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;
50 unsigned NumExitBlocks;
53 CodeExtractor(ETForest *ef = 0, DominatorTree* dt = 0, bool AggArgs = false)
54 : EF(ef), DT(dt), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
56 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
58 bool isEligible(const std::vector<BasicBlock*> &code);
61 /// definedInRegion - Return true if the specified value is defined in the
63 bool definedInRegion(Value *V) const {
64 if (Instruction *I = dyn_cast<Instruction>(V))
65 if (BlocksToExtract.count(I->getParent()))
70 /// definedInCaller - Return true if the specified value is defined in the
71 /// function being code extracted, but not in the region being extracted.
72 /// These values must be passed in as live-ins to the function.
73 bool definedInCaller(Value *V) const {
74 if (isa<Argument>(V)) return true;
75 if (Instruction *I = dyn_cast<Instruction>(V))
76 if (!BlocksToExtract.count(I->getParent()))
81 void severSplitPHINodes(BasicBlock *&Header);
82 void splitReturnBlocks();
83 void findInputsOutputs(Values &inputs, Values &outputs);
85 Function *constructFunction(const Values &inputs,
86 const Values &outputs,
88 BasicBlock *newRootNode, BasicBlock *newHeader,
89 Function *oldFunction, Module *M);
91 void moveCodeToFunction(Function *newFunction);
93 void emitCallAndSwitchStatement(Function *newFunction,
94 BasicBlock *newHeader,
101 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
102 /// region, we need to split the entry block of the region so that the PHI node
103 /// is easier to deal with.
104 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
105 bool HasPredsFromRegion = false;
106 unsigned NumPredsOutsideRegion = 0;
108 if (Header != &Header->getParent()->getEntryBlock()) {
109 PHINode *PN = dyn_cast<PHINode>(Header->begin());
110 if (!PN) return; // No PHI nodes.
112 // If the header node contains any PHI nodes, check to see if there is more
113 // than one entry from outside the region. If so, we need to sever the
114 // header block into two.
115 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
116 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
117 HasPredsFromRegion = true;
119 ++NumPredsOutsideRegion;
121 // If there is one (or fewer) predecessor from outside the region, we don't
122 // need to do anything special.
123 if (NumPredsOutsideRegion <= 1) return;
126 // Otherwise, we need to split the header block into two pieces: one
127 // containing PHI nodes merging values from outside of the region, and a
128 // second that contains all of the code for the block and merges back any
129 // incoming values from inside of the region.
130 BasicBlock::iterator AfterPHIs = Header->begin();
131 while (isa<PHINode>(AfterPHIs)) ++AfterPHIs;
132 BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
133 Header->getName()+".ce");
135 // We only want to code extract the second block now, and it becomes the new
136 // header of the region.
137 BasicBlock *OldPred = Header;
138 BlocksToExtract.erase(OldPred);
139 BlocksToExtract.insert(NewBB);
142 // Okay, update dominator sets. The blocks that dominate the new one are the
143 // blocks that dominate TIBB plus the new block itself.
145 DominatorTree::Node* idom = DT->getNode(OldPred)->getIDom();
146 DT->createNewNode(NewBB, idom);
147 EF->addNewBlock(NewBB, idom->getBlock());
149 // Additionally, NewBB replaces OldPred as the immediate dominator of blocks
150 Function *F = Header->getParent();
151 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
152 if (DT->getNode(I)->getIDom()->getBlock() == OldPred) {
153 DT->changeImmediateDominator(DT->getNode(I), DT->getNode(NewBB));
154 EF->setImmediateDominator(I, NewBB);
158 // Okay, now we need to adjust the PHI nodes and any branches from within the
159 // region to go to the new header block instead of the old header block.
160 if (HasPredsFromRegion) {
161 PHINode *PN = cast<PHINode>(OldPred->begin());
162 // Loop over all of the predecessors of OldPred that are in the region,
163 // changing them to branch to NewBB instead.
164 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
165 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
166 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
167 TI->replaceUsesOfWith(OldPred, NewBB);
170 // Okay, everthing within the region is now branching to the right block, we
171 // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
172 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
173 PHINode *PN = cast<PHINode>(AfterPHIs);
174 // Create a new PHI node in the new region, which has an incoming value
175 // from OldPred of PN.
176 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".ce",
178 NewPN->addIncoming(PN, OldPred);
180 // Loop over all of the incoming value in PN, moving them to NewPN if they
181 // are from the extracted region.
182 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
183 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
184 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
185 PN->removeIncomingValue(i);
193 void CodeExtractor::splitReturnBlocks() {
194 for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(),
195 E = BlocksToExtract.end(); I != E; ++I)
196 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator()))
197 (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
200 // findInputsOutputs - Find inputs to, outputs from the code region.
202 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
203 std::set<BasicBlock*> ExitBlocks;
204 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
205 ce = BlocksToExtract.end(); ci != ce; ++ci) {
206 BasicBlock *BB = *ci;
208 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
209 // If a used value is defined outside the region, it's an input. If an
210 // instruction is used outside the region, it's an output.
211 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
212 if (definedInCaller(*O))
213 inputs.push_back(*O);
215 // Consider uses of this instruction (outputs).
216 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
218 if (!definedInRegion(*UI)) {
219 outputs.push_back(I);
224 // Keep track of the exit blocks from the region.
225 TerminatorInst *TI = BB->getTerminator();
226 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
227 if (!BlocksToExtract.count(TI->getSuccessor(i)))
228 ExitBlocks.insert(TI->getSuccessor(i));
229 } // for: basic blocks
231 NumExitBlocks = ExitBlocks.size();
233 // Eliminate duplicates.
234 std::sort(inputs.begin(), inputs.end());
235 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
236 std::sort(outputs.begin(), outputs.end());
237 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
240 /// constructFunction - make a function based on inputs and outputs, as follows:
241 /// f(in0, ..., inN, out0, ..., outN)
243 Function *CodeExtractor::constructFunction(const Values &inputs,
244 const Values &outputs,
246 BasicBlock *newRootNode,
247 BasicBlock *newHeader,
248 Function *oldFunction,
250 DOUT << "inputs: " << inputs.size() << "\n";
251 DOUT << "outputs: " << outputs.size() << "\n";
253 // This function returns unsigned, outputs will go back by reference.
254 switch (NumExitBlocks) {
256 case 1: RetTy = Type::VoidTy; break;
257 case 2: RetTy = Type::Int1Ty; break;
258 default: RetTy = Type::Int16Ty; break;
261 std::vector<const Type*> paramTy;
263 // Add the types of the input values to the function's argument list
264 for (Values::const_iterator i = inputs.begin(),
265 e = inputs.end(); i != e; ++i) {
266 const Value *value = *i;
267 DOUT << "value used in func: " << *value << "\n";
268 paramTy.push_back(value->getType());
271 // Add the types of the output values to the function's argument list.
272 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
274 DOUT << "instr used in func: " << **I << "\n";
276 paramTy.push_back((*I)->getType());
278 paramTy.push_back(PointerType::get((*I)->getType()));
281 DOUT << "Function type: " << *RetTy << " f(";
282 for (std::vector<const Type*>::iterator i = paramTy.begin(),
283 e = paramTy.end(); i != e; ++i)
287 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
288 PointerType *StructPtr = PointerType::get(StructType::get(paramTy));
290 paramTy.push_back(StructPtr);
292 const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false);
294 // Create the new function
295 Function *newFunction = new Function(funcType,
296 GlobalValue::InternalLinkage,
297 oldFunction->getName() + "_" +
298 header->getName(), M);
299 newFunction->getBasicBlockList().push_back(newRootNode);
301 // Create an iterator to name all of the arguments we inserted.
302 Function::arg_iterator AI = newFunction->arg_begin();
304 // Rewrite all users of the inputs in the extracted region to use the
305 // arguments (or appropriate addressing into struct) instead.
306 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
309 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
310 Value *Idx1 = ConstantInt::get(Type::Int32Ty, i);
311 std::string GEPname = "gep_" + inputs[i]->getName();
312 TerminatorInst *TI = newFunction->begin()->getTerminator();
313 GetElementPtrInst *GEP = new GetElementPtrInst(AI, Idx0, Idx1,
315 RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
319 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
320 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
322 if (Instruction* inst = dyn_cast<Instruction>(*use))
323 if (BlocksToExtract.count(inst->getParent()))
324 inst->replaceUsesOfWith(inputs[i], RewriteVal);
327 // Set names for input and output arguments.
328 if (!AggregateArgs) {
329 AI = newFunction->arg_begin();
330 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
331 AI->setName(inputs[i]->getName());
332 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
333 AI->setName(outputs[i]->getName()+".out");
336 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
337 // within the new function. This must be done before we lose track of which
338 // blocks were originally in the code region.
339 std::vector<User*> Users(header->use_begin(), header->use_end());
340 for (unsigned i = 0, e = Users.size(); i != e; ++i)
341 // The BasicBlock which contains the branch is not in the region
342 // modify the branch target to a new block
343 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
344 if (!BlocksToExtract.count(TI->getParent()) &&
345 TI->getParent()->getParent() == oldFunction)
346 TI->replaceUsesOfWith(header, newHeader);
351 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
352 /// the call instruction, splitting any PHI nodes in the header block as
355 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
356 Values &inputs, Values &outputs) {
357 // Emit a call to the new function, passing in: *pointer to struct (if
358 // aggregating parameters), or plan inputs and allocated memory for outputs
359 std::vector<Value*> params, StructValues, ReloadOutputs;
361 // Add inputs as params, or to be filled into the struct
362 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
364 StructValues.push_back(*i);
366 params.push_back(*i);
368 // Create allocas for the outputs
369 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
371 StructValues.push_back(*i);
374 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
375 codeReplacer->getParent()->begin()->begin());
376 ReloadOutputs.push_back(alloca);
377 params.push_back(alloca);
381 AllocaInst *Struct = 0;
382 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
383 std::vector<const Type*> ArgTypes;
384 for (Values::iterator v = StructValues.begin(),
385 ve = StructValues.end(); v != ve; ++v)
386 ArgTypes.push_back((*v)->getType());
388 // Allocate a struct at the beginning of this function
389 Type *StructArgTy = StructType::get(ArgTypes);
391 new AllocaInst(StructArgTy, 0, "structArg",
392 codeReplacer->getParent()->begin()->begin());
393 params.push_back(Struct);
395 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
396 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
397 Value *Idx1 = ConstantInt::get(Type::Int32Ty, i);
398 GetElementPtrInst *GEP =
399 new GetElementPtrInst(Struct, Idx0, Idx1,
400 "gep_" + StructValues[i]->getName());
401 codeReplacer->getInstList().push_back(GEP);
402 StoreInst *SI = new StoreInst(StructValues[i], GEP);
403 codeReplacer->getInstList().push_back(SI);
407 // Emit the call to the function
408 CallInst *call = new CallInst(newFunction, ¶ms[0], params.size(),
409 NumExitBlocks > 1 ? "targetBlock" : "");
410 codeReplacer->getInstList().push_back(call);
412 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
413 unsigned FirstOut = inputs.size();
415 std::advance(OutputArgBegin, inputs.size());
417 // Reload the outputs passed in by reference
418 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
421 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
422 Value *Idx1 = ConstantInt::get(Type::Int32Ty, FirstOut + i);
423 GetElementPtrInst *GEP
424 = new GetElementPtrInst(Struct, Idx0, Idx1,
425 "gep_reload_" + outputs[i]->getName());
426 codeReplacer->getInstList().push_back(GEP);
429 Output = ReloadOutputs[i];
431 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
432 codeReplacer->getInstList().push_back(load);
433 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
434 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
435 Instruction *inst = cast<Instruction>(Users[u]);
436 if (!BlocksToExtract.count(inst->getParent()))
437 inst->replaceUsesOfWith(outputs[i], load);
441 // Now we can emit a switch statement using the call as a value.
442 SwitchInst *TheSwitch =
443 new SwitchInst(ConstantInt::getNullValue(Type::Int16Ty),
444 codeReplacer, 0, codeReplacer);
446 // Since there may be multiple exits from the original region, make the new
447 // function return an unsigned, switch on that number. This loop iterates
448 // over all of the blocks in the extracted region, updating any terminator
449 // instructions in the to-be-extracted region that branch to blocks that are
450 // not in the region to be extracted.
451 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
453 unsigned switchVal = 0;
454 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
455 e = BlocksToExtract.end(); i != e; ++i) {
456 TerminatorInst *TI = (*i)->getTerminator();
457 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
458 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
459 BasicBlock *OldTarget = TI->getSuccessor(i);
460 // add a new basic block which returns the appropriate value
461 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
463 // If we don't already have an exit stub for this non-extracted
464 // destination, create one now!
465 NewTarget = new BasicBlock(OldTarget->getName() + ".exitStub",
467 unsigned SuccNum = switchVal++;
470 switch (NumExitBlocks) {
472 case 1: break; // No value needed.
473 case 2: // Conditional branch, return a bool
474 brVal = ConstantInt::get(Type::Int1Ty, !SuccNum);
477 brVal = ConstantInt::get(Type::Int16Ty, SuccNum);
481 ReturnInst *NTRet = new ReturnInst(brVal, NewTarget);
483 // Update the switch instruction.
484 TheSwitch->addCase(ConstantInt::get(Type::Int16Ty, SuccNum),
487 // Restore values just before we exit
488 Function::arg_iterator OAI = OutputArgBegin;
489 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
490 // For an invoke, the normal destination is the only one that is
491 // dominated by the result of the invocation
492 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
494 bool DominatesDef = true;
496 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
497 DefBlock = Invoke->getNormalDest();
499 // Make sure we are looking at the original successor block, not
500 // at a newly inserted exit block, which won't be in the dominator
502 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
503 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
504 if (DefBlock == I->second) {
509 // In the extract block case, if the block we are extracting ends
510 // with an invoke instruction, make sure that we don't emit a
511 // store of the invoke value for the unwind block.
512 if (!EF && DefBlock != OldTarget)
513 DominatesDef = false;
517 DominatesDef = EF->dominates(DefBlock, OldTarget);
521 Value *Idx0 = Constant::getNullValue(Type::Int32Ty);
522 Value *Idx1 = ConstantInt::get(Type::Int32Ty,FirstOut+out);
523 GetElementPtrInst *GEP =
524 new GetElementPtrInst(OAI, Idx0, Idx1,
525 "gep_" + outputs[out]->getName(),
527 new StoreInst(outputs[out], GEP, NTRet);
529 new StoreInst(outputs[out], OAI, NTRet);
532 // Advance output iterator even if we don't emit a store
533 if (!AggregateArgs) ++OAI;
537 // rewrite the original branch instruction with this new target
538 TI->setSuccessor(i, NewTarget);
542 // Now that we've done the deed, simplify the switch instruction.
543 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
544 switch (NumExitBlocks) {
546 // There are no successors (the block containing the switch itself), which
547 // means that previously this was the last part of the function, and hence
548 // this should be rewritten as a `ret'
550 // Check if the function should return a value
551 if (OldFnRetTy == Type::VoidTy) {
552 new ReturnInst(0, TheSwitch); // Return void
553 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
554 // return what we have
555 new ReturnInst(TheSwitch->getCondition(), TheSwitch);
557 // Otherwise we must have code extracted an unwind or something, just
558 // return whatever we want.
559 new ReturnInst(Constant::getNullValue(OldFnRetTy), TheSwitch);
562 TheSwitch->getParent()->getInstList().erase(TheSwitch);
565 // Only a single destination, change the switch into an unconditional
567 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch);
568 TheSwitch->getParent()->getInstList().erase(TheSwitch);
571 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
573 TheSwitch->getParent()->getInstList().erase(TheSwitch);
576 // Otherwise, make the default destination of the switch instruction be one
577 // of the other successors.
578 TheSwitch->setOperand(0, call);
579 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
580 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
585 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
586 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
587 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
588 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
590 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
591 e = BlocksToExtract.end(); i != e; ++i) {
592 // Delete the basic block from the old function, and the list of blocks
593 oldBlocks.remove(*i);
595 // Insert this basic block into the new function
596 newBlocks.push_back(*i);
600 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
601 /// new function. Returns pointer to the new function.
605 /// find inputs and outputs for the region
607 /// for inputs: add to function as args, map input instr* to arg#
608 /// for outputs: add allocas for scalars,
609 /// add to func as args, map output instr* to arg#
611 /// rewrite func to use argument #s instead of instr*
613 /// for each scalar output in the function: at every exit, store intermediate
614 /// computed result back into memory.
616 Function *CodeExtractor::
617 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
618 if (!isEligible(code))
621 // 1) Find inputs, outputs
622 // 2) Construct new function
623 // * Add allocas for defs, pass as args by reference
624 // * Pass in uses as args
625 // 3) Move code region, add call instr to func
627 BlocksToExtract.insert(code.begin(), code.end());
629 Values inputs, outputs;
631 // Assumption: this is a single-entry code region, and the header is the first
632 // block in the region.
633 BasicBlock *header = code[0];
635 for (unsigned i = 1, e = code.size(); i != e; ++i)
636 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
638 assert(BlocksToExtract.count(*PI) &&
639 "No blocks in this region may have entries from outside the region"
640 " except for the first block!");
642 // If we have to split PHI nodes or the entry block, do so now.
643 severSplitPHINodes(header);
645 // If we have any return instructions in the region, split those blocks so
646 // that the return is not in the region.
649 Function *oldFunction = header->getParent();
651 // This takes place of the original loop
652 BasicBlock *codeReplacer = new BasicBlock("codeRepl", oldFunction, header);
654 // The new function needs a root node because other nodes can branch to the
655 // head of the region, but the entry node of a function cannot have preds.
656 BasicBlock *newFuncRoot = new BasicBlock("newFuncRoot");
657 newFuncRoot->getInstList().push_back(new BranchInst(header));
659 // Find inputs to, outputs from the code region.
660 findInputsOutputs(inputs, outputs);
662 // Construct new function based on inputs/outputs & add allocas for all defs.
663 Function *newFunction = constructFunction(inputs, outputs, header,
665 codeReplacer, oldFunction,
666 oldFunction->getParent());
668 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
670 moveCodeToFunction(newFunction);
672 // Loop over all of the PHI nodes in the header block, and change any
673 // references to the old incoming edge to be the new incoming edge.
674 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
675 PHINode *PN = cast<PHINode>(I);
676 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
677 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
678 PN->setIncomingBlock(i, newFuncRoot);
681 // Look at all successors of the codeReplacer block. If any of these blocks
682 // had PHI nodes in them, we need to update the "from" block to be the code
683 // replacer, not the original block in the extracted region.
684 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
685 succ_end(codeReplacer));
686 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
687 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
688 PHINode *PN = cast<PHINode>(I);
689 std::set<BasicBlock*> ProcessedPreds;
690 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
691 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
692 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
693 PN->setIncomingBlock(i, codeReplacer);
695 // There were multiple entries in the PHI for this block, now there
696 // is only one, so remove the duplicated entries.
697 PN->removeIncomingValue(i, false);
702 //cerr << "NEW FUNCTION: " << *newFunction;
703 // verifyFunction(*newFunction);
705 // cerr << "OLD FUNCTION: " << *oldFunction;
706 // verifyFunction(*oldFunction);
708 DEBUG(if (verifyFunction(*newFunction)) abort());
712 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
713 // Deny code region if it contains allocas or vastarts.
714 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
716 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
718 if (isa<AllocaInst>(*I))
720 else if (const CallInst *CI = dyn_cast<CallInst>(I))
721 if (const Function *F = CI->getCalledFunction())
722 if (F->getIntrinsicID() == Intrinsic::vastart)
728 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
731 Function* llvm::ExtractCodeRegion(ETForest &EF, DominatorTree &DT,
732 const std::vector<BasicBlock*> &code,
733 bool AggregateArgs) {
734 return CodeExtractor(&EF, &DT, AggregateArgs).ExtractCodeRegion(code);
737 /// ExtractBasicBlock - slurp a natural loop into a brand new function
739 Function* llvm::ExtractLoop(ETForest &EF, DominatorTree &DF, Loop *L, bool AggregateArgs) {
740 return CodeExtractor(&EF, &DF, AggregateArgs).ExtractCodeRegion(L->getBlocks());
743 /// ExtractBasicBlock - slurp a basic block into a brand new function
745 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
746 std::vector<BasicBlock*> Blocks;
747 Blocks.push_back(BB);
748 return CodeExtractor(0, 0, AggregateArgs).ExtractCodeRegion(Blocks);