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/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(DominatorTree* dt = 0, bool AggArgs = false)
53 : DT(dt), 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()->getEntryBlock()) {
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 DT->splitBlock(NewBB);
146 // Okay, now we need to adjust the PHI nodes and any branches from within the
147 // region to go to the new header block instead of the old header block.
148 if (HasPredsFromRegion) {
149 PHINode *PN = cast<PHINode>(OldPred->begin());
150 // Loop over all of the predecessors of OldPred that are in the region,
151 // changing them to branch to NewBB instead.
152 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
153 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
154 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
155 TI->replaceUsesOfWith(OldPred, NewBB);
158 // Okay, everthing within the region is now branching to the right block, we
159 // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
160 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
161 PHINode *PN = cast<PHINode>(AfterPHIs);
162 // Create a new PHI node in the new region, which has an incoming value
163 // from OldPred of PN.
164 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".ce",
166 NewPN->addIncoming(PN, OldPred);
168 // Loop over all of the incoming value in PN, moving them to NewPN if they
169 // are from the extracted region.
170 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
171 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
172 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
173 PN->removeIncomingValue(i);
181 void CodeExtractor::splitReturnBlocks() {
182 for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(),
183 E = BlocksToExtract.end(); I != E; ++I)
184 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator()))
185 (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
188 // findInputsOutputs - Find inputs to, outputs from the code region.
190 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
191 std::set<BasicBlock*> ExitBlocks;
192 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
193 ce = BlocksToExtract.end(); ci != ce; ++ci) {
194 BasicBlock *BB = *ci;
196 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
197 // If a used value is defined outside the region, it's an input. If an
198 // instruction is used outside the region, it's an output.
199 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
200 if (definedInCaller(*O))
201 inputs.push_back(*O);
203 // Consider uses of this instruction (outputs).
204 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
206 if (!definedInRegion(*UI)) {
207 outputs.push_back(I);
212 // Keep track of the exit blocks from the region.
213 TerminatorInst *TI = BB->getTerminator();
214 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
215 if (!BlocksToExtract.count(TI->getSuccessor(i)))
216 ExitBlocks.insert(TI->getSuccessor(i));
217 } // for: basic blocks
219 NumExitBlocks = ExitBlocks.size();
221 // Eliminate duplicates.
222 std::sort(inputs.begin(), inputs.end());
223 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
224 std::sort(outputs.begin(), outputs.end());
225 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
228 /// constructFunction - make a function based on inputs and outputs, as follows:
229 /// f(in0, ..., inN, out0, ..., outN)
231 Function *CodeExtractor::constructFunction(const Values &inputs,
232 const Values &outputs,
234 BasicBlock *newRootNode,
235 BasicBlock *newHeader,
236 Function *oldFunction,
238 DOUT << "inputs: " << inputs.size() << "\n";
239 DOUT << "outputs: " << outputs.size() << "\n";
241 // This function returns unsigned, outputs will go back by reference.
242 switch (NumExitBlocks) {
244 case 1: RetTy = Type::VoidTy; break;
245 case 2: RetTy = Type::Int1Ty; break;
246 default: RetTy = Type::Int16Ty; break;
249 std::vector<const Type*> paramTy;
251 // Add the types of the input values to the function's argument list
252 for (Values::const_iterator i = inputs.begin(),
253 e = inputs.end(); i != e; ++i) {
254 const Value *value = *i;
255 DOUT << "value used in func: " << *value << "\n";
256 paramTy.push_back(value->getType());
259 // Add the types of the output values to the function's argument list.
260 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
262 DOUT << "instr used in func: " << **I << "\n";
264 paramTy.push_back((*I)->getType());
266 paramTy.push_back(PointerType::getUnqual((*I)->getType()));
269 DOUT << "Function type: " << *RetTy << " f(";
270 for (std::vector<const Type*>::iterator i = paramTy.begin(),
271 e = paramTy.end(); i != e; ++i)
275 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
276 PointerType *StructPtr = PointerType::getUnqual(StructType::get(paramTy));
278 paramTy.push_back(StructPtr);
280 const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false);
282 // Create the new function
283 Function *newFunction = new Function(funcType,
284 GlobalValue::InternalLinkage,
285 oldFunction->getName() + "_" +
286 header->getName(), M);
287 newFunction->getBasicBlockList().push_back(newRootNode);
289 // Create an iterator to name all of the arguments we inserted.
290 Function::arg_iterator AI = newFunction->arg_begin();
292 // Rewrite all users of the inputs in the extracted region to use the
293 // arguments (or appropriate addressing into struct) instead.
294 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
298 Idx[0] = Constant::getNullValue(Type::Int32Ty);
299 Idx[1] = ConstantInt::get(Type::Int32Ty, i);
300 std::string GEPname = "gep_" + inputs[i]->getName();
301 TerminatorInst *TI = newFunction->begin()->getTerminator();
302 GetElementPtrInst *GEP = new GetElementPtrInst(AI, Idx, Idx+2,
304 RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
308 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
309 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
311 if (Instruction* inst = dyn_cast<Instruction>(*use))
312 if (BlocksToExtract.count(inst->getParent()))
313 inst->replaceUsesOfWith(inputs[i], RewriteVal);
316 // Set names for input and output arguments.
317 if (!AggregateArgs) {
318 AI = newFunction->arg_begin();
319 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
320 AI->setName(inputs[i]->getName());
321 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
322 AI->setName(outputs[i]->getName()+".out");
325 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
326 // within the new function. This must be done before we lose track of which
327 // blocks were originally in the code region.
328 std::vector<User*> Users(header->use_begin(), header->use_end());
329 for (unsigned i = 0, e = Users.size(); i != e; ++i)
330 // The BasicBlock which contains the branch is not in the region
331 // modify the branch target to a new block
332 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
333 if (!BlocksToExtract.count(TI->getParent()) &&
334 TI->getParent()->getParent() == oldFunction)
335 TI->replaceUsesOfWith(header, newHeader);
340 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
341 /// the call instruction, splitting any PHI nodes in the header block as
344 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
345 Values &inputs, Values &outputs) {
346 // Emit a call to the new function, passing in: *pointer to struct (if
347 // aggregating parameters), or plan inputs and allocated memory for outputs
348 std::vector<Value*> params, StructValues, ReloadOutputs;
350 // Add inputs as params, or to be filled into the struct
351 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
353 StructValues.push_back(*i);
355 params.push_back(*i);
357 // Create allocas for the outputs
358 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
360 StructValues.push_back(*i);
363 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
364 codeReplacer->getParent()->begin()->begin());
365 ReloadOutputs.push_back(alloca);
366 params.push_back(alloca);
370 AllocaInst *Struct = 0;
371 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
372 std::vector<const Type*> ArgTypes;
373 for (Values::iterator v = StructValues.begin(),
374 ve = StructValues.end(); v != ve; ++v)
375 ArgTypes.push_back((*v)->getType());
377 // Allocate a struct at the beginning of this function
378 Type *StructArgTy = StructType::get(ArgTypes);
380 new AllocaInst(StructArgTy, 0, "structArg",
381 codeReplacer->getParent()->begin()->begin());
382 params.push_back(Struct);
384 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
386 Idx[0] = Constant::getNullValue(Type::Int32Ty);
387 Idx[1] = ConstantInt::get(Type::Int32Ty, i);
388 GetElementPtrInst *GEP =
389 new GetElementPtrInst(Struct, Idx, Idx + 2,
390 "gep_" + StructValues[i]->getName());
391 codeReplacer->getInstList().push_back(GEP);
392 StoreInst *SI = new StoreInst(StructValues[i], GEP);
393 codeReplacer->getInstList().push_back(SI);
397 // Emit the call to the function
398 CallInst *call = new CallInst(newFunction, params.begin(), params.end(),
399 NumExitBlocks > 1 ? "targetBlock" : "");
400 codeReplacer->getInstList().push_back(call);
402 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
403 unsigned FirstOut = inputs.size();
405 std::advance(OutputArgBegin, inputs.size());
407 // Reload the outputs passed in by reference
408 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
412 Idx[0] = Constant::getNullValue(Type::Int32Ty);
413 Idx[1] = ConstantInt::get(Type::Int32Ty, FirstOut + i);
414 GetElementPtrInst *GEP
415 = new GetElementPtrInst(Struct, Idx, Idx + 2,
416 "gep_reload_" + outputs[i]->getName());
417 codeReplacer->getInstList().push_back(GEP);
420 Output = ReloadOutputs[i];
422 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
423 codeReplacer->getInstList().push_back(load);
424 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
425 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
426 Instruction *inst = cast<Instruction>(Users[u]);
427 if (!BlocksToExtract.count(inst->getParent()))
428 inst->replaceUsesOfWith(outputs[i], load);
432 // Now we can emit a switch statement using the call as a value.
433 SwitchInst *TheSwitch =
434 new SwitchInst(ConstantInt::getNullValue(Type::Int16Ty),
435 codeReplacer, 0, codeReplacer);
437 // Since there may be multiple exits from the original region, make the new
438 // function return an unsigned, switch on that number. This loop iterates
439 // over all of the blocks in the extracted region, updating any terminator
440 // instructions in the to-be-extracted region that branch to blocks that are
441 // not in the region to be extracted.
442 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
444 unsigned switchVal = 0;
445 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
446 e = BlocksToExtract.end(); i != e; ++i) {
447 TerminatorInst *TI = (*i)->getTerminator();
448 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
449 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
450 BasicBlock *OldTarget = TI->getSuccessor(i);
451 // add a new basic block which returns the appropriate value
452 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
454 // If we don't already have an exit stub for this non-extracted
455 // destination, create one now!
456 NewTarget = new BasicBlock(OldTarget->getName() + ".exitStub",
458 unsigned SuccNum = switchVal++;
461 switch (NumExitBlocks) {
463 case 1: break; // No value needed.
464 case 2: // Conditional branch, return a bool
465 brVal = ConstantInt::get(Type::Int1Ty, !SuccNum);
468 brVal = ConstantInt::get(Type::Int16Ty, SuccNum);
472 ReturnInst *NTRet = new ReturnInst(brVal, NewTarget);
474 // Update the switch instruction.
475 TheSwitch->addCase(ConstantInt::get(Type::Int16Ty, SuccNum),
478 // Restore values just before we exit
479 Function::arg_iterator OAI = OutputArgBegin;
480 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
481 // For an invoke, the normal destination is the only one that is
482 // dominated by the result of the invocation
483 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
485 bool DominatesDef = true;
487 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
488 DefBlock = Invoke->getNormalDest();
490 // Make sure we are looking at the original successor block, not
491 // at a newly inserted exit block, which won't be in the dominator
493 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
494 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
495 if (DefBlock == I->second) {
500 // In the extract block case, if the block we are extracting ends
501 // with an invoke instruction, make sure that we don't emit a
502 // store of the invoke value for the unwind block.
503 if (!DT && DefBlock != OldTarget)
504 DominatesDef = false;
508 DominatesDef = DT->dominates(DefBlock, OldTarget);
513 Idx[0] = Constant::getNullValue(Type::Int32Ty);
514 Idx[1] = ConstantInt::get(Type::Int32Ty,FirstOut+out);
515 GetElementPtrInst *GEP =
516 new GetElementPtrInst(OAI, Idx, Idx + 2,
517 "gep_" + outputs[out]->getName(),
519 new StoreInst(outputs[out], GEP, NTRet);
521 new StoreInst(outputs[out], OAI, NTRet);
524 // Advance output iterator even if we don't emit a store
525 if (!AggregateArgs) ++OAI;
529 // rewrite the original branch instruction with this new target
530 TI->setSuccessor(i, NewTarget);
534 // Now that we've done the deed, simplify the switch instruction.
535 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
536 switch (NumExitBlocks) {
538 // There are no successors (the block containing the switch itself), which
539 // means that previously this was the last part of the function, and hence
540 // this should be rewritten as a `ret'
542 // Check if the function should return a value
543 if (OldFnRetTy == Type::VoidTy) {
544 new ReturnInst(0, TheSwitch); // Return void
545 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
546 // return what we have
547 new ReturnInst(TheSwitch->getCondition(), TheSwitch);
549 // Otherwise we must have code extracted an unwind or something, just
550 // return whatever we want.
551 new ReturnInst(Constant::getNullValue(OldFnRetTy), TheSwitch);
554 TheSwitch->getParent()->getInstList().erase(TheSwitch);
557 // Only a single destination, change the switch into an unconditional
559 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch);
560 TheSwitch->getParent()->getInstList().erase(TheSwitch);
563 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
565 TheSwitch->getParent()->getInstList().erase(TheSwitch);
568 // Otherwise, make the default destination of the switch instruction be one
569 // of the other successors.
570 TheSwitch->setOperand(0, call);
571 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
572 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
577 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
578 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
579 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
580 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
582 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
583 e = BlocksToExtract.end(); i != e; ++i) {
584 // Delete the basic block from the old function, and the list of blocks
585 oldBlocks.remove(*i);
587 // Insert this basic block into the new function
588 newBlocks.push_back(*i);
592 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
593 /// new function. Returns pointer to the new function.
597 /// find inputs and outputs for the region
599 /// for inputs: add to function as args, map input instr* to arg#
600 /// for outputs: add allocas for scalars,
601 /// add to func as args, map output instr* to arg#
603 /// rewrite func to use argument #s instead of instr*
605 /// for each scalar output in the function: at every exit, store intermediate
606 /// computed result back into memory.
608 Function *CodeExtractor::
609 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
610 if (!isEligible(code))
613 // 1) Find inputs, outputs
614 // 2) Construct new function
615 // * Add allocas for defs, pass as args by reference
616 // * Pass in uses as args
617 // 3) Move code region, add call instr to func
619 BlocksToExtract.insert(code.begin(), code.end());
621 Values inputs, outputs;
623 // Assumption: this is a single-entry code region, and the header is the first
624 // block in the region.
625 BasicBlock *header = code[0];
627 for (unsigned i = 1, e = code.size(); i != e; ++i)
628 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
630 assert(BlocksToExtract.count(*PI) &&
631 "No blocks in this region may have entries from outside the region"
632 " except for the first block!");
634 // If we have to split PHI nodes or the entry block, do so now.
635 severSplitPHINodes(header);
637 // If we have any return instructions in the region, split those blocks so
638 // that the return is not in the region.
641 Function *oldFunction = header->getParent();
643 // This takes place of the original loop
644 BasicBlock *codeReplacer = new BasicBlock("codeRepl", oldFunction, header);
646 // The new function needs a root node because other nodes can branch to the
647 // head of the region, but the entry node of a function cannot have preds.
648 BasicBlock *newFuncRoot = new BasicBlock("newFuncRoot");
649 newFuncRoot->getInstList().push_back(new BranchInst(header));
651 // Find inputs to, outputs from the code region.
652 findInputsOutputs(inputs, outputs);
654 // Construct new function based on inputs/outputs & add allocas for all defs.
655 Function *newFunction = constructFunction(inputs, outputs, header,
657 codeReplacer, oldFunction,
658 oldFunction->getParent());
660 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
662 moveCodeToFunction(newFunction);
664 // Loop over all of the PHI nodes in the header block, and change any
665 // references to the old incoming edge to be the new incoming edge.
666 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
667 PHINode *PN = cast<PHINode>(I);
668 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
669 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
670 PN->setIncomingBlock(i, newFuncRoot);
673 // Look at all successors of the codeReplacer block. If any of these blocks
674 // had PHI nodes in them, we need to update the "from" block to be the code
675 // replacer, not the original block in the extracted region.
676 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
677 succ_end(codeReplacer));
678 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
679 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
680 PHINode *PN = cast<PHINode>(I);
681 std::set<BasicBlock*> ProcessedPreds;
682 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
683 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
684 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
685 PN->setIncomingBlock(i, codeReplacer);
687 // There were multiple entries in the PHI for this block, now there
688 // is only one, so remove the duplicated entries.
689 PN->removeIncomingValue(i, false);
695 //cerr << "NEW FUNCTION: " << *newFunction;
696 // verifyFunction(*newFunction);
698 // cerr << "OLD FUNCTION: " << *oldFunction;
699 // verifyFunction(*oldFunction);
701 DEBUG(if (verifyFunction(*newFunction)) abort());
705 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
706 // Deny code region if it contains allocas or vastarts.
707 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
709 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
711 if (isa<AllocaInst>(*I))
713 else if (const CallInst *CI = dyn_cast<CallInst>(I))
714 if (const Function *F = CI->getCalledFunction())
715 if (F->getIntrinsicID() == Intrinsic::vastart)
721 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
724 Function* llvm::ExtractCodeRegion(DominatorTree &DT,
725 const std::vector<BasicBlock*> &code,
726 bool AggregateArgs) {
727 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code);
730 /// ExtractBasicBlock - slurp a natural loop into a brand new function
732 Function* llvm::ExtractLoop(DominatorTree &DT, Loop *L, bool AggregateArgs) {
733 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(L->getBlocks());
736 /// ExtractBasicBlock - slurp a basic block into a brand new function
738 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
739 std::vector<BasicBlock*> Blocks;
740 Blocks.push_back(BB);
741 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);