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 assert(NameSuffix && "NameSuffix cannot be null!");
82 for (Function::const_arg_iterator I = OldFunc->arg_begin(),
83 E = OldFunc->arg_end(); I != E; ++I)
84 assert(VMap.count(I) && "No mapping from source argument specified!");
87 // Clone any attributes.
88 if (NewFunc->arg_size() == OldFunc->arg_size())
89 NewFunc->copyAttributesFrom(OldFunc);
91 //Some arguments were deleted with the VMap. Copy arguments one by one
92 for (Function::const_arg_iterator I = OldFunc->arg_begin(),
93 E = OldFunc->arg_end(); I != E; ++I)
94 if (Argument* Anew = dyn_cast<Argument>(VMap[I]))
95 Anew->addAttr( OldFunc->getAttributes()
96 .getParamAttributes(I->getArgNo() + 1));
97 NewFunc->setAttributes(NewFunc->getAttributes()
98 .addAttr(0, OldFunc->getAttributes()
99 .getRetAttributes()));
100 NewFunc->setAttributes(NewFunc->getAttributes()
101 .addAttr(~0, OldFunc->getAttributes()
102 .getFnAttributes()));
106 // Loop over all of the basic blocks in the function, cloning them as
107 // appropriate. Note that we save BE this way in order to handle cloning of
108 // recursive functions into themselves.
110 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
112 const BasicBlock &BB = *BI;
114 // Create a new basic block and copy instructions into it!
115 BasicBlock *CBB = CloneBasicBlock(&BB, VMap, NameSuffix, NewFunc,
117 VMap[&BB] = CBB; // Add basic block mapping.
119 if (ReturnInst *RI = dyn_cast<ReturnInst>(CBB->getTerminator()))
120 Returns.push_back(RI);
123 // Loop over all of the instructions in the function, fixing up operand
124 // references as we go. This uses VMap to do all the hard work.
126 for (Function::iterator BB = cast<BasicBlock>(VMap[OldFunc->begin()]),
127 BE = NewFunc->end(); BB != BE; ++BB)
128 // Loop over all instructions, fixing each one as we find it...
129 for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II)
130 RemapInstruction(II, VMap, ModuleLevelChanges);
133 /// CloneFunction - Return a copy of the specified function, but without
134 /// embedding the function into another module. Also, any references specified
135 /// in the VMap are changed to refer to their mapped value instead of the
136 /// original one. If any of the arguments to the function are in the VMap,
137 /// the arguments are deleted from the resultant function. The VMap is
138 /// updated to include mappings from all of the instructions and basicblocks in
139 /// the function from their old to new values.
141 Function *llvm::CloneFunction(const Function *F,
142 ValueToValueMapTy &VMap,
143 bool ModuleLevelChanges,
144 ClonedCodeInfo *CodeInfo) {
145 std::vector<const Type*> ArgTypes;
147 // The user might be deleting arguments to the function by specifying them in
148 // the VMap. If so, we need to not add the arguments to the arg ty vector
150 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
152 if (VMap.count(I) == 0) // Haven't mapped the argument to anything yet?
153 ArgTypes.push_back(I->getType());
155 // Create a new function type...
156 FunctionType *FTy = FunctionType::get(F->getFunctionType()->getReturnType(),
157 ArgTypes, F->getFunctionType()->isVarArg());
159 // Create the new function...
160 Function *NewF = Function::Create(FTy, F->getLinkage(), F->getName());
162 // Loop over the arguments, copying the names of the mapped arguments over...
163 Function::arg_iterator DestI = NewF->arg_begin();
164 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
166 if (VMap.count(I) == 0) { // Is this argument preserved?
167 DestI->setName(I->getName()); // Copy the name over...
168 VMap[I] = DestI++; // Add mapping to VMap
171 SmallVector<ReturnInst*, 8> Returns; // Ignore returns cloned.
172 CloneFunctionInto(NewF, F, VMap, ModuleLevelChanges, Returns, "", CodeInfo);
179 /// PruningFunctionCloner - This class is a private class used to implement
180 /// the CloneAndPruneFunctionInto method.
181 struct PruningFunctionCloner {
183 const Function *OldFunc;
184 ValueToValueMapTy &VMap;
185 bool ModuleLevelChanges;
186 SmallVectorImpl<ReturnInst*> &Returns;
187 const char *NameSuffix;
188 ClonedCodeInfo *CodeInfo;
189 const TargetData *TD;
191 PruningFunctionCloner(Function *newFunc, const Function *oldFunc,
192 ValueToValueMapTy &valueMap,
193 bool moduleLevelChanges,
194 SmallVectorImpl<ReturnInst*> &returns,
195 const char *nameSuffix,
196 ClonedCodeInfo *codeInfo,
197 const TargetData *td)
198 : NewFunc(newFunc), OldFunc(oldFunc),
199 VMap(valueMap), ModuleLevelChanges(moduleLevelChanges),
200 Returns(returns), NameSuffix(nameSuffix), CodeInfo(codeInfo), TD(td) {
203 /// CloneBlock - The specified block is found to be reachable, clone it and
204 /// anything that it can reach.
205 void CloneBlock(const BasicBlock *BB,
206 std::vector<const BasicBlock*> &ToClone);
209 /// ConstantFoldMappedInstruction - Constant fold the specified instruction,
210 /// mapping its operands through VMap if they are available.
211 Constant *ConstantFoldMappedInstruction(const Instruction *I);
215 /// CloneBlock - The specified block is found to be reachable, clone it and
216 /// anything that it can reach.
217 void PruningFunctionCloner::CloneBlock(const BasicBlock *BB,
218 std::vector<const BasicBlock*> &ToClone){
219 Value *&BBEntry = VMap[BB];
221 // Have we already cloned this block?
224 // Nope, clone it now.
226 BBEntry = NewBB = BasicBlock::Create(BB->getContext());
227 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
229 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
231 // Loop over all instructions, and copy them over, DCE'ing as we go. This
232 // loop doesn't include the terminator.
233 for (BasicBlock::const_iterator II = BB->begin(), IE = --BB->end();
235 // If this instruction constant folds, don't bother cloning the instruction,
236 // instead, just add the constant to the value map.
237 if (Constant *C = ConstantFoldMappedInstruction(II)) {
242 Instruction *NewInst = II->clone();
244 NewInst->setName(II->getName()+NameSuffix);
245 NewBB->getInstList().push_back(NewInst);
246 VMap[II] = NewInst; // Add instruction map to value.
248 hasCalls |= (isa<CallInst>(II) && !isa<DbgInfoIntrinsic>(II));
249 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
250 if (isa<ConstantInt>(AI->getArraySize()))
251 hasStaticAllocas = true;
253 hasDynamicAllocas = true;
257 // Finally, clone over the terminator.
258 const TerminatorInst *OldTI = BB->getTerminator();
259 bool TerminatorDone = false;
260 if (const BranchInst *BI = dyn_cast<BranchInst>(OldTI)) {
261 if (BI->isConditional()) {
262 // If the condition was a known constant in the callee...
263 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
264 // Or is a known constant in the caller...
266 Cond = dyn_cast_or_null<ConstantInt>(VMap[BI->getCondition()]);
268 // Constant fold to uncond branch!
270 BasicBlock *Dest = BI->getSuccessor(!Cond->getZExtValue());
271 VMap[OldTI] = BranchInst::Create(Dest, NewBB);
272 ToClone.push_back(Dest);
273 TerminatorDone = true;
276 } else if (const SwitchInst *SI = dyn_cast<SwitchInst>(OldTI)) {
277 // If switching on a value known constant in the caller.
278 ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition());
279 if (Cond == 0) // Or known constant after constant prop in the callee...
280 Cond = dyn_cast_or_null<ConstantInt>(VMap[SI->getCondition()]);
281 if (Cond) { // Constant fold to uncond branch!
282 BasicBlock *Dest = SI->getSuccessor(SI->findCaseValue(Cond));
283 VMap[OldTI] = BranchInst::Create(Dest, NewBB);
284 ToClone.push_back(Dest);
285 TerminatorDone = true;
289 if (!TerminatorDone) {
290 Instruction *NewInst = OldTI->clone();
291 if (OldTI->hasName())
292 NewInst->setName(OldTI->getName()+NameSuffix);
293 NewBB->getInstList().push_back(NewInst);
294 VMap[OldTI] = NewInst; // Add instruction map to value.
296 // Recursively clone any reachable successor blocks.
297 const TerminatorInst *TI = BB->getTerminator();
298 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
299 ToClone.push_back(TI->getSuccessor(i));
303 CodeInfo->ContainsCalls |= hasCalls;
304 CodeInfo->ContainsUnwinds |= isa<UnwindInst>(OldTI);
305 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
306 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
307 BB != &BB->getParent()->front();
310 if (ReturnInst *RI = dyn_cast<ReturnInst>(NewBB->getTerminator()))
311 Returns.push_back(RI);
314 /// ConstantFoldMappedInstruction - Constant fold the specified instruction,
315 /// mapping its operands through VMap if they are available.
316 Constant *PruningFunctionCloner::
317 ConstantFoldMappedInstruction(const Instruction *I) {
318 SmallVector<Constant*, 8> Ops;
319 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
320 if (Constant *Op = dyn_cast_or_null<Constant>(MapValue(I->getOperand(i),
321 VMap, ModuleLevelChanges)))
324 return 0; // All operands not constant!
326 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
327 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
330 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
331 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
332 if (!LI->isVolatile() && CE->getOpcode() == Instruction::GetElementPtr)
333 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
334 if (GV->isConstant() && GV->hasDefinitiveInitializer())
335 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(),
338 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), &Ops[0],
343 UpdateInlinedAtInfo(const DebugLoc &InsnDL, const DebugLoc &TheCallDL,
345 DebugLoc NewLoc = TheCallDL;
346 if (MDNode *IA = InsnDL.getInlinedAt(Ctx))
347 NewLoc = UpdateInlinedAtInfo(DebugLoc::getFromDILocation(IA), TheCallDL,
350 return DebugLoc::get(InsnDL.getLine(), InsnDL.getCol(),
351 InsnDL.getScope(Ctx), NewLoc.getAsMDNode(Ctx));
354 /// CloneAndPruneFunctionInto - This works exactly like CloneFunctionInto,
355 /// except that it does some simple constant prop and DCE on the fly. The
356 /// effect of this is to copy significantly less code in cases where (for
357 /// example) a function call with constant arguments is inlined, and those
358 /// constant arguments cause a significant amount of code in the callee to be
359 /// dead. Since this doesn't produce an exact copy of the input, it can't be
360 /// used for things like CloneFunction or CloneModule.
361 void llvm::CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc,
362 ValueToValueMapTy &VMap,
363 bool ModuleLevelChanges,
364 SmallVectorImpl<ReturnInst*> &Returns,
365 const char *NameSuffix,
366 ClonedCodeInfo *CodeInfo,
367 const TargetData *TD,
368 Instruction *TheCall) {
369 assert(NameSuffix && "NameSuffix cannot be null!");
372 for (Function::const_arg_iterator II = OldFunc->arg_begin(),
373 E = OldFunc->arg_end(); II != E; ++II)
374 assert(VMap.count(II) && "No mapping from source argument specified!");
377 PruningFunctionCloner PFC(NewFunc, OldFunc, VMap, ModuleLevelChanges,
378 Returns, NameSuffix, CodeInfo, TD);
380 // Clone the entry block, and anything recursively reachable from it.
381 std::vector<const BasicBlock*> CloneWorklist;
382 CloneWorklist.push_back(&OldFunc->getEntryBlock());
383 while (!CloneWorklist.empty()) {
384 const BasicBlock *BB = CloneWorklist.back();
385 CloneWorklist.pop_back();
386 PFC.CloneBlock(BB, CloneWorklist);
389 // Loop over all of the basic blocks in the old function. If the block was
390 // reachable, we have cloned it and the old block is now in the value map:
391 // insert it into the new function in the right order. If not, ignore it.
393 // Defer PHI resolution until rest of function is resolved.
394 SmallVector<const PHINode*, 16> PHIToResolve;
395 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
397 BasicBlock *NewBB = cast_or_null<BasicBlock>(VMap[BI]);
398 if (NewBB == 0) continue; // Dead block.
400 // Add the new block to the new function.
401 NewFunc->getBasicBlockList().push_back(NewBB);
403 // Loop over all of the instructions in the block, fixing up operand
404 // references as we go. This uses VMap to do all the hard work.
406 BasicBlock::iterator I = NewBB->begin();
410 TheCallDL = TheCall->getDebugLoc();
412 // Handle PHI nodes specially, as we have to remove references to dead
414 if (PHINode *PN = dyn_cast<PHINode>(I)) {
415 // Skip over all PHI nodes, remembering them for later.
416 BasicBlock::const_iterator OldI = BI->begin();
417 for (; (PN = dyn_cast<PHINode>(I)); ++I, ++OldI) {
418 if (I->hasMetadata()) {
419 if (!TheCallDL.isUnknown()) {
420 DebugLoc IDL = I->getDebugLoc();
421 if (!IDL.isUnknown()) {
422 DebugLoc NewDL = UpdateInlinedAtInfo(IDL, TheCallDL,
424 I->setDebugLoc(NewDL);
427 // The cloned instruction has dbg info but the call instruction
428 // does not have dbg info. Remove dbg info from cloned instruction.
429 I->setDebugLoc(DebugLoc());
432 PHIToResolve.push_back(cast<PHINode>(OldI));
438 // FIXME: Unclone all this metadata stuff.
442 // Otherwise, remap the rest of the instructions normally.
443 for (; I != NewBB->end(); ++I) {
444 if (I->hasMetadata()) {
445 if (!TheCallDL.isUnknown()) {
446 DebugLoc IDL = I->getDebugLoc();
447 if (!IDL.isUnknown()) {
448 DebugLoc NewDL = UpdateInlinedAtInfo(IDL, TheCallDL,
450 I->setDebugLoc(NewDL);
453 // The cloned instruction has dbg info but the call instruction
454 // does not have dbg info. Remove dbg info from cloned instruction.
455 I->setDebugLoc(DebugLoc());
458 RemapInstruction(I, VMap, ModuleLevelChanges);
462 // Defer PHI resolution until rest of function is resolved, PHI resolution
463 // requires the CFG to be up-to-date.
464 for (unsigned phino = 0, e = PHIToResolve.size(); phino != e; ) {
465 const PHINode *OPN = PHIToResolve[phino];
466 unsigned NumPreds = OPN->getNumIncomingValues();
467 const BasicBlock *OldBB = OPN->getParent();
468 BasicBlock *NewBB = cast<BasicBlock>(VMap[OldBB]);
470 // Map operands for blocks that are live and remove operands for blocks
472 for (; phino != PHIToResolve.size() &&
473 PHIToResolve[phino]->getParent() == OldBB; ++phino) {
474 OPN = PHIToResolve[phino];
475 PHINode *PN = cast<PHINode>(VMap[OPN]);
476 for (unsigned pred = 0, e = NumPreds; pred != e; ++pred) {
477 if (BasicBlock *MappedBlock =
478 cast_or_null<BasicBlock>(VMap[PN->getIncomingBlock(pred)])) {
479 Value *InVal = MapValue(PN->getIncomingValue(pred),
480 VMap, ModuleLevelChanges);
481 assert(InVal && "Unknown input value?");
482 PN->setIncomingValue(pred, InVal);
483 PN->setIncomingBlock(pred, MappedBlock);
485 PN->removeIncomingValue(pred, false);
486 --pred, --e; // Revisit the next entry.
491 // The loop above has removed PHI entries for those blocks that are dead
492 // and has updated others. However, if a block is live (i.e. copied over)
493 // but its terminator has been changed to not go to this block, then our
494 // phi nodes will have invalid entries. Update the PHI nodes in this
496 PHINode *PN = cast<PHINode>(NewBB->begin());
497 NumPreds = std::distance(pred_begin(NewBB), pred_end(NewBB));
498 if (NumPreds != PN->getNumIncomingValues()) {
499 assert(NumPreds < PN->getNumIncomingValues());
500 // Count how many times each predecessor comes to this block.
501 std::map<BasicBlock*, unsigned> PredCount;
502 for (pred_iterator PI = pred_begin(NewBB), E = pred_end(NewBB);
506 // Figure out how many entries to remove from each PHI.
507 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
508 ++PredCount[PN->getIncomingBlock(i)];
510 // At this point, the excess predecessor entries are positive in the
511 // map. Loop over all of the PHIs and remove excess predecessor
513 BasicBlock::iterator I = NewBB->begin();
514 for (; (PN = dyn_cast<PHINode>(I)); ++I) {
515 for (std::map<BasicBlock*, unsigned>::iterator PCI =PredCount.begin(),
516 E = PredCount.end(); PCI != E; ++PCI) {
517 BasicBlock *Pred = PCI->first;
518 for (unsigned NumToRemove = PCI->second; NumToRemove; --NumToRemove)
519 PN->removeIncomingValue(Pred, false);
524 // If the loops above have made these phi nodes have 0 or 1 operand,
525 // replace them with undef or the input value. We must do this for
526 // correctness, because 0-operand phis are not valid.
527 PN = cast<PHINode>(NewBB->begin());
528 if (PN->getNumIncomingValues() == 0) {
529 BasicBlock::iterator I = NewBB->begin();
530 BasicBlock::const_iterator OldI = OldBB->begin();
531 while ((PN = dyn_cast<PHINode>(I++))) {
532 Value *NV = UndefValue::get(PN->getType());
533 PN->replaceAllUsesWith(NV);
534 assert(VMap[OldI] == PN && "VMap mismatch");
536 PN->eraseFromParent();
540 // NOTE: We cannot eliminate single entry phi nodes here, because of
541 // VMap. Single entry phi nodes can have multiple VMap entries
542 // pointing at them. Thus, deleting one would require scanning the VMap
543 // to update any entries in it that would require that. This would be
547 // Now that the inlined function body has been fully constructed, go through
548 // and zap unconditional fall-through branches. This happen all the time when
549 // specializing code: code specialization turns conditional branches into
550 // uncond branches, and this code folds them.
551 Function::iterator I = cast<BasicBlock>(VMap[&OldFunc->getEntryBlock()]);
552 while (I != NewFunc->end()) {
553 BranchInst *BI = dyn_cast<BranchInst>(I->getTerminator());
554 if (!BI || BI->isConditional()) { ++I; continue; }
556 // Note that we can't eliminate uncond branches if the destination has
557 // single-entry PHI nodes. Eliminating the single-entry phi nodes would
558 // require scanning the VMap to update any entries that point to the phi
560 BasicBlock *Dest = BI->getSuccessor(0);
561 if (!Dest->getSinglePredecessor() || isa<PHINode>(Dest->begin())) {
565 // We know all single-entry PHI nodes in the inlined function have been
566 // removed, so we just need to splice the blocks.
567 BI->eraseFromParent();
569 // Move all the instructions in the succ to the pred.
570 I->getInstList().splice(I->end(), Dest->getInstList());
572 // Make all PHI nodes that referred to Dest now refer to I as their source.
573 Dest->replaceAllUsesWith(I);
575 // Remove the dest block.
576 Dest->eraseFromParent();
578 // Do not increment I, iteratively merge all things this block branches to.