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/Support/raw_ostream.h"
33 #include "llvm/ADT/StringExtras.h"
38 // Provide a command-line option to aggregate function arguments into a struct
39 // for functions produced by the code extractor. This is useful when converting
40 // extracted functions to pthread-based code, as only one argument (void*) can
41 // be passed in to pthread_create().
43 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
44 cl::desc("Aggregate arguments to code-extracted functions"));
47 class VISIBILITY_HIDDEN CodeExtractor {
48 typedef std::vector<Value*> Values;
49 std::set<BasicBlock*> BlocksToExtract;
52 unsigned NumExitBlocks;
55 CodeExtractor(DominatorTree* dt = 0, bool AggArgs = false)
56 : DT(dt), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
58 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
60 bool isEligible(const std::vector<BasicBlock*> &code);
63 /// definedInRegion - Return true if the specified value is defined in the
65 bool definedInRegion(Value *V) const {
66 if (Instruction *I = dyn_cast<Instruction>(V))
67 if (BlocksToExtract.count(I->getParent()))
72 /// definedInCaller - Return true if the specified value is defined in the
73 /// function being code extracted, but not in the region being extracted.
74 /// These values must be passed in as live-ins to the function.
75 bool definedInCaller(Value *V) const {
76 if (isa<Argument>(V)) return true;
77 if (Instruction *I = dyn_cast<Instruction>(V))
78 if (!BlocksToExtract.count(I->getParent()))
83 void severSplitPHINodes(BasicBlock *&Header);
84 void splitReturnBlocks();
85 void findInputsOutputs(Values &inputs, Values &outputs);
87 Function *constructFunction(const Values &inputs,
88 const Values &outputs,
90 BasicBlock *newRootNode, BasicBlock *newHeader,
91 Function *oldFunction, Module *M);
93 void moveCodeToFunction(Function *newFunction);
95 void emitCallAndSwitchStatement(Function *newFunction,
96 BasicBlock *newHeader,
103 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
104 /// region, we need to split the entry block of the region so that the PHI node
105 /// is easier to deal with.
106 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
107 bool HasPredsFromRegion = false;
108 unsigned NumPredsOutsideRegion = 0;
110 if (Header != &Header->getParent()->getEntryBlock()) {
111 PHINode *PN = dyn_cast<PHINode>(Header->begin());
112 if (!PN) return; // No PHI nodes.
114 // If the header node contains any PHI nodes, check to see if there is more
115 // than one entry from outside the region. If so, we need to sever the
116 // header block into two.
117 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
118 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
119 HasPredsFromRegion = true;
121 ++NumPredsOutsideRegion;
123 // If there is one (or fewer) predecessor from outside the region, we don't
124 // need to do anything special.
125 if (NumPredsOutsideRegion <= 1) return;
128 // Otherwise, we need to split the header block into two pieces: one
129 // containing PHI nodes merging values from outside of the region, and a
130 // second that contains all of the code for the block and merges back any
131 // incoming values from inside of the region.
132 BasicBlock::iterator AfterPHIs = Header->getFirstNonPHI();
133 BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
134 Header->getName()+".ce");
136 // We only want to code extract the second block now, and it becomes the new
137 // header of the region.
138 BasicBlock *OldPred = Header;
139 BlocksToExtract.erase(OldPred);
140 BlocksToExtract.insert(NewBB);
143 // Okay, update dominator sets. The blocks that dominate the new one are the
144 // blocks that dominate TIBB plus the new block itself.
146 DT->splitBlock(NewBB);
148 // Okay, now we need to adjust the PHI nodes and any branches from within the
149 // region to go to the new header block instead of the old header block.
150 if (HasPredsFromRegion) {
151 PHINode *PN = cast<PHINode>(OldPred->begin());
152 // Loop over all of the predecessors of OldPred that are in the region,
153 // changing them to branch to NewBB instead.
154 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
155 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
156 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
157 TI->replaceUsesOfWith(OldPred, NewBB);
160 // Okay, everthing within the region is now branching to the right block, we
161 // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
162 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
163 PHINode *PN = cast<PHINode>(AfterPHIs);
164 // Create a new PHI node in the new region, which has an incoming value
165 // from OldPred of PN.
166 PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName()+".ce",
168 NewPN->addIncoming(PN, OldPred);
170 // Loop over all of the incoming value in PN, moving them to NewPN if they
171 // are from the extracted region.
172 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
173 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
174 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
175 PN->removeIncomingValue(i);
183 void CodeExtractor::splitReturnBlocks() {
184 for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(),
185 E = BlocksToExtract.end(); I != E; ++I)
186 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator())) {
187 BasicBlock *New = (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
189 // Old dominates New. New node domiantes all other nodes dominated
191 DomTreeNode *OldNode = DT->getNode(*I);
192 SmallVector<DomTreeNode*, 8> Children;
193 for (DomTreeNode::iterator DI = OldNode->begin(), DE = OldNode->end();
195 Children.push_back(*DI);
197 DomTreeNode *NewNode = DT->addNewBlock(New, *I);
199 for (SmallVector<DomTreeNode*, 8>::iterator I = Children.begin(),
200 E = Children.end(); I != E; ++I)
201 DT->changeImmediateDominator(*I, NewNode);
206 // findInputsOutputs - Find inputs to, outputs from the code region.
208 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
209 std::set<BasicBlock*> ExitBlocks;
210 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
211 ce = BlocksToExtract.end(); ci != ce; ++ci) {
212 BasicBlock *BB = *ci;
214 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
215 // If a used value is defined outside the region, it's an input. If an
216 // instruction is used outside the region, it's an output.
217 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
218 if (definedInCaller(*O))
219 inputs.push_back(*O);
221 // Consider uses of this instruction (outputs).
222 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
224 if (!definedInRegion(*UI)) {
225 outputs.push_back(I);
230 // Keep track of the exit blocks from the region.
231 TerminatorInst *TI = BB->getTerminator();
232 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
233 if (!BlocksToExtract.count(TI->getSuccessor(i)))
234 ExitBlocks.insert(TI->getSuccessor(i));
235 } // for: basic blocks
237 NumExitBlocks = ExitBlocks.size();
239 // Eliminate duplicates.
240 std::sort(inputs.begin(), inputs.end());
241 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
242 std::sort(outputs.begin(), outputs.end());
243 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
246 /// constructFunction - make a function based on inputs and outputs, as follows:
247 /// f(in0, ..., inN, out0, ..., outN)
249 Function *CodeExtractor::constructFunction(const Values &inputs,
250 const Values &outputs,
252 BasicBlock *newRootNode,
253 BasicBlock *newHeader,
254 Function *oldFunction,
256 DEBUG(errs() << "inputs: " << inputs.size() << "\n");
257 DEBUG(errs() << "outputs: " << outputs.size() << "\n");
259 // This function returns unsigned, outputs will go back by reference.
260 switch (NumExitBlocks) {
262 case 1: RetTy = Type::getVoidTy(header->getContext()); break;
263 case 2: RetTy = Type::getInt1Ty(header->getContext()); break;
264 default: RetTy = Type::getInt16Ty(header->getContext()); break;
267 std::vector<const Type*> paramTy;
269 // Add the types of the input values to the function's argument list
270 for (Values::const_iterator i = inputs.begin(),
271 e = inputs.end(); i != e; ++i) {
272 const Value *value = *i;
273 DEBUG(errs() << "value used in func: " << *value << "\n");
274 paramTy.push_back(value->getType());
277 // Add the types of the output values to the function's argument list.
278 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
280 DEBUG(errs() << "instr used in func: " << **I << "\n");
282 paramTy.push_back((*I)->getType());
284 paramTy.push_back(PointerType::getUnqual((*I)->getType()));
287 DEBUG(errs() << "Function type: " << *RetTy << " f(");
288 for (std::vector<const Type*>::iterator i = paramTy.begin(),
289 e = paramTy.end(); i != e; ++i)
290 DEBUG(errs() << **i << ", ");
291 DEBUG(errs() << ")\n");
293 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
294 PointerType *StructPtr =
295 PointerType::getUnqual(StructType::get(M->getContext(), paramTy));
297 paramTy.push_back(StructPtr);
299 const FunctionType *funcType =
300 FunctionType::get(RetTy, paramTy, false);
302 // Create the new function
303 Function *newFunction = Function::Create(funcType,
304 GlobalValue::InternalLinkage,
305 oldFunction->getName() + "_" +
306 header->getName(), M);
307 // If the old function is no-throw, so is the new one.
308 if (oldFunction->doesNotThrow())
309 newFunction->setDoesNotThrow(true);
311 newFunction->getBasicBlockList().push_back(newRootNode);
313 // Create an iterator to name all of the arguments we inserted.
314 Function::arg_iterator AI = newFunction->arg_begin();
316 // Rewrite all users of the inputs in the extracted region to use the
317 // arguments (or appropriate addressing into struct) instead.
318 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
322 Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext()));
323 Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i);
324 TerminatorInst *TI = newFunction->begin()->getTerminator();
325 GetElementPtrInst *GEP =
326 GetElementPtrInst::Create(AI, Idx, Idx+2,
327 "gep_" + inputs[i]->getName(), TI);
328 RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI);
332 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
333 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
335 if (Instruction* inst = dyn_cast<Instruction>(*use))
336 if (BlocksToExtract.count(inst->getParent()))
337 inst->replaceUsesOfWith(inputs[i], RewriteVal);
340 // Set names for input and output arguments.
341 if (!AggregateArgs) {
342 AI = newFunction->arg_begin();
343 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
344 AI->setName(inputs[i]->getName());
345 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
346 AI->setName(outputs[i]->getName()+".out");
349 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
350 // within the new function. This must be done before we lose track of which
351 // blocks were originally in the code region.
352 std::vector<User*> Users(header->use_begin(), header->use_end());
353 for (unsigned i = 0, e = Users.size(); i != e; ++i)
354 // The BasicBlock which contains the branch is not in the region
355 // modify the branch target to a new block
356 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
357 if (!BlocksToExtract.count(TI->getParent()) &&
358 TI->getParent()->getParent() == oldFunction)
359 TI->replaceUsesOfWith(header, newHeader);
364 static BasicBlock* FindPhiPredForUseInBlock(Value* Used, BasicBlock* BB) {
365 for (Value::use_iterator UI = Used->use_begin(),
366 UE = Used->use_end(); UI != UE; ++UI) {
367 PHINode *P = dyn_cast<PHINode>(*UI);
368 if (P && P->getParent() == BB)
369 return P->getIncomingBlock(UI);
375 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
376 /// the call instruction, splitting any PHI nodes in the header block as
379 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
380 Values &inputs, Values &outputs) {
381 // Emit a call to the new function, passing in: *pointer to struct (if
382 // aggregating parameters), or plan inputs and allocated memory for outputs
383 std::vector<Value*> params, StructValues, ReloadOutputs, Reloads;
385 LLVMContext &Context = newFunction->getContext();
387 // Add inputs as params, or to be filled into the struct
388 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
390 StructValues.push_back(*i);
392 params.push_back(*i);
394 // Create allocas for the outputs
395 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
397 StructValues.push_back(*i);
400 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
401 codeReplacer->getParent()->begin()->begin());
402 ReloadOutputs.push_back(alloca);
403 params.push_back(alloca);
407 AllocaInst *Struct = 0;
408 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
409 std::vector<const Type*> ArgTypes;
410 for (Values::iterator v = StructValues.begin(),
411 ve = StructValues.end(); v != ve; ++v)
412 ArgTypes.push_back((*v)->getType());
414 // Allocate a struct at the beginning of this function
415 Type *StructArgTy = StructType::get(newFunction->getContext(), ArgTypes);
417 new AllocaInst(StructArgTy, 0, "structArg",
418 codeReplacer->getParent()->begin()->begin());
419 params.push_back(Struct);
421 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
423 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
424 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i);
425 GetElementPtrInst *GEP =
426 GetElementPtrInst::Create(Struct, Idx, Idx + 2,
427 "gep_" + StructValues[i]->getName());
428 codeReplacer->getInstList().push_back(GEP);
429 StoreInst *SI = new StoreInst(StructValues[i], GEP);
430 codeReplacer->getInstList().push_back(SI);
434 // Emit the call to the function
435 CallInst *call = CallInst::Create(newFunction, params.begin(), params.end(),
436 NumExitBlocks > 1 ? "targetBlock" : "");
437 codeReplacer->getInstList().push_back(call);
439 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
440 unsigned FirstOut = inputs.size();
442 std::advance(OutputArgBegin, inputs.size());
444 // Reload the outputs passed in by reference
445 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
449 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
450 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i);
451 GetElementPtrInst *GEP
452 = GetElementPtrInst::Create(Struct, Idx, Idx + 2,
453 "gep_reload_" + outputs[i]->getName());
454 codeReplacer->getInstList().push_back(GEP);
457 Output = ReloadOutputs[i];
459 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
460 Reloads.push_back(load);
461 codeReplacer->getInstList().push_back(load);
462 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
463 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
464 Instruction *inst = cast<Instruction>(Users[u]);
465 if (!BlocksToExtract.count(inst->getParent()))
466 inst->replaceUsesOfWith(outputs[i], load);
470 // Now we can emit a switch statement using the call as a value.
471 SwitchInst *TheSwitch =
472 SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)),
473 codeReplacer, 0, codeReplacer);
475 // Since there may be multiple exits from the original region, make the new
476 // function return an unsigned, switch on that number. This loop iterates
477 // over all of the blocks in the extracted region, updating any terminator
478 // instructions in the to-be-extracted region that branch to blocks that are
479 // not in the region to be extracted.
480 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
482 unsigned switchVal = 0;
483 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
484 e = BlocksToExtract.end(); i != e; ++i) {
485 TerminatorInst *TI = (*i)->getTerminator();
486 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
487 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
488 BasicBlock *OldTarget = TI->getSuccessor(i);
489 // add a new basic block which returns the appropriate value
490 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
492 // If we don't already have an exit stub for this non-extracted
493 // destination, create one now!
494 NewTarget = BasicBlock::Create(Context,
495 OldTarget->getName() + ".exitStub",
497 unsigned SuccNum = switchVal++;
500 switch (NumExitBlocks) {
502 case 1: break; // No value needed.
503 case 2: // Conditional branch, return a bool
504 brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum);
507 brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum);
511 ReturnInst *NTRet = ReturnInst::Create(Context, brVal, NewTarget);
513 // Update the switch instruction.
514 TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context),
518 // Restore values just before we exit
519 Function::arg_iterator OAI = OutputArgBegin;
520 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
521 // For an invoke, the normal destination is the only one that is
522 // dominated by the result of the invocation
523 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
525 bool DominatesDef = true;
527 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
528 DefBlock = Invoke->getNormalDest();
530 // Make sure we are looking at the original successor block, not
531 // at a newly inserted exit block, which won't be in the dominator
533 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
534 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
535 if (DefBlock == I->second) {
540 // In the extract block case, if the block we are extracting ends
541 // with an invoke instruction, make sure that we don't emit a
542 // store of the invoke value for the unwind block.
543 if (!DT && DefBlock != OldTarget)
544 DominatesDef = false;
548 DominatesDef = DT->dominates(DefBlock, OldTarget);
550 // If the output value is used by a phi in the target block,
551 // then we need to test for dominance of the phi's predecessor
552 // instead. Unfortunately, this a little complicated since we
553 // have already rewritten uses of the value to uses of the reload.
554 BasicBlock* pred = FindPhiPredForUseInBlock(Reloads[out],
556 if (pred && DT && DT->dominates(DefBlock, pred))
563 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
564 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context),
566 GetElementPtrInst *GEP =
567 GetElementPtrInst::Create(OAI, Idx, Idx + 2,
568 "gep_" + outputs[out]->getName(),
570 new StoreInst(outputs[out], GEP, NTRet);
572 new StoreInst(outputs[out], OAI, NTRet);
575 // Advance output iterator even if we don't emit a store
576 if (!AggregateArgs) ++OAI;
580 // rewrite the original branch instruction with this new target
581 TI->setSuccessor(i, NewTarget);
585 // Now that we've done the deed, simplify the switch instruction.
586 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
587 switch (NumExitBlocks) {
589 // There are no successors (the block containing the switch itself), which
590 // means that previously this was the last part of the function, and hence
591 // this should be rewritten as a `ret'
593 // Check if the function should return a value
594 if (OldFnRetTy == Type::getVoidTy(Context)) {
595 ReturnInst::Create(Context, 0, TheSwitch); // Return void
596 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
597 // return what we have
598 ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch);
600 // Otherwise we must have code extracted an unwind or something, just
601 // return whatever we want.
602 ReturnInst::Create(Context,
603 Constant::getNullValue(OldFnRetTy), TheSwitch);
606 TheSwitch->eraseFromParent();
609 // Only a single destination, change the switch into an unconditional
611 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
612 TheSwitch->eraseFromParent();
615 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
617 TheSwitch->eraseFromParent();
620 // Otherwise, make the default destination of the switch instruction be one
621 // of the other successors.
622 TheSwitch->setOperand(0, call);
623 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
624 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
629 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
630 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
631 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
632 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
634 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
635 e = BlocksToExtract.end(); i != e; ++i) {
636 // Delete the basic block from the old function, and the list of blocks
637 oldBlocks.remove(*i);
639 // Insert this basic block into the new function
640 newBlocks.push_back(*i);
644 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
645 /// new function. Returns pointer to the new function.
649 /// find inputs and outputs for the region
651 /// for inputs: add to function as args, map input instr* to arg#
652 /// for outputs: add allocas for scalars,
653 /// add to func as args, map output instr* to arg#
655 /// rewrite func to use argument #s instead of instr*
657 /// for each scalar output in the function: at every exit, store intermediate
658 /// computed result back into memory.
660 Function *CodeExtractor::
661 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
662 if (!isEligible(code))
665 // 1) Find inputs, outputs
666 // 2) Construct new function
667 // * Add allocas for defs, pass as args by reference
668 // * Pass in uses as args
669 // 3) Move code region, add call instr to func
671 BlocksToExtract.insert(code.begin(), code.end());
673 Values inputs, outputs;
675 // Assumption: this is a single-entry code region, and the header is the first
676 // block in the region.
677 BasicBlock *header = code[0];
679 for (unsigned i = 1, e = code.size(); i != e; ++i)
680 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
682 assert(BlocksToExtract.count(*PI) &&
683 "No blocks in this region may have entries from outside the region"
684 " except for the first block!");
686 // If we have to split PHI nodes or the entry block, do so now.
687 severSplitPHINodes(header);
689 // If we have any return instructions in the region, split those blocks so
690 // that the return is not in the region.
693 Function *oldFunction = header->getParent();
695 // This takes place of the original loop
696 BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(),
697 "codeRepl", oldFunction,
700 // The new function needs a root node because other nodes can branch to the
701 // head of the region, but the entry node of a function cannot have preds.
702 BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(),
704 newFuncRoot->getInstList().push_back(BranchInst::Create(header));
706 // Find inputs to, outputs from the code region.
707 findInputsOutputs(inputs, outputs);
709 // Construct new function based on inputs/outputs & add allocas for all defs.
710 Function *newFunction = constructFunction(inputs, outputs, header,
712 codeReplacer, oldFunction,
713 oldFunction->getParent());
715 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
717 moveCodeToFunction(newFunction);
719 // Loop over all of the PHI nodes in the header block, and change any
720 // references to the old incoming edge to be the new incoming edge.
721 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
722 PHINode *PN = cast<PHINode>(I);
723 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
724 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
725 PN->setIncomingBlock(i, newFuncRoot);
728 // Look at all successors of the codeReplacer block. If any of these blocks
729 // had PHI nodes in them, we need to update the "from" block to be the code
730 // replacer, not the original block in the extracted region.
731 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
732 succ_end(codeReplacer));
733 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
734 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
735 PHINode *PN = cast<PHINode>(I);
736 std::set<BasicBlock*> ProcessedPreds;
737 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
738 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
739 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
740 PN->setIncomingBlock(i, codeReplacer);
742 // There were multiple entries in the PHI for this block, now there
743 // is only one, so remove the duplicated entries.
744 PN->removeIncomingValue(i, false);
750 //cerr << "NEW FUNCTION: " << *newFunction;
751 // verifyFunction(*newFunction);
753 // cerr << "OLD FUNCTION: " << *oldFunction;
754 // verifyFunction(*oldFunction);
756 DEBUG(if (verifyFunction(*newFunction))
757 llvm_report_error("verifyFunction failed!"));
761 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
762 // Deny code region if it contains allocas or vastarts.
763 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
765 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
767 if (isa<AllocaInst>(*I))
769 else if (const CallInst *CI = dyn_cast<CallInst>(I))
770 if (const Function *F = CI->getCalledFunction())
771 if (F->getIntrinsicID() == Intrinsic::vastart)
777 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
780 Function* llvm::ExtractCodeRegion(DominatorTree &DT,
781 const std::vector<BasicBlock*> &code,
782 bool AggregateArgs) {
783 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code);
786 /// ExtractBasicBlock - slurp a natural loop into a brand new function
788 Function* llvm::ExtractLoop(DominatorTree &DT, Loop *L, bool AggregateArgs) {
789 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(L->getBlocks());
792 /// ExtractBasicBlock - slurp a basic block into a brand new function
794 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
795 std::vector<BasicBlock*> Blocks;
796 Blocks.push_back(BB);
797 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);