(const RuntimeCheckingPtrGroup *CG, 1576 Loop *TheLoop, Instruction *Loc, 1577 SCEVExpander &Exp) { 1578 LLVMContext &Ctx = Loc->getContext(); 1579 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, CG->AddressSpace); 1580 1581 Value *Start = nullptr, *End = nullptr; 1582 LLVM_DEBUG(dbgs() << "LAA: Adding RT check for range:\n"); 1583 Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc); 1584 End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc); 1585 if (CG->NeedsFreeze) { 1586 IRBuilder<> Builder(Loc); 1587 Start = Builder.CreateFreeze(Start, Start->getName() + ".fr"); 1588 End = Builder.CreateFreeze(End, End->getName() + ".fr"); 1589 } 1590 LLVM_DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n"); 1591 return {Start, End}; 1592 } 1593 1594 /// Turns a collection of checks into a collection of expanded upper and 1595 /// lower bounds for both pointers in the check. 1596 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> 1597 expandBounds(const SmallVectorImpl<RuntimePointerCheck> &PointerChecks, Loop *L, 1598 Instruction *Loc, SCEVExpander &Exp) { 1599 SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds; 1600 1601 // Here we're relying on the SCEV Expander's cache to only emit code for the 1602 // same bounds once. 1603 transform(PointerChecks, std::back_inserter(ChecksWithBounds), 1604 [&](const RuntimePointerCheck &Check) { 1605 PointerBounds First = expandBounds(Check.first, L, Loc, Exp), 1606 Second = expandBounds(Check.second, L, Loc, Exp); 1607 return std::make_pair(First, Second); 1608 }); 1609 1610 return ChecksWithBounds; 1611 } 1612 1613 Value *llvm::addRuntimeChecks( 1614 Instruction *Loc, Loop *TheLoop, 1615 const SmallVectorImpl<RuntimePointerCheck> &PointerChecks, 1616 SCEVExpander &Exp) { 1617 // TODO: Move noalias annotation code from LoopVersioning here and share with LV if possible. 1618 // TODO: Pass RtPtrChecking instead of PointerChecks and SE separately, if possible 1619 auto ExpandedChecks = expandBounds(PointerChecks, TheLoop, Loc, Exp); 1620 1621 LLVMContext &Ctx = Loc->getContext(); 1622 IRBuilder<InstSimplifyFolder> ChkBuilder(Ctx, 1623 Loc->getModule()->getDataLayout()); 1624 ChkBuilder.SetInsertPoint(Loc); 1625 // Our instructions might fold to a constant. 1626 Value *MemoryRuntimeCheck = nullptr; 1627 1628 for (const auto &Check : ExpandedChecks) { 1629 const PointerBounds &A = Check.first, &B = Check.second; 1630 // Check if two pointers (A and B) conflict where conflict is computed as: 1631 // start(A) <= end(B) && start(B) <= end(A) 1632 unsigned AS0 = A.Start->getType()->getPointerAddressSpace(); 1633 unsigned AS1 = B.Start->getType()->getPointerAddressSpace(); 1634 1635 assert((AS0 == B.End->getType()->getPointerAddressSpace()) && 1636 (AS1 == A.End->getType()->getPointerAddressSpace()) && 1637 "Trying to bounds check pointers with different address spaces"); 1638 1639 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0); 1640 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1); 1641 1642 Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc"); 1643 Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc"); 1644 Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc"); 1645 Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc"); 1646 1647 // [A|B].Start points to the first accessed byte under base [A|B]. 1648 // [A|B].End points to the last accessed byte, plus one. 1649 // There is no conflict when the intervals are disjoint: 1650 // NoConflict = (B.Start >= A.End) || (A.Start >= B.End) 1651 // 1652 // bound0 = (B.Start < A.End) 1653 // bound1 = (A.Start < B.End) 1654 // IsConflict = bound0 & bound1 1655 Value *Cmp0 = ChkBuilder.CreateICmpULT(Start0, End1, "bound0"); 1656 Value *Cmp1 = ChkBuilder.CreateICmpULT(Start1, End0, "bound1"); 1657 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict"); 1658 if (MemoryRuntimeCheck) { 1659 IsConflict = 1660 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx"); 1661 } 1662 MemoryRuntimeCheck = IsConflict; 1663 } 1664 1665 return MemoryRuntimeCheck; 1666 } 1667 1668 Value *llvm::addDiffRuntimeChecks( 1669 Instruction *Loc, ArrayRef<PointerDiffInfo> Checks, SCEVExpander &Expander, 1670 function_ref<Value *(IRBuilderBase &, unsigned)> GetVF, unsigned IC) { 1671 1672 LLVMContext &Ctx = Loc->getContext(); 1673 IRBuilder<InstSimplifyFolder> ChkBuilder(Ctx, 1674 Loc->getModule()->getDataLayout()); 1675 ChkBuilder.SetInsertPoint(Loc); 1676 // Our instructions might fold to a constant. 1677 Value *MemoryRuntimeCheck = nullptr; 1678 1679 for (const auto &C : Checks) { 1680 Type *Ty = C.SinkStart->getType(); 1681 // Compute VF * IC * AccessSize. 1682 auto *VFTimesUFTimesSize = 1683 ChkBuilder.CreateMul(GetVF(ChkBuilder, Ty->getScalarSizeInBits()), 1684 ConstantInt::get(Ty, IC * C.AccessSize)); 1685 Value *Sink = Expander.expandCodeFor(C.SinkStart, Ty, Loc); 1686 Value *Src = Expander.expandCodeFor(C.SrcStart, Ty, Loc); 1687 if (C.NeedsFreeze) { 1688 IRBuilder<> Builder(Loc); 1689 Sink = Builder.CreateFreeze(Sink, Sink->getName() + ".fr"); 1690 Src = Builder.CreateFreeze(Src, Src->getName() + ".fr"); 1691 } 1692 Value *Diff = ChkBuilder.CreateSub(Sink, Src); 1693 Value *IsConflict = 1694 ChkBuilder.CreateICmpULT(Diff, VFTimesUFTimesSize, "diff.check"); 1695 1696 if (MemoryRuntimeCheck) { 1697 IsConflict = 1698 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx"); 1699 } 1700 MemoryRuntimeCheck = IsConflict; 1701 } 1702 1703 return MemoryRuntimeCheck; 1704 } 1705 1706 std::optional<IVConditionInfo> 1707 llvm::hasPartialIVCondition(const Loop &L, unsigned MSSAThreshold, 1708 const MemorySSA &MSSA, AAResults &AA) { 1709 auto *TI = dyn_cast<BranchInst>(L.getHeader()->getTerminator()); 1710 if (!TI || !TI->isConditional()) 1711 return {}; 1712 1713 auto *CondI = dyn_cast<CmpInst>(TI->getCondition()); 1714 // The case with the condition outside the loop should already be handled 1715 // earlier. 1716 if (!CondI || !L.contains(CondI)) 1717 return {}; 1718 1719 SmallVector<Instruction *> InstToDuplicate; 1720 InstToDuplicate.push_back(CondI); 1721 1722 SmallVector<Value *, 4> WorkList; 1723 WorkList.append(CondI->op_begin(), CondI->op_end()); 1724 1725 SmallVector<MemoryAccess *, 4> AccessesToCheck; 1726 SmallVector<MemoryLocation, 4> AccessedLocs; 1727 while (!WorkList.empty()) { 1728 Instruction *I = dyn_cast<Instruction>(WorkList.pop_back_val()); 1729 if (!I || !L.contains(I)) 1730 continue; 1731 1732 // TODO: support additional instructions. 1733 if (!isa<LoadInst>(I) && !isa<GetElementPtrInst>(I)) 1734 return {}; 1735 1736 // Do not duplicate volatile and atomic loads. 1737 if (auto *LI = dyn_cast<LoadInst>(I)) 1738 if (LI->isVolatile() || LI->isAtomic()) 1739 return {}; 1740 1741 InstToDuplicate.push_back(I); 1742 if (MemoryAccess *MA = MSSA.getMemoryAccess(I)) { 1743 if (auto *MemUse = dyn_cast_or_null<MemoryUse>(MA)) { 1744 // Queue the defining access to check for alias checks. 1745 AccessesToCheck.push_back(MemUse->getDefiningAccess()); 1746 AccessedLocs.push_back(MemoryLocation::get(I)); 1747 } else { 1748 // MemoryDefs may clobber the location or may be atomic memory 1749 // operations. Bail out. 1750 return {}; 1751 } 1752 } 1753 WorkList.append(I->op_begin(), I->op_end()); 1754 } 1755 1756 if (InstToDuplicate.empty()) 1757 return {}; 1758 1759 SmallVector<BasicBlock *, 4> ExitingBlocks; 1760 L.getExitingBlocks(ExitingBlocks); 1761 auto HasNoClobbersOnPath = 1762 [&L, &AA, &AccessedLocs, &ExitingBlocks, &InstToDuplicate, 1763 MSSAThreshold](BasicBlock *Succ, BasicBlock *Header, 1764 SmallVector<MemoryAccess *, 4> AccessesToCheck) 1765 -> std::optional<IVConditionInfo> { 1766 IVConditionInfo Info; 1767 // First, collect all blocks in the loop that are on a patch from Succ 1768 // to the header. 1769 SmallVector<BasicBlock *, 4> WorkList; 1770 WorkList.push_back(Succ); 1771 WorkList.push_back(Header); 1772 SmallPtrSet<BasicBlock *, 4> Seen; 1773 Seen.insert(Header); 1774 Info.PathIsNoop &= 1775 all_of(*Header, [](Instruction &I) { return !I.mayHaveSideEffects(); }); 1776 1777 while (!WorkList.empty()) { 1778 BasicBlock *Current = WorkList.pop_back_val(); 1779 if (!L.contains(Current)) 1780 continue; 1781 const auto &SeenIns = Seen.insert(Current); 1782 if (!SeenIns.second) 1783 continue; 1784 1785 Info.PathIsNoop &= all_of( 1786 *Current, [](Instruction &I) { return !I.mayHaveSideEffects(); }); 1787 WorkList.append(succ_begin(Current), succ_end(Current)); 1788 } 1789 1790 // Require at least 2 blocks on a path through the loop. This skips 1791 // paths that directly exit the loop. 1792 if (Seen.size() < 2) 1793 return {}; 1794 1795 // Next, check if there are any MemoryDefs that are on the path through 1796 // the loop (in the Seen set) and they may-alias any of the locations in 1797 // AccessedLocs. If that is the case, they may modify the condition and 1798 // partial unswitching is not possible. 1799 SmallPtrSet<MemoryAccess *, 4> SeenAccesses; 1800 while (!AccessesToCheck.empty()) { 1801 MemoryAccess *Current = AccessesToCheck.pop_back_val(); 1802 auto SeenI = SeenAccesses.insert(Current); 1803 if (!SeenI.second || !Seen.contains(Current->getBlock())) 1804 continue; 1805 1806 // Bail out if exceeded the threshold. 1807 if (SeenAccesses.size() >= MSSAThreshold) 1808 return {}; 1809 1810 // MemoryUse are read-only accesses. 1811 if (isa<MemoryUse>(Current)) 1812 continue; 1813 1814 // For a MemoryDef, check if is aliases any of the location feeding 1815 // the original condition. 1816 if (auto *CurrentDef = dyn_cast<MemoryDef>(Current)) { 1817 if (any_of(AccessedLocs, [&AA, CurrentDef](MemoryLocation &Loc) { 1818 return isModSet( 1819 AA.getModRefInfo(CurrentDef->getMemoryInst(), Loc)); 1820 })) 1821 return {}; 1822 } 1823 1824 for (Use &U : Current->uses()) 1825 AccessesToCheck.push_back(cast<MemoryAccess>(U.getUser())); 1826 } 1827 1828 // We could also allow loops with known trip counts without mustprogress, 1829 // but ScalarEvolution may not be available. 1830 Info.PathIsNoop &= isMustProgress(&L); 1831 1832 // If the path is considered a no-op so far, check if it reaches a 1833 // single exit block without any phis. This ensures no values from the 1834 // loop are used outside of the loop. 1835 if (Info.PathIsNoop) { 1836 for (auto *Exiting : ExitingBlocks) { 1837 if (!Seen.contains(Exiting)) 1838 continue; 1839 for (auto *Succ : successors(Exiting)) { 1840 if (L.contains(Succ)) 1841 continue; 1842 1843 Info.PathIsNoop &= Succ->phis().empty() && 1844 (!Info.ExitForPath || Info.ExitForPath == Succ); 1845 if (!Info.PathIsNoop) 1846 break; 1847 assert((!Info.ExitForPath || Info.ExitForPath == Succ) && 1848 "cannot have multiple exit blocks"); 1849 Info.ExitForPath = Succ; 1850 } 1851 } 1852 } 1853 if (!Info.ExitForPath) 1854 Info.PathIsNoop = false; 1855 1856 Info.InstToDuplicate = InstToDuplicate; 1857 return Info; 1858 }; 1859 1860 // If we branch to the same successor, partial unswitching will not be 1861 // beneficial. 1862 if (TI->getSuccessor(0) == TI->getSuccessor(1)) 1863 return {}; 1864 1865 if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(0), L.getHeader(), 1866 AccessesToCheck)) { 1867 Info->KnownValue = ConstantInt::getTrue(TI->getContext()); 1868 return Info; 1869 } 1870 if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(1), L.getHeader(), 1871 AccessesToCheck)) { 1872 Info->KnownValue = ConstantInt::getFalse(TI->getContext()); 1873 return Info; 1874 } 1875 1876 return {}; 1877 } 1878