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/Optional.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/Analysis/AssumptionCache.h"
24 #include "llvm/Analysis/BlockFrequencyInfo.h"
25 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
26 #include "llvm/Analysis/BranchProbabilityInfo.h"
27 #include "llvm/Analysis/LoopInfo.h"
28 #include "llvm/IR/Argument.h"
29 #include "llvm/IR/Attributes.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/CFG.h"
32 #include "llvm/IR/Constant.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/DIBuilder.h"
35 #include "llvm/IR/DataLayout.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/Pass.h"
56 #include "llvm/Support/BlockFrequency.h"
57 #include "llvm/Support/BranchProbability.h"
58 #include "llvm/Support/Casting.h"
59 #include "llvm/Support/CommandLine.h"
60 #include "llvm/Support/Debug.h"
61 #include "llvm/Support/ErrorHandling.h"
62 #include "llvm/Support/raw_ostream.h"
63 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
64 #include "llvm/Transforms/Utils/Local.h"
65 #include <cassert>
66 #include <cstdint>
67 #include <iterator>
68 #include <map>
69 #include <set>
70 #include <utility>
71 #include <vector>
72 
73 using namespace llvm;
74 using namespace llvm::PatternMatch;
75 using ProfileCount = Function::ProfileCount;
76 
77 #define DEBUG_TYPE "code-extractor"
78 
79 // Provide a command-line option to aggregate function arguments into a struct
80 // for functions produced by the code extractor. This is useful when converting
81 // extracted functions to pthread-based code, as only one argument (void*) can
82 // be passed in to pthread_create().
83 static cl::opt<bool>
84 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
85                  cl::desc("Aggregate arguments to code-extracted functions"));
86 
87 /// Test whether a block is valid for extraction.
isBlockValidForExtraction(const BasicBlock & BB,const SetVector<BasicBlock * > & Result,bool AllowVarArgs,bool AllowAlloca)88 static bool isBlockValidForExtraction(const BasicBlock &BB,
89                                       const SetVector<BasicBlock *> &Result,
90                                       bool AllowVarArgs, bool AllowAlloca) {
91   // taking the address of a basic block moved to another function is illegal
92   if (BB.hasAddressTaken())
93     return false;
94 
95   // don't hoist code that uses another basicblock address, as it's likely to
96   // lead to unexpected behavior, like cross-function jumps
97   SmallPtrSet<User const *, 16> Visited;
98   SmallVector<User const *, 16> ToVisit;
99 
100   for (Instruction const &Inst : BB)
101     ToVisit.push_back(&Inst);
102 
103   while (!ToVisit.empty()) {
104     User const *Curr = ToVisit.pop_back_val();
105     if (!Visited.insert(Curr).second)
106       continue;
107     if (isa<BlockAddress const>(Curr))
108       return false; // even a reference to self is likely to be not compatible
109 
110     if (isa<Instruction>(Curr) && cast<Instruction>(Curr)->getParent() != &BB)
111       continue;
112 
113     for (auto const &U : Curr->operands()) {
114       if (auto *UU = dyn_cast<User>(U))
115         ToVisit.push_back(UU);
116     }
117   }
118 
119   // If explicitly requested, allow vastart and alloca. For invoke instructions
120   // verify that extraction is valid.
121   for (BasicBlock::const_iterator I = BB.begin(), E = BB.end(); I != E; ++I) {
122     if (isa<AllocaInst>(I)) {
123        if (!AllowAlloca)
124          return false;
125        continue;
126     }
127 
128     if (const auto *II = dyn_cast<InvokeInst>(I)) {
129       // Unwind destination (either a landingpad, catchswitch, or cleanuppad)
130       // must be a part of the subgraph which is being extracted.
131       if (auto *UBB = II->getUnwindDest())
132         if (!Result.count(UBB))
133           return false;
134       continue;
135     }
136 
137     // All catch handlers of a catchswitch instruction as well as the unwind
138     // destination must be in the subgraph.
139     if (const auto *CSI = dyn_cast<CatchSwitchInst>(I)) {
140       if (auto *UBB = CSI->getUnwindDest())
141         if (!Result.count(UBB))
142           return false;
143       for (auto *HBB : CSI->handlers())
144         if (!Result.count(const_cast<BasicBlock*>(HBB)))
145           return false;
146       continue;
147     }
148 
149     // Make sure that entire catch handler is within subgraph. It is sufficient
150     // to check that catch return's block is in the list.
151     if (const auto *CPI = dyn_cast<CatchPadInst>(I)) {
152       for (const auto *U : CPI->users())
153         if (const auto *CRI = dyn_cast<CatchReturnInst>(U))
154           if (!Result.count(const_cast<BasicBlock*>(CRI->getParent())))
155             return false;
156       continue;
157     }
158 
159     // And do similar checks for cleanup handler - the entire handler must be
160     // in subgraph which is going to be extracted. For cleanup return should
161     // additionally check that the unwind destination is also in the subgraph.
162     if (const auto *CPI = dyn_cast<CleanupPadInst>(I)) {
163       for (const auto *U : CPI->users())
164         if (const auto *CRI = dyn_cast<CleanupReturnInst>(U))
165           if (!Result.count(const_cast<BasicBlock*>(CRI->getParent())))
166             return false;
167       continue;
168     }
169     if (const auto *CRI = dyn_cast<CleanupReturnInst>(I)) {
170       if (auto *UBB = CRI->getUnwindDest())
171         if (!Result.count(UBB))
172           return false;
173       continue;
174     }
175 
176     if (const CallInst *CI = dyn_cast<CallInst>(I)) {
177       if (const Function *F = CI->getCalledFunction()) {
178         auto IID = F->getIntrinsicID();
179         if (IID == Intrinsic::vastart) {
180           if (AllowVarArgs)
181             continue;
182           else
183             return false;
184         }
185 
186         // Currently, we miscompile outlined copies of eh_typid_for. There are
187         // proposals for fixing this in llvm.org/PR39545.
188         if (IID == Intrinsic::eh_typeid_for)
189           return false;
190       }
191     }
192   }
193 
194   return true;
195 }
196 
197 /// Build a set of blocks to extract if the input blocks are viable.
198 static SetVector<BasicBlock *>
buildExtractionBlockSet(ArrayRef<BasicBlock * > BBs,DominatorTree * DT,bool AllowVarArgs,bool AllowAlloca)199 buildExtractionBlockSet(ArrayRef<BasicBlock *> BBs, DominatorTree *DT,
200                         bool AllowVarArgs, bool AllowAlloca) {
201   assert(!BBs.empty() && "The set of blocks to extract must be non-empty");
202   SetVector<BasicBlock *> Result;
203 
204   // Loop over the blocks, adding them to our set-vector, and aborting with an
205   // empty set if we encounter invalid blocks.
206   for (BasicBlock *BB : BBs) {
207     // If this block is dead, don't process it.
208     if (DT && !DT->isReachableFromEntry(BB))
209       continue;
210 
211     if (!Result.insert(BB))
212       llvm_unreachable("Repeated basic blocks in extraction input");
213   }
214 
215   LLVM_DEBUG(dbgs() << "Region front block: " << Result.front()->getName()
216                     << '\n');
217 
218   for (auto *BB : Result) {
219     if (!isBlockValidForExtraction(*BB, Result, AllowVarArgs, AllowAlloca))
220       return {};
221 
222     // Make sure that the first block is not a landing pad.
223     if (BB == Result.front()) {
224       if (BB->isEHPad()) {
225         LLVM_DEBUG(dbgs() << "The first block cannot be an unwind block\n");
226         return {};
227       }
228       continue;
229     }
230 
231     // All blocks other than the first must not have predecessors outside of
232     // the subgraph which is being extracted.
233     for (auto *PBB : predecessors(BB))
234       if (!Result.count(PBB)) {
235         LLVM_DEBUG(dbgs() << "No blocks in this region may have entries from "
236                              "outside the region except for the first block!\n"
237                           << "Problematic source BB: " << BB->getName() << "\n"
238                           << "Problematic destination BB: " << PBB->getName()
239                           << "\n");
240         return {};
241       }
242   }
243 
244   return Result;
245 }
246 
CodeExtractor(ArrayRef<BasicBlock * > BBs,DominatorTree * DT,bool AggregateArgs,BlockFrequencyInfo * BFI,BranchProbabilityInfo * BPI,AssumptionCache * AC,bool AllowVarArgs,bool AllowAlloca,std::string Suffix)247 CodeExtractor::CodeExtractor(ArrayRef<BasicBlock *> BBs, DominatorTree *DT,
248                              bool AggregateArgs, BlockFrequencyInfo *BFI,
249                              BranchProbabilityInfo *BPI, AssumptionCache *AC,
250                              bool AllowVarArgs, bool AllowAlloca,
251                              std::string Suffix)
252     : DT(DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI),
253       BPI(BPI), AC(AC), AllowVarArgs(AllowVarArgs),
254       Blocks(buildExtractionBlockSet(BBs, DT, AllowVarArgs, AllowAlloca)),
255       Suffix(Suffix) {}
256 
CodeExtractor(DominatorTree & DT,Loop & L,bool AggregateArgs,BlockFrequencyInfo * BFI,BranchProbabilityInfo * BPI,AssumptionCache * AC,std::string Suffix)257 CodeExtractor::CodeExtractor(DominatorTree &DT, Loop &L, bool AggregateArgs,
258                              BlockFrequencyInfo *BFI,
259                              BranchProbabilityInfo *BPI, AssumptionCache *AC,
260                              std::string Suffix)
261     : DT(&DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI),
262       BPI(BPI), AC(AC), AllowVarArgs(false),
263       Blocks(buildExtractionBlockSet(L.getBlocks(), &DT,
264                                      /* AllowVarArgs */ false,
265                                      /* AllowAlloca */ false)),
266       Suffix(Suffix) {}
267 
268 /// definedInRegion - Return true if the specified value is defined in the
269 /// extracted region.
definedInRegion(const SetVector<BasicBlock * > & Blocks,Value * V)270 static bool definedInRegion(const SetVector<BasicBlock *> &Blocks, Value *V) {
271   if (Instruction *I = dyn_cast<Instruction>(V))
272     if (Blocks.count(I->getParent()))
273       return true;
274   return false;
275 }
276 
277 /// definedInCaller - Return true if the specified value is defined in the
278 /// function being code extracted, but not in the region being extracted.
279 /// These values must be passed in as live-ins to the function.
definedInCaller(const SetVector<BasicBlock * > & Blocks,Value * V)280 static bool definedInCaller(const SetVector<BasicBlock *> &Blocks, Value *V) {
281   if (isa<Argument>(V)) return true;
282   if (Instruction *I = dyn_cast<Instruction>(V))
283     if (!Blocks.count(I->getParent()))
284       return true;
285   return false;
286 }
287 
getCommonExitBlock(const SetVector<BasicBlock * > & Blocks)288 static BasicBlock *getCommonExitBlock(const SetVector<BasicBlock *> &Blocks) {
289   BasicBlock *CommonExitBlock = nullptr;
290   auto hasNonCommonExitSucc = [&](BasicBlock *Block) {
291     for (auto *Succ : successors(Block)) {
292       // Internal edges, ok.
293       if (Blocks.count(Succ))
294         continue;
295       if (!CommonExitBlock) {
296         CommonExitBlock = Succ;
297         continue;
298       }
299       if (CommonExitBlock != Succ)
300         return true;
301     }
302     return false;
303   };
304 
305   if (any_of(Blocks, hasNonCommonExitSucc))
306     return nullptr;
307 
308   return CommonExitBlock;
309 }
310 
CodeExtractorAnalysisCache(Function & F)311 CodeExtractorAnalysisCache::CodeExtractorAnalysisCache(Function &F) {
312   for (BasicBlock &BB : F) {
313     for (Instruction &II : BB.instructionsWithoutDebug())
314       if (auto *AI = dyn_cast<AllocaInst>(&II))
315         Allocas.push_back(AI);
316 
317     findSideEffectInfoForBlock(BB);
318   }
319 }
320 
findSideEffectInfoForBlock(BasicBlock & BB)321 void CodeExtractorAnalysisCache::findSideEffectInfoForBlock(BasicBlock &BB) {
322   for (Instruction &II : BB.instructionsWithoutDebug()) {
323     unsigned Opcode = II.getOpcode();
324     Value *MemAddr = nullptr;
325     switch (Opcode) {
326     case Instruction::Store:
327     case Instruction::Load: {
328       if (Opcode == Instruction::Store) {
329         StoreInst *SI = cast<StoreInst>(&II);
330         MemAddr = SI->getPointerOperand();
331       } else {
332         LoadInst *LI = cast<LoadInst>(&II);
333         MemAddr = LI->getPointerOperand();
334       }
335       // Global variable can not be aliased with locals.
336       if (dyn_cast<Constant>(MemAddr))
337         break;
338       Value *Base = MemAddr->stripInBoundsConstantOffsets();
339       if (!isa<AllocaInst>(Base)) {
340         SideEffectingBlocks.insert(&BB);
341         return;
342       }
343       BaseMemAddrs[&BB].insert(Base);
344       break;
345     }
346     default: {
347       IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(&II);
348       if (IntrInst) {
349         if (IntrInst->isLifetimeStartOrEnd())
350           break;
351         SideEffectingBlocks.insert(&BB);
352         return;
353       }
354       // Treat all the other cases conservatively if it has side effects.
355       if (II.mayHaveSideEffects()) {
356         SideEffectingBlocks.insert(&BB);
357         return;
358       }
359     }
360     }
361   }
362 }
363 
doesBlockContainClobberOfAddr(BasicBlock & BB,AllocaInst * Addr) const364 bool CodeExtractorAnalysisCache::doesBlockContainClobberOfAddr(
365     BasicBlock &BB, AllocaInst *Addr) const {
366   if (SideEffectingBlocks.count(&BB))
367     return true;
368   auto It = BaseMemAddrs.find(&BB);
369   if (It != BaseMemAddrs.end())
370     return It->second.count(Addr);
371   return false;
372 }
373 
isLegalToShrinkwrapLifetimeMarkers(const CodeExtractorAnalysisCache & CEAC,Instruction * Addr) const374 bool CodeExtractor::isLegalToShrinkwrapLifetimeMarkers(
375     const CodeExtractorAnalysisCache &CEAC, Instruction *Addr) const {
376   AllocaInst *AI = cast<AllocaInst>(Addr->stripInBoundsConstantOffsets());
377   Function *Func = (*Blocks.begin())->getParent();
378   for (BasicBlock &BB : *Func) {
379     if (Blocks.count(&BB))
380       continue;
381     if (CEAC.doesBlockContainClobberOfAddr(BB, AI))
382       return false;
383   }
384   return true;
385 }
386 
387 BasicBlock *
findOrCreateBlockForHoisting(BasicBlock * CommonExitBlock)388 CodeExtractor::findOrCreateBlockForHoisting(BasicBlock *CommonExitBlock) {
389   BasicBlock *SinglePredFromOutlineRegion = nullptr;
390   assert(!Blocks.count(CommonExitBlock) &&
391          "Expect a block outside the region!");
392   for (auto *Pred : predecessors(CommonExitBlock)) {
393     if (!Blocks.count(Pred))
394       continue;
395     if (!SinglePredFromOutlineRegion) {
396       SinglePredFromOutlineRegion = Pred;
397     } else if (SinglePredFromOutlineRegion != Pred) {
398       SinglePredFromOutlineRegion = nullptr;
399       break;
400     }
401   }
402 
403   if (SinglePredFromOutlineRegion)
404     return SinglePredFromOutlineRegion;
405 
406 #ifndef NDEBUG
407   auto getFirstPHI = [](BasicBlock *BB) {
408     BasicBlock::iterator I = BB->begin();
409     PHINode *FirstPhi = nullptr;
410     while (I != BB->end()) {
411       PHINode *Phi = dyn_cast<PHINode>(I);
412       if (!Phi)
413         break;
414       if (!FirstPhi) {
415         FirstPhi = Phi;
416         break;
417       }
418     }
419     return FirstPhi;
420   };
421   // If there are any phi nodes, the single pred either exists or has already
422   // be created before code extraction.
423   assert(!getFirstPHI(CommonExitBlock) && "Phi not expected");
424 #endif
425 
426   BasicBlock *NewExitBlock = CommonExitBlock->splitBasicBlock(
427       CommonExitBlock->getFirstNonPHI()->getIterator());
428 
429   for (auto PI = pred_begin(CommonExitBlock), PE = pred_end(CommonExitBlock);
430        PI != PE;) {
431     BasicBlock *Pred = *PI++;
432     if (Blocks.count(Pred))
433       continue;
434     Pred->getTerminator()->replaceUsesOfWith(CommonExitBlock, NewExitBlock);
435   }
436   // Now add the old exit block to the outline region.
437   Blocks.insert(CommonExitBlock);
438   return CommonExitBlock;
439 }
440 
441 // Find the pair of life time markers for address 'Addr' that are either
442 // defined inside the outline region or can legally be shrinkwrapped into the
443 // outline region. If there are not other untracked uses of the address, return
444 // the pair of markers if found; otherwise return a pair of nullptr.
445 CodeExtractor::LifetimeMarkerInfo
getLifetimeMarkers(const CodeExtractorAnalysisCache & CEAC,Instruction * Addr,BasicBlock * ExitBlock) const446 CodeExtractor::getLifetimeMarkers(const CodeExtractorAnalysisCache &CEAC,
447                                   Instruction *Addr,
448                                   BasicBlock *ExitBlock) const {
449   LifetimeMarkerInfo Info;
450 
451   for (User *U : Addr->users()) {
452     IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(U);
453     if (IntrInst) {
454       // We don't model addresses with multiple start/end markers, but the
455       // markers do not need to be in the region.
456       if (IntrInst->getIntrinsicID() == Intrinsic::lifetime_start) {
457         if (Info.LifeStart)
458           return {};
459         Info.LifeStart = IntrInst;
460         continue;
461       }
462       if (IntrInst->getIntrinsicID() == Intrinsic::lifetime_end) {
463         if (Info.LifeEnd)
464           return {};
465         Info.LifeEnd = IntrInst;
466         continue;
467       }
468       // At this point, permit debug uses outside of the region.
469       // This is fixed in a later call to fixupDebugInfoPostExtraction().
470       if (isa<DbgInfoIntrinsic>(IntrInst))
471         continue;
472     }
473     // Find untracked uses of the address, bail.
474     if (!definedInRegion(Blocks, U))
475       return {};
476   }
477 
478   if (!Info.LifeStart || !Info.LifeEnd)
479     return {};
480 
481   Info.SinkLifeStart = !definedInRegion(Blocks, Info.LifeStart);
482   Info.HoistLifeEnd = !definedInRegion(Blocks, Info.LifeEnd);
483   // Do legality check.
484   if ((Info.SinkLifeStart || Info.HoistLifeEnd) &&
485       !isLegalToShrinkwrapLifetimeMarkers(CEAC, Addr))
486     return {};
487 
488   // Check to see if we have a place to do hoisting, if not, bail.
489   if (Info.HoistLifeEnd && !ExitBlock)
490     return {};
491 
492   return Info;
493 }
494 
findAllocas(const CodeExtractorAnalysisCache & CEAC,ValueSet & SinkCands,ValueSet & HoistCands,BasicBlock * & ExitBlock) const495 void CodeExtractor::findAllocas(const CodeExtractorAnalysisCache &CEAC,
496                                 ValueSet &SinkCands, ValueSet &HoistCands,
497                                 BasicBlock *&ExitBlock) const {
498   Function *Func = (*Blocks.begin())->getParent();
499   ExitBlock = getCommonExitBlock(Blocks);
500 
501   auto moveOrIgnoreLifetimeMarkers =
502       [&](const LifetimeMarkerInfo &LMI) -> bool {
503     if (!LMI.LifeStart)
504       return false;
505     if (LMI.SinkLifeStart) {
506       LLVM_DEBUG(dbgs() << "Sinking lifetime.start: " << *LMI.LifeStart
507                         << "\n");
508       SinkCands.insert(LMI.LifeStart);
509     }
510     if (LMI.HoistLifeEnd) {
511       LLVM_DEBUG(dbgs() << "Hoisting lifetime.end: " << *LMI.LifeEnd << "\n");
512       HoistCands.insert(LMI.LifeEnd);
513     }
514     return true;
515   };
516 
517   // Look up allocas in the original function in CodeExtractorAnalysisCache, as
518   // this is much faster than walking all the instructions.
519   for (AllocaInst *AI : CEAC.getAllocas()) {
520     BasicBlock *BB = AI->getParent();
521     if (Blocks.count(BB))
522       continue;
523 
524     // As a prior call to extractCodeRegion() may have shrinkwrapped the alloca,
525     // check whether it is actually still in the original function.
526     Function *AIFunc = BB->getParent();
527     if (AIFunc != Func)
528       continue;
529 
530     LifetimeMarkerInfo MarkerInfo = getLifetimeMarkers(CEAC, AI, ExitBlock);
531     bool Moved = moveOrIgnoreLifetimeMarkers(MarkerInfo);
532     if (Moved) {
533       LLVM_DEBUG(dbgs() << "Sinking alloca: " << *AI << "\n");
534       SinkCands.insert(AI);
535       continue;
536     }
537 
538     // Find bitcasts in the outlined region that have lifetime marker users
539     // outside that region. Replace the lifetime marker use with an
540     // outside region bitcast to avoid unnecessary alloca/reload instructions
541     // and extra lifetime markers.
542     SmallVector<Instruction *, 2> LifetimeBitcastUsers;
543     for (User *U : AI->users()) {
544       if (!definedInRegion(Blocks, U))
545         continue;
546 
547       if (U->stripInBoundsConstantOffsets() != AI)
548         continue;
549 
550       Instruction *Bitcast = cast<Instruction>(U);
551       for (User *BU : Bitcast->users()) {
552         IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(BU);
553         if (!IntrInst)
554           continue;
555 
556         if (!IntrInst->isLifetimeStartOrEnd())
557           continue;
558 
559         if (definedInRegion(Blocks, IntrInst))
560           continue;
561 
562         LLVM_DEBUG(dbgs() << "Replace use of extracted region bitcast"
563                           << *Bitcast << " in out-of-region lifetime marker "
564                           << *IntrInst << "\n");
565         LifetimeBitcastUsers.push_back(IntrInst);
566       }
567     }
568 
569     for (Instruction *I : LifetimeBitcastUsers) {
570       Module *M = AIFunc->getParent();
571       LLVMContext &Ctx = M->getContext();
572       auto *Int8PtrTy = Type::getInt8PtrTy(Ctx);
573       CastInst *CastI =
574           CastInst::CreatePointerCast(AI, Int8PtrTy, "lt.cast", I);
575       I->replaceUsesOfWith(I->getOperand(1), CastI);
576     }
577 
578     // Follow any bitcasts.
579     SmallVector<Instruction *, 2> Bitcasts;
580     SmallVector<LifetimeMarkerInfo, 2> BitcastLifetimeInfo;
581     for (User *U : AI->users()) {
582       if (U->stripInBoundsConstantOffsets() == AI) {
583         Instruction *Bitcast = cast<Instruction>(U);
584         LifetimeMarkerInfo LMI = getLifetimeMarkers(CEAC, Bitcast, ExitBlock);
585         if (LMI.LifeStart) {
586           Bitcasts.push_back(Bitcast);
587           BitcastLifetimeInfo.push_back(LMI);
588           continue;
589         }
590       }
591 
592       // Found unknown use of AI.
593       if (!definedInRegion(Blocks, U)) {
594         Bitcasts.clear();
595         break;
596       }
597     }
598 
599     // Either no bitcasts reference the alloca or there are unknown uses.
600     if (Bitcasts.empty())
601       continue;
602 
603     LLVM_DEBUG(dbgs() << "Sinking alloca (via bitcast): " << *AI << "\n");
604     SinkCands.insert(AI);
605     for (unsigned I = 0, E = Bitcasts.size(); I != E; ++I) {
606       Instruction *BitcastAddr = Bitcasts[I];
607       const LifetimeMarkerInfo &LMI = BitcastLifetimeInfo[I];
608       assert(LMI.LifeStart &&
609              "Unsafe to sink bitcast without lifetime markers");
610       moveOrIgnoreLifetimeMarkers(LMI);
611       if (!definedInRegion(Blocks, BitcastAddr)) {
612         LLVM_DEBUG(dbgs() << "Sinking bitcast-of-alloca: " << *BitcastAddr
613                           << "\n");
614         SinkCands.insert(BitcastAddr);
615       }
616     }
617   }
618 }
619 
isEligible() const620 bool CodeExtractor::isEligible() const {
621   if (Blocks.empty())
622     return false;
623   BasicBlock *Header = *Blocks.begin();
624   Function *F = Header->getParent();
625 
626   // For functions with varargs, check that varargs handling is only done in the
627   // outlined function, i.e vastart and vaend are only used in outlined blocks.
628   if (AllowVarArgs && F->getFunctionType()->isVarArg()) {
629     auto containsVarArgIntrinsic = [](const Instruction &I) {
630       if (const CallInst *CI = dyn_cast<CallInst>(&I))
631         if (const Function *Callee = CI->getCalledFunction())
632           return Callee->getIntrinsicID() == Intrinsic::vastart ||
633                  Callee->getIntrinsicID() == Intrinsic::vaend;
634       return false;
635     };
636 
637     for (auto &BB : *F) {
638       if (Blocks.count(&BB))
639         continue;
640       if (llvm::any_of(BB, containsVarArgIntrinsic))
641         return false;
642     }
643   }
644   return true;
645 }
646 
findInputsOutputs(ValueSet & Inputs,ValueSet & Outputs,const ValueSet & SinkCands) const647 void CodeExtractor::findInputsOutputs(ValueSet &Inputs, ValueSet &Outputs,
648                                       const ValueSet &SinkCands) const {
649   for (BasicBlock *BB : Blocks) {
650     // If a used value is defined outside the region, it's an input.  If an
651     // instruction is used outside the region, it's an output.
652     for (Instruction &II : *BB) {
653       for (auto &OI : II.operands()) {
654         Value *V = OI;
655         if (!SinkCands.count(V) && definedInCaller(Blocks, V))
656           Inputs.insert(V);
657       }
658 
659       for (User *U : II.users())
660         if (!definedInRegion(Blocks, U)) {
661           Outputs.insert(&II);
662           break;
663         }
664     }
665   }
666 }
667 
668 /// severSplitPHINodesOfEntry - If a PHI node has multiple inputs from outside
669 /// of the region, we need to split the entry block of the region so that the
670 /// PHI node is easier to deal with.
severSplitPHINodesOfEntry(BasicBlock * & Header)671 void CodeExtractor::severSplitPHINodesOfEntry(BasicBlock *&Header) {
672   unsigned NumPredsFromRegion = 0;
673   unsigned NumPredsOutsideRegion = 0;
674 
675   if (Header != &Header->getParent()->getEntryBlock()) {
676     PHINode *PN = dyn_cast<PHINode>(Header->begin());
677     if (!PN) return;  // No PHI nodes.
678 
679     // If the header node contains any PHI nodes, check to see if there is more
680     // than one entry from outside the region.  If so, we need to sever the
681     // header block into two.
682     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
683       if (Blocks.count(PN->getIncomingBlock(i)))
684         ++NumPredsFromRegion;
685       else
686         ++NumPredsOutsideRegion;
687 
688     // If there is one (or fewer) predecessor from outside the region, we don't
689     // need to do anything special.
690     if (NumPredsOutsideRegion <= 1) return;
691   }
692 
693   // Otherwise, we need to split the header block into two pieces: one
694   // containing PHI nodes merging values from outside of the region, and a
695   // second that contains all of the code for the block and merges back any
696   // incoming values from inside of the region.
697   BasicBlock *NewBB = SplitBlock(Header, Header->getFirstNonPHI(), DT);
698 
699   // We only want to code extract the second block now, and it becomes the new
700   // header of the region.
701   BasicBlock *OldPred = Header;
702   Blocks.remove(OldPred);
703   Blocks.insert(NewBB);
704   Header = NewBB;
705 
706   // Okay, now we need to adjust the PHI nodes and any branches from within the
707   // region to go to the new header block instead of the old header block.
708   if (NumPredsFromRegion) {
709     PHINode *PN = cast<PHINode>(OldPred->begin());
710     // Loop over all of the predecessors of OldPred that are in the region,
711     // changing them to branch to NewBB instead.
712     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
713       if (Blocks.count(PN->getIncomingBlock(i))) {
714         Instruction *TI = PN->getIncomingBlock(i)->getTerminator();
715         TI->replaceUsesOfWith(OldPred, NewBB);
716       }
717 
718     // Okay, everything within the region is now branching to the right block, we
719     // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
720     BasicBlock::iterator AfterPHIs;
721     for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
722       PHINode *PN = cast<PHINode>(AfterPHIs);
723       // Create a new PHI node in the new region, which has an incoming value
724       // from OldPred of PN.
725       PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion,
726                                        PN->getName() + ".ce", &NewBB->front());
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.
severSplitPHINodesOfExits(const SmallPtrSetImpl<BasicBlock * > & Exits)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         SmallVector<BasicBlock *, 4> Preds(predecessors(ExitBB));
772         for (BasicBlock *PredBB : Preds)
773           if (Blocks.count(PredBB))
774             PredBB->getTerminator()->replaceUsesOfWith(ExitBB, NewBB);
775         BranchInst::Create(ExitBB, NewBB);
776         Blocks.insert(NewBB);
777       }
778 
779       // Split this PHI.
780       PHINode *NewPN =
781           PHINode::Create(PN.getType(), IncomingVals.size(),
782                           PN.getName() + ".ce", NewBB->getFirstNonPHI());
783       for (unsigned i : IncomingVals)
784         NewPN->addIncoming(PN.getIncomingValue(i), PN.getIncomingBlock(i));
785       for (unsigned i : reverse(IncomingVals))
786         PN.removeIncomingValue(i, false);
787       PN.addIncoming(NewPN, NewBB);
788     }
789   }
790 }
791 
splitReturnBlocks()792 void CodeExtractor::splitReturnBlocks() {
793   for (BasicBlock *Block : Blocks)
794     if (ReturnInst *RI = dyn_cast<ReturnInst>(Block->getTerminator())) {
795       BasicBlock *New =
796           Block->splitBasicBlock(RI->getIterator(), Block->getName() + ".ret");
797       if (DT) {
798         // Old dominates New. New node dominates all other nodes dominated
799         // by Old.
800         DomTreeNode *OldNode = DT->getNode(Block);
801         SmallVector<DomTreeNode *, 8> Children(OldNode->begin(),
802                                                OldNode->end());
803 
804         DomTreeNode *NewNode = DT->addNewBlock(New, Block);
805 
806         for (DomTreeNode *I : Children)
807           DT->changeImmediateDominator(I, NewNode);
808       }
809     }
810 }
811 
812 /// constructFunction - make a function based on inputs and outputs, as follows:
813 /// f(in0, ..., inN, out0, ..., outN)
constructFunction(const ValueSet & inputs,const ValueSet & outputs,BasicBlock * header,BasicBlock * newRootNode,BasicBlock * newHeader,Function * oldFunction,Module * M)814 Function *CodeExtractor::constructFunction(const ValueSet &inputs,
815                                            const ValueSet &outputs,
816                                            BasicBlock *header,
817                                            BasicBlock *newRootNode,
818                                            BasicBlock *newHeader,
819                                            Function *oldFunction,
820                                            Module *M) {
821   LLVM_DEBUG(dbgs() << "inputs: " << inputs.size() << "\n");
822   LLVM_DEBUG(dbgs() << "outputs: " << outputs.size() << "\n");
823 
824   // This function returns unsigned, outputs will go back by reference.
825   switch (NumExitBlocks) {
826   case 0:
827   case 1: RetTy = Type::getVoidTy(header->getContext()); break;
828   case 2: RetTy = Type::getInt1Ty(header->getContext()); break;
829   default: RetTy = Type::getInt16Ty(header->getContext()); break;
830   }
831 
832   std::vector<Type *> paramTy;
833 
834   // Add the types of the input values to the function's argument list
835   for (Value *value : inputs) {
836     LLVM_DEBUG(dbgs() << "value used in func: " << *value << "\n");
837     paramTy.push_back(value->getType());
838   }
839 
840   // Add the types of the output values to the function's argument list.
841   for (Value *output : outputs) {
842     LLVM_DEBUG(dbgs() << "instr used in func: " << *output << "\n");
843     if (AggregateArgs)
844       paramTy.push_back(output->getType());
845     else
846       paramTy.push_back(PointerType::getUnqual(output->getType()));
847   }
848 
849   LLVM_DEBUG({
850     dbgs() << "Function type: " << *RetTy << " f(";
851     for (Type *i : paramTy)
852       dbgs() << *i << ", ";
853     dbgs() << ")\n";
854   });
855 
856   StructType *StructTy = nullptr;
857   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
858     StructTy = StructType::get(M->getContext(), paramTy);
859     paramTy.clear();
860     paramTy.push_back(PointerType::getUnqual(StructTy));
861   }
862   FunctionType *funcType =
863                   FunctionType::get(RetTy, paramTy,
864                                     AllowVarArgs && oldFunction->isVarArg());
865 
866   std::string SuffixToUse =
867       Suffix.empty()
868           ? (header->getName().empty() ? "extracted" : header->getName().str())
869           : Suffix;
870   // Create the new function
871   Function *newFunction = Function::Create(
872       funcType, GlobalValue::InternalLinkage, oldFunction->getAddressSpace(),
873       oldFunction->getName() + "." + SuffixToUse, M);
874   // If the old function is no-throw, so is the new one.
875   if (oldFunction->doesNotThrow())
876     newFunction->setDoesNotThrow();
877 
878   // Inherit the uwtable attribute if we need to.
879   if (oldFunction->hasUWTable())
880     newFunction->setHasUWTable();
881 
882   // Inherit all of the target dependent attributes and white-listed
883   // target independent attributes.
884   //  (e.g. If the extracted region contains a call to an x86.sse
885   //  instruction we need to make sure that the extracted region has the
886   //  "target-features" attribute allowing it to be lowered.
887   // FIXME: This should be changed to check to see if a specific
888   //           attribute can not be inherited.
889   for (const auto &Attr : oldFunction->getAttributes().getFnAttributes()) {
890     if (Attr.isStringAttribute()) {
891       if (Attr.getKindAsString() == "thunk")
892         continue;
893     } else
894       switch (Attr.getKindAsEnum()) {
895       // Those attributes cannot be propagated safely. Explicitly list them
896       // here so we get a warning if new attributes are added. This list also
897       // includes non-function attributes.
898       case Attribute::Alignment:
899       case Attribute::AllocSize:
900       case Attribute::ArgMemOnly:
901       case Attribute::Builtin:
902       case Attribute::ByVal:
903       case Attribute::Convergent:
904       case Attribute::Dereferenceable:
905       case Attribute::DereferenceableOrNull:
906       case Attribute::InAlloca:
907       case Attribute::InReg:
908       case Attribute::InaccessibleMemOnly:
909       case Attribute::InaccessibleMemOrArgMemOnly:
910       case Attribute::JumpTable:
911       case Attribute::Naked:
912       case Attribute::Nest:
913       case Attribute::NoAlias:
914       case Attribute::NoBuiltin:
915       case Attribute::NoCapture:
916       case Attribute::NoMerge:
917       case Attribute::NoReturn:
918       case Attribute::NoSync:
919       case Attribute::NoUndef:
920       case Attribute::None:
921       case Attribute::NonNull:
922       case Attribute::Preallocated:
923       case Attribute::ReadNone:
924       case Attribute::ReadOnly:
925       case Attribute::Returned:
926       case Attribute::ReturnsTwice:
927       case Attribute::SExt:
928       case Attribute::Speculatable:
929       case Attribute::StackAlignment:
930       case Attribute::StructRet:
931       case Attribute::SwiftError:
932       case Attribute::SwiftSelf:
933       case Attribute::WillReturn:
934       case Attribute::WriteOnly:
935       case Attribute::ZExt:
936       case Attribute::ImmArg:
937       case Attribute::ByRef:
938       case Attribute::EndAttrKinds:
939       case Attribute::EmptyKey:
940       case Attribute::TombstoneKey:
941         continue;
942       // Those attributes should be safe to propagate to the extracted function.
943       case Attribute::AlwaysInline:
944       case Attribute::Cold:
945       case Attribute::Hot:
946       case Attribute::NoRecurse:
947       case Attribute::InlineHint:
948       case Attribute::MinSize:
949       case Attribute::NoCallback:
950       case Attribute::NoDuplicate:
951       case Attribute::NoFree:
952       case Attribute::NoImplicitFloat:
953       case Attribute::NoInline:
954       case Attribute::NonLazyBind:
955       case Attribute::NoRedZone:
956       case Attribute::NoUnwind:
957       case Attribute::NullPointerIsValid:
958       case Attribute::OptForFuzzing:
959       case Attribute::OptimizeNone:
960       case Attribute::OptimizeForSize:
961       case Attribute::SafeStack:
962       case Attribute::ShadowCallStack:
963       case Attribute::SanitizeAddress:
964       case Attribute::SanitizeMemory:
965       case Attribute::SanitizeThread:
966       case Attribute::SanitizeHWAddress:
967       case Attribute::SanitizeMemTag:
968       case Attribute::SpeculativeLoadHardening:
969       case Attribute::StackProtect:
970       case Attribute::StackProtectReq:
971       case Attribute::StackProtectStrong:
972       case Attribute::StrictFP:
973       case Attribute::UWTable:
974       case Attribute::NoCfCheck:
975       case Attribute::MustProgress:
976       case Attribute::NoProfile:
977         break;
978       }
979 
980     newFunction->addFnAttr(Attr);
981   }
982   newFunction->getBasicBlockList().push_back(newRootNode);
983 
984   // Create an iterator to name all of the arguments we inserted.
985   Function::arg_iterator AI = newFunction->arg_begin();
986 
987   // Rewrite all users of the inputs in the extracted region to use the
988   // arguments (or appropriate addressing into struct) instead.
989   for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
990     Value *RewriteVal;
991     if (AggregateArgs) {
992       Value *Idx[2];
993       Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext()));
994       Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i);
995       Instruction *TI = newFunction->begin()->getTerminator();
996       GetElementPtrInst *GEP = GetElementPtrInst::Create(
997           StructTy, &*AI, Idx, "gep_" + inputs[i]->getName(), TI);
998       RewriteVal = new LoadInst(StructTy->getElementType(i), GEP,
999                                 "loadgep_" + inputs[i]->getName(), TI);
1000     } else
1001       RewriteVal = &*AI++;
1002 
1003     std::vector<User *> Users(inputs[i]->user_begin(), inputs[i]->user_end());
1004     for (User *use : Users)
1005       if (Instruction *inst = dyn_cast<Instruction>(use))
1006         if (Blocks.count(inst->getParent()))
1007           inst->replaceUsesOfWith(inputs[i], RewriteVal);
1008   }
1009 
1010   // Set names for input and output arguments.
1011   if (!AggregateArgs) {
1012     AI = newFunction->arg_begin();
1013     for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
1014       AI->setName(inputs[i]->getName());
1015     for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
1016       AI->setName(outputs[i]->getName()+".out");
1017   }
1018 
1019   // Rewrite branches to basic blocks outside of the loop to new dummy blocks
1020   // within the new function. This must be done before we lose track of which
1021   // blocks were originally in the code region.
1022   std::vector<User *> Users(header->user_begin(), header->user_end());
1023   for (auto &U : Users)
1024     // The BasicBlock which contains the branch is not in the region
1025     // modify the branch target to a new block
1026     if (Instruction *I = dyn_cast<Instruction>(U))
1027       if (I->isTerminator() && I->getFunction() == oldFunction &&
1028           !Blocks.count(I->getParent()))
1029         I->replaceUsesOfWith(header, newHeader);
1030 
1031   return newFunction;
1032 }
1033 
1034 /// Erase lifetime.start markers which reference inputs to the extraction
1035 /// region, and insert the referenced memory into \p LifetimesStart.
1036 ///
1037 /// The extraction region is defined by a set of blocks (\p Blocks), and a set
1038 /// of allocas which will be moved from the caller function into the extracted
1039 /// function (\p SunkAllocas).
eraseLifetimeMarkersOnInputs(const SetVector<BasicBlock * > & Blocks,const SetVector<Value * > & SunkAllocas,SetVector<Value * > & LifetimesStart)1040 static void eraseLifetimeMarkersOnInputs(const SetVector<BasicBlock *> &Blocks,
1041                                          const SetVector<Value *> &SunkAllocas,
1042                                          SetVector<Value *> &LifetimesStart) {
1043   for (BasicBlock *BB : Blocks) {
1044     for (auto It = BB->begin(), End = BB->end(); It != End;) {
1045       auto *II = dyn_cast<IntrinsicInst>(&*It);
1046       ++It;
1047       if (!II || !II->isLifetimeStartOrEnd())
1048         continue;
1049 
1050       // Get the memory operand of the lifetime marker. If the underlying
1051       // object is a sunk alloca, or is otherwise defined in the extraction
1052       // region, the lifetime marker must not be erased.
1053       Value *Mem = II->getOperand(1)->stripInBoundsOffsets();
1054       if (SunkAllocas.count(Mem) || definedInRegion(Blocks, Mem))
1055         continue;
1056 
1057       if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1058         LifetimesStart.insert(Mem);
1059       II->eraseFromParent();
1060     }
1061   }
1062 }
1063 
1064 /// Insert lifetime start/end markers surrounding the call to the new function
1065 /// for objects defined in the caller.
insertLifetimeMarkersSurroundingCall(Module * M,ArrayRef<Value * > LifetimesStart,ArrayRef<Value * > LifetimesEnd,CallInst * TheCall)1066 static void insertLifetimeMarkersSurroundingCall(
1067     Module *M, ArrayRef<Value *> LifetimesStart, ArrayRef<Value *> LifetimesEnd,
1068     CallInst *TheCall) {
1069   LLVMContext &Ctx = M->getContext();
1070   auto Int8PtrTy = Type::getInt8PtrTy(Ctx);
1071   auto NegativeOne = ConstantInt::getSigned(Type::getInt64Ty(Ctx), -1);
1072   Instruction *Term = TheCall->getParent()->getTerminator();
1073 
1074   // The memory argument to a lifetime marker must be a i8*. Cache any bitcasts
1075   // needed to satisfy this requirement so they may be reused.
1076   DenseMap<Value *, Value *> Bitcasts;
1077 
1078   // Emit lifetime markers for the pointers given in \p Objects. Insert the
1079   // markers before the call if \p InsertBefore, and after the call otherwise.
1080   auto insertMarkers = [&](Function *MarkerFunc, ArrayRef<Value *> Objects,
1081                            bool InsertBefore) {
1082     for (Value *Mem : Objects) {
1083       assert((!isa<Instruction>(Mem) || cast<Instruction>(Mem)->getFunction() ==
1084                                             TheCall->getFunction()) &&
1085              "Input memory not defined in original function");
1086       Value *&MemAsI8Ptr = Bitcasts[Mem];
1087       if (!MemAsI8Ptr) {
1088         if (Mem->getType() == Int8PtrTy)
1089           MemAsI8Ptr = Mem;
1090         else
1091           MemAsI8Ptr =
1092               CastInst::CreatePointerCast(Mem, Int8PtrTy, "lt.cast", TheCall);
1093       }
1094 
1095       auto Marker = CallInst::Create(MarkerFunc, {NegativeOne, MemAsI8Ptr});
1096       if (InsertBefore)
1097         Marker->insertBefore(TheCall);
1098       else
1099         Marker->insertBefore(Term);
1100     }
1101   };
1102 
1103   if (!LifetimesStart.empty()) {
1104     auto StartFn = llvm::Intrinsic::getDeclaration(
1105         M, llvm::Intrinsic::lifetime_start, Int8PtrTy);
1106     insertMarkers(StartFn, LifetimesStart, /*InsertBefore=*/true);
1107   }
1108 
1109   if (!LifetimesEnd.empty()) {
1110     auto EndFn = llvm::Intrinsic::getDeclaration(
1111         M, llvm::Intrinsic::lifetime_end, Int8PtrTy);
1112     insertMarkers(EndFn, LifetimesEnd, /*InsertBefore=*/false);
1113   }
1114 }
1115 
1116 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
1117 /// the call instruction, splitting any PHI nodes in the header block as
1118 /// necessary.
emitCallAndSwitchStatement(Function * newFunction,BasicBlock * codeReplacer,ValueSet & inputs,ValueSet & outputs)1119 CallInst *CodeExtractor::emitCallAndSwitchStatement(Function *newFunction,
1120                                                     BasicBlock *codeReplacer,
1121                                                     ValueSet &inputs,
1122                                                     ValueSet &outputs) {
1123   // Emit a call to the new function, passing in: *pointer to struct (if
1124   // aggregating parameters), or plan inputs and allocated memory for outputs
1125   std::vector<Value *> params, StructValues, ReloadOutputs, Reloads;
1126 
1127   Module *M = newFunction->getParent();
1128   LLVMContext &Context = M->getContext();
1129   const DataLayout &DL = M->getDataLayout();
1130   CallInst *call = nullptr;
1131 
1132   // Add inputs as params, or to be filled into the struct
1133   unsigned ArgNo = 0;
1134   SmallVector<unsigned, 1> SwiftErrorArgs;
1135   for (Value *input : inputs) {
1136     if (AggregateArgs)
1137       StructValues.push_back(input);
1138     else {
1139       params.push_back(input);
1140       if (input->isSwiftError())
1141         SwiftErrorArgs.push_back(ArgNo);
1142     }
1143     ++ArgNo;
1144   }
1145 
1146   // Create allocas for the outputs
1147   for (Value *output : outputs) {
1148     if (AggregateArgs) {
1149       StructValues.push_back(output);
1150     } else {
1151       AllocaInst *alloca =
1152         new AllocaInst(output->getType(), DL.getAllocaAddrSpace(),
1153                        nullptr, output->getName() + ".loc",
1154                        &codeReplacer->getParent()->front().front());
1155       ReloadOutputs.push_back(alloca);
1156       params.push_back(alloca);
1157     }
1158   }
1159 
1160   StructType *StructArgTy = nullptr;
1161   AllocaInst *Struct = nullptr;
1162   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
1163     std::vector<Type *> ArgTypes;
1164     for (ValueSet::iterator v = StructValues.begin(),
1165            ve = StructValues.end(); v != ve; ++v)
1166       ArgTypes.push_back((*v)->getType());
1167 
1168     // Allocate a struct at the beginning of this function
1169     StructArgTy = StructType::get(newFunction->getContext(), ArgTypes);
1170     Struct = new AllocaInst(StructArgTy, DL.getAllocaAddrSpace(), nullptr,
1171                             "structArg",
1172                             &codeReplacer->getParent()->front().front());
1173     params.push_back(Struct);
1174 
1175     for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
1176       Value *Idx[2];
1177       Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
1178       Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i);
1179       GetElementPtrInst *GEP = GetElementPtrInst::Create(
1180           StructArgTy, Struct, Idx, "gep_" + StructValues[i]->getName());
1181       codeReplacer->getInstList().push_back(GEP);
1182       new StoreInst(StructValues[i], GEP, codeReplacer);
1183     }
1184   }
1185 
1186   // Emit the call to the function
1187   call = CallInst::Create(newFunction, params,
1188                           NumExitBlocks > 1 ? "targetBlock" : "");
1189   // Add debug location to the new call, if the original function has debug
1190   // info. In that case, the terminator of the entry block of the extracted
1191   // function contains the first debug location of the extracted function,
1192   // set in extractCodeRegion.
1193   if (codeReplacer->getParent()->getSubprogram()) {
1194     if (auto DL = newFunction->getEntryBlock().getTerminator()->getDebugLoc())
1195       call->setDebugLoc(DL);
1196   }
1197   codeReplacer->getInstList().push_back(call);
1198 
1199   // Set swifterror parameter attributes.
1200   for (unsigned SwiftErrArgNo : SwiftErrorArgs) {
1201     call->addParamAttr(SwiftErrArgNo, Attribute::SwiftError);
1202     newFunction->addParamAttr(SwiftErrArgNo, Attribute::SwiftError);
1203   }
1204 
1205   Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
1206   unsigned FirstOut = inputs.size();
1207   if (!AggregateArgs)
1208     std::advance(OutputArgBegin, inputs.size());
1209 
1210   // Reload the outputs passed in by reference.
1211   for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
1212     Value *Output = nullptr;
1213     if (AggregateArgs) {
1214       Value *Idx[2];
1215       Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
1216       Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i);
1217       GetElementPtrInst *GEP = GetElementPtrInst::Create(
1218           StructArgTy, Struct, Idx, "gep_reload_" + outputs[i]->getName());
1219       codeReplacer->getInstList().push_back(GEP);
1220       Output = GEP;
1221     } else {
1222       Output = ReloadOutputs[i];
1223     }
1224     LoadInst *load = new LoadInst(outputs[i]->getType(), Output,
1225                                   outputs[i]->getName() + ".reload",
1226                                   codeReplacer);
1227     Reloads.push_back(load);
1228     std::vector<User *> Users(outputs[i]->user_begin(), outputs[i]->user_end());
1229     for (unsigned u = 0, e = Users.size(); u != e; ++u) {
1230       Instruction *inst = cast<Instruction>(Users[u]);
1231       if (!Blocks.count(inst->getParent()))
1232         inst->replaceUsesOfWith(outputs[i], load);
1233     }
1234   }
1235 
1236   // Now we can emit a switch statement using the call as a value.
1237   SwitchInst *TheSwitch =
1238       SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)),
1239                          codeReplacer, 0, codeReplacer);
1240 
1241   // Since there may be multiple exits from the original region, make the new
1242   // function return an unsigned, switch on that number.  This loop iterates
1243   // over all of the blocks in the extracted region, updating any terminator
1244   // instructions in the to-be-extracted region that branch to blocks that are
1245   // not in the region to be extracted.
1246   std::map<BasicBlock *, BasicBlock *> ExitBlockMap;
1247 
1248   unsigned switchVal = 0;
1249   for (BasicBlock *Block : Blocks) {
1250     Instruction *TI = Block->getTerminator();
1251     for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
1252       if (!Blocks.count(TI->getSuccessor(i))) {
1253         BasicBlock *OldTarget = TI->getSuccessor(i);
1254         // add a new basic block which returns the appropriate value
1255         BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
1256         if (!NewTarget) {
1257           // If we don't already have an exit stub for this non-extracted
1258           // destination, create one now!
1259           NewTarget = BasicBlock::Create(Context,
1260                                          OldTarget->getName() + ".exitStub",
1261                                          newFunction);
1262           unsigned SuccNum = switchVal++;
1263 
1264           Value *brVal = nullptr;
1265           switch (NumExitBlocks) {
1266           case 0:
1267           case 1: break;  // No value needed.
1268           case 2:         // Conditional branch, return a bool
1269             brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum);
1270             break;
1271           default:
1272             brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum);
1273             break;
1274           }
1275 
1276           ReturnInst::Create(Context, brVal, NewTarget);
1277 
1278           // Update the switch instruction.
1279           TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context),
1280                                               SuccNum),
1281                              OldTarget);
1282         }
1283 
1284         // rewrite the original branch instruction with this new target
1285         TI->setSuccessor(i, NewTarget);
1286       }
1287   }
1288 
1289   // Store the arguments right after the definition of output value.
1290   // This should be proceeded after creating exit stubs to be ensure that invoke
1291   // result restore will be placed in the outlined function.
1292   Function::arg_iterator OAI = OutputArgBegin;
1293   for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
1294     auto *OutI = dyn_cast<Instruction>(outputs[i]);
1295     if (!OutI)
1296       continue;
1297 
1298     // Find proper insertion point.
1299     BasicBlock::iterator InsertPt;
1300     // In case OutI is an invoke, we insert the store at the beginning in the
1301     // 'normal destination' BB. Otherwise we insert the store right after OutI.
1302     if (auto *InvokeI = dyn_cast<InvokeInst>(OutI))
1303       InsertPt = InvokeI->getNormalDest()->getFirstInsertionPt();
1304     else if (auto *Phi = dyn_cast<PHINode>(OutI))
1305       InsertPt = Phi->getParent()->getFirstInsertionPt();
1306     else
1307       InsertPt = std::next(OutI->getIterator());
1308 
1309     Instruction *InsertBefore = &*InsertPt;
1310     assert((InsertBefore->getFunction() == newFunction ||
1311             Blocks.count(InsertBefore->getParent())) &&
1312            "InsertPt should be in new function");
1313     assert(OAI != newFunction->arg_end() &&
1314            "Number of output arguments should match "
1315            "the amount of defined values");
1316     if (AggregateArgs) {
1317       Value *Idx[2];
1318       Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
1319       Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i);
1320       GetElementPtrInst *GEP = GetElementPtrInst::Create(
1321           StructArgTy, &*OAI, Idx, "gep_" + outputs[i]->getName(),
1322           InsertBefore);
1323       new StoreInst(outputs[i], GEP, InsertBefore);
1324       // Since there should be only one struct argument aggregating
1325       // all the output values, we shouldn't increment OAI, which always
1326       // points to the struct argument, in this case.
1327     } else {
1328       new StoreInst(outputs[i], &*OAI, InsertBefore);
1329       ++OAI;
1330     }
1331   }
1332 
1333   // Now that we've done the deed, simplify the switch instruction.
1334   Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
1335   switch (NumExitBlocks) {
1336   case 0:
1337     // There are no successors (the block containing the switch itself), which
1338     // means that previously this was the last part of the function, and hence
1339     // this should be rewritten as a `ret'
1340 
1341     // Check if the function should return a value
1342     if (OldFnRetTy->isVoidTy()) {
1343       ReturnInst::Create(Context, nullptr, TheSwitch);  // Return void
1344     } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
1345       // return what we have
1346       ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch);
1347     } else {
1348       // Otherwise we must have code extracted an unwind or something, just
1349       // return whatever we want.
1350       ReturnInst::Create(Context,
1351                          Constant::getNullValue(OldFnRetTy), TheSwitch);
1352     }
1353 
1354     TheSwitch->eraseFromParent();
1355     break;
1356   case 1:
1357     // Only a single destination, change the switch into an unconditional
1358     // branch.
1359     BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
1360     TheSwitch->eraseFromParent();
1361     break;
1362   case 2:
1363     BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
1364                        call, TheSwitch);
1365     TheSwitch->eraseFromParent();
1366     break;
1367   default:
1368     // Otherwise, make the default destination of the switch instruction be one
1369     // of the other successors.
1370     TheSwitch->setCondition(call);
1371     TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks));
1372     // Remove redundant case
1373     TheSwitch->removeCase(SwitchInst::CaseIt(TheSwitch, NumExitBlocks-1));
1374     break;
1375   }
1376 
1377   // Insert lifetime markers around the reloads of any output values. The
1378   // allocas output values are stored in are only in-use in the codeRepl block.
1379   insertLifetimeMarkersSurroundingCall(M, ReloadOutputs, ReloadOutputs, call);
1380 
1381   return call;
1382 }
1383 
moveCodeToFunction(Function * newFunction)1384 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
1385   Function *oldFunc = (*Blocks.begin())->getParent();
1386   Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
1387   Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
1388 
1389   for (BasicBlock *Block : Blocks) {
1390     // Delete the basic block from the old function, and the list of blocks
1391     oldBlocks.remove(Block);
1392 
1393     // Insert this basic block into the new function
1394     newBlocks.push_back(Block);
1395   }
1396 }
1397 
calculateNewCallTerminatorWeights(BasicBlock * CodeReplacer,DenseMap<BasicBlock *,BlockFrequency> & ExitWeights,BranchProbabilityInfo * BPI)1398 void CodeExtractor::calculateNewCallTerminatorWeights(
1399     BasicBlock *CodeReplacer,
1400     DenseMap<BasicBlock *, BlockFrequency> &ExitWeights,
1401     BranchProbabilityInfo *BPI) {
1402   using Distribution = BlockFrequencyInfoImplBase::Distribution;
1403   using BlockNode = BlockFrequencyInfoImplBase::BlockNode;
1404 
1405   // Update the branch weights for the exit block.
1406   Instruction *TI = CodeReplacer->getTerminator();
1407   SmallVector<unsigned, 8> BranchWeights(TI->getNumSuccessors(), 0);
1408 
1409   // Block Frequency distribution with dummy node.
1410   Distribution BranchDist;
1411 
1412   SmallVector<BranchProbability, 4> EdgeProbabilities(
1413       TI->getNumSuccessors(), BranchProbability::getUnknown());
1414 
1415   // Add each of the frequencies of the successors.
1416   for (unsigned i = 0, e = TI->getNumSuccessors(); i < e; ++i) {
1417     BlockNode ExitNode(i);
1418     uint64_t ExitFreq = ExitWeights[TI->getSuccessor(i)].getFrequency();
1419     if (ExitFreq != 0)
1420       BranchDist.addExit(ExitNode, ExitFreq);
1421     else
1422       EdgeProbabilities[i] = BranchProbability::getZero();
1423   }
1424 
1425   // Check for no total weight.
1426   if (BranchDist.Total == 0) {
1427     BPI->setEdgeProbability(CodeReplacer, EdgeProbabilities);
1428     return;
1429   }
1430 
1431   // Normalize the distribution so that they can fit in unsigned.
1432   BranchDist.normalize();
1433 
1434   // Create normalized branch weights and set the metadata.
1435   for (unsigned I = 0, E = BranchDist.Weights.size(); I < E; ++I) {
1436     const auto &Weight = BranchDist.Weights[I];
1437 
1438     // Get the weight and update the current BFI.
1439     BranchWeights[Weight.TargetNode.Index] = Weight.Amount;
1440     BranchProbability BP(Weight.Amount, BranchDist.Total);
1441     EdgeProbabilities[Weight.TargetNode.Index] = BP;
1442   }
1443   BPI->setEdgeProbability(CodeReplacer, EdgeProbabilities);
1444   TI->setMetadata(
1445       LLVMContext::MD_prof,
1446       MDBuilder(TI->getContext()).createBranchWeights(BranchWeights));
1447 }
1448 
1449 /// Erase debug info intrinsics which refer to values in \p F but aren't in
1450 /// \p F.
eraseDebugIntrinsicsWithNonLocalRefs(Function & F)1451 static void eraseDebugIntrinsicsWithNonLocalRefs(Function &F) {
1452   for (Instruction &I : instructions(F)) {
1453     SmallVector<DbgVariableIntrinsic *, 4> DbgUsers;
1454     findDbgUsers(DbgUsers, &I);
1455     for (DbgVariableIntrinsic *DVI : DbgUsers)
1456       if (DVI->getFunction() != &F)
1457         DVI->eraseFromParent();
1458   }
1459 }
1460 
1461 /// Fix up the debug info in the old and new functions by pointing line
1462 /// locations and debug intrinsics to the new subprogram scope, and by deleting
1463 /// intrinsics which point to values outside of the new function.
fixupDebugInfoPostExtraction(Function & OldFunc,Function & NewFunc,CallInst & TheCall)1464 static void fixupDebugInfoPostExtraction(Function &OldFunc, Function &NewFunc,
1465                                          CallInst &TheCall) {
1466   DISubprogram *OldSP = OldFunc.getSubprogram();
1467   LLVMContext &Ctx = OldFunc.getContext();
1468 
1469   if (!OldSP) {
1470     // Erase any debug info the new function contains.
1471     stripDebugInfo(NewFunc);
1472     // Make sure the old function doesn't contain any non-local metadata refs.
1473     eraseDebugIntrinsicsWithNonLocalRefs(NewFunc);
1474     return;
1475   }
1476 
1477   // Create a subprogram for the new function. Leave out a description of the
1478   // function arguments, as the parameters don't correspond to anything at the
1479   // source level.
1480   assert(OldSP->getUnit() && "Missing compile unit for subprogram");
1481   DIBuilder DIB(*OldFunc.getParent(), /*AllowUnresolved=*/false,
1482                 OldSP->getUnit());
1483   auto SPType = DIB.createSubroutineType(DIB.getOrCreateTypeArray(None));
1484   DISubprogram::DISPFlags SPFlags = DISubprogram::SPFlagDefinition |
1485                                     DISubprogram::SPFlagOptimized |
1486                                     DISubprogram::SPFlagLocalToUnit;
1487   auto NewSP = DIB.createFunction(
1488       OldSP->getUnit(), NewFunc.getName(), NewFunc.getName(), OldSP->getFile(),
1489       /*LineNo=*/0, SPType, /*ScopeLine=*/0, DINode::FlagZero, SPFlags);
1490   NewFunc.setSubprogram(NewSP);
1491 
1492   // Debug intrinsics in the new function need to be updated in one of two
1493   // ways:
1494   //  1) They need to be deleted, because they describe a value in the old
1495   //     function.
1496   //  2) They need to point to fresh metadata, e.g. because they currently
1497   //     point to a variable in the wrong scope.
1498   SmallDenseMap<DINode *, DINode *> RemappedMetadata;
1499   SmallVector<Instruction *, 4> DebugIntrinsicsToDelete;
1500   for (Instruction &I : instructions(NewFunc)) {
1501     auto *DII = dyn_cast<DbgInfoIntrinsic>(&I);
1502     if (!DII)
1503       continue;
1504 
1505     // Point the intrinsic to a fresh label within the new function.
1506     if (auto *DLI = dyn_cast<DbgLabelInst>(&I)) {
1507       DILabel *OldLabel = DLI->getLabel();
1508       DINode *&NewLabel = RemappedMetadata[OldLabel];
1509       if (!NewLabel)
1510         NewLabel = DILabel::get(Ctx, NewSP, OldLabel->getName(),
1511                                 OldLabel->getFile(), OldLabel->getLine());
1512       DLI->setArgOperand(0, MetadataAsValue::get(Ctx, NewLabel));
1513       continue;
1514     }
1515 
1516     // If the location isn't a constant or an instruction, delete the
1517     // intrinsic.
1518     auto *DVI = cast<DbgVariableIntrinsic>(DII);
1519     Value *Location = DVI->getVariableLocation();
1520     if (!Location ||
1521         (!isa<Constant>(Location) && !isa<Instruction>(Location))) {
1522       DebugIntrinsicsToDelete.push_back(DVI);
1523       continue;
1524     }
1525 
1526     // If the variable location is an instruction but isn't in the new
1527     // function, delete the intrinsic.
1528     Instruction *LocationInst = dyn_cast<Instruction>(Location);
1529     if (LocationInst && LocationInst->getFunction() != &NewFunc) {
1530       DebugIntrinsicsToDelete.push_back(DVI);
1531       continue;
1532     }
1533 
1534     // Point the intrinsic to a fresh variable within the new function.
1535     DILocalVariable *OldVar = DVI->getVariable();
1536     DINode *&NewVar = RemappedMetadata[OldVar];
1537     if (!NewVar)
1538       NewVar = DIB.createAutoVariable(
1539           NewSP, OldVar->getName(), OldVar->getFile(), OldVar->getLine(),
1540           OldVar->getType(), /*AlwaysPreserve=*/false, DINode::FlagZero,
1541           OldVar->getAlignInBits());
1542     DVI->setArgOperand(1, MetadataAsValue::get(Ctx, NewVar));
1543   }
1544   for (auto *DII : DebugIntrinsicsToDelete)
1545     DII->eraseFromParent();
1546   DIB.finalizeSubprogram(NewSP);
1547 
1548   // Fix up the scope information attached to the line locations in the new
1549   // function.
1550   for (Instruction &I : instructions(NewFunc)) {
1551     if (const DebugLoc &DL = I.getDebugLoc())
1552       I.setDebugLoc(DILocation::get(Ctx, DL.getLine(), DL.getCol(), NewSP));
1553 
1554     // Loop info metadata may contain line locations. Fix them up.
1555     auto updateLoopInfoLoc = [&Ctx,
1556                               NewSP](const DILocation &Loc) -> DILocation * {
1557       return DILocation::get(Ctx, Loc.getLine(), Loc.getColumn(), NewSP,
1558                              nullptr);
1559     };
1560     updateLoopMetadataDebugLocations(I, updateLoopInfoLoc);
1561   }
1562   if (!TheCall.getDebugLoc())
1563     TheCall.setDebugLoc(DILocation::get(Ctx, 0, 0, OldSP));
1564 
1565   eraseDebugIntrinsicsWithNonLocalRefs(NewFunc);
1566 }
1567 
1568 Function *
extractCodeRegion(const CodeExtractorAnalysisCache & CEAC)1569 CodeExtractor::extractCodeRegion(const CodeExtractorAnalysisCache &CEAC) {
1570   if (!isEligible())
1571     return nullptr;
1572 
1573   // Assumption: this is a single-entry code region, and the header is the first
1574   // block in the region.
1575   BasicBlock *header = *Blocks.begin();
1576   Function *oldFunction = header->getParent();
1577 
1578   // Calculate the entry frequency of the new function before we change the root
1579   //   block.
1580   BlockFrequency EntryFreq;
1581   if (BFI) {
1582     assert(BPI && "Both BPI and BFI are required to preserve profile info");
1583     for (BasicBlock *Pred : predecessors(header)) {
1584       if (Blocks.count(Pred))
1585         continue;
1586       EntryFreq +=
1587           BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, header);
1588     }
1589   }
1590 
1591   // Remove @llvm.assume calls that will be moved to the new function from the
1592   // old function's assumption cache.
1593   for (BasicBlock *Block : Blocks) {
1594     for (auto It = Block->begin(), End = Block->end(); It != End;) {
1595       Instruction *I = &*It;
1596       ++It;
1597 
1598       if (match(I, m_Intrinsic<Intrinsic::assume>())) {
1599         if (AC)
1600           AC->unregisterAssumption(cast<CallInst>(I));
1601         I->eraseFromParent();
1602       }
1603     }
1604   }
1605 
1606   // If we have any return instructions in the region, split those blocks so
1607   // that the return is not in the region.
1608   splitReturnBlocks();
1609 
1610   // Calculate the exit blocks for the extracted region and the total exit
1611   // weights for each of those blocks.
1612   DenseMap<BasicBlock *, BlockFrequency> ExitWeights;
1613   SmallPtrSet<BasicBlock *, 1> ExitBlocks;
1614   for (BasicBlock *Block : Blocks) {
1615     for (succ_iterator SI = succ_begin(Block), SE = succ_end(Block); SI != SE;
1616          ++SI) {
1617       if (!Blocks.count(*SI)) {
1618         // Update the branch weight for this successor.
1619         if (BFI) {
1620           BlockFrequency &BF = ExitWeights[*SI];
1621           BF += BFI->getBlockFreq(Block) * BPI->getEdgeProbability(Block, *SI);
1622         }
1623         ExitBlocks.insert(*SI);
1624       }
1625     }
1626   }
1627   NumExitBlocks = ExitBlocks.size();
1628 
1629   // If we have to split PHI nodes of the entry or exit blocks, do so now.
1630   severSplitPHINodesOfEntry(header);
1631   severSplitPHINodesOfExits(ExitBlocks);
1632 
1633   // This takes place of the original loop
1634   BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(),
1635                                                 "codeRepl", oldFunction,
1636                                                 header);
1637 
1638   // The new function needs a root node because other nodes can branch to the
1639   // head of the region, but the entry node of a function cannot have preds.
1640   BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(),
1641                                                "newFuncRoot");
1642   auto *BranchI = BranchInst::Create(header);
1643   // If the original function has debug info, we have to add a debug location
1644   // to the new branch instruction from the artificial entry block.
1645   // We use the debug location of the first instruction in the extracted
1646   // blocks, as there is no other equivalent line in the source code.
1647   if (oldFunction->getSubprogram()) {
1648     any_of(Blocks, [&BranchI](const BasicBlock *BB) {
1649       return any_of(*BB, [&BranchI](const Instruction &I) {
1650         if (!I.getDebugLoc())
1651           return false;
1652         BranchI->setDebugLoc(I.getDebugLoc());
1653         return true;
1654       });
1655     });
1656   }
1657   newFuncRoot->getInstList().push_back(BranchI);
1658 
1659   ValueSet inputs, outputs, SinkingCands, HoistingCands;
1660   BasicBlock *CommonExit = nullptr;
1661   findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit);
1662   assert(HoistingCands.empty() || CommonExit);
1663 
1664   // Find inputs to, outputs from the code region.
1665   findInputsOutputs(inputs, outputs, SinkingCands);
1666 
1667   // Now sink all instructions which only have non-phi uses inside the region.
1668   // Group the allocas at the start of the block, so that any bitcast uses of
1669   // the allocas are well-defined.
1670   AllocaInst *FirstSunkAlloca = nullptr;
1671   for (auto *II : SinkingCands) {
1672     if (auto *AI = dyn_cast<AllocaInst>(II)) {
1673       AI->moveBefore(*newFuncRoot, newFuncRoot->getFirstInsertionPt());
1674       if (!FirstSunkAlloca)
1675         FirstSunkAlloca = AI;
1676     }
1677   }
1678   assert((SinkingCands.empty() || FirstSunkAlloca) &&
1679          "Did not expect a sink candidate without any allocas");
1680   for (auto *II : SinkingCands) {
1681     if (!isa<AllocaInst>(II)) {
1682       cast<Instruction>(II)->moveAfter(FirstSunkAlloca);
1683     }
1684   }
1685 
1686   if (!HoistingCands.empty()) {
1687     auto *HoistToBlock = findOrCreateBlockForHoisting(CommonExit);
1688     Instruction *TI = HoistToBlock->getTerminator();
1689     for (auto *II : HoistingCands)
1690       cast<Instruction>(II)->moveBefore(TI);
1691   }
1692 
1693   // Collect objects which are inputs to the extraction region and also
1694   // referenced by lifetime start markers within it. The effects of these
1695   // markers must be replicated in the calling function to prevent the stack
1696   // coloring pass from merging slots which store input objects.
1697   ValueSet LifetimesStart;
1698   eraseLifetimeMarkersOnInputs(Blocks, SinkingCands, LifetimesStart);
1699 
1700   // Construct new function based on inputs/outputs & add allocas for all defs.
1701   Function *newFunction =
1702       constructFunction(inputs, outputs, header, newFuncRoot, codeReplacer,
1703                         oldFunction, oldFunction->getParent());
1704 
1705   // Update the entry count of the function.
1706   if (BFI) {
1707     auto Count = BFI->getProfileCountFromFreq(EntryFreq.getFrequency());
1708     if (Count.hasValue())
1709       newFunction->setEntryCount(
1710           ProfileCount(Count.getValue(), Function::PCT_Real)); // FIXME
1711     BFI->setBlockFreq(codeReplacer, EntryFreq.getFrequency());
1712   }
1713 
1714   CallInst *TheCall =
1715       emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
1716 
1717   moveCodeToFunction(newFunction);
1718 
1719   // Replicate the effects of any lifetime start/end markers which referenced
1720   // input objects in the extraction region by placing markers around the call.
1721   insertLifetimeMarkersSurroundingCall(
1722       oldFunction->getParent(), LifetimesStart.getArrayRef(), {}, TheCall);
1723 
1724   // Propagate personality info to the new function if there is one.
1725   if (oldFunction->hasPersonalityFn())
1726     newFunction->setPersonalityFn(oldFunction->getPersonalityFn());
1727 
1728   // Update the branch weights for the exit block.
1729   if (BFI && NumExitBlocks > 1)
1730     calculateNewCallTerminatorWeights(codeReplacer, ExitWeights, BPI);
1731 
1732   // Loop over all of the PHI nodes in the header and exit blocks, and change
1733   // any references to the old incoming edge to be the new incoming edge.
1734   for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
1735     PHINode *PN = cast<PHINode>(I);
1736     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1737       if (!Blocks.count(PN->getIncomingBlock(i)))
1738         PN->setIncomingBlock(i, newFuncRoot);
1739   }
1740 
1741   for (BasicBlock *ExitBB : ExitBlocks)
1742     for (PHINode &PN : ExitBB->phis()) {
1743       Value *IncomingCodeReplacerVal = nullptr;
1744       for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
1745         // Ignore incoming values from outside of the extracted region.
1746         if (!Blocks.count(PN.getIncomingBlock(i)))
1747           continue;
1748 
1749         // Ensure that there is only one incoming value from codeReplacer.
1750         if (!IncomingCodeReplacerVal) {
1751           PN.setIncomingBlock(i, codeReplacer);
1752           IncomingCodeReplacerVal = PN.getIncomingValue(i);
1753         } else
1754           assert(IncomingCodeReplacerVal == PN.getIncomingValue(i) &&
1755                  "PHI has two incompatbile incoming values from codeRepl");
1756       }
1757     }
1758 
1759   fixupDebugInfoPostExtraction(*oldFunction, *newFunction, *TheCall);
1760 
1761   // Mark the new function `noreturn` if applicable. Terminators which resume
1762   // exception propagation are treated as returning instructions. This is to
1763   // avoid inserting traps after calls to outlined functions which unwind.
1764   bool doesNotReturn = none_of(*newFunction, [](const BasicBlock &BB) {
1765     const Instruction *Term = BB.getTerminator();
1766     return isa<ReturnInst>(Term) || isa<ResumeInst>(Term);
1767   });
1768   if (doesNotReturn)
1769     newFunction->setDoesNotReturn();
1770 
1771   LLVM_DEBUG(if (verifyFunction(*newFunction, &errs())) {
1772     newFunction->dump();
1773     report_fatal_error("verification of newFunction failed!");
1774   });
1775   LLVM_DEBUG(if (verifyFunction(*oldFunction))
1776              report_fatal_error("verification of oldFunction failed!"));
1777   LLVM_DEBUG(if (AC && verifyAssumptionCache(*oldFunction, *newFunction, AC))
1778                  report_fatal_error("Stale Asumption cache for old Function!"));
1779   return newFunction;
1780 }
1781 
verifyAssumptionCache(const Function & OldFunc,const Function & NewFunc,AssumptionCache * AC)1782 bool CodeExtractor::verifyAssumptionCache(const Function &OldFunc,
1783                                           const Function &NewFunc,
1784                                           AssumptionCache *AC) {
1785   for (auto AssumeVH : AC->assumptions()) {
1786     auto *I = dyn_cast_or_null<CallInst>(AssumeVH);
1787     if (!I)
1788       continue;
1789 
1790     // There shouldn't be any llvm.assume intrinsics in the new function.
1791     if (I->getFunction() != &OldFunc)
1792       return true;
1793 
1794     // There shouldn't be any stale affected values in the assumption cache
1795     // that were previously in the old function, but that have now been moved
1796     // to the new function.
1797     for (auto AffectedValVH : AC->assumptionsFor(I->getOperand(0))) {
1798       auto *AffectedCI = dyn_cast_or_null<CallInst>(AffectedValVH);
1799       if (!AffectedCI)
1800         continue;
1801       if (AffectedCI->getFunction() != &OldFunc)
1802         return true;
1803       auto *AssumedInst = cast<Instruction>(AffectedCI->getOperand(0));
1804       if (AssumedInst->getFunction() != &OldFunc)
1805         return true;
1806     }
1807   }
1808   return false;
1809 }
1810