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/Debug.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/raw_ostream.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"));
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 BasicBlock *New = (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
188 // Old dominates New. New node domiantes all other nodes dominated
190 DomTreeNode *OldNode = DT->getNode(*I);
191 SmallVector<DomTreeNode*, 8> Children;
192 for (DomTreeNode::iterator DI = OldNode->begin(), DE = OldNode->end();
194 Children.push_back(*DI);
196 DomTreeNode *NewNode = DT->addNewBlock(New, *I);
198 for (SmallVector<DomTreeNode*, 8>::iterator I = Children.begin(),
199 E = Children.end(); I != E; ++I)
200 DT->changeImmediateDominator(*I, NewNode);
205 // findInputsOutputs - Find inputs to, outputs from the code region.
207 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
208 std::set<BasicBlock*> ExitBlocks;
209 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
210 ce = BlocksToExtract.end(); ci != ce; ++ci) {
211 BasicBlock *BB = *ci;
213 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
214 // If a used value is defined outside the region, it's an input. If an
215 // instruction is used outside the region, it's an output.
216 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
217 if (definedInCaller(*O))
218 inputs.push_back(*O);
220 // Consider uses of this instruction (outputs).
221 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
223 if (!definedInRegion(*UI)) {
224 outputs.push_back(I);
229 // Keep track of the exit blocks from the region.
230 TerminatorInst *TI = BB->getTerminator();
231 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
232 if (!BlocksToExtract.count(TI->getSuccessor(i)))
233 ExitBlocks.insert(TI->getSuccessor(i));
234 } // for: basic blocks
236 NumExitBlocks = ExitBlocks.size();
238 // Eliminate duplicates.
239 std::sort(inputs.begin(), inputs.end());
240 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
241 std::sort(outputs.begin(), outputs.end());
242 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
245 /// constructFunction - make a function based on inputs and outputs, as follows:
246 /// f(in0, ..., inN, out0, ..., outN)
248 Function *CodeExtractor::constructFunction(const Values &inputs,
249 const Values &outputs,
251 BasicBlock *newRootNode,
252 BasicBlock *newHeader,
253 Function *oldFunction,
255 DEBUG(errs() << "inputs: " << inputs.size() << "\n");
256 DEBUG(errs() << "outputs: " << outputs.size() << "\n");
258 // This function returns unsigned, outputs will go back by reference.
259 switch (NumExitBlocks) {
261 case 1: RetTy = Type::getVoidTy(header->getContext()); break;
262 case 2: RetTy = Type::getInt1Ty(header->getContext()); break;
263 default: RetTy = Type::getInt16Ty(header->getContext()); break;
266 std::vector<const Type*> paramTy;
268 // Add the types of the input values to the function's argument list
269 for (Values::const_iterator i = inputs.begin(),
270 e = inputs.end(); i != e; ++i) {
271 const Value *value = *i;
272 DEBUG(errs() << "value used in func: " << *value << "\n");
273 paramTy.push_back(value->getType());
276 // Add the types of the output values to the function's argument list.
277 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
279 DEBUG(errs() << "instr used in func: " << **I << "\n");
281 paramTy.push_back((*I)->getType());
283 paramTy.push_back(PointerType::getUnqual((*I)->getType()));
286 DEBUG(errs() << "Function type: " << *RetTy << " f(");
287 for (std::vector<const Type*>::iterator i = paramTy.begin(),
288 e = paramTy.end(); i != e; ++i)
289 DEBUG(errs() << **i << ", ");
290 DEBUG(errs() << ")\n");
292 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
293 PointerType *StructPtr =
294 PointerType::getUnqual(StructType::get(M->getContext(), paramTy));
296 paramTy.push_back(StructPtr);
298 const FunctionType *funcType =
299 FunctionType::get(RetTy, paramTy, false);
301 // Create the new function
302 Function *newFunction = Function::Create(funcType,
303 GlobalValue::InternalLinkage,
304 oldFunction->getName() + "_" +
305 header->getName(), M);
306 // If the old function is no-throw, so is the new one.
307 if (oldFunction->doesNotThrow())
308 newFunction->setDoesNotThrow(true);
310 newFunction->getBasicBlockList().push_back(newRootNode);
312 // Create an iterator to name all of the arguments we inserted.
313 Function::arg_iterator AI = newFunction->arg_begin();
315 // Rewrite all users of the inputs in the extracted region to use the
316 // arguments (or appropriate addressing into struct) instead.
317 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
321 Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext()));
322 Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i);
323 TerminatorInst *TI = newFunction->begin()->getTerminator();
324 GetElementPtrInst *GEP =
325 GetElementPtrInst::Create(AI, Idx, Idx+2,
326 "gep_" + inputs[i]->getName(), TI);
327 RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI);
331 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
332 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
334 if (Instruction* inst = dyn_cast<Instruction>(*use))
335 if (BlocksToExtract.count(inst->getParent()))
336 inst->replaceUsesOfWith(inputs[i], RewriteVal);
339 // Set names for input and output arguments.
340 if (!AggregateArgs) {
341 AI = newFunction->arg_begin();
342 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
343 AI->setName(inputs[i]->getName());
344 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
345 AI->setName(outputs[i]->getName()+".out");
348 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
349 // within the new function. This must be done before we lose track of which
350 // blocks were originally in the code region.
351 std::vector<User*> Users(header->use_begin(), header->use_end());
352 for (unsigned i = 0, e = Users.size(); i != e; ++i)
353 // The BasicBlock which contains the branch is not in the region
354 // modify the branch target to a new block
355 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
356 if (!BlocksToExtract.count(TI->getParent()) &&
357 TI->getParent()->getParent() == oldFunction)
358 TI->replaceUsesOfWith(header, newHeader);
363 /// FindPhiPredForUseInBlock - Given a value and a basic block, find a PHI
364 /// that uses the value within the basic block, and return the predecessor
365 /// block associated with that use, or return 0 if none is found.
366 static BasicBlock* FindPhiPredForUseInBlock(Value* Used, BasicBlock* BB) {
367 for (Value::use_iterator UI = Used->use_begin(),
368 UE = Used->use_end(); UI != UE; ++UI) {
369 PHINode *P = dyn_cast<PHINode>(*UI);
370 if (P && P->getParent() == BB)
371 return P->getIncomingBlock(UI);
377 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
378 /// the call instruction, splitting any PHI nodes in the header block as
381 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
382 Values &inputs, Values &outputs) {
383 // Emit a call to the new function, passing in: *pointer to struct (if
384 // aggregating parameters), or plan inputs and allocated memory for outputs
385 std::vector<Value*> params, StructValues, ReloadOutputs, Reloads;
387 LLVMContext &Context = newFunction->getContext();
389 // Add inputs as params, or to be filled into the struct
390 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
392 StructValues.push_back(*i);
394 params.push_back(*i);
396 // Create allocas for the outputs
397 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
399 StructValues.push_back(*i);
402 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
403 codeReplacer->getParent()->begin()->begin());
404 ReloadOutputs.push_back(alloca);
405 params.push_back(alloca);
409 AllocaInst *Struct = 0;
410 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
411 std::vector<const Type*> ArgTypes;
412 for (Values::iterator v = StructValues.begin(),
413 ve = StructValues.end(); v != ve; ++v)
414 ArgTypes.push_back((*v)->getType());
416 // Allocate a struct at the beginning of this function
417 Type *StructArgTy = StructType::get(newFunction->getContext(), ArgTypes);
419 new AllocaInst(StructArgTy, 0, "structArg",
420 codeReplacer->getParent()->begin()->begin());
421 params.push_back(Struct);
423 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
425 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
426 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i);
427 GetElementPtrInst *GEP =
428 GetElementPtrInst::Create(Struct, Idx, Idx + 2,
429 "gep_" + StructValues[i]->getName());
430 codeReplacer->getInstList().push_back(GEP);
431 StoreInst *SI = new StoreInst(StructValues[i], GEP);
432 codeReplacer->getInstList().push_back(SI);
436 // Emit the call to the function
437 CallInst *call = CallInst::Create(newFunction, params.begin(), params.end(),
438 NumExitBlocks > 1 ? "targetBlock" : "");
439 codeReplacer->getInstList().push_back(call);
441 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
442 unsigned FirstOut = inputs.size();
444 std::advance(OutputArgBegin, inputs.size());
446 // Reload the outputs passed in by reference
447 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
451 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
452 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i);
453 GetElementPtrInst *GEP
454 = GetElementPtrInst::Create(Struct, Idx, Idx + 2,
455 "gep_reload_" + outputs[i]->getName());
456 codeReplacer->getInstList().push_back(GEP);
459 Output = ReloadOutputs[i];
461 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
462 Reloads.push_back(load);
463 codeReplacer->getInstList().push_back(load);
464 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
465 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
466 Instruction *inst = cast<Instruction>(Users[u]);
467 if (!BlocksToExtract.count(inst->getParent()))
468 inst->replaceUsesOfWith(outputs[i], load);
472 // Now we can emit a switch statement using the call as a value.
473 SwitchInst *TheSwitch =
474 SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)),
475 codeReplacer, 0, codeReplacer);
477 // Since there may be multiple exits from the original region, make the new
478 // function return an unsigned, switch on that number. This loop iterates
479 // over all of the blocks in the extracted region, updating any terminator
480 // instructions in the to-be-extracted region that branch to blocks that are
481 // not in the region to be extracted.
482 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
484 unsigned switchVal = 0;
485 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
486 e = BlocksToExtract.end(); i != e; ++i) {
487 TerminatorInst *TI = (*i)->getTerminator();
488 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
489 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
490 BasicBlock *OldTarget = TI->getSuccessor(i);
491 // add a new basic block which returns the appropriate value
492 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
494 // If we don't already have an exit stub for this non-extracted
495 // destination, create one now!
496 NewTarget = BasicBlock::Create(Context,
497 OldTarget->getName() + ".exitStub",
499 unsigned SuccNum = switchVal++;
502 switch (NumExitBlocks) {
504 case 1: break; // No value needed.
505 case 2: // Conditional branch, return a bool
506 brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum);
509 brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum);
513 ReturnInst *NTRet = ReturnInst::Create(Context, brVal, NewTarget);
515 // Update the switch instruction.
516 TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context),
520 // Restore values just before we exit
521 Function::arg_iterator OAI = OutputArgBegin;
522 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
523 // For an invoke, the normal destination is the only one that is
524 // dominated by the result of the invocation
525 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
527 bool DominatesDef = true;
529 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
530 DefBlock = Invoke->getNormalDest();
532 // Make sure we are looking at the original successor block, not
533 // at a newly inserted exit block, which won't be in the dominator
535 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
536 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
537 if (DefBlock == I->second) {
542 // In the extract block case, if the block we are extracting ends
543 // with an invoke instruction, make sure that we don't emit a
544 // store of the invoke value for the unwind block.
545 if (!DT && DefBlock != OldTarget)
546 DominatesDef = false;
550 DominatesDef = DT->dominates(DefBlock, OldTarget);
552 // If the output value is used by a phi in the target block,
553 // then we need to test for dominance of the phi's predecessor
554 // instead. Unfortunately, this a little complicated since we
555 // have already rewritten uses of the value to uses of the reload.
556 BasicBlock* pred = FindPhiPredForUseInBlock(Reloads[out],
558 if (pred && DT && DT->dominates(DefBlock, pred))
565 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
566 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context),
568 GetElementPtrInst *GEP =
569 GetElementPtrInst::Create(OAI, Idx, Idx + 2,
570 "gep_" + outputs[out]->getName(),
572 new StoreInst(outputs[out], GEP, NTRet);
574 new StoreInst(outputs[out], OAI, NTRet);
577 // Advance output iterator even if we don't emit a store
578 if (!AggregateArgs) ++OAI;
582 // rewrite the original branch instruction with this new target
583 TI->setSuccessor(i, NewTarget);
587 // Now that we've done the deed, simplify the switch instruction.
588 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
589 switch (NumExitBlocks) {
591 // There are no successors (the block containing the switch itself), which
592 // means that previously this was the last part of the function, and hence
593 // this should be rewritten as a `ret'
595 // Check if the function should return a value
596 if (OldFnRetTy == Type::getVoidTy(Context)) {
597 ReturnInst::Create(Context, 0, TheSwitch); // Return void
598 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
599 // return what we have
600 ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch);
602 // Otherwise we must have code extracted an unwind or something, just
603 // return whatever we want.
604 ReturnInst::Create(Context,
605 Constant::getNullValue(OldFnRetTy), TheSwitch);
608 TheSwitch->eraseFromParent();
611 // Only a single destination, change the switch into an unconditional
613 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
614 TheSwitch->eraseFromParent();
617 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
619 TheSwitch->eraseFromParent();
622 // Otherwise, make the default destination of the switch instruction be one
623 // of the other successors.
624 TheSwitch->setOperand(0, call);
625 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
626 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
631 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
632 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
633 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
634 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
636 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
637 e = BlocksToExtract.end(); i != e; ++i) {
638 // Delete the basic block from the old function, and the list of blocks
639 oldBlocks.remove(*i);
641 // Insert this basic block into the new function
642 newBlocks.push_back(*i);
646 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
647 /// new function. Returns pointer to the new function.
651 /// find inputs and outputs for the region
653 /// for inputs: add to function as args, map input instr* to arg#
654 /// for outputs: add allocas for scalars,
655 /// add to func as args, map output instr* to arg#
657 /// rewrite func to use argument #s instead of instr*
659 /// for each scalar output in the function: at every exit, store intermediate
660 /// computed result back into memory.
662 Function *CodeExtractor::
663 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
664 if (!isEligible(code))
667 // 1) Find inputs, outputs
668 // 2) Construct new function
669 // * Add allocas for defs, pass as args by reference
670 // * Pass in uses as args
671 // 3) Move code region, add call instr to func
673 BlocksToExtract.insert(code.begin(), code.end());
675 Values inputs, outputs;
677 // Assumption: this is a single-entry code region, and the header is the first
678 // block in the region.
679 BasicBlock *header = code[0];
681 for (unsigned i = 1, e = code.size(); i != e; ++i)
682 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
684 assert(BlocksToExtract.count(*PI) &&
685 "No blocks in this region may have entries from outside the region"
686 " except for the first block!");
688 // If we have to split PHI nodes or the entry block, do so now.
689 severSplitPHINodes(header);
691 // If we have any return instructions in the region, split those blocks so
692 // that the return is not in the region.
695 Function *oldFunction = header->getParent();
697 // This takes place of the original loop
698 BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(),
699 "codeRepl", oldFunction,
702 // The new function needs a root node because other nodes can branch to the
703 // head of the region, but the entry node of a function cannot have preds.
704 BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(),
706 newFuncRoot->getInstList().push_back(BranchInst::Create(header));
708 // Find inputs to, outputs from the code region.
709 findInputsOutputs(inputs, outputs);
711 // Construct new function based on inputs/outputs & add allocas for all defs.
712 Function *newFunction = constructFunction(inputs, outputs, header,
714 codeReplacer, oldFunction,
715 oldFunction->getParent());
717 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
719 moveCodeToFunction(newFunction);
721 // Loop over all of the PHI nodes in the header block, and change any
722 // references to the old incoming edge to be the new incoming edge.
723 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
724 PHINode *PN = cast<PHINode>(I);
725 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
726 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
727 PN->setIncomingBlock(i, newFuncRoot);
730 // Look at all successors of the codeReplacer block. If any of these blocks
731 // had PHI nodes in them, we need to update the "from" block to be the code
732 // replacer, not the original block in the extracted region.
733 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
734 succ_end(codeReplacer));
735 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
736 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
737 PHINode *PN = cast<PHINode>(I);
738 std::set<BasicBlock*> ProcessedPreds;
739 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
740 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
741 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
742 PN->setIncomingBlock(i, codeReplacer);
744 // There were multiple entries in the PHI for this block, now there
745 // is only one, so remove the duplicated entries.
746 PN->removeIncomingValue(i, false);
752 //cerr << "NEW FUNCTION: " << *newFunction;
753 // verifyFunction(*newFunction);
755 // cerr << "OLD FUNCTION: " << *oldFunction;
756 // verifyFunction(*oldFunction);
758 DEBUG(if (verifyFunction(*newFunction))
759 llvm_report_error("verifyFunction failed!"));
763 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
764 // Deny code region if it contains allocas or vastarts.
765 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
767 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
769 if (isa<AllocaInst>(*I))
771 else if (const CallInst *CI = dyn_cast<CallInst>(I))
772 if (const Function *F = CI->getCalledFunction())
773 if (F->getIntrinsicID() == Intrinsic::vastart)
779 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
782 Function* llvm::ExtractCodeRegion(DominatorTree &DT,
783 const std::vector<BasicBlock*> &code,
784 bool AggregateArgs) {
785 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code);
788 /// ExtractBasicBlock - slurp a natural loop into a brand new function
790 Function* llvm::ExtractLoop(DominatorTree &DT, Loop *L, bool AggregateArgs) {
791 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(L->getBlocks());
794 /// ExtractBasicBlock - slurp a basic block into a brand new function
796 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
797 std::vector<BasicBlock*> Blocks;
798 Blocks.push_back(BB);
799 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);