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/Function.h"
21 #include "llvm/Support/CFG.h"
22 #include "llvm/Support/Compiler.h"
23 #include "llvm/Transforms/Utils/ValueMapper.h"
24 #include "llvm/Analysis/ConstantFolding.h"
25 #include "llvm/ADT/SmallVector.h"
29 // CloneBasicBlock - See comments in Cloning.h
30 BasicBlock *llvm::CloneBasicBlock(const BasicBlock *BB,
31 DenseMap<const Value*, Value*> &ValueMap,
32 const char *NameSuffix, Function *F,
33 ClonedCodeInfo *CodeInfo) {
34 BasicBlock *NewBB = new BasicBlock("", F);
35 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
37 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
39 // Loop over all instructions, and copy them over.
40 for (BasicBlock::const_iterator II = BB->begin(), IE = BB->end();
42 Instruction *NewInst = II->clone();
44 NewInst->setName(II->getName()+NameSuffix);
45 NewBB->getInstList().push_back(NewInst);
46 ValueMap[II] = NewInst; // Add instruction map to value.
48 hasCalls |= isa<CallInst>(II);
49 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
50 if (isa<ConstantInt>(AI->getArraySize()))
51 hasStaticAllocas = true;
53 hasDynamicAllocas = true;
58 CodeInfo->ContainsCalls |= hasCalls;
59 CodeInfo->ContainsUnwinds |= isa<UnwindInst>(BB->getTerminator());
60 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
61 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
62 BB != &BB->getParent()->getEntryBlock();
67 // Clone OldFunc into NewFunc, transforming the old arguments into references to
70 void llvm::CloneFunctionInto(Function *NewFunc, const Function *OldFunc,
71 DenseMap<const Value*, Value*> &ValueMap,
72 std::vector<ReturnInst*> &Returns,
73 const char *NameSuffix, ClonedCodeInfo *CodeInfo) {
74 assert(NameSuffix && "NameSuffix cannot be null!");
77 for (Function::const_arg_iterator I = OldFunc->arg_begin(),
78 E = OldFunc->arg_end(); I != E; ++I)
79 assert(ValueMap.count(I) && "No mapping from source argument specified!");
82 // Clone the parameter attributes
83 NewFunc->setParamAttrs(OldFunc->getParamAttrs());
85 // Loop over all of the basic blocks in the function, cloning them as
86 // appropriate. Note that we save BE this way in order to handle cloning of
87 // recursive functions into themselves.
89 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
91 const BasicBlock &BB = *BI;
93 // Create a new basic block and copy instructions into it!
94 BasicBlock *CBB = CloneBasicBlock(&BB, ValueMap, NameSuffix, NewFunc,
96 ValueMap[&BB] = CBB; // Add basic block mapping.
98 if (ReturnInst *RI = dyn_cast<ReturnInst>(CBB->getTerminator()))
99 Returns.push_back(RI);
102 // Loop over all of the instructions in the function, fixing up operand
103 // references as we go. This uses ValueMap to do all the hard work.
105 for (Function::iterator BB = cast<BasicBlock>(ValueMap[OldFunc->begin()]),
106 BE = NewFunc->end(); BB != BE; ++BB)
107 // Loop over all instructions, fixing each one as we find it...
108 for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II)
109 RemapInstruction(II, ValueMap);
112 /// CloneFunction - Return a copy of the specified function, but without
113 /// embedding the function into another module. Also, any references specified
114 /// in the ValueMap are changed to refer to their mapped value instead of the
115 /// original one. If any of the arguments to the function are in the ValueMap,
116 /// the arguments are deleted from the resultant function. The ValueMap is
117 /// updated to include mappings from all of the instructions and basicblocks in
118 /// the function from their old to new values.
120 Function *llvm::CloneFunction(const Function *F,
121 DenseMap<const Value*, Value*> &ValueMap,
122 ClonedCodeInfo *CodeInfo) {
123 std::vector<const Type*> ArgTypes;
125 // The user might be deleting arguments to the function by specifying them in
126 // the ValueMap. If so, we need to not add the arguments to the arg ty vector
128 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
130 if (ValueMap.count(I) == 0) // Haven't mapped the argument to anything yet?
131 ArgTypes.push_back(I->getType());
133 // Create a new function type...
134 FunctionType *FTy = FunctionType::get(F->getFunctionType()->getReturnType(),
135 ArgTypes, F->getFunctionType()->isVarArg());
137 // Create the new function...
138 Function *NewF = new Function(FTy, F->getLinkage(), F->getName());
140 // Loop over the arguments, copying the names of the mapped arguments over...
141 Function::arg_iterator DestI = NewF->arg_begin();
142 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
144 if (ValueMap.count(I) == 0) { // Is this argument preserved?
145 DestI->setName(I->getName()); // Copy the name over...
146 ValueMap[I] = DestI++; // Add mapping to ValueMap
149 std::vector<ReturnInst*> Returns; // Ignore returns cloned...
150 CloneFunctionInto(NewF, F, ValueMap, Returns, "", CodeInfo);
157 /// PruningFunctionCloner - This class is a private class used to implement
158 /// the CloneAndPruneFunctionInto method.
159 struct VISIBILITY_HIDDEN PruningFunctionCloner {
161 const Function *OldFunc;
162 DenseMap<const Value*, Value*> &ValueMap;
163 std::vector<ReturnInst*> &Returns;
164 const char *NameSuffix;
165 ClonedCodeInfo *CodeInfo;
166 const TargetData *TD;
169 PruningFunctionCloner(Function *newFunc, const Function *oldFunc,
170 DenseMap<const Value*, Value*> &valueMap,
171 std::vector<ReturnInst*> &returns,
172 const char *nameSuffix,
173 ClonedCodeInfo *codeInfo,
174 const TargetData *td)
175 : NewFunc(newFunc), OldFunc(oldFunc), ValueMap(valueMap), Returns(returns),
176 NameSuffix(nameSuffix), CodeInfo(codeInfo), TD(td) {
179 /// CloneBlock - The specified block is found to be reachable, clone it and
180 /// anything that it can reach.
181 void CloneBlock(const BasicBlock *BB,
182 std::vector<const BasicBlock*> &ToClone);
185 /// ConstantFoldMappedInstruction - Constant fold the specified instruction,
186 /// mapping its operands through ValueMap if they are available.
187 Constant *ConstantFoldMappedInstruction(const Instruction *I);
191 /// CloneBlock - The specified block is found to be reachable, clone it and
192 /// anything that it can reach.
193 void PruningFunctionCloner::CloneBlock(const BasicBlock *BB,
194 std::vector<const BasicBlock*> &ToClone){
195 Value *&BBEntry = ValueMap[BB];
197 // Have we already cloned this block?
200 // Nope, clone it now.
202 BBEntry = NewBB = new BasicBlock();
203 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
205 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
207 // Loop over all instructions, and copy them over, DCE'ing as we go. This
208 // loop doesn't include the terminator.
209 for (BasicBlock::const_iterator II = BB->begin(), IE = --BB->end();
211 // If this instruction constant folds, don't bother cloning the instruction,
212 // instead, just add the constant to the value map.
213 if (Constant *C = ConstantFoldMappedInstruction(II)) {
218 Instruction *NewInst = II->clone();
220 NewInst->setName(II->getName()+NameSuffix);
221 NewBB->getInstList().push_back(NewInst);
222 ValueMap[II] = NewInst; // Add instruction map to value.
224 hasCalls |= isa<CallInst>(II);
225 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
226 if (isa<ConstantInt>(AI->getArraySize()))
227 hasStaticAllocas = true;
229 hasDynamicAllocas = true;
233 // Finally, clone over the terminator.
234 const TerminatorInst *OldTI = BB->getTerminator();
235 bool TerminatorDone = false;
236 if (const BranchInst *BI = dyn_cast<BranchInst>(OldTI)) {
237 if (BI->isConditional()) {
238 // If the condition was a known constant in the callee...
239 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
240 // Or is a known constant in the caller...
242 Cond = dyn_cast_or_null<ConstantInt>(ValueMap[BI->getCondition()]);
244 // Constant fold to uncond branch!
246 BasicBlock *Dest = BI->getSuccessor(!Cond->getZExtValue());
247 ValueMap[OldTI] = new BranchInst(Dest, NewBB);
248 ToClone.push_back(Dest);
249 TerminatorDone = true;
252 } else if (const SwitchInst *SI = dyn_cast<SwitchInst>(OldTI)) {
253 // If switching on a value known constant in the caller.
254 ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition());
255 if (Cond == 0) // Or known constant after constant prop in the callee...
256 Cond = dyn_cast_or_null<ConstantInt>(ValueMap[SI->getCondition()]);
257 if (Cond) { // Constant fold to uncond branch!
258 BasicBlock *Dest = SI->getSuccessor(SI->findCaseValue(Cond));
259 ValueMap[OldTI] = new BranchInst(Dest, NewBB);
260 ToClone.push_back(Dest);
261 TerminatorDone = true;
265 if (!TerminatorDone) {
266 Instruction *NewInst = OldTI->clone();
267 if (OldTI->hasName())
268 NewInst->setName(OldTI->getName()+NameSuffix);
269 NewBB->getInstList().push_back(NewInst);
270 ValueMap[OldTI] = NewInst; // Add instruction map to value.
272 // Recursively clone any reachable successor blocks.
273 const TerminatorInst *TI = BB->getTerminator();
274 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
275 ToClone.push_back(TI->getSuccessor(i));
279 CodeInfo->ContainsCalls |= hasCalls;
280 CodeInfo->ContainsUnwinds |= isa<UnwindInst>(OldTI);
281 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
282 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
283 BB != &BB->getParent()->front();
286 if (ReturnInst *RI = dyn_cast<ReturnInst>(NewBB->getTerminator()))
287 Returns.push_back(RI);
290 /// ConstantFoldMappedInstruction - Constant fold the specified instruction,
291 /// mapping its operands through ValueMap if they are available.
292 Constant *PruningFunctionCloner::
293 ConstantFoldMappedInstruction(const Instruction *I) {
294 SmallVector<Constant*, 8> Ops;
295 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
296 if (Constant *Op = dyn_cast_or_null<Constant>(MapValue(I->getOperand(i),
300 return 0; // All operands not constant!
303 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
304 return ConstantFoldCompareInstOperands(CI->getPredicate(),
305 &Ops[0], Ops.size(), TD);
307 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
308 &Ops[0], Ops.size(), TD);
311 /// CloneAndPruneFunctionInto - This works exactly like CloneFunctionInto,
312 /// except that it does some simple constant prop and DCE on the fly. The
313 /// effect of this is to copy significantly less code in cases where (for
314 /// example) a function call with constant arguments is inlined, and those
315 /// constant arguments cause a significant amount of code in the callee to be
316 /// dead. Since this doesn't produce an exact copy of the input, it can't be
317 /// used for things like CloneFunction or CloneModule.
318 void llvm::CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc,
319 DenseMap<const Value*, Value*> &ValueMap,
320 std::vector<ReturnInst*> &Returns,
321 const char *NameSuffix,
322 ClonedCodeInfo *CodeInfo,
323 const TargetData *TD) {
324 assert(NameSuffix && "NameSuffix cannot be null!");
327 for (Function::const_arg_iterator II = OldFunc->arg_begin(),
328 E = OldFunc->arg_end(); II != E; ++II)
329 assert(ValueMap.count(II) && "No mapping from source argument specified!");
332 PruningFunctionCloner PFC(NewFunc, OldFunc, ValueMap, Returns,
333 NameSuffix, CodeInfo, TD);
335 // Clone the entry block, and anything recursively reachable from it.
336 std::vector<const BasicBlock*> CloneWorklist;
337 CloneWorklist.push_back(&OldFunc->getEntryBlock());
338 while (!CloneWorklist.empty()) {
339 const BasicBlock *BB = CloneWorklist.back();
340 CloneWorklist.pop_back();
341 PFC.CloneBlock(BB, CloneWorklist);
344 // Loop over all of the basic blocks in the old function. If the block was
345 // reachable, we have cloned it and the old block is now in the value map:
346 // insert it into the new function in the right order. If not, ignore it.
348 // Defer PHI resolution until rest of function is resolved.
349 std::vector<const PHINode*> PHIToResolve;
350 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
352 BasicBlock *NewBB = cast_or_null<BasicBlock>(ValueMap[BI]);
353 if (NewBB == 0) continue; // Dead block.
355 // Add the new block to the new function.
356 NewFunc->getBasicBlockList().push_back(NewBB);
358 // Loop over all of the instructions in the block, fixing up operand
359 // references as we go. This uses ValueMap to do all the hard work.
361 BasicBlock::iterator I = NewBB->begin();
363 // Handle PHI nodes specially, as we have to remove references to dead
365 if (PHINode *PN = dyn_cast<PHINode>(I)) {
366 // Skip over all PHI nodes, remembering them for later.
367 BasicBlock::const_iterator OldI = BI->begin();
368 for (; (PN = dyn_cast<PHINode>(I)); ++I, ++OldI)
369 PHIToResolve.push_back(cast<PHINode>(OldI));
372 // Otherwise, remap the rest of the instructions normally.
373 for (; I != NewBB->end(); ++I)
374 RemapInstruction(I, ValueMap);
377 // Defer PHI resolution until rest of function is resolved, PHI resolution
378 // requires the CFG to be up-to-date.
379 for (unsigned phino = 0, e = PHIToResolve.size(); phino != e; ) {
380 const PHINode *OPN = PHIToResolve[phino];
381 unsigned NumPreds = OPN->getNumIncomingValues();
382 const BasicBlock *OldBB = OPN->getParent();
383 BasicBlock *NewBB = cast<BasicBlock>(ValueMap[OldBB]);
385 // Map operands for blocks that are live and remove operands for blocks
387 for (; phino != PHIToResolve.size() &&
388 PHIToResolve[phino]->getParent() == OldBB; ++phino) {
389 OPN = PHIToResolve[phino];
390 PHINode *PN = cast<PHINode>(ValueMap[OPN]);
391 for (unsigned pred = 0, e = NumPreds; pred != e; ++pred) {
392 if (BasicBlock *MappedBlock =
393 cast_or_null<BasicBlock>(ValueMap[PN->getIncomingBlock(pred)])) {
394 Value *InVal = MapValue(PN->getIncomingValue(pred), ValueMap);
395 assert(InVal && "Unknown input value?");
396 PN->setIncomingValue(pred, InVal);
397 PN->setIncomingBlock(pred, MappedBlock);
399 PN->removeIncomingValue(pred, false);
400 --pred, --e; // Revisit the next entry.
405 // The loop above has removed PHI entries for those blocks that are dead
406 // and has updated others. However, if a block is live (i.e. copied over)
407 // but its terminator has been changed to not go to this block, then our
408 // phi nodes will have invalid entries. Update the PHI nodes in this
410 PHINode *PN = cast<PHINode>(NewBB->begin());
411 NumPreds = std::distance(pred_begin(NewBB), pred_end(NewBB));
412 if (NumPreds != PN->getNumIncomingValues()) {
413 assert(NumPreds < PN->getNumIncomingValues());
414 // Count how many times each predecessor comes to this block.
415 std::map<BasicBlock*, unsigned> PredCount;
416 for (pred_iterator PI = pred_begin(NewBB), E = pred_end(NewBB);
420 // Figure out how many entries to remove from each PHI.
421 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
422 ++PredCount[PN->getIncomingBlock(i)];
424 // At this point, the excess predecessor entries are positive in the
425 // map. Loop over all of the PHIs and remove excess predecessor
427 BasicBlock::iterator I = NewBB->begin();
428 for (; (PN = dyn_cast<PHINode>(I)); ++I) {
429 for (std::map<BasicBlock*, unsigned>::iterator PCI =PredCount.begin(),
430 E = PredCount.end(); PCI != E; ++PCI) {
431 BasicBlock *Pred = PCI->first;
432 for (unsigned NumToRemove = PCI->second; NumToRemove; --NumToRemove)
433 PN->removeIncomingValue(Pred, false);
438 // If the loops above have made these phi nodes have 0 or 1 operand,
439 // replace them with undef or the input value. We must do this for
440 // correctness, because 0-operand phis are not valid.
441 PN = cast<PHINode>(NewBB->begin());
442 if (PN->getNumIncomingValues() == 0) {
443 BasicBlock::iterator I = NewBB->begin();
444 BasicBlock::const_iterator OldI = OldBB->begin();
445 while ((PN = dyn_cast<PHINode>(I++))) {
446 Value *NV = UndefValue::get(PN->getType());
447 PN->replaceAllUsesWith(NV);
448 assert(ValueMap[OldI] == PN && "ValueMap mismatch");
450 PN->eraseFromParent();
454 // NOTE: We cannot eliminate single entry phi nodes here, because of
455 // ValueMap. Single entry phi nodes can have multiple ValueMap entries
456 // pointing at them. Thus, deleting one would require scanning the ValueMap
457 // to update any entries in it that would require that. This would be
461 // Now that the inlined function body has been fully constructed, go through
462 // and zap unconditional fall-through branches. This happen all the time when
463 // specializing code: code specialization turns conditional branches into
464 // uncond branches, and this code folds them.
465 Function::iterator I = cast<BasicBlock>(ValueMap[&OldFunc->getEntryBlock()]);
466 while (I != NewFunc->end()) {
467 BranchInst *BI = dyn_cast<BranchInst>(I->getTerminator());
468 if (!BI || BI->isConditional()) { ++I; continue; }
470 // Note that we can't eliminate uncond branches if the destination has
471 // single-entry PHI nodes. Eliminating the single-entry phi nodes would
472 // require scanning the ValueMap to update any entries that point to the phi
474 BasicBlock *Dest = BI->getSuccessor(0);
475 if (!Dest->getSinglePredecessor() || isa<PHINode>(Dest->begin())) {
479 // We know all single-entry PHI nodes in the inlined function have been
480 // removed, so we just need to splice the blocks.
481 BI->eraseFromParent();
483 // Move all the instructions in the succ to the pred.
484 I->getInstList().splice(I->end(), Dest->getInstList());
486 // Make all PHI nodes that referred to Dest now refer to I as their source.
487 Dest->replaceAllUsesWith(I);
489 // Remove the dest block.
490 Dest->eraseFromParent();
492 // Do not increment I, iteratively merge all things this block branches to.