1 //===- CodeExtractor.cpp - Pull code region into a new function -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Transforms/Utils/CodeExtractor.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/Analysis/Dominators.h"
21 #include "llvm/Analysis/LoopInfo.h"
22 #include "llvm/Analysis/RegionInfo.h"
23 #include "llvm/Analysis/RegionIterator.h"
24 #include "llvm/Analysis/Verifier.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/IR/Module.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"
41 // Provide a command-line option to aggregate function arguments into a struct
42 // for functions produced by the code extractor. This is useful when converting
43 // extracted functions to pthread-based code, as only one argument (void*) can
44 // be passed in to pthread_create().
46 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
47 cl::desc("Aggregate arguments to code-extracted functions"));
49 /// \brief Test whether a block is valid for extraction.
50 static bool isBlockValidForExtraction(const BasicBlock &BB) {
51 // Landing pads must be in the function where they were inserted for cleanup.
52 if (BB.isLandingPad())
55 // Don't hoist code containing allocas, invokes, or vastarts.
56 for (BasicBlock::const_iterator I = BB.begin(), E = BB.end(); I != E; ++I) {
57 if (isa<AllocaInst>(I) || isa<InvokeInst>(I))
59 if (const CallInst *CI = dyn_cast<CallInst>(I))
60 if (const Function *F = CI->getCalledFunction())
61 if (F->getIntrinsicID() == Intrinsic::vastart)
68 /// \brief Build a set of blocks to extract if the input blocks are viable.
69 template <typename IteratorT>
70 static SetVector<BasicBlock *> buildExtractionBlockSet(IteratorT BBBegin,
72 SetVector<BasicBlock *> Result;
74 assert(BBBegin != BBEnd);
76 // Loop over the blocks, adding them to our set-vector, and aborting with an
77 // empty set if we encounter invalid blocks.
78 for (IteratorT I = BBBegin, E = BBEnd; I != E; ++I) {
79 if (!Result.insert(*I))
80 llvm_unreachable("Repeated basic blocks in extraction input");
82 if (!isBlockValidForExtraction(**I)) {
89 for (SetVector<BasicBlock *>::iterator I = llvm::next(Result.begin()),
92 for (pred_iterator PI = pred_begin(*I), PE = pred_end(*I);
94 assert(Result.count(*PI) &&
95 "No blocks in this region may have entries from outside the region"
96 " except for the first block!");
102 /// \brief Helper to call buildExtractionBlockSet with an ArrayRef.
103 static SetVector<BasicBlock *>
104 buildExtractionBlockSet(ArrayRef<BasicBlock *> BBs) {
105 return buildExtractionBlockSet(BBs.begin(), BBs.end());
108 /// \brief Helper to call buildExtractionBlockSet with a RegionNode.
109 static SetVector<BasicBlock *>
110 buildExtractionBlockSet(const RegionNode &RN) {
111 if (!RN.isSubRegion())
112 // Just a single BasicBlock.
113 return buildExtractionBlockSet(RN.getNodeAs<BasicBlock>());
115 const Region &R = *RN.getNodeAs<Region>();
117 return buildExtractionBlockSet(R.block_begin(), R.block_end());
120 CodeExtractor::CodeExtractor(BasicBlock *BB, bool AggregateArgs)
121 : DT(0), AggregateArgs(AggregateArgs||AggregateArgsOpt),
122 Blocks(buildExtractionBlockSet(BB)), NumExitBlocks(~0U) {}
124 CodeExtractor::CodeExtractor(ArrayRef<BasicBlock *> BBs, DominatorTree *DT,
126 : DT(DT), AggregateArgs(AggregateArgs||AggregateArgsOpt),
127 Blocks(buildExtractionBlockSet(BBs)), NumExitBlocks(~0U) {}
129 CodeExtractor::CodeExtractor(DominatorTree &DT, Loop &L, bool AggregateArgs)
130 : DT(&DT), AggregateArgs(AggregateArgs||AggregateArgsOpt),
131 Blocks(buildExtractionBlockSet(L.getBlocks())), NumExitBlocks(~0U) {}
133 CodeExtractor::CodeExtractor(DominatorTree &DT, const RegionNode &RN,
135 : DT(&DT), AggregateArgs(AggregateArgs||AggregateArgsOpt),
136 Blocks(buildExtractionBlockSet(RN)), NumExitBlocks(~0U) {}
138 /// definedInRegion - Return true if the specified value is defined in the
139 /// extracted region.
140 static bool definedInRegion(const SetVector<BasicBlock *> &Blocks, Value *V) {
141 if (Instruction *I = dyn_cast<Instruction>(V))
142 if (Blocks.count(I->getParent()))
147 /// definedInCaller - Return true if the specified value is defined in the
148 /// function being code extracted, but not in the region being extracted.
149 /// These values must be passed in as live-ins to the function.
150 static bool definedInCaller(const SetVector<BasicBlock *> &Blocks, Value *V) {
151 if (isa<Argument>(V)) return true;
152 if (Instruction *I = dyn_cast<Instruction>(V))
153 if (!Blocks.count(I->getParent()))
158 void CodeExtractor::findInputsOutputs(ValueSet &Inputs,
159 ValueSet &Outputs) const {
160 for (SetVector<BasicBlock *>::const_iterator I = Blocks.begin(),
165 // If a used value is defined outside the region, it's an input. If an
166 // instruction is used outside the region, it's an output.
167 for (BasicBlock::iterator II = BB->begin(), IE = BB->end();
169 for (User::op_iterator OI = II->op_begin(), OE = II->op_end();
171 if (definedInCaller(Blocks, *OI))
174 for (Value::use_iterator UI = II->use_begin(), UE = II->use_end();
176 if (!definedInRegion(Blocks, *UI)) {
184 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
185 /// region, we need to split the entry block of the region so that the PHI node
186 /// is easier to deal with.
187 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
188 unsigned NumPredsFromRegion = 0;
189 unsigned NumPredsOutsideRegion = 0;
191 if (Header != &Header->getParent()->getEntryBlock()) {
192 PHINode *PN = dyn_cast<PHINode>(Header->begin());
193 if (!PN) return; // No PHI nodes.
195 // If the header node contains any PHI nodes, check to see if there is more
196 // than one entry from outside the region. If so, we need to sever the
197 // header block into two.
198 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
199 if (Blocks.count(PN->getIncomingBlock(i)))
200 ++NumPredsFromRegion;
202 ++NumPredsOutsideRegion;
204 // If there is one (or fewer) predecessor from outside the region, we don't
205 // need to do anything special.
206 if (NumPredsOutsideRegion <= 1) return;
209 // Otherwise, we need to split the header block into two pieces: one
210 // containing PHI nodes merging values from outside of the region, and a
211 // second that contains all of the code for the block and merges back any
212 // incoming values from inside of the region.
213 BasicBlock::iterator AfterPHIs = Header->getFirstNonPHI();
214 BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
215 Header->getName()+".ce");
217 // We only want to code extract the second block now, and it becomes the new
218 // header of the region.
219 BasicBlock *OldPred = Header;
220 Blocks.remove(OldPred);
221 Blocks.insert(NewBB);
224 // Okay, update dominator sets. The blocks that dominate the new one are the
225 // blocks that dominate TIBB plus the new block itself.
227 DT->splitBlock(NewBB);
229 // Okay, now we need to adjust the PHI nodes and any branches from within the
230 // region to go to the new header block instead of the old header block.
231 if (NumPredsFromRegion) {
232 PHINode *PN = cast<PHINode>(OldPred->begin());
233 // Loop over all of the predecessors of OldPred that are in the region,
234 // changing them to branch to NewBB instead.
235 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
236 if (Blocks.count(PN->getIncomingBlock(i))) {
237 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
238 TI->replaceUsesOfWith(OldPred, NewBB);
241 // Okay, everything within the region is now branching to the right block, we
242 // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
243 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
244 PHINode *PN = cast<PHINode>(AfterPHIs);
245 // Create a new PHI node in the new region, which has an incoming value
246 // from OldPred of PN.
247 PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion,
248 PN->getName()+".ce", NewBB->begin());
249 NewPN->addIncoming(PN, OldPred);
251 // Loop over all of the incoming value in PN, moving them to NewPN if they
252 // are from the extracted region.
253 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
254 if (Blocks.count(PN->getIncomingBlock(i))) {
255 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
256 PN->removeIncomingValue(i);
264 void CodeExtractor::splitReturnBlocks() {
265 for (SetVector<BasicBlock *>::iterator I = Blocks.begin(), E = Blocks.end();
267 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator())) {
268 BasicBlock *New = (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
270 // Old dominates New. New node dominates all other nodes dominated
272 DomTreeNode *OldNode = DT->getNode(*I);
273 SmallVector<DomTreeNode*, 8> Children;
274 for (DomTreeNode::iterator DI = OldNode->begin(), DE = OldNode->end();
276 Children.push_back(*DI);
278 DomTreeNode *NewNode = DT->addNewBlock(New, *I);
280 for (SmallVector<DomTreeNode*, 8>::iterator I = Children.begin(),
281 E = Children.end(); I != E; ++I)
282 DT->changeImmediateDominator(*I, NewNode);
287 /// constructFunction - make a function based on inputs and outputs, as follows:
288 /// f(in0, ..., inN, out0, ..., outN)
290 Function *CodeExtractor::constructFunction(const ValueSet &inputs,
291 const ValueSet &outputs,
293 BasicBlock *newRootNode,
294 BasicBlock *newHeader,
295 Function *oldFunction,
297 DEBUG(dbgs() << "inputs: " << inputs.size() << "\n");
298 DEBUG(dbgs() << "outputs: " << outputs.size() << "\n");
300 // This function returns unsigned, outputs will go back by reference.
301 switch (NumExitBlocks) {
303 case 1: RetTy = Type::getVoidTy(header->getContext()); break;
304 case 2: RetTy = Type::getInt1Ty(header->getContext()); break;
305 default: RetTy = Type::getInt16Ty(header->getContext()); break;
308 std::vector<Type*> paramTy;
310 // Add the types of the input values to the function's argument list
311 for (ValueSet::const_iterator i = inputs.begin(), e = inputs.end();
313 const Value *value = *i;
314 DEBUG(dbgs() << "value used in func: " << *value << "\n");
315 paramTy.push_back(value->getType());
318 // Add the types of the output values to the function's argument list.
319 for (ValueSet::const_iterator I = outputs.begin(), E = outputs.end();
321 DEBUG(dbgs() << "instr used in func: " << **I << "\n");
323 paramTy.push_back((*I)->getType());
325 paramTy.push_back(PointerType::getUnqual((*I)->getType()));
328 DEBUG(dbgs() << "Function type: " << *RetTy << " f(");
329 for (std::vector<Type*>::iterator i = paramTy.begin(),
330 e = paramTy.end(); i != e; ++i)
331 DEBUG(dbgs() << **i << ", ");
332 DEBUG(dbgs() << ")\n");
334 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
335 PointerType *StructPtr =
336 PointerType::getUnqual(StructType::get(M->getContext(), paramTy));
338 paramTy.push_back(StructPtr);
340 FunctionType *funcType =
341 FunctionType::get(RetTy, paramTy, false);
343 // Create the new function
344 Function *newFunction = Function::Create(funcType,
345 GlobalValue::InternalLinkage,
346 oldFunction->getName() + "_" +
347 header->getName(), M);
348 // If the old function is no-throw, so is the new one.
349 if (oldFunction->doesNotThrow())
350 newFunction->setDoesNotThrow();
352 newFunction->getBasicBlockList().push_back(newRootNode);
354 // Create an iterator to name all of the arguments we inserted.
355 Function::arg_iterator AI = newFunction->arg_begin();
357 // Rewrite all users of the inputs in the extracted region to use the
358 // arguments (or appropriate addressing into struct) instead.
359 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
363 Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext()));
364 Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i);
365 TerminatorInst *TI = newFunction->begin()->getTerminator();
366 GetElementPtrInst *GEP =
367 GetElementPtrInst::Create(AI, Idx, "gep_" + inputs[i]->getName(), TI);
368 RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI);
372 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
373 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
375 if (Instruction* inst = dyn_cast<Instruction>(*use))
376 if (Blocks.count(inst->getParent()))
377 inst->replaceUsesOfWith(inputs[i], RewriteVal);
380 // Set names for input and output arguments.
381 if (!AggregateArgs) {
382 AI = newFunction->arg_begin();
383 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
384 AI->setName(inputs[i]->getName());
385 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
386 AI->setName(outputs[i]->getName()+".out");
389 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
390 // within the new function. This must be done before we lose track of which
391 // blocks were originally in the code region.
392 std::vector<User*> Users(header->use_begin(), header->use_end());
393 for (unsigned i = 0, e = Users.size(); i != e; ++i)
394 // The BasicBlock which contains the branch is not in the region
395 // modify the branch target to a new block
396 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
397 if (!Blocks.count(TI->getParent()) &&
398 TI->getParent()->getParent() == oldFunction)
399 TI->replaceUsesOfWith(header, newHeader);
404 /// FindPhiPredForUseInBlock - Given a value and a basic block, find a PHI
405 /// that uses the value within the basic block, and return the predecessor
406 /// block associated with that use, or return 0 if none is found.
407 static BasicBlock* FindPhiPredForUseInBlock(Value* Used, BasicBlock* BB) {
408 for (Value::use_iterator UI = Used->use_begin(),
409 UE = Used->use_end(); UI != UE; ++UI) {
410 PHINode *P = dyn_cast<PHINode>(*UI);
411 if (P && P->getParent() == BB)
412 return P->getIncomingBlock(UI);
418 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
419 /// the call instruction, splitting any PHI nodes in the header block as
422 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
423 ValueSet &inputs, ValueSet &outputs) {
424 // Emit a call to the new function, passing in: *pointer to struct (if
425 // aggregating parameters), or plan inputs and allocated memory for outputs
426 std::vector<Value*> params, StructValues, ReloadOutputs, Reloads;
428 LLVMContext &Context = newFunction->getContext();
430 // Add inputs as params, or to be filled into the struct
431 for (ValueSet::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
433 StructValues.push_back(*i);
435 params.push_back(*i);
437 // Create allocas for the outputs
438 for (ValueSet::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
440 StructValues.push_back(*i);
443 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
444 codeReplacer->getParent()->begin()->begin());
445 ReloadOutputs.push_back(alloca);
446 params.push_back(alloca);
450 AllocaInst *Struct = 0;
451 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
452 std::vector<Type*> ArgTypes;
453 for (ValueSet::iterator v = StructValues.begin(),
454 ve = StructValues.end(); v != ve; ++v)
455 ArgTypes.push_back((*v)->getType());
457 // Allocate a struct at the beginning of this function
458 Type *StructArgTy = StructType::get(newFunction->getContext(), ArgTypes);
460 new AllocaInst(StructArgTy, 0, "structArg",
461 codeReplacer->getParent()->begin()->begin());
462 params.push_back(Struct);
464 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
466 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
467 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i);
468 GetElementPtrInst *GEP =
469 GetElementPtrInst::Create(Struct, Idx,
470 "gep_" + StructValues[i]->getName());
471 codeReplacer->getInstList().push_back(GEP);
472 StoreInst *SI = new StoreInst(StructValues[i], GEP);
473 codeReplacer->getInstList().push_back(SI);
477 // Emit the call to the function
478 CallInst *call = CallInst::Create(newFunction, params,
479 NumExitBlocks > 1 ? "targetBlock" : "");
480 codeReplacer->getInstList().push_back(call);
482 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
483 unsigned FirstOut = inputs.size();
485 std::advance(OutputArgBegin, inputs.size());
487 // Reload the outputs passed in by reference
488 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
492 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
493 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i);
494 GetElementPtrInst *GEP
495 = GetElementPtrInst::Create(Struct, Idx,
496 "gep_reload_" + outputs[i]->getName());
497 codeReplacer->getInstList().push_back(GEP);
500 Output = ReloadOutputs[i];
502 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
503 Reloads.push_back(load);
504 codeReplacer->getInstList().push_back(load);
505 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
506 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
507 Instruction *inst = cast<Instruction>(Users[u]);
508 if (!Blocks.count(inst->getParent()))
509 inst->replaceUsesOfWith(outputs[i], load);
513 // Now we can emit a switch statement using the call as a value.
514 SwitchInst *TheSwitch =
515 SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)),
516 codeReplacer, 0, codeReplacer);
518 // Since there may be multiple exits from the original region, make the new
519 // function return an unsigned, switch on that number. This loop iterates
520 // over all of the blocks in the extracted region, updating any terminator
521 // instructions in the to-be-extracted region that branch to blocks that are
522 // not in the region to be extracted.
523 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
525 unsigned switchVal = 0;
526 for (SetVector<BasicBlock*>::const_iterator i = Blocks.begin(),
527 e = Blocks.end(); i != e; ++i) {
528 TerminatorInst *TI = (*i)->getTerminator();
529 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
530 if (!Blocks.count(TI->getSuccessor(i))) {
531 BasicBlock *OldTarget = TI->getSuccessor(i);
532 // add a new basic block which returns the appropriate value
533 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
535 // If we don't already have an exit stub for this non-extracted
536 // destination, create one now!
537 NewTarget = BasicBlock::Create(Context,
538 OldTarget->getName() + ".exitStub",
540 unsigned SuccNum = switchVal++;
543 switch (NumExitBlocks) {
545 case 1: break; // No value needed.
546 case 2: // Conditional branch, return a bool
547 brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum);
550 brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum);
554 ReturnInst *NTRet = ReturnInst::Create(Context, brVal, NewTarget);
556 // Update the switch instruction.
557 TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context),
561 // Restore values just before we exit
562 Function::arg_iterator OAI = OutputArgBegin;
563 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
564 // For an invoke, the normal destination is the only one that is
565 // dominated by the result of the invocation
566 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
568 bool DominatesDef = true;
570 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
571 DefBlock = Invoke->getNormalDest();
573 // Make sure we are looking at the original successor block, not
574 // at a newly inserted exit block, which won't be in the dominator
576 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
577 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
578 if (DefBlock == I->second) {
583 // In the extract block case, if the block we are extracting ends
584 // with an invoke instruction, make sure that we don't emit a
585 // store of the invoke value for the unwind block.
586 if (!DT && DefBlock != OldTarget)
587 DominatesDef = false;
591 DominatesDef = DT->dominates(DefBlock, OldTarget);
593 // If the output value is used by a phi in the target block,
594 // then we need to test for dominance of the phi's predecessor
595 // instead. Unfortunately, this a little complicated since we
596 // have already rewritten uses of the value to uses of the reload.
597 BasicBlock* pred = FindPhiPredForUseInBlock(Reloads[out],
599 if (pred && DT && DT->dominates(DefBlock, pred))
606 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
607 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context),
609 GetElementPtrInst *GEP =
610 GetElementPtrInst::Create(OAI, Idx,
611 "gep_" + outputs[out]->getName(),
613 new StoreInst(outputs[out], GEP, NTRet);
615 new StoreInst(outputs[out], OAI, NTRet);
618 // Advance output iterator even if we don't emit a store
619 if (!AggregateArgs) ++OAI;
623 // rewrite the original branch instruction with this new target
624 TI->setSuccessor(i, NewTarget);
628 // Now that we've done the deed, simplify the switch instruction.
629 Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
630 switch (NumExitBlocks) {
632 // There are no successors (the block containing the switch itself), which
633 // means that previously this was the last part of the function, and hence
634 // this should be rewritten as a `ret'
636 // Check if the function should return a value
637 if (OldFnRetTy->isVoidTy()) {
638 ReturnInst::Create(Context, 0, TheSwitch); // Return void
639 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
640 // return what we have
641 ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch);
643 // Otherwise we must have code extracted an unwind or something, just
644 // return whatever we want.
645 ReturnInst::Create(Context,
646 Constant::getNullValue(OldFnRetTy), TheSwitch);
649 TheSwitch->eraseFromParent();
652 // Only a single destination, change the switch into an unconditional
654 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
655 TheSwitch->eraseFromParent();
658 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
660 TheSwitch->eraseFromParent();
663 // Otherwise, make the default destination of the switch instruction be one
664 // of the other successors.
665 TheSwitch->setCondition(call);
666 TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks));
667 // Remove redundant case
668 SwitchInst::CaseIt ToBeRemoved(TheSwitch, NumExitBlocks-1);
669 TheSwitch->removeCase(ToBeRemoved);
674 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
675 Function *oldFunc = (*Blocks.begin())->getParent();
676 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
677 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
679 for (SetVector<BasicBlock*>::const_iterator i = Blocks.begin(),
680 e = Blocks.end(); i != e; ++i) {
681 // Delete the basic block from the old function, and the list of blocks
682 oldBlocks.remove(*i);
684 // Insert this basic block into the new function
685 newBlocks.push_back(*i);
689 Function *CodeExtractor::extractCodeRegion() {
693 ValueSet inputs, outputs;
695 // Assumption: this is a single-entry code region, and the header is the first
696 // block in the region.
697 BasicBlock *header = *Blocks.begin();
699 // If we have to split PHI nodes or the entry block, do so now.
700 severSplitPHINodes(header);
702 // If we have any return instructions in the region, split those blocks so
703 // that the return is not in the region.
706 Function *oldFunction = header->getParent();
708 // This takes place of the original loop
709 BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(),
710 "codeRepl", oldFunction,
713 // The new function needs a root node because other nodes can branch to the
714 // head of the region, but the entry node of a function cannot have preds.
715 BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(),
717 newFuncRoot->getInstList().push_back(BranchInst::Create(header));
719 // Find inputs to, outputs from the code region.
720 findInputsOutputs(inputs, outputs);
722 SmallPtrSet<BasicBlock *, 1> ExitBlocks;
723 for (SetVector<BasicBlock *>::iterator I = Blocks.begin(), E = Blocks.end();
725 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI)
726 if (!Blocks.count(*SI))
727 ExitBlocks.insert(*SI);
728 NumExitBlocks = ExitBlocks.size();
730 // Construct new function based on inputs/outputs & add allocas for all defs.
731 Function *newFunction = constructFunction(inputs, outputs, header,
733 codeReplacer, oldFunction,
734 oldFunction->getParent());
736 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
738 moveCodeToFunction(newFunction);
740 // Loop over all of the PHI nodes in the header block, and change any
741 // references to the old incoming edge to be the new incoming edge.
742 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
743 PHINode *PN = cast<PHINode>(I);
744 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
745 if (!Blocks.count(PN->getIncomingBlock(i)))
746 PN->setIncomingBlock(i, newFuncRoot);
749 // Look at all successors of the codeReplacer block. If any of these blocks
750 // had PHI nodes in them, we need to update the "from" block to be the code
751 // replacer, not the original block in the extracted region.
752 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
753 succ_end(codeReplacer));
754 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
755 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
756 PHINode *PN = cast<PHINode>(I);
757 std::set<BasicBlock*> ProcessedPreds;
758 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
759 if (Blocks.count(PN->getIncomingBlock(i))) {
760 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
761 PN->setIncomingBlock(i, codeReplacer);
763 // There were multiple entries in the PHI for this block, now there
764 // is only one, so remove the duplicated entries.
765 PN->removeIncomingValue(i, false);
771 //cerr << "NEW FUNCTION: " << *newFunction;
772 // verifyFunction(*newFunction);
774 // cerr << "OLD FUNCTION: " << *oldFunction;
775 // verifyFunction(*oldFunction);
777 DEBUG(if (verifyFunction(*newFunction))
778 report_fatal_error("verifyFunction failed!"));