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/Module.h"
21 #include "llvm/Pass.h"
22 #include "llvm/Analysis/Dominators.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/Verifier.h"
25 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
26 #include "Support/CommandLine.h"
27 #include "Support/Debug.h"
28 #include "Support/StringExtras.h"
33 // Provide a command-line option to aggregate function arguments into a struct
34 // for functions produced by the code extrator. This is useful when converting
35 // extracted functions to pthread-based code, as only one argument (void*) can
36 // be passed in to pthread_create().
38 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
39 cl::desc("Aggregate arguments to code-extracted functions"));
43 typedef std::vector<Value*> Values;
44 std::set<BasicBlock*> BlocksToExtract;
47 unsigned NumExitBlocks;
50 CodeExtractor(DominatorSet *ds = 0, bool AggArgs = false)
51 : DS(ds), AggregateArgs(AggregateArgsOpt), NumExitBlocks(~0U) {}
53 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
55 bool isEligible(const std::vector<BasicBlock*> &code);
58 void findInputsOutputs(Values &inputs, Values &outputs,
59 BasicBlock *newHeader,
60 BasicBlock *newRootNode);
62 Function *constructFunction(const Values &inputs,
63 const Values &outputs,
65 BasicBlock *newRootNode, BasicBlock *newHeader,
66 Function *oldFunction, Module *M);
68 void moveCodeToFunction(Function *newFunction);
70 void emitCallAndSwitchStatement(Function *newFunction,
71 BasicBlock *newHeader,
78 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs,
79 BasicBlock *newHeader,
80 BasicBlock *newRootNode) {
81 std::set<BasicBlock*> ExitBlocks;
82 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
83 ce = BlocksToExtract.end(); ci != ce; ++ci) {
85 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
86 // If a used value is defined outside the region, it's an input. If an
87 // instruction is used outside the region, it's an output.
88 if (PHINode *PN = dyn_cast<PHINode>(I)) {
89 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
90 Value *V = PN->getIncomingValue(i);
91 if (!BlocksToExtract.count(PN->getIncomingBlock(i)) &&
92 (isa<Instruction>(V) || isa<Argument>(V)))
94 else if (Instruction *opI = dyn_cast<Instruction>(V)) {
95 if (!BlocksToExtract.count(opI->getParent()))
96 inputs.push_back(opI);
97 } else if (isa<Argument>(V))
101 // All other instructions go through the generic input finder
102 // Loop over the operands of each instruction (inputs)
103 for (User::op_iterator op = I->op_begin(), opE = I->op_end();
105 if (Instruction *opI = dyn_cast<Instruction>(*op)) {
106 // Check if definition of this operand is within the loop
107 if (!BlocksToExtract.count(opI->getParent()))
108 inputs.push_back(opI);
109 } else if (isa<Argument>(*op)) {
110 inputs.push_back(*op);
114 // Consider uses of this instruction (outputs)
115 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
117 if (!BlocksToExtract.count(cast<Instruction>(*UI)->getParent())) {
118 outputs.push_back(I);
123 TerminatorInst *TI = BB->getTerminator();
124 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
125 if (!BlocksToExtract.count(TI->getSuccessor(i)))
126 ExitBlocks.insert(TI->getSuccessor(i));
127 } // for: basic blocks
129 NumExitBlocks = ExitBlocks.size();
132 /// constructFunction - make a function based on inputs and outputs, as follows:
133 /// f(in0, ..., inN, out0, ..., outN)
135 Function *CodeExtractor::constructFunction(const Values &inputs,
136 const Values &outputs,
138 BasicBlock *newRootNode,
139 BasicBlock *newHeader,
140 Function *oldFunction,
142 DEBUG(std::cerr << "inputs: " << inputs.size() << "\n");
143 DEBUG(std::cerr << "outputs: " << outputs.size() << "\n");
145 // This function returns unsigned, outputs will go back by reference.
146 switch (NumExitBlocks) {
148 case 1: RetTy = Type::VoidTy; break;
149 case 2: RetTy = Type::BoolTy; break;
150 default: RetTy = Type::UShortTy; break;
153 std::vector<const Type*> paramTy;
155 // Add the types of the input values to the function's argument list
156 for (Values::const_iterator i = inputs.begin(),
157 e = inputs.end(); i != e; ++i) {
158 const Value *value = *i;
159 DEBUG(std::cerr << "value used in func: " << value << "\n");
160 paramTy.push_back(value->getType());
163 // Add the types of the output values to the function's argument list.
164 for (Values::const_iterator I = outputs.begin(), E = outputs.end();
166 DEBUG(std::cerr << "instr used in func: " << *I << "\n");
168 paramTy.push_back((*I)->getType());
170 paramTy.push_back(PointerType::get((*I)->getType()));
173 DEBUG(std::cerr << "Function type: " << RetTy << " f(");
174 DEBUG(for (std::vector<const Type*>::iterator i = paramTy.begin(),
175 e = paramTy.end(); i != e; ++i) std::cerr << *i << ", ");
176 DEBUG(std::cerr << ")\n");
178 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
179 PointerType *StructPtr = PointerType::get(StructType::get(paramTy));
181 paramTy.push_back(StructPtr);
183 const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false);
185 // Create the new function
186 Function *newFunction = new Function(funcType,
187 GlobalValue::InternalLinkage,
188 oldFunction->getName() + "_code", M);
189 newFunction->getBasicBlockList().push_back(newRootNode);
191 // Create an iterator to name all of the arguments we inserted.
192 Function::aiterator AI = newFunction->abegin();
194 // Rewrite all users of the inputs in the extracted region to use the
195 // arguments (or appropriate addressing into struct) instead.
196 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
199 std::vector<Value*> Indices;
200 Indices.push_back(Constant::getNullValue(Type::UIntTy));
201 Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
202 std::string GEPname = "gep_" + inputs[i]->getName();
203 TerminatorInst *TI = newFunction->begin()->getTerminator();
204 GetElementPtrInst *GEP = new GetElementPtrInst(AI, Indices, GEPname, TI);
205 RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
209 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
210 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
212 if (Instruction* inst = dyn_cast<Instruction>(*use))
213 if (BlocksToExtract.count(inst->getParent()))
214 inst->replaceUsesOfWith(inputs[i], RewriteVal);
217 // Set names for input and output arguments.
218 if (!AggregateArgs) {
219 AI = newFunction->abegin();
220 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
221 AI->setName(inputs[i]->getName());
222 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
223 AI->setName(outputs[i]->getName()+".out");
226 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
227 // within the new function. This must be done before we lose track of which
228 // blocks were originally in the code region.
229 std::vector<User*> Users(header->use_begin(), header->use_end());
230 for (unsigned i = 0, e = Users.size(); i != e; ++i)
231 // The BasicBlock which contains the branch is not in the region
232 // modify the branch target to a new block
233 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
234 if (!BlocksToExtract.count(TI->getParent()) &&
235 TI->getParent()->getParent() == oldFunction)
236 TI->replaceUsesOfWith(header, newHeader);
241 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
242 Function *oldFunc = (*BlocksToExtract.begin())->getParent();
243 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
244 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
246 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
247 e = BlocksToExtract.end(); i != e; ++i) {
248 // Delete the basic block from the old function, and the list of blocks
249 oldBlocks.remove(*i);
251 // Insert this basic block into the new function
252 newBlocks.push_back(*i);
257 CodeExtractor::emitCallAndSwitchStatement(Function *newFunction,
258 BasicBlock *codeReplacer,
262 // Emit a call to the new function, passing in:
263 // *pointer to struct (if aggregating parameters), or
264 // plan inputs and allocated memory for outputs
265 std::vector<Value*> params, StructValues, ReloadOutputs;
267 // Add inputs as params, or to be filled into the struct
268 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
270 StructValues.push_back(*i);
272 params.push_back(*i);
274 // Create allocas for the outputs
275 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
277 StructValues.push_back(*i);
280 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
281 codeReplacer->getParent()->begin()->begin());
282 ReloadOutputs.push_back(alloca);
283 params.push_back(alloca);
287 AllocaInst *Struct = 0;
288 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
289 std::vector<const Type*> ArgTypes;
290 for (Values::iterator v = StructValues.begin(),
291 ve = StructValues.end(); v != ve; ++v)
292 ArgTypes.push_back((*v)->getType());
294 // Allocate a struct at the beginning of this function
295 Type *StructArgTy = StructType::get(ArgTypes);
297 new AllocaInst(StructArgTy, 0, "structArg",
298 codeReplacer->getParent()->begin()->begin());
299 params.push_back(Struct);
301 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
302 std::vector<Value*> Indices;
303 Indices.push_back(Constant::getNullValue(Type::UIntTy));
304 Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
305 GetElementPtrInst *GEP =
306 new GetElementPtrInst(Struct, Indices,
307 "gep_" + StructValues[i]->getName(), 0);
308 codeReplacer->getInstList().push_back(GEP);
309 StoreInst *SI = new StoreInst(StructValues[i], GEP);
310 codeReplacer->getInstList().push_back(SI);
314 // Emit the call to the function
315 CallInst *call = new CallInst(newFunction, params,
316 NumExitBlocks > 1 ? "targetBlock": "");
317 codeReplacer->getInstList().push_back(call);
319 Function::aiterator OutputArgBegin = newFunction->abegin();
320 unsigned FirstOut = inputs.size();
322 std::advance(OutputArgBegin, inputs.size());
324 // Reload the outputs passed in by reference
325 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
328 std::vector<Value*> Indices;
329 Indices.push_back(Constant::getNullValue(Type::UIntTy));
330 Indices.push_back(ConstantUInt::get(Type::UIntTy, FirstOut + i));
331 GetElementPtrInst *GEP
332 = new GetElementPtrInst(Struct, Indices,
333 "gep_reload_" + outputs[i]->getName(), 0);
334 codeReplacer->getInstList().push_back(GEP);
337 Output = ReloadOutputs[i];
339 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
340 codeReplacer->getInstList().push_back(load);
341 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
342 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
343 Instruction *inst = cast<Instruction>(Users[u]);
344 if (!BlocksToExtract.count(inst->getParent()))
345 inst->replaceUsesOfWith(outputs[i], load);
349 // Now we can emit a switch statement using the call as a value.
350 SwitchInst *TheSwitch =
351 new SwitchInst(ConstantUInt::getNullValue(Type::UShortTy),
352 codeReplacer, codeReplacer);
354 // Since there may be multiple exits from the original region, make the new
355 // function return an unsigned, switch on that number. This loop iterates
356 // over all of the blocks in the extracted region, updating any terminator
357 // instructions in the to-be-extracted region that branch to blocks that are
358 // not in the region to be extracted.
359 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
361 unsigned switchVal = 0;
362 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
363 e = BlocksToExtract.end(); i != e; ++i) {
364 TerminatorInst *TI = (*i)->getTerminator();
365 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
366 if (!BlocksToExtract.count(TI->getSuccessor(i))) {
367 BasicBlock *OldTarget = TI->getSuccessor(i);
368 // add a new basic block which returns the appropriate value
369 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
371 // If we don't already have an exit stub for this non-extracted
372 // destination, create one now!
373 NewTarget = new BasicBlock(OldTarget->getName() + ".exitStub",
375 unsigned SuccNum = switchVal++;
378 switch (NumExitBlocks) {
380 case 1: break; // No value needed.
381 case 2: // Conditional branch, return a bool
382 brVal = SuccNum ? ConstantBool::False : ConstantBool::True;
385 brVal = ConstantUInt::get(Type::UShortTy, SuccNum);
389 ReturnInst *NTRet = new ReturnInst(brVal, NewTarget);
391 // Update the switch instruction.
392 TheSwitch->addCase(ConstantUInt::get(Type::UShortTy, SuccNum),
395 // Restore values just before we exit
396 Function::aiterator OAI = OutputArgBegin;
397 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
398 // For an invoke, the normal destination is the only one that is
399 // dominated by the result of the invocation
400 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
401 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out]))
402 DefBlock = Invoke->getNormalDest();
403 if (!DS || DS->dominates(DefBlock, TI->getParent()))
405 std::vector<Value*> Indices;
406 Indices.push_back(Constant::getNullValue(Type::UIntTy));
407 Indices.push_back(ConstantUInt::get(Type::UIntTy,FirstOut+out));
408 GetElementPtrInst *GEP =
409 new GetElementPtrInst(OAI, Indices,
410 "gep_" + outputs[out]->getName(),
412 new StoreInst(outputs[out], GEP, NTRet);
414 new StoreInst(outputs[out], OAI, NTRet);
415 // Advance output iterator even if we don't emit a store
416 if (!AggregateArgs) ++OAI;
420 // rewrite the original branch instruction with this new target
421 TI->setSuccessor(i, NewTarget);
425 // Now that we've done the deed, simplify the switch instruction.
426 switch (NumExitBlocks) {
428 // There is only 1 successor (the block containing the switch itself), which
429 // means that previously this was the last part of the function, and hence
430 // this should be rewritten as a `ret'
432 // Check if the function should return a value
433 if (TheSwitch->getParent()->getParent()->getReturnType() != Type::VoidTy &&
434 TheSwitch->getParent()->getParent()->getReturnType() ==
435 TheSwitch->getCondition()->getType())
436 // return what we have
437 new ReturnInst(TheSwitch->getCondition(), TheSwitch);
440 new ReturnInst(0, TheSwitch);
442 TheSwitch->getParent()->getInstList().erase(TheSwitch);
445 // Only a single destination, change the switch into an unconditional
447 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch);
448 TheSwitch->getParent()->getInstList().erase(TheSwitch);
451 new BranchInst(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
453 TheSwitch->getParent()->getInstList().erase(TheSwitch);
456 // Otherwise, make the default destination of the switch instruction be one
457 // of the other successors.
458 TheSwitch->setOperand(0, call);
459 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
460 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case
466 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
467 /// new function. Returns pointer to the new function.
471 /// find inputs and outputs for the region
473 /// for inputs: add to function as args, map input instr* to arg#
474 /// for outputs: add allocas for scalars,
475 /// add to func as args, map output instr* to arg#
477 /// rewrite func to use argument #s instead of instr*
479 /// for each scalar output in the function: at every exit, store intermediate
480 /// computed result back into memory.
482 Function *CodeExtractor::ExtractCodeRegion(const std::vector<BasicBlock*> &code)
484 if (!isEligible(code))
487 // 1) Find inputs, outputs
488 // 2) Construct new function
489 // * Add allocas for defs, pass as args by reference
490 // * Pass in uses as args
491 // 3) Move code region, add call instr to func
493 BlocksToExtract.insert(code.begin(), code.end());
495 Values inputs, outputs;
497 // Assumption: this is a single-entry code region, and the header is the first
498 // block in the region.
499 BasicBlock *header = code[0];
500 for (unsigned i = 1, e = code.size(); i != e; ++i)
501 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
503 assert(BlocksToExtract.count(*PI) &&
504 "No blocks in this region may have entries from outside the region"
505 " except for the first block!");
507 Function *oldFunction = header->getParent();
509 // This takes place of the original loop
510 BasicBlock *codeReplacer = new BasicBlock("codeRepl", oldFunction);
512 // The new function needs a root node because other nodes can branch to the
513 // head of the loop, and the root cannot have predecessors
514 BasicBlock *newFuncRoot = new BasicBlock("newFuncRoot");
515 newFuncRoot->getInstList().push_back(new BranchInst(header));
517 // Find inputs to, outputs from the code region
519 // If one of the inputs is coming from a different basic block and it's in a
520 // phi node, we need to rewrite the phi node:
522 // * All the inputs which involve basic blocks OUTSIDE of this region go into
523 // a NEW phi node that takes care of finding which value really came in.
524 // The result of this phi is passed to the function as an argument.
526 // * All the other phi values stay.
528 // FIXME: PHI nodes' incoming blocks aren't being rewritten to accomodate for
529 // blocks moving to a new function.
530 // SOLUTION: move Phi nodes out of the loop header into the codeReplacer, pass
531 // the values as parameters to the function
532 findInputsOutputs(inputs, outputs, codeReplacer, newFuncRoot);
534 // Step 2: Construct new function based on inputs/outputs,
535 // Add allocas for all defs
536 Function *newFunction = constructFunction(inputs, outputs, code[0],
538 codeReplacer, oldFunction,
539 oldFunction->getParent());
541 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
543 moveCodeToFunction(newFunction);
545 // Loop over all of the PHI nodes in the entry block (code[0]), and change any
546 // references to the old incoming edge to be the new incoming edge.
547 for (BasicBlock::iterator I = code[0]->begin();
548 PHINode *PN = dyn_cast<PHINode>(I); ++I)
549 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
550 if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
551 PN->setIncomingBlock(i, newFuncRoot);
553 // Look at all successors of the codeReplacer block. If any of these blocks
554 // had PHI nodes in them, we need to update the "from" block to be the code
555 // replacer, not the original block in the extracted region.
556 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
557 succ_end(codeReplacer));
558 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
559 for (BasicBlock::iterator I = Succs[i]->begin();
560 PHINode *PN = dyn_cast<PHINode>(I); ++I)
561 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
562 if (BlocksToExtract.count(PN->getIncomingBlock(i)))
563 PN->setIncomingBlock(i, codeReplacer);
566 DEBUG(if (verifyFunction(*newFunction)) abort());
570 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
571 // Deny code region if it contains allocas
572 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
574 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
576 if (isa<AllocaInst>(*I))
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);