1 //===- CodeExtractor.cpp - Pull code region into a new function -----------===// 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 implements the interface to tear out a code region, such as an 10 // individual loop or a parallel section, into a new function, replacing it with 11 // a call to the new function. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/Utils/CodeExtractor.h" 16 #include "llvm/ADT/ArrayRef.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SetVector.h" 20 #include "llvm/ADT/SmallPtrSet.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/Analysis/AssumptionCache.h" 23 #include "llvm/Analysis/BlockFrequencyInfo.h" 24 #include "llvm/Analysis/BlockFrequencyInfoImpl.h" 25 #include "llvm/Analysis/BranchProbabilityInfo.h" 26 #include "llvm/Analysis/LoopInfo.h" 27 #include "llvm/IR/Argument.h" 28 #include "llvm/IR/Attributes.h" 29 #include "llvm/IR/BasicBlock.h" 30 #include "llvm/IR/CFG.h" 31 #include "llvm/IR/Constant.h" 32 #include "llvm/IR/Constants.h" 33 #include "llvm/IR/DIBuilder.h" 34 #include "llvm/IR/DataLayout.h" 35 #include "llvm/IR/DebugInfo.h" 36 #include "llvm/IR/DebugInfoMetadata.h" 37 #include "llvm/IR/DerivedTypes.h" 38 #include "llvm/IR/Dominators.h" 39 #include "llvm/IR/Function.h" 40 #include "llvm/IR/GlobalValue.h" 41 #include "llvm/IR/InstIterator.h" 42 #include "llvm/IR/InstrTypes.h" 43 #include "llvm/IR/Instruction.h" 44 #include "llvm/IR/Instructions.h" 45 #include "llvm/IR/IntrinsicInst.h" 46 #include "llvm/IR/Intrinsics.h" 47 #include "llvm/IR/LLVMContext.h" 48 #include "llvm/IR/MDBuilder.h" 49 #include "llvm/IR/Module.h" 50 #include "llvm/IR/PatternMatch.h" 51 #include "llvm/IR/Type.h" 52 #include "llvm/IR/User.h" 53 #include "llvm/IR/Value.h" 54 #include "llvm/IR/Verifier.h" 55 #include "llvm/Support/BlockFrequency.h" 56 #include "llvm/Support/BranchProbability.h" 57 #include "llvm/Support/Casting.h" 58 #include "llvm/Support/CommandLine.h" 59 #include "llvm/Support/Debug.h" 60 #include "llvm/Support/ErrorHandling.h" 61 #include "llvm/Support/raw_ostream.h" 62 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 63 #include <cassert> 64 #include <cstdint> 65 #include <iterator> 66 #include <map> 67 #include <utility> 68 #include <vector> 69 70 using namespace llvm; 71 using namespace llvm::PatternMatch; 72 using ProfileCount = Function::ProfileCount; 73 74 #define DEBUG_TYPE "code-extractor" 75 76 // Provide a command-line option to aggregate function arguments into a struct 77 // for functions produced by the code extractor. This is useful when converting 78 // extracted functions to pthread-based code, as only one argument (void*) can 79 // be passed in to pthread_create(). 80 static cl::opt<bool> 81 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden, 82 cl::desc("Aggregate arguments to code-extracted functions")); 83 84 /// Test whether a block is valid for extraction. 85 static bool isBlockValidForExtraction(const BasicBlock &BB, 86 const SetVector<BasicBlock *> &Result, 87 bool AllowVarArgs, bool AllowAlloca) { 88 // taking the address of a basic block moved to another function is illegal 89 if (BB.hasAddressTaken()) 90 return false; 91 92 // don't hoist code that uses another basicblock address, as it's likely to 93 // lead to unexpected behavior, like cross-function jumps 94 SmallPtrSet<User const *, 16> Visited; 95 SmallVector<User const *, 16> ToVisit; 96 97 for (Instruction const &Inst : BB) 98 ToVisit.push_back(&Inst); 99 100 while (!ToVisit.empty()) { 101 User const *Curr = ToVisit.pop_back_val(); 102 if (!Visited.insert(Curr).second) 103 continue; 104 if (isa<BlockAddress const>(Curr)) 105 return false; // even a reference to self is likely to be not compatible 106 107 if (isa<Instruction>(Curr) && cast<Instruction>(Curr)->getParent() != &BB) 108 continue; 109 110 for (auto const &U : Curr->operands()) { 111 if (auto *UU = dyn_cast<User>(U)) 112 ToVisit.push_back(UU); 113 } 114 } 115 116 // If explicitly requested, allow vastart and alloca. For invoke instructions 117 // verify that extraction is valid. 118 for (BasicBlock::const_iterator I = BB.begin(), E = BB.end(); I != E; ++I) { 119 if (isa<AllocaInst>(I)) { 120 if (!AllowAlloca) 121 return false; 122 continue; 123 } 124 125 if (const auto *II = dyn_cast<InvokeInst>(I)) { 126 // Unwind destination (either a landingpad, catchswitch, or cleanuppad) 127 // must be a part of the subgraph which is being extracted. 128 if (auto *UBB = II->getUnwindDest()) 129 if (!Result.count(UBB)) 130 return false; 131 continue; 132 } 133 134 // All catch handlers of a catchswitch instruction as well as the unwind 135 // destination must be in the subgraph. 136 if (const auto *CSI = dyn_cast<CatchSwitchInst>(I)) { 137 if (auto *UBB = CSI->getUnwindDest()) 138 if (!Result.count(UBB)) 139 return false; 140 for (const auto *HBB : CSI->handlers()) 141 if (!Result.count(const_cast<BasicBlock*>(HBB))) 142 return false; 143 continue; 144 } 145 146 // Make sure that entire catch handler is within subgraph. It is sufficient 147 // to check that catch return's block is in the list. 148 if (const auto *CPI = dyn_cast<CatchPadInst>(I)) { 149 for (const auto *U : CPI->users()) 150 if (const auto *CRI = dyn_cast<CatchReturnInst>(U)) 151 if (!Result.count(const_cast<BasicBlock*>(CRI->getParent()))) 152 return false; 153 continue; 154 } 155 156 // And do similar checks for cleanup handler - the entire handler must be 157 // in subgraph which is going to be extracted. For cleanup return should 158 // additionally check that the unwind destination is also in the subgraph. 159 if (const auto *CPI = dyn_cast<CleanupPadInst>(I)) { 160 for (const auto *U : CPI->users()) 161 if (const auto *CRI = dyn_cast<CleanupReturnInst>(U)) 162 if (!Result.count(const_cast<BasicBlock*>(CRI->getParent()))) 163 return false; 164 continue; 165 } 166 if (const auto *CRI = dyn_cast<CleanupReturnInst>(I)) { 167 if (auto *UBB = CRI->getUnwindDest()) 168 if (!Result.count(UBB)) 169 return false; 170 continue; 171 } 172 173 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 174 if (const Function *F = CI->getCalledFunction()) { 175 auto IID = F->getIntrinsicID(); 176 if (IID == Intrinsic::vastart) { 177 if (AllowVarArgs) 178 continue; 179 else 180 return false; 181 } 182 183 // Currently, we miscompile outlined copies of eh_typid_for. There are 184 // proposals for fixing this in llvm.org/PR39545. 185 if (IID == Intrinsic::eh_typeid_for) 186 return false; 187 } 188 } 189 } 190 191 return true; 192 } 193 194 /// Build a set of blocks to extract if the input blocks are viable. 195 static SetVector<BasicBlock *> 196 buildExtractionBlockSet(ArrayRef<BasicBlock *> BBs, DominatorTree *DT, 197 bool AllowVarArgs, bool AllowAlloca) { 198 assert(!BBs.empty() && "The set of blocks to extract must be non-empty"); 199 SetVector<BasicBlock *> Result; 200 201 // Loop over the blocks, adding them to our set-vector, and aborting with an 202 // empty set if we encounter invalid blocks. 203 for (BasicBlock *BB : BBs) { 204 // If this block is dead, don't process it. 205 if (DT && !DT->isReachableFromEntry(BB)) 206 continue; 207 208 if (!Result.insert(BB)) 209 llvm_unreachable("Repeated basic blocks in extraction input"); 210 } 211 212 LLVM_DEBUG(dbgs() << "Region front block: " << Result.front()->getName() 213 << '\n'); 214 215 for (auto *BB : Result) { 216 if (!isBlockValidForExtraction(*BB, Result, AllowVarArgs, AllowAlloca)) 217 return {}; 218 219 // Make sure that the first block is not a landing pad. 220 if (BB == Result.front()) { 221 if (BB->isEHPad()) { 222 LLVM_DEBUG(dbgs() << "The first block cannot be an unwind block\n"); 223 return {}; 224 } 225 continue; 226 } 227 228 // All blocks other than the first must not have predecessors outside of 229 // the subgraph which is being extracted. 230 for (auto *PBB : predecessors(BB)) 231 if (!Result.count(PBB)) { 232 LLVM_DEBUG(dbgs() << "No blocks in this region may have entries from " 233 "outside the region except for the first block!\n" 234 << "Problematic source BB: " << BB->getName() << "\n" 235 << "Problematic destination BB: " << PBB->getName() 236 << "\n"); 237 return {}; 238 } 239 } 240 241 return Result; 242 } 243 244 CodeExtractor::CodeExtractor(ArrayRef<BasicBlock *> BBs, DominatorTree *DT, 245 bool AggregateArgs, BlockFrequencyInfo *BFI, 246 BranchProbabilityInfo *BPI, AssumptionCache *AC, 247 bool AllowVarArgs, bool AllowAlloca, 248 BasicBlock *AllocationBlock, std::string Suffix, 249 bool ArgsInZeroAddressSpace) 250 : DT(DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI), 251 BPI(BPI), AC(AC), AllocationBlock(AllocationBlock), 252 AllowVarArgs(AllowVarArgs), 253 Blocks(buildExtractionBlockSet(BBs, DT, AllowVarArgs, AllowAlloca)), 254 Suffix(Suffix), ArgsInZeroAddressSpace(ArgsInZeroAddressSpace) {} 255 256 CodeExtractor::CodeExtractor(DominatorTree &DT, Loop &L, bool AggregateArgs, 257 BlockFrequencyInfo *BFI, 258 BranchProbabilityInfo *BPI, AssumptionCache *AC, 259 std::string Suffix) 260 : DT(&DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI), 261 BPI(BPI), AC(AC), AllocationBlock(nullptr), AllowVarArgs(false), 262 Blocks(buildExtractionBlockSet(L.getBlocks(), &DT, 263 /* AllowVarArgs */ false, 264 /* AllowAlloca */ false)), 265 Suffix(Suffix) {} 266 267 /// definedInRegion - Return true if the specified value is defined in the 268 /// extracted region. 269 static bool definedInRegion(const SetVector<BasicBlock *> &Blocks, Value *V) { 270 if (Instruction *I = dyn_cast<Instruction>(V)) 271 if (Blocks.count(I->getParent())) 272 return true; 273 return false; 274 } 275 276 /// definedInCaller - Return true if the specified value is defined in the 277 /// function being code extracted, but not in the region being extracted. 278 /// These values must be passed in as live-ins to the function. 279 static bool definedInCaller(const SetVector<BasicBlock *> &Blocks, Value *V) { 280 if (isa<Argument>(V)) return true; 281 if (Instruction *I = dyn_cast<Instruction>(V)) 282 if (!Blocks.count(I->getParent())) 283 return true; 284 return false; 285 } 286 287 static BasicBlock *getCommonExitBlock(const SetVector<BasicBlock *> &Blocks) { 288 BasicBlock *CommonExitBlock = nullptr; 289 auto hasNonCommonExitSucc = [&](BasicBlock *Block) { 290 for (auto *Succ : successors(Block)) { 291 // Internal edges, ok. 292 if (Blocks.count(Succ)) 293 continue; 294 if (!CommonExitBlock) { 295 CommonExitBlock = Succ; 296 continue; 297 } 298 if (CommonExitBlock != Succ) 299 return true; 300 } 301 return false; 302 }; 303 304 if (any_of(Blocks, hasNonCommonExitSucc)) 305 return nullptr; 306 307 return CommonExitBlock; 308 } 309 310 CodeExtractorAnalysisCache::CodeExtractorAnalysisCache(Function &F) { 311 for (BasicBlock &BB : F) { 312 for (Instruction &II : BB.instructionsWithoutDebug()) 313 if (auto *AI = dyn_cast<AllocaInst>(&II)) 314 Allocas.push_back(AI); 315 316 findSideEffectInfoForBlock(BB); 317 } 318 } 319 320 void CodeExtractorAnalysisCache::findSideEffectInfoForBlock(BasicBlock &BB) { 321 for (Instruction &II : BB.instructionsWithoutDebug()) { 322 unsigned Opcode = II.getOpcode(); 323 Value *MemAddr = nullptr; 324 switch (Opcode) { 325 case Instruction::Store: 326 case Instruction::Load: { 327 if (Opcode == Instruction::Store) { 328 StoreInst *SI = cast<StoreInst>(&II); 329 MemAddr = SI->getPointerOperand(); 330 } else { 331 LoadInst *LI = cast<LoadInst>(&II); 332 MemAddr = LI->getPointerOperand(); 333 } 334 // Global variable can not be aliased with locals. 335 if (isa<Constant>(MemAddr)) 336 break; 337 Value *Base = MemAddr->stripInBoundsConstantOffsets(); 338 if (!isa<AllocaInst>(Base)) { 339 SideEffectingBlocks.insert(&BB); 340 return; 341 } 342 BaseMemAddrs[&BB].insert(Base); 343 break; 344 } 345 default: { 346 IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(&II); 347 if (IntrInst) { 348 if (IntrInst->isLifetimeStartOrEnd()) 349 break; 350 SideEffectingBlocks.insert(&BB); 351 return; 352 } 353 // Treat all the other cases conservatively if it has side effects. 354 if (II.mayHaveSideEffects()) { 355 SideEffectingBlocks.insert(&BB); 356 return; 357 } 358 } 359 } 360 } 361 } 362 363 bool CodeExtractorAnalysisCache::doesBlockContainClobberOfAddr( 364 BasicBlock &BB, AllocaInst *Addr) const { 365 if (SideEffectingBlocks.count(&BB)) 366 return true; 367 auto It = BaseMemAddrs.find(&BB); 368 if (It != BaseMemAddrs.end()) 369 return It->second.count(Addr); 370 return false; 371 } 372 373 bool CodeExtractor::isLegalToShrinkwrapLifetimeMarkers( 374 const CodeExtractorAnalysisCache &CEAC, Instruction *Addr) const { 375 AllocaInst *AI = cast<AllocaInst>(Addr->stripInBoundsConstantOffsets()); 376 Function *Func = (*Blocks.begin())->getParent(); 377 for (BasicBlock &BB : *Func) { 378 if (Blocks.count(&BB)) 379 continue; 380 if (CEAC.doesBlockContainClobberOfAddr(BB, AI)) 381 return false; 382 } 383 return true; 384 } 385 386 BasicBlock * 387 CodeExtractor::findOrCreateBlockForHoisting(BasicBlock *CommonExitBlock) { 388 BasicBlock *SinglePredFromOutlineRegion = nullptr; 389 assert(!Blocks.count(CommonExitBlock) && 390 "Expect a block outside the region!"); 391 for (auto *Pred : predecessors(CommonExitBlock)) { 392 if (!Blocks.count(Pred)) 393 continue; 394 if (!SinglePredFromOutlineRegion) { 395 SinglePredFromOutlineRegion = Pred; 396 } else if (SinglePredFromOutlineRegion != Pred) { 397 SinglePredFromOutlineRegion = nullptr; 398 break; 399 } 400 } 401 402 if (SinglePredFromOutlineRegion) 403 return SinglePredFromOutlineRegion; 404 405 #ifndef NDEBUG 406 auto getFirstPHI = [](BasicBlock *BB) { 407 BasicBlock::iterator I = BB->begin(); 408 PHINode *FirstPhi = nullptr; 409 while (I != BB->end()) { 410 PHINode *Phi = dyn_cast<PHINode>(I); 411 if (!Phi) 412 break; 413 if (!FirstPhi) { 414 FirstPhi = Phi; 415 break; 416 } 417 } 418 return FirstPhi; 419 }; 420 // If there are any phi nodes, the single pred either exists or has already 421 // be created before code extraction. 422 assert(!getFirstPHI(CommonExitBlock) && "Phi not expected"); 423 #endif 424 425 BasicBlock *NewExitBlock = CommonExitBlock->splitBasicBlock( 426 CommonExitBlock->getFirstNonPHI()->getIterator()); 427 428 for (BasicBlock *Pred : 429 llvm::make_early_inc_range(predecessors(CommonExitBlock))) { 430 if (Blocks.count(Pred)) 431 continue; 432 Pred->getTerminator()->replaceUsesOfWith(CommonExitBlock, NewExitBlock); 433 } 434 // Now add the old exit block to the outline region. 435 Blocks.insert(CommonExitBlock); 436 OldTargets.push_back(NewExitBlock); 437 return CommonExitBlock; 438 } 439 440 // Find the pair of life time markers for address 'Addr' that are either 441 // defined inside the outline region or can legally be shrinkwrapped into the 442 // outline region. If there are not other untracked uses of the address, return 443 // the pair of markers if found; otherwise return a pair of nullptr. 444 CodeExtractor::LifetimeMarkerInfo 445 CodeExtractor::getLifetimeMarkers(const CodeExtractorAnalysisCache &CEAC, 446 Instruction *Addr, 447 BasicBlock *ExitBlock) const { 448 LifetimeMarkerInfo Info; 449 450 for (User *U : Addr->users()) { 451 IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(U); 452 if (IntrInst) { 453 // We don't model addresses with multiple start/end markers, but the 454 // markers do not need to be in the region. 455 if (IntrInst->getIntrinsicID() == Intrinsic::lifetime_start) { 456 if (Info.LifeStart) 457 return {}; 458 Info.LifeStart = IntrInst; 459 continue; 460 } 461 if (IntrInst->getIntrinsicID() == Intrinsic::lifetime_end) { 462 if (Info.LifeEnd) 463 return {}; 464 Info.LifeEnd = IntrInst; 465 continue; 466 } 467 // At this point, permit debug uses outside of the region. 468 // This is fixed in a later call to fixupDebugInfoPostExtraction(). 469 if (isa<DbgInfoIntrinsic>(IntrInst)) 470 continue; 471 } 472 // Find untracked uses of the address, bail. 473 if (!definedInRegion(Blocks, U)) 474 return {}; 475 } 476 477 if (!Info.LifeStart || !Info.LifeEnd) 478 return {}; 479 480 Info.SinkLifeStart = !definedInRegion(Blocks, Info.LifeStart); 481 Info.HoistLifeEnd = !definedInRegion(Blocks, Info.LifeEnd); 482 // Do legality check. 483 if ((Info.SinkLifeStart || Info.HoistLifeEnd) && 484 !isLegalToShrinkwrapLifetimeMarkers(CEAC, Addr)) 485 return {}; 486 487 // Check to see if we have a place to do hoisting, if not, bail. 488 if (Info.HoistLifeEnd && !ExitBlock) 489 return {}; 490 491 return Info; 492 } 493 494 void CodeExtractor::findAllocas(const CodeExtractorAnalysisCache &CEAC, 495 ValueSet &SinkCands, ValueSet &HoistCands, 496 BasicBlock *&ExitBlock) const { 497 Function *Func = (*Blocks.begin())->getParent(); 498 ExitBlock = getCommonExitBlock(Blocks); 499 500 auto moveOrIgnoreLifetimeMarkers = 501 [&](const LifetimeMarkerInfo &LMI) -> bool { 502 if (!LMI.LifeStart) 503 return false; 504 if (LMI.SinkLifeStart) { 505 LLVM_DEBUG(dbgs() << "Sinking lifetime.start: " << *LMI.LifeStart 506 << "\n"); 507 SinkCands.insert(LMI.LifeStart); 508 } 509 if (LMI.HoistLifeEnd) { 510 LLVM_DEBUG(dbgs() << "Hoisting lifetime.end: " << *LMI.LifeEnd << "\n"); 511 HoistCands.insert(LMI.LifeEnd); 512 } 513 return true; 514 }; 515 516 // Look up allocas in the original function in CodeExtractorAnalysisCache, as 517 // this is much faster than walking all the instructions. 518 for (AllocaInst *AI : CEAC.getAllocas()) { 519 BasicBlock *BB = AI->getParent(); 520 if (Blocks.count(BB)) 521 continue; 522 523 // As a prior call to extractCodeRegion() may have shrinkwrapped the alloca, 524 // check whether it is actually still in the original function. 525 Function *AIFunc = BB->getParent(); 526 if (AIFunc != Func) 527 continue; 528 529 LifetimeMarkerInfo MarkerInfo = getLifetimeMarkers(CEAC, AI, ExitBlock); 530 bool Moved = moveOrIgnoreLifetimeMarkers(MarkerInfo); 531 if (Moved) { 532 LLVM_DEBUG(dbgs() << "Sinking alloca: " << *AI << "\n"); 533 SinkCands.insert(AI); 534 continue; 535 } 536 537 // Find bitcasts in the outlined region that have lifetime marker users 538 // outside that region. Replace the lifetime marker use with an 539 // outside region bitcast to avoid unnecessary alloca/reload instructions 540 // and extra lifetime markers. 541 SmallVector<Instruction *, 2> LifetimeBitcastUsers; 542 for (User *U : AI->users()) { 543 if (!definedInRegion(Blocks, U)) 544 continue; 545 546 if (U->stripInBoundsConstantOffsets() != AI) 547 continue; 548 549 Instruction *Bitcast = cast<Instruction>(U); 550 for (User *BU : Bitcast->users()) { 551 IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(BU); 552 if (!IntrInst) 553 continue; 554 555 if (!IntrInst->isLifetimeStartOrEnd()) 556 continue; 557 558 if (definedInRegion(Blocks, IntrInst)) 559 continue; 560 561 LLVM_DEBUG(dbgs() << "Replace use of extracted region bitcast" 562 << *Bitcast << " in out-of-region lifetime marker " 563 << *IntrInst << "\n"); 564 LifetimeBitcastUsers.push_back(IntrInst); 565 } 566 } 567 568 for (Instruction *I : LifetimeBitcastUsers) { 569 Module *M = AIFunc->getParent(); 570 LLVMContext &Ctx = M->getContext(); 571 auto *Int8PtrTy = PointerType::getUnqual(Ctx); 572 CastInst *CastI = 573 CastInst::CreatePointerCast(AI, Int8PtrTy, "lt.cast", I); 574 I->replaceUsesOfWith(I->getOperand(1), CastI); 575 } 576 577 // Follow any bitcasts. 578 SmallVector<Instruction *, 2> Bitcasts; 579 SmallVector<LifetimeMarkerInfo, 2> BitcastLifetimeInfo; 580 for (User *U : AI->users()) { 581 if (U->stripInBoundsConstantOffsets() == AI) { 582 Instruction *Bitcast = cast<Instruction>(U); 583 LifetimeMarkerInfo LMI = getLifetimeMarkers(CEAC, Bitcast, ExitBlock); 584 if (LMI.LifeStart) { 585 Bitcasts.push_back(Bitcast); 586 BitcastLifetimeInfo.push_back(LMI); 587 continue; 588 } 589 } 590 591 // Found unknown use of AI. 592 if (!definedInRegion(Blocks, U)) { 593 Bitcasts.clear(); 594 break; 595 } 596 } 597 598 // Either no bitcasts reference the alloca or there are unknown uses. 599 if (Bitcasts.empty()) 600 continue; 601 602 LLVM_DEBUG(dbgs() << "Sinking alloca (via bitcast): " << *AI << "\n"); 603 SinkCands.insert(AI); 604 for (unsigned I = 0, E = Bitcasts.size(); I != E; ++I) { 605 Instruction *BitcastAddr = Bitcasts[I]; 606 const LifetimeMarkerInfo &LMI = BitcastLifetimeInfo[I]; 607 assert(LMI.LifeStart && 608 "Unsafe to sink bitcast without lifetime markers"); 609 moveOrIgnoreLifetimeMarkers(LMI); 610 if (!definedInRegion(Blocks, BitcastAddr)) { 611 LLVM_DEBUG(dbgs() << "Sinking bitcast-of-alloca: " << *BitcastAddr 612 << "\n"); 613 SinkCands.insert(BitcastAddr); 614 } 615 } 616 } 617 } 618 619 bool CodeExtractor::isEligible() const { 620 if (Blocks.empty()) 621 return false; 622 BasicBlock *Header = *Blocks.begin(); 623 Function *F = Header->getParent(); 624 625 // For functions with varargs, check that varargs handling is only done in the 626 // outlined function, i.e vastart and vaend are only used in outlined blocks. 627 if (AllowVarArgs && F->getFunctionType()->isVarArg()) { 628 auto containsVarArgIntrinsic = [](const Instruction &I) { 629 if (const CallInst *CI = dyn_cast<CallInst>(&I)) 630 if (const Function *Callee = CI->getCalledFunction()) 631 return Callee->getIntrinsicID() == Intrinsic::vastart || 632 Callee->getIntrinsicID() == Intrinsic::vaend; 633 return false; 634 }; 635 636 for (auto &BB : *F) { 637 if (Blocks.count(&BB)) 638 continue; 639 if (llvm::any_of(BB, containsVarArgIntrinsic)) 640 return false; 641 } 642 } 643 return true; 644 } 645 646 void CodeExtractor::findInputsOutputs(ValueSet &Inputs, ValueSet &Outputs, 647 const ValueSet &SinkCands) const { 648 for (BasicBlock *BB : Blocks) { 649 // If a used value is defined outside the region, it's an input. If an 650 // instruction is used outside the region, it's an output. 651 for (Instruction &II : *BB) { 652 for (auto &OI : II.operands()) { 653 Value *V = OI; 654 if (!SinkCands.count(V) && definedInCaller(Blocks, V)) 655 Inputs.insert(V); 656 } 657 658 for (User *U : II.users()) 659 if (!definedInRegion(Blocks, U)) { 660 Outputs.insert(&II); 661 break; 662 } 663 } 664 } 665 } 666 667 /// severSplitPHINodesOfEntry - If a PHI node has multiple inputs from outside 668 /// of the region, we need to split the entry block of the region so that the 669 /// PHI node is easier to deal with. 670 void CodeExtractor::severSplitPHINodesOfEntry(BasicBlock *&Header) { 671 unsigned NumPredsFromRegion = 0; 672 unsigned NumPredsOutsideRegion = 0; 673 674 if (Header != &Header->getParent()->getEntryBlock()) { 675 PHINode *PN = dyn_cast<PHINode>(Header->begin()); 676 if (!PN) return; // No PHI nodes. 677 678 // If the header node contains any PHI nodes, check to see if there is more 679 // than one entry from outside the region. If so, we need to sever the 680 // header block into two. 681 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 682 if (Blocks.count(PN->getIncomingBlock(i))) 683 ++NumPredsFromRegion; 684 else 685 ++NumPredsOutsideRegion; 686 687 // If there is one (or fewer) predecessor from outside the region, we don't 688 // need to do anything special. 689 if (NumPredsOutsideRegion <= 1) return; 690 } 691 692 // Otherwise, we need to split the header block into two pieces: one 693 // containing PHI nodes merging values from outside of the region, and a 694 // second that contains all of the code for the block and merges back any 695 // incoming values from inside of the region. 696 BasicBlock *NewBB = SplitBlock(Header, Header->getFirstNonPHI(), DT); 697 698 // We only want to code extract the second block now, and it becomes the new 699 // header of the region. 700 BasicBlock *OldPred = Header; 701 Blocks.remove(OldPred); 702 Blocks.insert(NewBB); 703 Header = NewBB; 704 705 // Okay, now we need to adjust the PHI nodes and any branches from within the 706 // region to go to the new header block instead of the old header block. 707 if (NumPredsFromRegion) { 708 PHINode *PN = cast<PHINode>(OldPred->begin()); 709 // Loop over all of the predecessors of OldPred that are in the region, 710 // changing them to branch to NewBB instead. 711 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 712 if (Blocks.count(PN->getIncomingBlock(i))) { 713 Instruction *TI = PN->getIncomingBlock(i)->getTerminator(); 714 TI->replaceUsesOfWith(OldPred, NewBB); 715 } 716 717 // Okay, everything within the region is now branching to the right block, we 718 // just have to update the PHI nodes now, inserting PHI nodes into NewBB. 719 BasicBlock::iterator AfterPHIs; 720 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) { 721 PHINode *PN = cast<PHINode>(AfterPHIs); 722 // Create a new PHI node in the new region, which has an incoming value 723 // from OldPred of PN. 724 PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion, 725 PN->getName() + ".ce"); 726 NewPN->insertBefore(NewBB->begin()); 727 PN->replaceAllUsesWith(NewPN); 728 NewPN->addIncoming(PN, OldPred); 729 730 // Loop over all of the incoming value in PN, moving them to NewPN if they 731 // are from the extracted region. 732 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) { 733 if (Blocks.count(PN->getIncomingBlock(i))) { 734 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i)); 735 PN->removeIncomingValue(i); 736 --i; 737 } 738 } 739 } 740 } 741 } 742 743 /// severSplitPHINodesOfExits - if PHI nodes in exit blocks have inputs from 744 /// outlined region, we split these PHIs on two: one with inputs from region 745 /// and other with remaining incoming blocks; then first PHIs are placed in 746 /// outlined region. 747 void CodeExtractor::severSplitPHINodesOfExits( 748 const SmallPtrSetImpl<BasicBlock *> &Exits) { 749 for (BasicBlock *ExitBB : Exits) { 750 BasicBlock *NewBB = nullptr; 751 752 for (PHINode &PN : ExitBB->phis()) { 753 // Find all incoming values from the outlining region. 754 SmallVector<unsigned, 2> IncomingVals; 755 for (unsigned i = 0; i < PN.getNumIncomingValues(); ++i) 756 if (Blocks.count(PN.getIncomingBlock(i))) 757 IncomingVals.push_back(i); 758 759 // Do not process PHI if there is one (or fewer) predecessor from region. 760 // If PHI has exactly one predecessor from region, only this one incoming 761 // will be replaced on codeRepl block, so it should be safe to skip PHI. 762 if (IncomingVals.size() <= 1) 763 continue; 764 765 // Create block for new PHIs and add it to the list of outlined if it 766 // wasn't done before. 767 if (!NewBB) { 768 NewBB = BasicBlock::Create(ExitBB->getContext(), 769 ExitBB->getName() + ".split", 770 ExitBB->getParent(), ExitBB); 771 NewBB->IsNewDbgInfoFormat = ExitBB->IsNewDbgInfoFormat; 772 SmallVector<BasicBlock *, 4> Preds(predecessors(ExitBB)); 773 for (BasicBlock *PredBB : Preds) 774 if (Blocks.count(PredBB)) 775 PredBB->getTerminator()->replaceUsesOfWith(ExitBB, NewBB); 776 BranchInst::Create(ExitBB, NewBB); 777 Blocks.insert(NewBB); 778 } 779 780 // Split this PHI. 781 PHINode *NewPN = PHINode::Create(PN.getType(), IncomingVals.size(), 782 PN.getName() + ".ce"); 783 NewPN->insertBefore(NewBB->getFirstNonPHIIt()); 784 for (unsigned i : IncomingVals) 785 NewPN->addIncoming(PN.getIncomingValue(i), PN.getIncomingBlock(i)); 786 for (unsigned i : reverse(IncomingVals)) 787 PN.removeIncomingValue(i, false); 788 PN.addIncoming(NewPN, NewBB); 789 } 790 } 791 } 792 793 void CodeExtractor::splitReturnBlocks() { 794 for (BasicBlock *Block : Blocks) 795 if (ReturnInst *RI = dyn_cast<ReturnInst>(Block->getTerminator())) { 796 BasicBlock *New = 797 Block->splitBasicBlock(RI->getIterator(), Block->getName() + ".ret"); 798 if (DT) { 799 // Old dominates New. New node dominates all other nodes dominated 800 // by Old. 801 DomTreeNode *OldNode = DT->getNode(Block); 802 SmallVector<DomTreeNode *, 8> Children(OldNode->begin(), 803 OldNode->end()); 804 805 DomTreeNode *NewNode = DT->addNewBlock(New, Block); 806 807 for (DomTreeNode *I : Children) 808 DT->changeImmediateDominator(I, NewNode); 809 } 810 } 811 } 812 813 /// constructFunction - make a function based on inputs and outputs, as follows: 814 /// f(in0, ..., inN, out0, ..., outN) 815 Function *CodeExtractor::constructFunction(const ValueSet &inputs, 816 const ValueSet &outputs, 817 BasicBlock *header, 818 BasicBlock *newRootNode, 819 BasicBlock *newHeader, 820 Function *oldFunction, 821 Module *M) { 822 LLVM_DEBUG(dbgs() << "inputs: " << inputs.size() << "\n"); 823 LLVM_DEBUG(dbgs() << "outputs: " << outputs.size() << "\n"); 824 825 // This function returns unsigned, outputs will go back by reference. 826 switch (NumExitBlocks) { 827 case 0: 828 case 1: RetTy = Type::getVoidTy(header->getContext()); break; 829 case 2: RetTy = Type::getInt1Ty(header->getContext()); break; 830 default: RetTy = Type::getInt16Ty(header->getContext()); break; 831 } 832 833 std::vector<Type *> ParamTy; 834 std::vector<Type *> AggParamTy; 835 ValueSet StructValues; 836 const DataLayout &DL = M->getDataLayout(); 837 838 // Add the types of the input values to the function's argument list 839 for (Value *value : inputs) { 840 LLVM_DEBUG(dbgs() << "value used in func: " << *value << "\n"); 841 if (AggregateArgs && !ExcludeArgsFromAggregate.contains(value)) { 842 AggParamTy.push_back(value->getType()); 843 StructValues.insert(value); 844 } else 845 ParamTy.push_back(value->getType()); 846 } 847 848 // Add the types of the output values to the function's argument list. 849 for (Value *output : outputs) { 850 LLVM_DEBUG(dbgs() << "instr used in func: " << *output << "\n"); 851 if (AggregateArgs && !ExcludeArgsFromAggregate.contains(output)) { 852 AggParamTy.push_back(output->getType()); 853 StructValues.insert(output); 854 } else 855 ParamTy.push_back( 856 PointerType::get(output->getType(), DL.getAllocaAddrSpace())); 857 } 858 859 assert( 860 (ParamTy.size() + AggParamTy.size()) == 861 (inputs.size() + outputs.size()) && 862 "Number of scalar and aggregate params does not match inputs, outputs"); 863 assert((StructValues.empty() || AggregateArgs) && 864 "Expeced StructValues only with AggregateArgs set"); 865 866 // Concatenate scalar and aggregate params in ParamTy. 867 size_t NumScalarParams = ParamTy.size(); 868 StructType *StructTy = nullptr; 869 if (AggregateArgs && !AggParamTy.empty()) { 870 StructTy = StructType::get(M->getContext(), AggParamTy); 871 ParamTy.push_back(PointerType::get( 872 StructTy, ArgsInZeroAddressSpace ? 0 : DL.getAllocaAddrSpace())); 873 } 874 875 LLVM_DEBUG({ 876 dbgs() << "Function type: " << *RetTy << " f("; 877 for (Type *i : ParamTy) 878 dbgs() << *i << ", "; 879 dbgs() << ")\n"; 880 }); 881 882 FunctionType *funcType = FunctionType::get( 883 RetTy, ParamTy, AllowVarArgs && oldFunction->isVarArg()); 884 885 std::string SuffixToUse = 886 Suffix.empty() 887 ? (header->getName().empty() ? "extracted" : header->getName().str()) 888 : Suffix; 889 // Create the new function 890 Function *newFunction = Function::Create( 891 funcType, GlobalValue::InternalLinkage, oldFunction->getAddressSpace(), 892 oldFunction->getName() + "." + SuffixToUse, M); 893 newFunction->IsNewDbgInfoFormat = oldFunction->IsNewDbgInfoFormat; 894 895 // Inherit all of the target dependent attributes and white-listed 896 // target independent attributes. 897 // (e.g. If the extracted region contains a call to an x86.sse 898 // instruction we need to make sure that the extracted region has the 899 // "target-features" attribute allowing it to be lowered. 900 // FIXME: This should be changed to check to see if a specific 901 // attribute can not be inherited. 902 for (const auto &Attr : oldFunction->getAttributes().getFnAttrs()) { 903 if (Attr.isStringAttribute()) { 904 if (Attr.getKindAsString() == "thunk") 905 continue; 906 } else 907 switch (Attr.getKindAsEnum()) { 908 // Those attributes cannot be propagated safely. Explicitly list them 909 // here so we get a warning if new attributes are added. 910 case Attribute::AllocSize: 911 case Attribute::Builtin: 912 case Attribute::Convergent: 913 case Attribute::JumpTable: 914 case Attribute::Naked: 915 case Attribute::NoBuiltin: 916 case Attribute::NoMerge: 917 case Attribute::NoReturn: 918 case Attribute::NoSync: 919 case Attribute::ReturnsTwice: 920 case Attribute::Speculatable: 921 case Attribute::StackAlignment: 922 case Attribute::WillReturn: 923 case Attribute::AllocKind: 924 case Attribute::PresplitCoroutine: 925 case Attribute::Memory: 926 case Attribute::NoFPClass: 927 case Attribute::CoroDestroyOnlyWhenComplete: 928 continue; 929 // Those attributes should be safe to propagate to the extracted function. 930 case Attribute::AlwaysInline: 931 case Attribute::Cold: 932 case Attribute::DisableSanitizerInstrumentation: 933 case Attribute::FnRetThunkExtern: 934 case Attribute::Hot: 935 case Attribute::NoRecurse: 936 case Attribute::InlineHint: 937 case Attribute::MinSize: 938 case Attribute::NoCallback: 939 case Attribute::NoDuplicate: 940 case Attribute::NoFree: 941 case Attribute::NoImplicitFloat: 942 case Attribute::NoInline: 943 case Attribute::NonLazyBind: 944 case Attribute::NoRedZone: 945 case Attribute::NoUnwind: 946 case Attribute::NoSanitizeBounds: 947 case Attribute::NoSanitizeCoverage: 948 case Attribute::NullPointerIsValid: 949 case Attribute::OptimizeForDebugging: 950 case Attribute::OptForFuzzing: 951 case Attribute::OptimizeNone: 952 case Attribute::OptimizeForSize: 953 case Attribute::SafeStack: 954 case Attribute::ShadowCallStack: 955 case Attribute::SanitizeAddress: 956 case Attribute::SanitizeMemory: 957 case Attribute::SanitizeThread: 958 case Attribute::SanitizeHWAddress: 959 case Attribute::SanitizeMemTag: 960 case Attribute::SpeculativeLoadHardening: 961 case Attribute::StackProtect: 962 case Attribute::StackProtectReq: 963 case Attribute::StackProtectStrong: 964 case Attribute::StrictFP: 965 case Attribute::UWTable: 966 case Attribute::VScaleRange: 967 case Attribute::NoCfCheck: 968 case Attribute::MustProgress: 969 case Attribute::NoProfile: 970 case Attribute::SkipProfile: 971 break; 972 // These attributes cannot be applied to functions. 973 case Attribute::Alignment: 974 case Attribute::AllocatedPointer: 975 case Attribute::AllocAlign: 976 case Attribute::ByVal: 977 case Attribute::Dereferenceable: 978 case Attribute::DereferenceableOrNull: 979 case Attribute::ElementType: 980 case Attribute::InAlloca: 981 case Attribute::InReg: 982 case Attribute::Nest: 983 case Attribute::NoAlias: 984 case Attribute::NoCapture: 985 case Attribute::NoUndef: 986 case Attribute::NonNull: 987 case Attribute::Preallocated: 988 case Attribute::ReadNone: 989 case Attribute::ReadOnly: 990 case Attribute::Returned: 991 case Attribute::SExt: 992 case Attribute::StructRet: 993 case Attribute::SwiftError: 994 case Attribute::SwiftSelf: 995 case Attribute::SwiftAsync: 996 case Attribute::ZExt: 997 case Attribute::ImmArg: 998 case Attribute::ByRef: 999 case Attribute::WriteOnly: 1000 case Attribute::Writable: 1001 case Attribute::DeadOnUnwind: 1002 // These are not really attributes. 1003 case Attribute::None: 1004 case Attribute::EndAttrKinds: 1005 case Attribute::EmptyKey: 1006 case Attribute::TombstoneKey: 1007 llvm_unreachable("Not a function attribute"); 1008 } 1009 1010 newFunction->addFnAttr(Attr); 1011 } 1012 newFunction->insert(newFunction->end(), newRootNode); 1013 1014 // Create scalar and aggregate iterators to name all of the arguments we 1015 // inserted. 1016 Function::arg_iterator ScalarAI = newFunction->arg_begin(); 1017 Function::arg_iterator AggAI = std::next(ScalarAI, NumScalarParams); 1018 1019 // Rewrite all users of the inputs in the extracted region to use the 1020 // arguments (or appropriate addressing into struct) instead. 1021 for (unsigned i = 0, e = inputs.size(), aggIdx = 0; i != e; ++i) { 1022 Value *RewriteVal; 1023 if (AggregateArgs && StructValues.contains(inputs[i])) { 1024 Value *Idx[2]; 1025 Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext())); 1026 Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), aggIdx); 1027 Instruction *TI = newFunction->begin()->getTerminator(); 1028 GetElementPtrInst *GEP = GetElementPtrInst::Create( 1029 StructTy, &*AggAI, Idx, "gep_" + inputs[i]->getName(), TI); 1030 RewriteVal = new LoadInst(StructTy->getElementType(aggIdx), GEP, 1031 "loadgep_" + inputs[i]->getName(), TI); 1032 ++aggIdx; 1033 } else 1034 RewriteVal = &*ScalarAI++; 1035 1036 std::vector<User *> Users(inputs[i]->user_begin(), inputs[i]->user_end()); 1037 for (User *use : Users) 1038 if (Instruction *inst = dyn_cast<Instruction>(use)) 1039 if (Blocks.count(inst->getParent())) 1040 inst->replaceUsesOfWith(inputs[i], RewriteVal); 1041 } 1042 1043 // Set names for input and output arguments. 1044 if (NumScalarParams) { 1045 ScalarAI = newFunction->arg_begin(); 1046 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++ScalarAI) 1047 if (!StructValues.contains(inputs[i])) 1048 ScalarAI->setName(inputs[i]->getName()); 1049 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++ScalarAI) 1050 if (!StructValues.contains(outputs[i])) 1051 ScalarAI->setName(outputs[i]->getName() + ".out"); 1052 } 1053 1054 // Rewrite branches to basic blocks outside of the loop to new dummy blocks 1055 // within the new function. This must be done before we lose track of which 1056 // blocks were originally in the code region. 1057 std::vector<User *> Users(header->user_begin(), header->user_end()); 1058 for (auto &U : Users) 1059 // The BasicBlock which contains the branch is not in the region 1060 // modify the branch target to a new block 1061 if (Instruction *I = dyn_cast<Instruction>(U)) 1062 if (I->isTerminator() && I->getFunction() == oldFunction && 1063 !Blocks.count(I->getParent())) 1064 I->replaceUsesOfWith(header, newHeader); 1065 1066 return newFunction; 1067 } 1068 1069 /// Erase lifetime.start markers which reference inputs to the extraction 1070 /// region, and insert the referenced memory into \p LifetimesStart. 1071 /// 1072 /// The extraction region is defined by a set of blocks (\p Blocks), and a set 1073 /// of allocas which will be moved from the caller function into the extracted 1074 /// function (\p SunkAllocas). 1075 static void eraseLifetimeMarkersOnInputs(const SetVector<BasicBlock *> &Blocks, 1076 const SetVector<Value *> &SunkAllocas, 1077 SetVector<Value *> &LifetimesStart) { 1078 for (BasicBlock *BB : Blocks) { 1079 for (Instruction &I : llvm::make_early_inc_range(*BB)) { 1080 auto *II = dyn_cast<IntrinsicInst>(&I); 1081 if (!II || !II->isLifetimeStartOrEnd()) 1082 continue; 1083 1084 // Get the memory operand of the lifetime marker. If the underlying 1085 // object is a sunk alloca, or is otherwise defined in the extraction 1086 // region, the lifetime marker must not be erased. 1087 Value *Mem = II->getOperand(1)->stripInBoundsOffsets(); 1088 if (SunkAllocas.count(Mem) || definedInRegion(Blocks, Mem)) 1089 continue; 1090 1091 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 1092 LifetimesStart.insert(Mem); 1093 II->eraseFromParent(); 1094 } 1095 } 1096 } 1097 1098 /// Insert lifetime start/end markers surrounding the call to the new function 1099 /// for objects defined in the caller. 1100 static void insertLifetimeMarkersSurroundingCall( 1101 Module *M, ArrayRef<Value *> LifetimesStart, ArrayRef<Value *> LifetimesEnd, 1102 CallInst *TheCall) { 1103 LLVMContext &Ctx = M->getContext(); 1104 auto NegativeOne = ConstantInt::getSigned(Type::getInt64Ty(Ctx), -1); 1105 Instruction *Term = TheCall->getParent()->getTerminator(); 1106 1107 // Emit lifetime markers for the pointers given in \p Objects. Insert the 1108 // markers before the call if \p InsertBefore, and after the call otherwise. 1109 auto insertMarkers = [&](Intrinsic::ID MarkerFunc, ArrayRef<Value *> Objects, 1110 bool InsertBefore) { 1111 for (Value *Mem : Objects) { 1112 assert((!isa<Instruction>(Mem) || cast<Instruction>(Mem)->getFunction() == 1113 TheCall->getFunction()) && 1114 "Input memory not defined in original function"); 1115 1116 Function *Func = Intrinsic::getDeclaration(M, MarkerFunc, Mem->getType()); 1117 auto Marker = CallInst::Create(Func, {NegativeOne, Mem}); 1118 if (InsertBefore) 1119 Marker->insertBefore(TheCall); 1120 else 1121 Marker->insertBefore(Term); 1122 } 1123 }; 1124 1125 if (!LifetimesStart.empty()) { 1126 insertMarkers(Intrinsic::lifetime_start, LifetimesStart, 1127 /*InsertBefore=*/true); 1128 } 1129 1130 if (!LifetimesEnd.empty()) { 1131 insertMarkers(Intrinsic::lifetime_end, LifetimesEnd, 1132 /*InsertBefore=*/false); 1133 } 1134 } 1135 1136 /// emitCallAndSwitchStatement - This method sets up the caller side by adding 1137 /// the call instruction, splitting any PHI nodes in the header block as 1138 /// necessary. 1139 CallInst *CodeExtractor::emitCallAndSwitchStatement(Function *newFunction, 1140 BasicBlock *codeReplacer, 1141 ValueSet &inputs, 1142 ValueSet &outputs) { 1143 // Emit a call to the new function, passing in: *pointer to struct (if 1144 // aggregating parameters), or plan inputs and allocated memory for outputs 1145 std::vector<Value *> params, ReloadOutputs, Reloads; 1146 ValueSet StructValues; 1147 1148 Module *M = newFunction->getParent(); 1149 LLVMContext &Context = M->getContext(); 1150 const DataLayout &DL = M->getDataLayout(); 1151 CallInst *call = nullptr; 1152 1153 // Add inputs as params, or to be filled into the struct 1154 unsigned ScalarInputArgNo = 0; 1155 SmallVector<unsigned, 1> SwiftErrorArgs; 1156 for (Value *input : inputs) { 1157 if (AggregateArgs && !ExcludeArgsFromAggregate.contains(input)) 1158 StructValues.insert(input); 1159 else { 1160 params.push_back(input); 1161 if (input->isSwiftError()) 1162 SwiftErrorArgs.push_back(ScalarInputArgNo); 1163 } 1164 ++ScalarInputArgNo; 1165 } 1166 1167 // Create allocas for the outputs 1168 unsigned ScalarOutputArgNo = 0; 1169 for (Value *output : outputs) { 1170 if (AggregateArgs && !ExcludeArgsFromAggregate.contains(output)) { 1171 StructValues.insert(output); 1172 } else { 1173 AllocaInst *alloca = 1174 new AllocaInst(output->getType(), DL.getAllocaAddrSpace(), 1175 nullptr, output->getName() + ".loc", 1176 &codeReplacer->getParent()->front().front()); 1177 ReloadOutputs.push_back(alloca); 1178 params.push_back(alloca); 1179 ++ScalarOutputArgNo; 1180 } 1181 } 1182 1183 StructType *StructArgTy = nullptr; 1184 AllocaInst *Struct = nullptr; 1185 unsigned NumAggregatedInputs = 0; 1186 if (AggregateArgs && !StructValues.empty()) { 1187 std::vector<Type *> ArgTypes; 1188 for (Value *V : StructValues) 1189 ArgTypes.push_back(V->getType()); 1190 1191 // Allocate a struct at the beginning of this function 1192 StructArgTy = StructType::get(newFunction->getContext(), ArgTypes); 1193 Struct = new AllocaInst( 1194 StructArgTy, DL.getAllocaAddrSpace(), nullptr, "structArg", 1195 AllocationBlock ? &*AllocationBlock->getFirstInsertionPt() 1196 : &codeReplacer->getParent()->front().front()); 1197 1198 if (ArgsInZeroAddressSpace && DL.getAllocaAddrSpace() != 0) { 1199 auto *StructSpaceCast = new AddrSpaceCastInst( 1200 Struct, PointerType ::get(Context, 0), "structArg.ascast"); 1201 StructSpaceCast->insertAfter(Struct); 1202 params.push_back(StructSpaceCast); 1203 } else { 1204 params.push_back(Struct); 1205 } 1206 // Store aggregated inputs in the struct. 1207 for (unsigned i = 0, e = StructValues.size(); i != e; ++i) { 1208 if (inputs.contains(StructValues[i])) { 1209 Value *Idx[2]; 1210 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 1211 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i); 1212 GetElementPtrInst *GEP = GetElementPtrInst::Create( 1213 StructArgTy, Struct, Idx, "gep_" + StructValues[i]->getName()); 1214 GEP->insertInto(codeReplacer, codeReplacer->end()); 1215 new StoreInst(StructValues[i], GEP, codeReplacer); 1216 NumAggregatedInputs++; 1217 } 1218 } 1219 } 1220 1221 // Emit the call to the function 1222 call = CallInst::Create(newFunction, params, 1223 NumExitBlocks > 1 ? "targetBlock" : ""); 1224 // Add debug location to the new call, if the original function has debug 1225 // info. In that case, the terminator of the entry block of the extracted 1226 // function contains the first debug location of the extracted function, 1227 // set in extractCodeRegion. 1228 if (codeReplacer->getParent()->getSubprogram()) { 1229 if (auto DL = newFunction->getEntryBlock().getTerminator()->getDebugLoc()) 1230 call->setDebugLoc(DL); 1231 } 1232 call->insertInto(codeReplacer, codeReplacer->end()); 1233 1234 // Set swifterror parameter attributes. 1235 for (unsigned SwiftErrArgNo : SwiftErrorArgs) { 1236 call->addParamAttr(SwiftErrArgNo, Attribute::SwiftError); 1237 newFunction->addParamAttr(SwiftErrArgNo, Attribute::SwiftError); 1238 } 1239 1240 // Reload the outputs passed in by reference, use the struct if output is in 1241 // the aggregate or reload from the scalar argument. 1242 for (unsigned i = 0, e = outputs.size(), scalarIdx = 0, 1243 aggIdx = NumAggregatedInputs; 1244 i != e; ++i) { 1245 Value *Output = nullptr; 1246 if (AggregateArgs && StructValues.contains(outputs[i])) { 1247 Value *Idx[2]; 1248 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 1249 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), aggIdx); 1250 GetElementPtrInst *GEP = GetElementPtrInst::Create( 1251 StructArgTy, Struct, Idx, "gep_reload_" + outputs[i]->getName()); 1252 GEP->insertInto(codeReplacer, codeReplacer->end()); 1253 Output = GEP; 1254 ++aggIdx; 1255 } else { 1256 Output = ReloadOutputs[scalarIdx]; 1257 ++scalarIdx; 1258 } 1259 LoadInst *load = new LoadInst(outputs[i]->getType(), Output, 1260 outputs[i]->getName() + ".reload", 1261 codeReplacer); 1262 Reloads.push_back(load); 1263 std::vector<User *> Users(outputs[i]->user_begin(), outputs[i]->user_end()); 1264 for (User *U : Users) { 1265 Instruction *inst = cast<Instruction>(U); 1266 if (!Blocks.count(inst->getParent())) 1267 inst->replaceUsesOfWith(outputs[i], load); 1268 } 1269 } 1270 1271 // Now we can emit a switch statement using the call as a value. 1272 SwitchInst *TheSwitch = 1273 SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)), 1274 codeReplacer, 0, codeReplacer); 1275 1276 // Since there may be multiple exits from the original region, make the new 1277 // function return an unsigned, switch on that number. This loop iterates 1278 // over all of the blocks in the extracted region, updating any terminator 1279 // instructions in the to-be-extracted region that branch to blocks that are 1280 // not in the region to be extracted. 1281 std::map<BasicBlock *, BasicBlock *> ExitBlockMap; 1282 1283 // Iterate over the previously collected targets, and create new blocks inside 1284 // the function to branch to. 1285 unsigned switchVal = 0; 1286 for (BasicBlock *OldTarget : OldTargets) { 1287 if (Blocks.count(OldTarget)) 1288 continue; 1289 BasicBlock *&NewTarget = ExitBlockMap[OldTarget]; 1290 if (NewTarget) 1291 continue; 1292 1293 // If we don't already have an exit stub for this non-extracted 1294 // destination, create one now! 1295 NewTarget = BasicBlock::Create(Context, 1296 OldTarget->getName() + ".exitStub", 1297 newFunction); 1298 unsigned SuccNum = switchVal++; 1299 1300 Value *brVal = nullptr; 1301 assert(NumExitBlocks < 0xffff && "too many exit blocks for switch"); 1302 switch (NumExitBlocks) { 1303 case 0: 1304 case 1: break; // No value needed. 1305 case 2: // Conditional branch, return a bool 1306 brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum); 1307 break; 1308 default: 1309 brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum); 1310 break; 1311 } 1312 1313 ReturnInst::Create(Context, brVal, NewTarget); 1314 1315 // Update the switch instruction. 1316 TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context), 1317 SuccNum), 1318 OldTarget); 1319 } 1320 1321 for (BasicBlock *Block : Blocks) { 1322 Instruction *TI = Block->getTerminator(); 1323 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 1324 if (Blocks.count(TI->getSuccessor(i))) 1325 continue; 1326 BasicBlock *OldTarget = TI->getSuccessor(i); 1327 // add a new basic block which returns the appropriate value 1328 BasicBlock *NewTarget = ExitBlockMap[OldTarget]; 1329 assert(NewTarget && "Unknown target block!"); 1330 1331 // rewrite the original branch instruction with this new target 1332 TI->setSuccessor(i, NewTarget); 1333 } 1334 } 1335 1336 // Store the arguments right after the definition of output value. 1337 // This should be proceeded after creating exit stubs to be ensure that invoke 1338 // result restore will be placed in the outlined function. 1339 Function::arg_iterator ScalarOutputArgBegin = newFunction->arg_begin(); 1340 std::advance(ScalarOutputArgBegin, ScalarInputArgNo); 1341 Function::arg_iterator AggOutputArgBegin = newFunction->arg_begin(); 1342 std::advance(AggOutputArgBegin, ScalarInputArgNo + ScalarOutputArgNo); 1343 1344 for (unsigned i = 0, e = outputs.size(), aggIdx = NumAggregatedInputs; i != e; 1345 ++i) { 1346 auto *OutI = dyn_cast<Instruction>(outputs[i]); 1347 if (!OutI) 1348 continue; 1349 1350 // Find proper insertion point. 1351 BasicBlock::iterator InsertPt; 1352 // In case OutI is an invoke, we insert the store at the beginning in the 1353 // 'normal destination' BB. Otherwise we insert the store right after OutI. 1354 if (auto *InvokeI = dyn_cast<InvokeInst>(OutI)) 1355 InsertPt = InvokeI->getNormalDest()->getFirstInsertionPt(); 1356 else if (auto *Phi = dyn_cast<PHINode>(OutI)) 1357 InsertPt = Phi->getParent()->getFirstInsertionPt(); 1358 else 1359 InsertPt = std::next(OutI->getIterator()); 1360 1361 Instruction *InsertBefore = &*InsertPt; 1362 assert((InsertBefore->getFunction() == newFunction || 1363 Blocks.count(InsertBefore->getParent())) && 1364 "InsertPt should be in new function"); 1365 if (AggregateArgs && StructValues.contains(outputs[i])) { 1366 assert(AggOutputArgBegin != newFunction->arg_end() && 1367 "Number of aggregate output arguments should match " 1368 "the number of defined values"); 1369 Value *Idx[2]; 1370 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 1371 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), aggIdx); 1372 GetElementPtrInst *GEP = GetElementPtrInst::Create( 1373 StructArgTy, &*AggOutputArgBegin, Idx, "gep_" + outputs[i]->getName(), 1374 InsertBefore); 1375 new StoreInst(outputs[i], GEP, InsertBefore); 1376 ++aggIdx; 1377 // Since there should be only one struct argument aggregating 1378 // all the output values, we shouldn't increment AggOutputArgBegin, which 1379 // always points to the struct argument, in this case. 1380 } else { 1381 assert(ScalarOutputArgBegin != newFunction->arg_end() && 1382 "Number of scalar output arguments should match " 1383 "the number of defined values"); 1384 new StoreInst(outputs[i], &*ScalarOutputArgBegin, InsertBefore); 1385 ++ScalarOutputArgBegin; 1386 } 1387 } 1388 1389 // Now that we've done the deed, simplify the switch instruction. 1390 Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType(); 1391 switch (NumExitBlocks) { 1392 case 0: 1393 // There are no successors (the block containing the switch itself), which 1394 // means that previously this was the last part of the function, and hence 1395 // this should be rewritten as a `ret' 1396 1397 // Check if the function should return a value 1398 if (OldFnRetTy->isVoidTy()) { 1399 ReturnInst::Create(Context, nullptr, TheSwitch); // Return void 1400 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) { 1401 // return what we have 1402 ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch); 1403 } else { 1404 // Otherwise we must have code extracted an unwind or something, just 1405 // return whatever we want. 1406 ReturnInst::Create(Context, 1407 Constant::getNullValue(OldFnRetTy), TheSwitch); 1408 } 1409 1410 TheSwitch->eraseFromParent(); 1411 break; 1412 case 1: 1413 // Only a single destination, change the switch into an unconditional 1414 // branch. 1415 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch); 1416 TheSwitch->eraseFromParent(); 1417 break; 1418 case 2: 1419 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2), 1420 call, TheSwitch); 1421 TheSwitch->eraseFromParent(); 1422 break; 1423 default: 1424 // Otherwise, make the default destination of the switch instruction be one 1425 // of the other successors. 1426 TheSwitch->setCondition(call); 1427 TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks)); 1428 // Remove redundant case 1429 TheSwitch->removeCase(SwitchInst::CaseIt(TheSwitch, NumExitBlocks-1)); 1430 break; 1431 } 1432 1433 // Insert lifetime markers around the reloads of any output values. The 1434 // allocas output values are stored in are only in-use in the codeRepl block. 1435 insertLifetimeMarkersSurroundingCall(M, ReloadOutputs, ReloadOutputs, call); 1436 1437 return call; 1438 } 1439 1440 void CodeExtractor::moveCodeToFunction(Function *newFunction) { 1441 auto newFuncIt = newFunction->front().getIterator(); 1442 for (BasicBlock *Block : Blocks) { 1443 // Delete the basic block from the old function, and the list of blocks 1444 Block->removeFromParent(); 1445 1446 // Insert this basic block into the new function 1447 // Insert the original blocks after the entry block created 1448 // for the new function. The entry block may be followed 1449 // by a set of exit blocks at this point, but these exit 1450 // blocks better be placed at the end of the new function. 1451 newFuncIt = newFunction->insert(std::next(newFuncIt), Block); 1452 } 1453 } 1454 1455 void CodeExtractor::calculateNewCallTerminatorWeights( 1456 BasicBlock *CodeReplacer, 1457 DenseMap<BasicBlock *, BlockFrequency> &ExitWeights, 1458 BranchProbabilityInfo *BPI) { 1459 using Distribution = BlockFrequencyInfoImplBase::Distribution; 1460 using BlockNode = BlockFrequencyInfoImplBase::BlockNode; 1461 1462 // Update the branch weights for the exit block. 1463 Instruction *TI = CodeReplacer->getTerminator(); 1464 SmallVector<unsigned, 8> BranchWeights(TI->getNumSuccessors(), 0); 1465 1466 // Block Frequency distribution with dummy node. 1467 Distribution BranchDist; 1468 1469 SmallVector<BranchProbability, 4> EdgeProbabilities( 1470 TI->getNumSuccessors(), BranchProbability::getUnknown()); 1471 1472 // Add each of the frequencies of the successors. 1473 for (unsigned i = 0, e = TI->getNumSuccessors(); i < e; ++i) { 1474 BlockNode ExitNode(i); 1475 uint64_t ExitFreq = ExitWeights[TI->getSuccessor(i)].getFrequency(); 1476 if (ExitFreq != 0) 1477 BranchDist.addExit(ExitNode, ExitFreq); 1478 else 1479 EdgeProbabilities[i] = BranchProbability::getZero(); 1480 } 1481 1482 // Check for no total weight. 1483 if (BranchDist.Total == 0) { 1484 BPI->setEdgeProbability(CodeReplacer, EdgeProbabilities); 1485 return; 1486 } 1487 1488 // Normalize the distribution so that they can fit in unsigned. 1489 BranchDist.normalize(); 1490 1491 // Create normalized branch weights and set the metadata. 1492 for (unsigned I = 0, E = BranchDist.Weights.size(); I < E; ++I) { 1493 const auto &Weight = BranchDist.Weights[I]; 1494 1495 // Get the weight and update the current BFI. 1496 BranchWeights[Weight.TargetNode.Index] = Weight.Amount; 1497 BranchProbability BP(Weight.Amount, BranchDist.Total); 1498 EdgeProbabilities[Weight.TargetNode.Index] = BP; 1499 } 1500 BPI->setEdgeProbability(CodeReplacer, EdgeProbabilities); 1501 TI->setMetadata( 1502 LLVMContext::MD_prof, 1503 MDBuilder(TI->getContext()).createBranchWeights(BranchWeights)); 1504 } 1505 1506 /// Erase debug info intrinsics which refer to values in \p F but aren't in 1507 /// \p F. 1508 static void eraseDebugIntrinsicsWithNonLocalRefs(Function &F) { 1509 for (Instruction &I : instructions(F)) { 1510 SmallVector<DbgVariableIntrinsic *, 4> DbgUsers; 1511 SmallVector<DPValue *, 4> DPValues; 1512 findDbgUsers(DbgUsers, &I, &DPValues); 1513 for (DbgVariableIntrinsic *DVI : DbgUsers) 1514 if (DVI->getFunction() != &F) 1515 DVI->eraseFromParent(); 1516 for (DPValue *DPV : DPValues) 1517 if (DPV->getFunction() != &F) 1518 DPV->eraseFromParent(); 1519 } 1520 } 1521 1522 /// Fix up the debug info in the old and new functions by pointing line 1523 /// locations and debug intrinsics to the new subprogram scope, and by deleting 1524 /// intrinsics which point to values outside of the new function. 1525 static void fixupDebugInfoPostExtraction(Function &OldFunc, Function &NewFunc, 1526 CallInst &TheCall) { 1527 DISubprogram *OldSP = OldFunc.getSubprogram(); 1528 LLVMContext &Ctx = OldFunc.getContext(); 1529 1530 if (!OldSP) { 1531 // Erase any debug info the new function contains. 1532 stripDebugInfo(NewFunc); 1533 // Make sure the old function doesn't contain any non-local metadata refs. 1534 eraseDebugIntrinsicsWithNonLocalRefs(NewFunc); 1535 return; 1536 } 1537 1538 // Create a subprogram for the new function. Leave out a description of the 1539 // function arguments, as the parameters don't correspond to anything at the 1540 // source level. 1541 assert(OldSP->getUnit() && "Missing compile unit for subprogram"); 1542 DIBuilder DIB(*OldFunc.getParent(), /*AllowUnresolved=*/false, 1543 OldSP->getUnit()); 1544 auto SPType = 1545 DIB.createSubroutineType(DIB.getOrCreateTypeArray(std::nullopt)); 1546 DISubprogram::DISPFlags SPFlags = DISubprogram::SPFlagDefinition | 1547 DISubprogram::SPFlagOptimized | 1548 DISubprogram::SPFlagLocalToUnit; 1549 auto NewSP = DIB.createFunction( 1550 OldSP->getUnit(), NewFunc.getName(), NewFunc.getName(), OldSP->getFile(), 1551 /*LineNo=*/0, SPType, /*ScopeLine=*/0, DINode::FlagZero, SPFlags); 1552 NewFunc.setSubprogram(NewSP); 1553 1554 auto IsInvalidLocation = [&NewFunc](Value *Location) { 1555 // Location is invalid if it isn't a constant or an instruction, or is an 1556 // instruction but isn't in the new function. 1557 if (!Location || 1558 (!isa<Constant>(Location) && !isa<Instruction>(Location))) 1559 return true; 1560 Instruction *LocationInst = dyn_cast<Instruction>(Location); 1561 return LocationInst && LocationInst->getFunction() != &NewFunc; 1562 }; 1563 1564 // Debug intrinsics in the new function need to be updated in one of two 1565 // ways: 1566 // 1) They need to be deleted, because they describe a value in the old 1567 // function. 1568 // 2) They need to point to fresh metadata, e.g. because they currently 1569 // point to a variable in the wrong scope. 1570 SmallDenseMap<DINode *, DINode *> RemappedMetadata; 1571 SmallVector<Instruction *, 4> DebugIntrinsicsToDelete; 1572 SmallVector<DPValue *, 4> DPVsToDelete; 1573 DenseMap<const MDNode *, MDNode *> Cache; 1574 1575 auto GetUpdatedDIVariable = [&](DILocalVariable *OldVar) { 1576 DINode *&NewVar = RemappedMetadata[OldVar]; 1577 if (!NewVar) { 1578 DILocalScope *NewScope = DILocalScope::cloneScopeForSubprogram( 1579 *OldVar->getScope(), *NewSP, Ctx, Cache); 1580 NewVar = DIB.createAutoVariable( 1581 NewScope, OldVar->getName(), OldVar->getFile(), OldVar->getLine(), 1582 OldVar->getType(), /*AlwaysPreserve=*/false, DINode::FlagZero, 1583 OldVar->getAlignInBits()); 1584 } 1585 return cast<DILocalVariable>(NewVar); 1586 }; 1587 1588 auto UpdateDPValuesOnInst = [&](Instruction &I) -> void { 1589 for (auto &DPV : I.getDbgValueRange()) { 1590 // Apply the two updates that dbg.values get: invalid operands, and 1591 // variable metadata fixup. 1592 if (any_of(DPV.location_ops(), IsInvalidLocation)) { 1593 DPVsToDelete.push_back(&DPV); 1594 continue; 1595 } 1596 if (DPV.isDbgAssign() && IsInvalidLocation(DPV.getAddress())) { 1597 DPVsToDelete.push_back(&DPV); 1598 continue; 1599 } 1600 if (!DPV.getDebugLoc().getInlinedAt()) 1601 DPV.setVariable(GetUpdatedDIVariable(DPV.getVariable())); 1602 DPV.setDebugLoc(DebugLoc::replaceInlinedAtSubprogram(DPV.getDebugLoc(), 1603 *NewSP, Ctx, Cache)); 1604 } 1605 }; 1606 1607 for (Instruction &I : instructions(NewFunc)) { 1608 UpdateDPValuesOnInst(I); 1609 1610 auto *DII = dyn_cast<DbgInfoIntrinsic>(&I); 1611 if (!DII) 1612 continue; 1613 1614 // Point the intrinsic to a fresh label within the new function if the 1615 // intrinsic was not inlined from some other function. 1616 if (auto *DLI = dyn_cast<DbgLabelInst>(&I)) { 1617 if (DLI->getDebugLoc().getInlinedAt()) 1618 continue; 1619 DILabel *OldLabel = DLI->getLabel(); 1620 DINode *&NewLabel = RemappedMetadata[OldLabel]; 1621 if (!NewLabel) { 1622 DILocalScope *NewScope = DILocalScope::cloneScopeForSubprogram( 1623 *OldLabel->getScope(), *NewSP, Ctx, Cache); 1624 NewLabel = DILabel::get(Ctx, NewScope, OldLabel->getName(), 1625 OldLabel->getFile(), OldLabel->getLine()); 1626 } 1627 DLI->setArgOperand(0, MetadataAsValue::get(Ctx, NewLabel)); 1628 continue; 1629 } 1630 1631 auto *DVI = cast<DbgVariableIntrinsic>(DII); 1632 // If any of the used locations are invalid, delete the intrinsic. 1633 if (any_of(DVI->location_ops(), IsInvalidLocation)) { 1634 DebugIntrinsicsToDelete.push_back(DVI); 1635 continue; 1636 } 1637 // DbgAssign intrinsics have an extra Value argument: 1638 if (auto *DAI = dyn_cast<DbgAssignIntrinsic>(DVI); 1639 DAI && IsInvalidLocation(DAI->getAddress())) { 1640 DebugIntrinsicsToDelete.push_back(DVI); 1641 continue; 1642 } 1643 // If the variable was in the scope of the old function, i.e. it was not 1644 // inlined, point the intrinsic to a fresh variable within the new function. 1645 if (!DVI->getDebugLoc().getInlinedAt()) 1646 DVI->setVariable(GetUpdatedDIVariable(DVI->getVariable())); 1647 } 1648 1649 for (auto *DII : DebugIntrinsicsToDelete) 1650 DII->eraseFromParent(); 1651 for (auto *DPV : DPVsToDelete) 1652 DPV->getMarker()->MarkedInstr->dropOneDbgValue(DPV); 1653 DIB.finalizeSubprogram(NewSP); 1654 1655 // Fix up the scope information attached to the line locations in the new 1656 // function. 1657 for (Instruction &I : instructions(NewFunc)) { 1658 if (const DebugLoc &DL = I.getDebugLoc()) 1659 I.setDebugLoc( 1660 DebugLoc::replaceInlinedAtSubprogram(DL, *NewSP, Ctx, Cache)); 1661 1662 // Loop info metadata may contain line locations. Fix them up. 1663 auto updateLoopInfoLoc = [&Ctx, &Cache, NewSP](Metadata *MD) -> Metadata * { 1664 if (auto *Loc = dyn_cast_or_null<DILocation>(MD)) 1665 return DebugLoc::replaceInlinedAtSubprogram(Loc, *NewSP, Ctx, Cache); 1666 return MD; 1667 }; 1668 updateLoopMetadataDebugLocations(I, updateLoopInfoLoc); 1669 } 1670 if (!TheCall.getDebugLoc()) 1671 TheCall.setDebugLoc(DILocation::get(Ctx, 0, 0, OldSP)); 1672 1673 eraseDebugIntrinsicsWithNonLocalRefs(NewFunc); 1674 } 1675 1676 Function * 1677 CodeExtractor::extractCodeRegion(const CodeExtractorAnalysisCache &CEAC) { 1678 ValueSet Inputs, Outputs; 1679 return extractCodeRegion(CEAC, Inputs, Outputs); 1680 } 1681 1682 Function * 1683 CodeExtractor::extractCodeRegion(const CodeExtractorAnalysisCache &CEAC, 1684 ValueSet &inputs, ValueSet &outputs) { 1685 if (!isEligible()) 1686 return nullptr; 1687 1688 // Assumption: this is a single-entry code region, and the header is the first 1689 // block in the region. 1690 BasicBlock *header = *Blocks.begin(); 1691 Function *oldFunction = header->getParent(); 1692 1693 // Calculate the entry frequency of the new function before we change the root 1694 // block. 1695 BlockFrequency EntryFreq; 1696 if (BFI) { 1697 assert(BPI && "Both BPI and BFI are required to preserve profile info"); 1698 for (BasicBlock *Pred : predecessors(header)) { 1699 if (Blocks.count(Pred)) 1700 continue; 1701 EntryFreq += 1702 BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, header); 1703 } 1704 } 1705 1706 // Remove @llvm.assume calls that will be moved to the new function from the 1707 // old function's assumption cache. 1708 for (BasicBlock *Block : Blocks) { 1709 for (Instruction &I : llvm::make_early_inc_range(*Block)) { 1710 if (auto *AI = dyn_cast<AssumeInst>(&I)) { 1711 if (AC) 1712 AC->unregisterAssumption(AI); 1713 AI->eraseFromParent(); 1714 } 1715 } 1716 } 1717 1718 // If we have any return instructions in the region, split those blocks so 1719 // that the return is not in the region. 1720 splitReturnBlocks(); 1721 1722 // Calculate the exit blocks for the extracted region and the total exit 1723 // weights for each of those blocks. 1724 DenseMap<BasicBlock *, BlockFrequency> ExitWeights; 1725 SmallPtrSet<BasicBlock *, 1> ExitBlocks; 1726 for (BasicBlock *Block : Blocks) { 1727 for (BasicBlock *Succ : successors(Block)) { 1728 if (!Blocks.count(Succ)) { 1729 // Update the branch weight for this successor. 1730 if (BFI) { 1731 BlockFrequency &BF = ExitWeights[Succ]; 1732 BF += BFI->getBlockFreq(Block) * BPI->getEdgeProbability(Block, Succ); 1733 } 1734 ExitBlocks.insert(Succ); 1735 } 1736 } 1737 } 1738 NumExitBlocks = ExitBlocks.size(); 1739 1740 for (BasicBlock *Block : Blocks) { 1741 for (BasicBlock *OldTarget : successors(Block)) 1742 if (!Blocks.contains(OldTarget)) 1743 OldTargets.push_back(OldTarget); 1744 } 1745 1746 // If we have to split PHI nodes of the entry or exit blocks, do so now. 1747 severSplitPHINodesOfEntry(header); 1748 severSplitPHINodesOfExits(ExitBlocks); 1749 1750 // This takes place of the original loop 1751 BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(), 1752 "codeRepl", oldFunction, 1753 header); 1754 codeReplacer->IsNewDbgInfoFormat = oldFunction->IsNewDbgInfoFormat; 1755 1756 // The new function needs a root node because other nodes can branch to the 1757 // head of the region, but the entry node of a function cannot have preds. 1758 BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(), 1759 "newFuncRoot"); 1760 newFuncRoot->IsNewDbgInfoFormat = oldFunction->IsNewDbgInfoFormat; 1761 1762 auto *BranchI = BranchInst::Create(header); 1763 // If the original function has debug info, we have to add a debug location 1764 // to the new branch instruction from the artificial entry block. 1765 // We use the debug location of the first instruction in the extracted 1766 // blocks, as there is no other equivalent line in the source code. 1767 if (oldFunction->getSubprogram()) { 1768 any_of(Blocks, [&BranchI](const BasicBlock *BB) { 1769 return any_of(*BB, [&BranchI](const Instruction &I) { 1770 if (!I.getDebugLoc()) 1771 return false; 1772 BranchI->setDebugLoc(I.getDebugLoc()); 1773 return true; 1774 }); 1775 }); 1776 } 1777 BranchI->insertInto(newFuncRoot, newFuncRoot->end()); 1778 1779 ValueSet SinkingCands, HoistingCands; 1780 BasicBlock *CommonExit = nullptr; 1781 findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit); 1782 assert(HoistingCands.empty() || CommonExit); 1783 1784 // Find inputs to, outputs from the code region. 1785 findInputsOutputs(inputs, outputs, SinkingCands); 1786 1787 // Now sink all instructions which only have non-phi uses inside the region. 1788 // Group the allocas at the start of the block, so that any bitcast uses of 1789 // the allocas are well-defined. 1790 AllocaInst *FirstSunkAlloca = nullptr; 1791 for (auto *II : SinkingCands) { 1792 if (auto *AI = dyn_cast<AllocaInst>(II)) { 1793 AI->moveBefore(*newFuncRoot, newFuncRoot->getFirstInsertionPt()); 1794 if (!FirstSunkAlloca) 1795 FirstSunkAlloca = AI; 1796 } 1797 } 1798 assert((SinkingCands.empty() || FirstSunkAlloca) && 1799 "Did not expect a sink candidate without any allocas"); 1800 for (auto *II : SinkingCands) { 1801 if (!isa<AllocaInst>(II)) { 1802 cast<Instruction>(II)->moveAfter(FirstSunkAlloca); 1803 } 1804 } 1805 1806 if (!HoistingCands.empty()) { 1807 auto *HoistToBlock = findOrCreateBlockForHoisting(CommonExit); 1808 Instruction *TI = HoistToBlock->getTerminator(); 1809 for (auto *II : HoistingCands) 1810 cast<Instruction>(II)->moveBefore(TI); 1811 } 1812 1813 // Collect objects which are inputs to the extraction region and also 1814 // referenced by lifetime start markers within it. The effects of these 1815 // markers must be replicated in the calling function to prevent the stack 1816 // coloring pass from merging slots which store input objects. 1817 ValueSet LifetimesStart; 1818 eraseLifetimeMarkersOnInputs(Blocks, SinkingCands, LifetimesStart); 1819 1820 // Construct new function based on inputs/outputs & add allocas for all defs. 1821 Function *newFunction = 1822 constructFunction(inputs, outputs, header, newFuncRoot, codeReplacer, 1823 oldFunction, oldFunction->getParent()); 1824 1825 // Update the entry count of the function. 1826 if (BFI) { 1827 auto Count = BFI->getProfileCountFromFreq(EntryFreq); 1828 if (Count) 1829 newFunction->setEntryCount( 1830 ProfileCount(*Count, Function::PCT_Real)); // FIXME 1831 BFI->setBlockFreq(codeReplacer, EntryFreq); 1832 } 1833 1834 CallInst *TheCall = 1835 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs); 1836 1837 moveCodeToFunction(newFunction); 1838 1839 // Replicate the effects of any lifetime start/end markers which referenced 1840 // input objects in the extraction region by placing markers around the call. 1841 insertLifetimeMarkersSurroundingCall( 1842 oldFunction->getParent(), LifetimesStart.getArrayRef(), {}, TheCall); 1843 1844 // Propagate personality info to the new function if there is one. 1845 if (oldFunction->hasPersonalityFn()) 1846 newFunction->setPersonalityFn(oldFunction->getPersonalityFn()); 1847 1848 // Update the branch weights for the exit block. 1849 if (BFI && NumExitBlocks > 1) 1850 calculateNewCallTerminatorWeights(codeReplacer, ExitWeights, BPI); 1851 1852 // Loop over all of the PHI nodes in the header and exit blocks, and change 1853 // any references to the old incoming edge to be the new incoming edge. 1854 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) { 1855 PHINode *PN = cast<PHINode>(I); 1856 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1857 if (!Blocks.count(PN->getIncomingBlock(i))) 1858 PN->setIncomingBlock(i, newFuncRoot); 1859 } 1860 1861 for (BasicBlock *ExitBB : ExitBlocks) 1862 for (PHINode &PN : ExitBB->phis()) { 1863 Value *IncomingCodeReplacerVal = nullptr; 1864 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 1865 // Ignore incoming values from outside of the extracted region. 1866 if (!Blocks.count(PN.getIncomingBlock(i))) 1867 continue; 1868 1869 // Ensure that there is only one incoming value from codeReplacer. 1870 if (!IncomingCodeReplacerVal) { 1871 PN.setIncomingBlock(i, codeReplacer); 1872 IncomingCodeReplacerVal = PN.getIncomingValue(i); 1873 } else 1874 assert(IncomingCodeReplacerVal == PN.getIncomingValue(i) && 1875 "PHI has two incompatbile incoming values from codeRepl"); 1876 } 1877 } 1878 1879 fixupDebugInfoPostExtraction(*oldFunction, *newFunction, *TheCall); 1880 1881 // Mark the new function `noreturn` if applicable. Terminators which resume 1882 // exception propagation are treated as returning instructions. This is to 1883 // avoid inserting traps after calls to outlined functions which unwind. 1884 bool doesNotReturn = none_of(*newFunction, [](const BasicBlock &BB) { 1885 const Instruction *Term = BB.getTerminator(); 1886 return isa<ReturnInst>(Term) || isa<ResumeInst>(Term); 1887 }); 1888 if (doesNotReturn) 1889 newFunction->setDoesNotReturn(); 1890 1891 LLVM_DEBUG(if (verifyFunction(*newFunction, &errs())) { 1892 newFunction->dump(); 1893 report_fatal_error("verification of newFunction failed!"); 1894 }); 1895 LLVM_DEBUG(if (verifyFunction(*oldFunction)) 1896 report_fatal_error("verification of oldFunction failed!")); 1897 LLVM_DEBUG(if (AC && verifyAssumptionCache(*oldFunction, *newFunction, AC)) 1898 report_fatal_error("Stale Asumption cache for old Function!")); 1899 return newFunction; 1900 } 1901 1902 bool CodeExtractor::verifyAssumptionCache(const Function &OldFunc, 1903 const Function &NewFunc, 1904 AssumptionCache *AC) { 1905 for (auto AssumeVH : AC->assumptions()) { 1906 auto *I = dyn_cast_or_null<CallInst>(AssumeVH); 1907 if (!I) 1908 continue; 1909 1910 // There shouldn't be any llvm.assume intrinsics in the new function. 1911 if (I->getFunction() != &OldFunc) 1912 return true; 1913 1914 // There shouldn't be any stale affected values in the assumption cache 1915 // that were previously in the old function, but that have now been moved 1916 // to the new function. 1917 for (auto AffectedValVH : AC->assumptionsFor(I->getOperand(0))) { 1918 auto *AffectedCI = dyn_cast_or_null<CallInst>(AffectedValVH); 1919 if (!AffectedCI) 1920 continue; 1921 if (AffectedCI->getFunction() != &OldFunc) 1922 return true; 1923 auto *AssumedInst = cast<Instruction>(AffectedCI->getOperand(0)); 1924 if (AssumedInst->getFunction() != &OldFunc) 1925 return true; 1926 } 1927 } 1928 return false; 1929 } 1930 1931 void CodeExtractor::excludeArgFromAggregate(Value *Arg) { 1932 ExcludeArgsFromAggregate.insert(Arg); 1933 } 1934