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 "Support/CommandLine.h"
28 #include "Support/Debug.h"
29 #include "Support/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 findInputsOutputs(Values &inputs, Values &outputs);
82 Function *constructFunction(const Values &inputs,
83 const Values &outputs,
85 BasicBlock *newRootNode, BasicBlock *newHeader,
86 Function *oldFunction, Module *M);
88 void moveCodeToFunction(Function *newFunction);
90 void emitCallAndSwitchStatement(Function *newFunction,
91 BasicBlock *newHeader,
98 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
99 /// region, we need to split the entry block of the region so that the PHI node
100 /// is easier to deal with.
101 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
107 // findInputsOutputs - Find inputs to, outputs from the code region.
109 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
110 std::set<BasicBlock*> ExitBlocks;
111 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
112 ce = BlocksToExtract.end(); ci != ce; ++ci) {
113 BasicBlock *BB = *ci;
115 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
116 // If a used value is defined outside the region, it's an input. If an
117 // instruction is used outside the region, it's an output.
118 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
119 if (definedInCaller(*O))
120 inputs.push_back(*O);
122 // Consider uses of this instruction (outputs).
123 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
125 if (!definedInRegion(*UI)) {
126 outputs.push_back(I);
131 // Keep track of the exit blocks from the region.
132 TerminatorInst *TI = BB->getTerminator();
133 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
134 if (!BlocksToExtract.count(TI->getSuccessor(i)))
135 ExitBlocks.insert(TI->getSuccessor(i));
136 } // for: basic blocks
138 NumExitBlocks = ExitBlocks.size();
141 /// constructFunction - make a function based on inputs and outputs, as follows:
142 /// f(in0, ..., inN, out0, ..., outN)
144 Function *CodeExtractor::constructFunction(const Values &inputs,
145 const Values &outputs,
147 BasicBlock *newRootNode,
148 BasicBlock *newHeader,
149 Function *oldFunction,
151 DEBUG(std::cerr << "inputs: " << inputs.size() << "\n");
152 DEBUG(std::cerr << "outputs: " << outputs.size() << "\n");
154 // This function returns unsigned, outputs will go back by reference.
155 switch (NumExitBlocks) {
157 case 1: RetTy = Type::VoidTy; break;
158 case 2: RetTy = Type::BoolTy; break;
159 default: RetTy = Type::UShortTy; break;
162 std::vector<const Type*> paramTy;
164 // Add the types of the input values to the function's argument list
165 for (Values::const_iterator i = inputs.begin(),
166 e = inputs.end(); i != e; ++i) {
167 const Value *value = *i;
168 DEBUG(std::cerr << "value used in func: " << value << "\n");
169 paramTy.push_back(value->getType());
172 // Add the types of the output values to the function's argument list.
173 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
175 DEBUG(std::cerr << "instr used in func: " << *I << "\n");
177 paramTy.push_back((*I)->getType());
179 paramTy.push_back(PointerType::get((*I)->getType()));
182 DEBUG(std::cerr << "Function type: " << RetTy << " f(");
183 DEBUG(for (std::vector<const Type*>::iterator i = paramTy.begin(),
184 e = paramTy.end(); i != e; ++i) std::cerr << *i << ", ");
185 DEBUG(std::cerr << ")\n");
187 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
188 PointerType *StructPtr = PointerType::get(StructType::get(paramTy));
190 paramTy.push_back(StructPtr);
192 const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false);
194 // Create the new function
195 Function *newFunction = new Function(funcType,
196 GlobalValue::InternalLinkage,
197 oldFunction->getName() + "_code", M);
198 newFunction->getBasicBlockList().push_back(newRootNode);
200 // Create an iterator to name all of the arguments we inserted.
201 Function::aiterator AI = newFunction->abegin();
203 // Rewrite all users of the inputs in the extracted region to use the
204 // arguments (or appropriate addressing into struct) instead.
205 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
208 std::vector<Value*> Indices;
209 Indices.push_back(Constant::getNullValue(Type::UIntTy));
210 Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
211 std::string GEPname = "gep_" + inputs[i]->getName();
212 TerminatorInst *TI = newFunction->begin()->getTerminator();
213 GetElementPtrInst *GEP = new GetElementPtrInst(AI, Indices, GEPname, TI);
214 RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
218 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
219 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
221 if (Instruction* inst = dyn_cast<Instruction>(*use))
222 if (BlocksToExtract.count(inst->getParent()))
223 inst->replaceUsesOfWith(inputs[i], RewriteVal);
226 // Set names for input and output arguments.
227 if (!AggregateArgs) {
228 AI = newFunction->abegin();
229 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
230 AI->setName(inputs[i]->getName());
231 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
232 AI->setName(outputs[i]->getName()+".out");
235 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
236 // within the new function. This must be done before we lose track of which
237 // blocks were originally in the code region.
238 std::vector<User*> Users(header->use_begin(), header->use_end());
239 for (unsigned i = 0, e = Users.size(); i != e; ++i)
240 // The BasicBlock which contains the branch is not in the region
241 // modify the branch target to a new block
242 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
243 if (!BlocksToExtract.count(TI->getParent()) &&
244 TI->getParent()->getParent() == oldFunction)
245 TI->replaceUsesOfWith(header, newHeader);
250 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
251 /// the call instruction, splitting any PHI nodes in the header block as
254 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
255 Values &inputs, Values &outputs) {
256 // Emit a call to the new function, passing in: *pointer to struct (if
257 // aggregating parameters), or plan inputs and allocated memory for outputs
258 std::vector<Value*> params, StructValues, ReloadOutputs;
260 // Add inputs as params, or to be filled into the struct
261 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
263 StructValues.push_back(*i);
265 params.push_back(*i);
267 // Create allocas for the outputs
268 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
270 StructValues.push_back(*i);
273 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
274 codeReplacer->getParent()->begin()->begin());
275 ReloadOutputs.push_back(alloca);
276 params.push_back(alloca);
280 AllocaInst *Struct = 0;
281 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
282 std::vector<const Type*> ArgTypes;
283 for (Values::iterator v = StructValues.begin(),
284 ve = StructValues.end(); v != ve; ++v)
285 ArgTypes.push_back((*v)->getType());
287 // Allocate a struct at the beginning of this function
288 Type *StructArgTy = StructType::get(ArgTypes);
290 new AllocaInst(StructArgTy, 0, "structArg",
291 codeReplacer->getParent()->begin()->begin());
292 params.push_back(Struct);
294 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
295 std::vector<Value*> Indices;
296 Indices.push_back(Constant::getNullValue(Type::UIntTy));
297 Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
298 GetElementPtrInst *GEP =
299 new GetElementPtrInst(Struct, Indices,
300 "gep_" + StructValues[i]->getName(), 0);
301 codeReplacer->getInstList().push_back(GEP);
302 StoreInst *SI = new StoreInst(StructValues[i], GEP);
303 codeReplacer->getInstList().push_back(SI);
307 // Emit the call to the function
308 CallInst *call = new CallInst(newFunction, params,
309 NumExitBlocks > 1 ? "targetBlock": "");
310 codeReplacer->getInstList().push_back(call);
312 Function::aiterator OutputArgBegin = newFunction->abegin();
313 unsigned FirstOut = inputs.size();
315 std::advance(OutputArgBegin, inputs.size());
317 // Reload the outputs passed in by reference
318 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
321 std::vector<Value*> Indices;
322 Indices.push_back(Constant::getNullValue(Type::UIntTy));
323 Indices.push_back(ConstantUInt::get(Type::UIntTy, FirstOut + i));
324 GetElementPtrInst *GEP
325 = new GetElementPtrInst(Struct, Indices,
326 "gep_reload_" + outputs[i]->getName(), 0);
327 codeReplacer->getInstList().push_back(GEP);
330 Output = ReloadOutputs[i];
332 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
333 codeReplacer->getInstList().push_back(load);
334 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
335 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
336 Instruction *inst = cast<Instruction>(Users[u]);
337 if (!BlocksToExtract.count(inst->getParent()))
338 inst->replaceUsesOfWith(outputs[i], load);
342 // Now we can emit a switch statement using the call as a value.
343 SwitchInst *TheSwitch =
344 new SwitchInst(ConstantUInt::getNullValue(Type::UShortTy),
345 codeReplacer, codeReplacer);
347 // Since there may be multiple exits from the original region, make the new
348 // function return an unsigned, switch on that number. This loop iterates
349 // over all of the blocks in the extracted region, updating any terminator
350 // instructions in the to-be-extracted region that branch to blocks that are
351 // not in the region to be extracted.
352 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
354 unsigned switchVal = 0;
355 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
356 e = BlocksToExtract.end(); i != e; ++i) {
357 TerminatorInst *TI = (*i)->getTerminator();
358 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
359 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
360 BasicBlock *OldTarget = TI->getSuccessor(i);
361 // add a new basic block which returns the appropriate value
362 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
364 // If we don't already have an exit stub for this non-extracted
365 // destination, create one now!
366 NewTarget = new BasicBlock(OldTarget->getName() + ".exitStub",
368 unsigned SuccNum = switchVal++;
371 switch (NumExitBlocks) {
373 case 1: break; // No value needed.
374 case 2: // Conditional branch, return a bool
375 brVal = SuccNum ? ConstantBool::False : ConstantBool::True;
378 brVal = ConstantUInt::get(Type::UShortTy, SuccNum);
382 ReturnInst *NTRet = new ReturnInst(brVal, NewTarget);
384 // Update the switch instruction.
385 TheSwitch->addCase(ConstantUInt::get(Type::UShortTy, SuccNum),
388 // Restore values just before we exit
389 Function::aiterator OAI = OutputArgBegin;
390 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
391 // For an invoke, the normal destination is the only one that is
392 // dominated by the result of the invocation
393 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
394 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out]))
395 DefBlock = Invoke->getNormalDest();
396 if (!DS || DS->dominates(DefBlock, TI->getParent()))
398 std::vector<Value*> Indices;
399 Indices.push_back(Constant::getNullValue(Type::UIntTy));
400 Indices.push_back(ConstantUInt::get(Type::UIntTy,FirstOut+out));
401 GetElementPtrInst *GEP =
402 new GetElementPtrInst(OAI, Indices,
403 "gep_" + outputs[out]->getName(),
405 new StoreInst(outputs[out], GEP, NTRet);
407 new StoreInst(outputs[out], OAI, NTRet);
408 // Advance output iterator even if we don't emit a store
409 if (!AggregateArgs) ++OAI;
413 // rewrite the original branch instruction with this new target
414 TI->setSuccessor(i, NewTarget);
418 // Now that we've done the deed, simplify the switch instruction.
419 switch (NumExitBlocks) {
421 // There is only 1 successor (the block containing the switch itself), which
422 // means that previously this was the last part of the function, and hence
423 // this should be rewritten as a `ret'
425 // Check if the function should return a value
426 if (TheSwitch->getParent()->getParent()->getReturnType() != Type::VoidTy &&
427 TheSwitch->getParent()->getParent()->getReturnType() ==
428 TheSwitch->getCondition()->getType())
429 // return what we have
430 new ReturnInst(TheSwitch->getCondition(), TheSwitch);
433 new ReturnInst(0, TheSwitch);
435 TheSwitch->getParent()->getInstList().erase(TheSwitch);
438 // Only a single destination, change the switch into an unconditional
440 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch);
441 TheSwitch->getParent()->getInstList().erase(TheSwitch);
444 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
446 TheSwitch->getParent()->getInstList().erase(TheSwitch);
449 // Otherwise, make the default destination of the switch instruction be one
450 // of the other successors.
451 TheSwitch->setOperand(0, call);
452 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
453 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
458 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
459 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
460 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
461 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
463 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
464 e = BlocksToExtract.end(); i != e; ++i) {
465 // Delete the basic block from the old function, and the list of blocks
466 oldBlocks.remove(*i);
468 // Insert this basic block into the new function
469 newBlocks.push_back(*i);
473 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
474 /// new function. Returns pointer to the new function.
478 /// find inputs and outputs for the region
480 /// for inputs: add to function as args, map input instr* to arg#
481 /// for outputs: add allocas for scalars,
482 /// add to func as args, map output instr* to arg#
484 /// rewrite func to use argument #s instead of instr*
486 /// for each scalar output in the function: at every exit, store intermediate
487 /// computed result back into memory.
489 Function *CodeExtractor::ExtractCodeRegion(const std::vector<BasicBlock*> &code)
491 if (!isEligible(code))
494 // 1) Find inputs, outputs
495 // 2) Construct new function
496 // * Add allocas for defs, pass as args by reference
497 // * Pass in uses as args
498 // 3) Move code region, add call instr to func
500 BlocksToExtract.insert(code.begin(), code.end());
502 Values inputs, outputs;
504 // Assumption: this is a single-entry code region, and the header is the first
505 // block in the region.
506 BasicBlock *header = code[0];
508 // If we have to split PHI nodes, do so now.
509 severSplitPHINodes(header);
511 for (unsigned i = 1, e = code.size(); i != e; ++i)
512 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
514 assert(BlocksToExtract.count(*PI) &&
515 "No blocks in this region may have entries from outside the region"
516 " except for the first block!");
518 Function *oldFunction = header->getParent();
520 // This takes place of the original loop
521 BasicBlock *codeReplacer = new BasicBlock("codeRepl", oldFunction);
523 // The new function needs a root node because other nodes can branch to the
524 // head of the region, but the entry node of a function cannot have preds.
525 BasicBlock *newFuncRoot = new BasicBlock("newFuncRoot");
526 newFuncRoot->getInstList().push_back(new BranchInst(header));
528 // Find inputs to, outputs from the code region.
529 findInputsOutputs(inputs, outputs);
531 // Construct new function based on inputs/outputs & add allocas for all defs.
532 Function *newFunction = constructFunction(inputs, outputs, code[0],
534 codeReplacer, oldFunction,
535 oldFunction->getParent());
537 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
539 moveCodeToFunction(newFunction);
541 // Loop over all of the PHI nodes in the entry block (code[0]), and change any
542 // references to the old incoming edge to be the new incoming edge.
543 for (BasicBlock::iterator I = code[0]->begin();
544 PHINode *PN = dyn_cast<PHINode>(I); ++I)
545 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
546 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
547 PN->setIncomingBlock(i, newFuncRoot);
549 // Look at all successors of the codeReplacer block. If any of these blocks
550 // had PHI nodes in them, we need to update the "from" block to be the code
551 // replacer, not the original block in the extracted region.
552 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
553 succ_end(codeReplacer));
554 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
555 for (BasicBlock::iterator I = Succs[i]->begin();
556 PHINode *PN = dyn_cast<PHINode>(I); ++I)
557 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
558 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
559 PN->setIncomingBlock(i, codeReplacer);
562 DEBUG(if (verifyFunction(*newFunction)) abort());
566 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
567 // Deny code region if it contains allocas or vastarts.
568 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
570 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
572 if (isa<AllocaInst>(*I))
574 else if (const CallInst *CI = dyn_cast<CallInst>(I))
575 if (const Function *F = CI->getCalledFunction())
576 if (F->getIntrinsicID() == Intrinsic::vastart)
582 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
585 Function* llvm::ExtractCodeRegion(DominatorSet &DS,
586 const std::vector<BasicBlock*> &code,
587 bool AggregateArgs) {
588 return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(code);
591 /// ExtractBasicBlock - slurp a natural loop into a brand new function
593 Function* llvm::ExtractLoop(DominatorSet &DS, Loop *L, bool AggregateArgs) {
594 return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(L->getBlocks());
597 /// ExtractBasicBlock - slurp a basic block into a brand new function
599 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
600 std::vector<BasicBlock*> Blocks;
601 Blocks.push_back(BB);
602 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);