1 //===- CodeExtractor.cpp - Pull code region into a new function -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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/Debug.h"
29 #include "llvm/ADT/StringExtras.h"
34 // Provide a command-line option to aggregate function arguments into a struct
35 // for functions produced by the code extrator. This is useful when converting
36 // extracted functions to pthread-based code, as only one argument (void*) can
37 // be passed in to pthread_create().
39 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
40 cl::desc("Aggregate arguments to code-extracted functions"));
44 typedef std::vector<Value*> Values;
45 std::set<BasicBlock*> BlocksToExtract;
48 unsigned NumExitBlocks;
51 CodeExtractor(DominatorSet *ds = 0, bool AggArgs = false)
52 : DS(ds), AggregateArgs(AggregateArgsOpt), NumExitBlocks(~0U) {}
54 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
56 bool isEligible(const std::vector<BasicBlock*> &code);
59 /// definedInRegion - Return true if the specified value is defined in the
61 bool definedInRegion(Value *V) const {
62 if (Instruction *I = dyn_cast<Instruction>(V))
63 if (BlocksToExtract.count(I->getParent()))
68 /// definedInCaller - Return true if the specified value is defined in the
69 /// function being code extracted, but not in the region being extracted.
70 /// These values must be passed in as live-ins to the function.
71 bool definedInCaller(Value *V) const {
72 if (isa<Argument>(V)) return true;
73 if (Instruction *I = dyn_cast<Instruction>(V))
74 if (!BlocksToExtract.count(I->getParent()))
79 void severSplitPHINodes(BasicBlock *&Header);
80 void splitReturnBlocks();
81 void findInputsOutputs(Values &inputs, Values &outputs);
83 Function *constructFunction(const Values &inputs,
84 const Values &outputs,
86 BasicBlock *newRootNode, BasicBlock *newHeader,
87 Function *oldFunction, Module *M);
89 void moveCodeToFunction(Function *newFunction);
91 void emitCallAndSwitchStatement(Function *newFunction,
92 BasicBlock *newHeader,
99 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
100 /// region, we need to split the entry block of the region so that the PHI node
101 /// is easier to deal with.
102 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
103 bool HasPredsFromRegion = false;
104 unsigned NumPredsOutsideRegion = 0;
106 if (Header != &Header->getParent()->front()) {
107 PHINode *PN = dyn_cast<PHINode>(Header->begin());
108 if (!PN) return; // No PHI nodes.
110 // If the header node contains any PHI nodes, check to see if there is more
111 // than one entry from outside the region. If so, we need to sever the
112 // header block into two.
113 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
114 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
115 HasPredsFromRegion = true;
117 ++NumPredsOutsideRegion;
119 // If there is one (or fewer) predecessor from outside the region, we don't
120 // need to do anything special.
121 if (NumPredsOutsideRegion <= 1) return;
124 // Otherwise, we need to split the header block into two pieces: one
125 // containing PHI nodes merging values from outside of the region, and a
126 // second that contains all of the code for the block and merges back any
127 // incoming values from inside of the region.
128 BasicBlock::iterator AfterPHIs = Header->begin();
129 while (isa<PHINode>(AfterPHIs)) ++AfterPHIs;
130 BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
131 Header->getName()+".ce");
133 // We only want to code extract the second block now, and it becomes the new
134 // header of the region.
135 BasicBlock *OldPred = Header;
136 BlocksToExtract.erase(OldPred);
137 BlocksToExtract.insert(NewBB);
140 // Okay, update dominator sets. The blocks that dominate the new one are the
141 // blocks that dominate TIBB plus the new block itself.
143 DominatorSet::DomSetType DomSet = DS->getDominators(OldPred);
144 DomSet.insert(NewBB); // A block always dominates itself.
145 DS->addBasicBlock(NewBB, DomSet);
147 // Additionally, NewBB dominates all blocks in the function that are
148 // dominated by OldPred.
149 Function *F = Header->getParent();
150 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
151 if (DS->properlyDominates(OldPred, I))
152 DS->addDominator(I, NewBB);
155 // Okay, now we need to adjust the PHI nodes and any branches from within the
156 // region to go to the new header block instead of the old header block.
157 if (HasPredsFromRegion) {
158 PHINode *PN = cast<PHINode>(OldPred->begin());
159 // Loop over all of the predecessors of OldPred that are in the region,
160 // changing them to branch to NewBB instead.
161 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
162 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
163 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
164 TI->replaceUsesOfWith(OldPred, NewBB);
167 // Okay, everthing within the region is now branching to the right block, we
168 // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
169 for (AfterPHIs = OldPred->begin();
170 PHINode *PN = dyn_cast<PHINode>(AfterPHIs); ++AfterPHIs) {
171 // Create a new PHI node in the new region, which has an incoming value
172 // from OldPred of PN.
173 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".ce",
175 NewPN->addIncoming(PN, OldPred);
177 // Loop over all of the incoming value in PN, moving them to NewPN if they
178 // are from the extracted region.
179 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
180 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
181 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
182 PN->removeIncomingValue(i);
190 void CodeExtractor::splitReturnBlocks() {
191 for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(),
192 E = BlocksToExtract.end(); I != E; ++I)
193 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator()))
194 (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
197 // findInputsOutputs - Find inputs to, outputs from the code region.
199 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
200 std::set<BasicBlock*> ExitBlocks;
201 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
202 ce = BlocksToExtract.end(); ci != ce; ++ci) {
203 BasicBlock *BB = *ci;
205 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
206 // If a used value is defined outside the region, it's an input. If an
207 // instruction is used outside the region, it's an output.
208 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
209 if (definedInCaller(*O))
210 inputs.push_back(*O);
212 // Consider uses of this instruction (outputs).
213 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
215 if (!definedInRegion(*UI)) {
216 outputs.push_back(I);
221 // Keep track of the exit blocks from the region.
222 TerminatorInst *TI = BB->getTerminator();
223 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
224 if (!BlocksToExtract.count(TI->getSuccessor(i)))
225 ExitBlocks.insert(TI->getSuccessor(i));
226 } // for: basic blocks
228 NumExitBlocks = ExitBlocks.size();
230 // Eliminate duplicates.
231 std::sort(inputs.begin(), inputs.end());
232 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
233 std::sort(outputs.begin(), outputs.end());
234 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
237 /// constructFunction - make a function based on inputs and outputs, as follows:
238 /// f(in0, ..., inN, out0, ..., outN)
240 Function *CodeExtractor::constructFunction(const Values &inputs,
241 const Values &outputs,
243 BasicBlock *newRootNode,
244 BasicBlock *newHeader,
245 Function *oldFunction,
247 DEBUG(std::cerr << "inputs: " << inputs.size() << "\n");
248 DEBUG(std::cerr << "outputs: " << outputs.size() << "\n");
250 // This function returns unsigned, outputs will go back by reference.
251 switch (NumExitBlocks) {
253 case 1: RetTy = Type::VoidTy; break;
254 case 2: RetTy = Type::BoolTy; break;
255 default: RetTy = Type::UShortTy; break;
258 std::vector<const Type*> paramTy;
260 // Add the types of the input values to the function's argument list
261 for (Values::const_iterator i = inputs.begin(),
262 e = inputs.end(); i != e; ++i) {
263 const Value *value = *i;
264 DEBUG(std::cerr << "value used in func: " << *value << "\n");
265 paramTy.push_back(value->getType());
268 // Add the types of the output values to the function's argument list.
269 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
271 DEBUG(std::cerr << "instr used in func: " << **I << "\n");
273 paramTy.push_back((*I)->getType());
275 paramTy.push_back(PointerType::get((*I)->getType()));
278 DEBUG(std::cerr << "Function type: " << *RetTy << " f(");
279 DEBUG(for (std::vector<const Type*>::iterator i = paramTy.begin(),
280 e = paramTy.end(); i != e; ++i) std::cerr << **i << ", ");
281 DEBUG(std::cerr << ")\n");
283 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
284 PointerType *StructPtr = PointerType::get(StructType::get(paramTy));
286 paramTy.push_back(StructPtr);
288 const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false);
290 // Create the new function
291 Function *newFunction = new Function(funcType,
292 GlobalValue::InternalLinkage,
293 oldFunction->getName() + "_" +
294 header->getName(), M);
295 newFunction->getBasicBlockList().push_back(newRootNode);
297 // Create an iterator to name all of the arguments we inserted.
298 Function::aiterator AI = newFunction->abegin();
300 // Rewrite all users of the inputs in the extracted region to use the
301 // arguments (or appropriate addressing into struct) instead.
302 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
305 std::vector<Value*> Indices;
306 Indices.push_back(Constant::getNullValue(Type::UIntTy));
307 Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
308 std::string GEPname = "gep_" + inputs[i]->getName();
309 TerminatorInst *TI = newFunction->begin()->getTerminator();
310 GetElementPtrInst *GEP = new GetElementPtrInst(AI, Indices, GEPname, TI);
311 RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
315 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
316 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
318 if (Instruction* inst = dyn_cast<Instruction>(*use))
319 if (BlocksToExtract.count(inst->getParent()))
320 inst->replaceUsesOfWith(inputs[i], RewriteVal);
323 // Set names for input and output arguments.
324 if (!AggregateArgs) {
325 AI = newFunction->abegin();
326 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
327 AI->setName(inputs[i]->getName());
328 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
329 AI->setName(outputs[i]->getName()+".out");
332 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
333 // within the new function. This must be done before we lose track of which
334 // blocks were originally in the code region.
335 std::vector<User*> Users(header->use_begin(), header->use_end());
336 for (unsigned i = 0, e = Users.size(); i != e; ++i)
337 // The BasicBlock which contains the branch is not in the region
338 // modify the branch target to a new block
339 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
340 if (!BlocksToExtract.count(TI->getParent()) &&
341 TI->getParent()->getParent() == oldFunction)
342 TI->replaceUsesOfWith(header, newHeader);
347 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
348 /// the call instruction, splitting any PHI nodes in the header block as
351 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
352 Values &inputs, Values &outputs) {
353 // Emit a call to the new function, passing in: *pointer to struct (if
354 // aggregating parameters), or plan inputs and allocated memory for outputs
355 std::vector<Value*> params, StructValues, ReloadOutputs;
357 // Add inputs as params, or to be filled into the struct
358 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
360 StructValues.push_back(*i);
362 params.push_back(*i);
364 // Create allocas for the outputs
365 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
367 StructValues.push_back(*i);
370 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
371 codeReplacer->getParent()->begin()->begin());
372 ReloadOutputs.push_back(alloca);
373 params.push_back(alloca);
377 AllocaInst *Struct = 0;
378 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
379 std::vector<const Type*> ArgTypes;
380 for (Values::iterator v = StructValues.begin(),
381 ve = StructValues.end(); v != ve; ++v)
382 ArgTypes.push_back((*v)->getType());
384 // Allocate a struct at the beginning of this function
385 Type *StructArgTy = StructType::get(ArgTypes);
387 new AllocaInst(StructArgTy, 0, "structArg",
388 codeReplacer->getParent()->begin()->begin());
389 params.push_back(Struct);
391 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
392 std::vector<Value*> Indices;
393 Indices.push_back(Constant::getNullValue(Type::UIntTy));
394 Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
395 GetElementPtrInst *GEP =
396 new GetElementPtrInst(Struct, Indices,
397 "gep_" + StructValues[i]->getName());
398 codeReplacer->getInstList().push_back(GEP);
399 StoreInst *SI = new StoreInst(StructValues[i], GEP);
400 codeReplacer->getInstList().push_back(SI);
404 // Emit the call to the function
405 CallInst *call = new CallInst(newFunction, params,
406 NumExitBlocks > 1 ? "targetBlock": "");
407 codeReplacer->getInstList().push_back(call);
409 Function::aiterator OutputArgBegin = newFunction->abegin();
410 unsigned FirstOut = inputs.size();
412 std::advance(OutputArgBegin, inputs.size());
414 // Reload the outputs passed in by reference
415 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
418 std::vector<Value*> Indices;
419 Indices.push_back(Constant::getNullValue(Type::UIntTy));
420 Indices.push_back(ConstantUInt::get(Type::UIntTy, FirstOut + i));
421 GetElementPtrInst *GEP
422 = new GetElementPtrInst(Struct, Indices,
423 "gep_reload_" + outputs[i]->getName());
424 codeReplacer->getInstList().push_back(GEP);
427 Output = ReloadOutputs[i];
429 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
430 codeReplacer->getInstList().push_back(load);
431 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
432 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
433 Instruction *inst = cast<Instruction>(Users[u]);
434 if (!BlocksToExtract.count(inst->getParent()))
435 inst->replaceUsesOfWith(outputs[i], load);
439 // Now we can emit a switch statement using the call as a value.
440 SwitchInst *TheSwitch =
441 new SwitchInst(ConstantUInt::getNullValue(Type::UShortTy),
442 codeReplacer, codeReplacer);
444 // Since there may be multiple exits from the original region, make the new
445 // function return an unsigned, switch on that number. This loop iterates
446 // over all of the blocks in the extracted region, updating any terminator
447 // instructions in the to-be-extracted region that branch to blocks that are
448 // not in the region to be extracted.
449 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
451 unsigned switchVal = 0;
452 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
453 e = BlocksToExtract.end(); i != e; ++i) {
454 TerminatorInst *TI = (*i)->getTerminator();
455 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
456 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
457 BasicBlock *OldTarget = TI->getSuccessor(i);
458 // add a new basic block which returns the appropriate value
459 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
461 // If we don't already have an exit stub for this non-extracted
462 // destination, create one now!
463 NewTarget = new BasicBlock(OldTarget->getName() + ".exitStub",
465 unsigned SuccNum = switchVal++;
468 switch (NumExitBlocks) {
470 case 1: break; // No value needed.
471 case 2: // Conditional branch, return a bool
472 brVal = SuccNum ? ConstantBool::False : ConstantBool::True;
475 brVal = ConstantUInt::get(Type::UShortTy, SuccNum);
479 ReturnInst *NTRet = new ReturnInst(brVal, NewTarget);
481 // Update the switch instruction.
482 TheSwitch->addCase(ConstantUInt::get(Type::UShortTy, SuccNum),
485 // Restore values just before we exit
486 Function::aiterator OAI = OutputArgBegin;
487 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
488 // For an invoke, the normal destination is the only one that is
489 // dominated by the result of the invocation
490 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
491 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out]))
492 DefBlock = Invoke->getNormalDest();
493 if (!DS || DS->dominates(DefBlock, TI->getParent()))
495 std::vector<Value*> Indices;
496 Indices.push_back(Constant::getNullValue(Type::UIntTy));
497 Indices.push_back(ConstantUInt::get(Type::UIntTy,FirstOut+out));
498 GetElementPtrInst *GEP =
499 new GetElementPtrInst(OAI, Indices,
500 "gep_" + outputs[out]->getName(),
502 new StoreInst(outputs[out], GEP, NTRet);
504 new StoreInst(outputs[out], OAI, NTRet);
505 // Advance output iterator even if we don't emit a store
506 if (!AggregateArgs) ++OAI;
510 // rewrite the original branch instruction with this new target
511 TI->setSuccessor(i, NewTarget);
515 // Now that we've done the deed, simplify the switch instruction.
516 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
517 switch (NumExitBlocks) {
519 // There are no successors (the block containing the switch itself), which
520 // means that previously this was the last part of the function, and hence
521 // this should be rewritten as a `ret'
523 // Check if the function should return a value
524 if (OldFnRetTy == Type::VoidTy) {
525 new ReturnInst(0, TheSwitch); // Return void
526 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
527 // return what we have
528 new ReturnInst(TheSwitch->getCondition(), TheSwitch);
530 // Otherwise we must have code extracted an unwind or something, just
531 // return whatever we want.
532 new ReturnInst(Constant::getNullValue(OldFnRetTy), TheSwitch);
535 TheSwitch->getParent()->getInstList().erase(TheSwitch);
538 // Only a single destination, change the switch into an unconditional
540 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch);
541 TheSwitch->getParent()->getInstList().erase(TheSwitch);
544 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
546 TheSwitch->getParent()->getInstList().erase(TheSwitch);
549 // Otherwise, make the default destination of the switch instruction be one
550 // of the other successors.
551 TheSwitch->setOperand(0, call);
552 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
553 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
558 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
559 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
560 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
561 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
563 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
564 e = BlocksToExtract.end(); i != e; ++i) {
565 // Delete the basic block from the old function, and the list of blocks
566 oldBlocks.remove(*i);
568 // Insert this basic block into the new function
569 newBlocks.push_back(*i);
573 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
574 /// new function. Returns pointer to the new function.
578 /// find inputs and outputs for the region
580 /// for inputs: add to function as args, map input instr* to arg#
581 /// for outputs: add allocas for scalars,
582 /// add to func as args, map output instr* to arg#
584 /// rewrite func to use argument #s instead of instr*
586 /// for each scalar output in the function: at every exit, store intermediate
587 /// computed result back into memory.
589 Function *CodeExtractor::
590 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
591 if (!isEligible(code))
594 // 1) Find inputs, outputs
595 // 2) Construct new function
596 // * Add allocas for defs, pass as args by reference
597 // * Pass in uses as args
598 // 3) Move code region, add call instr to func
600 BlocksToExtract.insert(code.begin(), code.end());
602 Values inputs, outputs;
604 // Assumption: this is a single-entry code region, and the header is the first
605 // block in the region.
606 BasicBlock *header = code[0];
608 for (unsigned i = 1, e = code.size(); i != e; ++i)
609 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
611 assert(BlocksToExtract.count(*PI) &&
612 "No blocks in this region may have entries from outside the region"
613 " except for the first block!");
615 // If we have to split PHI nodes or the entry block, do so now.
616 severSplitPHINodes(header);
618 // If we have any return instructions in the region, split those blocks so
619 // that the return is not in the region.
622 Function *oldFunction = header->getParent();
624 // This takes place of the original loop
625 BasicBlock *codeReplacer = new BasicBlock("codeRepl", oldFunction, header);
627 // The new function needs a root node because other nodes can branch to the
628 // head of the region, but the entry node of a function cannot have preds.
629 BasicBlock *newFuncRoot = new BasicBlock("newFuncRoot");
630 newFuncRoot->getInstList().push_back(new BranchInst(header));
632 // Find inputs to, outputs from the code region.
633 findInputsOutputs(inputs, outputs);
635 // Construct new function based on inputs/outputs & add allocas for all defs.
636 Function *newFunction = constructFunction(inputs, outputs, header,
638 codeReplacer, oldFunction,
639 oldFunction->getParent());
641 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
643 moveCodeToFunction(newFunction);
645 // Loop over all of the PHI nodes in the header block, and change any
646 // references to the old incoming edge to be the new incoming edge.
647 for (BasicBlock::iterator I = header->begin();
648 PHINode *PN = dyn_cast<PHINode>(I); ++I)
649 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
650 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
651 PN->setIncomingBlock(i, newFuncRoot);
653 // Look at all successors of the codeReplacer block. If any of these blocks
654 // had PHI nodes in them, we need to update the "from" block to be the code
655 // replacer, not the original block in the extracted region.
656 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
657 succ_end(codeReplacer));
658 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
659 for (BasicBlock::iterator I = Succs[i]->begin();
660 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
661 std::set<BasicBlock*> ProcessedPreds;
662 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
663 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
664 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
665 PN->setIncomingBlock(i, codeReplacer);
667 // There were multiple entries in the PHI for this block, now there
668 // is only one, so remove the duplicated entries.
669 PN->removeIncomingValue(i, false);
674 //std::cerr << "NEW FUNCTION: " << *newFunction;
675 // verifyFunction(*newFunction);
677 // std::cerr << "OLD FUNCTION: " << *oldFunction;
678 // verifyFunction(*oldFunction);
680 DEBUG(if (verifyFunction(*newFunction)) abort());
684 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
685 // Deny code region if it contains allocas or vastarts.
686 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
688 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
690 if (isa<AllocaInst>(*I))
692 else if (const CallInst *CI = dyn_cast<CallInst>(I))
693 if (const Function *F = CI->getCalledFunction())
694 if (F->getIntrinsicID() == Intrinsic::vastart)
700 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
703 Function* llvm::ExtractCodeRegion(DominatorSet &DS,
704 const std::vector<BasicBlock*> &code,
705 bool AggregateArgs) {
706 return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(code);
709 /// ExtractBasicBlock - slurp a natural loop into a brand new function
711 Function* llvm::ExtractLoop(DominatorSet &DS, Loop *L, bool AggregateArgs) {
712 return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(L->getBlocks());
715 /// ExtractBasicBlock - slurp a basic block into a brand new function
717 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
718 std::vector<BasicBlock*> Blocks;
719 Blocks.push_back(BB);
720 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);