1 //===- CloneFunction.cpp - Clone a function into another function ---------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the CloneFunctionInto interface, which is used as the
10 // low-level function cloner. This is used by the CloneFunction and function
11 // inliner to do the dirty work of copying the body of a function around.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/Analysis/ConstantFolding.h"
18 #include "llvm/Analysis/DomTreeUpdater.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/IR/CFG.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DebugInfo.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/GlobalVariable.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/IR/MDBuilder.h"
31 #include "llvm/IR/Metadata.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
34 #include "llvm/Transforms/Utils/Cloning.h"
35 #include "llvm/Transforms/Utils/Local.h"
36 #include "llvm/Transforms/Utils/ValueMapper.h"
37 #include <map>
38 using namespace llvm;
39
40 #define DEBUG_TYPE "clone-function"
41
42 /// See comments in Cloning.h.
CloneBasicBlock(const BasicBlock * BB,ValueToValueMapTy & VMap,const Twine & NameSuffix,Function * F,ClonedCodeInfo * CodeInfo,DebugInfoFinder * DIFinder)43 BasicBlock *llvm::CloneBasicBlock(const BasicBlock *BB, ValueToValueMapTy &VMap,
44 const Twine &NameSuffix, Function *F,
45 ClonedCodeInfo *CodeInfo,
46 DebugInfoFinder *DIFinder) {
47 DenseMap<const MDNode *, MDNode *> Cache;
48 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "", F);
49 if (BB->hasName())
50 NewBB->setName(BB->getName() + NameSuffix);
51
52 bool hasCalls = false, hasDynamicAllocas = false;
53 Module *TheModule = F ? F->getParent() : nullptr;
54
55 // Loop over all instructions, and copy them over.
56 for (const Instruction &I : *BB) {
57 if (DIFinder && TheModule)
58 DIFinder->processInstruction(*TheModule, I);
59
60 Instruction *NewInst = I.clone();
61 if (I.hasName())
62 NewInst->setName(I.getName() + NameSuffix);
63 NewBB->getInstList().push_back(NewInst);
64 VMap[&I] = NewInst; // Add instruction map to value.
65
66 hasCalls |= (isa<CallInst>(I) && !isa<DbgInfoIntrinsic>(I));
67 if (const AllocaInst *AI = dyn_cast<AllocaInst>(&I)) {
68 if (!AI->isStaticAlloca()) {
69 hasDynamicAllocas = true;
70 }
71 }
72 }
73
74 if (CodeInfo) {
75 CodeInfo->ContainsCalls |= hasCalls;
76 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
77 }
78 return NewBB;
79 }
80
81 // Clone OldFunc into NewFunc, transforming the old arguments into references to
82 // VMap values.
83 //
CloneFunctionInto(Function * NewFunc,const Function * OldFunc,ValueToValueMapTy & VMap,bool ModuleLevelChanges,SmallVectorImpl<ReturnInst * > & Returns,const char * NameSuffix,ClonedCodeInfo * CodeInfo,ValueMapTypeRemapper * TypeMapper,ValueMaterializer * Materializer)84 void llvm::CloneFunctionInto(Function *NewFunc, const Function *OldFunc,
85 ValueToValueMapTy &VMap,
86 bool ModuleLevelChanges,
87 SmallVectorImpl<ReturnInst*> &Returns,
88 const char *NameSuffix, ClonedCodeInfo *CodeInfo,
89 ValueMapTypeRemapper *TypeMapper,
90 ValueMaterializer *Materializer) {
91 assert(NameSuffix && "NameSuffix cannot be null!");
92
93 #ifndef NDEBUG
94 for (const Argument &I : OldFunc->args())
95 assert(VMap.count(&I) && "No mapping from source argument specified!");
96 #endif
97
98 // Copy all attributes other than those stored in the AttributeList. We need
99 // to remap the parameter indices of the AttributeList.
100 AttributeList NewAttrs = NewFunc->getAttributes();
101 NewFunc->copyAttributesFrom(OldFunc);
102 NewFunc->setAttributes(NewAttrs);
103
104 // Fix up the personality function that got copied over.
105 if (OldFunc->hasPersonalityFn())
106 NewFunc->setPersonalityFn(
107 MapValue(OldFunc->getPersonalityFn(), VMap,
108 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges,
109 TypeMapper, Materializer));
110
111 SmallVector<AttributeSet, 4> NewArgAttrs(NewFunc->arg_size());
112 AttributeList OldAttrs = OldFunc->getAttributes();
113
114 // Clone any argument attributes that are present in the VMap.
115 for (const Argument &OldArg : OldFunc->args()) {
116 if (Argument *NewArg = dyn_cast<Argument>(VMap[&OldArg])) {
117 NewArgAttrs[NewArg->getArgNo()] =
118 OldAttrs.getParamAttributes(OldArg.getArgNo());
119 }
120 }
121
122 NewFunc->setAttributes(
123 AttributeList::get(NewFunc->getContext(), OldAttrs.getFnAttributes(),
124 OldAttrs.getRetAttributes(), NewArgAttrs));
125
126 bool MustCloneSP =
127 OldFunc->getParent() && OldFunc->getParent() == NewFunc->getParent();
128 DISubprogram *SP = OldFunc->getSubprogram();
129 if (SP) {
130 assert(!MustCloneSP || ModuleLevelChanges);
131 // Add mappings for some DebugInfo nodes that we don't want duplicated
132 // even if they're distinct.
133 auto &MD = VMap.MD();
134 MD[SP->getUnit()].reset(SP->getUnit());
135 MD[SP->getType()].reset(SP->getType());
136 MD[SP->getFile()].reset(SP->getFile());
137 // If we're not cloning into the same module, no need to clone the
138 // subprogram
139 if (!MustCloneSP)
140 MD[SP].reset(SP);
141 }
142
143 // Everything else beyond this point deals with function instructions,
144 // so if we are dealing with a function declaration, we're done.
145 if (OldFunc->isDeclaration())
146 return;
147
148 // When we remap instructions, we want to avoid duplicating inlined
149 // DISubprograms, so record all subprograms we find as we duplicate
150 // instructions and then freeze them in the MD map.
151 // We also record information about dbg.value and dbg.declare to avoid
152 // duplicating the types.
153 DebugInfoFinder DIFinder;
154
155 // Loop over all of the basic blocks in the function, cloning them as
156 // appropriate. Note that we save BE this way in order to handle cloning of
157 // recursive functions into themselves.
158 for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
159 BI != BE; ++BI) {
160 const BasicBlock &BB = *BI;
161
162 // Create a new basic block and copy instructions into it!
163 BasicBlock *CBB = CloneBasicBlock(&BB, VMap, NameSuffix, NewFunc, CodeInfo,
164 ModuleLevelChanges ? &DIFinder : nullptr);
165
166 // Add basic block mapping.
167 VMap[&BB] = CBB;
168
169 // It is only legal to clone a function if a block address within that
170 // function is never referenced outside of the function. Given that, we
171 // want to map block addresses from the old function to block addresses in
172 // the clone. (This is different from the generic ValueMapper
173 // implementation, which generates an invalid blockaddress when
174 // cloning a function.)
175 if (BB.hasAddressTaken()) {
176 Constant *OldBBAddr = BlockAddress::get(const_cast<Function*>(OldFunc),
177 const_cast<BasicBlock*>(&BB));
178 VMap[OldBBAddr] = BlockAddress::get(NewFunc, CBB);
179 }
180
181 // Note return instructions for the caller.
182 if (ReturnInst *RI = dyn_cast<ReturnInst>(CBB->getTerminator()))
183 Returns.push_back(RI);
184 }
185
186 for (DISubprogram *ISP : DIFinder.subprograms())
187 if (ISP != SP)
188 VMap.MD()[ISP].reset(ISP);
189
190 for (DICompileUnit *CU : DIFinder.compile_units())
191 VMap.MD()[CU].reset(CU);
192
193 for (DIType *Type : DIFinder.types())
194 VMap.MD()[Type].reset(Type);
195
196 // Duplicate the metadata that is attached to the cloned function.
197 // Subprograms/CUs/types that were already mapped to themselves won't be
198 // duplicated.
199 SmallVector<std::pair<unsigned, MDNode *>, 1> MDs;
200 OldFunc->getAllMetadata(MDs);
201 for (auto MD : MDs) {
202 NewFunc->addMetadata(
203 MD.first,
204 *MapMetadata(MD.second, VMap,
205 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges,
206 TypeMapper, Materializer));
207 }
208
209 // Loop over all of the instructions in the function, fixing up operand
210 // references as we go. This uses VMap to do all the hard work.
211 for (Function::iterator BB =
212 cast<BasicBlock>(VMap[&OldFunc->front()])->getIterator(),
213 BE = NewFunc->end();
214 BB != BE; ++BB)
215 // Loop over all instructions, fixing each one as we find it...
216 for (Instruction &II : *BB)
217 RemapInstruction(&II, VMap,
218 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges,
219 TypeMapper, Materializer);
220
221 // Register all DICompileUnits of the old parent module in the new parent module
222 auto* OldModule = OldFunc->getParent();
223 auto* NewModule = NewFunc->getParent();
224 if (OldModule && NewModule && OldModule != NewModule && DIFinder.compile_unit_count()) {
225 auto* NMD = NewModule->getOrInsertNamedMetadata("llvm.dbg.cu");
226 // Avoid multiple insertions of the same DICompileUnit to NMD.
227 SmallPtrSet<const void*, 8> Visited;
228 for (auto* Operand : NMD->operands())
229 Visited.insert(Operand);
230 for (auto* Unit : DIFinder.compile_units())
231 // VMap.MD()[Unit] == Unit
232 if (Visited.insert(Unit).second)
233 NMD->addOperand(Unit);
234 }
235 }
236
237 /// Return a copy of the specified function and add it to that function's
238 /// module. Also, any references specified in the VMap are changed to refer to
239 /// their mapped value instead of the original one. If any of the arguments to
240 /// the function are in the VMap, the arguments are deleted from the resultant
241 /// function. The VMap is updated to include mappings from all of the
242 /// instructions and basicblocks in the function from their old to new values.
243 ///
CloneFunction(Function * F,ValueToValueMapTy & VMap,ClonedCodeInfo * CodeInfo)244 Function *llvm::CloneFunction(Function *F, ValueToValueMapTy &VMap,
245 ClonedCodeInfo *CodeInfo) {
246 std::vector<Type*> ArgTypes;
247
248 // The user might be deleting arguments to the function by specifying them in
249 // the VMap. If so, we need to not add the arguments to the arg ty vector
250 //
251 for (const Argument &I : F->args())
252 if (VMap.count(&I) == 0) // Haven't mapped the argument to anything yet?
253 ArgTypes.push_back(I.getType());
254
255 // Create a new function type...
256 FunctionType *FTy = FunctionType::get(F->getFunctionType()->getReturnType(),
257 ArgTypes, F->getFunctionType()->isVarArg());
258
259 // Create the new function...
260 Function *NewF = Function::Create(FTy, F->getLinkage(), F->getAddressSpace(),
261 F->getName(), F->getParent());
262
263 // Loop over the arguments, copying the names of the mapped arguments over...
264 Function::arg_iterator DestI = NewF->arg_begin();
265 for (const Argument & I : F->args())
266 if (VMap.count(&I) == 0) { // Is this argument preserved?
267 DestI->setName(I.getName()); // Copy the name over...
268 VMap[&I] = &*DestI++; // Add mapping to VMap
269 }
270
271 SmallVector<ReturnInst*, 8> Returns; // Ignore returns cloned.
272 CloneFunctionInto(NewF, F, VMap, F->getSubprogram() != nullptr, Returns, "",
273 CodeInfo);
274
275 return NewF;
276 }
277
278
279
280 namespace {
281 /// This is a private class used to implement CloneAndPruneFunctionInto.
282 struct PruningFunctionCloner {
283 Function *NewFunc;
284 const Function *OldFunc;
285 ValueToValueMapTy &VMap;
286 bool ModuleLevelChanges;
287 const char *NameSuffix;
288 ClonedCodeInfo *CodeInfo;
289
290 public:
PruningFunctionCloner__anon4f261df10111::PruningFunctionCloner291 PruningFunctionCloner(Function *newFunc, const Function *oldFunc,
292 ValueToValueMapTy &valueMap, bool moduleLevelChanges,
293 const char *nameSuffix, ClonedCodeInfo *codeInfo)
294 : NewFunc(newFunc), OldFunc(oldFunc), VMap(valueMap),
295 ModuleLevelChanges(moduleLevelChanges), NameSuffix(nameSuffix),
296 CodeInfo(codeInfo) {}
297
298 /// The specified block is found to be reachable, clone it and
299 /// anything that it can reach.
300 void CloneBlock(const BasicBlock *BB,
301 BasicBlock::const_iterator StartingInst,
302 std::vector<const BasicBlock*> &ToClone);
303 };
304 }
305
306 /// The specified block is found to be reachable, clone it and
307 /// anything that it can reach.
CloneBlock(const BasicBlock * BB,BasicBlock::const_iterator StartingInst,std::vector<const BasicBlock * > & ToClone)308 void PruningFunctionCloner::CloneBlock(const BasicBlock *BB,
309 BasicBlock::const_iterator StartingInst,
310 std::vector<const BasicBlock*> &ToClone){
311 WeakTrackingVH &BBEntry = VMap[BB];
312
313 // Have we already cloned this block?
314 if (BBEntry) return;
315
316 // Nope, clone it now.
317 BasicBlock *NewBB;
318 BBEntry = NewBB = BasicBlock::Create(BB->getContext());
319 if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix);
320
321 // It is only legal to clone a function if a block address within that
322 // function is never referenced outside of the function. Given that, we
323 // want to map block addresses from the old function to block addresses in
324 // the clone. (This is different from the generic ValueMapper
325 // implementation, which generates an invalid blockaddress when
326 // cloning a function.)
327 //
328 // Note that we don't need to fix the mapping for unreachable blocks;
329 // the default mapping there is safe.
330 if (BB->hasAddressTaken()) {
331 Constant *OldBBAddr = BlockAddress::get(const_cast<Function*>(OldFunc),
332 const_cast<BasicBlock*>(BB));
333 VMap[OldBBAddr] = BlockAddress::get(NewFunc, NewBB);
334 }
335
336 bool hasCalls = false, hasDynamicAllocas = false, hasStaticAllocas = false;
337
338 // Loop over all instructions, and copy them over, DCE'ing as we go. This
339 // loop doesn't include the terminator.
340 for (BasicBlock::const_iterator II = StartingInst, IE = --BB->end();
341 II != IE; ++II) {
342
343 Instruction *NewInst = II->clone();
344
345 // Eagerly remap operands to the newly cloned instruction, except for PHI
346 // nodes for which we defer processing until we update the CFG.
347 if (!isa<PHINode>(NewInst)) {
348 RemapInstruction(NewInst, VMap,
349 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges);
350
351 // If we can simplify this instruction to some other value, simply add
352 // a mapping to that value rather than inserting a new instruction into
353 // the basic block.
354 if (Value *V =
355 SimplifyInstruction(NewInst, BB->getModule()->getDataLayout())) {
356 // On the off-chance that this simplifies to an instruction in the old
357 // function, map it back into the new function.
358 if (NewFunc != OldFunc)
359 if (Value *MappedV = VMap.lookup(V))
360 V = MappedV;
361
362 if (!NewInst->mayHaveSideEffects()) {
363 VMap[&*II] = V;
364 NewInst->deleteValue();
365 continue;
366 }
367 }
368 }
369
370 if (II->hasName())
371 NewInst->setName(II->getName()+NameSuffix);
372 VMap[&*II] = NewInst; // Add instruction map to value.
373 NewBB->getInstList().push_back(NewInst);
374 hasCalls |= (isa<CallInst>(II) && !isa<DbgInfoIntrinsic>(II));
375
376 if (CodeInfo)
377 if (auto *CB = dyn_cast<CallBase>(&*II))
378 if (CB->hasOperandBundles())
379 CodeInfo->OperandBundleCallSites.push_back(NewInst);
380
381 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
382 if (isa<ConstantInt>(AI->getArraySize()))
383 hasStaticAllocas = true;
384 else
385 hasDynamicAllocas = true;
386 }
387 }
388
389 // Finally, clone over the terminator.
390 const Instruction *OldTI = BB->getTerminator();
391 bool TerminatorDone = false;
392 if (const BranchInst *BI = dyn_cast<BranchInst>(OldTI)) {
393 if (BI->isConditional()) {
394 // If the condition was a known constant in the callee...
395 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
396 // Or is a known constant in the caller...
397 if (!Cond) {
398 Value *V = VMap.lookup(BI->getCondition());
399 Cond = dyn_cast_or_null<ConstantInt>(V);
400 }
401
402 // Constant fold to uncond branch!
403 if (Cond) {
404 BasicBlock *Dest = BI->getSuccessor(!Cond->getZExtValue());
405 VMap[OldTI] = BranchInst::Create(Dest, NewBB);
406 ToClone.push_back(Dest);
407 TerminatorDone = true;
408 }
409 }
410 } else if (const SwitchInst *SI = dyn_cast<SwitchInst>(OldTI)) {
411 // If switching on a value known constant in the caller.
412 ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition());
413 if (!Cond) { // Or known constant after constant prop in the callee...
414 Value *V = VMap.lookup(SI->getCondition());
415 Cond = dyn_cast_or_null<ConstantInt>(V);
416 }
417 if (Cond) { // Constant fold to uncond branch!
418 SwitchInst::ConstCaseHandle Case = *SI->findCaseValue(Cond);
419 BasicBlock *Dest = const_cast<BasicBlock*>(Case.getCaseSuccessor());
420 VMap[OldTI] = BranchInst::Create(Dest, NewBB);
421 ToClone.push_back(Dest);
422 TerminatorDone = true;
423 }
424 }
425
426 if (!TerminatorDone) {
427 Instruction *NewInst = OldTI->clone();
428 if (OldTI->hasName())
429 NewInst->setName(OldTI->getName()+NameSuffix);
430 NewBB->getInstList().push_back(NewInst);
431 VMap[OldTI] = NewInst; // Add instruction map to value.
432
433 if (CodeInfo)
434 if (auto *CB = dyn_cast<CallBase>(OldTI))
435 if (CB->hasOperandBundles())
436 CodeInfo->OperandBundleCallSites.push_back(NewInst);
437
438 // Recursively clone any reachable successor blocks.
439 append_range(ToClone, successors(BB->getTerminator()));
440 }
441
442 if (CodeInfo) {
443 CodeInfo->ContainsCalls |= hasCalls;
444 CodeInfo->ContainsDynamicAllocas |= hasDynamicAllocas;
445 CodeInfo->ContainsDynamicAllocas |= hasStaticAllocas &&
446 BB != &BB->getParent()->front();
447 }
448 }
449
450 /// This works like CloneAndPruneFunctionInto, except that it does not clone the
451 /// entire function. Instead it starts at an instruction provided by the caller
452 /// and copies (and prunes) only the code reachable from that instruction.
CloneAndPruneIntoFromInst(Function * NewFunc,const Function * OldFunc,const Instruction * StartingInst,ValueToValueMapTy & VMap,bool ModuleLevelChanges,SmallVectorImpl<ReturnInst * > & Returns,const char * NameSuffix,ClonedCodeInfo * CodeInfo)453 void llvm::CloneAndPruneIntoFromInst(Function *NewFunc, const Function *OldFunc,
454 const Instruction *StartingInst,
455 ValueToValueMapTy &VMap,
456 bool ModuleLevelChanges,
457 SmallVectorImpl<ReturnInst *> &Returns,
458 const char *NameSuffix,
459 ClonedCodeInfo *CodeInfo) {
460 assert(NameSuffix && "NameSuffix cannot be null!");
461
462 ValueMapTypeRemapper *TypeMapper = nullptr;
463 ValueMaterializer *Materializer = nullptr;
464
465 #ifndef NDEBUG
466 // If the cloning starts at the beginning of the function, verify that
467 // the function arguments are mapped.
468 if (!StartingInst)
469 for (const Argument &II : OldFunc->args())
470 assert(VMap.count(&II) && "No mapping from source argument specified!");
471 #endif
472
473 PruningFunctionCloner PFC(NewFunc, OldFunc, VMap, ModuleLevelChanges,
474 NameSuffix, CodeInfo);
475 const BasicBlock *StartingBB;
476 if (StartingInst)
477 StartingBB = StartingInst->getParent();
478 else {
479 StartingBB = &OldFunc->getEntryBlock();
480 StartingInst = &StartingBB->front();
481 }
482
483 // Clone the entry block, and anything recursively reachable from it.
484 std::vector<const BasicBlock*> CloneWorklist;
485 PFC.CloneBlock(StartingBB, StartingInst->getIterator(), CloneWorklist);
486 while (!CloneWorklist.empty()) {
487 const BasicBlock *BB = CloneWorklist.back();
488 CloneWorklist.pop_back();
489 PFC.CloneBlock(BB, BB->begin(), CloneWorklist);
490 }
491
492 // Loop over all of the basic blocks in the old function. If the block was
493 // reachable, we have cloned it and the old block is now in the value map:
494 // insert it into the new function in the right order. If not, ignore it.
495 //
496 // Defer PHI resolution until rest of function is resolved.
497 SmallVector<const PHINode*, 16> PHIToResolve;
498 for (const BasicBlock &BI : *OldFunc) {
499 Value *V = VMap.lookup(&BI);
500 BasicBlock *NewBB = cast_or_null<BasicBlock>(V);
501 if (!NewBB) continue; // Dead block.
502
503 // Add the new block to the new function.
504 NewFunc->getBasicBlockList().push_back(NewBB);
505
506 // Handle PHI nodes specially, as we have to remove references to dead
507 // blocks.
508 for (const PHINode &PN : BI.phis()) {
509 // PHI nodes may have been remapped to non-PHI nodes by the caller or
510 // during the cloning process.
511 if (isa<PHINode>(VMap[&PN]))
512 PHIToResolve.push_back(&PN);
513 else
514 break;
515 }
516
517 // Finally, remap the terminator instructions, as those can't be remapped
518 // until all BBs are mapped.
519 RemapInstruction(NewBB->getTerminator(), VMap,
520 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges,
521 TypeMapper, Materializer);
522 }
523
524 // Defer PHI resolution until rest of function is resolved, PHI resolution
525 // requires the CFG to be up-to-date.
526 for (unsigned phino = 0, e = PHIToResolve.size(); phino != e; ) {
527 const PHINode *OPN = PHIToResolve[phino];
528 unsigned NumPreds = OPN->getNumIncomingValues();
529 const BasicBlock *OldBB = OPN->getParent();
530 BasicBlock *NewBB = cast<BasicBlock>(VMap[OldBB]);
531
532 // Map operands for blocks that are live and remove operands for blocks
533 // that are dead.
534 for (; phino != PHIToResolve.size() &&
535 PHIToResolve[phino]->getParent() == OldBB; ++phino) {
536 OPN = PHIToResolve[phino];
537 PHINode *PN = cast<PHINode>(VMap[OPN]);
538 for (unsigned pred = 0, e = NumPreds; pred != e; ++pred) {
539 Value *V = VMap.lookup(PN->getIncomingBlock(pred));
540 if (BasicBlock *MappedBlock = cast_or_null<BasicBlock>(V)) {
541 Value *InVal = MapValue(PN->getIncomingValue(pred),
542 VMap,
543 ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges);
544 assert(InVal && "Unknown input value?");
545 PN->setIncomingValue(pred, InVal);
546 PN->setIncomingBlock(pred, MappedBlock);
547 } else {
548 PN->removeIncomingValue(pred, false);
549 --pred; // Revisit the next entry.
550 --e;
551 }
552 }
553 }
554
555 // The loop above has removed PHI entries for those blocks that are dead
556 // and has updated others. However, if a block is live (i.e. copied over)
557 // but its terminator has been changed to not go to this block, then our
558 // phi nodes will have invalid entries. Update the PHI nodes in this
559 // case.
560 PHINode *PN = cast<PHINode>(NewBB->begin());
561 NumPreds = pred_size(NewBB);
562 if (NumPreds != PN->getNumIncomingValues()) {
563 assert(NumPreds < PN->getNumIncomingValues());
564 // Count how many times each predecessor comes to this block.
565 std::map<BasicBlock*, unsigned> PredCount;
566 for (pred_iterator PI = pred_begin(NewBB), E = pred_end(NewBB);
567 PI != E; ++PI)
568 --PredCount[*PI];
569
570 // Figure out how many entries to remove from each PHI.
571 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
572 ++PredCount[PN->getIncomingBlock(i)];
573
574 // At this point, the excess predecessor entries are positive in the
575 // map. Loop over all of the PHIs and remove excess predecessor
576 // entries.
577 BasicBlock::iterator I = NewBB->begin();
578 for (; (PN = dyn_cast<PHINode>(I)); ++I) {
579 for (const auto &PCI : PredCount) {
580 BasicBlock *Pred = PCI.first;
581 for (unsigned NumToRemove = PCI.second; NumToRemove; --NumToRemove)
582 PN->removeIncomingValue(Pred, false);
583 }
584 }
585 }
586
587 // If the loops above have made these phi nodes have 0 or 1 operand,
588 // replace them with undef or the input value. We must do this for
589 // correctness, because 0-operand phis are not valid.
590 PN = cast<PHINode>(NewBB->begin());
591 if (PN->getNumIncomingValues() == 0) {
592 BasicBlock::iterator I = NewBB->begin();
593 BasicBlock::const_iterator OldI = OldBB->begin();
594 while ((PN = dyn_cast<PHINode>(I++))) {
595 Value *NV = UndefValue::get(PN->getType());
596 PN->replaceAllUsesWith(NV);
597 assert(VMap[&*OldI] == PN && "VMap mismatch");
598 VMap[&*OldI] = NV;
599 PN->eraseFromParent();
600 ++OldI;
601 }
602 }
603 }
604
605 // Make a second pass over the PHINodes now that all of them have been
606 // remapped into the new function, simplifying the PHINode and performing any
607 // recursive simplifications exposed. This will transparently update the
608 // WeakTrackingVH in the VMap. Notably, we rely on that so that if we coalesce
609 // two PHINodes, the iteration over the old PHIs remains valid, and the
610 // mapping will just map us to the new node (which may not even be a PHI
611 // node).
612 const DataLayout &DL = NewFunc->getParent()->getDataLayout();
613 SmallSetVector<const Value *, 8> Worklist;
614 for (unsigned Idx = 0, Size = PHIToResolve.size(); Idx != Size; ++Idx)
615 if (isa<PHINode>(VMap[PHIToResolve[Idx]]))
616 Worklist.insert(PHIToResolve[Idx]);
617
618 // Note that we must test the size on each iteration, the worklist can grow.
619 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
620 const Value *OrigV = Worklist[Idx];
621 auto *I = dyn_cast_or_null<Instruction>(VMap.lookup(OrigV));
622 if (!I)
623 continue;
624
625 // Skip over non-intrinsic callsites, we don't want to remove any nodes from
626 // the CGSCC.
627 CallBase *CB = dyn_cast<CallBase>(I);
628 if (CB && CB->getCalledFunction() &&
629 !CB->getCalledFunction()->isIntrinsic())
630 continue;
631
632 // See if this instruction simplifies.
633 Value *SimpleV = SimplifyInstruction(I, DL);
634 if (!SimpleV)
635 continue;
636
637 // Stash away all the uses of the old instruction so we can check them for
638 // recursive simplifications after a RAUW. This is cheaper than checking all
639 // uses of To on the recursive step in most cases.
640 for (const User *U : OrigV->users())
641 Worklist.insert(cast<Instruction>(U));
642
643 // Replace the instruction with its simplified value.
644 I->replaceAllUsesWith(SimpleV);
645
646 // If the original instruction had no side effects, remove it.
647 if (isInstructionTriviallyDead(I))
648 I->eraseFromParent();
649 else
650 VMap[OrigV] = I;
651 }
652
653 // Now that the inlined function body has been fully constructed, go through
654 // and zap unconditional fall-through branches. This happens all the time when
655 // specializing code: code specialization turns conditional branches into
656 // uncond branches, and this code folds them.
657 Function::iterator Begin = cast<BasicBlock>(VMap[StartingBB])->getIterator();
658 Function::iterator I = Begin;
659 while (I != NewFunc->end()) {
660 // We need to simplify conditional branches and switches with a constant
661 // operand. We try to prune these out when cloning, but if the
662 // simplification required looking through PHI nodes, those are only
663 // available after forming the full basic block. That may leave some here,
664 // and we still want to prune the dead code as early as possible.
665 //
666 // Do the folding before we check if the block is dead since we want code
667 // like
668 // bb:
669 // br i1 undef, label %bb, label %bb
670 // to be simplified to
671 // bb:
672 // br label %bb
673 // before we call I->getSinglePredecessor().
674 ConstantFoldTerminator(&*I);
675
676 // Check if this block has become dead during inlining or other
677 // simplifications. Note that the first block will appear dead, as it has
678 // not yet been wired up properly.
679 if (I != Begin && (pred_empty(&*I) || I->getSinglePredecessor() == &*I)) {
680 BasicBlock *DeadBB = &*I++;
681 DeleteDeadBlock(DeadBB);
682 continue;
683 }
684
685 BranchInst *BI = dyn_cast<BranchInst>(I->getTerminator());
686 if (!BI || BI->isConditional()) { ++I; continue; }
687
688 BasicBlock *Dest = BI->getSuccessor(0);
689 if (!Dest->getSinglePredecessor()) {
690 ++I; continue;
691 }
692
693 // We shouldn't be able to get single-entry PHI nodes here, as instsimplify
694 // above should have zapped all of them..
695 assert(!isa<PHINode>(Dest->begin()));
696
697 // We know all single-entry PHI nodes in the inlined function have been
698 // removed, so we just need to splice the blocks.
699 BI->eraseFromParent();
700
701 // Make all PHI nodes that referred to Dest now refer to I as their source.
702 Dest->replaceAllUsesWith(&*I);
703
704 // Move all the instructions in the succ to the pred.
705 I->getInstList().splice(I->end(), Dest->getInstList());
706
707 // Remove the dest block.
708 Dest->eraseFromParent();
709
710 // Do not increment I, iteratively merge all things this block branches to.
711 }
712
713 // Make a final pass over the basic blocks from the old function to gather
714 // any return instructions which survived folding. We have to do this here
715 // because we can iteratively remove and merge returns above.
716 for (Function::iterator I = cast<BasicBlock>(VMap[StartingBB])->getIterator(),
717 E = NewFunc->end();
718 I != E; ++I)
719 if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator()))
720 Returns.push_back(RI);
721 }
722
723
724 /// This works exactly like CloneFunctionInto,
725 /// except that it does some simple constant prop and DCE on the fly. The
726 /// effect of this is to copy significantly less code in cases where (for
727 /// example) a function call with constant arguments is inlined, and those
728 /// constant arguments cause a significant amount of code in the callee to be
729 /// dead. Since this doesn't produce an exact copy of the input, it can't be
730 /// used for things like CloneFunction or CloneModule.
CloneAndPruneFunctionInto(Function * NewFunc,const Function * OldFunc,ValueToValueMapTy & VMap,bool ModuleLevelChanges,SmallVectorImpl<ReturnInst * > & Returns,const char * NameSuffix,ClonedCodeInfo * CodeInfo,Instruction * TheCall)731 void llvm::CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc,
732 ValueToValueMapTy &VMap,
733 bool ModuleLevelChanges,
734 SmallVectorImpl<ReturnInst*> &Returns,
735 const char *NameSuffix,
736 ClonedCodeInfo *CodeInfo,
737 Instruction *TheCall) {
738 CloneAndPruneIntoFromInst(NewFunc, OldFunc, &OldFunc->front().front(), VMap,
739 ModuleLevelChanges, Returns, NameSuffix, CodeInfo);
740 }
741
742 /// Remaps instructions in \p Blocks using the mapping in \p VMap.
remapInstructionsInBlocks(const SmallVectorImpl<BasicBlock * > & Blocks,ValueToValueMapTy & VMap)743 void llvm::remapInstructionsInBlocks(
744 const SmallVectorImpl<BasicBlock *> &Blocks, ValueToValueMapTy &VMap) {
745 // Rewrite the code to refer to itself.
746 for (auto *BB : Blocks)
747 for (auto &Inst : *BB)
748 RemapInstruction(&Inst, VMap,
749 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
750 }
751
752 /// Clones a loop \p OrigLoop. Returns the loop and the blocks in \p
753 /// Blocks.
754 ///
755 /// Updates LoopInfo and DominatorTree assuming the loop is dominated by block
756 /// \p LoopDomBB. Insert the new blocks before block specified in \p Before.
cloneLoopWithPreheader(BasicBlock * Before,BasicBlock * LoopDomBB,Loop * OrigLoop,ValueToValueMapTy & VMap,const Twine & NameSuffix,LoopInfo * LI,DominatorTree * DT,SmallVectorImpl<BasicBlock * > & Blocks)757 Loop *llvm::cloneLoopWithPreheader(BasicBlock *Before, BasicBlock *LoopDomBB,
758 Loop *OrigLoop, ValueToValueMapTy &VMap,
759 const Twine &NameSuffix, LoopInfo *LI,
760 DominatorTree *DT,
761 SmallVectorImpl<BasicBlock *> &Blocks) {
762 Function *F = OrigLoop->getHeader()->getParent();
763 Loop *ParentLoop = OrigLoop->getParentLoop();
764 DenseMap<Loop *, Loop *> LMap;
765
766 Loop *NewLoop = LI->AllocateLoop();
767 LMap[OrigLoop] = NewLoop;
768 if (ParentLoop)
769 ParentLoop->addChildLoop(NewLoop);
770 else
771 LI->addTopLevelLoop(NewLoop);
772
773 BasicBlock *OrigPH = OrigLoop->getLoopPreheader();
774 assert(OrigPH && "No preheader");
775 BasicBlock *NewPH = CloneBasicBlock(OrigPH, VMap, NameSuffix, F);
776 // To rename the loop PHIs.
777 VMap[OrigPH] = NewPH;
778 Blocks.push_back(NewPH);
779
780 // Update LoopInfo.
781 if (ParentLoop)
782 ParentLoop->addBasicBlockToLoop(NewPH, *LI);
783
784 // Update DominatorTree.
785 DT->addNewBlock(NewPH, LoopDomBB);
786
787 for (Loop *CurLoop : OrigLoop->getLoopsInPreorder()) {
788 Loop *&NewLoop = LMap[CurLoop];
789 if (!NewLoop) {
790 NewLoop = LI->AllocateLoop();
791
792 // Establish the parent/child relationship.
793 Loop *OrigParent = CurLoop->getParentLoop();
794 assert(OrigParent && "Could not find the original parent loop");
795 Loop *NewParentLoop = LMap[OrigParent];
796 assert(NewParentLoop && "Could not find the new parent loop");
797
798 NewParentLoop->addChildLoop(NewLoop);
799 }
800 }
801
802 for (BasicBlock *BB : OrigLoop->getBlocks()) {
803 Loop *CurLoop = LI->getLoopFor(BB);
804 Loop *&NewLoop = LMap[CurLoop];
805 assert(NewLoop && "Expecting new loop to be allocated");
806
807 BasicBlock *NewBB = CloneBasicBlock(BB, VMap, NameSuffix, F);
808 VMap[BB] = NewBB;
809
810 // Update LoopInfo.
811 NewLoop->addBasicBlockToLoop(NewBB, *LI);
812
813 // Add DominatorTree node. After seeing all blocks, update to correct
814 // IDom.
815 DT->addNewBlock(NewBB, NewPH);
816
817 Blocks.push_back(NewBB);
818 }
819
820 for (BasicBlock *BB : OrigLoop->getBlocks()) {
821 // Update loop headers.
822 Loop *CurLoop = LI->getLoopFor(BB);
823 if (BB == CurLoop->getHeader())
824 LMap[CurLoop]->moveToHeader(cast<BasicBlock>(VMap[BB]));
825
826 // Update DominatorTree.
827 BasicBlock *IDomBB = DT->getNode(BB)->getIDom()->getBlock();
828 DT->changeImmediateDominator(cast<BasicBlock>(VMap[BB]),
829 cast<BasicBlock>(VMap[IDomBB]));
830 }
831
832 // Move them physically from the end of the block list.
833 F->getBasicBlockList().splice(Before->getIterator(), F->getBasicBlockList(),
834 NewPH);
835 F->getBasicBlockList().splice(Before->getIterator(), F->getBasicBlockList(),
836 NewLoop->getHeader()->getIterator(), F->end());
837
838 return NewLoop;
839 }
840
841 /// Duplicate non-Phi instructions from the beginning of block up to
842 /// StopAt instruction into a split block between BB and its predecessor.
DuplicateInstructionsInSplitBetween(BasicBlock * BB,BasicBlock * PredBB,Instruction * StopAt,ValueToValueMapTy & ValueMapping,DomTreeUpdater & DTU)843 BasicBlock *llvm::DuplicateInstructionsInSplitBetween(
844 BasicBlock *BB, BasicBlock *PredBB, Instruction *StopAt,
845 ValueToValueMapTy &ValueMapping, DomTreeUpdater &DTU) {
846
847 assert(count(successors(PredBB), BB) == 1 &&
848 "There must be a single edge between PredBB and BB!");
849 // We are going to have to map operands from the original BB block to the new
850 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
851 // account for entry from PredBB.
852 BasicBlock::iterator BI = BB->begin();
853 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
854 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
855
856 BasicBlock *NewBB = SplitEdge(PredBB, BB);
857 NewBB->setName(PredBB->getName() + ".split");
858 Instruction *NewTerm = NewBB->getTerminator();
859
860 // FIXME: SplitEdge does not yet take a DTU, so we include the split edge
861 // in the update set here.
862 DTU.applyUpdates({{DominatorTree::Delete, PredBB, BB},
863 {DominatorTree::Insert, PredBB, NewBB},
864 {DominatorTree::Insert, NewBB, BB}});
865
866 // Clone the non-phi instructions of BB into NewBB, keeping track of the
867 // mapping and using it to remap operands in the cloned instructions.
868 // Stop once we see the terminator too. This covers the case where BB's
869 // terminator gets replaced and StopAt == BB's terminator.
870 for (; StopAt != &*BI && BB->getTerminator() != &*BI; ++BI) {
871 Instruction *New = BI->clone();
872 New->setName(BI->getName());
873 New->insertBefore(NewTerm);
874 ValueMapping[&*BI] = New;
875
876 // Remap operands to patch up intra-block references.
877 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
878 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
879 auto I = ValueMapping.find(Inst);
880 if (I != ValueMapping.end())
881 New->setOperand(i, I->second);
882 }
883 }
884
885 return NewBB;
886 }
887
cloneNoAliasScopes(ArrayRef<MDNode * > NoAliasDeclScopes,DenseMap<MDNode *,MDNode * > & ClonedScopes,StringRef Ext,LLVMContext & Context)888 void llvm::cloneNoAliasScopes(
889 ArrayRef<MDNode *> NoAliasDeclScopes,
890 DenseMap<MDNode *, MDNode *> &ClonedScopes,
891 StringRef Ext, LLVMContext &Context) {
892 MDBuilder MDB(Context);
893
894 for (auto *ScopeList : NoAliasDeclScopes) {
895 for (auto &MDOperand : ScopeList->operands()) {
896 if (MDNode *MD = dyn_cast<MDNode>(MDOperand)) {
897 AliasScopeNode SNANode(MD);
898
899 std::string Name;
900 auto ScopeName = SNANode.getName();
901 if (!ScopeName.empty())
902 Name = (Twine(ScopeName) + ":" + Ext).str();
903 else
904 Name = std::string(Ext);
905
906 MDNode *NewScope = MDB.createAnonymousAliasScope(
907 const_cast<MDNode *>(SNANode.getDomain()), Name);
908 ClonedScopes.insert(std::make_pair(MD, NewScope));
909 }
910 }
911 }
912 }
913
adaptNoAliasScopes(Instruction * I,const DenseMap<MDNode *,MDNode * > & ClonedScopes,LLVMContext & Context)914 void llvm::adaptNoAliasScopes(
915 Instruction *I, const DenseMap<MDNode *, MDNode *> &ClonedScopes,
916 LLVMContext &Context) {
917 auto CloneScopeList = [&](const MDNode *ScopeList) -> MDNode * {
918 bool NeedsReplacement = false;
919 SmallVector<Metadata *, 8> NewScopeList;
920 for (auto &MDOp : ScopeList->operands()) {
921 if (MDNode *MD = dyn_cast<MDNode>(MDOp)) {
922 if (auto *NewMD = ClonedScopes.lookup(MD)) {
923 NewScopeList.push_back(NewMD);
924 NeedsReplacement = true;
925 continue;
926 }
927 NewScopeList.push_back(MD);
928 }
929 }
930 if (NeedsReplacement)
931 return MDNode::get(Context, NewScopeList);
932 return nullptr;
933 };
934
935 if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(I))
936 if (auto *NewScopeList = CloneScopeList(Decl->getScopeList()))
937 Decl->setScopeList(NewScopeList);
938
939 auto replaceWhenNeeded = [&](unsigned MD_ID) {
940 if (const MDNode *CSNoAlias = I->getMetadata(MD_ID))
941 if (auto *NewScopeList = CloneScopeList(CSNoAlias))
942 I->setMetadata(MD_ID, NewScopeList);
943 };
944 replaceWhenNeeded(LLVMContext::MD_noalias);
945 replaceWhenNeeded(LLVMContext::MD_alias_scope);
946 }
947
cloneAndAdaptNoAliasScopes(ArrayRef<MDNode * > NoAliasDeclScopes,ArrayRef<BasicBlock * > NewBlocks,LLVMContext & Context,StringRef Ext)948 void llvm::cloneAndAdaptNoAliasScopes(
949 ArrayRef<MDNode *> NoAliasDeclScopes,
950 ArrayRef<BasicBlock *> NewBlocks, LLVMContext &Context, StringRef Ext) {
951 if (NoAliasDeclScopes.empty())
952 return;
953
954 DenseMap<MDNode *, MDNode *> ClonedScopes;
955 LLVM_DEBUG(dbgs() << "cloneAndAdaptNoAliasScopes: cloning "
956 << NoAliasDeclScopes.size() << " node(s)\n");
957
958 cloneNoAliasScopes(NoAliasDeclScopes, ClonedScopes, Ext, Context);
959 // Identify instructions using metadata that needs adaptation
960 for (BasicBlock *NewBlock : NewBlocks)
961 for (Instruction &I : *NewBlock)
962 adaptNoAliasScopes(&I, ClonedScopes, Context);
963 }
964
cloneAndAdaptNoAliasScopes(ArrayRef<MDNode * > NoAliasDeclScopes,Instruction * IStart,Instruction * IEnd,LLVMContext & Context,StringRef Ext)965 void llvm::cloneAndAdaptNoAliasScopes(
966 ArrayRef<MDNode *> NoAliasDeclScopes, Instruction *IStart,
967 Instruction *IEnd, LLVMContext &Context, StringRef Ext) {
968 if (NoAliasDeclScopes.empty())
969 return;
970
971 DenseMap<MDNode *, MDNode *> ClonedScopes;
972 LLVM_DEBUG(dbgs() << "cloneAndAdaptNoAliasScopes: cloning "
973 << NoAliasDeclScopes.size() << " node(s)\n");
974
975 cloneNoAliasScopes(NoAliasDeclScopes, ClonedScopes, Ext, Context);
976 // Identify instructions using metadata that needs adaptation
977 assert(IStart->getParent() == IEnd->getParent() && "different basic block ?");
978 auto ItStart = IStart->getIterator();
979 auto ItEnd = IEnd->getIterator();
980 ++ItEnd; // IEnd is included, increment ItEnd to get the end of the range
981 for (auto &I : llvm::make_range(ItStart, ItEnd))
982 adaptNoAliasScopes(&I, ClonedScopes, Context);
983 }
984
identifyNoAliasScopesToClone(ArrayRef<BasicBlock * > BBs,SmallVectorImpl<MDNode * > & NoAliasDeclScopes)985 void llvm::identifyNoAliasScopesToClone(
986 ArrayRef<BasicBlock *> BBs, SmallVectorImpl<MDNode *> &NoAliasDeclScopes) {
987 for (BasicBlock *BB : BBs)
988 for (Instruction &I : *BB)
989 if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
990 NoAliasDeclScopes.push_back(Decl->getScopeList());
991 }
992
identifyNoAliasScopesToClone(BasicBlock::iterator Start,BasicBlock::iterator End,SmallVectorImpl<MDNode * > & NoAliasDeclScopes)993 void llvm::identifyNoAliasScopesToClone(
994 BasicBlock::iterator Start, BasicBlock::iterator End,
995 SmallVectorImpl<MDNode *> &NoAliasDeclScopes) {
996 for (Instruction &I : make_range(Start, End))
997 if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
998 NoAliasDeclScopes.push_back(Decl->getScopeList());
999 }
1000