1 //===- CodeExtractor.cpp - Pull code region into a new function -----------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the interface to tear out a code region, such as an
11 // individual loop or a parallel section, into a new function, replacing it with
12 // a call to the new function.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/Transforms/Utils/CodeExtractor.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/Analysis/RegionInfo.h"
22 #include "llvm/Analysis/RegionIterator.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/Verifier.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include <algorithm>
38 #include <set>
39 using namespace llvm;
40 
41 #define DEBUG_TYPE "code-extractor"
42 
43 // Provide a command-line option to aggregate function arguments into a struct
44 // for functions produced by the code extractor. This is useful when converting
45 // extracted functions to pthread-based code, as only one argument (void*) can
46 // be passed in to pthread_create().
47 static cl::opt<bool>
48 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
49                  cl::desc("Aggregate arguments to code-extracted functions"));
50 
51 /// \brief Test whether a block is valid for extraction.
isBlockValidForExtraction(const BasicBlock & BB)52 static bool isBlockValidForExtraction(const BasicBlock &BB) {
53   // Landing pads must be in the function where they were inserted for cleanup.
54   if (BB.isEHPad())
55     return false;
56 
57   // Don't hoist code containing allocas, invokes, or vastarts.
58   for (BasicBlock::const_iterator I = BB.begin(), E = BB.end(); I != E; ++I) {
59     if (isa<AllocaInst>(I) || isa<InvokeInst>(I))
60       return false;
61     if (const CallInst *CI = dyn_cast<CallInst>(I))
62       if (const Function *F = CI->getCalledFunction())
63         if (F->getIntrinsicID() == Intrinsic::vastart)
64           return false;
65   }
66 
67   return true;
68 }
69 
70 /// \brief Build a set of blocks to extract if the input blocks are viable.
71 template <typename IteratorT>
buildExtractionBlockSet(IteratorT BBBegin,IteratorT BBEnd)72 static SetVector<BasicBlock *> buildExtractionBlockSet(IteratorT BBBegin,
73                                                        IteratorT BBEnd) {
74   SetVector<BasicBlock *> Result;
75 
76   assert(BBBegin != BBEnd);
77 
78   // Loop over the blocks, adding them to our set-vector, and aborting with an
79   // empty set if we encounter invalid blocks.
80   for (IteratorT I = BBBegin, E = BBEnd; I != E; ++I) {
81     if (!Result.insert(*I))
82       llvm_unreachable("Repeated basic blocks in extraction input");
83 
84     if (!isBlockValidForExtraction(**I)) {
85       Result.clear();
86       return Result;
87     }
88   }
89 
90 #ifndef NDEBUG
91   for (SetVector<BasicBlock *>::iterator I = std::next(Result.begin()),
92                                          E = Result.end();
93        I != E; ++I)
94     for (pred_iterator PI = pred_begin(*I), PE = pred_end(*I);
95          PI != PE; ++PI)
96       assert(Result.count(*PI) &&
97              "No blocks in this region may have entries from outside the region"
98              " except for the first block!");
99 #endif
100 
101   return Result;
102 }
103 
104 /// \brief Helper to call buildExtractionBlockSet with an ArrayRef.
105 static SetVector<BasicBlock *>
buildExtractionBlockSet(ArrayRef<BasicBlock * > BBs)106 buildExtractionBlockSet(ArrayRef<BasicBlock *> BBs) {
107   return buildExtractionBlockSet(BBs.begin(), BBs.end());
108 }
109 
110 /// \brief Helper to call buildExtractionBlockSet with a RegionNode.
111 static SetVector<BasicBlock *>
buildExtractionBlockSet(const RegionNode & RN)112 buildExtractionBlockSet(const RegionNode &RN) {
113   if (!RN.isSubRegion())
114     // Just a single BasicBlock.
115     return buildExtractionBlockSet(RN.getNodeAs<BasicBlock>());
116 
117   const Region &R = *RN.getNodeAs<Region>();
118 
119   return buildExtractionBlockSet(R.block_begin(), R.block_end());
120 }
121 
CodeExtractor(BasicBlock * BB,bool AggregateArgs)122 CodeExtractor::CodeExtractor(BasicBlock *BB, bool AggregateArgs)
123   : DT(nullptr), AggregateArgs(AggregateArgs||AggregateArgsOpt),
124     Blocks(buildExtractionBlockSet(BB)), NumExitBlocks(~0U) {}
125 
CodeExtractor(ArrayRef<BasicBlock * > BBs,DominatorTree * DT,bool AggregateArgs)126 CodeExtractor::CodeExtractor(ArrayRef<BasicBlock *> BBs, DominatorTree *DT,
127                              bool AggregateArgs)
128   : DT(DT), AggregateArgs(AggregateArgs||AggregateArgsOpt),
129     Blocks(buildExtractionBlockSet(BBs)), NumExitBlocks(~0U) {}
130 
CodeExtractor(DominatorTree & DT,Loop & L,bool AggregateArgs)131 CodeExtractor::CodeExtractor(DominatorTree &DT, Loop &L, bool AggregateArgs)
132   : DT(&DT), AggregateArgs(AggregateArgs||AggregateArgsOpt),
133     Blocks(buildExtractionBlockSet(L.getBlocks())), NumExitBlocks(~0U) {}
134 
CodeExtractor(DominatorTree & DT,const RegionNode & RN,bool AggregateArgs)135 CodeExtractor::CodeExtractor(DominatorTree &DT, const RegionNode &RN,
136                              bool AggregateArgs)
137   : DT(&DT), AggregateArgs(AggregateArgs||AggregateArgsOpt),
138     Blocks(buildExtractionBlockSet(RN)), NumExitBlocks(~0U) {}
139 
140 /// definedInRegion - Return true if the specified value is defined in the
141 /// extracted region.
definedInRegion(const SetVector<BasicBlock * > & Blocks,Value * V)142 static bool definedInRegion(const SetVector<BasicBlock *> &Blocks, Value *V) {
143   if (Instruction *I = dyn_cast<Instruction>(V))
144     if (Blocks.count(I->getParent()))
145       return true;
146   return false;
147 }
148 
149 /// definedInCaller - Return true if the specified value is defined in the
150 /// function being code extracted, but not in the region being extracted.
151 /// These values must be passed in as live-ins to the function.
definedInCaller(const SetVector<BasicBlock * > & Blocks,Value * V)152 static bool definedInCaller(const SetVector<BasicBlock *> &Blocks, Value *V) {
153   if (isa<Argument>(V)) return true;
154   if (Instruction *I = dyn_cast<Instruction>(V))
155     if (!Blocks.count(I->getParent()))
156       return true;
157   return false;
158 }
159 
findInputsOutputs(ValueSet & Inputs,ValueSet & Outputs) const160 void CodeExtractor::findInputsOutputs(ValueSet &Inputs,
161                                       ValueSet &Outputs) const {
162   for (SetVector<BasicBlock *>::const_iterator I = Blocks.begin(),
163                                                E = Blocks.end();
164        I != E; ++I) {
165     BasicBlock *BB = *I;
166 
167     // If a used value is defined outside the region, it's an input.  If an
168     // instruction is used outside the region, it's an output.
169     for (BasicBlock::iterator II = BB->begin(), IE = BB->end();
170          II != IE; ++II) {
171       for (User::op_iterator OI = II->op_begin(), OE = II->op_end();
172            OI != OE; ++OI)
173         if (definedInCaller(Blocks, *OI))
174           Inputs.insert(*OI);
175 
176       for (User *U : II->users())
177         if (!definedInRegion(Blocks, U)) {
178           Outputs.insert(&*II);
179           break;
180         }
181     }
182   }
183 }
184 
185 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
186 /// region, we need to split the entry block of the region so that the PHI node
187 /// is easier to deal with.
severSplitPHINodes(BasicBlock * & Header)188 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
189   unsigned NumPredsFromRegion = 0;
190   unsigned NumPredsOutsideRegion = 0;
191 
192   if (Header != &Header->getParent()->getEntryBlock()) {
193     PHINode *PN = dyn_cast<PHINode>(Header->begin());
194     if (!PN) return;  // No PHI nodes.
195 
196     // If the header node contains any PHI nodes, check to see if there is more
197     // than one entry from outside the region.  If so, we need to sever the
198     // header block into two.
199     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
200       if (Blocks.count(PN->getIncomingBlock(i)))
201         ++NumPredsFromRegion;
202       else
203         ++NumPredsOutsideRegion;
204 
205     // If there is one (or fewer) predecessor from outside the region, we don't
206     // need to do anything special.
207     if (NumPredsOutsideRegion <= 1) return;
208   }
209 
210   // Otherwise, we need to split the header block into two pieces: one
211   // containing PHI nodes merging values from outside of the region, and a
212   // second that contains all of the code for the block and merges back any
213   // incoming values from inside of the region.
214   BasicBlock::iterator AfterPHIs = Header->getFirstNonPHI()->getIterator();
215   BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
216                                               Header->getName()+".ce");
217 
218   // We only want to code extract the second block now, and it becomes the new
219   // header of the region.
220   BasicBlock *OldPred = Header;
221   Blocks.remove(OldPred);
222   Blocks.insert(NewBB);
223   Header = NewBB;
224 
225   // Okay, update dominator sets. The blocks that dominate the new one are the
226   // blocks that dominate TIBB plus the new block itself.
227   if (DT)
228     DT->splitBlock(NewBB);
229 
230   // Okay, now we need to adjust the PHI nodes and any branches from within the
231   // region to go to the new header block instead of the old header block.
232   if (NumPredsFromRegion) {
233     PHINode *PN = cast<PHINode>(OldPred->begin());
234     // Loop over all of the predecessors of OldPred that are in the region,
235     // changing them to branch to NewBB instead.
236     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
237       if (Blocks.count(PN->getIncomingBlock(i))) {
238         TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
239         TI->replaceUsesOfWith(OldPred, NewBB);
240       }
241 
242     // Okay, everything within the region is now branching to the right block, we
243     // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
244     for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
245       PHINode *PN = cast<PHINode>(AfterPHIs);
246       // Create a new PHI node in the new region, which has an incoming value
247       // from OldPred of PN.
248       PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion,
249                                        PN->getName() + ".ce", &NewBB->front());
250       NewPN->addIncoming(PN, OldPred);
251 
252       // Loop over all of the incoming value in PN, moving them to NewPN if they
253       // are from the extracted region.
254       for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
255         if (Blocks.count(PN->getIncomingBlock(i))) {
256           NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
257           PN->removeIncomingValue(i);
258           --i;
259         }
260       }
261     }
262   }
263 }
264 
splitReturnBlocks()265 void CodeExtractor::splitReturnBlocks() {
266   for (SetVector<BasicBlock *>::iterator I = Blocks.begin(), E = Blocks.end();
267        I != E; ++I)
268     if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator())) {
269       BasicBlock *New =
270           (*I)->splitBasicBlock(RI->getIterator(), (*I)->getName() + ".ret");
271       if (DT) {
272         // Old dominates New. New node dominates all other nodes dominated
273         // by Old.
274         DomTreeNode *OldNode = DT->getNode(*I);
275         SmallVector<DomTreeNode*, 8> Children;
276         for (DomTreeNode::iterator DI = OldNode->begin(), DE = OldNode->end();
277              DI != DE; ++DI)
278           Children.push_back(*DI);
279 
280         DomTreeNode *NewNode = DT->addNewBlock(New, *I);
281 
282         for (SmallVectorImpl<DomTreeNode *>::iterator I = Children.begin(),
283                E = Children.end(); I != E; ++I)
284           DT->changeImmediateDominator(*I, NewNode);
285       }
286     }
287 }
288 
289 /// constructFunction - make a function based on inputs and outputs, as follows:
290 /// f(in0, ..., inN, out0, ..., outN)
291 ///
constructFunction(const ValueSet & inputs,const ValueSet & outputs,BasicBlock * header,BasicBlock * newRootNode,BasicBlock * newHeader,Function * oldFunction,Module * M)292 Function *CodeExtractor::constructFunction(const ValueSet &inputs,
293                                            const ValueSet &outputs,
294                                            BasicBlock *header,
295                                            BasicBlock *newRootNode,
296                                            BasicBlock *newHeader,
297                                            Function *oldFunction,
298                                            Module *M) {
299   DEBUG(dbgs() << "inputs: " << inputs.size() << "\n");
300   DEBUG(dbgs() << "outputs: " << outputs.size() << "\n");
301 
302   // This function returns unsigned, outputs will go back by reference.
303   switch (NumExitBlocks) {
304   case 0:
305   case 1: RetTy = Type::getVoidTy(header->getContext()); break;
306   case 2: RetTy = Type::getInt1Ty(header->getContext()); break;
307   default: RetTy = Type::getInt16Ty(header->getContext()); break;
308   }
309 
310   std::vector<Type*> paramTy;
311 
312   // Add the types of the input values to the function's argument list
313   for (ValueSet::const_iterator i = inputs.begin(), e = inputs.end();
314        i != e; ++i) {
315     const Value *value = *i;
316     DEBUG(dbgs() << "value used in func: " << *value << "\n");
317     paramTy.push_back(value->getType());
318   }
319 
320   // Add the types of the output values to the function's argument list.
321   for (ValueSet::const_iterator I = outputs.begin(), E = outputs.end();
322        I != E; ++I) {
323     DEBUG(dbgs() << "instr used in func: " << **I << "\n");
324     if (AggregateArgs)
325       paramTy.push_back((*I)->getType());
326     else
327       paramTy.push_back(PointerType::getUnqual((*I)->getType()));
328   }
329 
330   DEBUG(dbgs() << "Function type: " << *RetTy << " f(");
331   for (std::vector<Type*>::iterator i = paramTy.begin(),
332          e = paramTy.end(); i != e; ++i)
333     DEBUG(dbgs() << **i << ", ");
334   DEBUG(dbgs() << ")\n");
335 
336   StructType *StructTy;
337   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
338     StructTy = StructType::get(M->getContext(), paramTy);
339     paramTy.clear();
340     paramTy.push_back(PointerType::getUnqual(StructTy));
341   }
342   FunctionType *funcType =
343                   FunctionType::get(RetTy, paramTy, false);
344 
345   // Create the new function
346   Function *newFunction = Function::Create(funcType,
347                                            GlobalValue::InternalLinkage,
348                                            oldFunction->getName() + "_" +
349                                            header->getName(), M);
350   // If the old function is no-throw, so is the new one.
351   if (oldFunction->doesNotThrow())
352     newFunction->setDoesNotThrow();
353 
354   newFunction->getBasicBlockList().push_back(newRootNode);
355 
356   // Create an iterator to name all of the arguments we inserted.
357   Function::arg_iterator AI = newFunction->arg_begin();
358 
359   // Rewrite all users of the inputs in the extracted region to use the
360   // arguments (or appropriate addressing into struct) instead.
361   for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
362     Value *RewriteVal;
363     if (AggregateArgs) {
364       Value *Idx[2];
365       Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext()));
366       Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i);
367       TerminatorInst *TI = newFunction->begin()->getTerminator();
368       GetElementPtrInst *GEP = GetElementPtrInst::Create(
369           StructTy, &*AI, Idx, "gep_" + inputs[i]->getName(), TI);
370       RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI);
371     } else
372       RewriteVal = &*AI++;
373 
374     std::vector<User*> Users(inputs[i]->user_begin(), inputs[i]->user_end());
375     for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
376          use != useE; ++use)
377       if (Instruction* inst = dyn_cast<Instruction>(*use))
378         if (Blocks.count(inst->getParent()))
379           inst->replaceUsesOfWith(inputs[i], RewriteVal);
380   }
381 
382   // Set names for input and output arguments.
383   if (!AggregateArgs) {
384     AI = newFunction->arg_begin();
385     for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
386       AI->setName(inputs[i]->getName());
387     for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
388       AI->setName(outputs[i]->getName()+".out");
389   }
390 
391   // Rewrite branches to basic blocks outside of the loop to new dummy blocks
392   // within the new function. This must be done before we lose track of which
393   // blocks were originally in the code region.
394   std::vector<User*> Users(header->user_begin(), header->user_end());
395   for (unsigned i = 0, e = Users.size(); i != e; ++i)
396     // The BasicBlock which contains the branch is not in the region
397     // modify the branch target to a new block
398     if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
399       if (!Blocks.count(TI->getParent()) &&
400           TI->getParent()->getParent() == oldFunction)
401         TI->replaceUsesOfWith(header, newHeader);
402 
403   return newFunction;
404 }
405 
406 /// FindPhiPredForUseInBlock - Given a value and a basic block, find a PHI
407 /// that uses the value within the basic block, and return the predecessor
408 /// block associated with that use, or return 0 if none is found.
FindPhiPredForUseInBlock(Value * Used,BasicBlock * BB)409 static BasicBlock* FindPhiPredForUseInBlock(Value* Used, BasicBlock* BB) {
410   for (Use &U : Used->uses()) {
411      PHINode *P = dyn_cast<PHINode>(U.getUser());
412      if (P && P->getParent() == BB)
413        return P->getIncomingBlock(U);
414   }
415 
416   return nullptr;
417 }
418 
419 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
420 /// the call instruction, splitting any PHI nodes in the header block as
421 /// necessary.
422 void CodeExtractor::
emitCallAndSwitchStatement(Function * newFunction,BasicBlock * codeReplacer,ValueSet & inputs,ValueSet & outputs)423 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
424                            ValueSet &inputs, ValueSet &outputs) {
425   // Emit a call to the new function, passing in: *pointer to struct (if
426   // aggregating parameters), or plan inputs and allocated memory for outputs
427   std::vector<Value*> params, StructValues, ReloadOutputs, Reloads;
428 
429   LLVMContext &Context = newFunction->getContext();
430 
431   // Add inputs as params, or to be filled into the struct
432   for (ValueSet::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
433     if (AggregateArgs)
434       StructValues.push_back(*i);
435     else
436       params.push_back(*i);
437 
438   // Create allocas for the outputs
439   for (ValueSet::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
440     if (AggregateArgs) {
441       StructValues.push_back(*i);
442     } else {
443       AllocaInst *alloca =
444           new AllocaInst((*i)->getType(), nullptr, (*i)->getName() + ".loc",
445                          &codeReplacer->getParent()->front().front());
446       ReloadOutputs.push_back(alloca);
447       params.push_back(alloca);
448     }
449   }
450 
451   StructType *StructArgTy = nullptr;
452   AllocaInst *Struct = nullptr;
453   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
454     std::vector<Type*> ArgTypes;
455     for (ValueSet::iterator v = StructValues.begin(),
456            ve = StructValues.end(); v != ve; ++v)
457       ArgTypes.push_back((*v)->getType());
458 
459     // Allocate a struct at the beginning of this function
460     StructArgTy = StructType::get(newFunction->getContext(), ArgTypes);
461     Struct = new AllocaInst(StructArgTy, nullptr, "structArg",
462                             &codeReplacer->getParent()->front().front());
463     params.push_back(Struct);
464 
465     for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
466       Value *Idx[2];
467       Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
468       Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i);
469       GetElementPtrInst *GEP = GetElementPtrInst::Create(
470           StructArgTy, Struct, Idx, "gep_" + StructValues[i]->getName());
471       codeReplacer->getInstList().push_back(GEP);
472       StoreInst *SI = new StoreInst(StructValues[i], GEP);
473       codeReplacer->getInstList().push_back(SI);
474     }
475   }
476 
477   // Emit the call to the function
478   CallInst *call = CallInst::Create(newFunction, params,
479                                     NumExitBlocks > 1 ? "targetBlock" : "");
480   codeReplacer->getInstList().push_back(call);
481 
482   Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
483   unsigned FirstOut = inputs.size();
484   if (!AggregateArgs)
485     std::advance(OutputArgBegin, inputs.size());
486 
487   // Reload the outputs passed in by reference
488   for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
489     Value *Output = nullptr;
490     if (AggregateArgs) {
491       Value *Idx[2];
492       Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
493       Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i);
494       GetElementPtrInst *GEP = GetElementPtrInst::Create(
495           StructArgTy, Struct, Idx, "gep_reload_" + outputs[i]->getName());
496       codeReplacer->getInstList().push_back(GEP);
497       Output = GEP;
498     } else {
499       Output = ReloadOutputs[i];
500     }
501     LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
502     Reloads.push_back(load);
503     codeReplacer->getInstList().push_back(load);
504     std::vector<User*> Users(outputs[i]->user_begin(), outputs[i]->user_end());
505     for (unsigned u = 0, e = Users.size(); u != e; ++u) {
506       Instruction *inst = cast<Instruction>(Users[u]);
507       if (!Blocks.count(inst->getParent()))
508         inst->replaceUsesOfWith(outputs[i], load);
509     }
510   }
511 
512   // Now we can emit a switch statement using the call as a value.
513   SwitchInst *TheSwitch =
514       SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)),
515                          codeReplacer, 0, codeReplacer);
516 
517   // Since there may be multiple exits from the original region, make the new
518   // function return an unsigned, switch on that number.  This loop iterates
519   // over all of the blocks in the extracted region, updating any terminator
520   // instructions in the to-be-extracted region that branch to blocks that are
521   // not in the region to be extracted.
522   std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
523 
524   unsigned switchVal = 0;
525   for (SetVector<BasicBlock*>::const_iterator i = Blocks.begin(),
526          e = Blocks.end(); i != e; ++i) {
527     TerminatorInst *TI = (*i)->getTerminator();
528     for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
529       if (!Blocks.count(TI->getSuccessor(i))) {
530         BasicBlock *OldTarget = TI->getSuccessor(i);
531         // add a new basic block which returns the appropriate value
532         BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
533         if (!NewTarget) {
534           // If we don't already have an exit stub for this non-extracted
535           // destination, create one now!
536           NewTarget = BasicBlock::Create(Context,
537                                          OldTarget->getName() + ".exitStub",
538                                          newFunction);
539           unsigned SuccNum = switchVal++;
540 
541           Value *brVal = nullptr;
542           switch (NumExitBlocks) {
543           case 0:
544           case 1: break;  // No value needed.
545           case 2:         // Conditional branch, return a bool
546             brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum);
547             break;
548           default:
549             brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum);
550             break;
551           }
552 
553           ReturnInst *NTRet = ReturnInst::Create(Context, brVal, NewTarget);
554 
555           // Update the switch instruction.
556           TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context),
557                                               SuccNum),
558                              OldTarget);
559 
560           // Restore values just before we exit
561           Function::arg_iterator OAI = OutputArgBegin;
562           for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
563             // For an invoke, the normal destination is the only one that is
564             // dominated by the result of the invocation
565             BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
566 
567             bool DominatesDef = true;
568 
569             BasicBlock *NormalDest = nullptr;
570             if (auto *Invoke = dyn_cast<InvokeInst>(outputs[out]))
571               NormalDest = Invoke->getNormalDest();
572 
573             if (NormalDest) {
574               DefBlock = NormalDest;
575 
576               // Make sure we are looking at the original successor block, not
577               // at a newly inserted exit block, which won't be in the dominator
578               // info.
579               for (std::map<BasicBlock*, BasicBlock*>::iterator I =
580                      ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
581                 if (DefBlock == I->second) {
582                   DefBlock = I->first;
583                   break;
584                 }
585 
586               // In the extract block case, if the block we are extracting ends
587               // with an invoke instruction, make sure that we don't emit a
588               // store of the invoke value for the unwind block.
589               if (!DT && DefBlock != OldTarget)
590                 DominatesDef = false;
591             }
592 
593             if (DT) {
594               DominatesDef = DT->dominates(DefBlock, OldTarget);
595 
596               // If the output value is used by a phi in the target block,
597               // then we need to test for dominance of the phi's predecessor
598               // instead.  Unfortunately, this a little complicated since we
599               // have already rewritten uses of the value to uses of the reload.
600               BasicBlock* pred = FindPhiPredForUseInBlock(Reloads[out],
601                                                           OldTarget);
602               if (pred && DT && DT->dominates(DefBlock, pred))
603                 DominatesDef = true;
604             }
605 
606             if (DominatesDef) {
607               if (AggregateArgs) {
608                 Value *Idx[2];
609                 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
610                 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context),
611                                           FirstOut+out);
612                 GetElementPtrInst *GEP = GetElementPtrInst::Create(
613                     StructArgTy, &*OAI, Idx, "gep_" + outputs[out]->getName(),
614                     NTRet);
615                 new StoreInst(outputs[out], GEP, NTRet);
616               } else {
617                 new StoreInst(outputs[out], &*OAI, NTRet);
618               }
619             }
620             // Advance output iterator even if we don't emit a store
621             if (!AggregateArgs) ++OAI;
622           }
623         }
624 
625         // rewrite the original branch instruction with this new target
626         TI->setSuccessor(i, NewTarget);
627       }
628   }
629 
630   // Now that we've done the deed, simplify the switch instruction.
631   Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
632   switch (NumExitBlocks) {
633   case 0:
634     // There are no successors (the block containing the switch itself), which
635     // means that previously this was the last part of the function, and hence
636     // this should be rewritten as a `ret'
637 
638     // Check if the function should return a value
639     if (OldFnRetTy->isVoidTy()) {
640       ReturnInst::Create(Context, nullptr, TheSwitch);  // Return void
641     } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
642       // return what we have
643       ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch);
644     } else {
645       // Otherwise we must have code extracted an unwind or something, just
646       // return whatever we want.
647       ReturnInst::Create(Context,
648                          Constant::getNullValue(OldFnRetTy), TheSwitch);
649     }
650 
651     TheSwitch->eraseFromParent();
652     break;
653   case 1:
654     // Only a single destination, change the switch into an unconditional
655     // branch.
656     BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
657     TheSwitch->eraseFromParent();
658     break;
659   case 2:
660     BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
661                        call, TheSwitch);
662     TheSwitch->eraseFromParent();
663     break;
664   default:
665     // Otherwise, make the default destination of the switch instruction be one
666     // of the other successors.
667     TheSwitch->setCondition(call);
668     TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks));
669     // Remove redundant case
670     TheSwitch->removeCase(SwitchInst::CaseIt(TheSwitch, NumExitBlocks-1));
671     break;
672   }
673 }
674 
moveCodeToFunction(Function * newFunction)675 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
676   Function *oldFunc = (*Blocks.begin())->getParent();
677   Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
678   Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
679 
680   for (SetVector<BasicBlock*>::const_iterator i = Blocks.begin(),
681          e = Blocks.end(); i != e; ++i) {
682     // Delete the basic block from the old function, and the list of blocks
683     oldBlocks.remove(*i);
684 
685     // Insert this basic block into the new function
686     newBlocks.push_back(*i);
687   }
688 }
689 
extractCodeRegion()690 Function *CodeExtractor::extractCodeRegion() {
691   if (!isEligible())
692     return nullptr;
693 
694   ValueSet inputs, outputs;
695 
696   // Assumption: this is a single-entry code region, and the header is the first
697   // block in the region.
698   BasicBlock *header = *Blocks.begin();
699 
700   // If we have to split PHI nodes or the entry block, do so now.
701   severSplitPHINodes(header);
702 
703   // If we have any return instructions in the region, split those blocks so
704   // that the return is not in the region.
705   splitReturnBlocks();
706 
707   Function *oldFunction = header->getParent();
708 
709   // This takes place of the original loop
710   BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(),
711                                                 "codeRepl", oldFunction,
712                                                 header);
713 
714   // The new function needs a root node because other nodes can branch to the
715   // head of the region, but the entry node of a function cannot have preds.
716   BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(),
717                                                "newFuncRoot");
718   newFuncRoot->getInstList().push_back(BranchInst::Create(header));
719 
720   // Find inputs to, outputs from the code region.
721   findInputsOutputs(inputs, outputs);
722 
723   SmallPtrSet<BasicBlock *, 1> ExitBlocks;
724   for (SetVector<BasicBlock *>::iterator I = Blocks.begin(), E = Blocks.end();
725        I != E; ++I)
726     for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI)
727       if (!Blocks.count(*SI))
728         ExitBlocks.insert(*SI);
729   NumExitBlocks = ExitBlocks.size();
730 
731   // Construct new function based on inputs/outputs & add allocas for all defs.
732   Function *newFunction = constructFunction(inputs, outputs, header,
733                                             newFuncRoot,
734                                             codeReplacer, oldFunction,
735                                             oldFunction->getParent());
736 
737   emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
738 
739   moveCodeToFunction(newFunction);
740 
741   // Loop over all of the PHI nodes in the header block, and change any
742   // references to the old incoming edge to be the new incoming edge.
743   for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
744     PHINode *PN = cast<PHINode>(I);
745     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
746       if (!Blocks.count(PN->getIncomingBlock(i)))
747         PN->setIncomingBlock(i, newFuncRoot);
748   }
749 
750   // Look at all successors of the codeReplacer block.  If any of these blocks
751   // had PHI nodes in them, we need to update the "from" block to be the code
752   // replacer, not the original block in the extracted region.
753   std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
754                                  succ_end(codeReplacer));
755   for (unsigned i = 0, e = Succs.size(); i != e; ++i)
756     for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
757       PHINode *PN = cast<PHINode>(I);
758       std::set<BasicBlock*> ProcessedPreds;
759       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
760         if (Blocks.count(PN->getIncomingBlock(i))) {
761           if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
762             PN->setIncomingBlock(i, codeReplacer);
763           else {
764             // There were multiple entries in the PHI for this block, now there
765             // is only one, so remove the duplicated entries.
766             PN->removeIncomingValue(i, false);
767             --i; --e;
768           }
769         }
770     }
771 
772   //cerr << "NEW FUNCTION: " << *newFunction;
773   //  verifyFunction(*newFunction);
774 
775   //  cerr << "OLD FUNCTION: " << *oldFunction;
776   //  verifyFunction(*oldFunction);
777 
778   DEBUG(if (verifyFunction(*newFunction))
779         report_fatal_error("verifyFunction failed!"));
780   return newFunction;
781 }
782