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/Constants.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/IntrinsicInst.h"
21 #include "llvm/GlobalVariable.h"
22 #include "llvm/Function.h"
23 #include "llvm/LLVMContext.h"
24 #include "llvm/Metadata.h"
25 #include "llvm/Support/CFG.h"
26 #include "llvm/Transforms/Utils/ValueMapper.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/DebugInfo.h"
29 #include "llvm/ADT/SmallVector.h"
33 // CloneBasicBlock - See comments in Cloning.h
34 BasicBlock *llvm::CloneBasicBlock(const BasicBlock *BB,
35 ValueToValueMapTy &VMap,
36 const Twine &NameSuffix, Function *F,
37 ClonedCodeInfo *CodeInfo) {
38 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "", F);
39 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
41 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
43 // Loop over all instructions, and copy them over.
44 for (BasicBlock::const_iterator II = BB->begin(), IE = BB->end();
46 Instruction *NewInst = II->clone();
48 NewInst->setName(II->getName()+NameSuffix);
49 NewBB->getInstList().push_back(NewInst);
50 VMap[II] = NewInst; // Add instruction map to value.
52 hasCalls |= (isa<CallInst>(II) && !isa<DbgInfoIntrinsic>(II));
53 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
54 if (isa<ConstantInt>(AI->getArraySize()))
55 hasStaticAllocas = true;
57 hasDynamicAllocas = true;
62 CodeInfo->ContainsCalls |= hasCalls;
63 CodeInfo->ContainsUnwinds |= isa<UnwindInst>(BB->getTerminator());
64 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
65 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
66 BB != &BB->getParent()->getEntryBlock();
71 // Clone OldFunc into NewFunc, transforming the old arguments into references to
74 void llvm::CloneFunctionInto(Function *NewFunc, const Function *OldFunc,
75 ValueToValueMapTy &VMap,
76 bool ModuleLevelChanges,
77 SmallVectorImpl<ReturnInst*> &Returns,
78 const char *NameSuffix, ClonedCodeInfo *CodeInfo,
79 ValueMapTypeRemapper *TypeMapper) {
80 assert(NameSuffix && "NameSuffix cannot be null!");
83 for (Function::const_arg_iterator I = OldFunc->arg_begin(),
84 E = OldFunc->arg_end(); I != E; ++I)
85 assert(VMap.count(I) && "No mapping from source argument specified!");
88 // Clone any attributes.
89 if (NewFunc->arg_size() == OldFunc->arg_size())
90 NewFunc->copyAttributesFrom(OldFunc);
92 //Some arguments were deleted with the VMap. Copy arguments one by one
93 for (Function::const_arg_iterator I = OldFunc->arg_begin(),
94 E = OldFunc->arg_end(); I != E; ++I)
95 if (Argument* Anew = dyn_cast<Argument>(VMap[I]))
96 Anew->addAttr( OldFunc->getAttributes()
97 .getParamAttributes(I->getArgNo() + 1));
98 NewFunc->setAttributes(NewFunc->getAttributes()
99 .addAttr(0, OldFunc->getAttributes()
100 .getRetAttributes()));
101 NewFunc->setAttributes(NewFunc->getAttributes()
102 .addAttr(~0, OldFunc->getAttributes()
103 .getFnAttributes()));
107 // Loop over all of the basic blocks in the function, cloning them as
108 // appropriate. Note that we save BE this way in order to handle cloning of
109 // recursive functions into themselves.
111 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
113 const BasicBlock &BB = *BI;
115 // Create a new basic block and copy instructions into it!
116 BasicBlock *CBB = CloneBasicBlock(&BB, VMap, NameSuffix, NewFunc, CodeInfo);
118 // Add basic block mapping.
121 // It is only legal to clone a function if a block address within that
122 // function is never referenced outside of the function. Given that, we
123 // want to map block addresses from the old function to block addresses in
124 // the clone. (This is different from the generic ValueMapper
125 // implementation, which generates an invalid blockaddress when
126 // cloning a function.)
127 if (BB.hasAddressTaken()) {
128 Constant *OldBBAddr = BlockAddress::get(const_cast<Function*>(OldFunc),
129 const_cast<BasicBlock*>(&BB));
130 VMap[OldBBAddr] = BlockAddress::get(NewFunc, CBB);
133 // Note return instructions for the caller.
134 if (ReturnInst *RI = dyn_cast<ReturnInst>(CBB->getTerminator()))
135 Returns.push_back(RI);
138 // Loop over all of the instructions in the function, fixing up operand
139 // references as we go. This uses VMap to do all the hard work.
140 for (Function::iterator BB = cast<BasicBlock>(VMap[OldFunc->begin()]),
141 BE = NewFunc->end(); BB != BE; ++BB)
142 // Loop over all instructions, fixing each one as we find it...
143 for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II)
144 RemapInstruction(II, VMap,
145 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges,
149 /// CloneFunction - Return a copy of the specified function, but without
150 /// embedding the function into another module. Also, any references specified
151 /// in the VMap are changed to refer to their mapped value instead of the
152 /// original one. If any of the arguments to the function are in the VMap,
153 /// the arguments are deleted from the resultant function. The VMap is
154 /// updated to include mappings from all of the instructions and basicblocks in
155 /// the function from their old to new values.
157 Function *llvm::CloneFunction(const Function *F, ValueToValueMapTy &VMap,
158 bool ModuleLevelChanges,
159 ClonedCodeInfo *CodeInfo) {
160 std::vector<Type*> ArgTypes;
162 // The user might be deleting arguments to the function by specifying them in
163 // the VMap. If so, we need to not add the arguments to the arg ty vector
165 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
167 if (VMap.count(I) == 0) // Haven't mapped the argument to anything yet?
168 ArgTypes.push_back(I->getType());
170 // Create a new function type...
171 FunctionType *FTy = FunctionType::get(F->getFunctionType()->getReturnType(),
172 ArgTypes, F->getFunctionType()->isVarArg());
174 // Create the new function...
175 Function *NewF = Function::Create(FTy, F->getLinkage(), F->getName());
177 // Loop over the arguments, copying the names of the mapped arguments over...
178 Function::arg_iterator DestI = NewF->arg_begin();
179 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
181 if (VMap.count(I) == 0) { // Is this argument preserved?
182 DestI->setName(I->getName()); // Copy the name over...
183 VMap[I] = DestI++; // Add mapping to VMap
186 SmallVector<ReturnInst*, 8> Returns; // Ignore returns cloned.
187 CloneFunctionInto(NewF, F, VMap, ModuleLevelChanges, Returns, "", CodeInfo);
194 /// PruningFunctionCloner - This class is a private class used to implement
195 /// the CloneAndPruneFunctionInto method.
196 struct PruningFunctionCloner {
198 const Function *OldFunc;
199 ValueToValueMapTy &VMap;
200 bool ModuleLevelChanges;
201 SmallVectorImpl<ReturnInst*> &Returns;
202 const char *NameSuffix;
203 ClonedCodeInfo *CodeInfo;
204 const TargetData *TD;
206 PruningFunctionCloner(Function *newFunc, const Function *oldFunc,
207 ValueToValueMapTy &valueMap,
208 bool moduleLevelChanges,
209 SmallVectorImpl<ReturnInst*> &returns,
210 const char *nameSuffix,
211 ClonedCodeInfo *codeInfo,
212 const TargetData *td)
213 : NewFunc(newFunc), OldFunc(oldFunc),
214 VMap(valueMap), ModuleLevelChanges(moduleLevelChanges),
215 Returns(returns), NameSuffix(nameSuffix), CodeInfo(codeInfo), TD(td) {
218 /// CloneBlock - The specified block is found to be reachable, clone it and
219 /// anything that it can reach.
220 void CloneBlock(const BasicBlock *BB,
221 std::vector<const BasicBlock*> &ToClone);
224 /// ConstantFoldMappedInstruction - Constant fold the specified instruction,
225 /// mapping its operands through VMap if they are available.
226 Constant *ConstantFoldMappedInstruction(const Instruction *I);
230 /// CloneBlock - The specified block is found to be reachable, clone it and
231 /// anything that it can reach.
232 void PruningFunctionCloner::CloneBlock(const BasicBlock *BB,
233 std::vector<const BasicBlock*> &ToClone){
234 TrackingVH<Value> &BBEntry = VMap[BB];
236 // Have we already cloned this block?
239 // Nope, clone it now.
241 BBEntry = NewBB = BasicBlock::Create(BB->getContext());
242 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
244 // It is only legal to clone a function if a block address within that
245 // function is never referenced outside of the function. Given that, we
246 // want to map block addresses from the old function to block addresses in
247 // the clone. (This is different from the generic ValueMapper
248 // implementation, which generates an invalid blockaddress when
249 // cloning a function.)
251 // Note that we don't need to fix the mapping for unreachable blocks;
252 // the default mapping there is safe.
253 if (BB->hasAddressTaken()) {
254 Constant *OldBBAddr = BlockAddress::get(const_cast<Function*>(OldFunc),
255 const_cast<BasicBlock*>(BB));
256 VMap[OldBBAddr] = BlockAddress::get(NewFunc, NewBB);
260 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
262 // Loop over all instructions, and copy them over, DCE'ing as we go. This
263 // loop doesn't include the terminator.
264 for (BasicBlock::const_iterator II = BB->begin(), IE = --BB->end();
266 // If this instruction constant folds, don't bother cloning the instruction,
267 // instead, just add the constant to the value map.
268 if (Constant *C = ConstantFoldMappedInstruction(II)) {
273 Instruction *NewInst = II->clone();
275 NewInst->setName(II->getName()+NameSuffix);
276 NewBB->getInstList().push_back(NewInst);
277 VMap[II] = NewInst; // Add instruction map to value.
279 hasCalls |= (isa<CallInst>(II) && !isa<DbgInfoIntrinsic>(II));
280 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
281 if (isa<ConstantInt>(AI->getArraySize()))
282 hasStaticAllocas = true;
284 hasDynamicAllocas = true;
288 // Finally, clone over the terminator.
289 const TerminatorInst *OldTI = BB->getTerminator();
290 bool TerminatorDone = false;
291 if (const BranchInst *BI = dyn_cast<BranchInst>(OldTI)) {
292 if (BI->isConditional()) {
293 // If the condition was a known constant in the callee...
294 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
295 // Or is a known constant in the caller...
297 Value *V = VMap[BI->getCondition()];
298 Cond = dyn_cast_or_null<ConstantInt>(V);
301 // Constant fold to uncond branch!
303 BasicBlock *Dest = BI->getSuccessor(!Cond->getZExtValue());
304 VMap[OldTI] = BranchInst::Create(Dest, NewBB);
305 ToClone.push_back(Dest);
306 TerminatorDone = true;
309 } else if (const SwitchInst *SI = dyn_cast<SwitchInst>(OldTI)) {
310 // If switching on a value known constant in the caller.
311 ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition());
312 if (Cond == 0) { // Or known constant after constant prop in the callee...
313 Value *V = VMap[SI->getCondition()];
314 Cond = dyn_cast_or_null<ConstantInt>(V);
316 if (Cond) { // Constant fold to uncond branch!
317 BasicBlock *Dest = SI->getSuccessor(SI->findCaseValue(Cond));
318 VMap[OldTI] = BranchInst::Create(Dest, NewBB);
319 ToClone.push_back(Dest);
320 TerminatorDone = true;
324 if (!TerminatorDone) {
325 Instruction *NewInst = OldTI->clone();
326 if (OldTI->hasName())
327 NewInst->setName(OldTI->getName()+NameSuffix);
328 NewBB->getInstList().push_back(NewInst);
329 VMap[OldTI] = NewInst; // Add instruction map to value.
331 // Recursively clone any reachable successor blocks.
332 const TerminatorInst *TI = BB->getTerminator();
333 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
334 ToClone.push_back(TI->getSuccessor(i));
338 CodeInfo->ContainsCalls |= hasCalls;
339 CodeInfo->ContainsUnwinds |= isa<UnwindInst>(OldTI);
340 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
341 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
342 BB != &BB->getParent()->front();
345 if (ReturnInst *RI = dyn_cast<ReturnInst>(NewBB->getTerminator()))
346 Returns.push_back(RI);
349 /// ConstantFoldMappedInstruction - Constant fold the specified instruction,
350 /// mapping its operands through VMap if they are available.
351 Constant *PruningFunctionCloner::
352 ConstantFoldMappedInstruction(const Instruction *I) {
353 SmallVector<Constant*, 8> Ops;
354 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
355 if (Constant *Op = dyn_cast_or_null<Constant>(MapValue(I->getOperand(i),
357 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges)))
360 return 0; // All operands not constant!
362 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
363 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
366 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
367 if (!LI->isVolatile())
368 return ConstantFoldLoadFromConstPtr(Ops[0], TD);
370 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD);
373 /// CloneAndPruneFunctionInto - This works exactly like CloneFunctionInto,
374 /// except that it does some simple constant prop and DCE on the fly. The
375 /// effect of this is to copy significantly less code in cases where (for
376 /// example) a function call with constant arguments is inlined, and those
377 /// constant arguments cause a significant amount of code in the callee to be
378 /// dead. Since this doesn't produce an exact copy of the input, it can't be
379 /// used for things like CloneFunction or CloneModule.
380 void llvm::CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc,
381 ValueToValueMapTy &VMap,
382 bool ModuleLevelChanges,
383 SmallVectorImpl<ReturnInst*> &Returns,
384 const char *NameSuffix,
385 ClonedCodeInfo *CodeInfo,
386 const TargetData *TD,
387 Instruction *TheCall) {
388 assert(NameSuffix && "NameSuffix cannot be null!");
391 for (Function::const_arg_iterator II = OldFunc->arg_begin(),
392 E = OldFunc->arg_end(); II != E; ++II)
393 assert(VMap.count(II) && "No mapping from source argument specified!");
396 PruningFunctionCloner PFC(NewFunc, OldFunc, VMap, ModuleLevelChanges,
397 Returns, NameSuffix, CodeInfo, TD);
399 // Clone the entry block, and anything recursively reachable from it.
400 std::vector<const BasicBlock*> CloneWorklist;
401 CloneWorklist.push_back(&OldFunc->getEntryBlock());
402 while (!CloneWorklist.empty()) {
403 const BasicBlock *BB = CloneWorklist.back();
404 CloneWorklist.pop_back();
405 PFC.CloneBlock(BB, CloneWorklist);
408 // Loop over all of the basic blocks in the old function. If the block was
409 // reachable, we have cloned it and the old block is now in the value map:
410 // insert it into the new function in the right order. If not, ignore it.
412 // Defer PHI resolution until rest of function is resolved.
413 SmallVector<const PHINode*, 16> PHIToResolve;
414 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
417 BasicBlock *NewBB = cast_or_null<BasicBlock>(V);
418 if (NewBB == 0) continue; // Dead block.
420 // Add the new block to the new function.
421 NewFunc->getBasicBlockList().push_back(NewBB);
423 // Loop over all of the instructions in the block, fixing up operand
424 // references as we go. This uses VMap to do all the hard work.
426 BasicBlock::iterator I = NewBB->begin();
430 TheCallDL = TheCall->getDebugLoc();
432 // Handle PHI nodes specially, as we have to remove references to dead
434 if (PHINode *PN = dyn_cast<PHINode>(I)) {
435 // Skip over all PHI nodes, remembering them for later.
436 BasicBlock::const_iterator OldI = BI->begin();
437 for (; (PN = dyn_cast<PHINode>(I)); ++I, ++OldI)
438 PHIToResolve.push_back(cast<PHINode>(OldI));
441 // Otherwise, remap the rest of the instructions normally.
442 for (; I != NewBB->end(); ++I)
443 RemapInstruction(I, VMap,
444 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges);
447 // Defer PHI resolution until rest of function is resolved, PHI resolution
448 // requires the CFG to be up-to-date.
449 for (unsigned phino = 0, e = PHIToResolve.size(); phino != e; ) {
450 const PHINode *OPN = PHIToResolve[phino];
451 unsigned NumPreds = OPN->getNumIncomingValues();
452 const BasicBlock *OldBB = OPN->getParent();
453 BasicBlock *NewBB = cast<BasicBlock>(VMap[OldBB]);
455 // Map operands for blocks that are live and remove operands for blocks
457 for (; phino != PHIToResolve.size() &&
458 PHIToResolve[phino]->getParent() == OldBB; ++phino) {
459 OPN = PHIToResolve[phino];
460 PHINode *PN = cast<PHINode>(VMap[OPN]);
461 for (unsigned pred = 0, e = NumPreds; pred != e; ++pred) {
462 Value *V = VMap[PN->getIncomingBlock(pred)];
463 if (BasicBlock *MappedBlock = cast_or_null<BasicBlock>(V)) {
464 Value *InVal = MapValue(PN->getIncomingValue(pred),
466 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges);
467 assert(InVal && "Unknown input value?");
468 PN->setIncomingValue(pred, InVal);
469 PN->setIncomingBlock(pred, MappedBlock);
471 PN->removeIncomingValue(pred, false);
472 --pred, --e; // Revisit the next entry.
477 // The loop above has removed PHI entries for those blocks that are dead
478 // and has updated others. However, if a block is live (i.e. copied over)
479 // but its terminator has been changed to not go to this block, then our
480 // phi nodes will have invalid entries. Update the PHI nodes in this
482 PHINode *PN = cast<PHINode>(NewBB->begin());
483 NumPreds = std::distance(pred_begin(NewBB), pred_end(NewBB));
484 if (NumPreds != PN->getNumIncomingValues()) {
485 assert(NumPreds < PN->getNumIncomingValues());
486 // Count how many times each predecessor comes to this block.
487 std::map<BasicBlock*, unsigned> PredCount;
488 for (pred_iterator PI = pred_begin(NewBB), E = pred_end(NewBB);
492 // Figure out how many entries to remove from each PHI.
493 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
494 ++PredCount[PN->getIncomingBlock(i)];
496 // At this point, the excess predecessor entries are positive in the
497 // map. Loop over all of the PHIs and remove excess predecessor
499 BasicBlock::iterator I = NewBB->begin();
500 for (; (PN = dyn_cast<PHINode>(I)); ++I) {
501 for (std::map<BasicBlock*, unsigned>::iterator PCI =PredCount.begin(),
502 E = PredCount.end(); PCI != E; ++PCI) {
503 BasicBlock *Pred = PCI->first;
504 for (unsigned NumToRemove = PCI->second; NumToRemove; --NumToRemove)
505 PN->removeIncomingValue(Pred, false);
510 // If the loops above have made these phi nodes have 0 or 1 operand,
511 // replace them with undef or the input value. We must do this for
512 // correctness, because 0-operand phis are not valid.
513 PN = cast<PHINode>(NewBB->begin());
514 if (PN->getNumIncomingValues() == 0) {
515 BasicBlock::iterator I = NewBB->begin();
516 BasicBlock::const_iterator OldI = OldBB->begin();
517 while ((PN = dyn_cast<PHINode>(I++))) {
518 Value *NV = UndefValue::get(PN->getType());
519 PN->replaceAllUsesWith(NV);
520 assert(VMap[OldI] == PN && "VMap mismatch");
522 PN->eraseFromParent();
526 // NOTE: We cannot eliminate single entry phi nodes here, because of
527 // VMap. Single entry phi nodes can have multiple VMap entries
528 // pointing at them. Thus, deleting one would require scanning the VMap
529 // to update any entries in it that would require that. This would be
533 // Now that the inlined function body has been fully constructed, go through
534 // and zap unconditional fall-through branches. This happen all the time when
535 // specializing code: code specialization turns conditional branches into
536 // uncond branches, and this code folds them.
537 Function::iterator I = cast<BasicBlock>(VMap[&OldFunc->getEntryBlock()]);
538 while (I != NewFunc->end()) {
539 BranchInst *BI = dyn_cast<BranchInst>(I->getTerminator());
540 if (!BI || BI->isConditional()) { ++I; continue; }
542 // Note that we can't eliminate uncond branches if the destination has
543 // single-entry PHI nodes. Eliminating the single-entry phi nodes would
544 // require scanning the VMap to update any entries that point to the phi
546 BasicBlock *Dest = BI->getSuccessor(0);
547 if (!Dest->getSinglePredecessor() || isa<PHINode>(Dest->begin())) {
551 // We know all single-entry PHI nodes in the inlined function have been
552 // removed, so we just need to splice the blocks.
553 BI->eraseFromParent();
555 // Make all PHI nodes that referred to Dest now refer to I as their source.
556 Dest->replaceAllUsesWith(I);
558 // Move all the instructions in the succ to the pred.
559 I->getInstList().splice(I->end(), Dest->getInstList());
561 // Remove the dest block.
562 Dest->eraseFromParent();
564 // Do not increment I, iteratively merge all things this block branches to.