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/CodeExtractor.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/Analysis/BlockFrequencyInfo.h"
21 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
22 #include "llvm/Analysis/BranchProbabilityInfo.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/RegionInfo.h"
25 #include "llvm/Analysis/RegionIterator.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/IR/MDBuilder.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/Verifier.h"
35 #include "llvm/Pass.h"
36 #include "llvm/Support/BlockFrequency.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #define DEBUG_TYPE "code-extractor"
48 // Provide a command-line option to aggregate function arguments into a struct
49 // for functions produced by the code extractor. This is useful when converting
50 // extracted functions to pthread-based code, as only one argument (void*) can
51 // be passed in to pthread_create().
53 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
54 cl::desc("Aggregate arguments to code-extracted functions"));
56 /// \brief Test whether a block is valid for extraction.
57 bool CodeExtractor::isBlockValidForExtraction(const BasicBlock &BB) {
58 // Landing pads must be in the function where they were inserted for cleanup.
62 // Don't hoist code containing allocas, invokes, or vastarts.
63 for (BasicBlock::const_iterator I = BB.begin(), E = BB.end(); I != E; ++I) {
64 if (isa<AllocaInst>(I) || isa<InvokeInst>(I))
66 if (const CallInst *CI = dyn_cast<CallInst>(I))
67 if (const Function *F = CI->getCalledFunction())
68 if (F->getIntrinsicID() == Intrinsic::vastart)
75 /// \brief Build a set of blocks to extract if the input blocks are viable.
76 template <typename IteratorT>
77 static SetVector<BasicBlock *> buildExtractionBlockSet(IteratorT BBBegin,
79 SetVector<BasicBlock *> Result;
81 assert(BBBegin != BBEnd);
83 // Loop over the blocks, adding them to our set-vector, and aborting with an
84 // empty set if we encounter invalid blocks.
86 if (!Result.insert(*BBBegin))
87 llvm_unreachable("Repeated basic blocks in extraction input");
89 if (!CodeExtractor::isBlockValidForExtraction(**BBBegin)) {
93 } while (++BBBegin != BBEnd);
96 for (SetVector<BasicBlock *>::iterator I = std::next(Result.begin()),
99 for (pred_iterator PI = pred_begin(*I), PE = pred_end(*I);
101 assert(Result.count(*PI) &&
102 "No blocks in this region may have entries from outside the region"
103 " except for the first block!");
109 /// \brief Helper to call buildExtractionBlockSet with an ArrayRef.
110 static SetVector<BasicBlock *>
111 buildExtractionBlockSet(ArrayRef<BasicBlock *> BBs) {
112 return buildExtractionBlockSet(BBs.begin(), BBs.end());
115 /// \brief Helper to call buildExtractionBlockSet with a RegionNode.
116 static SetVector<BasicBlock *>
117 buildExtractionBlockSet(const RegionNode &RN) {
118 if (!RN.isSubRegion())
119 // Just a single BasicBlock.
120 return buildExtractionBlockSet(RN.getNodeAs<BasicBlock>());
122 const Region &R = *RN.getNodeAs<Region>();
124 return buildExtractionBlockSet(R.block_begin(), R.block_end());
127 CodeExtractor::CodeExtractor(BasicBlock *BB, bool AggregateArgs,
128 BlockFrequencyInfo *BFI,
129 BranchProbabilityInfo *BPI)
130 : DT(nullptr), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI),
131 BPI(BPI), Blocks(buildExtractionBlockSet(BB)), NumExitBlocks(~0U) {}
133 CodeExtractor::CodeExtractor(ArrayRef<BasicBlock *> BBs, DominatorTree *DT,
134 bool AggregateArgs, BlockFrequencyInfo *BFI,
135 BranchProbabilityInfo *BPI)
136 : DT(DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI),
137 BPI(BPI), Blocks(buildExtractionBlockSet(BBs)), NumExitBlocks(~0U) {}
139 CodeExtractor::CodeExtractor(DominatorTree &DT, Loop &L, bool AggregateArgs,
140 BlockFrequencyInfo *BFI,
141 BranchProbabilityInfo *BPI)
142 : DT(&DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI),
143 BPI(BPI), Blocks(buildExtractionBlockSet(L.getBlocks())),
144 NumExitBlocks(~0U) {}
146 CodeExtractor::CodeExtractor(DominatorTree &DT, const RegionNode &RN,
147 bool AggregateArgs, BlockFrequencyInfo *BFI,
148 BranchProbabilityInfo *BPI)
149 : DT(&DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI),
150 BPI(BPI), Blocks(buildExtractionBlockSet(RN)), NumExitBlocks(~0U) {}
152 /// definedInRegion - Return true if the specified value is defined in the
153 /// extracted region.
154 static bool definedInRegion(const SetVector<BasicBlock *> &Blocks, Value *V) {
155 if (Instruction *I = dyn_cast<Instruction>(V))
156 if (Blocks.count(I->getParent()))
161 /// definedInCaller - Return true if the specified value is defined in the
162 /// function being code extracted, but not in the region being extracted.
163 /// These values must be passed in as live-ins to the function.
164 static bool definedInCaller(const SetVector<BasicBlock *> &Blocks, Value *V) {
165 if (isa<Argument>(V)) return true;
166 if (Instruction *I = dyn_cast<Instruction>(V))
167 if (!Blocks.count(I->getParent()))
172 void CodeExtractor::findInputsOutputs(ValueSet &Inputs,
173 ValueSet &Outputs) const {
174 for (BasicBlock *BB : Blocks) {
175 // If a used value is defined outside the region, it's an input. If an
176 // instruction is used outside the region, it's an output.
177 for (Instruction &II : *BB) {
178 for (User::op_iterator OI = II.op_begin(), OE = II.op_end(); OI != OE;
180 if (definedInCaller(Blocks, *OI))
183 for (User *U : II.users())
184 if (!definedInRegion(Blocks, U)) {
192 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
193 /// region, we need to split the entry block of the region so that the PHI node
194 /// is easier to deal with.
195 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
196 unsigned NumPredsFromRegion = 0;
197 unsigned NumPredsOutsideRegion = 0;
199 if (Header != &Header->getParent()->getEntryBlock()) {
200 PHINode *PN = dyn_cast<PHINode>(Header->begin());
201 if (!PN) return; // No PHI nodes.
203 // If the header node contains any PHI nodes, check to see if there is more
204 // than one entry from outside the region. If so, we need to sever the
205 // header block into two.
206 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
207 if (Blocks.count(PN->getIncomingBlock(i)))
208 ++NumPredsFromRegion;
210 ++NumPredsOutsideRegion;
212 // If there is one (or fewer) predecessor from outside the region, we don't
213 // need to do anything special.
214 if (NumPredsOutsideRegion <= 1) return;
217 // Otherwise, we need to split the header block into two pieces: one
218 // containing PHI nodes merging values from outside of the region, and a
219 // second that contains all of the code for the block and merges back any
220 // incoming values from inside of the region.
221 BasicBlock::iterator AfterPHIs = Header->getFirstNonPHI()->getIterator();
222 BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
223 Header->getName()+".ce");
225 // We only want to code extract the second block now, and it becomes the new
226 // header of the region.
227 BasicBlock *OldPred = Header;
228 Blocks.remove(OldPred);
229 Blocks.insert(NewBB);
232 // Okay, update dominator sets. The blocks that dominate the new one are the
233 // blocks that dominate TIBB plus the new block itself.
235 DT->splitBlock(NewBB);
237 // Okay, now we need to adjust the PHI nodes and any branches from within the
238 // region to go to the new header block instead of the old header block.
239 if (NumPredsFromRegion) {
240 PHINode *PN = cast<PHINode>(OldPred->begin());
241 // Loop over all of the predecessors of OldPred that are in the region,
242 // changing them to branch to NewBB instead.
243 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
244 if (Blocks.count(PN->getIncomingBlock(i))) {
245 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
246 TI->replaceUsesOfWith(OldPred, NewBB);
249 // Okay, everything within the region is now branching to the right block, we
250 // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
251 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
252 PHINode *PN = cast<PHINode>(AfterPHIs);
253 // Create a new PHI node in the new region, which has an incoming value
254 // from OldPred of PN.
255 PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion,
256 PN->getName() + ".ce", &NewBB->front());
257 NewPN->addIncoming(PN, OldPred);
259 // Loop over all of the incoming value in PN, moving them to NewPN if they
260 // are from the extracted region.
261 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
262 if (Blocks.count(PN->getIncomingBlock(i))) {
263 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
264 PN->removeIncomingValue(i);
272 void CodeExtractor::splitReturnBlocks() {
273 for (BasicBlock *Block : Blocks)
274 if (ReturnInst *RI = dyn_cast<ReturnInst>(Block->getTerminator())) {
276 Block->splitBasicBlock(RI->getIterator(), Block->getName() + ".ret");
278 // Old dominates New. New node dominates all other nodes dominated
280 DomTreeNode *OldNode = DT->getNode(Block);
281 SmallVector<DomTreeNode *, 8> Children(OldNode->begin(),
284 DomTreeNode *NewNode = DT->addNewBlock(New, Block);
286 for (DomTreeNode *I : Children)
287 DT->changeImmediateDominator(I, NewNode);
292 /// constructFunction - make a function based on inputs and outputs, as follows:
293 /// f(in0, ..., inN, out0, ..., outN)
295 Function *CodeExtractor::constructFunction(const ValueSet &inputs,
296 const ValueSet &outputs,
298 BasicBlock *newRootNode,
299 BasicBlock *newHeader,
300 Function *oldFunction,
302 DEBUG(dbgs() << "inputs: " << inputs.size() << "\n");
303 DEBUG(dbgs() << "outputs: " << outputs.size() << "\n");
305 // This function returns unsigned, outputs will go back by reference.
306 switch (NumExitBlocks) {
308 case 1: RetTy = Type::getVoidTy(header->getContext()); break;
309 case 2: RetTy = Type::getInt1Ty(header->getContext()); break;
310 default: RetTy = Type::getInt16Ty(header->getContext()); break;
313 std::vector<Type*> paramTy;
315 // Add the types of the input values to the function's argument list
316 for (Value *value : inputs) {
317 DEBUG(dbgs() << "value used in func: " << *value << "\n");
318 paramTy.push_back(value->getType());
321 // Add the types of the output values to the function's argument list.
322 for (Value *output : outputs) {
323 DEBUG(dbgs() << "instr used in func: " << *output << "\n");
325 paramTy.push_back(output->getType());
327 paramTy.push_back(PointerType::getUnqual(output->getType()));
331 dbgs() << "Function type: " << *RetTy << " f(";
332 for (Type *i : paramTy)
333 dbgs() << *i << ", ";
337 StructType *StructTy;
338 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
339 StructTy = StructType::get(M->getContext(), paramTy);
341 paramTy.push_back(PointerType::getUnqual(StructTy));
343 FunctionType *funcType =
344 FunctionType::get(RetTy, paramTy, false);
346 // Create the new function
347 Function *newFunction = Function::Create(funcType,
348 GlobalValue::InternalLinkage,
349 oldFunction->getName() + "_" +
350 header->getName(), M);
351 // If the old function is no-throw, so is the new one.
352 if (oldFunction->doesNotThrow())
353 newFunction->setDoesNotThrow();
355 // Inherit the uwtable attribute if we need to.
356 if (oldFunction->hasUWTable())
357 newFunction->setHasUWTable();
359 // Inherit all of the target dependent attributes.
360 // (e.g. If the extracted region contains a call to an x86.sse
361 // instruction we need to make sure that the extracted region has the
362 // "target-features" attribute allowing it to be lowered.
363 // FIXME: This should be changed to check to see if a specific
364 // attribute can not be inherited.
365 AttrBuilder AB(oldFunction->getAttributes().getFnAttributes());
366 for (const auto &Attr : AB.td_attrs())
367 newFunction->addFnAttr(Attr.first, Attr.second);
369 newFunction->getBasicBlockList().push_back(newRootNode);
371 // Create an iterator to name all of the arguments we inserted.
372 Function::arg_iterator AI = newFunction->arg_begin();
374 // Rewrite all users of the inputs in the extracted region to use the
375 // arguments (or appropriate addressing into struct) instead.
376 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
380 Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext()));
381 Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i);
382 TerminatorInst *TI = newFunction->begin()->getTerminator();
383 GetElementPtrInst *GEP = GetElementPtrInst::Create(
384 StructTy, &*AI, Idx, "gep_" + inputs[i]->getName(), TI);
385 RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI);
389 std::vector<User*> Users(inputs[i]->user_begin(), inputs[i]->user_end());
390 for (User *use : Users)
391 if (Instruction *inst = dyn_cast<Instruction>(use))
392 if (Blocks.count(inst->getParent()))
393 inst->replaceUsesOfWith(inputs[i], RewriteVal);
396 // Set names for input and output arguments.
397 if (!AggregateArgs) {
398 AI = newFunction->arg_begin();
399 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
400 AI->setName(inputs[i]->getName());
401 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
402 AI->setName(outputs[i]->getName()+".out");
405 // Rewrite branches to basic blocks outside of the loop to new dummy blocks
406 // within the new function. This must be done before we lose track of which
407 // blocks were originally in the code region.
408 std::vector<User*> Users(header->user_begin(), header->user_end());
409 for (unsigned i = 0, e = Users.size(); i != e; ++i)
410 // The BasicBlock which contains the branch is not in the region
411 // modify the branch target to a new block
412 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
413 if (!Blocks.count(TI->getParent()) &&
414 TI->getParent()->getParent() == oldFunction)
415 TI->replaceUsesOfWith(header, newHeader);
420 /// FindPhiPredForUseInBlock - Given a value and a basic block, find a PHI
421 /// that uses the value within the basic block, and return the predecessor
422 /// block associated with that use, or return 0 if none is found.
423 static BasicBlock* FindPhiPredForUseInBlock(Value* Used, BasicBlock* BB) {
424 for (Use &U : Used->uses()) {
425 PHINode *P = dyn_cast<PHINode>(U.getUser());
426 if (P && P->getParent() == BB)
427 return P->getIncomingBlock(U);
433 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
434 /// the call instruction, splitting any PHI nodes in the header block as
437 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
438 ValueSet &inputs, ValueSet &outputs) {
439 // Emit a call to the new function, passing in: *pointer to struct (if
440 // aggregating parameters), or plan inputs and allocated memory for outputs
441 std::vector<Value*> params, StructValues, ReloadOutputs, Reloads;
443 Module *M = newFunction->getParent();
444 LLVMContext &Context = M->getContext();
445 const DataLayout &DL = M->getDataLayout();
447 // Add inputs as params, or to be filled into the struct
448 for (Value *input : inputs)
450 StructValues.push_back(input);
452 params.push_back(input);
454 // Create allocas for the outputs
455 for (Value *output : outputs) {
457 StructValues.push_back(output);
460 new AllocaInst(output->getType(), DL.getAllocaAddrSpace(),
461 nullptr, output->getName() + ".loc",
462 &codeReplacer->getParent()->front().front());
463 ReloadOutputs.push_back(alloca);
464 params.push_back(alloca);
468 StructType *StructArgTy = nullptr;
469 AllocaInst *Struct = nullptr;
470 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
471 std::vector<Type*> ArgTypes;
472 for (ValueSet::iterator v = StructValues.begin(),
473 ve = StructValues.end(); v != ve; ++v)
474 ArgTypes.push_back((*v)->getType());
476 // Allocate a struct at the beginning of this function
477 StructArgTy = StructType::get(newFunction->getContext(), ArgTypes);
478 Struct = new AllocaInst(StructArgTy, DL.getAllocaAddrSpace(), nullptr,
480 &codeReplacer->getParent()->front().front());
481 params.push_back(Struct);
483 for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
485 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
486 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i);
487 GetElementPtrInst *GEP = GetElementPtrInst::Create(
488 StructArgTy, Struct, Idx, "gep_" + StructValues[i]->getName());
489 codeReplacer->getInstList().push_back(GEP);
490 StoreInst *SI = new StoreInst(StructValues[i], GEP);
491 codeReplacer->getInstList().push_back(SI);
495 // Emit the call to the function
496 CallInst *call = CallInst::Create(newFunction, params,
497 NumExitBlocks > 1 ? "targetBlock" : "");
498 codeReplacer->getInstList().push_back(call);
500 Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
501 unsigned FirstOut = inputs.size();
503 std::advance(OutputArgBegin, inputs.size());
505 // Reload the outputs passed in by reference
506 for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
507 Value *Output = nullptr;
510 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
511 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i);
512 GetElementPtrInst *GEP = GetElementPtrInst::Create(
513 StructArgTy, Struct, Idx, "gep_reload_" + outputs[i]->getName());
514 codeReplacer->getInstList().push_back(GEP);
517 Output = ReloadOutputs[i];
519 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
520 Reloads.push_back(load);
521 codeReplacer->getInstList().push_back(load);
522 std::vector<User*> Users(outputs[i]->user_begin(), outputs[i]->user_end());
523 for (unsigned u = 0, e = Users.size(); u != e; ++u) {
524 Instruction *inst = cast<Instruction>(Users[u]);
525 if (!Blocks.count(inst->getParent()))
526 inst->replaceUsesOfWith(outputs[i], load);
530 // Now we can emit a switch statement using the call as a value.
531 SwitchInst *TheSwitch =
532 SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)),
533 codeReplacer, 0, codeReplacer);
535 // Since there may be multiple exits from the original region, make the new
536 // function return an unsigned, switch on that number. This loop iterates
537 // over all of the blocks in the extracted region, updating any terminator
538 // instructions in the to-be-extracted region that branch to blocks that are
539 // not in the region to be extracted.
540 std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
542 unsigned switchVal = 0;
543 for (BasicBlock *Block : Blocks) {
544 TerminatorInst *TI = Block->getTerminator();
545 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
546 if (!Blocks.count(TI->getSuccessor(i))) {
547 BasicBlock *OldTarget = TI->getSuccessor(i);
548 // add a new basic block which returns the appropriate value
549 BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
551 // If we don't already have an exit stub for this non-extracted
552 // destination, create one now!
553 NewTarget = BasicBlock::Create(Context,
554 OldTarget->getName() + ".exitStub",
556 unsigned SuccNum = switchVal++;
558 Value *brVal = nullptr;
559 switch (NumExitBlocks) {
561 case 1: break; // No value needed.
562 case 2: // Conditional branch, return a bool
563 brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum);
566 brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum);
570 ReturnInst *NTRet = ReturnInst::Create(Context, brVal, NewTarget);
572 // Update the switch instruction.
573 TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context),
577 // Restore values just before we exit
578 Function::arg_iterator OAI = OutputArgBegin;
579 for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
580 // For an invoke, the normal destination is the only one that is
581 // dominated by the result of the invocation
582 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
584 bool DominatesDef = true;
586 BasicBlock *NormalDest = nullptr;
587 if (auto *Invoke = dyn_cast<InvokeInst>(outputs[out]))
588 NormalDest = Invoke->getNormalDest();
591 DefBlock = NormalDest;
593 // Make sure we are looking at the original successor block, not
594 // at a newly inserted exit block, which won't be in the dominator
596 for (const auto &I : ExitBlockMap)
597 if (DefBlock == I.second) {
602 // In the extract block case, if the block we are extracting ends
603 // with an invoke instruction, make sure that we don't emit a
604 // store of the invoke value for the unwind block.
605 if (!DT && DefBlock != OldTarget)
606 DominatesDef = false;
610 DominatesDef = DT->dominates(DefBlock, OldTarget);
612 // If the output value is used by a phi in the target block,
613 // then we need to test for dominance of the phi's predecessor
614 // instead. Unfortunately, this a little complicated since we
615 // have already rewritten uses of the value to uses of the reload.
616 BasicBlock* pred = FindPhiPredForUseInBlock(Reloads[out],
618 if (pred && DT && DT->dominates(DefBlock, pred))
625 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
626 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context),
628 GetElementPtrInst *GEP = GetElementPtrInst::Create(
629 StructArgTy, &*OAI, Idx, "gep_" + outputs[out]->getName(),
631 new StoreInst(outputs[out], GEP, NTRet);
633 new StoreInst(outputs[out], &*OAI, NTRet);
636 // Advance output iterator even if we don't emit a store
637 if (!AggregateArgs) ++OAI;
641 // rewrite the original branch instruction with this new target
642 TI->setSuccessor(i, NewTarget);
646 // Now that we've done the deed, simplify the switch instruction.
647 Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
648 switch (NumExitBlocks) {
650 // There are no successors (the block containing the switch itself), which
651 // means that previously this was the last part of the function, and hence
652 // this should be rewritten as a `ret'
654 // Check if the function should return a value
655 if (OldFnRetTy->isVoidTy()) {
656 ReturnInst::Create(Context, nullptr, TheSwitch); // Return void
657 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
658 // return what we have
659 ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch);
661 // Otherwise we must have code extracted an unwind or something, just
662 // return whatever we want.
663 ReturnInst::Create(Context,
664 Constant::getNullValue(OldFnRetTy), TheSwitch);
667 TheSwitch->eraseFromParent();
670 // Only a single destination, change the switch into an unconditional
672 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
673 TheSwitch->eraseFromParent();
676 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
678 TheSwitch->eraseFromParent();
681 // Otherwise, make the default destination of the switch instruction be one
682 // of the other successors.
683 TheSwitch->setCondition(call);
684 TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks));
685 // Remove redundant case
686 TheSwitch->removeCase(SwitchInst::CaseIt(TheSwitch, NumExitBlocks-1));
691 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
692 Function *oldFunc = (*Blocks.begin())->getParent();
693 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
694 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
696 for (BasicBlock *Block : Blocks) {
697 // Delete the basic block from the old function, and the list of blocks
698 oldBlocks.remove(Block);
700 // Insert this basic block into the new function
701 newBlocks.push_back(Block);
705 void CodeExtractor::calculateNewCallTerminatorWeights(
706 BasicBlock *CodeReplacer,
707 DenseMap<BasicBlock *, BlockFrequency> &ExitWeights,
708 BranchProbabilityInfo *BPI) {
709 typedef BlockFrequencyInfoImplBase::Distribution Distribution;
710 typedef BlockFrequencyInfoImplBase::BlockNode BlockNode;
712 // Update the branch weights for the exit block.
713 TerminatorInst *TI = CodeReplacer->getTerminator();
714 SmallVector<unsigned, 8> BranchWeights(TI->getNumSuccessors(), 0);
716 // Block Frequency distribution with dummy node.
717 Distribution BranchDist;
719 // Add each of the frequencies of the successors.
720 for (unsigned i = 0, e = TI->getNumSuccessors(); i < e; ++i) {
721 BlockNode ExitNode(i);
722 uint64_t ExitFreq = ExitWeights[TI->getSuccessor(i)].getFrequency();
724 BranchDist.addExit(ExitNode, ExitFreq);
726 BPI->setEdgeProbability(CodeReplacer, i, BranchProbability::getZero());
729 // Check for no total weight.
730 if (BranchDist.Total == 0)
733 // Normalize the distribution so that they can fit in unsigned.
734 BranchDist.normalize();
736 // Create normalized branch weights and set the metadata.
737 for (unsigned I = 0, E = BranchDist.Weights.size(); I < E; ++I) {
738 const auto &Weight = BranchDist.Weights[I];
740 // Get the weight and update the current BFI.
741 BranchWeights[Weight.TargetNode.Index] = Weight.Amount;
742 BranchProbability BP(Weight.Amount, BranchDist.Total);
743 BPI->setEdgeProbability(CodeReplacer, Weight.TargetNode.Index, BP);
746 LLVMContext::MD_prof,
747 MDBuilder(TI->getContext()).createBranchWeights(BranchWeights));
750 Function *CodeExtractor::extractCodeRegion() {
754 ValueSet inputs, outputs;
756 // Assumption: this is a single-entry code region, and the header is the first
757 // block in the region.
758 BasicBlock *header = *Blocks.begin();
760 // Calculate the entry frequency of the new function before we change the root
762 BlockFrequency EntryFreq;
764 assert(BPI && "Both BPI and BFI are required to preserve profile info");
765 for (BasicBlock *Pred : predecessors(header)) {
766 if (Blocks.count(Pred))
769 BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, header);
773 // If we have to split PHI nodes or the entry block, do so now.
774 severSplitPHINodes(header);
776 // If we have any return instructions in the region, split those blocks so
777 // that the return is not in the region.
780 Function *oldFunction = header->getParent();
782 // This takes place of the original loop
783 BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(),
784 "codeRepl", oldFunction,
787 // The new function needs a root node because other nodes can branch to the
788 // head of the region, but the entry node of a function cannot have preds.
789 BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(),
791 newFuncRoot->getInstList().push_back(BranchInst::Create(header));
793 // Find inputs to, outputs from the code region.
794 findInputsOutputs(inputs, outputs);
796 // Calculate the exit blocks for the extracted region and the total exit
797 // weights for each of those blocks.
798 DenseMap<BasicBlock *, BlockFrequency> ExitWeights;
799 SmallPtrSet<BasicBlock *, 1> ExitBlocks;
800 for (BasicBlock *Block : Blocks) {
801 for (succ_iterator SI = succ_begin(Block), SE = succ_end(Block); SI != SE;
803 if (!Blocks.count(*SI)) {
804 // Update the branch weight for this successor.
806 BlockFrequency &BF = ExitWeights[*SI];
807 BF += BFI->getBlockFreq(Block) * BPI->getEdgeProbability(Block, *SI);
809 ExitBlocks.insert(*SI);
813 NumExitBlocks = ExitBlocks.size();
815 // Construct new function based on inputs/outputs & add allocas for all defs.
816 Function *newFunction = constructFunction(inputs, outputs, header,
818 codeReplacer, oldFunction,
819 oldFunction->getParent());
821 // Update the entry count of the function.
823 Optional<uint64_t> EntryCount =
824 BFI->getProfileCountFromFreq(EntryFreq.getFrequency());
825 if (EntryCount.hasValue())
826 newFunction->setEntryCount(EntryCount.getValue());
827 BFI->setBlockFreq(codeReplacer, EntryFreq.getFrequency());
830 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
832 moveCodeToFunction(newFunction);
834 // Update the branch weights for the exit block.
835 if (BFI && NumExitBlocks > 1)
836 calculateNewCallTerminatorWeights(codeReplacer, ExitWeights, BPI);
838 // Loop over all of the PHI nodes in the header block, and change any
839 // references to the old incoming edge to be the new incoming edge.
840 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
841 PHINode *PN = cast<PHINode>(I);
842 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
843 if (!Blocks.count(PN->getIncomingBlock(i)))
844 PN->setIncomingBlock(i, newFuncRoot);
847 // Look at all successors of the codeReplacer block. If any of these blocks
848 // had PHI nodes in them, we need to update the "from" block to be the code
849 // replacer, not the original block in the extracted region.
850 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
851 succ_end(codeReplacer));
852 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
853 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
854 PHINode *PN = cast<PHINode>(I);
855 std::set<BasicBlock*> ProcessedPreds;
856 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
857 if (Blocks.count(PN->getIncomingBlock(i))) {
858 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
859 PN->setIncomingBlock(i, codeReplacer);
861 // There were multiple entries in the PHI for this block, now there
862 // is only one, so remove the duplicated entries.
863 PN->removeIncomingValue(i, false);
869 //cerr << "NEW FUNCTION: " << *newFunction;
870 // verifyFunction(*newFunction);
872 // cerr << "OLD FUNCTION: " << *oldFunction;
873 // verifyFunction(*oldFunction);
875 DEBUG(if (verifyFunction(*newFunction))
876 report_fatal_error("verifyFunction failed!"));