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