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/GlobalVariable.h"
21 #include "llvm/Function.h"
22 #include "llvm/Support/CFG.h"
23 #include "llvm/Support/Compiler.h"
24 #include "llvm/Transforms/Utils/ValueMapper.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/ADT/SmallVector.h"
30 // CloneBasicBlock - See comments in Cloning.h
31 BasicBlock *llvm::CloneBasicBlock(const BasicBlock *BB,
32 DenseMap<const Value*, Value*> &ValueMap,
33 const char *NameSuffix, Function *F,
34 ClonedCodeInfo *CodeInfo) {
35 BasicBlock *NewBB = BasicBlock::Create("", F);
36 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
38 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
40 // Loop over all instructions, and copy them over.
41 for (BasicBlock::const_iterator II = BB->begin(), IE = BB->end();
43 Instruction *NewInst = II->clone();
45 NewInst->setName(II->getName()+NameSuffix);
46 NewBB->getInstList().push_back(NewInst);
47 ValueMap[II] = NewInst; // Add instruction map to value.
49 hasCalls |= isa<CallInst>(II);
50 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
51 if (isa<ConstantInt>(AI->getArraySize()))
52 hasStaticAllocas = true;
54 hasDynamicAllocas = true;
59 CodeInfo->ContainsCalls |= hasCalls;
60 CodeInfo->ContainsUnwinds |= isa<UnwindInst>(BB->getTerminator());
61 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
62 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
63 BB != &BB->getParent()->getEntryBlock();
68 // Clone OldFunc into NewFunc, transforming the old arguments into references to
71 void llvm::CloneFunctionInto(Function *NewFunc, const Function *OldFunc,
72 DenseMap<const Value*, Value*> &ValueMap,
73 std::vector<ReturnInst*> &Returns,
74 const char *NameSuffix, ClonedCodeInfo *CodeInfo) {
75 assert(NameSuffix && "NameSuffix cannot be null!");
78 for (Function::const_arg_iterator I = OldFunc->arg_begin(),
79 E = OldFunc->arg_end(); I != E; ++I)
80 assert(ValueMap.count(I) && "No mapping from source argument specified!");
83 // Clone the parameter attributes
84 NewFunc->setParamAttrs(OldFunc->getParamAttrs());
86 // Clone the calling convention
87 NewFunc->setCallingConv(OldFunc->getCallingConv());
89 // Loop over all of the basic blocks in the function, cloning them as
90 // appropriate. Note that we save BE this way in order to handle cloning of
91 // recursive functions into themselves.
93 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
95 const BasicBlock &BB = *BI;
97 // Create a new basic block and copy instructions into it!
98 BasicBlock *CBB = CloneBasicBlock(&BB, ValueMap, NameSuffix, NewFunc,
100 ValueMap[&BB] = CBB; // Add basic block mapping.
102 if (ReturnInst *RI = dyn_cast<ReturnInst>(CBB->getTerminator()))
103 Returns.push_back(RI);
106 // Loop over all of the instructions in the function, fixing up operand
107 // references as we go. This uses ValueMap to do all the hard work.
109 for (Function::iterator BB = cast<BasicBlock>(ValueMap[OldFunc->begin()]),
110 BE = NewFunc->end(); BB != BE; ++BB)
111 // Loop over all instructions, fixing each one as we find it...
112 for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II)
113 RemapInstruction(II, ValueMap);
116 /// CloneFunction - Return a copy of the specified function, but without
117 /// embedding the function into another module. Also, any references specified
118 /// in the ValueMap are changed to refer to their mapped value instead of the
119 /// original one. If any of the arguments to the function are in the ValueMap,
120 /// the arguments are deleted from the resultant function. The ValueMap is
121 /// updated to include mappings from all of the instructions and basicblocks in
122 /// the function from their old to new values.
124 Function *llvm::CloneFunction(const Function *F,
125 DenseMap<const Value*, Value*> &ValueMap,
126 ClonedCodeInfo *CodeInfo) {
127 std::vector<const Type*> ArgTypes;
129 // The user might be deleting arguments to the function by specifying them in
130 // the ValueMap. If so, we need to not add the arguments to the arg ty vector
132 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
134 if (ValueMap.count(I) == 0) // Haven't mapped the argument to anything yet?
135 ArgTypes.push_back(I->getType());
137 // Create a new function type...
138 FunctionType *FTy = FunctionType::get(F->getFunctionType()->getReturnType(),
139 ArgTypes, F->getFunctionType()->isVarArg());
141 // Create the new function...
142 Function *NewF = Function::Create(FTy, F->getLinkage(), F->getName());
144 // Loop over the arguments, copying the names of the mapped arguments over...
145 Function::arg_iterator DestI = NewF->arg_begin();
146 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
148 if (ValueMap.count(I) == 0) { // Is this argument preserved?
149 DestI->setName(I->getName()); // Copy the name over...
150 ValueMap[I] = DestI++; // Add mapping to ValueMap
153 std::vector<ReturnInst*> Returns; // Ignore returns cloned...
154 CloneFunctionInto(NewF, F, ValueMap, Returns, "", CodeInfo);
161 /// PruningFunctionCloner - This class is a private class used to implement
162 /// the CloneAndPruneFunctionInto method.
163 struct VISIBILITY_HIDDEN PruningFunctionCloner {
165 const Function *OldFunc;
166 DenseMap<const Value*, Value*> &ValueMap;
167 std::vector<ReturnInst*> &Returns;
168 const char *NameSuffix;
169 ClonedCodeInfo *CodeInfo;
170 const TargetData *TD;
173 PruningFunctionCloner(Function *newFunc, const Function *oldFunc,
174 DenseMap<const Value*, Value*> &valueMap,
175 std::vector<ReturnInst*> &returns,
176 const char *nameSuffix,
177 ClonedCodeInfo *codeInfo,
178 const TargetData *td)
179 : NewFunc(newFunc), OldFunc(oldFunc), ValueMap(valueMap), Returns(returns),
180 NameSuffix(nameSuffix), CodeInfo(codeInfo), TD(td) {
183 /// CloneBlock - The specified block is found to be reachable, clone it and
184 /// anything that it can reach.
185 void CloneBlock(const BasicBlock *BB,
186 std::vector<const BasicBlock*> &ToClone);
189 /// ConstantFoldMappedInstruction - Constant fold the specified instruction,
190 /// mapping its operands through ValueMap if they are available.
191 Constant *ConstantFoldMappedInstruction(const Instruction *I);
195 /// CloneBlock - The specified block is found to be reachable, clone it and
196 /// anything that it can reach.
197 void PruningFunctionCloner::CloneBlock(const BasicBlock *BB,
198 std::vector<const BasicBlock*> &ToClone){
199 Value *&BBEntry = ValueMap[BB];
201 // Have we already cloned this block?
204 // Nope, clone it now.
206 BBEntry = NewBB = BasicBlock::Create();
207 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
209 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
211 // Loop over all instructions, and copy them over, DCE'ing as we go. This
212 // loop doesn't include the terminator.
213 for (BasicBlock::const_iterator II = BB->begin(), IE = --BB->end();
215 // If this instruction constant folds, don't bother cloning the instruction,
216 // instead, just add the constant to the value map.
217 if (Constant *C = ConstantFoldMappedInstruction(II)) {
222 Instruction *NewInst = II->clone();
224 NewInst->setName(II->getName()+NameSuffix);
225 NewBB->getInstList().push_back(NewInst);
226 ValueMap[II] = NewInst; // Add instruction map to value.
228 hasCalls |= isa<CallInst>(II);
229 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
230 if (isa<ConstantInt>(AI->getArraySize()))
231 hasStaticAllocas = true;
233 hasDynamicAllocas = true;
237 // Finally, clone over the terminator.
238 const TerminatorInst *OldTI = BB->getTerminator();
239 bool TerminatorDone = false;
240 if (const BranchInst *BI = dyn_cast<BranchInst>(OldTI)) {
241 if (BI->isConditional()) {
242 // If the condition was a known constant in the callee...
243 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
244 // Or is a known constant in the caller...
246 Cond = dyn_cast_or_null<ConstantInt>(ValueMap[BI->getCondition()]);
248 // Constant fold to uncond branch!
250 BasicBlock *Dest = BI->getSuccessor(!Cond->getZExtValue());
251 ValueMap[OldTI] = BranchInst::Create(Dest, NewBB);
252 ToClone.push_back(Dest);
253 TerminatorDone = true;
256 } else if (const SwitchInst *SI = dyn_cast<SwitchInst>(OldTI)) {
257 // If switching on a value known constant in the caller.
258 ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition());
259 if (Cond == 0) // Or known constant after constant prop in the callee...
260 Cond = dyn_cast_or_null<ConstantInt>(ValueMap[SI->getCondition()]);
261 if (Cond) { // Constant fold to uncond branch!
262 BasicBlock *Dest = SI->getSuccessor(SI->findCaseValue(Cond));
263 ValueMap[OldTI] = BranchInst::Create(Dest, NewBB);
264 ToClone.push_back(Dest);
265 TerminatorDone = true;
269 if (!TerminatorDone) {
270 Instruction *NewInst = OldTI->clone();
271 if (OldTI->hasName())
272 NewInst->setName(OldTI->getName()+NameSuffix);
273 NewBB->getInstList().push_back(NewInst);
274 ValueMap[OldTI] = NewInst; // Add instruction map to value.
276 // Recursively clone any reachable successor blocks.
277 const TerminatorInst *TI = BB->getTerminator();
278 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
279 ToClone.push_back(TI->getSuccessor(i));
283 CodeInfo->ContainsCalls |= hasCalls;
284 CodeInfo->ContainsUnwinds |= isa<UnwindInst>(OldTI);
285 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
286 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
287 BB != &BB->getParent()->front();
290 if (ReturnInst *RI = dyn_cast<ReturnInst>(NewBB->getTerminator()))
291 Returns.push_back(RI);
294 /// ConstantFoldMappedInstruction - Constant fold the specified instruction,
295 /// mapping its operands through ValueMap if they are available.
296 Constant *PruningFunctionCloner::
297 ConstantFoldMappedInstruction(const Instruction *I) {
298 SmallVector<Constant*, 8> Ops;
299 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
300 if (Constant *Op = dyn_cast_or_null<Constant>(MapValue(I->getOperand(i),
304 return 0; // All operands not constant!
306 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
307 return ConstantFoldCompareInstOperands(CI->getPredicate(),
308 &Ops[0], Ops.size(), TD);
310 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
311 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
312 if (!LI->isVolatile() && CE->getOpcode() == Instruction::GetElementPtr)
313 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
314 if (GV->isConstant() && !GV->isDeclaration())
315 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(),
318 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), &Ops[0],
322 /// CloneAndPruneFunctionInto - This works exactly like CloneFunctionInto,
323 /// except that it does some simple constant prop and DCE on the fly. The
324 /// effect of this is to copy significantly less code in cases where (for
325 /// example) a function call with constant arguments is inlined, and those
326 /// constant arguments cause a significant amount of code in the callee to be
327 /// dead. Since this doesn't produce an exact copy of the input, it can't be
328 /// used for things like CloneFunction or CloneModule.
329 void llvm::CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc,
330 DenseMap<const Value*, Value*> &ValueMap,
331 std::vector<ReturnInst*> &Returns,
332 const char *NameSuffix,
333 ClonedCodeInfo *CodeInfo,
334 const TargetData *TD) {
335 assert(NameSuffix && "NameSuffix cannot be null!");
338 for (Function::const_arg_iterator II = OldFunc->arg_begin(),
339 E = OldFunc->arg_end(); II != E; ++II)
340 assert(ValueMap.count(II) && "No mapping from source argument specified!");
343 PruningFunctionCloner PFC(NewFunc, OldFunc, ValueMap, Returns,
344 NameSuffix, CodeInfo, TD);
346 // Clone the entry block, and anything recursively reachable from it.
347 std::vector<const BasicBlock*> CloneWorklist;
348 CloneWorklist.push_back(&OldFunc->getEntryBlock());
349 while (!CloneWorklist.empty()) {
350 const BasicBlock *BB = CloneWorklist.back();
351 CloneWorklist.pop_back();
352 PFC.CloneBlock(BB, CloneWorklist);
355 // Loop over all of the basic blocks in the old function. If the block was
356 // reachable, we have cloned it and the old block is now in the value map:
357 // insert it into the new function in the right order. If not, ignore it.
359 // Defer PHI resolution until rest of function is resolved.
360 std::vector<const PHINode*> PHIToResolve;
361 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
363 BasicBlock *NewBB = cast_or_null<BasicBlock>(ValueMap[BI]);
364 if (NewBB == 0) continue; // Dead block.
366 // Add the new block to the new function.
367 NewFunc->getBasicBlockList().push_back(NewBB);
369 // Loop over all of the instructions in the block, fixing up operand
370 // references as we go. This uses ValueMap to do all the hard work.
372 BasicBlock::iterator I = NewBB->begin();
374 // Handle PHI nodes specially, as we have to remove references to dead
376 if (PHINode *PN = dyn_cast<PHINode>(I)) {
377 // Skip over all PHI nodes, remembering them for later.
378 BasicBlock::const_iterator OldI = BI->begin();
379 for (; (PN = dyn_cast<PHINode>(I)); ++I, ++OldI)
380 PHIToResolve.push_back(cast<PHINode>(OldI));
383 // Otherwise, remap the rest of the instructions normally.
384 for (; I != NewBB->end(); ++I)
385 RemapInstruction(I, ValueMap);
388 // Defer PHI resolution until rest of function is resolved, PHI resolution
389 // requires the CFG to be up-to-date.
390 for (unsigned phino = 0, e = PHIToResolve.size(); phino != e; ) {
391 const PHINode *OPN = PHIToResolve[phino];
392 unsigned NumPreds = OPN->getNumIncomingValues();
393 const BasicBlock *OldBB = OPN->getParent();
394 BasicBlock *NewBB = cast<BasicBlock>(ValueMap[OldBB]);
396 // Map operands for blocks that are live and remove operands for blocks
398 for (; phino != PHIToResolve.size() &&
399 PHIToResolve[phino]->getParent() == OldBB; ++phino) {
400 OPN = PHIToResolve[phino];
401 PHINode *PN = cast<PHINode>(ValueMap[OPN]);
402 for (unsigned pred = 0, e = NumPreds; pred != e; ++pred) {
403 if (BasicBlock *MappedBlock =
404 cast_or_null<BasicBlock>(ValueMap[PN->getIncomingBlock(pred)])) {
405 Value *InVal = MapValue(PN->getIncomingValue(pred), ValueMap);
406 assert(InVal && "Unknown input value?");
407 PN->setIncomingValue(pred, InVal);
408 PN->setIncomingBlock(pred, MappedBlock);
410 PN->removeIncomingValue(pred, false);
411 --pred, --e; // Revisit the next entry.
416 // The loop above has removed PHI entries for those blocks that are dead
417 // and has updated others. However, if a block is live (i.e. copied over)
418 // but its terminator has been changed to not go to this block, then our
419 // phi nodes will have invalid entries. Update the PHI nodes in this
421 PHINode *PN = cast<PHINode>(NewBB->begin());
422 NumPreds = std::distance(pred_begin(NewBB), pred_end(NewBB));
423 if (NumPreds != PN->getNumIncomingValues()) {
424 assert(NumPreds < PN->getNumIncomingValues());
425 // Count how many times each predecessor comes to this block.
426 std::map<BasicBlock*, unsigned> PredCount;
427 for (pred_iterator PI = pred_begin(NewBB), E = pred_end(NewBB);
431 // Figure out how many entries to remove from each PHI.
432 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
433 ++PredCount[PN->getIncomingBlock(i)];
435 // At this point, the excess predecessor entries are positive in the
436 // map. Loop over all of the PHIs and remove excess predecessor
438 BasicBlock::iterator I = NewBB->begin();
439 for (; (PN = dyn_cast<PHINode>(I)); ++I) {
440 for (std::map<BasicBlock*, unsigned>::iterator PCI =PredCount.begin(),
441 E = PredCount.end(); PCI != E; ++PCI) {
442 BasicBlock *Pred = PCI->first;
443 for (unsigned NumToRemove = PCI->second; NumToRemove; --NumToRemove)
444 PN->removeIncomingValue(Pred, false);
449 // If the loops above have made these phi nodes have 0 or 1 operand,
450 // replace them with undef or the input value. We must do this for
451 // correctness, because 0-operand phis are not valid.
452 PN = cast<PHINode>(NewBB->begin());
453 if (PN->getNumIncomingValues() == 0) {
454 BasicBlock::iterator I = NewBB->begin();
455 BasicBlock::const_iterator OldI = OldBB->begin();
456 while ((PN = dyn_cast<PHINode>(I++))) {
457 Value *NV = UndefValue::get(PN->getType());
458 PN->replaceAllUsesWith(NV);
459 assert(ValueMap[OldI] == PN && "ValueMap mismatch");
461 PN->eraseFromParent();
465 // NOTE: We cannot eliminate single entry phi nodes here, because of
466 // ValueMap. Single entry phi nodes can have multiple ValueMap entries
467 // pointing at them. Thus, deleting one would require scanning the ValueMap
468 // to update any entries in it that would require that. This would be
472 // Now that the inlined function body has been fully constructed, go through
473 // and zap unconditional fall-through branches. This happen all the time when
474 // specializing code: code specialization turns conditional branches into
475 // uncond branches, and this code folds them.
476 Function::iterator I = cast<BasicBlock>(ValueMap[&OldFunc->getEntryBlock()]);
477 while (I != NewFunc->end()) {
478 BranchInst *BI = dyn_cast<BranchInst>(I->getTerminator());
479 if (!BI || BI->isConditional()) { ++I; continue; }
481 // Note that we can't eliminate uncond branches if the destination has
482 // single-entry PHI nodes. Eliminating the single-entry phi nodes would
483 // require scanning the ValueMap to update any entries that point to the phi
485 BasicBlock *Dest = BI->getSuccessor(0);
486 if (!Dest->getSinglePredecessor() || isa<PHINode>(Dest->begin())) {
490 // We know all single-entry PHI nodes in the inlined function have been
491 // removed, so we just need to splice the blocks.
492 BI->eraseFromParent();
494 // Move all the instructions in the succ to the pred.
495 I->getInstList().splice(I->end(), Dest->getInstList());
497 // Make all PHI nodes that referred to Dest now refer to I as their source.
498 Dest->replaceAllUsesWith(I);
500 // Remove the dest block.
501 Dest->eraseFromParent();
503 // Do not increment I, iteratively merge all things this block branches to.