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/SetVector.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"));
48 typedef SetVector<Value*> Values;
49 SetVector<BasicBlock*> BlocksToExtract;
52 unsigned NumExitBlocks;
55 CodeExtractor(DominatorTree* dt = 0, bool AggArgs = false)
56 : DT(dt), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
58 Function *ExtractCodeRegion(ArrayRef<BasicBlock*> code);
60 bool isEligible(ArrayRef<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 unsigned NumPredsFromRegion = 0;
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 ++NumPredsFromRegion;
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.remove(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 (NumPredsFromRegion) {
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, everything 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(), 1 + NumPredsFromRegion,
167 PN->getName()+".ce", NewBB->begin());
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 (SetVector<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 dominates 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 (SetVector<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))
221 // Consider uses of this instruction (outputs).
222 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
224 if (!definedInRegion(*UI)) {
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();
240 /// constructFunction - make a function based on inputs and outputs, as follows:
241 /// f(in0, ..., inN, out0, ..., outN)
243 Function *CodeExtractor::constructFunction(const Values &inputs,
244 const Values &outputs,
246 BasicBlock *newRootNode,
247 BasicBlock *newHeader,
248 Function *oldFunction,
250 DEBUG(dbgs() << "inputs: " << inputs.size() << "\n");
251 DEBUG(dbgs() << "outputs: " << outputs.size() << "\n");
253 // This function returns unsigned, outputs will go back by reference.
254 switch (NumExitBlocks) {
256 case 1: RetTy = Type::getVoidTy(header->getContext()); break;
257 case 2: RetTy = Type::getInt1Ty(header->getContext()); break;
258 default: RetTy = Type::getInt16Ty(header->getContext()); break;
261 std::vector<Type*> paramTy;
263 // Add the types of the input values to the function's argument list
264 for (Values::const_iterator i = inputs.begin(),
265 e = inputs.end(); i != e; ++i) {
266 const Value *value = *i;
267 DEBUG(dbgs() << "value used in func: " << *value << "\n");
268 paramTy.push_back(value->getType());
271 // Add the types of the output values to the function's argument list.
272 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
274 DEBUG(dbgs() << "instr used in func: " << **I << "\n");
276 paramTy.push_back((*I)->getType());
278 paramTy.push_back(PointerType::getUnqual((*I)->getType()));
281 DEBUG(dbgs() << "Function type: " << *RetTy << " f(");
282 for (std::vector<Type*>::iterator i = paramTy.begin(),
283 e = paramTy.end(); i != e; ++i)
284 DEBUG(dbgs() << **i << ", ");
285 DEBUG(dbgs() << ")\n");
287 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
288 PointerType *StructPtr =
289 PointerType::getUnqual(StructType::get(M->getContext(), paramTy));
291 paramTy.push_back(StructPtr);
293 FunctionType *funcType =
294 FunctionType::get(RetTy, paramTy, false);
296 // Create the new function
297 Function *newFunction = Function::Create(funcType,
298 GlobalValue::InternalLinkage,
299 oldFunction->getName() + "_" +
300 header->getName(), M);
301 // If the old function is no-throw, so is the new one.
302 if (oldFunction->doesNotThrow())
303 newFunction->setDoesNotThrow(true);
305 newFunction->getBasicBlockList().push_back(newRootNode);
307 // Create an iterator to name all of the arguments we inserted.
308 Function::arg_iterator AI = newFunction->arg_begin();
310 // Rewrite all users of the inputs in the extracted region to use the
311 // arguments (or appropriate addressing into struct) instead.
312 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
316 Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext()));
317 Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i);
318 TerminatorInst *TI = newFunction->begin()->getTerminator();
319 GetElementPtrInst *GEP =
320 GetElementPtrInst::Create(AI, Idx, "gep_" + inputs[i]->getName(), TI);
321 RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI);
325 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
326 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
328 if (Instruction* inst = dyn_cast<Instruction>(*use))
329 if (BlocksToExtract.count(inst->getParent()))
330 inst->replaceUsesOfWith(inputs[i], RewriteVal);
333 // Set names for input and output arguments.
334 if (!AggregateArgs) {
335 AI = newFunction->arg_begin();
336 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
337 AI->setName(inputs[i]->getName());
338 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
339 AI->setName(outputs[i]->getName()+".out");
342 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
343 // within the new function. This must be done before we lose track of which
344 // blocks were originally in the code region.
345 std::vector<User*> Users(header->use_begin(), header->use_end());
346 for (unsigned i = 0, e = Users.size(); i != e; ++i)
347 // The BasicBlock which contains the branch is not in the region
348 // modify the branch target to a new block
349 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
350 if (!BlocksToExtract.count(TI->getParent()) &&
351 TI->getParent()->getParent() == oldFunction)
352 TI->replaceUsesOfWith(header, newHeader);
357 /// FindPhiPredForUseInBlock - Given a value and a basic block, find a PHI
358 /// that uses the value within the basic block, and return the predecessor
359 /// block associated with that use, or return 0 if none is found.
360 static BasicBlock* FindPhiPredForUseInBlock(Value* Used, BasicBlock* BB) {
361 for (Value::use_iterator UI = Used->use_begin(),
362 UE = Used->use_end(); UI != UE; ++UI) {
363 PHINode *P = dyn_cast<PHINode>(*UI);
364 if (P && P->getParent() == BB)
365 return P->getIncomingBlock(UI);
371 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
372 /// the call instruction, splitting any PHI nodes in the header block as
375 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
376 Values &inputs, Values &outputs) {
377 // Emit a call to the new function, passing in: *pointer to struct (if
378 // aggregating parameters), or plan inputs and allocated memory for outputs
379 std::vector<Value*> params, StructValues, ReloadOutputs, Reloads;
381 LLVMContext &Context = newFunction->getContext();
383 // Add inputs as params, or to be filled into the struct
384 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
386 StructValues.push_back(*i);
388 params.push_back(*i);
390 // Create allocas for the outputs
391 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
393 StructValues.push_back(*i);
396 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
397 codeReplacer->getParent()->begin()->begin());
398 ReloadOutputs.push_back(alloca);
399 params.push_back(alloca);
403 AllocaInst *Struct = 0;
404 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
405 std::vector<Type*> ArgTypes;
406 for (Values::iterator v = StructValues.begin(),
407 ve = StructValues.end(); v != ve; ++v)
408 ArgTypes.push_back((*v)->getType());
410 // Allocate a struct at the beginning of this function
411 Type *StructArgTy = StructType::get(newFunction->getContext(), ArgTypes);
413 new AllocaInst(StructArgTy, 0, "structArg",
414 codeReplacer->getParent()->begin()->begin());
415 params.push_back(Struct);
417 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
419 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
420 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i);
421 GetElementPtrInst *GEP =
422 GetElementPtrInst::Create(Struct, Idx,
423 "gep_" + StructValues[i]->getName());
424 codeReplacer->getInstList().push_back(GEP);
425 StoreInst *SI = new StoreInst(StructValues[i], GEP);
426 codeReplacer->getInstList().push_back(SI);
430 // Emit the call to the function
431 CallInst *call = CallInst::Create(newFunction, params,
432 NumExitBlocks > 1 ? "targetBlock" : "");
433 codeReplacer->getInstList().push_back(call);
435 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
436 unsigned FirstOut = inputs.size();
438 std::advance(OutputArgBegin, inputs.size());
440 // Reload the outputs passed in by reference
441 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
445 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
446 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i);
447 GetElementPtrInst *GEP
448 = GetElementPtrInst::Create(Struct, Idx,
449 "gep_reload_" + outputs[i]->getName());
450 codeReplacer->getInstList().push_back(GEP);
453 Output = ReloadOutputs[i];
455 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
456 Reloads.push_back(load);
457 codeReplacer->getInstList().push_back(load);
458 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
459 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
460 Instruction *inst = cast<Instruction>(Users[u]);
461 if (!BlocksToExtract.count(inst->getParent()))
462 inst->replaceUsesOfWith(outputs[i], load);
466 // Now we can emit a switch statement using the call as a value.
467 SwitchInst *TheSwitch =
468 SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)),
469 codeReplacer, 0, codeReplacer);
471 // Since there may be multiple exits from the original region, make the new
472 // function return an unsigned, switch on that number. This loop iterates
473 // over all of the blocks in the extracted region, updating any terminator
474 // instructions in the to-be-extracted region that branch to blocks that are
475 // not in the region to be extracted.
476 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
478 unsigned switchVal = 0;
479 for (SetVector<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
480 e = BlocksToExtract.end(); i != e; ++i) {
481 TerminatorInst *TI = (*i)->getTerminator();
482 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
483 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
484 BasicBlock *OldTarget = TI->getSuccessor(i);
485 // add a new basic block which returns the appropriate value
486 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
488 // If we don't already have an exit stub for this non-extracted
489 // destination, create one now!
490 NewTarget = BasicBlock::Create(Context,
491 OldTarget->getName() + ".exitStub",
493 unsigned SuccNum = switchVal++;
496 switch (NumExitBlocks) {
498 case 1: break; // No value needed.
499 case 2: // Conditional branch, return a bool
500 brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum);
503 brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum);
507 ReturnInst *NTRet = ReturnInst::Create(Context, brVal, NewTarget);
509 // Update the switch instruction.
510 TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context),
514 // Restore values just before we exit
515 Function::arg_iterator OAI = OutputArgBegin;
516 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
517 // For an invoke, the normal destination is the only one that is
518 // dominated by the result of the invocation
519 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
521 bool DominatesDef = true;
523 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
524 DefBlock = Invoke->getNormalDest();
526 // Make sure we are looking at the original successor block, not
527 // at a newly inserted exit block, which won't be in the dominator
529 for (std::map<BasicBlock*, BasicBlock*>::iterator I =
530 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
531 if (DefBlock == I->second) {
536 // In the extract block case, if the block we are extracting ends
537 // with an invoke instruction, make sure that we don't emit a
538 // store of the invoke value for the unwind block.
539 if (!DT && DefBlock != OldTarget)
540 DominatesDef = false;
544 DominatesDef = DT->dominates(DefBlock, OldTarget);
546 // If the output value is used by a phi in the target block,
547 // then we need to test for dominance of the phi's predecessor
548 // instead. Unfortunately, this a little complicated since we
549 // have already rewritten uses of the value to uses of the reload.
550 BasicBlock* pred = FindPhiPredForUseInBlock(Reloads[out],
552 if (pred && DT && DT->dominates(DefBlock, pred))
559 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
560 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context),
562 GetElementPtrInst *GEP =
563 GetElementPtrInst::Create(OAI, Idx,
564 "gep_" + outputs[out]->getName(),
566 new StoreInst(outputs[out], GEP, NTRet);
568 new StoreInst(outputs[out], OAI, NTRet);
571 // Advance output iterator even if we don't emit a store
572 if (!AggregateArgs) ++OAI;
576 // rewrite the original branch instruction with this new target
577 TI->setSuccessor(i, NewTarget);
581 // Now that we've done the deed, simplify the switch instruction.
582 Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
583 switch (NumExitBlocks) {
585 // There are no successors (the block containing the switch itself), which
586 // means that previously this was the last part of the function, and hence
587 // this should be rewritten as a `ret'
589 // Check if the function should return a value
590 if (OldFnRetTy->isVoidTy()) {
591 ReturnInst::Create(Context, 0, TheSwitch); // Return void
592 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
593 // return what we have
594 ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch);
596 // Otherwise we must have code extracted an unwind or something, just
597 // return whatever we want.
598 ReturnInst::Create(Context,
599 Constant::getNullValue(OldFnRetTy), TheSwitch);
602 TheSwitch->eraseFromParent();
605 // Only a single destination, change the switch into an unconditional
607 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
608 TheSwitch->eraseFromParent();
611 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
613 TheSwitch->eraseFromParent();
616 // Otherwise, make the default destination of the switch instruction be one
617 // of the other successors.
618 TheSwitch->setOperand(0, call);
619 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
620 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
625 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
626 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
627 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
628 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
630 for (SetVector<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
631 e = BlocksToExtract.end(); i != e; ++i) {
632 // Delete the basic block from the old function, and the list of blocks
633 oldBlocks.remove(*i);
635 // Insert this basic block into the new function
636 newBlocks.push_back(*i);
640 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
641 /// new function. Returns pointer to the new function.
645 /// find inputs and outputs for the region
647 /// for inputs: add to function as args, map input instr* to arg#
648 /// for outputs: add allocas for scalars,
649 /// add to func as args, map output instr* to arg#
651 /// rewrite func to use argument #s instead of instr*
653 /// for each scalar output in the function: at every exit, store intermediate
654 /// computed result back into memory.
656 Function *CodeExtractor::
657 ExtractCodeRegion(ArrayRef<BasicBlock*> code) {
658 if (!isEligible(code))
661 // 1) Find inputs, outputs
662 // 2) Construct new function
663 // * Add allocas for defs, pass as args by reference
664 // * Pass in uses as args
665 // 3) Move code region, add call instr to func
667 BlocksToExtract.insert(code.begin(), code.end());
669 Values inputs, outputs;
671 // Assumption: this is a single-entry code region, and the header is the first
672 // block in the region.
673 BasicBlock *header = code[0];
675 for (unsigned i = 1, e = code.size(); i != e; ++i)
676 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
678 assert(BlocksToExtract.count(*PI) &&
679 "No blocks in this region may have entries from outside the region"
680 " except for the first block!");
682 // If we have to split PHI nodes or the entry block, do so now.
683 severSplitPHINodes(header);
685 // If we have any return instructions in the region, split those blocks so
686 // that the return is not in the region.
689 Function *oldFunction = header->getParent();
691 // This takes place of the original loop
692 BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(),
693 "codeRepl", oldFunction,
696 // The new function needs a root node because other nodes can branch to the
697 // head of the region, but the entry node of a function cannot have preds.
698 BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(),
700 newFuncRoot->getInstList().push_back(BranchInst::Create(header));
702 // Find inputs to, outputs from the code region.
703 findInputsOutputs(inputs, outputs);
705 // Construct new function based on inputs/outputs & add allocas for all defs.
706 Function *newFunction = constructFunction(inputs, outputs, header,
708 codeReplacer, oldFunction,
709 oldFunction->getParent());
711 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
713 moveCodeToFunction(newFunction);
715 // Loop over all of the PHI nodes in the header block, and change any
716 // references to the old incoming edge to be the new incoming edge.
717 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
718 PHINode *PN = cast<PHINode>(I);
719 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
720 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
721 PN->setIncomingBlock(i, newFuncRoot);
724 // Look at all successors of the codeReplacer block. If any of these blocks
725 // had PHI nodes in them, we need to update the "from" block to be the code
726 // replacer, not the original block in the extracted region.
727 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
728 succ_end(codeReplacer));
729 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
730 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
731 PHINode *PN = cast<PHINode>(I);
732 std::set<BasicBlock*> ProcessedPreds;
733 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
734 if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
735 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
736 PN->setIncomingBlock(i, codeReplacer);
738 // There were multiple entries in the PHI for this block, now there
739 // is only one, so remove the duplicated entries.
740 PN->removeIncomingValue(i, false);
746 //cerr << "NEW FUNCTION: " << *newFunction;
747 // verifyFunction(*newFunction);
749 // cerr << "OLD FUNCTION: " << *oldFunction;
750 // verifyFunction(*oldFunction);
752 DEBUG(if (verifyFunction(*newFunction))
753 report_fatal_error("verifyFunction failed!"));
757 bool CodeExtractor::isEligible(ArrayRef<BasicBlock*> code) {
758 // Deny a single basic block that's a landing pad block.
759 if (code.size() == 1 && code[0]->isLandingPad())
762 // Deny code region if it contains allocas or vastarts.
763 for (ArrayRef<BasicBlock*>::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 ArrayRef<BasicBlock*> code,
782 bool AggregateArgs) {
783 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code);
786 /// ExtractLoop - 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(ArrayRef<BasicBlock*> BBs, bool AggregateArgs){
795 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(BBs);