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/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/LLVMContext.h"
22 #include "llvm/Module.h"
23 #include "llvm/Pass.h"
24 #include "llvm/Analysis/Dominators.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/Verifier.h"
27 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Compiler.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/ADT/StringExtras.h"
37 // Provide a command-line option to aggregate function arguments into a struct
38 // for functions produced by the code extractor. This is useful when converting
39 // extracted functions to pthread-based code, as only one argument (void*) can
40 // be passed in to pthread_create().
42 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
43 cl::desc("Aggregate arguments to code-extracted functions"));
46 class VISIBILITY_HIDDEN CodeExtractor {
47 typedef std::vector<Value*> Values;
48 std::set<BasicBlock*> BlocksToExtract;
51 unsigned NumExitBlocks;
54 CodeExtractor(DominatorTree* dt = 0, bool AggArgs = false)
55 : DT(dt), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
57 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
59 bool isEligible(const std::vector<BasicBlock*> &code);
62 /// definedInRegion - Return true if the specified value is defined in the
64 bool definedInRegion(Value *V) const {
65 if (Instruction *I = dyn_cast<Instruction>(V))
66 if (BlocksToExtract.count(I->getParent()))
71 /// definedInCaller - Return true if the specified value is defined in the
72 /// function being code extracted, but not in the region being extracted.
73 /// These values must be passed in as live-ins to the function.
74 bool definedInCaller(Value *V) const {
75 if (isa<Argument>(V)) return true;
76 if (Instruction *I = dyn_cast<Instruction>(V))
77 if (!BlocksToExtract.count(I->getParent()))
82 void severSplitPHINodes(BasicBlock *&Header);
83 void splitReturnBlocks();
84 void findInputsOutputs(Values &inputs, Values &outputs);
86 Function *constructFunction(const Values &inputs,
87 const Values &outputs,
89 BasicBlock *newRootNode, BasicBlock *newHeader,
90 Function *oldFunction, Module *M);
92 void moveCodeToFunction(Function *newFunction);
94 void emitCallAndSwitchStatement(Function *newFunction,
95 BasicBlock *newHeader,
102 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
103 /// region, we need to split the entry block of the region so that the PHI node
104 /// is easier to deal with.
105 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
106 bool HasPredsFromRegion = false;
107 unsigned NumPredsOutsideRegion = 0;
109 if (Header != &Header->getParent()->getEntryBlock()) {
110 PHINode *PN = dyn_cast<PHINode>(Header->begin());
111 if (!PN) return; // No PHI nodes.
113 // If the header node contains any PHI nodes, check to see if there is more
114 // than one entry from outside the region. If so, we need to sever the
115 // header block into two.
116 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
117 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
118 HasPredsFromRegion = true;
120 ++NumPredsOutsideRegion;
122 // If there is one (or fewer) predecessor from outside the region, we don't
123 // need to do anything special.
124 if (NumPredsOutsideRegion <= 1) return;
127 // Otherwise, we need to split the header block into two pieces: one
128 // containing PHI nodes merging values from outside of the region, and a
129 // second that contains all of the code for the block and merges back any
130 // incoming values from inside of the region.
131 BasicBlock::iterator AfterPHIs = Header->getFirstNonPHI();
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 DT->splitBlock(NewBB);
147 // Okay, now we need to adjust the PHI nodes and any branches from within the
148 // region to go to the new header block instead of the old header block.
149 if (HasPredsFromRegion) {
150 PHINode *PN = cast<PHINode>(OldPred->begin());
151 // Loop over all of the predecessors of OldPred that are in the region,
152 // changing them to branch to NewBB instead.
153 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
154 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
155 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
156 TI->replaceUsesOfWith(OldPred, NewBB);
159 // Okay, everthing within the region is now branching to the right block, we
160 // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
161 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
162 PHINode *PN = cast<PHINode>(AfterPHIs);
163 // Create a new PHI node in the new region, which has an incoming value
164 // from OldPred of PN.
165 PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName()+".ce",
167 NewPN->addIncoming(PN, OldPred);
169 // Loop over all of the incoming value in PN, moving them to NewPN if they
170 // are from the extracted region.
171 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
172 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
173 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
174 PN->removeIncomingValue(i);
182 void CodeExtractor::splitReturnBlocks() {
183 for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(),
184 E = BlocksToExtract.end(); I != E; ++I)
185 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator()))
186 (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
189 // findInputsOutputs - Find inputs to, outputs from the code region.
191 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
192 std::set<BasicBlock*> ExitBlocks;
193 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
194 ce = BlocksToExtract.end(); ci != ce; ++ci) {
195 BasicBlock *BB = *ci;
197 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
198 // If a used value is defined outside the region, it's an input. If an
199 // instruction is used outside the region, it's an output.
200 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
201 if (definedInCaller(*O))
202 inputs.push_back(*O);
204 // Consider uses of this instruction (outputs).
205 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
207 if (!definedInRegion(*UI)) {
208 outputs.push_back(I);
213 // Keep track of the exit blocks from the region.
214 TerminatorInst *TI = BB->getTerminator();
215 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
216 if (!BlocksToExtract.count(TI->getSuccessor(i)))
217 ExitBlocks.insert(TI->getSuccessor(i));
218 } // for: basic blocks
220 NumExitBlocks = ExitBlocks.size();
222 // Eliminate duplicates.
223 std::sort(inputs.begin(), inputs.end());
224 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
225 std::sort(outputs.begin(), outputs.end());
226 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
229 /// constructFunction - make a function based on inputs and outputs, as follows:
230 /// f(in0, ..., inN, out0, ..., outN)
232 Function *CodeExtractor::constructFunction(const Values &inputs,
233 const Values &outputs,
235 BasicBlock *newRootNode,
236 BasicBlock *newHeader,
237 Function *oldFunction,
239 DOUT << "inputs: " << inputs.size() << "\n";
240 DOUT << "outputs: " << outputs.size() << "\n";
242 LLVMContext &Context = header->getContext();
244 // This function returns unsigned, outputs will go back by reference.
245 switch (NumExitBlocks) {
247 case 1: RetTy = Type::VoidTy; break;
248 case 2: RetTy = Type::Int1Ty; break;
249 default: RetTy = Type::Int16Ty; break;
252 std::vector<const Type*> paramTy;
254 // Add the types of the input values to the function's argument list
255 for (Values::const_iterator i = inputs.begin(),
256 e = inputs.end(); i != e; ++i) {
257 const Value *value = *i;
258 DOUT << "value used in func: " << *value << "\n";
259 paramTy.push_back(value->getType());
262 // Add the types of the output values to the function's argument list.
263 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
265 DOUT << "instr used in func: " << **I << "\n";
267 paramTy.push_back((*I)->getType());
270 header->getContext().getPointerTypeUnqual((*I)->getType()));
273 DOUT << "Function type: " << *RetTy << " f(";
274 for (std::vector<const Type*>::iterator i = paramTy.begin(),
275 e = paramTy.end(); i != e; ++i)
279 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
280 PointerType *StructPtr =
281 Context.getPointerTypeUnqual(Context.getStructType(paramTy));
283 paramTy.push_back(StructPtr);
285 const FunctionType *funcType =
286 Context.getFunctionType(RetTy, paramTy, false);
288 // Create the new function
289 Function *newFunction = Function::Create(funcType,
290 GlobalValue::InternalLinkage,
291 oldFunction->getName() + "_" +
292 header->getName(), M);
293 // If the old function is no-throw, so is the new one.
294 if (oldFunction->doesNotThrow())
295 newFunction->setDoesNotThrow(true);
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) {
308 Idx[0] = Context.getNullValue(Type::Int32Ty);
309 Idx[1] = ConstantInt::get(Type::Int32Ty, i);
310 TerminatorInst *TI = newFunction->begin()->getTerminator();
311 GetElementPtrInst *GEP =
312 GetElementPtrInst::Create(AI, Idx, Idx+2,
313 "gep_" + inputs[i]->getName(), TI);
314 RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI);
318 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
319 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
321 if (Instruction* inst = dyn_cast<Instruction>(*use))
322 if (BlocksToExtract.count(inst->getParent()))
323 inst->replaceUsesOfWith(inputs[i], RewriteVal);
326 // Set names for input and output arguments.
327 if (!AggregateArgs) {
328 AI = newFunction->arg_begin();
329 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
330 AI->setName(inputs[i]->getName());
331 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
332 AI->setName(outputs[i]->getName()+".out");
335 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
336 // within the new function. This must be done before we lose track of which
337 // blocks were originally in the code region.
338 std::vector<User*> Users(header->use_begin(), header->use_end());
339 for (unsigned i = 0, e = Users.size(); i != e; ++i)
340 // The BasicBlock which contains the branch is not in the region
341 // modify the branch target to a new block
342 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
343 if (!BlocksToExtract.count(TI->getParent()) &&
344 TI->getParent()->getParent() == oldFunction)
345 TI->replaceUsesOfWith(header, newHeader);
350 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
351 /// the call instruction, splitting any PHI nodes in the header block as
354 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
355 Values &inputs, Values &outputs) {
356 LLVMContext &Context = codeReplacer->getContext();
358 // Emit a call to the new function, passing in: *pointer to struct (if
359 // aggregating parameters), or plan inputs and allocated memory for outputs
360 std::vector<Value*> params, StructValues, ReloadOutputs;
362 // Add inputs as params, or to be filled into the struct
363 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
365 StructValues.push_back(*i);
367 params.push_back(*i);
369 // Create allocas for the outputs
370 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
372 StructValues.push_back(*i);
375 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
376 codeReplacer->getParent()->begin()->begin());
377 ReloadOutputs.push_back(alloca);
378 params.push_back(alloca);
382 AllocaInst *Struct = 0;
383 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
384 std::vector<const Type*> ArgTypes;
385 for (Values::iterator v = StructValues.begin(),
386 ve = StructValues.end(); v != ve; ++v)
387 ArgTypes.push_back((*v)->getType());
389 // Allocate a struct at the beginning of this function
390 Type *StructArgTy = Context.getStructType(ArgTypes);
392 new AllocaInst(StructArgTy, 0, "structArg",
393 codeReplacer->getParent()->begin()->begin());
394 params.push_back(Struct);
396 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
398 Idx[0] = Context.getNullValue(Type::Int32Ty);
399 Idx[1] = ConstantInt::get(Type::Int32Ty, i);
400 GetElementPtrInst *GEP =
401 GetElementPtrInst::Create(Struct, Idx, Idx + 2,
402 "gep_" + StructValues[i]->getName());
403 codeReplacer->getInstList().push_back(GEP);
404 StoreInst *SI = new StoreInst(StructValues[i], GEP);
405 codeReplacer->getInstList().push_back(SI);
409 // Emit the call to the function
410 CallInst *call = CallInst::Create(newFunction, params.begin(), params.end(),
411 NumExitBlocks > 1 ? "targetBlock" : "");
412 codeReplacer->getInstList().push_back(call);
414 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
415 unsigned FirstOut = inputs.size();
417 std::advance(OutputArgBegin, inputs.size());
419 // Reload the outputs passed in by reference
420 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
424 Idx[0] = Context.getNullValue(Type::Int32Ty);
425 Idx[1] = ConstantInt::get(Type::Int32Ty, FirstOut + i);
426 GetElementPtrInst *GEP
427 = GetElementPtrInst::Create(Struct, Idx, Idx + 2,
428 "gep_reload_" + outputs[i]->getName());
429 codeReplacer->getInstList().push_back(GEP);
432 Output = ReloadOutputs[i];
434 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
435 codeReplacer->getInstList().push_back(load);
436 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
437 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
438 Instruction *inst = cast<Instruction>(Users[u]);
439 if (!BlocksToExtract.count(inst->getParent()))
440 inst->replaceUsesOfWith(outputs[i], load);
444 // Now we can emit a switch statement using the call as a value.
445 SwitchInst *TheSwitch =
446 SwitchInst::Create(Context.getNullValue(Type::Int16Ty),
447 codeReplacer, 0, codeReplacer);
449 // Since there may be multiple exits from the original region, make the new
450 // function return an unsigned, switch on that number. This loop iterates
451 // over all of the blocks in the extracted region, updating any terminator
452 // instructions in the to-be-extracted region that branch to blocks that are
453 // not in the region to be extracted.
454 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
456 unsigned switchVal = 0;
457 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
458 e = BlocksToExtract.end(); i != e; ++i) {
459 TerminatorInst *TI = (*i)->getTerminator();
460 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
461 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
462 BasicBlock *OldTarget = TI->getSuccessor(i);
463 // add a new basic block which returns the appropriate value
464 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
466 // If we don't already have an exit stub for this non-extracted
467 // destination, create one now!
468 NewTarget = BasicBlock::Create(OldTarget->getName() + ".exitStub",
470 unsigned SuccNum = switchVal++;
473 switch (NumExitBlocks) {
475 case 1: break; // No value needed.
476 case 2: // Conditional branch, return a bool
477 brVal = ConstantInt::get(Type::Int1Ty, !SuccNum);
480 brVal = ConstantInt::get(Type::Int16Ty, SuccNum);
484 ReturnInst *NTRet = ReturnInst::Create(brVal, NewTarget);
486 // Update the switch instruction.
487 TheSwitch->addCase(ConstantInt::get(Type::Int16Ty, SuccNum),
490 // Restore values just before we exit
491 Function::arg_iterator OAI = OutputArgBegin;
492 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
493 // For an invoke, the normal destination is the only one that is
494 // dominated by the result of the invocation
495 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
497 bool DominatesDef = true;
499 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
500 DefBlock = Invoke->getNormalDest();
502 // Make sure we are looking at the original successor block, not
503 // at a newly inserted exit block, which won't be in the dominator
505 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
506 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
507 if (DefBlock == I->second) {
512 // In the extract block case, if the block we are extracting ends
513 // with an invoke instruction, make sure that we don't emit a
514 // store of the invoke value for the unwind block.
515 if (!DT && DefBlock != OldTarget)
516 DominatesDef = false;
520 DominatesDef = DT->dominates(DefBlock, OldTarget);
525 Idx[0] = Context.getNullValue(Type::Int32Ty);
526 Idx[1] = ConstantInt::get(Type::Int32Ty,FirstOut+out);
527 GetElementPtrInst *GEP =
528 GetElementPtrInst::Create(OAI, Idx, Idx + 2,
529 "gep_" + outputs[out]->getName(),
531 new StoreInst(outputs[out], GEP, NTRet);
533 new StoreInst(outputs[out], OAI, NTRet);
536 // Advance output iterator even if we don't emit a store
537 if (!AggregateArgs) ++OAI;
541 // rewrite the original branch instruction with this new target
542 TI->setSuccessor(i, NewTarget);
546 // Now that we've done the deed, simplify the switch instruction.
547 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
548 switch (NumExitBlocks) {
550 // There are no successors (the block containing the switch itself), which
551 // means that previously this was the last part of the function, and hence
552 // this should be rewritten as a `ret'
554 // Check if the function should return a value
555 if (OldFnRetTy == Type::VoidTy) {
556 ReturnInst::Create(0, TheSwitch); // Return void
557 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
558 // return what we have
559 ReturnInst::Create(TheSwitch->getCondition(), TheSwitch);
561 // Otherwise we must have code extracted an unwind or something, just
562 // return whatever we want.
563 ReturnInst::Create(Context.getNullValue(OldFnRetTy), TheSwitch);
566 TheSwitch->eraseFromParent();
569 // Only a single destination, change the switch into an unconditional
571 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
572 TheSwitch->eraseFromParent();
575 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
577 TheSwitch->eraseFromParent();
580 // Otherwise, make the default destination of the switch instruction be one
581 // of the other successors.
582 TheSwitch->setOperand(0, call);
583 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
584 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
589 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
590 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
591 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
592 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
594 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
595 e = BlocksToExtract.end(); i != e; ++i) {
596 // Delete the basic block from the old function, and the list of blocks
597 oldBlocks.remove(*i);
599 // Insert this basic block into the new function
600 newBlocks.push_back(*i);
604 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
605 /// new function. Returns pointer to the new function.
609 /// find inputs and outputs for the region
611 /// for inputs: add to function as args, map input instr* to arg#
612 /// for outputs: add allocas for scalars,
613 /// add to func as args, map output instr* to arg#
615 /// rewrite func to use argument #s instead of instr*
617 /// for each scalar output in the function: at every exit, store intermediate
618 /// computed result back into memory.
620 Function *CodeExtractor::
621 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
622 if (!isEligible(code))
625 // 1) Find inputs, outputs
626 // 2) Construct new function
627 // * Add allocas for defs, pass as args by reference
628 // * Pass in uses as args
629 // 3) Move code region, add call instr to func
631 BlocksToExtract.insert(code.begin(), code.end());
633 Values inputs, outputs;
635 // Assumption: this is a single-entry code region, and the header is the first
636 // block in the region.
637 BasicBlock *header = code[0];
639 for (unsigned i = 1, e = code.size(); i != e; ++i)
640 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
642 assert(BlocksToExtract.count(*PI) &&
643 "No blocks in this region may have entries from outside the region"
644 " except for the first block!");
646 // If we have to split PHI nodes or the entry block, do so now.
647 severSplitPHINodes(header);
649 // If we have any return instructions in the region, split those blocks so
650 // that the return is not in the region.
653 Function *oldFunction = header->getParent();
655 // This takes place of the original loop
656 BasicBlock *codeReplacer = BasicBlock::Create("codeRepl", oldFunction,
659 // The new function needs a root node because other nodes can branch to the
660 // head of the region, but the entry node of a function cannot have preds.
661 BasicBlock *newFuncRoot = BasicBlock::Create("newFuncRoot");
662 newFuncRoot->getInstList().push_back(BranchInst::Create(header));
664 // Find inputs to, outputs from the code region.
665 findInputsOutputs(inputs, outputs);
667 // Construct new function based on inputs/outputs & add allocas for all defs.
668 Function *newFunction = constructFunction(inputs, outputs, header,
670 codeReplacer, oldFunction,
671 oldFunction->getParent());
673 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
675 moveCodeToFunction(newFunction);
677 // Loop over all of the PHI nodes in the header block, and change any
678 // references to the old incoming edge to be the new incoming edge.
679 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
680 PHINode *PN = cast<PHINode>(I);
681 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
682 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
683 PN->setIncomingBlock(i, newFuncRoot);
686 // Look at all successors of the codeReplacer block. If any of these blocks
687 // had PHI nodes in them, we need to update the "from" block to be the code
688 // replacer, not the original block in the extracted region.
689 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
690 succ_end(codeReplacer));
691 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
692 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
693 PHINode *PN = cast<PHINode>(I);
694 std::set<BasicBlock*> ProcessedPreds;
695 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
696 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
697 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
698 PN->setIncomingBlock(i, codeReplacer);
700 // There were multiple entries in the PHI for this block, now there
701 // is only one, so remove the duplicated entries.
702 PN->removeIncomingValue(i, false);
708 //cerr << "NEW FUNCTION: " << *newFunction;
709 // verifyFunction(*newFunction);
711 // cerr << "OLD FUNCTION: " << *oldFunction;
712 // verifyFunction(*oldFunction);
714 DEBUG(if (verifyFunction(*newFunction))
715 llvm_report_error("verifyFunction failed!"));
719 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
720 // Deny code region if it contains allocas or vastarts.
721 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
723 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
725 if (isa<AllocaInst>(*I))
727 else if (const CallInst *CI = dyn_cast<CallInst>(I))
728 if (const Function *F = CI->getCalledFunction())
729 if (F->getIntrinsicID() == Intrinsic::vastart)
735 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
738 Function* llvm::ExtractCodeRegion(DominatorTree &DT,
739 const std::vector<BasicBlock*> &code,
740 bool AggregateArgs) {
741 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code);
744 /// ExtractBasicBlock - slurp a natural loop into a brand new function
746 Function* llvm::ExtractLoop(DominatorTree &DT, Loop *L, bool AggregateArgs) {
747 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(L->getBlocks());
750 /// ExtractBasicBlock - slurp a basic block into a brand new function
752 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
753 std::vector<BasicBlock*> Blocks;
754 Blocks.push_back(BB);
755 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);