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/ADT/StringExtras.h"
36 // Provide a command-line option to aggregate function arguments into a struct
37 // for functions produced by the code extractor. This is useful when converting
38 // extracted functions to pthread-based code, as only one argument (void*) can
39 // be passed in to pthread_create().
41 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
42 cl::desc("Aggregate arguments to code-extracted functions"));
45 class VISIBILITY_HIDDEN CodeExtractor {
46 typedef std::vector<Value*> Values;
47 std::set<BasicBlock*> BlocksToExtract;
50 unsigned NumExitBlocks;
53 CodeExtractor(DominatorTree* dt = 0, bool AggArgs = false)
54 : DT(dt), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
56 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
58 bool isEligible(const std::vector<BasicBlock*> &code);
61 /// definedInRegion - Return true if the specified value is defined in the
63 bool definedInRegion(Value *V) const {
64 if (Instruction *I = dyn_cast<Instruction>(V))
65 if (BlocksToExtract.count(I->getParent()))
70 /// definedInCaller - Return true if the specified value is defined in the
71 /// function being code extracted, but not in the region being extracted.
72 /// These values must be passed in as live-ins to the function.
73 bool definedInCaller(Value *V) const {
74 if (isa<Argument>(V)) return true;
75 if (Instruction *I = dyn_cast<Instruction>(V))
76 if (!BlocksToExtract.count(I->getParent()))
81 void severSplitPHINodes(BasicBlock *&Header);
82 void splitReturnBlocks();
83 void findInputsOutputs(Values &inputs, Values &outputs);
85 Function *constructFunction(const Values &inputs,
86 const Values &outputs,
88 BasicBlock *newRootNode, BasicBlock *newHeader,
89 Function *oldFunction, Module *M);
91 void moveCodeToFunction(Function *newFunction);
93 void emitCallAndSwitchStatement(Function *newFunction,
94 BasicBlock *newHeader,
101 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
102 /// region, we need to split the entry block of the region so that the PHI node
103 /// is easier to deal with.
104 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
105 bool HasPredsFromRegion = false;
106 unsigned NumPredsOutsideRegion = 0;
108 if (Header != &Header->getParent()->getEntryBlock()) {
109 PHINode *PN = dyn_cast<PHINode>(Header->begin());
110 if (!PN) return; // No PHI nodes.
112 // If the header node contains any PHI nodes, check to see if there is more
113 // than one entry from outside the region. If so, we need to sever the
114 // header block into two.
115 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
116 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
117 HasPredsFromRegion = true;
119 ++NumPredsOutsideRegion;
121 // If there is one (or fewer) predecessor from outside the region, we don't
122 // need to do anything special.
123 if (NumPredsOutsideRegion <= 1) return;
126 // Otherwise, we need to split the header block into two pieces: one
127 // containing PHI nodes merging values from outside of the region, and a
128 // second that contains all of the code for the block and merges back any
129 // incoming values from inside of the region.
130 BasicBlock::iterator AfterPHIs = Header->getFirstNonPHI();
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 = PHINode::Create(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 LLVMContext *Context = header->getContext();
243 // This function returns unsigned, outputs will go back by reference.
244 switch (NumExitBlocks) {
246 case 1: RetTy = Type::VoidTy; break;
247 case 2: RetTy = Type::Int1Ty; break;
248 default: RetTy = Type::Int16Ty; break;
251 std::vector<const Type*> paramTy;
253 // Add the types of the input values to the function's argument list
254 for (Values::const_iterator i = inputs.begin(),
255 e = inputs.end(); i != e; ++i) {
256 const Value *value = *i;
257 DOUT << "value used in func: " << *value << "\n";
258 paramTy.push_back(value->getType());
261 // Add the types of the output values to the function's argument list.
262 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
264 DOUT << "instr used in func: " << **I << "\n";
266 paramTy.push_back((*I)->getType());
269 header->getContext()->getPointerTypeUnqual((*I)->getType()));
272 DOUT << "Function type: " << *RetTy << " f(";
273 for (std::vector<const Type*>::iterator i = paramTy.begin(),
274 e = paramTy.end(); i != e; ++i)
278 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
279 PointerType *StructPtr =
280 Context->getPointerTypeUnqual(Context->getStructType(paramTy));
282 paramTy.push_back(StructPtr);
284 const FunctionType *funcType =
285 Context->getFunctionType(RetTy, paramTy, false);
287 // Create the new function
288 Function *newFunction = Function::Create(funcType,
289 GlobalValue::InternalLinkage,
290 oldFunction->getName() + "_" +
291 header->getName(), M);
292 // If the old function is no-throw, so is the new one.
293 if (oldFunction->doesNotThrow())
294 newFunction->setDoesNotThrow(true);
296 newFunction->getBasicBlockList().push_back(newRootNode);
298 // Create an iterator to name all of the arguments we inserted.
299 Function::arg_iterator AI = newFunction->arg_begin();
301 // Rewrite all users of the inputs in the extracted region to use the
302 // arguments (or appropriate addressing into struct) instead.
303 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
307 Idx[0] = Context->getNullValue(Type::Int32Ty);
308 Idx[1] = Context->getConstantInt(Type::Int32Ty, i);
309 std::string GEPname = "gep_" + inputs[i]->getName();
310 TerminatorInst *TI = newFunction->begin()->getTerminator();
311 GetElementPtrInst *GEP = GetElementPtrInst::Create(AI, Idx, Idx+2,
313 RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
317 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
318 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
320 if (Instruction* inst = dyn_cast<Instruction>(*use))
321 if (BlocksToExtract.count(inst->getParent()))
322 inst->replaceUsesOfWith(inputs[i], RewriteVal);
325 // Set names for input and output arguments.
326 if (!AggregateArgs) {
327 AI = newFunction->arg_begin();
328 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
329 AI->setName(inputs[i]->getName());
330 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
331 AI->setName(outputs[i]->getName()+".out");
334 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
335 // within the new function. This must be done before we lose track of which
336 // blocks were originally in the code region.
337 std::vector<User*> Users(header->use_begin(), header->use_end());
338 for (unsigned i = 0, e = Users.size(); i != e; ++i)
339 // The BasicBlock which contains the branch is not in the region
340 // modify the branch target to a new block
341 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
342 if (!BlocksToExtract.count(TI->getParent()) &&
343 TI->getParent()->getParent() == oldFunction)
344 TI->replaceUsesOfWith(header, newHeader);
349 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
350 /// the call instruction, splitting any PHI nodes in the header block as
353 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
354 Values &inputs, Values &outputs) {
355 LLVMContext *Context = codeReplacer->getContext();
357 // Emit a call to the new function, passing in: *pointer to struct (if
358 // aggregating parameters), or plan inputs and allocated memory for outputs
359 std::vector<Value*> params, StructValues, ReloadOutputs;
361 // Add inputs as params, or to be filled into the struct
362 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
364 StructValues.push_back(*i);
366 params.push_back(*i);
368 // Create allocas for the outputs
369 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
371 StructValues.push_back(*i);
374 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
375 codeReplacer->getParent()->begin()->begin());
376 ReloadOutputs.push_back(alloca);
377 params.push_back(alloca);
381 AllocaInst *Struct = 0;
382 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
383 std::vector<const Type*> ArgTypes;
384 for (Values::iterator v = StructValues.begin(),
385 ve = StructValues.end(); v != ve; ++v)
386 ArgTypes.push_back((*v)->getType());
388 // Allocate a struct at the beginning of this function
389 Type *StructArgTy = Context->getStructType(ArgTypes);
391 new AllocaInst(StructArgTy, 0, "structArg",
392 codeReplacer->getParent()->begin()->begin());
393 params.push_back(Struct);
395 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
397 Idx[0] = Context->getNullValue(Type::Int32Ty);
398 Idx[1] = Context->getConstantInt(Type::Int32Ty, i);
399 GetElementPtrInst *GEP =
400 GetElementPtrInst::Create(Struct, Idx, Idx + 2,
401 "gep_" + StructValues[i]->getName());
402 codeReplacer->getInstList().push_back(GEP);
403 StoreInst *SI = new StoreInst(StructValues[i], GEP);
404 codeReplacer->getInstList().push_back(SI);
408 // Emit the call to the function
409 CallInst *call = CallInst::Create(newFunction, params.begin(), params.end(),
410 NumExitBlocks > 1 ? "targetBlock" : "");
411 codeReplacer->getInstList().push_back(call);
413 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
414 unsigned FirstOut = inputs.size();
416 std::advance(OutputArgBegin, inputs.size());
418 // Reload the outputs passed in by reference
419 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
423 Idx[0] = Context->getNullValue(Type::Int32Ty);
424 Idx[1] = Context->getConstantInt(Type::Int32Ty, FirstOut + i);
425 GetElementPtrInst *GEP
426 = GetElementPtrInst::Create(Struct, Idx, Idx + 2,
427 "gep_reload_" + outputs[i]->getName());
428 codeReplacer->getInstList().push_back(GEP);
431 Output = ReloadOutputs[i];
433 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
434 codeReplacer->getInstList().push_back(load);
435 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
436 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
437 Instruction *inst = cast<Instruction>(Users[u]);
438 if (!BlocksToExtract.count(inst->getParent()))
439 inst->replaceUsesOfWith(outputs[i], load);
443 // Now we can emit a switch statement using the call as a value.
444 SwitchInst *TheSwitch =
445 SwitchInst::Create(Context->getNullValue(Type::Int16Ty),
446 codeReplacer, 0, codeReplacer);
448 // Since there may be multiple exits from the original region, make the new
449 // function return an unsigned, switch on that number. This loop iterates
450 // over all of the blocks in the extracted region, updating any terminator
451 // instructions in the to-be-extracted region that branch to blocks that are
452 // not in the region to be extracted.
453 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
455 unsigned switchVal = 0;
456 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
457 e = BlocksToExtract.end(); i != e; ++i) {
458 TerminatorInst *TI = (*i)->getTerminator();
459 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
460 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
461 BasicBlock *OldTarget = TI->getSuccessor(i);
462 // add a new basic block which returns the appropriate value
463 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
465 // If we don't already have an exit stub for this non-extracted
466 // destination, create one now!
467 NewTarget = BasicBlock::Create(OldTarget->getName() + ".exitStub",
469 unsigned SuccNum = switchVal++;
472 switch (NumExitBlocks) {
474 case 1: break; // No value needed.
475 case 2: // Conditional branch, return a bool
476 brVal = Context->getConstantInt(Type::Int1Ty, !SuccNum);
479 brVal = Context->getConstantInt(Type::Int16Ty, SuccNum);
483 ReturnInst *NTRet = ReturnInst::Create(brVal, NewTarget);
485 // Update the switch instruction.
486 TheSwitch->addCase(Context->getConstantInt(Type::Int16Ty, SuccNum),
489 // Restore values just before we exit
490 Function::arg_iterator OAI = OutputArgBegin;
491 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
492 // For an invoke, the normal destination is the only one that is
493 // dominated by the result of the invocation
494 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
496 bool DominatesDef = true;
498 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
499 DefBlock = Invoke->getNormalDest();
501 // Make sure we are looking at the original successor block, not
502 // at a newly inserted exit block, which won't be in the dominator
504 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
505 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
506 if (DefBlock == I->second) {
511 // In the extract block case, if the block we are extracting ends
512 // with an invoke instruction, make sure that we don't emit a
513 // store of the invoke value for the unwind block.
514 if (!DT && DefBlock != OldTarget)
515 DominatesDef = false;
519 DominatesDef = DT->dominates(DefBlock, OldTarget);
524 Idx[0] = Context->getNullValue(Type::Int32Ty);
525 Idx[1] = Context->getConstantInt(Type::Int32Ty,FirstOut+out);
526 GetElementPtrInst *GEP =
527 GetElementPtrInst::Create(OAI, Idx, Idx + 2,
528 "gep_" + outputs[out]->getName(),
530 new StoreInst(outputs[out], GEP, NTRet);
532 new StoreInst(outputs[out], OAI, NTRet);
535 // Advance output iterator even if we don't emit a store
536 if (!AggregateArgs) ++OAI;
540 // rewrite the original branch instruction with this new target
541 TI->setSuccessor(i, NewTarget);
545 // Now that we've done the deed, simplify the switch instruction.
546 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
547 switch (NumExitBlocks) {
549 // There are no successors (the block containing the switch itself), which
550 // means that previously this was the last part of the function, and hence
551 // this should be rewritten as a `ret'
553 // Check if the function should return a value
554 if (OldFnRetTy == Type::VoidTy) {
555 ReturnInst::Create(0, TheSwitch); // Return void
556 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
557 // return what we have
558 ReturnInst::Create(TheSwitch->getCondition(), TheSwitch);
560 // Otherwise we must have code extracted an unwind or something, just
561 // return whatever we want.
562 ReturnInst::Create(Context->getNullValue(OldFnRetTy), TheSwitch);
565 TheSwitch->eraseFromParent();
568 // Only a single destination, change the switch into an unconditional
570 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
571 TheSwitch->eraseFromParent();
574 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
576 TheSwitch->eraseFromParent();
579 // Otherwise, make the default destination of the switch instruction be one
580 // of the other successors.
581 TheSwitch->setOperand(0, call);
582 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
583 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
588 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
589 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
590 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
591 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
593 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
594 e = BlocksToExtract.end(); i != e; ++i) {
595 // Delete the basic block from the old function, and the list of blocks
596 oldBlocks.remove(*i);
598 // Insert this basic block into the new function
599 newBlocks.push_back(*i);
603 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
604 /// new function. Returns pointer to the new function.
608 /// find inputs and outputs for the region
610 /// for inputs: add to function as args, map input instr* to arg#
611 /// for outputs: add allocas for scalars,
612 /// add to func as args, map output instr* to arg#
614 /// rewrite func to use argument #s instead of instr*
616 /// for each scalar output in the function: at every exit, store intermediate
617 /// computed result back into memory.
619 Function *CodeExtractor::
620 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
621 if (!isEligible(code))
624 // 1) Find inputs, outputs
625 // 2) Construct new function
626 // * Add allocas for defs, pass as args by reference
627 // * Pass in uses as args
628 // 3) Move code region, add call instr to func
630 BlocksToExtract.insert(code.begin(), code.end());
632 Values inputs, outputs;
634 // Assumption: this is a single-entry code region, and the header is the first
635 // block in the region.
636 BasicBlock *header = code[0];
638 for (unsigned i = 1, e = code.size(); i != e; ++i)
639 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
641 assert(BlocksToExtract.count(*PI) &&
642 "No blocks in this region may have entries from outside the region"
643 " except for the first block!");
645 // If we have to split PHI nodes or the entry block, do so now.
646 severSplitPHINodes(header);
648 // If we have any return instructions in the region, split those blocks so
649 // that the return is not in the region.
652 Function *oldFunction = header->getParent();
654 // This takes place of the original loop
655 BasicBlock *codeReplacer = BasicBlock::Create("codeRepl", oldFunction,
658 // The new function needs a root node because other nodes can branch to the
659 // head of the region, but the entry node of a function cannot have preds.
660 BasicBlock *newFuncRoot = BasicBlock::Create("newFuncRoot");
661 newFuncRoot->getInstList().push_back(BranchInst::Create(header));
663 // Find inputs to, outputs from the code region.
664 findInputsOutputs(inputs, outputs);
666 // Construct new function based on inputs/outputs & add allocas for all defs.
667 Function *newFunction = constructFunction(inputs, outputs, header,
669 codeReplacer, oldFunction,
670 oldFunction->getParent());
672 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
674 moveCodeToFunction(newFunction);
676 // Loop over all of the PHI nodes in the header block, and change any
677 // references to the old incoming edge to be the new incoming edge.
678 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
679 PHINode *PN = cast<PHINode>(I);
680 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
681 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
682 PN->setIncomingBlock(i, newFuncRoot);
685 // Look at all successors of the codeReplacer block. If any of these blocks
686 // had PHI nodes in them, we need to update the "from" block to be the code
687 // replacer, not the original block in the extracted region.
688 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
689 succ_end(codeReplacer));
690 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
691 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
692 PHINode *PN = cast<PHINode>(I);
693 std::set<BasicBlock*> ProcessedPreds;
694 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
695 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
696 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
697 PN->setIncomingBlock(i, codeReplacer);
699 // There were multiple entries in the PHI for this block, now there
700 // is only one, so remove the duplicated entries.
701 PN->removeIncomingValue(i, false);
707 //cerr << "NEW FUNCTION: " << *newFunction;
708 // verifyFunction(*newFunction);
710 // cerr << "OLD FUNCTION: " << *oldFunction;
711 // verifyFunction(*oldFunction);
713 DEBUG(if (verifyFunction(*newFunction)) abort());
717 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
718 // Deny code region if it contains allocas or vastarts.
719 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
721 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
723 if (isa<AllocaInst>(*I))
725 else if (const CallInst *CI = dyn_cast<CallInst>(I))
726 if (const Function *F = CI->getCalledFunction())
727 if (F->getIntrinsicID() == Intrinsic::vastart)
733 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
736 Function* llvm::ExtractCodeRegion(DominatorTree &DT,
737 const std::vector<BasicBlock*> &code,
738 bool AggregateArgs) {
739 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code);
742 /// ExtractBasicBlock - slurp a natural loop into a brand new function
744 Function* llvm::ExtractLoop(DominatorTree &DT, Loop *L, bool AggregateArgs) {
745 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(L->getBlocks());
748 /// ExtractBasicBlock - slurp a basic block into a brand new function
750 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
751 std::vector<BasicBlock*> Blocks;
752 Blocks.push_back(BB);
753 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);