1 //===- CloneFunction.cpp - Clone a function into another 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 CloneFunctionInto interface, which is used as the
11 // low-level function cloner. This is used by the CloneFunction and function
12 // inliner to do the dirty work of copying the body of a function around.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Transforms/Utils/Cloning.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/Analysis/ConstantFolding.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/DebugInfo.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/DerivedTypes.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/GlobalVariable.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/LLVMContext.h"
28 #include "llvm/IR/Metadata.h"
29 #include "llvm/Support/CFG.h"
30 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Transforms/Utils/ValueMapper.h"
36 // CloneBasicBlock - See comments in Cloning.h
37 BasicBlock *llvm::CloneBasicBlock(const BasicBlock *BB,
38 ValueToValueMapTy &VMap,
39 const Twine &NameSuffix, Function *F,
40 ClonedCodeInfo *CodeInfo) {
41 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "", F);
42 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
44 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
46 // Loop over all instructions, and copy them over.
47 for (BasicBlock::const_iterator II = BB->begin(), IE = BB->end();
49 Instruction *NewInst = II->clone();
51 NewInst->setName(II->getName()+NameSuffix);
52 NewBB->getInstList().push_back(NewInst);
53 VMap[II] = NewInst; // Add instruction map to value.
55 hasCalls |= (isa<CallInst>(II) && !isa<DbgInfoIntrinsic>(II));
56 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
57 if (isa<ConstantInt>(AI->getArraySize()))
58 hasStaticAllocas = true;
60 hasDynamicAllocas = true;
65 CodeInfo->ContainsCalls |= hasCalls;
66 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
67 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
68 BB != &BB->getParent()->getEntryBlock();
73 // Clone OldFunc into NewFunc, transforming the old arguments into references to
76 void llvm::CloneFunctionInto(Function *NewFunc, const Function *OldFunc,
77 ValueToValueMapTy &VMap,
78 bool ModuleLevelChanges,
79 SmallVectorImpl<ReturnInst*> &Returns,
80 const char *NameSuffix, ClonedCodeInfo *CodeInfo,
81 ValueMapTypeRemapper *TypeMapper,
82 ValueMaterializer *Materializer) {
83 assert(NameSuffix && "NameSuffix cannot be null!");
86 for (Function::const_arg_iterator I = OldFunc->arg_begin(),
87 E = OldFunc->arg_end(); I != E; ++I)
88 assert(VMap.count(I) && "No mapping from source argument specified!");
91 AttributeSet OldAttrs = OldFunc->getAttributes();
92 // Clone any argument attributes that are present in the VMap.
93 for (Function::const_arg_iterator I = OldFunc->arg_begin(),
94 E = OldFunc->arg_end();
96 if (Argument *Anew = dyn_cast<Argument>(VMap[I])) {
98 OldAttrs.getParamAttributes(I->getArgNo() + 1);
99 if (attrs.getNumSlots() > 0)
100 Anew->addAttr(attrs);
103 NewFunc->setAttributes(NewFunc->getAttributes()
104 .addAttributes(NewFunc->getContext(),
105 AttributeSet::ReturnIndex,
106 OldAttrs.getRetAttributes()));
107 NewFunc->setAttributes(NewFunc->getAttributes()
108 .addAttributes(NewFunc->getContext(),
109 AttributeSet::FunctionIndex,
110 OldAttrs.getFnAttributes()));
112 // Loop over all of the basic blocks in the function, cloning them as
113 // appropriate. Note that we save BE this way in order to handle cloning of
114 // recursive functions into themselves.
116 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
118 const BasicBlock &BB = *BI;
120 // Create a new basic block and copy instructions into it!
121 BasicBlock *CBB = CloneBasicBlock(&BB, VMap, NameSuffix, NewFunc, CodeInfo);
123 // Add basic block mapping.
126 // It is only legal to clone a function if a block address within that
127 // function is never referenced outside of the function. Given that, we
128 // want to map block addresses from the old function to block addresses in
129 // the clone. (This is different from the generic ValueMapper
130 // implementation, which generates an invalid blockaddress when
131 // cloning a function.)
132 if (BB.hasAddressTaken()) {
133 Constant *OldBBAddr = BlockAddress::get(const_cast<Function*>(OldFunc),
134 const_cast<BasicBlock*>(&BB));
135 VMap[OldBBAddr] = BlockAddress::get(NewFunc, CBB);
138 // Note return instructions for the caller.
139 if (ReturnInst *RI = dyn_cast<ReturnInst>(CBB->getTerminator()))
140 Returns.push_back(RI);
143 // Loop over all of the instructions in the function, fixing up operand
144 // references as we go. This uses VMap to do all the hard work.
145 for (Function::iterator BB = cast<BasicBlock>(VMap[OldFunc->begin()]),
146 BE = NewFunc->end(); BB != BE; ++BB)
147 // Loop over all instructions, fixing each one as we find it...
148 for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II)
149 RemapInstruction(II, VMap,
150 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges,
151 TypeMapper, Materializer);
154 /// CloneFunction - Return a copy of the specified function, but without
155 /// embedding the function into another module. Also, any references specified
156 /// in the VMap are changed to refer to their mapped value instead of the
157 /// original one. If any of the arguments to the function are in the VMap,
158 /// the arguments are deleted from the resultant function. The VMap is
159 /// updated to include mappings from all of the instructions and basicblocks in
160 /// the function from their old to new values.
162 Function *llvm::CloneFunction(const Function *F, ValueToValueMapTy &VMap,
163 bool ModuleLevelChanges,
164 ClonedCodeInfo *CodeInfo) {
165 std::vector<Type*> ArgTypes;
167 // The user might be deleting arguments to the function by specifying them in
168 // the VMap. If so, we need to not add the arguments to the arg ty vector
170 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
172 if (VMap.count(I) == 0) // Haven't mapped the argument to anything yet?
173 ArgTypes.push_back(I->getType());
175 // Create a new function type...
176 FunctionType *FTy = FunctionType::get(F->getFunctionType()->getReturnType(),
177 ArgTypes, F->getFunctionType()->isVarArg());
179 // Create the new function...
180 Function *NewF = Function::Create(FTy, F->getLinkage(), F->getName());
182 // Loop over the arguments, copying the names of the mapped arguments over...
183 Function::arg_iterator DestI = NewF->arg_begin();
184 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
186 if (VMap.count(I) == 0) { // Is this argument preserved?
187 DestI->setName(I->getName()); // Copy the name over...
188 VMap[I] = DestI++; // Add mapping to VMap
191 SmallVector<ReturnInst*, 8> Returns; // Ignore returns cloned.
192 CloneFunctionInto(NewF, F, VMap, ModuleLevelChanges, Returns, "", CodeInfo);
199 /// PruningFunctionCloner - This class is a private class used to implement
200 /// the CloneAndPruneFunctionInto method.
201 struct PruningFunctionCloner {
203 const Function *OldFunc;
204 ValueToValueMapTy &VMap;
205 bool ModuleLevelChanges;
206 const char *NameSuffix;
207 ClonedCodeInfo *CodeInfo;
208 const DataLayout *TD;
210 PruningFunctionCloner(Function *newFunc, const Function *oldFunc,
211 ValueToValueMapTy &valueMap,
212 bool moduleLevelChanges,
213 const char *nameSuffix,
214 ClonedCodeInfo *codeInfo,
215 const DataLayout *td)
216 : NewFunc(newFunc), OldFunc(oldFunc),
217 VMap(valueMap), ModuleLevelChanges(moduleLevelChanges),
218 NameSuffix(nameSuffix), CodeInfo(codeInfo), TD(td) {
221 /// CloneBlock - The specified block is found to be reachable, clone it and
222 /// anything that it can reach.
223 void CloneBlock(const BasicBlock *BB,
224 std::vector<const BasicBlock*> &ToClone);
228 /// CloneBlock - The specified block is found to be reachable, clone it and
229 /// anything that it can reach.
230 void PruningFunctionCloner::CloneBlock(const BasicBlock *BB,
231 std::vector<const BasicBlock*> &ToClone){
232 WeakVH &BBEntry = VMap[BB];
234 // Have we already cloned this block?
237 // Nope, clone it now.
239 BBEntry = NewBB = BasicBlock::Create(BB->getContext());
240 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
242 // It is only legal to clone a function if a block address within that
243 // function is never referenced outside of the function. Given that, we
244 // want to map block addresses from the old function to block addresses in
245 // the clone. (This is different from the generic ValueMapper
246 // implementation, which generates an invalid blockaddress when
247 // cloning a function.)
249 // Note that we don't need to fix the mapping for unreachable blocks;
250 // the default mapping there is safe.
251 if (BB->hasAddressTaken()) {
252 Constant *OldBBAddr = BlockAddress::get(const_cast<Function*>(OldFunc),
253 const_cast<BasicBlock*>(BB));
254 VMap[OldBBAddr] = BlockAddress::get(NewFunc, NewBB);
258 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
260 // Loop over all instructions, and copy them over, DCE'ing as we go. This
261 // loop doesn't include the terminator.
262 for (BasicBlock::const_iterator II = BB->begin(), IE = --BB->end();
264 Instruction *NewInst = II->clone();
266 // Eagerly remap operands to the newly cloned instruction, except for PHI
267 // nodes for which we defer processing until we update the CFG.
268 if (!isa<PHINode>(NewInst)) {
269 RemapInstruction(NewInst, VMap,
270 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges);
272 // If we can simplify this instruction to some other value, simply add
273 // a mapping to that value rather than inserting a new instruction into
275 if (Value *V = SimplifyInstruction(NewInst, TD)) {
276 // On the off-chance that this simplifies to an instruction in the old
277 // function, map it back into the new function.
278 if (Value *MappedV = VMap.lookup(V))
288 NewInst->setName(II->getName()+NameSuffix);
289 VMap[II] = NewInst; // Add instruction map to value.
290 NewBB->getInstList().push_back(NewInst);
291 hasCalls |= (isa<CallInst>(II) && !isa<DbgInfoIntrinsic>(II));
292 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
293 if (isa<ConstantInt>(AI->getArraySize()))
294 hasStaticAllocas = true;
296 hasDynamicAllocas = true;
300 // Finally, clone over the terminator.
301 const TerminatorInst *OldTI = BB->getTerminator();
302 bool TerminatorDone = false;
303 if (const BranchInst *BI = dyn_cast<BranchInst>(OldTI)) {
304 if (BI->isConditional()) {
305 // If the condition was a known constant in the callee...
306 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
307 // Or is a known constant in the caller...
309 Value *V = VMap[BI->getCondition()];
310 Cond = dyn_cast_or_null<ConstantInt>(V);
313 // Constant fold to uncond branch!
315 BasicBlock *Dest = BI->getSuccessor(!Cond->getZExtValue());
316 VMap[OldTI] = BranchInst::Create(Dest, NewBB);
317 ToClone.push_back(Dest);
318 TerminatorDone = true;
321 } else if (const SwitchInst *SI = dyn_cast<SwitchInst>(OldTI)) {
322 // If switching on a value known constant in the caller.
323 ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition());
324 if (Cond == 0) { // Or known constant after constant prop in the callee...
325 Value *V = VMap[SI->getCondition()];
326 Cond = dyn_cast_or_null<ConstantInt>(V);
328 if (Cond) { // Constant fold to uncond branch!
329 SwitchInst::ConstCaseIt Case = SI->findCaseValue(Cond);
330 BasicBlock *Dest = const_cast<BasicBlock*>(Case.getCaseSuccessor());
331 VMap[OldTI] = BranchInst::Create(Dest, NewBB);
332 ToClone.push_back(Dest);
333 TerminatorDone = true;
337 if (!TerminatorDone) {
338 Instruction *NewInst = OldTI->clone();
339 if (OldTI->hasName())
340 NewInst->setName(OldTI->getName()+NameSuffix);
341 NewBB->getInstList().push_back(NewInst);
342 VMap[OldTI] = NewInst; // Add instruction map to value.
344 // Recursively clone any reachable successor blocks.
345 const TerminatorInst *TI = BB->getTerminator();
346 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
347 ToClone.push_back(TI->getSuccessor(i));
351 CodeInfo->ContainsCalls |= hasCalls;
352 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
353 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
354 BB != &BB->getParent()->front();
358 /// CloneAndPruneFunctionInto - This works exactly like CloneFunctionInto,
359 /// except that it does some simple constant prop and DCE on the fly. The
360 /// effect of this is to copy significantly less code in cases where (for
361 /// example) a function call with constant arguments is inlined, and those
362 /// constant arguments cause a significant amount of code in the callee to be
363 /// dead. Since this doesn't produce an exact copy of the input, it can't be
364 /// used for things like CloneFunction or CloneModule.
365 void llvm::CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc,
366 ValueToValueMapTy &VMap,
367 bool ModuleLevelChanges,
368 SmallVectorImpl<ReturnInst*> &Returns,
369 const char *NameSuffix,
370 ClonedCodeInfo *CodeInfo,
371 const DataLayout *TD,
372 Instruction *TheCall) {
373 assert(NameSuffix && "NameSuffix cannot be null!");
376 for (Function::const_arg_iterator II = OldFunc->arg_begin(),
377 E = OldFunc->arg_end(); II != E; ++II)
378 assert(VMap.count(II) && "No mapping from source argument specified!");
381 PruningFunctionCloner PFC(NewFunc, OldFunc, VMap, ModuleLevelChanges,
382 NameSuffix, CodeInfo, TD);
384 // Clone the entry block, and anything recursively reachable from it.
385 std::vector<const BasicBlock*> CloneWorklist;
386 CloneWorklist.push_back(&OldFunc->getEntryBlock());
387 while (!CloneWorklist.empty()) {
388 const BasicBlock *BB = CloneWorklist.back();
389 CloneWorklist.pop_back();
390 PFC.CloneBlock(BB, CloneWorklist);
393 // Loop over all of the basic blocks in the old function. If the block was
394 // reachable, we have cloned it and the old block is now in the value map:
395 // insert it into the new function in the right order. If not, ignore it.
397 // Defer PHI resolution until rest of function is resolved.
398 SmallVector<const PHINode*, 16> PHIToResolve;
399 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
402 BasicBlock *NewBB = cast_or_null<BasicBlock>(V);
403 if (NewBB == 0) continue; // Dead block.
405 // Add the new block to the new function.
406 NewFunc->getBasicBlockList().push_back(NewBB);
408 // Handle PHI nodes specially, as we have to remove references to dead
410 for (BasicBlock::const_iterator I = BI->begin(), E = BI->end(); I != E; ++I)
411 if (const PHINode *PN = dyn_cast<PHINode>(I))
412 PHIToResolve.push_back(PN);
416 // Finally, remap the terminator instructions, as those can't be remapped
417 // until all BBs are mapped.
418 RemapInstruction(NewBB->getTerminator(), VMap,
419 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges);
422 // Defer PHI resolution until rest of function is resolved, PHI resolution
423 // requires the CFG to be up-to-date.
424 for (unsigned phino = 0, e = PHIToResolve.size(); phino != e; ) {
425 const PHINode *OPN = PHIToResolve[phino];
426 unsigned NumPreds = OPN->getNumIncomingValues();
427 const BasicBlock *OldBB = OPN->getParent();
428 BasicBlock *NewBB = cast<BasicBlock>(VMap[OldBB]);
430 // Map operands for blocks that are live and remove operands for blocks
432 for (; phino != PHIToResolve.size() &&
433 PHIToResolve[phino]->getParent() == OldBB; ++phino) {
434 OPN = PHIToResolve[phino];
435 PHINode *PN = cast<PHINode>(VMap[OPN]);
436 for (unsigned pred = 0, e = NumPreds; pred != e; ++pred) {
437 Value *V = VMap[PN->getIncomingBlock(pred)];
438 if (BasicBlock *MappedBlock = cast_or_null<BasicBlock>(V)) {
439 Value *InVal = MapValue(PN->getIncomingValue(pred),
441 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges);
442 assert(InVal && "Unknown input value?");
443 PN->setIncomingValue(pred, InVal);
444 PN->setIncomingBlock(pred, MappedBlock);
446 PN->removeIncomingValue(pred, false);
447 --pred, --e; // Revisit the next entry.
452 // The loop above has removed PHI entries for those blocks that are dead
453 // and has updated others. However, if a block is live (i.e. copied over)
454 // but its terminator has been changed to not go to this block, then our
455 // phi nodes will have invalid entries. Update the PHI nodes in this
457 PHINode *PN = cast<PHINode>(NewBB->begin());
458 NumPreds = std::distance(pred_begin(NewBB), pred_end(NewBB));
459 if (NumPreds != PN->getNumIncomingValues()) {
460 assert(NumPreds < PN->getNumIncomingValues());
461 // Count how many times each predecessor comes to this block.
462 std::map<BasicBlock*, unsigned> PredCount;
463 for (pred_iterator PI = pred_begin(NewBB), E = pred_end(NewBB);
467 // Figure out how many entries to remove from each PHI.
468 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
469 ++PredCount[PN->getIncomingBlock(i)];
471 // At this point, the excess predecessor entries are positive in the
472 // map. Loop over all of the PHIs and remove excess predecessor
474 BasicBlock::iterator I = NewBB->begin();
475 for (; (PN = dyn_cast<PHINode>(I)); ++I) {
476 for (std::map<BasicBlock*, unsigned>::iterator PCI =PredCount.begin(),
477 E = PredCount.end(); PCI != E; ++PCI) {
478 BasicBlock *Pred = PCI->first;
479 for (unsigned NumToRemove = PCI->second; NumToRemove; --NumToRemove)
480 PN->removeIncomingValue(Pred, false);
485 // If the loops above have made these phi nodes have 0 or 1 operand,
486 // replace them with undef or the input value. We must do this for
487 // correctness, because 0-operand phis are not valid.
488 PN = cast<PHINode>(NewBB->begin());
489 if (PN->getNumIncomingValues() == 0) {
490 BasicBlock::iterator I = NewBB->begin();
491 BasicBlock::const_iterator OldI = OldBB->begin();
492 while ((PN = dyn_cast<PHINode>(I++))) {
493 Value *NV = UndefValue::get(PN->getType());
494 PN->replaceAllUsesWith(NV);
495 assert(VMap[OldI] == PN && "VMap mismatch");
497 PN->eraseFromParent();
503 // Make a second pass over the PHINodes now that all of them have been
504 // remapped into the new function, simplifying the PHINode and performing any
505 // recursive simplifications exposed. This will transparently update the
506 // WeakVH in the VMap. Notably, we rely on that so that if we coalesce
507 // two PHINodes, the iteration over the old PHIs remains valid, and the
508 // mapping will just map us to the new node (which may not even be a PHI
510 for (unsigned Idx = 0, Size = PHIToResolve.size(); Idx != Size; ++Idx)
511 if (PHINode *PN = dyn_cast<PHINode>(VMap[PHIToResolve[Idx]]))
512 recursivelySimplifyInstruction(PN, TD);
514 // Now that the inlined function body has been fully constructed, go through
515 // and zap unconditional fall-through branches. This happen all the time when
516 // specializing code: code specialization turns conditional branches into
517 // uncond branches, and this code folds them.
518 Function::iterator Begin = cast<BasicBlock>(VMap[&OldFunc->getEntryBlock()]);
519 Function::iterator I = Begin;
520 while (I != NewFunc->end()) {
521 // Check if this block has become dead during inlining or other
522 // simplifications. Note that the first block will appear dead, as it has
523 // not yet been wired up properly.
524 if (I != Begin && (pred_begin(I) == pred_end(I) ||
525 I->getSinglePredecessor() == I)) {
526 BasicBlock *DeadBB = I++;
527 DeleteDeadBlock(DeadBB);
531 // We need to simplify conditional branches and switches with a constant
532 // operand. We try to prune these out when cloning, but if the
533 // simplification required looking through PHI nodes, those are only
534 // available after forming the full basic block. That may leave some here,
535 // and we still want to prune the dead code as early as possible.
536 ConstantFoldTerminator(I);
538 BranchInst *BI = dyn_cast<BranchInst>(I->getTerminator());
539 if (!BI || BI->isConditional()) { ++I; continue; }
541 BasicBlock *Dest = BI->getSuccessor(0);
542 if (!Dest->getSinglePredecessor()) {
546 // We shouldn't be able to get single-entry PHI nodes here, as instsimplify
547 // above should have zapped all of them..
548 assert(!isa<PHINode>(Dest->begin()));
550 // We know all single-entry PHI nodes in the inlined function have been
551 // removed, so we just need to splice the blocks.
552 BI->eraseFromParent();
554 // Make all PHI nodes that referred to Dest now refer to I as their source.
555 Dest->replaceAllUsesWith(I);
557 // Move all the instructions in the succ to the pred.
558 I->getInstList().splice(I->end(), Dest->getInstList());
560 // Remove the dest block.
561 Dest->eraseFromParent();
563 // Do not increment I, iteratively merge all things this block branches to.
566 // Make a final pass over the basic blocks from theh old function to gather
567 // any return instructions which survived folding. We have to do this here
568 // because we can iteratively remove and merge returns above.
569 for (Function::iterator I = cast<BasicBlock>(VMap[&OldFunc->getEntryBlock()]),
572 if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator()))
573 Returns.push_back(RI);