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/Debug.h"
29 #include "llvm/ADT/StringExtras.h"
34 // Provide a command-line option to aggregate function arguments into a struct
35 // for functions produced by the code extrator. This is useful when converting
36 // extracted functions to pthread-based code, as only one argument (void*) can
37 // be passed in to pthread_create().
39 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
40 cl::desc("Aggregate arguments to code-extracted functions"));
44 typedef std::vector<Value*> Values;
45 std::set<BasicBlock*> BlocksToExtract;
48 unsigned NumExitBlocks;
51 CodeExtractor(DominatorSet *ds = 0, bool AggArgs = false)
52 : DS(ds), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
54 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
56 bool isEligible(const std::vector<BasicBlock*> &code);
59 /// definedInRegion - Return true if the specified value is defined in the
61 bool definedInRegion(Value *V) const {
62 if (Instruction *I = dyn_cast<Instruction>(V))
63 if (BlocksToExtract.count(I->getParent()))
68 /// definedInCaller - Return true if the specified value is defined in the
69 /// function being code extracted, but not in the region being extracted.
70 /// These values must be passed in as live-ins to the function.
71 bool definedInCaller(Value *V) const {
72 if (isa<Argument>(V)) return true;
73 if (Instruction *I = dyn_cast<Instruction>(V))
74 if (!BlocksToExtract.count(I->getParent()))
79 void severSplitPHINodes(BasicBlock *&Header);
80 void splitReturnBlocks();
81 void findInputsOutputs(Values &inputs, Values &outputs);
83 Function *constructFunction(const Values &inputs,
84 const Values &outputs,
86 BasicBlock *newRootNode, BasicBlock *newHeader,
87 Function *oldFunction, Module *M);
89 void moveCodeToFunction(Function *newFunction);
91 void emitCallAndSwitchStatement(Function *newFunction,
92 BasicBlock *newHeader,
99 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
100 /// region, we need to split the entry block of the region so that the PHI node
101 /// is easier to deal with.
102 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
103 bool HasPredsFromRegion = false;
104 unsigned NumPredsOutsideRegion = 0;
106 if (Header != &Header->getParent()->front()) {
107 PHINode *PN = dyn_cast<PHINode>(Header->begin());
108 if (!PN) return; // No PHI nodes.
110 // If the header node contains any PHI nodes, check to see if there is more
111 // than one entry from outside the region. If so, we need to sever the
112 // header block into two.
113 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
114 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
115 HasPredsFromRegion = true;
117 ++NumPredsOutsideRegion;
119 // If there is one (or fewer) predecessor from outside the region, we don't
120 // need to do anything special.
121 if (NumPredsOutsideRegion <= 1) return;
124 // Otherwise, we need to split the header block into two pieces: one
125 // containing PHI nodes merging values from outside of the region, and a
126 // second that contains all of the code for the block and merges back any
127 // incoming values from inside of the region.
128 BasicBlock::iterator AfterPHIs = Header->begin();
129 while (isa<PHINode>(AfterPHIs)) ++AfterPHIs;
130 BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
131 Header->getName()+".ce");
133 // We only want to code extract the second block now, and it becomes the new
134 // header of the region.
135 BasicBlock *OldPred = Header;
136 BlocksToExtract.erase(OldPred);
137 BlocksToExtract.insert(NewBB);
140 // Okay, update dominator sets. The blocks that dominate the new one are the
141 // blocks that dominate TIBB plus the new block itself.
143 DominatorSet::DomSetType DomSet = DS->getDominators(OldPred);
144 DomSet.insert(NewBB); // A block always dominates itself.
145 DS->addBasicBlock(NewBB, DomSet);
147 // Additionally, NewBB dominates all blocks in the function that are
148 // dominated by OldPred.
149 Function *F = Header->getParent();
150 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
151 if (DS->properlyDominates(OldPred, I))
152 DS->addDominator(I, NewBB);
155 // Okay, now we need to adjust the PHI nodes and any branches from within the
156 // region to go to the new header block instead of the old header block.
157 if (HasPredsFromRegion) {
158 PHINode *PN = cast<PHINode>(OldPred->begin());
159 // Loop over all of the predecessors of OldPred that are in the region,
160 // changing them to branch to NewBB instead.
161 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
162 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
163 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
164 TI->replaceUsesOfWith(OldPred, NewBB);
167 // Okay, everthing within the region is now branching to the right block, we
168 // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
169 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
170 PHINode *PN = cast<PHINode>(AfterPHIs);
171 // Create a new PHI node in the new region, which has an incoming value
172 // from OldPred of PN.
173 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".ce",
175 NewPN->addIncoming(PN, OldPred);
177 // Loop over all of the incoming value in PN, moving them to NewPN if they
178 // are from the extracted region.
179 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
180 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
181 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
182 PN->removeIncomingValue(i);
190 void CodeExtractor::splitReturnBlocks() {
191 for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(),
192 E = BlocksToExtract.end(); I != E; ++I)
193 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator()))
194 (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
197 // findInputsOutputs - Find inputs to, outputs from the code region.
199 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
200 std::set<BasicBlock*> ExitBlocks;
201 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
202 ce = BlocksToExtract.end(); ci != ce; ++ci) {
203 BasicBlock *BB = *ci;
205 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
206 // If a used value is defined outside the region, it's an input. If an
207 // instruction is used outside the region, it's an output.
208 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
209 if (definedInCaller(*O))
210 inputs.push_back(*O);
212 // Consider uses of this instruction (outputs).
213 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
215 if (!definedInRegion(*UI)) {
216 outputs.push_back(I);
221 // Keep track of the exit blocks from the region.
222 TerminatorInst *TI = BB->getTerminator();
223 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
224 if (!BlocksToExtract.count(TI->getSuccessor(i)))
225 ExitBlocks.insert(TI->getSuccessor(i));
226 } // for: basic blocks
228 NumExitBlocks = ExitBlocks.size();
230 // Eliminate duplicates.
231 std::sort(inputs.begin(), inputs.end());
232 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
233 std::sort(outputs.begin(), outputs.end());
234 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
237 /// constructFunction - make a function based on inputs and outputs, as follows:
238 /// f(in0, ..., inN, out0, ..., outN)
240 Function *CodeExtractor::constructFunction(const Values &inputs,
241 const Values &outputs,
243 BasicBlock *newRootNode,
244 BasicBlock *newHeader,
245 Function *oldFunction,
247 DOUT << "inputs: " << inputs.size() << "\n";
248 DOUT << "outputs: " << outputs.size() << "\n";
250 // This function returns unsigned, outputs will go back by reference.
251 switch (NumExitBlocks) {
253 case 1: RetTy = Type::VoidTy; break;
254 case 2: RetTy = Type::Int1Ty; break;
255 default: RetTy = Type::Int16Ty; break;
258 std::vector<const Type*> paramTy;
260 // Add the types of the input values to the function's argument list
261 for (Values::const_iterator i = inputs.begin(),
262 e = inputs.end(); i != e; ++i) {
263 const Value *value = *i;
264 DOUT << "value used in func: " << *value << "\n";
265 paramTy.push_back(value->getType());
268 // Add the types of the output values to the function's argument list.
269 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
271 DOUT << "instr used in func: " << **I << "\n";
273 paramTy.push_back((*I)->getType());
275 paramTy.push_back(PointerType::get((*I)->getType()));
278 DOUT << "Function type: " << *RetTy << " f(";
279 for (std::vector<const Type*>::iterator i = paramTy.begin(),
280 e = paramTy.end(); i != e; ++i)
284 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
285 PointerType *StructPtr = PointerType::get(StructType::get(paramTy));
287 paramTy.push_back(StructPtr);
289 const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false);
291 // Create the new function
292 Function *newFunction = new Function(funcType,
293 GlobalValue::InternalLinkage,
294 oldFunction->getName() + "_" +
295 header->getName(), M);
296 newFunction->getBasicBlockList().push_back(newRootNode);
298 // Create an iterator to name all of the arguments we inserted.
299 Function::arg_iterator AI = newFunction->arg_begin();
301 // Rewrite all users of the inputs in the extracted region to use the
302 // arguments (or appropriate addressing into struct) instead.
303 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
306 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
307 Value *Idx1 = ConstantInt::get(Type::Int32Ty, i);
308 std::string GEPname = "gep_" + inputs[i]->getName();
309 TerminatorInst *TI = newFunction->begin()->getTerminator();
310 GetElementPtrInst *GEP = new GetElementPtrInst(AI, Idx0, Idx1,
312 RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
316 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
317 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
319 if (Instruction* inst = dyn_cast<Instruction>(*use))
320 if (BlocksToExtract.count(inst->getParent()))
321 inst->replaceUsesOfWith(inputs[i], RewriteVal);
324 // Set names for input and output arguments.
325 if (!AggregateArgs) {
326 AI = newFunction->arg_begin();
327 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
328 AI->setName(inputs[i]->getName());
329 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
330 AI->setName(outputs[i]->getName()+".out");
333 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
334 // within the new function. This must be done before we lose track of which
335 // blocks were originally in the code region.
336 std::vector<User*> Users(header->use_begin(), header->use_end());
337 for (unsigned i = 0, e = Users.size(); i != e; ++i)
338 // The BasicBlock which contains the branch is not in the region
339 // modify the branch target to a new block
340 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
341 if (!BlocksToExtract.count(TI->getParent()) &&
342 TI->getParent()->getParent() == oldFunction)
343 TI->replaceUsesOfWith(header, newHeader);
348 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
349 /// the call instruction, splitting any PHI nodes in the header block as
352 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
353 Values &inputs, Values &outputs) {
354 // Emit a call to the new function, passing in: *pointer to struct (if
355 // aggregating parameters), or plan inputs and allocated memory for outputs
356 std::vector<Value*> params, StructValues, ReloadOutputs;
358 // Add inputs as params, or to be filled into the struct
359 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
361 StructValues.push_back(*i);
363 params.push_back(*i);
365 // Create allocas for the outputs
366 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
368 StructValues.push_back(*i);
371 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
372 codeReplacer->getParent()->begin()->begin());
373 ReloadOutputs.push_back(alloca);
374 params.push_back(alloca);
378 AllocaInst *Struct = 0;
379 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
380 std::vector<const Type*> ArgTypes;
381 for (Values::iterator v = StructValues.begin(),
382 ve = StructValues.end(); v != ve; ++v)
383 ArgTypes.push_back((*v)->getType());
385 // Allocate a struct at the beginning of this function
386 Type *StructArgTy = StructType::get(ArgTypes);
388 new AllocaInst(StructArgTy, 0, "structArg",
389 codeReplacer->getParent()->begin()->begin());
390 params.push_back(Struct);
392 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
393 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
394 Value *Idx1 = ConstantInt::get(Type::Int32Ty, i);
395 GetElementPtrInst *GEP =
396 new GetElementPtrInst(Struct, Idx0, Idx1,
397 "gep_" + StructValues[i]->getName());
398 codeReplacer->getInstList().push_back(GEP);
399 StoreInst *SI = new StoreInst(StructValues[i], GEP);
400 codeReplacer->getInstList().push_back(SI);
404 // Emit the call to the function
405 CallInst *call = new CallInst(newFunction, params,
406 NumExitBlocks > 1 ? "targetBlock" : "");
407 codeReplacer->getInstList().push_back(call);
409 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
410 unsigned FirstOut = inputs.size();
412 std::advance(OutputArgBegin, inputs.size());
414 // Reload the outputs passed in by reference
415 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
418 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
419 Value *Idx1 = ConstantInt::get(Type::Int32Ty, FirstOut + i);
420 GetElementPtrInst *GEP
421 = new GetElementPtrInst(Struct, Idx0, Idx1,
422 "gep_reload_" + outputs[i]->getName());
423 codeReplacer->getInstList().push_back(GEP);
426 Output = ReloadOutputs[i];
428 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
429 codeReplacer->getInstList().push_back(load);
430 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
431 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
432 Instruction *inst = cast<Instruction>(Users[u]);
433 if (!BlocksToExtract.count(inst->getParent()))
434 inst->replaceUsesOfWith(outputs[i], load);
438 // Now we can emit a switch statement using the call as a value.
439 SwitchInst *TheSwitch =
440 new SwitchInst(ConstantInt::getNullValue(Type::Int16Ty),
441 codeReplacer, 0, codeReplacer);
443 // Since there may be multiple exits from the original region, make the new
444 // function return an unsigned, switch on that number. This loop iterates
445 // over all of the blocks in the extracted region, updating any terminator
446 // instructions in the to-be-extracted region that branch to blocks that are
447 // not in the region to be extracted.
448 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
450 unsigned switchVal = 0;
451 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
452 e = BlocksToExtract.end(); i != e; ++i) {
453 TerminatorInst *TI = (*i)->getTerminator();
454 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
455 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
456 BasicBlock *OldTarget = TI->getSuccessor(i);
457 // add a new basic block which returns the appropriate value
458 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
460 // If we don't already have an exit stub for this non-extracted
461 // destination, create one now!
462 NewTarget = new BasicBlock(OldTarget->getName() + ".exitStub",
464 unsigned SuccNum = switchVal++;
467 switch (NumExitBlocks) {
469 case 1: break; // No value needed.
470 case 2: // Conditional branch, return a bool
471 brVal = ConstantInt::get(Type::Int1Ty, !SuccNum);
474 brVal = ConstantInt::get(Type::Int16Ty, SuccNum);
478 ReturnInst *NTRet = new ReturnInst(brVal, NewTarget);
480 // Update the switch instruction.
481 TheSwitch->addCase(ConstantInt::get(Type::Int16Ty, SuccNum),
484 // Restore values just before we exit
485 Function::arg_iterator OAI = OutputArgBegin;
486 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
487 // For an invoke, the normal destination is the only one that is
488 // dominated by the result of the invocation
489 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
491 bool DominatesDef = true;
493 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
494 DefBlock = Invoke->getNormalDest();
496 // Make sure we are looking at the original successor block, not
497 // at a newly inserted exit block, which won't be in the dominator
499 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
500 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
501 if (DefBlock == I->second) {
506 // In the extract block case, if the block we are extracting ends
507 // with an invoke instruction, make sure that we don't emit a
508 // store of the invoke value for the unwind block.
509 if (!DS && DefBlock != OldTarget)
510 DominatesDef = false;
514 DominatesDef = DS->dominates(DefBlock, OldTarget);
518 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
519 Value *Idx1 = ConstantInt::get(Type::Int32Ty,FirstOut+out);
520 GetElementPtrInst *GEP =
521 new GetElementPtrInst(OAI, Idx0, Idx1,
522 "gep_" + outputs[out]->getName(),
524 new StoreInst(outputs[out], GEP, NTRet);
526 new StoreInst(outputs[out], OAI, NTRet);
529 // Advance output iterator even if we don't emit a store
530 if (!AggregateArgs) ++OAI;
534 // rewrite the original branch instruction with this new target
535 TI->setSuccessor(i, NewTarget);
539 // Now that we've done the deed, simplify the switch instruction.
540 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
541 switch (NumExitBlocks) {
543 // There are no successors (the block containing the switch itself), which
544 // means that previously this was the last part of the function, and hence
545 // this should be rewritten as a `ret'
547 // Check if the function should return a value
548 if (OldFnRetTy == Type::VoidTy) {
549 new ReturnInst(0, TheSwitch); // Return void
550 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
551 // return what we have
552 new ReturnInst(TheSwitch->getCondition(), TheSwitch);
554 // Otherwise we must have code extracted an unwind or something, just
555 // return whatever we want.
556 new ReturnInst(Constant::getNullValue(OldFnRetTy), TheSwitch);
559 TheSwitch->getParent()->getInstList().erase(TheSwitch);
562 // Only a single destination, change the switch into an unconditional
564 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch);
565 TheSwitch->getParent()->getInstList().erase(TheSwitch);
568 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
570 TheSwitch->getParent()->getInstList().erase(TheSwitch);
573 // Otherwise, make the default destination of the switch instruction be one
574 // of the other successors.
575 TheSwitch->setOperand(0, call);
576 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
577 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
582 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
583 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
584 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
585 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
587 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
588 e = BlocksToExtract.end(); i != e; ++i) {
589 // Delete the basic block from the old function, and the list of blocks
590 oldBlocks.remove(*i);
592 // Insert this basic block into the new function
593 newBlocks.push_back(*i);
597 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
598 /// new function. Returns pointer to the new function.
602 /// find inputs and outputs for the region
604 /// for inputs: add to function as args, map input instr* to arg#
605 /// for outputs: add allocas for scalars,
606 /// add to func as args, map output instr* to arg#
608 /// rewrite func to use argument #s instead of instr*
610 /// for each scalar output in the function: at every exit, store intermediate
611 /// computed result back into memory.
613 Function *CodeExtractor::
614 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
615 if (!isEligible(code))
618 // 1) Find inputs, outputs
619 // 2) Construct new function
620 // * Add allocas for defs, pass as args by reference
621 // * Pass in uses as args
622 // 3) Move code region, add call instr to func
624 BlocksToExtract.insert(code.begin(), code.end());
626 Values inputs, outputs;
628 // Assumption: this is a single-entry code region, and the header is the first
629 // block in the region.
630 BasicBlock *header = code[0];
632 for (unsigned i = 1, e = code.size(); i != e; ++i)
633 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
635 assert(BlocksToExtract.count(*PI) &&
636 "No blocks in this region may have entries from outside the region"
637 " except for the first block!");
639 // If we have to split PHI nodes or the entry block, do so now.
640 severSplitPHINodes(header);
642 // If we have any return instructions in the region, split those blocks so
643 // that the return is not in the region.
646 Function *oldFunction = header->getParent();
648 // This takes place of the original loop
649 BasicBlock *codeReplacer = new BasicBlock("codeRepl", oldFunction, header);
651 // The new function needs a root node because other nodes can branch to the
652 // head of the region, but the entry node of a function cannot have preds.
653 BasicBlock *newFuncRoot = new BasicBlock("newFuncRoot");
654 newFuncRoot->getInstList().push_back(new BranchInst(header));
656 // Find inputs to, outputs from the code region.
657 findInputsOutputs(inputs, outputs);
659 // Construct new function based on inputs/outputs & add allocas for all defs.
660 Function *newFunction = constructFunction(inputs, outputs, header,
662 codeReplacer, oldFunction,
663 oldFunction->getParent());
665 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
667 moveCodeToFunction(newFunction);
669 // Loop over all of the PHI nodes in the header block, and change any
670 // references to the old incoming edge to be the new incoming edge.
671 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
672 PHINode *PN = cast<PHINode>(I);
673 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
674 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
675 PN->setIncomingBlock(i, newFuncRoot);
678 // Look at all successors of the codeReplacer block. If any of these blocks
679 // had PHI nodes in them, we need to update the "from" block to be the code
680 // replacer, not the original block in the extracted region.
681 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
682 succ_end(codeReplacer));
683 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
684 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
685 PHINode *PN = cast<PHINode>(I);
686 std::set<BasicBlock*> ProcessedPreds;
687 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
688 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
689 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
690 PN->setIncomingBlock(i, codeReplacer);
692 // There were multiple entries in the PHI for this block, now there
693 // is only one, so remove the duplicated entries.
694 PN->removeIncomingValue(i, false);
699 //cerr << "NEW FUNCTION: " << *newFunction;
700 // verifyFunction(*newFunction);
702 // cerr << "OLD FUNCTION: " << *oldFunction;
703 // verifyFunction(*oldFunction);
705 DEBUG(if (verifyFunction(*newFunction)) abort());
709 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
710 // Deny code region if it contains allocas or vastarts.
711 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
713 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
715 if (isa<AllocaInst>(*I))
717 else if (const CallInst *CI = dyn_cast<CallInst>(I))
718 if (const Function *F = CI->getCalledFunction())
719 if (F->getIntrinsicID() == Intrinsic::vastart)
725 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
728 Function* llvm::ExtractCodeRegion(DominatorSet &DS,
729 const std::vector<BasicBlock*> &code,
730 bool AggregateArgs) {
731 return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(code);
734 /// ExtractBasicBlock - slurp a natural loop into a brand new function
736 Function* llvm::ExtractLoop(DominatorSet &DS, Loop *L, bool AggregateArgs) {
737 return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(L->getBlocks());
740 /// ExtractBasicBlock - slurp a basic block into a brand new function
742 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
743 std::vector<BasicBlock*> Blocks;
744 Blocks.push_back(BB);
745 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);