1 //===- CodeExtractor.cpp - Pull code region into a new function -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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/FunctionUtils.h"
17 #include "llvm/Constants.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/Intrinsics.h"
21 #include "llvm/Module.h"
22 #include "llvm/Pass.h"
23 #include "llvm/Analysis/Dominators.h"
24 #include "llvm/Analysis/LoopInfo.h"
25 #include "llvm/Analysis/Verifier.h"
26 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/ADT/StringExtras.h"
35 // Provide a command-line option to aggregate function arguments into a struct
36 // for functions produced by the code extrator. This is useful when converting
37 // extracted functions to pthread-based code, as only one argument (void*) can
38 // be passed in to pthread_create().
40 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
41 cl::desc("Aggregate arguments to code-extracted functions"));
44 class VISIBILITY_HIDDEN CodeExtractor {
45 typedef std::vector<Value*> Values;
46 std::set<BasicBlock*> BlocksToExtract;
49 unsigned NumExitBlocks;
52 CodeExtractor(DominatorSet *ds = 0, bool AggArgs = false)
53 : DS(ds), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
55 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
57 bool isEligible(const std::vector<BasicBlock*> &code);
60 /// definedInRegion - Return true if the specified value is defined in the
62 bool definedInRegion(Value *V) const {
63 if (Instruction *I = dyn_cast<Instruction>(V))
64 if (BlocksToExtract.count(I->getParent()))
69 /// definedInCaller - Return true if the specified value is defined in the
70 /// function being code extracted, but not in the region being extracted.
71 /// These values must be passed in as live-ins to the function.
72 bool definedInCaller(Value *V) const {
73 if (isa<Argument>(V)) return true;
74 if (Instruction *I = dyn_cast<Instruction>(V))
75 if (!BlocksToExtract.count(I->getParent()))
80 void severSplitPHINodes(BasicBlock *&Header);
81 void splitReturnBlocks();
82 void findInputsOutputs(Values &inputs, Values &outputs);
84 Function *constructFunction(const Values &inputs,
85 const Values &outputs,
87 BasicBlock *newRootNode, BasicBlock *newHeader,
88 Function *oldFunction, Module *M);
90 void moveCodeToFunction(Function *newFunction);
92 void emitCallAndSwitchStatement(Function *newFunction,
93 BasicBlock *newHeader,
100 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
101 /// region, we need to split the entry block of the region so that the PHI node
102 /// is easier to deal with.
103 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
104 bool HasPredsFromRegion = false;
105 unsigned NumPredsOutsideRegion = 0;
107 if (Header != &Header->getParent()->front()) {
108 PHINode *PN = dyn_cast<PHINode>(Header->begin());
109 if (!PN) return; // No PHI nodes.
111 // If the header node contains any PHI nodes, check to see if there is more
112 // than one entry from outside the region. If so, we need to sever the
113 // header block into two.
114 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
115 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
116 HasPredsFromRegion = true;
118 ++NumPredsOutsideRegion;
120 // If there is one (or fewer) predecessor from outside the region, we don't
121 // need to do anything special.
122 if (NumPredsOutsideRegion <= 1) return;
125 // Otherwise, we need to split the header block into two pieces: one
126 // containing PHI nodes merging values from outside of the region, and a
127 // second that contains all of the code for the block and merges back any
128 // incoming values from inside of the region.
129 BasicBlock::iterator AfterPHIs = Header->begin();
130 while (isa<PHINode>(AfterPHIs)) ++AfterPHIs;
131 BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
132 Header->getName()+".ce");
134 // We only want to code extract the second block now, and it becomes the new
135 // header of the region.
136 BasicBlock *OldPred = Header;
137 BlocksToExtract.erase(OldPred);
138 BlocksToExtract.insert(NewBB);
141 // Okay, update dominator sets. The blocks that dominate the new one are the
142 // blocks that dominate TIBB plus the new block itself.
144 DominatorSet::DomSetType DomSet = DS->getDominators(OldPred);
145 DomSet.insert(NewBB); // A block always dominates itself.
146 DS->addBasicBlock(NewBB, DomSet);
148 // Additionally, NewBB dominates all blocks in the function that are
149 // dominated by OldPred.
150 Function *F = Header->getParent();
151 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
152 if (DS->properlyDominates(OldPred, I))
153 DS->addDominator(I, NewBB);
156 // Okay, now we need to adjust the PHI nodes and any branches from within the
157 // region to go to the new header block instead of the old header block.
158 if (HasPredsFromRegion) {
159 PHINode *PN = cast<PHINode>(OldPred->begin());
160 // Loop over all of the predecessors of OldPred that are in the region,
161 // changing them to branch to NewBB instead.
162 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
163 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
164 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
165 TI->replaceUsesOfWith(OldPred, NewBB);
168 // Okay, everthing within the region is now branching to the right block, we
169 // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
170 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
171 PHINode *PN = cast<PHINode>(AfterPHIs);
172 // Create a new PHI node in the new region, which has an incoming value
173 // from OldPred of PN.
174 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".ce",
176 NewPN->addIncoming(PN, OldPred);
178 // Loop over all of the incoming value in PN, moving them to NewPN if they
179 // are from the extracted region.
180 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
181 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
182 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
183 PN->removeIncomingValue(i);
191 void CodeExtractor::splitReturnBlocks() {
192 for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(),
193 E = BlocksToExtract.end(); I != E; ++I)
194 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator()))
195 (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
198 // findInputsOutputs - Find inputs to, outputs from the code region.
200 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
201 std::set<BasicBlock*> ExitBlocks;
202 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
203 ce = BlocksToExtract.end(); ci != ce; ++ci) {
204 BasicBlock *BB = *ci;
206 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
207 // If a used value is defined outside the region, it's an input. If an
208 // instruction is used outside the region, it's an output.
209 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
210 if (definedInCaller(*O))
211 inputs.push_back(*O);
213 // Consider uses of this instruction (outputs).
214 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
216 if (!definedInRegion(*UI)) {
217 outputs.push_back(I);
222 // Keep track of the exit blocks from the region.
223 TerminatorInst *TI = BB->getTerminator();
224 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
225 if (!BlocksToExtract.count(TI->getSuccessor(i)))
226 ExitBlocks.insert(TI->getSuccessor(i));
227 } // for: basic blocks
229 NumExitBlocks = ExitBlocks.size();
231 // Eliminate duplicates.
232 std::sort(inputs.begin(), inputs.end());
233 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
234 std::sort(outputs.begin(), outputs.end());
235 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
238 /// constructFunction - make a function based on inputs and outputs, as follows:
239 /// f(in0, ..., inN, out0, ..., outN)
241 Function *CodeExtractor::constructFunction(const Values &inputs,
242 const Values &outputs,
244 BasicBlock *newRootNode,
245 BasicBlock *newHeader,
246 Function *oldFunction,
248 DOUT << "inputs: " << inputs.size() << "\n";
249 DOUT << "outputs: " << outputs.size() << "\n";
251 // This function returns unsigned, outputs will go back by reference.
252 switch (NumExitBlocks) {
254 case 1: RetTy = Type::VoidTy; break;
255 case 2: RetTy = Type::Int1Ty; break;
256 default: RetTy = Type::Int16Ty; break;
259 std::vector<const Type*> paramTy;
261 // Add the types of the input values to the function's argument list
262 for (Values::const_iterator i = inputs.begin(),
263 e = inputs.end(); i != e; ++i) {
264 const Value *value = *i;
265 DOUT << "value used in func: " << *value << "\n";
266 paramTy.push_back(value->getType());
269 // Add the types of the output values to the function's argument list.
270 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
272 DOUT << "instr used in func: " << **I << "\n";
274 paramTy.push_back((*I)->getType());
276 paramTy.push_back(PointerType::get((*I)->getType()));
279 DOUT << "Function type: " << *RetTy << " f(";
280 for (std::vector<const Type*>::iterator i = paramTy.begin(),
281 e = paramTy.end(); i != e; ++i)
285 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
286 PointerType *StructPtr = PointerType::get(StructType::get(paramTy));
288 paramTy.push_back(StructPtr);
290 const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false);
292 // Create the new function
293 Function *newFunction = new Function(funcType,
294 GlobalValue::InternalLinkage,
295 oldFunction->getName() + "_" +
296 header->getName(), M);
297 newFunction->getBasicBlockList().push_back(newRootNode);
299 // Create an iterator to name all of the arguments we inserted.
300 Function::arg_iterator AI = newFunction->arg_begin();
302 // Rewrite all users of the inputs in the extracted region to use the
303 // arguments (or appropriate addressing into struct) instead.
304 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
307 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
308 Value *Idx1 = ConstantInt::get(Type::Int32Ty, i);
309 std::string GEPname = "gep_" + inputs[i]->getName();
310 TerminatorInst *TI = newFunction->begin()->getTerminator();
311 GetElementPtrInst *GEP = new GetElementPtrInst(AI, Idx0, Idx1,
313 RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
317 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
318 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
320 if (Instruction* inst = dyn_cast<Instruction>(*use))
321 if (BlocksToExtract.count(inst->getParent()))
322 inst->replaceUsesOfWith(inputs[i], RewriteVal);
325 // Set names for input and output arguments.
326 if (!AggregateArgs) {
327 AI = newFunction->arg_begin();
328 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
329 AI->setName(inputs[i]->getName());
330 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
331 AI->setName(outputs[i]->getName()+".out");
334 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
335 // within the new function. This must be done before we lose track of which
336 // blocks were originally in the code region.
337 std::vector<User*> Users(header->use_begin(), header->use_end());
338 for (unsigned i = 0, e = Users.size(); i != e; ++i)
339 // The BasicBlock which contains the branch is not in the region
340 // modify the branch target to a new block
341 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
342 if (!BlocksToExtract.count(TI->getParent()) &&
343 TI->getParent()->getParent() == oldFunction)
344 TI->replaceUsesOfWith(header, newHeader);
349 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
350 /// the call instruction, splitting any PHI nodes in the header block as
353 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
354 Values &inputs, Values &outputs) {
355 // Emit a call to the new function, passing in: *pointer to struct (if
356 // aggregating parameters), or plan inputs and allocated memory for outputs
357 std::vector<Value*> params, StructValues, ReloadOutputs;
359 // Add inputs as params, or to be filled into the struct
360 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
362 StructValues.push_back(*i);
364 params.push_back(*i);
366 // Create allocas for the outputs
367 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
369 StructValues.push_back(*i);
372 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
373 codeReplacer->getParent()->begin()->begin());
374 ReloadOutputs.push_back(alloca);
375 params.push_back(alloca);
379 AllocaInst *Struct = 0;
380 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
381 std::vector<const Type*> ArgTypes;
382 for (Values::iterator v = StructValues.begin(),
383 ve = StructValues.end(); v != ve; ++v)
384 ArgTypes.push_back((*v)->getType());
386 // Allocate a struct at the beginning of this function
387 Type *StructArgTy = StructType::get(ArgTypes);
389 new AllocaInst(StructArgTy, 0, "structArg",
390 codeReplacer->getParent()->begin()->begin());
391 params.push_back(Struct);
393 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
394 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
395 Value *Idx1 = ConstantInt::get(Type::Int32Ty, i);
396 GetElementPtrInst *GEP =
397 new GetElementPtrInst(Struct, Idx0, Idx1,
398 "gep_" + StructValues[i]->getName());
399 codeReplacer->getInstList().push_back(GEP);
400 StoreInst *SI = new StoreInst(StructValues[i], GEP);
401 codeReplacer->getInstList().push_back(SI);
405 // Emit the call to the function
406 CallInst *call = new CallInst(newFunction, params,
407 NumExitBlocks > 1 ? "targetBlock" : "");
408 codeReplacer->getInstList().push_back(call);
410 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
411 unsigned FirstOut = inputs.size();
413 std::advance(OutputArgBegin, inputs.size());
415 // Reload the outputs passed in by reference
416 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
419 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
420 Value *Idx1 = ConstantInt::get(Type::Int32Ty, FirstOut + i);
421 GetElementPtrInst *GEP
422 = new GetElementPtrInst(Struct, Idx0, Idx1,
423 "gep_reload_" + outputs[i]->getName());
424 codeReplacer->getInstList().push_back(GEP);
427 Output = ReloadOutputs[i];
429 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
430 codeReplacer->getInstList().push_back(load);
431 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
432 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
433 Instruction *inst = cast<Instruction>(Users[u]);
434 if (!BlocksToExtract.count(inst->getParent()))
435 inst->replaceUsesOfWith(outputs[i], load);
439 // Now we can emit a switch statement using the call as a value.
440 SwitchInst *TheSwitch =
441 new SwitchInst(ConstantInt::getNullValue(Type::Int16Ty),
442 codeReplacer, 0, codeReplacer);
444 // Since there may be multiple exits from the original region, make the new
445 // function return an unsigned, switch on that number. This loop iterates
446 // over all of the blocks in the extracted region, updating any terminator
447 // instructions in the to-be-extracted region that branch to blocks that are
448 // not in the region to be extracted.
449 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
451 unsigned switchVal = 0;
452 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
453 e = BlocksToExtract.end(); i != e; ++i) {
454 TerminatorInst *TI = (*i)->getTerminator();
455 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
456 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
457 BasicBlock *OldTarget = TI->getSuccessor(i);
458 // add a new basic block which returns the appropriate value
459 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
461 // If we don't already have an exit stub for this non-extracted
462 // destination, create one now!
463 NewTarget = new BasicBlock(OldTarget->getName() + ".exitStub",
465 unsigned SuccNum = switchVal++;
468 switch (NumExitBlocks) {
470 case 1: break; // No value needed.
471 case 2: // Conditional branch, return a bool
472 brVal = ConstantInt::get(Type::Int1Ty, !SuccNum);
475 brVal = ConstantInt::get(Type::Int16Ty, SuccNum);
479 ReturnInst *NTRet = new ReturnInst(brVal, NewTarget);
481 // Update the switch instruction.
482 TheSwitch->addCase(ConstantInt::get(Type::Int16Ty, SuccNum),
485 // Restore values just before we exit
486 Function::arg_iterator OAI = OutputArgBegin;
487 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
488 // For an invoke, the normal destination is the only one that is
489 // dominated by the result of the invocation
490 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
492 bool DominatesDef = true;
494 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
495 DefBlock = Invoke->getNormalDest();
497 // Make sure we are looking at the original successor block, not
498 // at a newly inserted exit block, which won't be in the dominator
500 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
501 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
502 if (DefBlock == I->second) {
507 // In the extract block case, if the block we are extracting ends
508 // with an invoke instruction, make sure that we don't emit a
509 // store of the invoke value for the unwind block.
510 if (!DS && DefBlock != OldTarget)
511 DominatesDef = false;
515 DominatesDef = DS->dominates(DefBlock, OldTarget);
519 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
520 Value *Idx1 = ConstantInt::get(Type::Int32Ty,FirstOut+out);
521 GetElementPtrInst *GEP =
522 new GetElementPtrInst(OAI, Idx0, Idx1,
523 "gep_" + outputs[out]->getName(),
525 new StoreInst(outputs[out], GEP, NTRet);
527 new StoreInst(outputs[out], OAI, NTRet);
530 // Advance output iterator even if we don't emit a store
531 if (!AggregateArgs) ++OAI;
535 // rewrite the original branch instruction with this new target
536 TI->setSuccessor(i, NewTarget);
540 // Now that we've done the deed, simplify the switch instruction.
541 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
542 switch (NumExitBlocks) {
544 // There are no successors (the block containing the switch itself), which
545 // means that previously this was the last part of the function, and hence
546 // this should be rewritten as a `ret'
548 // Check if the function should return a value
549 if (OldFnRetTy == Type::VoidTy) {
550 new ReturnInst(0, TheSwitch); // Return void
551 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
552 // return what we have
553 new ReturnInst(TheSwitch->getCondition(), TheSwitch);
555 // Otherwise we must have code extracted an unwind or something, just
556 // return whatever we want.
557 new ReturnInst(Constant::getNullValue(OldFnRetTy), TheSwitch);
560 TheSwitch->getParent()->getInstList().erase(TheSwitch);
563 // Only a single destination, change the switch into an unconditional
565 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch);
566 TheSwitch->getParent()->getInstList().erase(TheSwitch);
569 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
571 TheSwitch->getParent()->getInstList().erase(TheSwitch);
574 // Otherwise, make the default destination of the switch instruction be one
575 // of the other successors.
576 TheSwitch->setOperand(0, call);
577 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
578 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
583 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
584 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
585 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
586 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
588 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
589 e = BlocksToExtract.end(); i != e; ++i) {
590 // Delete the basic block from the old function, and the list of blocks
591 oldBlocks.remove(*i);
593 // Insert this basic block into the new function
594 newBlocks.push_back(*i);
598 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
599 /// new function. Returns pointer to the new function.
603 /// find inputs and outputs for the region
605 /// for inputs: add to function as args, map input instr* to arg#
606 /// for outputs: add allocas for scalars,
607 /// add to func as args, map output instr* to arg#
609 /// rewrite func to use argument #s instead of instr*
611 /// for each scalar output in the function: at every exit, store intermediate
612 /// computed result back into memory.
614 Function *CodeExtractor::
615 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
616 if (!isEligible(code))
619 // 1) Find inputs, outputs
620 // 2) Construct new function
621 // * Add allocas for defs, pass as args by reference
622 // * Pass in uses as args
623 // 3) Move code region, add call instr to func
625 BlocksToExtract.insert(code.begin(), code.end());
627 Values inputs, outputs;
629 // Assumption: this is a single-entry code region, and the header is the first
630 // block in the region.
631 BasicBlock *header = code[0];
633 for (unsigned i = 1, e = code.size(); i != e; ++i)
634 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
636 assert(BlocksToExtract.count(*PI) &&
637 "No blocks in this region may have entries from outside the region"
638 " except for the first block!");
640 // If we have to split PHI nodes or the entry block, do so now.
641 severSplitPHINodes(header);
643 // If we have any return instructions in the region, split those blocks so
644 // that the return is not in the region.
647 Function *oldFunction = header->getParent();
649 // This takes place of the original loop
650 BasicBlock *codeReplacer = new BasicBlock("codeRepl", oldFunction, header);
652 // The new function needs a root node because other nodes can branch to the
653 // head of the region, but the entry node of a function cannot have preds.
654 BasicBlock *newFuncRoot = new BasicBlock("newFuncRoot");
655 newFuncRoot->getInstList().push_back(new BranchInst(header));
657 // Find inputs to, outputs from the code region.
658 findInputsOutputs(inputs, outputs);
660 // Construct new function based on inputs/outputs & add allocas for all defs.
661 Function *newFunction = constructFunction(inputs, outputs, header,
663 codeReplacer, oldFunction,
664 oldFunction->getParent());
666 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
668 moveCodeToFunction(newFunction);
670 // Loop over all of the PHI nodes in the header block, and change any
671 // references to the old incoming edge to be the new incoming edge.
672 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
673 PHINode *PN = cast<PHINode>(I);
674 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
675 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
676 PN->setIncomingBlock(i, newFuncRoot);
679 // Look at all successors of the codeReplacer block. If any of these blocks
680 // had PHI nodes in them, we need to update the "from" block to be the code
681 // replacer, not the original block in the extracted region.
682 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
683 succ_end(codeReplacer));
684 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
685 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
686 PHINode *PN = cast<PHINode>(I);
687 std::set<BasicBlock*> ProcessedPreds;
688 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
689 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
690 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
691 PN->setIncomingBlock(i, codeReplacer);
693 // There were multiple entries in the PHI for this block, now there
694 // is only one, so remove the duplicated entries.
695 PN->removeIncomingValue(i, false);
700 //cerr << "NEW FUNCTION: " << *newFunction;
701 // verifyFunction(*newFunction);
703 // cerr << "OLD FUNCTION: " << *oldFunction;
704 // verifyFunction(*oldFunction);
706 DEBUG(if (verifyFunction(*newFunction)) abort());
710 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
711 // Deny code region if it contains allocas or vastarts.
712 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
714 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
716 if (isa<AllocaInst>(*I))
718 else if (const CallInst *CI = dyn_cast<CallInst>(I))
719 if (const Function *F = CI->getCalledFunction())
720 if (F->getIntrinsicID() == Intrinsic::vastart)
726 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
729 Function* llvm::ExtractCodeRegion(DominatorSet &DS,
730 const std::vector<BasicBlock*> &code,
731 bool AggregateArgs) {
732 return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(code);
735 /// ExtractBasicBlock - slurp a natural loop into a brand new function
737 Function* llvm::ExtractLoop(DominatorSet &DS, Loop *L, bool AggregateArgs) {
738 return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(L->getBlocks());
741 /// ExtractBasicBlock - slurp a basic block into a brand new function
743 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
744 std::vector<BasicBlock*> Blocks;
745 Blocks.push_back(BB);
746 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);