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 : DT(DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI), 250 BPI(BPI), AC(AC), AllocationBlock(AllocationBlock), 251 AllowVarArgs(AllowVarArgs), 252 Blocks(buildExtractionBlockSet(BBs, DT, AllowVarArgs, AllowAlloca)), 253 Suffix(Suffix) {} 254 255 CodeExtractor::CodeExtractor(DominatorTree &DT, Loop &L, bool AggregateArgs, 256 BlockFrequencyInfo *BFI, 257 BranchProbabilityInfo *BPI, AssumptionCache *AC, 258 std::string Suffix) 259 : DT(&DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI), 260 BPI(BPI), AC(AC), AllocationBlock(nullptr), AllowVarArgs(false), 261 Blocks(buildExtractionBlockSet(L.getBlocks(), &DT, 262 /* AllowVarArgs */ false, 263 /* AllowAlloca */ false)), 264 Suffix(Suffix) {} 265 266 /// definedInRegion - Return true if the specified value is defined in the 267 /// extracted region. 268 static bool definedInRegion(const SetVector<BasicBlock *> &Blocks, Value *V) { 269 if (Instruction *I = dyn_cast<Instruction>(V)) 270 if (Blocks.count(I->getParent())) 271 return true; 272 return false; 273 } 274 275 /// definedInCaller - Return true if the specified value is defined in the 276 /// function being code extracted, but not in the region being extracted. 277 /// These values must be passed in as live-ins to the function. 278 static bool definedInCaller(const SetVector<BasicBlock *> &Blocks, Value *V) { 279 if (isa<Argument>(V)) return true; 280 if (Instruction *I = dyn_cast<Instruction>(V)) 281 if (!Blocks.count(I->getParent())) 282 return true; 283 return false; 284 } 285 286 static BasicBlock *getCommonExitBlock(const SetVector<BasicBlock *> &Blocks) { 287 BasicBlock *CommonExitBlock = nullptr; 288 auto hasNonCommonExitSucc = [&](BasicBlock *Block) { 289 for (auto *Succ : successors(Block)) { 290 // Internal edges, ok. 291 if (Blocks.count(Succ)) 292 continue; 293 if (!CommonExitBlock) { 294 CommonExitBlock = Succ; 295 continue; 296 } 297 if (CommonExitBlock != Succ) 298 return true; 299 } 300 return false; 301 }; 302 303 if (any_of(Blocks, hasNonCommonExitSucc)) 304 return nullptr; 305 306 return CommonExitBlock; 307 } 308 309 CodeExtractorAnalysisCache::CodeExtractorAnalysisCache(Function &F) { 310 for (BasicBlock &BB : F) { 311 for (Instruction &II : BB.instructionsWithoutDebug()) 312 if (auto *AI = dyn_cast<AllocaInst>(&II)) 313 Allocas.push_back(AI); 314 315 findSideEffectInfoForBlock(BB); 316 } 317 } 318 319 void CodeExtractorAnalysisCache::findSideEffectInfoForBlock(BasicBlock &BB) { 320 for (Instruction &II : BB.instructionsWithoutDebug()) { 321 unsigned Opcode = II.getOpcode(); 322 Value *MemAddr = nullptr; 323 switch (Opcode) { 324 case Instruction::Store: 325 case Instruction::Load: { 326 if (Opcode == Instruction::Store) { 327 StoreInst *SI = cast<StoreInst>(&II); 328 MemAddr = SI->getPointerOperand(); 329 } else { 330 LoadInst *LI = cast<LoadInst>(&II); 331 MemAddr = LI->getPointerOperand(); 332 } 333 // Global variable can not be aliased with locals. 334 if (isa<Constant>(MemAddr)) 335 break; 336 Value *Base = MemAddr->stripInBoundsConstantOffsets(); 337 if (!isa<AllocaInst>(Base)) { 338 SideEffectingBlocks.insert(&BB); 339 return; 340 } 341 BaseMemAddrs[&BB].insert(Base); 342 break; 343 } 344 default: { 345 IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(&II); 346 if (IntrInst) { 347 if (IntrInst->isLifetimeStartOrEnd()) 348 break; 349 SideEffectingBlocks.insert(&BB); 350 return; 351 } 352 // Treat all the other cases conservatively if it has side effects. 353 if (II.mayHaveSideEffects()) { 354 SideEffectingBlocks.insert(&BB); 355 return; 356 } 357 } 358 } 359 } 360 } 361 362 bool CodeExtractorAnalysisCache::doesBlockContainClobberOfAddr( 363 BasicBlock &BB, AllocaInst *Addr) const { 364 if (SideEffectingBlocks.count(&BB)) 365 return true; 366 auto It = BaseMemAddrs.find(&BB); 367 if (It != BaseMemAddrs.end()) 368 return It->second.count(Addr); 369 return false; 370 } 371 372 bool CodeExtractor::isLegalToShrinkwrapLifetimeMarkers( 373 const CodeExtractorAnalysisCache &CEAC, Instruction *Addr) const { 374 AllocaInst *AI = cast<AllocaInst>(Addr->stripInBoundsConstantOffsets()); 375 Function *Func = (*Blocks.begin())->getParent(); 376 for (BasicBlock &BB : *Func) { 377 if (Blocks.count(&BB)) 378 continue; 379 if (CEAC.doesBlockContainClobberOfAddr(BB, AI)) 380 return false; 381 } 382 return true; 383 } 384 385 BasicBlock * 386 CodeExtractor::findOrCreateBlockForHoisting(BasicBlock *CommonExitBlock) { 387 BasicBlock *SinglePredFromOutlineRegion = nullptr; 388 assert(!Blocks.count(CommonExitBlock) && 389 "Expect a block outside the region!"); 390 for (auto *Pred : predecessors(CommonExitBlock)) { 391 if (!Blocks.count(Pred)) 392 continue; 393 if (!SinglePredFromOutlineRegion) { 394 SinglePredFromOutlineRegion = Pred; 395 } else if (SinglePredFromOutlineRegion != Pred) { 396 SinglePredFromOutlineRegion = nullptr; 397 break; 398 } 399 } 400 401 if (SinglePredFromOutlineRegion) 402 return SinglePredFromOutlineRegion; 403 404 #ifndef NDEBUG 405 auto getFirstPHI = [](BasicBlock *BB) { 406 BasicBlock::iterator I = BB->begin(); 407 PHINode *FirstPhi = nullptr; 408 while (I != BB->end()) { 409 PHINode *Phi = dyn_cast<PHINode>(I); 410 if (!Phi) 411 break; 412 if (!FirstPhi) { 413 FirstPhi = Phi; 414 break; 415 } 416 } 417 return FirstPhi; 418 }; 419 // If there are any phi nodes, the single pred either exists or has already 420 // be created before code extraction. 421 assert(!getFirstPHI(CommonExitBlock) && "Phi not expected"); 422 #endif 423 424 BasicBlock *NewExitBlock = CommonExitBlock->splitBasicBlock( 425 CommonExitBlock->getFirstNonPHI()->getIterator()); 426 427 for (BasicBlock *Pred : 428 llvm::make_early_inc_range(predecessors(CommonExitBlock))) { 429 if (Blocks.count(Pred)) 430 continue; 431 Pred->getTerminator()->replaceUsesOfWith(CommonExitBlock, NewExitBlock); 432 } 433 // Now add the old exit block to the outline region. 434 Blocks.insert(CommonExitBlock); 435 OldTargets.push_back(NewExitBlock); 436 return CommonExitBlock; 437 } 438 439 // Find the pair of life time markers for address 'Addr' that are either 440 // defined inside the outline region or can legally be shrinkwrapped into the 441 // outline region. If there are not other untracked uses of the address, return 442 // the pair of markers if found; otherwise return a pair of nullptr. 443 CodeExtractor::LifetimeMarkerInfo 444 CodeExtractor::getLifetimeMarkers(const CodeExtractorAnalysisCache &CEAC, 445 Instruction *Addr, 446 BasicBlock *ExitBlock) const { 447 LifetimeMarkerInfo Info; 448 449 for (User *U : Addr->users()) { 450 IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(U); 451 if (IntrInst) { 452 // We don't model addresses with multiple start/end markers, but the 453 // markers do not need to be in the region. 454 if (IntrInst->getIntrinsicID() == Intrinsic::lifetime_start) { 455 if (Info.LifeStart) 456 return {}; 457 Info.LifeStart = IntrInst; 458 continue; 459 } 460 if (IntrInst->getIntrinsicID() == Intrinsic::lifetime_end) { 461 if (Info.LifeEnd) 462 return {}; 463 Info.LifeEnd = IntrInst; 464 continue; 465 } 466 // At this point, permit debug uses outside of the region. 467 // This is fixed in a later call to fixupDebugInfoPostExtraction(). 468 if (isa<DbgInfoIntrinsic>(IntrInst)) 469 continue; 470 } 471 // Find untracked uses of the address, bail. 472 if (!definedInRegion(Blocks, U)) 473 return {}; 474 } 475 476 if (!Info.LifeStart || !Info.LifeEnd) 477 return {}; 478 479 Info.SinkLifeStart = !definedInRegion(Blocks, Info.LifeStart); 480 Info.HoistLifeEnd = !definedInRegion(Blocks, Info.LifeEnd); 481 // Do legality check. 482 if ((Info.SinkLifeStart || Info.HoistLifeEnd) && 483 !isLegalToShrinkwrapLifetimeMarkers(CEAC, Addr)) 484 return {}; 485 486 // Check to see if we have a place to do hoisting, if not, bail. 487 if (Info.HoistLifeEnd && !ExitBlock) 488 return {}; 489 490 return Info; 491 } 492 493 void CodeExtractor::findAllocas(const CodeExtractorAnalysisCache &CEAC, 494 ValueSet &SinkCands, ValueSet &HoistCands, 495 BasicBlock *&ExitBlock) const { 496 Function *Func = (*Blocks.begin())->getParent(); 497 ExitBlock = getCommonExitBlock(Blocks); 498 499 auto moveOrIgnoreLifetimeMarkers = 500 [&](const LifetimeMarkerInfo &LMI) -> bool { 501 if (!LMI.LifeStart) 502 return false; 503 if (LMI.SinkLifeStart) { 504 LLVM_DEBUG(dbgs() << "Sinking lifetime.start: " << *LMI.LifeStart 505 << "\n"); 506 SinkCands.insert(LMI.LifeStart); 507 } 508 if (LMI.HoistLifeEnd) { 509 LLVM_DEBUG(dbgs() << "Hoisting lifetime.end: " << *LMI.LifeEnd << "\n"); 510 HoistCands.insert(LMI.LifeEnd); 511 } 512 return true; 513 }; 514 515 // Look up allocas in the original function in CodeExtractorAnalysisCache, as 516 // this is much faster than walking all the instructions. 517 for (AllocaInst *AI : CEAC.getAllocas()) { 518 BasicBlock *BB = AI->getParent(); 519 if (Blocks.count(BB)) 520 continue; 521 522 // As a prior call to extractCodeRegion() may have shrinkwrapped the alloca, 523 // check whether it is actually still in the original function. 524 Function *AIFunc = BB->getParent(); 525 if (AIFunc != Func) 526 continue; 527 528 LifetimeMarkerInfo MarkerInfo = getLifetimeMarkers(CEAC, AI, ExitBlock); 529 bool Moved = moveOrIgnoreLifetimeMarkers(MarkerInfo); 530 if (Moved) { 531 LLVM_DEBUG(dbgs() << "Sinking alloca: " << *AI << "\n"); 532 SinkCands.insert(AI); 533 continue; 534 } 535 536 // Find bitcasts in the outlined region that have lifetime marker users 537 // outside that region. Replace the lifetime marker use with an 538 // outside region bitcast to avoid unnecessary alloca/reload instructions 539 // and extra lifetime markers. 540 SmallVector<Instruction *, 2> LifetimeBitcastUsers; 541 for (User *U : AI->users()) { 542 if (!definedInRegion(Blocks, U)) 543 continue; 544 545 if (U->stripInBoundsConstantOffsets() != AI) 546 continue; 547 548 Instruction *Bitcast = cast<Instruction>(U); 549 for (User *BU : Bitcast->users()) { 550 IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(BU); 551 if (!IntrInst) 552 continue; 553 554 if (!IntrInst->isLifetimeStartOrEnd()) 555 continue; 556 557 if (definedInRegion(Blocks, IntrInst)) 558 continue; 559 560 LLVM_DEBUG(dbgs() << "Replace use of extracted region bitcast" 561 << *Bitcast << " in out-of-region lifetime marker " 562 << *IntrInst << "\n"); 563 LifetimeBitcastUsers.push_back(IntrInst); 564 } 565 } 566 567 for (Instruction *I : LifetimeBitcastUsers) { 568 Module *M = AIFunc->getParent(); 569 LLVMContext &Ctx = M->getContext(); 570 auto *Int8PtrTy = Type::getInt8PtrTy(Ctx); 571 CastInst *CastI = 572 CastInst::CreatePointerCast(AI, Int8PtrTy, "lt.cast", I); 573 I->replaceUsesOfWith(I->getOperand(1), CastI); 574 } 575 576 // Follow any bitcasts. 577 SmallVector<Instruction *, 2> Bitcasts; 578 SmallVector<LifetimeMarkerInfo, 2> BitcastLifetimeInfo; 579 for (User *U : AI->users()) { 580 if (U->stripInBoundsConstantOffsets() == AI) { 581 Instruction *Bitcast = cast<Instruction>(U); 582 LifetimeMarkerInfo LMI = getLifetimeMarkers(CEAC, Bitcast, ExitBlock); 583 if (LMI.LifeStart) { 584 Bitcasts.push_back(Bitcast); 585 BitcastLifetimeInfo.push_back(LMI); 586 continue; 587 } 588 } 589 590 // Found unknown use of AI. 591 if (!definedInRegion(Blocks, U)) { 592 Bitcasts.clear(); 593 break; 594 } 595 } 596 597 // Either no bitcasts reference the alloca or there are unknown uses. 598 if (Bitcasts.empty()) 599 continue; 600 601 LLVM_DEBUG(dbgs() << "Sinking alloca (via bitcast): " << *AI << "\n"); 602 SinkCands.insert(AI); 603 for (unsigned I = 0, E = Bitcasts.size(); I != E; ++I) { 604 Instruction *BitcastAddr = Bitcasts[I]; 605 const LifetimeMarkerInfo &LMI = BitcastLifetimeInfo[I]; 606 assert(LMI.LifeStart && 607 "Unsafe to sink bitcast without lifetime markers"); 608 moveOrIgnoreLifetimeMarkers(LMI); 609 if (!definedInRegion(Blocks, BitcastAddr)) { 610 LLVM_DEBUG(dbgs() << "Sinking bitcast-of-alloca: " << *BitcastAddr 611 << "\n"); 612 SinkCands.insert(BitcastAddr); 613 } 614 } 615 } 616 } 617 618 bool CodeExtractor::isEligible() const { 619 if (Blocks.empty()) 620 return false; 621 BasicBlock *Header = *Blocks.begin(); 622 Function *F = Header->getParent(); 623 624 // For functions with varargs, check that varargs handling is only done in the 625 // outlined function, i.e vastart and vaend are only used in outlined blocks. 626 if (AllowVarArgs && F->getFunctionType()->isVarArg()) { 627 auto containsVarArgIntrinsic = [](const Instruction &I) { 628 if (const CallInst *CI = dyn_cast<CallInst>(&I)) 629 if (const Function *Callee = CI->getCalledFunction()) 630 return Callee->getIntrinsicID() == Intrinsic::vastart || 631 Callee->getIntrinsicID() == Intrinsic::vaend; 632 return false; 633 }; 634 635 for (auto &BB : *F) { 636 if (Blocks.count(&BB)) 637 continue; 638 if (llvm::any_of(BB, containsVarArgIntrinsic)) 639 return false; 640 } 641 } 642 return true; 643 } 644 645 void CodeExtractor::findInputsOutputs(ValueSet &Inputs, ValueSet &Outputs, 646 const ValueSet &SinkCands) const { 647 for (BasicBlock *BB : Blocks) { 648 // If a used value is defined outside the region, it's an input. If an 649 // instruction is used outside the region, it's an output. 650 for (Instruction &II : *BB) { 651 for (auto &OI : II.operands()) { 652 Value *V = OI; 653 if (!SinkCands.count(V) && definedInCaller(Blocks, V)) 654 Inputs.insert(V); 655 } 656 657 for (User *U : II.users()) 658 if (!definedInRegion(Blocks, U)) { 659 Outputs.insert(&II); 660 break; 661 } 662 } 663 } 664 } 665 666 /// severSplitPHINodesOfEntry - If a PHI node has multiple inputs from outside 667 /// of the region, we need to split the entry block of the region so that the 668 /// PHI node is easier to deal with. 669 void CodeExtractor::severSplitPHINodesOfEntry(BasicBlock *&Header) { 670 unsigned NumPredsFromRegion = 0; 671 unsigned NumPredsOutsideRegion = 0; 672 673 if (Header != &Header->getParent()->getEntryBlock()) { 674 PHINode *PN = dyn_cast<PHINode>(Header->begin()); 675 if (!PN) return; // No PHI nodes. 676 677 // If the header node contains any PHI nodes, check to see if there is more 678 // than one entry from outside the region. If so, we need to sever the 679 // header block into two. 680 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 681 if (Blocks.count(PN->getIncomingBlock(i))) 682 ++NumPredsFromRegion; 683 else 684 ++NumPredsOutsideRegion; 685 686 // If there is one (or fewer) predecessor from outside the region, we don't 687 // need to do anything special. 688 if (NumPredsOutsideRegion <= 1) return; 689 } 690 691 // Otherwise, we need to split the header block into two pieces: one 692 // containing PHI nodes merging values from outside of the region, and a 693 // second that contains all of the code for the block and merges back any 694 // incoming values from inside of the region. 695 BasicBlock *NewBB = SplitBlock(Header, Header->getFirstNonPHI(), DT); 696 697 // We only want to code extract the second block now, and it becomes the new 698 // header of the region. 699 BasicBlock *OldPred = Header; 700 Blocks.remove(OldPred); 701 Blocks.insert(NewBB); 702 Header = NewBB; 703 704 // Okay, now we need to adjust the PHI nodes and any branches from within the 705 // region to go to the new header block instead of the old header block. 706 if (NumPredsFromRegion) { 707 PHINode *PN = cast<PHINode>(OldPred->begin()); 708 // Loop over all of the predecessors of OldPred that are in the region, 709 // changing them to branch to NewBB instead. 710 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 711 if (Blocks.count(PN->getIncomingBlock(i))) { 712 Instruction *TI = PN->getIncomingBlock(i)->getTerminator(); 713 TI->replaceUsesOfWith(OldPred, NewBB); 714 } 715 716 // Okay, everything within the region is now branching to the right block, we 717 // just have to update the PHI nodes now, inserting PHI nodes into NewBB. 718 BasicBlock::iterator AfterPHIs; 719 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) { 720 PHINode *PN = cast<PHINode>(AfterPHIs); 721 // Create a new PHI node in the new region, which has an incoming value 722 // from OldPred of PN. 723 PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion, 724 PN->getName() + ".ce", &NewBB->front()); 725 PN->replaceAllUsesWith(NewPN); 726 NewPN->addIncoming(PN, OldPred); 727 728 // Loop over all of the incoming value in PN, moving them to NewPN if they 729 // are from the extracted region. 730 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) { 731 if (Blocks.count(PN->getIncomingBlock(i))) { 732 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i)); 733 PN->removeIncomingValue(i); 734 --i; 735 } 736 } 737 } 738 } 739 } 740 741 /// severSplitPHINodesOfExits - if PHI nodes in exit blocks have inputs from 742 /// outlined region, we split these PHIs on two: one with inputs from region 743 /// and other with remaining incoming blocks; then first PHIs are placed in 744 /// outlined region. 745 void CodeExtractor::severSplitPHINodesOfExits( 746 const SmallPtrSetImpl<BasicBlock *> &Exits) { 747 for (BasicBlock *ExitBB : Exits) { 748 BasicBlock *NewBB = nullptr; 749 750 for (PHINode &PN : ExitBB->phis()) { 751 // Find all incoming values from the outlining region. 752 SmallVector<unsigned, 2> IncomingVals; 753 for (unsigned i = 0; i < PN.getNumIncomingValues(); ++i) 754 if (Blocks.count(PN.getIncomingBlock(i))) 755 IncomingVals.push_back(i); 756 757 // Do not process PHI if there is one (or fewer) predecessor from region. 758 // If PHI has exactly one predecessor from region, only this one incoming 759 // will be replaced on codeRepl block, so it should be safe to skip PHI. 760 if (IncomingVals.size() <= 1) 761 continue; 762 763 // Create block for new PHIs and add it to the list of outlined if it 764 // wasn't done before. 765 if (!NewBB) { 766 NewBB = BasicBlock::Create(ExitBB->getContext(), 767 ExitBB->getName() + ".split", 768 ExitBB->getParent(), ExitBB); 769 SmallVector<BasicBlock *, 4> Preds(predecessors(ExitBB)); 770 for (BasicBlock *PredBB : Preds) 771 if (Blocks.count(PredBB)) 772 PredBB->getTerminator()->replaceUsesOfWith(ExitBB, NewBB); 773 BranchInst::Create(ExitBB, NewBB); 774 Blocks.insert(NewBB); 775 } 776 777 // Split this PHI. 778 PHINode *NewPN = 779 PHINode::Create(PN.getType(), IncomingVals.size(), 780 PN.getName() + ".ce", NewBB->getFirstNonPHI()); 781 for (unsigned i : IncomingVals) 782 NewPN->addIncoming(PN.getIncomingValue(i), PN.getIncomingBlock(i)); 783 for (unsigned i : reverse(IncomingVals)) 784 PN.removeIncomingValue(i, false); 785 PN.addIncoming(NewPN, NewBB); 786 } 787 } 788 } 789 790 void CodeExtractor::splitReturnBlocks() { 791 for (BasicBlock *Block : Blocks) 792 if (ReturnInst *RI = dyn_cast<ReturnInst>(Block->getTerminator())) { 793 BasicBlock *New = 794 Block->splitBasicBlock(RI->getIterator(), Block->getName() + ".ret"); 795 if (DT) { 796 // Old dominates New. New node dominates all other nodes dominated 797 // by Old. 798 DomTreeNode *OldNode = DT->getNode(Block); 799 SmallVector<DomTreeNode *, 8> Children(OldNode->begin(), 800 OldNode->end()); 801 802 DomTreeNode *NewNode = DT->addNewBlock(New, Block); 803 804 for (DomTreeNode *I : Children) 805 DT->changeImmediateDominator(I, NewNode); 806 } 807 } 808 } 809 810 /// constructFunction - make a function based on inputs and outputs, as follows: 811 /// f(in0, ..., inN, out0, ..., outN) 812 Function *CodeExtractor::constructFunction(const ValueSet &inputs, 813 const ValueSet &outputs, 814 BasicBlock *header, 815 BasicBlock *newRootNode, 816 BasicBlock *newHeader, 817 Function *oldFunction, 818 Module *M) { 819 LLVM_DEBUG(dbgs() << "inputs: " << inputs.size() << "\n"); 820 LLVM_DEBUG(dbgs() << "outputs: " << outputs.size() << "\n"); 821 822 // This function returns unsigned, outputs will go back by reference. 823 switch (NumExitBlocks) { 824 case 0: 825 case 1: RetTy = Type::getVoidTy(header->getContext()); break; 826 case 2: RetTy = Type::getInt1Ty(header->getContext()); break; 827 default: RetTy = Type::getInt16Ty(header->getContext()); break; 828 } 829 830 std::vector<Type *> ParamTy; 831 std::vector<Type *> AggParamTy; 832 ValueSet StructValues; 833 const DataLayout &DL = M->getDataLayout(); 834 835 // Add the types of the input values to the function's argument list 836 for (Value *value : inputs) { 837 LLVM_DEBUG(dbgs() << "value used in func: " << *value << "\n"); 838 if (AggregateArgs && !ExcludeArgsFromAggregate.contains(value)) { 839 AggParamTy.push_back(value->getType()); 840 StructValues.insert(value); 841 } else 842 ParamTy.push_back(value->getType()); 843 } 844 845 // Add the types of the output values to the function's argument list. 846 for (Value *output : outputs) { 847 LLVM_DEBUG(dbgs() << "instr used in func: " << *output << "\n"); 848 if (AggregateArgs && !ExcludeArgsFromAggregate.contains(output)) { 849 AggParamTy.push_back(output->getType()); 850 StructValues.insert(output); 851 } else 852 ParamTy.push_back( 853 PointerType::get(output->getType(), DL.getAllocaAddrSpace())); 854 } 855 856 assert( 857 (ParamTy.size() + AggParamTy.size()) == 858 (inputs.size() + outputs.size()) && 859 "Number of scalar and aggregate params does not match inputs, outputs"); 860 assert((StructValues.empty() || AggregateArgs) && 861 "Expeced StructValues only with AggregateArgs set"); 862 863 // Concatenate scalar and aggregate params in ParamTy. 864 size_t NumScalarParams = ParamTy.size(); 865 StructType *StructTy = nullptr; 866 if (AggregateArgs && !AggParamTy.empty()) { 867 StructTy = StructType::get(M->getContext(), AggParamTy); 868 ParamTy.push_back(PointerType::get(StructTy, DL.getAllocaAddrSpace())); 869 } 870 871 LLVM_DEBUG({ 872 dbgs() << "Function type: " << *RetTy << " f("; 873 for (Type *i : ParamTy) 874 dbgs() << *i << ", "; 875 dbgs() << ")\n"; 876 }); 877 878 FunctionType *funcType = FunctionType::get( 879 RetTy, ParamTy, AllowVarArgs && oldFunction->isVarArg()); 880 881 std::string SuffixToUse = 882 Suffix.empty() 883 ? (header->getName().empty() ? "extracted" : header->getName().str()) 884 : Suffix; 885 // Create the new function 886 Function *newFunction = Function::Create( 887 funcType, GlobalValue::InternalLinkage, oldFunction->getAddressSpace(), 888 oldFunction->getName() + "." + SuffixToUse, M); 889 890 // Inherit all of the target dependent attributes and white-listed 891 // target independent attributes. 892 // (e.g. If the extracted region contains a call to an x86.sse 893 // instruction we need to make sure that the extracted region has the 894 // "target-features" attribute allowing it to be lowered. 895 // FIXME: This should be changed to check to see if a specific 896 // attribute can not be inherited. 897 for (const auto &Attr : oldFunction->getAttributes().getFnAttrs()) { 898 if (Attr.isStringAttribute()) { 899 if (Attr.getKindAsString() == "thunk") 900 continue; 901 } else 902 switch (Attr.getKindAsEnum()) { 903 // Those attributes cannot be propagated safely. Explicitly list them 904 // here so we get a warning if new attributes are added. 905 case Attribute::AllocSize: 906 case Attribute::Builtin: 907 case Attribute::Convergent: 908 case Attribute::JumpTable: 909 case Attribute::Naked: 910 case Attribute::NoBuiltin: 911 case Attribute::NoMerge: 912 case Attribute::NoReturn: 913 case Attribute::NoSync: 914 case Attribute::ReturnsTwice: 915 case Attribute::Speculatable: 916 case Attribute::StackAlignment: 917 case Attribute::WillReturn: 918 case Attribute::AllocKind: 919 case Attribute::PresplitCoroutine: 920 case Attribute::Memory: 921 case Attribute::NoFPClass: 922 continue; 923 // Those attributes should be safe to propagate to the extracted function. 924 case Attribute::AlwaysInline: 925 case Attribute::Cold: 926 case Attribute::DisableSanitizerInstrumentation: 927 case Attribute::FnRetThunkExtern: 928 case Attribute::Hot: 929 case Attribute::NoRecurse: 930 case Attribute::InlineHint: 931 case Attribute::MinSize: 932 case Attribute::NoCallback: 933 case Attribute::NoDuplicate: 934 case Attribute::NoFree: 935 case Attribute::NoImplicitFloat: 936 case Attribute::NoInline: 937 case Attribute::NonLazyBind: 938 case Attribute::NoRedZone: 939 case Attribute::NoUnwind: 940 case Attribute::NoSanitizeBounds: 941 case Attribute::NoSanitizeCoverage: 942 case Attribute::NullPointerIsValid: 943 case Attribute::OptForFuzzing: 944 case Attribute::OptimizeNone: 945 case Attribute::OptimizeForSize: 946 case Attribute::SafeStack: 947 case Attribute::ShadowCallStack: 948 case Attribute::SanitizeAddress: 949 case Attribute::SanitizeMemory: 950 case Attribute::SanitizeThread: 951 case Attribute::SanitizeHWAddress: 952 case Attribute::SanitizeMemTag: 953 case Attribute::SpeculativeLoadHardening: 954 case Attribute::StackProtect: 955 case Attribute::StackProtectReq: 956 case Attribute::StackProtectStrong: 957 case Attribute::StrictFP: 958 case Attribute::UWTable: 959 case Attribute::VScaleRange: 960 case Attribute::NoCfCheck: 961 case Attribute::MustProgress: 962 case Attribute::NoProfile: 963 case Attribute::SkipProfile: 964 break; 965 // These attributes cannot be applied to functions. 966 case Attribute::Alignment: 967 case Attribute::AllocatedPointer: 968 case Attribute::AllocAlign: 969 case Attribute::ByVal: 970 case Attribute::Dereferenceable: 971 case Attribute::DereferenceableOrNull: 972 case Attribute::ElementType: 973 case Attribute::InAlloca: 974 case Attribute::InReg: 975 case Attribute::Nest: 976 case Attribute::NoAlias: 977 case Attribute::NoCapture: 978 case Attribute::NoUndef: 979 case Attribute::NonNull: 980 case Attribute::Preallocated: 981 case Attribute::ReadNone: 982 case Attribute::ReadOnly: 983 case Attribute::Returned: 984 case Attribute::SExt: 985 case Attribute::StructRet: 986 case Attribute::SwiftError: 987 case Attribute::SwiftSelf: 988 case Attribute::SwiftAsync: 989 case Attribute::ZExt: 990 case Attribute::ImmArg: 991 case Attribute::ByRef: 992 case Attribute::WriteOnly: 993 // These are not really attributes. 994 case Attribute::None: 995 case Attribute::EndAttrKinds: 996 case Attribute::EmptyKey: 997 case Attribute::TombstoneKey: 998 llvm_unreachable("Not a function attribute"); 999 } 1000 1001 newFunction->addFnAttr(Attr); 1002 } 1003 newFunction->insert(newFunction->end(), newRootNode); 1004 1005 // Create scalar and aggregate iterators to name all of the arguments we 1006 // inserted. 1007 Function::arg_iterator ScalarAI = newFunction->arg_begin(); 1008 Function::arg_iterator AggAI = std::next(ScalarAI, NumScalarParams); 1009 1010 // Rewrite all users of the inputs in the extracted region to use the 1011 // arguments (or appropriate addressing into struct) instead. 1012 for (unsigned i = 0, e = inputs.size(), aggIdx = 0; i != e; ++i) { 1013 Value *RewriteVal; 1014 if (AggregateArgs && StructValues.contains(inputs[i])) { 1015 Value *Idx[2]; 1016 Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext())); 1017 Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), aggIdx); 1018 Instruction *TI = newFunction->begin()->getTerminator(); 1019 GetElementPtrInst *GEP = GetElementPtrInst::Create( 1020 StructTy, &*AggAI, Idx, "gep_" + inputs[i]->getName(), TI); 1021 RewriteVal = new LoadInst(StructTy->getElementType(aggIdx), GEP, 1022 "loadgep_" + inputs[i]->getName(), TI); 1023 ++aggIdx; 1024 } else 1025 RewriteVal = &*ScalarAI++; 1026 1027 std::vector<User *> Users(inputs[i]->user_begin(), inputs[i]->user_end()); 1028 for (User *use : Users) 1029 if (Instruction *inst = dyn_cast<Instruction>(use)) 1030 if (Blocks.count(inst->getParent())) 1031 inst->replaceUsesOfWith(inputs[i], RewriteVal); 1032 } 1033 1034 // Set names for input and output arguments. 1035 if (NumScalarParams) { 1036 ScalarAI = newFunction->arg_begin(); 1037 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++ScalarAI) 1038 if (!StructValues.contains(inputs[i])) 1039 ScalarAI->setName(inputs[i]->getName()); 1040 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++ScalarAI) 1041 if (!StructValues.contains(outputs[i])) 1042 ScalarAI->setName(outputs[i]->getName() + ".out"); 1043 } 1044 1045 // Rewrite branches to basic blocks outside of the loop to new dummy blocks 1046 // within the new function. This must be done before we lose track of which 1047 // blocks were originally in the code region. 1048 std::vector<User *> Users(header->user_begin(), header->user_end()); 1049 for (auto &U : Users) 1050 // The BasicBlock which contains the branch is not in the region 1051 // modify the branch target to a new block 1052 if (Instruction *I = dyn_cast<Instruction>(U)) 1053 if (I->isTerminator() && I->getFunction() == oldFunction && 1054 !Blocks.count(I->getParent())) 1055 I->replaceUsesOfWith(header, newHeader); 1056 1057 return newFunction; 1058 } 1059 1060 /// Erase lifetime.start markers which reference inputs to the extraction 1061 /// region, and insert the referenced memory into \p LifetimesStart. 1062 /// 1063 /// The extraction region is defined by a set of blocks (\p Blocks), and a set 1064 /// of allocas which will be moved from the caller function into the extracted 1065 /// function (\p SunkAllocas). 1066 static void eraseLifetimeMarkersOnInputs(const SetVector<BasicBlock *> &Blocks, 1067 const SetVector<Value *> &SunkAllocas, 1068 SetVector<Value *> &LifetimesStart) { 1069 for (BasicBlock *BB : Blocks) { 1070 for (Instruction &I : llvm::make_early_inc_range(*BB)) { 1071 auto *II = dyn_cast<IntrinsicInst>(&I); 1072 if (!II || !II->isLifetimeStartOrEnd()) 1073 continue; 1074 1075 // Get the memory operand of the lifetime marker. If the underlying 1076 // object is a sunk alloca, or is otherwise defined in the extraction 1077 // region, the lifetime marker must not be erased. 1078 Value *Mem = II->getOperand(1)->stripInBoundsOffsets(); 1079 if (SunkAllocas.count(Mem) || definedInRegion(Blocks, Mem)) 1080 continue; 1081 1082 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 1083 LifetimesStart.insert(Mem); 1084 II->eraseFromParent(); 1085 } 1086 } 1087 } 1088 1089 /// Insert lifetime start/end markers surrounding the call to the new function 1090 /// for objects defined in the caller. 1091 static void insertLifetimeMarkersSurroundingCall( 1092 Module *M, ArrayRef<Value *> LifetimesStart, ArrayRef<Value *> LifetimesEnd, 1093 CallInst *TheCall) { 1094 LLVMContext &Ctx = M->getContext(); 1095 auto NegativeOne = ConstantInt::getSigned(Type::getInt64Ty(Ctx), -1); 1096 Instruction *Term = TheCall->getParent()->getTerminator(); 1097 1098 // Emit lifetime markers for the pointers given in \p Objects. Insert the 1099 // markers before the call if \p InsertBefore, and after the call otherwise. 1100 auto insertMarkers = [&](Intrinsic::ID MarkerFunc, ArrayRef<Value *> Objects, 1101 bool InsertBefore) { 1102 for (Value *Mem : Objects) { 1103 assert((!isa<Instruction>(Mem) || cast<Instruction>(Mem)->getFunction() == 1104 TheCall->getFunction()) && 1105 "Input memory not defined in original function"); 1106 1107 Function *Func = Intrinsic::getDeclaration(M, MarkerFunc, Mem->getType()); 1108 auto Marker = CallInst::Create(Func, {NegativeOne, Mem}); 1109 if (InsertBefore) 1110 Marker->insertBefore(TheCall); 1111 else 1112 Marker->insertBefore(Term); 1113 } 1114 }; 1115 1116 if (!LifetimesStart.empty()) { 1117 insertMarkers(Intrinsic::lifetime_start, LifetimesStart, 1118 /*InsertBefore=*/true); 1119 } 1120 1121 if (!LifetimesEnd.empty()) { 1122 insertMarkers(Intrinsic::lifetime_end, LifetimesEnd, 1123 /*InsertBefore=*/false); 1124 } 1125 } 1126 1127 /// emitCallAndSwitchStatement - This method sets up the caller side by adding 1128 /// the call instruction, splitting any PHI nodes in the header block as 1129 /// necessary. 1130 CallInst *CodeExtractor::emitCallAndSwitchStatement(Function *newFunction, 1131 BasicBlock *codeReplacer, 1132 ValueSet &inputs, 1133 ValueSet &outputs) { 1134 // Emit a call to the new function, passing in: *pointer to struct (if 1135 // aggregating parameters), or plan inputs and allocated memory for outputs 1136 std::vector<Value *> params, ReloadOutputs, Reloads; 1137 ValueSet StructValues; 1138 1139 Module *M = newFunction->getParent(); 1140 LLVMContext &Context = M->getContext(); 1141 const DataLayout &DL = M->getDataLayout(); 1142 CallInst *call = nullptr; 1143 1144 // Add inputs as params, or to be filled into the struct 1145 unsigned ScalarInputArgNo = 0; 1146 SmallVector<unsigned, 1> SwiftErrorArgs; 1147 for (Value *input : inputs) { 1148 if (AggregateArgs && !ExcludeArgsFromAggregate.contains(input)) 1149 StructValues.insert(input); 1150 else { 1151 params.push_back(input); 1152 if (input->isSwiftError()) 1153 SwiftErrorArgs.push_back(ScalarInputArgNo); 1154 } 1155 ++ScalarInputArgNo; 1156 } 1157 1158 // Create allocas for the outputs 1159 unsigned ScalarOutputArgNo = 0; 1160 for (Value *output : outputs) { 1161 if (AggregateArgs && !ExcludeArgsFromAggregate.contains(output)) { 1162 StructValues.insert(output); 1163 } else { 1164 AllocaInst *alloca = 1165 new AllocaInst(output->getType(), DL.getAllocaAddrSpace(), 1166 nullptr, output->getName() + ".loc", 1167 &codeReplacer->getParent()->front().front()); 1168 ReloadOutputs.push_back(alloca); 1169 params.push_back(alloca); 1170 ++ScalarOutputArgNo; 1171 } 1172 } 1173 1174 StructType *StructArgTy = nullptr; 1175 AllocaInst *Struct = nullptr; 1176 unsigned NumAggregatedInputs = 0; 1177 if (AggregateArgs && !StructValues.empty()) { 1178 std::vector<Type *> ArgTypes; 1179 for (Value *V : StructValues) 1180 ArgTypes.push_back(V->getType()); 1181 1182 // Allocate a struct at the beginning of this function 1183 StructArgTy = StructType::get(newFunction->getContext(), ArgTypes); 1184 Struct = new AllocaInst( 1185 StructArgTy, DL.getAllocaAddrSpace(), nullptr, "structArg", 1186 AllocationBlock ? &*AllocationBlock->getFirstInsertionPt() 1187 : &codeReplacer->getParent()->front().front()); 1188 params.push_back(Struct); 1189 1190 // Store aggregated inputs in the struct. 1191 for (unsigned i = 0, e = StructValues.size(); i != e; ++i) { 1192 if (inputs.contains(StructValues[i])) { 1193 Value *Idx[2]; 1194 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 1195 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i); 1196 GetElementPtrInst *GEP = GetElementPtrInst::Create( 1197 StructArgTy, Struct, Idx, "gep_" + StructValues[i]->getName()); 1198 GEP->insertInto(codeReplacer, codeReplacer->end()); 1199 new StoreInst(StructValues[i], GEP, codeReplacer); 1200 NumAggregatedInputs++; 1201 } 1202 } 1203 } 1204 1205 // Emit the call to the function 1206 call = CallInst::Create(newFunction, params, 1207 NumExitBlocks > 1 ? "targetBlock" : ""); 1208 // Add debug location to the new call, if the original function has debug 1209 // info. In that case, the terminator of the entry block of the extracted 1210 // function contains the first debug location of the extracted function, 1211 // set in extractCodeRegion. 1212 if (codeReplacer->getParent()->getSubprogram()) { 1213 if (auto DL = newFunction->getEntryBlock().getTerminator()->getDebugLoc()) 1214 call->setDebugLoc(DL); 1215 } 1216 call->insertInto(codeReplacer, codeReplacer->end()); 1217 1218 // Set swifterror parameter attributes. 1219 for (unsigned SwiftErrArgNo : SwiftErrorArgs) { 1220 call->addParamAttr(SwiftErrArgNo, Attribute::SwiftError); 1221 newFunction->addParamAttr(SwiftErrArgNo, Attribute::SwiftError); 1222 } 1223 1224 // Reload the outputs passed in by reference, use the struct if output is in 1225 // the aggregate or reload from the scalar argument. 1226 for (unsigned i = 0, e = outputs.size(), scalarIdx = 0, 1227 aggIdx = NumAggregatedInputs; 1228 i != e; ++i) { 1229 Value *Output = nullptr; 1230 if (AggregateArgs && StructValues.contains(outputs[i])) { 1231 Value *Idx[2]; 1232 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 1233 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), aggIdx); 1234 GetElementPtrInst *GEP = GetElementPtrInst::Create( 1235 StructArgTy, Struct, Idx, "gep_reload_" + outputs[i]->getName()); 1236 GEP->insertInto(codeReplacer, codeReplacer->end()); 1237 Output = GEP; 1238 ++aggIdx; 1239 } else { 1240 Output = ReloadOutputs[scalarIdx]; 1241 ++scalarIdx; 1242 } 1243 LoadInst *load = new LoadInst(outputs[i]->getType(), Output, 1244 outputs[i]->getName() + ".reload", 1245 codeReplacer); 1246 Reloads.push_back(load); 1247 std::vector<User *> Users(outputs[i]->user_begin(), outputs[i]->user_end()); 1248 for (User *U : Users) { 1249 Instruction *inst = cast<Instruction>(U); 1250 if (!Blocks.count(inst->getParent())) 1251 inst->replaceUsesOfWith(outputs[i], load); 1252 } 1253 } 1254 1255 // Now we can emit a switch statement using the call as a value. 1256 SwitchInst *TheSwitch = 1257 SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)), 1258 codeReplacer, 0, codeReplacer); 1259 1260 // Since there may be multiple exits from the original region, make the new 1261 // function return an unsigned, switch on that number. This loop iterates 1262 // over all of the blocks in the extracted region, updating any terminator 1263 // instructions in the to-be-extracted region that branch to blocks that are 1264 // not in the region to be extracted. 1265 std::map<BasicBlock *, BasicBlock *> ExitBlockMap; 1266 1267 // Iterate over the previously collected targets, and create new blocks inside 1268 // the function to branch to. 1269 unsigned switchVal = 0; 1270 for (BasicBlock *OldTarget : OldTargets) { 1271 if (Blocks.count(OldTarget)) 1272 continue; 1273 BasicBlock *&NewTarget = ExitBlockMap[OldTarget]; 1274 if (NewTarget) 1275 continue; 1276 1277 // If we don't already have an exit stub for this non-extracted 1278 // destination, create one now! 1279 NewTarget = BasicBlock::Create(Context, 1280 OldTarget->getName() + ".exitStub", 1281 newFunction); 1282 unsigned SuccNum = switchVal++; 1283 1284 Value *brVal = nullptr; 1285 assert(NumExitBlocks < 0xffff && "too many exit blocks for switch"); 1286 switch (NumExitBlocks) { 1287 case 0: 1288 case 1: break; // No value needed. 1289 case 2: // Conditional branch, return a bool 1290 brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum); 1291 break; 1292 default: 1293 brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum); 1294 break; 1295 } 1296 1297 ReturnInst::Create(Context, brVal, NewTarget); 1298 1299 // Update the switch instruction. 1300 TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context), 1301 SuccNum), 1302 OldTarget); 1303 } 1304 1305 for (BasicBlock *Block : Blocks) { 1306 Instruction *TI = Block->getTerminator(); 1307 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 1308 if (Blocks.count(TI->getSuccessor(i))) 1309 continue; 1310 BasicBlock *OldTarget = TI->getSuccessor(i); 1311 // add a new basic block which returns the appropriate value 1312 BasicBlock *NewTarget = ExitBlockMap[OldTarget]; 1313 assert(NewTarget && "Unknown target block!"); 1314 1315 // rewrite the original branch instruction with this new target 1316 TI->setSuccessor(i, NewTarget); 1317 } 1318 } 1319 1320 // Store the arguments right after the definition of output value. 1321 // This should be proceeded after creating exit stubs to be ensure that invoke 1322 // result restore will be placed in the outlined function. 1323 Function::arg_iterator ScalarOutputArgBegin = newFunction->arg_begin(); 1324 std::advance(ScalarOutputArgBegin, ScalarInputArgNo); 1325 Function::arg_iterator AggOutputArgBegin = newFunction->arg_begin(); 1326 std::advance(AggOutputArgBegin, ScalarInputArgNo + ScalarOutputArgNo); 1327 1328 for (unsigned i = 0, e = outputs.size(), aggIdx = NumAggregatedInputs; i != e; 1329 ++i) { 1330 auto *OutI = dyn_cast<Instruction>(outputs[i]); 1331 if (!OutI) 1332 continue; 1333 1334 // Find proper insertion point. 1335 BasicBlock::iterator InsertPt; 1336 // In case OutI is an invoke, we insert the store at the beginning in the 1337 // 'normal destination' BB. Otherwise we insert the store right after OutI. 1338 if (auto *InvokeI = dyn_cast<InvokeInst>(OutI)) 1339 InsertPt = InvokeI->getNormalDest()->getFirstInsertionPt(); 1340 else if (auto *Phi = dyn_cast<PHINode>(OutI)) 1341 InsertPt = Phi->getParent()->getFirstInsertionPt(); 1342 else 1343 InsertPt = std::next(OutI->getIterator()); 1344 1345 Instruction *InsertBefore = &*InsertPt; 1346 assert((InsertBefore->getFunction() == newFunction || 1347 Blocks.count(InsertBefore->getParent())) && 1348 "InsertPt should be in new function"); 1349 if (AggregateArgs && StructValues.contains(outputs[i])) { 1350 assert(AggOutputArgBegin != newFunction->arg_end() && 1351 "Number of aggregate output arguments should match " 1352 "the number of defined values"); 1353 Value *Idx[2]; 1354 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 1355 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), aggIdx); 1356 GetElementPtrInst *GEP = GetElementPtrInst::Create( 1357 StructArgTy, &*AggOutputArgBegin, Idx, "gep_" + outputs[i]->getName(), 1358 InsertBefore); 1359 new StoreInst(outputs[i], GEP, InsertBefore); 1360 ++aggIdx; 1361 // Since there should be only one struct argument aggregating 1362 // all the output values, we shouldn't increment AggOutputArgBegin, which 1363 // always points to the struct argument, in this case. 1364 } else { 1365 assert(ScalarOutputArgBegin != newFunction->arg_end() && 1366 "Number of scalar output arguments should match " 1367 "the number of defined values"); 1368 new StoreInst(outputs[i], &*ScalarOutputArgBegin, InsertBefore); 1369 ++ScalarOutputArgBegin; 1370 } 1371 } 1372 1373 // Now that we've done the deed, simplify the switch instruction. 1374 Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType(); 1375 switch (NumExitBlocks) { 1376 case 0: 1377 // There are no successors (the block containing the switch itself), which 1378 // means that previously this was the last part of the function, and hence 1379 // this should be rewritten as a `ret' 1380 1381 // Check if the function should return a value 1382 if (OldFnRetTy->isVoidTy()) { 1383 ReturnInst::Create(Context, nullptr, TheSwitch); // Return void 1384 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) { 1385 // return what we have 1386 ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch); 1387 } else { 1388 // Otherwise we must have code extracted an unwind or something, just 1389 // return whatever we want. 1390 ReturnInst::Create(Context, 1391 Constant::getNullValue(OldFnRetTy), TheSwitch); 1392 } 1393 1394 TheSwitch->eraseFromParent(); 1395 break; 1396 case 1: 1397 // Only a single destination, change the switch into an unconditional 1398 // branch. 1399 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch); 1400 TheSwitch->eraseFromParent(); 1401 break; 1402 case 2: 1403 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2), 1404 call, TheSwitch); 1405 TheSwitch->eraseFromParent(); 1406 break; 1407 default: 1408 // Otherwise, make the default destination of the switch instruction be one 1409 // of the other successors. 1410 TheSwitch->setCondition(call); 1411 TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks)); 1412 // Remove redundant case 1413 TheSwitch->removeCase(SwitchInst::CaseIt(TheSwitch, NumExitBlocks-1)); 1414 break; 1415 } 1416 1417 // Insert lifetime markers around the reloads of any output values. The 1418 // allocas output values are stored in are only in-use in the codeRepl block. 1419 insertLifetimeMarkersSurroundingCall(M, ReloadOutputs, ReloadOutputs, call); 1420 1421 return call; 1422 } 1423 1424 void CodeExtractor::moveCodeToFunction(Function *newFunction) { 1425 auto newFuncIt = newFunction->front().getIterator(); 1426 for (BasicBlock *Block : Blocks) { 1427 // Delete the basic block from the old function, and the list of blocks 1428 Block->removeFromParent(); 1429 1430 // Insert this basic block into the new function 1431 // Insert the original blocks after the entry block created 1432 // for the new function. The entry block may be followed 1433 // by a set of exit blocks at this point, but these exit 1434 // blocks better be placed at the end of the new function. 1435 newFuncIt = newFunction->insert(std::next(newFuncIt), Block); 1436 } 1437 } 1438 1439 void CodeExtractor::calculateNewCallTerminatorWeights( 1440 BasicBlock *CodeReplacer, 1441 DenseMap<BasicBlock *, BlockFrequency> &ExitWeights, 1442 BranchProbabilityInfo *BPI) { 1443 using Distribution = BlockFrequencyInfoImplBase::Distribution; 1444 using BlockNode = BlockFrequencyInfoImplBase::BlockNode; 1445 1446 // Update the branch weights for the exit block. 1447 Instruction *TI = CodeReplacer->getTerminator(); 1448 SmallVector<unsigned, 8> BranchWeights(TI->getNumSuccessors(), 0); 1449 1450 // Block Frequency distribution with dummy node. 1451 Distribution BranchDist; 1452 1453 SmallVector<BranchProbability, 4> EdgeProbabilities( 1454 TI->getNumSuccessors(), BranchProbability::getUnknown()); 1455 1456 // Add each of the frequencies of the successors. 1457 for (unsigned i = 0, e = TI->getNumSuccessors(); i < e; ++i) { 1458 BlockNode ExitNode(i); 1459 uint64_t ExitFreq = ExitWeights[TI->getSuccessor(i)].getFrequency(); 1460 if (ExitFreq != 0) 1461 BranchDist.addExit(ExitNode, ExitFreq); 1462 else 1463 EdgeProbabilities[i] = BranchProbability::getZero(); 1464 } 1465 1466 // Check for no total weight. 1467 if (BranchDist.Total == 0) { 1468 BPI->setEdgeProbability(CodeReplacer, EdgeProbabilities); 1469 return; 1470 } 1471 1472 // Normalize the distribution so that they can fit in unsigned. 1473 BranchDist.normalize(); 1474 1475 // Create normalized branch weights and set the metadata. 1476 for (unsigned I = 0, E = BranchDist.Weights.size(); I < E; ++I) { 1477 const auto &Weight = BranchDist.Weights[I]; 1478 1479 // Get the weight and update the current BFI. 1480 BranchWeights[Weight.TargetNode.Index] = Weight.Amount; 1481 BranchProbability BP(Weight.Amount, BranchDist.Total); 1482 EdgeProbabilities[Weight.TargetNode.Index] = BP; 1483 } 1484 BPI->setEdgeProbability(CodeReplacer, EdgeProbabilities); 1485 TI->setMetadata( 1486 LLVMContext::MD_prof, 1487 MDBuilder(TI->getContext()).createBranchWeights(BranchWeights)); 1488 } 1489 1490 /// Erase debug info intrinsics which refer to values in \p F but aren't in 1491 /// \p F. 1492 static void eraseDebugIntrinsicsWithNonLocalRefs(Function &F) { 1493 for (Instruction &I : instructions(F)) { 1494 SmallVector<DbgVariableIntrinsic *, 4> DbgUsers; 1495 findDbgUsers(DbgUsers, &I); 1496 for (DbgVariableIntrinsic *DVI : DbgUsers) 1497 if (DVI->getFunction() != &F) 1498 DVI->eraseFromParent(); 1499 } 1500 } 1501 1502 /// Fix up the debug info in the old and new functions by pointing line 1503 /// locations and debug intrinsics to the new subprogram scope, and by deleting 1504 /// intrinsics which point to values outside of the new function. 1505 static void fixupDebugInfoPostExtraction(Function &OldFunc, Function &NewFunc, 1506 CallInst &TheCall) { 1507 DISubprogram *OldSP = OldFunc.getSubprogram(); 1508 LLVMContext &Ctx = OldFunc.getContext(); 1509 1510 if (!OldSP) { 1511 // Erase any debug info the new function contains. 1512 stripDebugInfo(NewFunc); 1513 // Make sure the old function doesn't contain any non-local metadata refs. 1514 eraseDebugIntrinsicsWithNonLocalRefs(NewFunc); 1515 return; 1516 } 1517 1518 // Create a subprogram for the new function. Leave out a description of the 1519 // function arguments, as the parameters don't correspond to anything at the 1520 // source level. 1521 assert(OldSP->getUnit() && "Missing compile unit for subprogram"); 1522 DIBuilder DIB(*OldFunc.getParent(), /*AllowUnresolved=*/false, 1523 OldSP->getUnit()); 1524 auto SPType = 1525 DIB.createSubroutineType(DIB.getOrCreateTypeArray(std::nullopt)); 1526 DISubprogram::DISPFlags SPFlags = DISubprogram::SPFlagDefinition | 1527 DISubprogram::SPFlagOptimized | 1528 DISubprogram::SPFlagLocalToUnit; 1529 auto NewSP = DIB.createFunction( 1530 OldSP->getUnit(), NewFunc.getName(), NewFunc.getName(), OldSP->getFile(), 1531 /*LineNo=*/0, SPType, /*ScopeLine=*/0, DINode::FlagZero, SPFlags); 1532 NewFunc.setSubprogram(NewSP); 1533 1534 // Debug intrinsics in the new function need to be updated in one of two 1535 // ways: 1536 // 1) They need to be deleted, because they describe a value in the old 1537 // function. 1538 // 2) They need to point to fresh metadata, e.g. because they currently 1539 // point to a variable in the wrong scope. 1540 SmallDenseMap<DINode *, DINode *> RemappedMetadata; 1541 SmallVector<Instruction *, 4> DebugIntrinsicsToDelete; 1542 DenseMap<const MDNode *, MDNode *> Cache; 1543 for (Instruction &I : instructions(NewFunc)) { 1544 auto *DII = dyn_cast<DbgInfoIntrinsic>(&I); 1545 if (!DII) 1546 continue; 1547 1548 // Point the intrinsic to a fresh label within the new function if the 1549 // intrinsic was not inlined from some other function. 1550 if (auto *DLI = dyn_cast<DbgLabelInst>(&I)) { 1551 if (DLI->getDebugLoc().getInlinedAt()) 1552 continue; 1553 DILabel *OldLabel = DLI->getLabel(); 1554 DINode *&NewLabel = RemappedMetadata[OldLabel]; 1555 if (!NewLabel) { 1556 DILocalScope *NewScope = DILocalScope::cloneScopeForSubprogram( 1557 *OldLabel->getScope(), *NewSP, Ctx, Cache); 1558 NewLabel = DILabel::get(Ctx, NewScope, OldLabel->getName(), 1559 OldLabel->getFile(), OldLabel->getLine()); 1560 } 1561 DLI->setArgOperand(0, MetadataAsValue::get(Ctx, NewLabel)); 1562 continue; 1563 } 1564 1565 auto IsInvalidLocation = [&NewFunc](Value *Location) { 1566 // Location is invalid if it isn't a constant or an instruction, or is an 1567 // instruction but isn't in the new function. 1568 if (!Location || 1569 (!isa<Constant>(Location) && !isa<Instruction>(Location))) 1570 return true; 1571 Instruction *LocationInst = dyn_cast<Instruction>(Location); 1572 return LocationInst && LocationInst->getFunction() != &NewFunc; 1573 }; 1574 1575 auto *DVI = cast<DbgVariableIntrinsic>(DII); 1576 // If any of the used locations are invalid, delete the intrinsic. 1577 if (any_of(DVI->location_ops(), IsInvalidLocation)) { 1578 DebugIntrinsicsToDelete.push_back(DVI); 1579 continue; 1580 } 1581 // If the variable was in the scope of the old function, i.e. it was not 1582 // inlined, point the intrinsic to a fresh variable within the new function. 1583 if (!DVI->getDebugLoc().getInlinedAt()) { 1584 DILocalVariable *OldVar = DVI->getVariable(); 1585 DINode *&NewVar = RemappedMetadata[OldVar]; 1586 if (!NewVar) { 1587 DILocalScope *NewScope = DILocalScope::cloneScopeForSubprogram( 1588 *OldVar->getScope(), *NewSP, Ctx, Cache); 1589 NewVar = DIB.createAutoVariable( 1590 NewScope, OldVar->getName(), OldVar->getFile(), OldVar->getLine(), 1591 OldVar->getType(), /*AlwaysPreserve=*/false, DINode::FlagZero, 1592 OldVar->getAlignInBits()); 1593 } 1594 DVI->setVariable(cast<DILocalVariable>(NewVar)); 1595 } 1596 } 1597 1598 for (auto *DII : DebugIntrinsicsToDelete) 1599 DII->eraseFromParent(); 1600 DIB.finalizeSubprogram(NewSP); 1601 1602 // Fix up the scope information attached to the line locations in the new 1603 // function. 1604 for (Instruction &I : instructions(NewFunc)) { 1605 if (const DebugLoc &DL = I.getDebugLoc()) 1606 I.setDebugLoc( 1607 DebugLoc::replaceInlinedAtSubprogram(DL, *NewSP, Ctx, Cache)); 1608 1609 // Loop info metadata may contain line locations. Fix them up. 1610 auto updateLoopInfoLoc = [&Ctx, &Cache, NewSP](Metadata *MD) -> Metadata * { 1611 if (auto *Loc = dyn_cast_or_null<DILocation>(MD)) 1612 return DebugLoc::replaceInlinedAtSubprogram(Loc, *NewSP, Ctx, Cache); 1613 return MD; 1614 }; 1615 updateLoopMetadataDebugLocations(I, updateLoopInfoLoc); 1616 } 1617 if (!TheCall.getDebugLoc()) 1618 TheCall.setDebugLoc(DILocation::get(Ctx, 0, 0, OldSP)); 1619 1620 eraseDebugIntrinsicsWithNonLocalRefs(NewFunc); 1621 } 1622 1623 Function * 1624 CodeExtractor::extractCodeRegion(const CodeExtractorAnalysisCache &CEAC) { 1625 ValueSet Inputs, Outputs; 1626 return extractCodeRegion(CEAC, Inputs, Outputs); 1627 } 1628 1629 Function * 1630 CodeExtractor::extractCodeRegion(const CodeExtractorAnalysisCache &CEAC, 1631 ValueSet &inputs, ValueSet &outputs) { 1632 if (!isEligible()) 1633 return nullptr; 1634 1635 // Assumption: this is a single-entry code region, and the header is the first 1636 // block in the region. 1637 BasicBlock *header = *Blocks.begin(); 1638 Function *oldFunction = header->getParent(); 1639 1640 // Calculate the entry frequency of the new function before we change the root 1641 // block. 1642 BlockFrequency EntryFreq; 1643 if (BFI) { 1644 assert(BPI && "Both BPI and BFI are required to preserve profile info"); 1645 for (BasicBlock *Pred : predecessors(header)) { 1646 if (Blocks.count(Pred)) 1647 continue; 1648 EntryFreq += 1649 BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, header); 1650 } 1651 } 1652 1653 // Remove @llvm.assume calls that will be moved to the new function from the 1654 // old function's assumption cache. 1655 for (BasicBlock *Block : Blocks) { 1656 for (Instruction &I : llvm::make_early_inc_range(*Block)) { 1657 if (auto *AI = dyn_cast<AssumeInst>(&I)) { 1658 if (AC) 1659 AC->unregisterAssumption(AI); 1660 AI->eraseFromParent(); 1661 } 1662 } 1663 } 1664 1665 // If we have any return instructions in the region, split those blocks so 1666 // that the return is not in the region. 1667 splitReturnBlocks(); 1668 1669 // Calculate the exit blocks for the extracted region and the total exit 1670 // weights for each of those blocks. 1671 DenseMap<BasicBlock *, BlockFrequency> ExitWeights; 1672 SmallPtrSet<BasicBlock *, 1> ExitBlocks; 1673 for (BasicBlock *Block : Blocks) { 1674 for (BasicBlock *Succ : successors(Block)) { 1675 if (!Blocks.count(Succ)) { 1676 // Update the branch weight for this successor. 1677 if (BFI) { 1678 BlockFrequency &BF = ExitWeights[Succ]; 1679 BF += BFI->getBlockFreq(Block) * BPI->getEdgeProbability(Block, Succ); 1680 } 1681 ExitBlocks.insert(Succ); 1682 } 1683 } 1684 } 1685 NumExitBlocks = ExitBlocks.size(); 1686 1687 for (BasicBlock *Block : Blocks) { 1688 Instruction *TI = Block->getTerminator(); 1689 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 1690 if (Blocks.count(TI->getSuccessor(i))) 1691 continue; 1692 BasicBlock *OldTarget = TI->getSuccessor(i); 1693 OldTargets.push_back(OldTarget); 1694 } 1695 } 1696 1697 // If we have to split PHI nodes of the entry or exit blocks, do so now. 1698 severSplitPHINodesOfEntry(header); 1699 severSplitPHINodesOfExits(ExitBlocks); 1700 1701 // This takes place of the original loop 1702 BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(), 1703 "codeRepl", oldFunction, 1704 header); 1705 1706 // The new function needs a root node because other nodes can branch to the 1707 // head of the region, but the entry node of a function cannot have preds. 1708 BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(), 1709 "newFuncRoot"); 1710 auto *BranchI = BranchInst::Create(header); 1711 // If the original function has debug info, we have to add a debug location 1712 // to the new branch instruction from the artificial entry block. 1713 // We use the debug location of the first instruction in the extracted 1714 // blocks, as there is no other equivalent line in the source code. 1715 if (oldFunction->getSubprogram()) { 1716 any_of(Blocks, [&BranchI](const BasicBlock *BB) { 1717 return any_of(*BB, [&BranchI](const Instruction &I) { 1718 if (!I.getDebugLoc()) 1719 return false; 1720 BranchI->setDebugLoc(I.getDebugLoc()); 1721 return true; 1722 }); 1723 }); 1724 } 1725 BranchI->insertInto(newFuncRoot, newFuncRoot->end()); 1726 1727 ValueSet SinkingCands, HoistingCands; 1728 BasicBlock *CommonExit = nullptr; 1729 findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit); 1730 assert(HoistingCands.empty() || CommonExit); 1731 1732 // Find inputs to, outputs from the code region. 1733 findInputsOutputs(inputs, outputs, SinkingCands); 1734 1735 // Now sink all instructions which only have non-phi uses inside the region. 1736 // Group the allocas at the start of the block, so that any bitcast uses of 1737 // the allocas are well-defined. 1738 AllocaInst *FirstSunkAlloca = nullptr; 1739 for (auto *II : SinkingCands) { 1740 if (auto *AI = dyn_cast<AllocaInst>(II)) { 1741 AI->moveBefore(*newFuncRoot, newFuncRoot->getFirstInsertionPt()); 1742 if (!FirstSunkAlloca) 1743 FirstSunkAlloca = AI; 1744 } 1745 } 1746 assert((SinkingCands.empty() || FirstSunkAlloca) && 1747 "Did not expect a sink candidate without any allocas"); 1748 for (auto *II : SinkingCands) { 1749 if (!isa<AllocaInst>(II)) { 1750 cast<Instruction>(II)->moveAfter(FirstSunkAlloca); 1751 } 1752 } 1753 1754 if (!HoistingCands.empty()) { 1755 auto *HoistToBlock = findOrCreateBlockForHoisting(CommonExit); 1756 Instruction *TI = HoistToBlock->getTerminator(); 1757 for (auto *II : HoistingCands) 1758 cast<Instruction>(II)->moveBefore(TI); 1759 } 1760 1761 // Collect objects which are inputs to the extraction region and also 1762 // referenced by lifetime start markers within it. The effects of these 1763 // markers must be replicated in the calling function to prevent the stack 1764 // coloring pass from merging slots which store input objects. 1765 ValueSet LifetimesStart; 1766 eraseLifetimeMarkersOnInputs(Blocks, SinkingCands, LifetimesStart); 1767 1768 // Construct new function based on inputs/outputs & add allocas for all defs. 1769 Function *newFunction = 1770 constructFunction(inputs, outputs, header, newFuncRoot, codeReplacer, 1771 oldFunction, oldFunction->getParent()); 1772 1773 // Update the entry count of the function. 1774 if (BFI) { 1775 auto Count = BFI->getProfileCountFromFreq(EntryFreq.getFrequency()); 1776 if (Count) 1777 newFunction->setEntryCount( 1778 ProfileCount(*Count, Function::PCT_Real)); // FIXME 1779 BFI->setBlockFreq(codeReplacer, EntryFreq.getFrequency()); 1780 } 1781 1782 CallInst *TheCall = 1783 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs); 1784 1785 moveCodeToFunction(newFunction); 1786 1787 // Replicate the effects of any lifetime start/end markers which referenced 1788 // input objects in the extraction region by placing markers around the call. 1789 insertLifetimeMarkersSurroundingCall( 1790 oldFunction->getParent(), LifetimesStart.getArrayRef(), {}, TheCall); 1791 1792 // Propagate personality info to the new function if there is one. 1793 if (oldFunction->hasPersonalityFn()) 1794 newFunction->setPersonalityFn(oldFunction->getPersonalityFn()); 1795 1796 // Update the branch weights for the exit block. 1797 if (BFI && NumExitBlocks > 1) 1798 calculateNewCallTerminatorWeights(codeReplacer, ExitWeights, BPI); 1799 1800 // Loop over all of the PHI nodes in the header and exit blocks, and change 1801 // any references to the old incoming edge to be the new incoming edge. 1802 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) { 1803 PHINode *PN = cast<PHINode>(I); 1804 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1805 if (!Blocks.count(PN->getIncomingBlock(i))) 1806 PN->setIncomingBlock(i, newFuncRoot); 1807 } 1808 1809 for (BasicBlock *ExitBB : ExitBlocks) 1810 for (PHINode &PN : ExitBB->phis()) { 1811 Value *IncomingCodeReplacerVal = nullptr; 1812 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 1813 // Ignore incoming values from outside of the extracted region. 1814 if (!Blocks.count(PN.getIncomingBlock(i))) 1815 continue; 1816 1817 // Ensure that there is only one incoming value from codeReplacer. 1818 if (!IncomingCodeReplacerVal) { 1819 PN.setIncomingBlock(i, codeReplacer); 1820 IncomingCodeReplacerVal = PN.getIncomingValue(i); 1821 } else 1822 assert(IncomingCodeReplacerVal == PN.getIncomingValue(i) && 1823 "PHI has two incompatbile incoming values from codeRepl"); 1824 } 1825 } 1826 1827 fixupDebugInfoPostExtraction(*oldFunction, *newFunction, *TheCall); 1828 1829 // Mark the new function `noreturn` if applicable. Terminators which resume 1830 // exception propagation are treated as returning instructions. This is to 1831 // avoid inserting traps after calls to outlined functions which unwind. 1832 bool doesNotReturn = none_of(*newFunction, [](const BasicBlock &BB) { 1833 const Instruction *Term = BB.getTerminator(); 1834 return isa<ReturnInst>(Term) || isa<ResumeInst>(Term); 1835 }); 1836 if (doesNotReturn) 1837 newFunction->setDoesNotReturn(); 1838 1839 LLVM_DEBUG(if (verifyFunction(*newFunction, &errs())) { 1840 newFunction->dump(); 1841 report_fatal_error("verification of newFunction failed!"); 1842 }); 1843 LLVM_DEBUG(if (verifyFunction(*oldFunction)) 1844 report_fatal_error("verification of oldFunction failed!")); 1845 LLVM_DEBUG(if (AC && verifyAssumptionCache(*oldFunction, *newFunction, AC)) 1846 report_fatal_error("Stale Asumption cache for old Function!")); 1847 return newFunction; 1848 } 1849 1850 bool CodeExtractor::verifyAssumptionCache(const Function &OldFunc, 1851 const Function &NewFunc, 1852 AssumptionCache *AC) { 1853 for (auto AssumeVH : AC->assumptions()) { 1854 auto *I = dyn_cast_or_null<CallInst>(AssumeVH); 1855 if (!I) 1856 continue; 1857 1858 // There shouldn't be any llvm.assume intrinsics in the new function. 1859 if (I->getFunction() != &OldFunc) 1860 return true; 1861 1862 // There shouldn't be any stale affected values in the assumption cache 1863 // that were previously in the old function, but that have now been moved 1864 // to the new function. 1865 for (auto AffectedValVH : AC->assumptionsFor(I->getOperand(0))) { 1866 auto *AffectedCI = dyn_cast_or_null<CallInst>(AffectedValVH); 1867 if (!AffectedCI) 1868 continue; 1869 if (AffectedCI->getFunction() != &OldFunc) 1870 return true; 1871 auto *AssumedInst = cast<Instruction>(AffectedCI->getOperand(0)); 1872 if (AssumedInst->getFunction() != &OldFunc) 1873 return true; 1874 } 1875 } 1876 return false; 1877 } 1878 1879 void CodeExtractor::excludeArgFromAggregate(Value *Arg) { 1880 ExcludeArgsFromAggregate.insert(Arg); 1881 } 1882