73861866956df2ce1c37a0f3ea88f4a8717fd06b
1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements inlining of a function into a call site, resolving
11 // parameters and the return value as appropriate.
12 //
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/Cloning.h"
16 #include "llvm/ADT/SmallSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/CallGraph.h"
23 #include "llvm/Analysis/CaptureTracking.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/Attributes.h"
27 #include "llvm/IR/CallSite.h"
28 #include "llvm/IR/CFG.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DebugInfo.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Dominators.h"
34 #include "llvm/IR/IRBuilder.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/Intrinsics.h"
38 #include "llvm/IR/MDBuilder.h"
39 #include "llvm/IR/Module.h"
40 #include "llvm/Transforms/Utils/Local.h"
41 #include "llvm/Support/CommandLine.h"
42 #include <algorithm>
43 using namespace llvm;
45 static cl::opt<bool>
46 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
47 cl::Hidden,
48 cl::desc("Convert noalias attributes to metadata during inlining."));
50 static cl::opt<bool>
51 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
52 cl::init(true), cl::Hidden,
53 cl::desc("Convert align attributes to assumptions during inlining."));
55 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
56 bool InsertLifetime) {
57 return InlineFunction(CallSite(CI), IFI, InsertLifetime);
58 }
59 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
60 bool InsertLifetime) {
61 return InlineFunction(CallSite(II), IFI, InsertLifetime);
62 }
64 namespace {
65 /// A class for recording information about inlining through an invoke.
66 class InvokeInliningInfo {
67 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
68 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
69 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke.
70 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts.
71 SmallVector<Value*, 8> UnwindDestPHIValues;
73 public:
74 InvokeInliningInfo(InvokeInst *II)
75 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
76 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
77 // If there are PHI nodes in the unwind destination block, we need to keep
78 // track of which values came into them from the invoke before removing
79 // the edge from this block.
80 llvm::BasicBlock *InvokeBB = II->getParent();
81 BasicBlock::iterator I = OuterResumeDest->begin();
82 for (; isa<PHINode>(I); ++I) {
83 // Save the value to use for this edge.
84 PHINode *PHI = cast<PHINode>(I);
85 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
86 }
88 CallerLPad = cast<LandingPadInst>(I);
89 }
91 /// getOuterResumeDest - The outer unwind destination is the target of
92 /// unwind edges introduced for calls within the inlined function.
93 BasicBlock *getOuterResumeDest() const {
94 return OuterResumeDest;
95 }
97 BasicBlock *getInnerResumeDest();
99 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
101 /// forwardResume - Forward the 'resume' instruction to the caller's landing
102 /// pad block. When the landing pad block has only one predecessor, this is
103 /// a simple branch. When there is more than one predecessor, we need to
104 /// split the landing pad block after the landingpad instruction and jump
105 /// to there.
106 void forwardResume(ResumeInst *RI,
107 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads);
109 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
110 /// destination block for the given basic block, using the values for the
111 /// original invoke's source block.
112 void addIncomingPHIValuesFor(BasicBlock *BB) const {
113 addIncomingPHIValuesForInto(BB, OuterResumeDest);
114 }
116 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
117 BasicBlock::iterator I = dest->begin();
118 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
119 PHINode *phi = cast<PHINode>(I);
120 phi->addIncoming(UnwindDestPHIValues[i], src);
121 }
122 }
123 };
124 }
126 /// getInnerResumeDest - Get or create a target for the branch from ResumeInsts.
127 BasicBlock *InvokeInliningInfo::getInnerResumeDest() {
128 if (InnerResumeDest) return InnerResumeDest;
130 // Split the landing pad.
131 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
132 InnerResumeDest =
133 OuterResumeDest->splitBasicBlock(SplitPoint,
134 OuterResumeDest->getName() + ".body");
136 // The number of incoming edges we expect to the inner landing pad.
137 const unsigned PHICapacity = 2;
139 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
140 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
141 BasicBlock::iterator I = OuterResumeDest->begin();
142 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
143 PHINode *OuterPHI = cast<PHINode>(I);
144 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
145 OuterPHI->getName() + ".lpad-body",
146 InsertPoint);
147 OuterPHI->replaceAllUsesWith(InnerPHI);
148 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
149 }
151 // Create a PHI for the exception values.
152 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
153 "eh.lpad-body", InsertPoint);
154 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
155 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
157 // All done.
158 return InnerResumeDest;
159 }
161 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
162 /// block. When the landing pad block has only one predecessor, this is a simple
163 /// branch. When there is more than one predecessor, we need to split the
164 /// landing pad block after the landingpad instruction and jump to there.
165 void InvokeInliningInfo::forwardResume(ResumeInst *RI,
166 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads) {
167 BasicBlock *Dest = getInnerResumeDest();
168 BasicBlock *Src = RI->getParent();
170 BranchInst::Create(Dest, Src);
172 // Update the PHIs in the destination. They were inserted in an order which
173 // makes this work.
174 addIncomingPHIValuesForInto(Src, Dest);
176 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
177 RI->eraseFromParent();
178 }
180 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
181 /// an invoke, we have to turn all of the calls that can throw into
182 /// invokes. This function analyze BB to see if there are any calls, and if so,
183 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
184 /// nodes in that block with the values specified in InvokeDestPHIValues.
185 static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
186 InvokeInliningInfo &Invoke) {
187 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
188 Instruction *I = BBI++;
190 // We only need to check for function calls: inlined invoke
191 // instructions require no special handling.
192 CallInst *CI = dyn_cast<CallInst>(I);
194 // If this call cannot unwind, don't convert it to an invoke.
195 // Inline asm calls cannot throw.
196 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
197 continue;
199 // Convert this function call into an invoke instruction. First, split the
200 // basic block.
201 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
203 // Delete the unconditional branch inserted by splitBasicBlock
204 BB->getInstList().pop_back();
206 // Create the new invoke instruction.
207 ImmutableCallSite CS(CI);
208 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
209 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
210 Invoke.getOuterResumeDest(),
211 InvokeArgs, CI->getName(), BB);
212 II->setDebugLoc(CI->getDebugLoc());
213 II->setCallingConv(CI->getCallingConv());
214 II->setAttributes(CI->getAttributes());
216 // Make sure that anything using the call now uses the invoke! This also
217 // updates the CallGraph if present, because it uses a WeakVH.
218 CI->replaceAllUsesWith(II);
220 // Delete the original call
221 Split->getInstList().pop_front();
223 // Update any PHI nodes in the exceptional block to indicate that there is
224 // now a new entry in them.
225 Invoke.addIncomingPHIValuesFor(BB);
226 return;
227 }
228 }
230 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
231 /// in the body of the inlined function into invokes.
232 ///
233 /// II is the invoke instruction being inlined. FirstNewBlock is the first
234 /// block of the inlined code (the last block is the end of the function),
235 /// and InlineCodeInfo is information about the code that got inlined.
236 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
237 ClonedCodeInfo &InlinedCodeInfo) {
238 BasicBlock *InvokeDest = II->getUnwindDest();
240 Function *Caller = FirstNewBlock->getParent();
242 // The inlined code is currently at the end of the function, scan from the
243 // start of the inlined code to its end, checking for stuff we need to
244 // rewrite.
245 InvokeInliningInfo Invoke(II);
247 // Get all of the inlined landing pad instructions.
248 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
249 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I)
250 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
251 InlinedLPads.insert(II->getLandingPadInst());
253 // Append the clauses from the outer landing pad instruction into the inlined
254 // landing pad instructions.
255 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
256 for (LandingPadInst *InlinedLPad : InlinedLPads) {
257 unsigned OuterNum = OuterLPad->getNumClauses();
258 InlinedLPad->reserveClauses(OuterNum);
259 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
260 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
261 if (OuterLPad->isCleanup())
262 InlinedLPad->setCleanup(true);
263 }
265 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
266 if (InlinedCodeInfo.ContainsCalls)
267 HandleCallsInBlockInlinedThroughInvoke(BB, Invoke);
269 // Forward any resumes that are remaining here.
270 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
271 Invoke.forwardResume(RI, InlinedLPads);
272 }
274 // Now that everything is happy, we have one final detail. The PHI nodes in
275 // the exception destination block still have entries due to the original
276 // invoke instruction. Eliminate these entries (which might even delete the
277 // PHI node) now.
278 InvokeDest->removePredecessor(II->getParent());
279 }
281 /// CloneAliasScopeMetadata - When inlining a function that contains noalias
282 /// scope metadata, this metadata needs to be cloned so that the inlined blocks
283 /// have different "unqiue scopes" at every call site. Were this not done, then
284 /// aliasing scopes from a function inlined into a caller multiple times could
285 /// not be differentiated (and this would lead to miscompiles because the
286 /// non-aliasing property communicated by the metadata could have
287 /// call-site-specific control dependencies).
288 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
289 const Function *CalledFunc = CS.getCalledFunction();
290 SetVector<const MDNode *> MD;
292 // Note: We could only clone the metadata if it is already used in the
293 // caller. I'm omitting that check here because it might confuse
294 // inter-procedural alias analysis passes. We can revisit this if it becomes
295 // an efficiency or overhead problem.
297 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
298 I != IE; ++I)
299 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
300 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
301 MD.insert(M);
302 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
303 MD.insert(M);
304 }
306 if (MD.empty())
307 return;
309 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
310 // the set.
311 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
312 while (!Queue.empty()) {
313 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
314 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
315 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
316 if (MD.insert(M1))
317 Queue.push_back(M1);
318 }
320 // Now we have a complete set of all metadata in the chains used to specify
321 // the noalias scopes and the lists of those scopes.
322 SmallVector<TempMDTuple, 16> DummyNodes;
323 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
324 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
325 I != IE; ++I) {
326 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
327 MDMap[*I].reset(DummyNodes.back().get());
328 }
330 // Create new metadata nodes to replace the dummy nodes, replacing old
331 // metadata references with either a dummy node or an already-created new
332 // node.
333 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
334 I != IE; ++I) {
335 SmallVector<Metadata *, 4> NewOps;
336 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
337 const Metadata *V = (*I)->getOperand(i);
338 if (const MDNode *M = dyn_cast<MDNode>(V))
339 NewOps.push_back(MDMap[M]);
340 else
341 NewOps.push_back(const_cast<Metadata *>(V));
342 }
344 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
345 MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
346 assert(TempM->isTemporary() && "Expected temporary node");
348 TempM->replaceAllUsesWith(NewM);
349 }
351 // Now replace the metadata in the new inlined instructions with the
352 // repacements from the map.
353 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
354 VMI != VMIE; ++VMI) {
355 if (!VMI->second)
356 continue;
358 Instruction *NI = dyn_cast<Instruction>(VMI->second);
359 if (!NI)
360 continue;
362 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
363 MDNode *NewMD = MDMap[M];
364 // If the call site also had alias scope metadata (a list of scopes to
365 // which instructions inside it might belong), propagate those scopes to
366 // the inlined instructions.
367 if (MDNode *CSM =
368 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
369 NewMD = MDNode::concatenate(NewMD, CSM);
370 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
371 } else if (NI->mayReadOrWriteMemory()) {
372 if (MDNode *M =
373 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
374 NI->setMetadata(LLVMContext::MD_alias_scope, M);
375 }
377 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
378 MDNode *NewMD = MDMap[M];
379 // If the call site also had noalias metadata (a list of scopes with
380 // which instructions inside it don't alias), propagate those scopes to
381 // the inlined instructions.
382 if (MDNode *CSM =
383 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
384 NewMD = MDNode::concatenate(NewMD, CSM);
385 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
386 } else if (NI->mayReadOrWriteMemory()) {
387 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
388 NI->setMetadata(LLVMContext::MD_noalias, M);
389 }
390 }
391 }
393 /// AddAliasScopeMetadata - If the inlined function has noalias arguments, then
394 /// add new alias scopes for each noalias argument, tag the mapped noalias
395 /// parameters with noalias metadata specifying the new scope, and tag all
396 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
397 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
398 const DataLayout *DL, AliasAnalysis *AA) {
399 if (!EnableNoAliasConversion)
400 return;
402 const Function *CalledFunc = CS.getCalledFunction();
403 SmallVector<const Argument *, 4> NoAliasArgs;
405 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
406 E = CalledFunc->arg_end(); I != E; ++I) {
407 if (I->hasNoAliasAttr() && !I->hasNUses(0))
408 NoAliasArgs.push_back(I);
409 }
411 if (NoAliasArgs.empty())
412 return;
414 // To do a good job, if a noalias variable is captured, we need to know if
415 // the capture point dominates the particular use we're considering.
416 DominatorTree DT;
417 DT.recalculate(const_cast<Function&>(*CalledFunc));
419 // noalias indicates that pointer values based on the argument do not alias
420 // pointer values which are not based on it. So we add a new "scope" for each
421 // noalias function argument. Accesses using pointers based on that argument
422 // become part of that alias scope, accesses using pointers not based on that
423 // argument are tagged as noalias with that scope.
425 DenseMap<const Argument *, MDNode *> NewScopes;
426 MDBuilder MDB(CalledFunc->getContext());
428 // Create a new scope domain for this function.
429 MDNode *NewDomain =
430 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
431 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
432 const Argument *A = NoAliasArgs[i];
434 std::string Name = CalledFunc->getName();
435 if (A->hasName()) {
436 Name += ": %";
437 Name += A->getName();
438 } else {
439 Name += ": argument ";
440 Name += utostr(i);
441 }
443 // Note: We always create a new anonymous root here. This is true regardless
444 // of the linkage of the callee because the aliasing "scope" is not just a
445 // property of the callee, but also all control dependencies in the caller.
446 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
447 NewScopes.insert(std::make_pair(A, NewScope));
448 }
450 // Iterate over all new instructions in the map; for all memory-access
451 // instructions, add the alias scope metadata.
452 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
453 VMI != VMIE; ++VMI) {
454 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
455 if (!VMI->second)
456 continue;
458 Instruction *NI = dyn_cast<Instruction>(VMI->second);
459 if (!NI)
460 continue;
462 bool IsArgMemOnlyCall = false, IsFuncCall = false;
463 SmallVector<const Value *, 2> PtrArgs;
465 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
466 PtrArgs.push_back(LI->getPointerOperand());
467 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
468 PtrArgs.push_back(SI->getPointerOperand());
469 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
470 PtrArgs.push_back(VAAI->getPointerOperand());
471 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
472 PtrArgs.push_back(CXI->getPointerOperand());
473 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
474 PtrArgs.push_back(RMWI->getPointerOperand());
475 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
476 // If we know that the call does not access memory, then we'll still
477 // know that about the inlined clone of this call site, and we don't
478 // need to add metadata.
479 if (ICS.doesNotAccessMemory())
480 continue;
482 IsFuncCall = true;
483 if (AA) {
484 AliasAnalysis::ModRefBehavior MRB = AA->getModRefBehavior(ICS);
485 if (MRB == AliasAnalysis::OnlyAccessesArgumentPointees ||
486 MRB == AliasAnalysis::OnlyReadsArgumentPointees)
487 IsArgMemOnlyCall = true;
488 }
490 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
491 AE = ICS.arg_end(); AI != AE; ++AI) {
492 // We need to check the underlying objects of all arguments, not just
493 // the pointer arguments, because we might be passing pointers as
494 // integers, etc.
495 // However, if we know that the call only accesses pointer arguments,
496 // then we only need to check the pointer arguments.
497 if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
498 continue;
500 PtrArgs.push_back(*AI);
501 }
502 }
504 // If we found no pointers, then this instruction is not suitable for
505 // pairing with an instruction to receive aliasing metadata.
506 // However, if this is a call, this we might just alias with none of the
507 // noalias arguments.
508 if (PtrArgs.empty() && !IsFuncCall)
509 continue;
511 // It is possible that there is only one underlying object, but you
512 // need to go through several PHIs to see it, and thus could be
513 // repeated in the Objects list.
514 SmallPtrSet<const Value *, 4> ObjSet;
515 SmallVector<Metadata *, 4> Scopes, NoAliases;
517 SmallSetVector<const Argument *, 4> NAPtrArgs;
518 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
519 SmallVector<Value *, 4> Objects;
520 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
521 Objects, DL, /* MaxLookup = */ 0);
523 for (Value *O : Objects)
524 ObjSet.insert(O);
525 }
527 // Figure out if we're derived from anything that is not a noalias
528 // argument.
529 bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
530 for (const Value *V : ObjSet) {
531 // Is this value a constant that cannot be derived from any pointer
532 // value (we need to exclude constant expressions, for example, that
533 // are formed from arithmetic on global symbols).
534 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
535 isa<ConstantPointerNull>(V) ||
536 isa<ConstantDataVector>(V) || isa<UndefValue>(V);
537 if (IsNonPtrConst)
538 continue;
540 // If this is anything other than a noalias argument, then we cannot
541 // completely describe the aliasing properties using alias.scope
542 // metadata (and, thus, won't add any).
543 if (const Argument *A = dyn_cast<Argument>(V)) {
544 if (!A->hasNoAliasAttr())
545 UsesAliasingPtr = true;
546 } else {
547 UsesAliasingPtr = true;
548 }
550 // If this is not some identified function-local object (which cannot
551 // directly alias a noalias argument), or some other argument (which,
552 // by definition, also cannot alias a noalias argument), then we could
553 // alias a noalias argument that has been captured).
554 if (!isa<Argument>(V) &&
555 !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
556 CanDeriveViaCapture = true;
557 }
559 // A function call can always get captured noalias pointers (via other
560 // parameters, globals, etc.).
561 if (IsFuncCall && !IsArgMemOnlyCall)
562 CanDeriveViaCapture = true;
564 // First, we want to figure out all of the sets with which we definitely
565 // don't alias. Iterate over all noalias set, and add those for which:
566 // 1. The noalias argument is not in the set of objects from which we
567 // definitely derive.
568 // 2. The noalias argument has not yet been captured.
569 // An arbitrary function that might load pointers could see captured
570 // noalias arguments via other noalias arguments or globals, and so we
571 // must always check for prior capture.
572 for (const Argument *A : NoAliasArgs) {
573 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
574 // It might be tempting to skip the
575 // PointerMayBeCapturedBefore check if
576 // A->hasNoCaptureAttr() is true, but this is
577 // incorrect because nocapture only guarantees
578 // that no copies outlive the function, not
579 // that the value cannot be locally captured.
580 !PointerMayBeCapturedBefore(A,
581 /* ReturnCaptures */ false,
582 /* StoreCaptures */ false, I, &DT)))
583 NoAliases.push_back(NewScopes[A]);
584 }
586 if (!NoAliases.empty())
587 NI->setMetadata(LLVMContext::MD_noalias,
588 MDNode::concatenate(
589 NI->getMetadata(LLVMContext::MD_noalias),
590 MDNode::get(CalledFunc->getContext(), NoAliases)));
592 // Next, we want to figure out all of the sets to which we might belong.
593 // We might belong to a set if the noalias argument is in the set of
594 // underlying objects. If there is some non-noalias argument in our list
595 // of underlying objects, then we cannot add a scope because the fact
596 // that some access does not alias with any set of our noalias arguments
597 // cannot itself guarantee that it does not alias with this access
598 // (because there is some pointer of unknown origin involved and the
599 // other access might also depend on this pointer). We also cannot add
600 // scopes to arbitrary functions unless we know they don't access any
601 // non-parameter pointer-values.
602 bool CanAddScopes = !UsesAliasingPtr;
603 if (CanAddScopes && IsFuncCall)
604 CanAddScopes = IsArgMemOnlyCall;
606 if (CanAddScopes)
607 for (const Argument *A : NoAliasArgs) {
608 if (ObjSet.count(A))
609 Scopes.push_back(NewScopes[A]);
610 }
612 if (!Scopes.empty())
613 NI->setMetadata(
614 LLVMContext::MD_alias_scope,
615 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
616 MDNode::get(CalledFunc->getContext(), Scopes)));
617 }
618 }
619 }
621 /// If the inlined function has non-byval align arguments, then
622 /// add @llvm.assume-based alignment assumptions to preserve this information.
623 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
624 if (!PreserveAlignmentAssumptions || !IFI.DL)
625 return;
627 // To avoid inserting redundant assumptions, we should check for assumptions
628 // already in the caller. To do this, we might need a DT of the caller.
629 DominatorTree DT;
630 bool DTCalculated = false;
632 Function *CalledFunc = CS.getCalledFunction();
633 for (Function::arg_iterator I = CalledFunc->arg_begin(),
634 E = CalledFunc->arg_end();
635 I != E; ++I) {
636 unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
637 if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
638 if (!DTCalculated) {
639 DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
640 ->getParent()));
641 DTCalculated = true;
642 }
644 // If we can already prove the asserted alignment in the context of the
645 // caller, then don't bother inserting the assumption.
646 Value *Arg = CS.getArgument(I->getArgNo());
647 if (getKnownAlignment(Arg, IFI.DL,
648 &IFI.ACT->getAssumptionCache(*CalledFunc),
649 CS.getInstruction(), &DT) >= Align)
650 continue;
652 IRBuilder<>(CS.getInstruction()).CreateAlignmentAssumption(*IFI.DL, Arg,
653 Align);
654 }
655 }
656 }
658 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
659 /// into the caller, update the specified callgraph to reflect the changes we
660 /// made. Note that it's possible that not all code was copied over, so only
661 /// some edges of the callgraph may remain.
662 static void UpdateCallGraphAfterInlining(CallSite CS,
663 Function::iterator FirstNewBlock,
664 ValueToValueMapTy &VMap,
665 InlineFunctionInfo &IFI) {
666 CallGraph &CG = *IFI.CG;
667 const Function *Caller = CS.getInstruction()->getParent()->getParent();
668 const Function *Callee = CS.getCalledFunction();
669 CallGraphNode *CalleeNode = CG[Callee];
670 CallGraphNode *CallerNode = CG[Caller];
672 // Since we inlined some uninlined call sites in the callee into the caller,
673 // add edges from the caller to all of the callees of the callee.
674 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
676 // Consider the case where CalleeNode == CallerNode.
677 CallGraphNode::CalledFunctionsVector CallCache;
678 if (CalleeNode == CallerNode) {
679 CallCache.assign(I, E);
680 I = CallCache.begin();
681 E = CallCache.end();
682 }
684 for (; I != E; ++I) {
685 const Value *OrigCall = I->first;
687 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
688 // Only copy the edge if the call was inlined!
689 if (VMI == VMap.end() || VMI->second == nullptr)
690 continue;
692 // If the call was inlined, but then constant folded, there is no edge to
693 // add. Check for this case.
694 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
695 if (!NewCall) continue;
697 // Remember that this call site got inlined for the client of
698 // InlineFunction.
699 IFI.InlinedCalls.push_back(NewCall);
701 // It's possible that inlining the callsite will cause it to go from an
702 // indirect to a direct call by resolving a function pointer. If this
703 // happens, set the callee of the new call site to a more precise
704 // destination. This can also happen if the call graph node of the caller
705 // was just unnecessarily imprecise.
706 if (!I->second->getFunction())
707 if (Function *F = CallSite(NewCall).getCalledFunction()) {
708 // Indirect call site resolved to direct call.
709 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
711 continue;
712 }
714 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
715 }
717 // Update the call graph by deleting the edge from Callee to Caller. We must
718 // do this after the loop above in case Caller and Callee are the same.
719 CallerNode->removeCallEdgeFor(CS);
720 }
722 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
723 BasicBlock *InsertBlock,
724 InlineFunctionInfo &IFI) {
725 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
726 IRBuilder<> Builder(InsertBlock->begin());
728 Value *Size;
729 if (IFI.DL == nullptr)
730 Size = ConstantExpr::getSizeOf(AggTy);
731 else
732 Size = Builder.getInt64(IFI.DL->getTypeStoreSize(AggTy));
734 // Always generate a memcpy of alignment 1 here because we don't know
735 // the alignment of the src pointer. Other optimizations can infer
736 // better alignment.
737 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
738 }
740 /// HandleByValArgument - When inlining a call site that has a byval argument,
741 /// we have to make the implicit memcpy explicit by adding it.
742 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
743 const Function *CalledFunc,
744 InlineFunctionInfo &IFI,
745 unsigned ByValAlignment) {
746 PointerType *ArgTy = cast<PointerType>(Arg->getType());
747 Type *AggTy = ArgTy->getElementType();
749 Function *Caller = TheCall->getParent()->getParent();
751 // If the called function is readonly, then it could not mutate the caller's
752 // copy of the byval'd memory. In this case, it is safe to elide the copy and
753 // temporary.
754 if (CalledFunc->onlyReadsMemory()) {
755 // If the byval argument has a specified alignment that is greater than the
756 // passed in pointer, then we either have to round up the input pointer or
757 // give up on this transformation.
758 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
759 return Arg;
761 // If the pointer is already known to be sufficiently aligned, or if we can
762 // round it up to a larger alignment, then we don't need a temporary.
763 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, IFI.DL,
764 &IFI.ACT->getAssumptionCache(*Caller),
765 TheCall) >= ByValAlignment)
766 return Arg;
768 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
769 // for code quality, but rarely happens and is required for correctness.
770 }
772 // Create the alloca. If we have DataLayout, use nice alignment.
773 unsigned Align = 1;
774 if (IFI.DL)
775 Align = IFI.DL->getPrefTypeAlignment(AggTy);
777 // If the byval had an alignment specified, we *must* use at least that
778 // alignment, as it is required by the byval argument (and uses of the
779 // pointer inside the callee).
780 Align = std::max(Align, ByValAlignment);
782 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
783 &*Caller->begin()->begin());
784 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
786 // Uses of the argument in the function should use our new alloca
787 // instead.
788 return NewAlloca;
789 }
791 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
792 // intrinsic.
793 static bool isUsedByLifetimeMarker(Value *V) {
794 for (User *U : V->users()) {
795 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
796 switch (II->getIntrinsicID()) {
797 default: break;
798 case Intrinsic::lifetime_start:
799 case Intrinsic::lifetime_end:
800 return true;
801 }
802 }
803 }
804 return false;
805 }
807 // hasLifetimeMarkers - Check whether the given alloca already has
808 // lifetime.start or lifetime.end intrinsics.
809 static bool hasLifetimeMarkers(AllocaInst *AI) {
810 Type *Ty = AI->getType();
811 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
812 Ty->getPointerAddressSpace());
813 if (Ty == Int8PtrTy)
814 return isUsedByLifetimeMarker(AI);
816 // Do a scan to find all the casts to i8*.
817 for (User *U : AI->users()) {
818 if (U->getType() != Int8PtrTy) continue;
819 if (U->stripPointerCasts() != AI) continue;
820 if (isUsedByLifetimeMarker(U))
821 return true;
822 }
823 return false;
824 }
826 /// Rebuild the entire inlined-at chain for this instruction so that the top of
827 /// the chain now is inlined-at the new call site.
828 static DebugLoc
829 updateInlinedAtInfo(DebugLoc DL, MDLocation *InlinedAtNode,
830 LLVMContext &Ctx,
831 DenseMap<const MDLocation *, MDLocation *> &IANodes) {
832 SmallVector<MDLocation*, 3> InlinedAtLocations;
833 MDLocation *Last = InlinedAtNode;
834 DebugLoc CurInlinedAt = DL;
836 // Gather all the inlined-at nodes
837 while (MDLocation *IA =
838 cast_or_null<MDLocation>(CurInlinedAt.getInlinedAt(Ctx))) {
839 // Skip any we've already built nodes for
840 if (MDLocation *Found = IANodes[IA]) {
841 Last = Found;
842 break;
843 }
845 InlinedAtLocations.push_back(IA);
846 CurInlinedAt = DebugLoc::getFromDILocation(IA);
847 }
849 // Starting from the top, rebuild the nodes to point to the new inlined-at
850 // location (then rebuilding the rest of the chain behind it) and update the
851 // map of already-constructed inlined-at nodes.
852 for (auto I = InlinedAtLocations.rbegin(), E = InlinedAtLocations.rend();
853 I != E; ++I) {
854 const MDLocation *MD = *I;
855 Last = IANodes[MD] = MDLocation::getDistinct(
856 Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
857 }
859 // And finally create the normal location for this instruction, referring to
860 // the new inlined-at chain.
861 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), Last);
862 }
864 /// fixupLineNumbers - Update inlined instructions' line numbers to
865 /// to encode location where these instructions are inlined.
866 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
867 Instruction *TheCall) {
868 DebugLoc TheCallDL = TheCall->getDebugLoc();
869 if (TheCallDL.isUnknown())
870 return;
872 auto &Ctx = Fn->getContext();
873 auto *InlinedAtNode = cast<MDLocation>(TheCallDL.getAsMDNode(Ctx));
875 // Create a unique call site, not to be confused with any other call from the
876 // same location.
877 InlinedAtNode = MDLocation::getDistinct(
878 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
879 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
881 // Cache the inlined-at nodes as they're built so they are reused, without
882 // this every instruction's inlined-at chain would become distinct from each
883 // other.
884 DenseMap<const MDLocation *, MDLocation *> IANodes;
886 for (; FI != Fn->end(); ++FI) {
887 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
888 BI != BE; ++BI) {
889 DebugLoc DL = BI->getDebugLoc();
890 if (DL.isUnknown()) {
891 // If the inlined instruction has no line number, make it look as if it
892 // originates from the call location. This is important for
893 // ((__always_inline__, __nodebug__)) functions which must use caller
894 // location for all instructions in their function body.
896 // Don't update static allocas, as they may get moved later.
897 if (auto *AI = dyn_cast<AllocaInst>(BI))
898 if (isa<Constant>(AI->getArraySize()))
899 continue;
901 BI->setDebugLoc(TheCallDL);
902 } else {
903 BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
904 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
905 LLVMContext &Ctx = BI->getContext();
906 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
907 DVI->setOperand(2, MetadataAsValue::get(
908 Ctx, createInlinedVariable(DVI->getVariable(),
909 InlinedAt, Ctx)));
910 } else if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(BI)) {
911 LLVMContext &Ctx = BI->getContext();
912 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
913 DDI->setOperand(1, MetadataAsValue::get(
914 Ctx, createInlinedVariable(DDI->getVariable(),
915 InlinedAt, Ctx)));
916 }
917 }
918 }
919 }
920 }
922 /// InlineFunction - This function inlines the called function into the basic
923 /// block of the caller. This returns false if it is not possible to inline
924 /// this call. The program is still in a well defined state if this occurs
925 /// though.
926 ///
927 /// Note that this only does one level of inlining. For example, if the
928 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
929 /// exists in the instruction stream. Similarly this will inline a recursive
930 /// function by one level.
931 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
932 bool InsertLifetime) {
933 Instruction *TheCall = CS.getInstruction();
934 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
935 "Instruction not in function!");
937 // If IFI has any state in it, zap it before we fill it in.
938 IFI.reset();
940 const Function *CalledFunc = CS.getCalledFunction();
941 if (!CalledFunc || // Can't inline external function or indirect
942 CalledFunc->isDeclaration() || // call, or call to a vararg function!
943 CalledFunc->getFunctionType()->isVarArg()) return false;
945 // If the call to the callee cannot throw, set the 'nounwind' flag on any
946 // calls that we inline.
947 bool MarkNoUnwind = CS.doesNotThrow();
949 BasicBlock *OrigBB = TheCall->getParent();
950 Function *Caller = OrigBB->getParent();
952 // GC poses two hazards to inlining, which only occur when the callee has GC:
953 // 1. If the caller has no GC, then the callee's GC must be propagated to the
954 // caller.
955 // 2. If the caller has a differing GC, it is invalid to inline.
956 if (CalledFunc->hasGC()) {
957 if (!Caller->hasGC())
958 Caller->setGC(CalledFunc->getGC());
959 else if (CalledFunc->getGC() != Caller->getGC())
960 return false;
961 }
963 // Get the personality function from the callee if it contains a landing pad.
964 Value *CalleePersonality = nullptr;
965 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
966 I != E; ++I)
967 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
968 const BasicBlock *BB = II->getUnwindDest();
969 const LandingPadInst *LP = BB->getLandingPadInst();
970 CalleePersonality = LP->getPersonalityFn();
971 break;
972 }
974 // Find the personality function used by the landing pads of the caller. If it
975 // exists, then check to see that it matches the personality function used in
976 // the callee.
977 if (CalleePersonality) {
978 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
979 I != E; ++I)
980 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
981 const BasicBlock *BB = II->getUnwindDest();
982 const LandingPadInst *LP = BB->getLandingPadInst();
984 // If the personality functions match, then we can perform the
985 // inlining. Otherwise, we can't inline.
986 // TODO: This isn't 100% true. Some personality functions are proper
987 // supersets of others and can be used in place of the other.
988 if (LP->getPersonalityFn() != CalleePersonality)
989 return false;
991 break;
992 }
993 }
995 // Get an iterator to the last basic block in the function, which will have
996 // the new function inlined after it.
997 Function::iterator LastBlock = &Caller->back();
999 // Make sure to capture all of the return instructions from the cloned
1000 // function.
1001 SmallVector<ReturnInst*, 8> Returns;
1002 ClonedCodeInfo InlinedFunctionInfo;
1003 Function::iterator FirstNewBlock;
1005 { // Scope to destroy VMap after cloning.
1006 ValueToValueMapTy VMap;
1007 // Keep a list of pair (dst, src) to emit byval initializations.
1008 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
1010 assert(CalledFunc->arg_size() == CS.arg_size() &&
1011 "No varargs calls can be inlined!");
1013 // Calculate the vector of arguments to pass into the function cloner, which
1014 // matches up the formal to the actual argument values.
1015 CallSite::arg_iterator AI = CS.arg_begin();
1016 unsigned ArgNo = 0;
1017 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
1018 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
1019 Value *ActualArg = *AI;
1021 // When byval arguments actually inlined, we need to make the copy implied
1022 // by them explicit. However, we don't do this if the callee is readonly
1023 // or readnone, because the copy would be unneeded: the callee doesn't
1024 // modify the struct.
1025 if (CS.isByValArgument(ArgNo)) {
1026 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
1027 CalledFunc->getParamAlignment(ArgNo+1));
1028 if (ActualArg != *AI)
1029 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
1030 }
1032 VMap[I] = ActualArg;
1033 }
1035 // Add alignment assumptions if necessary. We do this before the inlined
1036 // instructions are actually cloned into the caller so that we can easily
1037 // check what will be known at the start of the inlined code.
1038 AddAlignmentAssumptions(CS, IFI);
1040 // We want the inliner to prune the code as it copies. We would LOVE to
1041 // have no dead or constant instructions leftover after inlining occurs
1042 // (which can happen, e.g., because an argument was constant), but we'll be
1043 // happy with whatever the cloner can do.
1044 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1045 /*ModuleLevelChanges=*/false, Returns, ".i",
1046 &InlinedFunctionInfo, IFI.DL, TheCall);
1048 // Remember the first block that is newly cloned over.
1049 FirstNewBlock = LastBlock; ++FirstNewBlock;
1051 // Inject byval arguments initialization.
1052 for (std::pair<Value*, Value*> &Init : ByValInit)
1053 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1054 FirstNewBlock, IFI);
1056 // Update the callgraph if requested.
1057 if (IFI.CG)
1058 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1060 // Update inlined instructions' line number information.
1061 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1063 // Clone existing noalias metadata if necessary.
1064 CloneAliasScopeMetadata(CS, VMap);
1066 // Add noalias metadata if necessary.
1067 AddAliasScopeMetadata(CS, VMap, IFI.DL, IFI.AA);
1069 // FIXME: We could register any cloned assumptions instead of clearing the
1070 // whole function's cache.
1071 if (IFI.ACT)
1072 IFI.ACT->getAssumptionCache(*Caller).clear();
1073 }
1075 // If there are any alloca instructions in the block that used to be the entry
1076 // block for the callee, move them to the entry block of the caller. First
1077 // calculate which instruction they should be inserted before. We insert the
1078 // instructions at the end of the current alloca list.
1079 {
1080 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1081 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1082 E = FirstNewBlock->end(); I != E; ) {
1083 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1084 if (!AI) continue;
1086 // If the alloca is now dead, remove it. This often occurs due to code
1087 // specialization.
1088 if (AI->use_empty()) {
1089 AI->eraseFromParent();
1090 continue;
1091 }
1093 if (!isa<Constant>(AI->getArraySize()))
1094 continue;
1096 // Keep track of the static allocas that we inline into the caller.
1097 IFI.StaticAllocas.push_back(AI);
1099 // Scan for the block of allocas that we can move over, and move them
1100 // all at once.
1101 while (isa<AllocaInst>(I) &&
1102 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1103 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1104 ++I;
1105 }
1107 // Transfer all of the allocas over in a block. Using splice means
1108 // that the instructions aren't removed from the symbol table, then
1109 // reinserted.
1110 Caller->getEntryBlock().getInstList().splice(InsertPoint,
1111 FirstNewBlock->getInstList(),
1112 AI, I);
1113 }
1114 }
1116 bool InlinedMustTailCalls = false;
1117 if (InlinedFunctionInfo.ContainsCalls) {
1118 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1119 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1120 CallSiteTailKind = CI->getTailCallKind();
1122 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1123 ++BB) {
1124 for (Instruction &I : *BB) {
1125 CallInst *CI = dyn_cast<CallInst>(&I);
1126 if (!CI)
1127 continue;
1129 // We need to reduce the strength of any inlined tail calls. For
1130 // musttail, we have to avoid introducing potential unbounded stack
1131 // growth. For example, if functions 'f' and 'g' are mutually recursive
1132 // with musttail, we can inline 'g' into 'f' so long as we preserve
1133 // musttail on the cloned call to 'f'. If either the inlined call site
1134 // or the cloned call site is *not* musttail, the program already has
1135 // one frame of stack growth, so it's safe to remove musttail. Here is
1136 // a table of example transformations:
1137 //
1138 // f -> musttail g -> musttail f ==> f -> musttail f
1139 // f -> musttail g -> tail f ==> f -> tail f
1140 // f -> g -> musttail f ==> f -> f
1141 // f -> g -> tail f ==> f -> f
1142 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1143 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1144 CI->setTailCallKind(ChildTCK);
1145 InlinedMustTailCalls |= CI->isMustTailCall();
1147 // Calls inlined through a 'nounwind' call site should be marked
1148 // 'nounwind'.
1149 if (MarkNoUnwind)
1150 CI->setDoesNotThrow();
1151 }
1152 }
1153 }
1155 // Leave lifetime markers for the static alloca's, scoping them to the
1156 // function we just inlined.
1157 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1158 IRBuilder<> builder(FirstNewBlock->begin());
1159 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1160 AllocaInst *AI = IFI.StaticAllocas[ai];
1162 // If the alloca is already scoped to something smaller than the whole
1163 // function then there's no need to add redundant, less accurate markers.
1164 if (hasLifetimeMarkers(AI))
1165 continue;
1167 // Try to determine the size of the allocation.
1168 ConstantInt *AllocaSize = nullptr;
1169 if (ConstantInt *AIArraySize =
1170 dyn_cast<ConstantInt>(AI->getArraySize())) {
1171 if (IFI.DL) {
1172 Type *AllocaType = AI->getAllocatedType();
1173 uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType);
1174 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1175 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
1176 // Check that array size doesn't saturate uint64_t and doesn't
1177 // overflow when it's multiplied by type size.
1178 if (AllocaArraySize != ~0ULL &&
1179 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1180 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1181 AllocaArraySize * AllocaTypeSize);
1182 }
1183 }
1184 }
1186 builder.CreateLifetimeStart(AI, AllocaSize);
1187 for (ReturnInst *RI : Returns) {
1188 // Don't insert llvm.lifetime.end calls between a musttail call and a
1189 // return. The return kills all local allocas.
1190 if (InlinedMustTailCalls &&
1191 RI->getParent()->getTerminatingMustTailCall())
1192 continue;
1193 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1194 }
1195 }
1196 }
1198 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1199 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1200 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1201 Module *M = Caller->getParent();
1202 // Get the two intrinsics we care about.
1203 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1204 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1206 // Insert the llvm.stacksave.
1207 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1208 .CreateCall(StackSave, "savedstack");
1210 // Insert a call to llvm.stackrestore before any return instructions in the
1211 // inlined function.
1212 for (ReturnInst *RI : Returns) {
1213 // Don't insert llvm.stackrestore calls between a musttail call and a
1214 // return. The return will restore the stack pointer.
1215 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1216 continue;
1217 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1218 }
1219 }
1221 // If we are inlining for an invoke instruction, we must make sure to rewrite
1222 // any call instructions into invoke instructions.
1223 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1224 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
1226 // Handle any inlined musttail call sites. In order for a new call site to be
1227 // musttail, the source of the clone and the inlined call site must have been
1228 // musttail. Therefore it's safe to return without merging control into the
1229 // phi below.
1230 if (InlinedMustTailCalls) {
1231 // Check if we need to bitcast the result of any musttail calls.
1232 Type *NewRetTy = Caller->getReturnType();
1233 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1235 // Handle the returns preceded by musttail calls separately.
1236 SmallVector<ReturnInst *, 8> NormalReturns;
1237 for (ReturnInst *RI : Returns) {
1238 CallInst *ReturnedMustTail =
1239 RI->getParent()->getTerminatingMustTailCall();
1240 if (!ReturnedMustTail) {
1241 NormalReturns.push_back(RI);
1242 continue;
1243 }
1244 if (!NeedBitCast)
1245 continue;
1247 // Delete the old return and any preceding bitcast.
1248 BasicBlock *CurBB = RI->getParent();
1249 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1250 RI->eraseFromParent();
1251 if (OldCast)
1252 OldCast->eraseFromParent();
1254 // Insert a new bitcast and return with the right type.
1255 IRBuilder<> Builder(CurBB);
1256 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1257 }
1259 // Leave behind the normal returns so we can merge control flow.
1260 std::swap(Returns, NormalReturns);
1261 }
1263 // If we cloned in _exactly one_ basic block, and if that block ends in a
1264 // return instruction, we splice the body of the inlined callee directly into
1265 // the calling basic block.
1266 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1267 // Move all of the instructions right before the call.
1268 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1269 FirstNewBlock->begin(), FirstNewBlock->end());
1270 // Remove the cloned basic block.
1271 Caller->getBasicBlockList().pop_back();
1273 // If the call site was an invoke instruction, add a branch to the normal
1274 // destination.
1275 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1276 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1277 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1278 }
1280 // If the return instruction returned a value, replace uses of the call with
1281 // uses of the returned value.
1282 if (!TheCall->use_empty()) {
1283 ReturnInst *R = Returns[0];
1284 if (TheCall == R->getReturnValue())
1285 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1286 else
1287 TheCall->replaceAllUsesWith(R->getReturnValue());
1288 }
1289 // Since we are now done with the Call/Invoke, we can delete it.
1290 TheCall->eraseFromParent();
1292 // Since we are now done with the return instruction, delete it also.
1293 Returns[0]->eraseFromParent();
1295 // We are now done with the inlining.
1296 return true;
1297 }
1299 // Otherwise, we have the normal case, of more than one block to inline or
1300 // multiple return sites.
1302 // We want to clone the entire callee function into the hole between the
1303 // "starter" and "ender" blocks. How we accomplish this depends on whether
1304 // this is an invoke instruction or a call instruction.
1305 BasicBlock *AfterCallBB;
1306 BranchInst *CreatedBranchToNormalDest = nullptr;
1307 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1309 // Add an unconditional branch to make this look like the CallInst case...
1310 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1312 // Split the basic block. This guarantees that no PHI nodes will have to be
1313 // updated due to new incoming edges, and make the invoke case more
1314 // symmetric to the call case.
1315 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
1316 CalledFunc->getName()+".exit");
1318 } else { // It's a call
1319 // If this is a call instruction, we need to split the basic block that
1320 // the call lives in.
1321 //
1322 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1323 CalledFunc->getName()+".exit");
1324 }
1326 // Change the branch that used to go to AfterCallBB to branch to the first
1327 // basic block of the inlined function.
1328 //
1329 TerminatorInst *Br = OrigBB->getTerminator();
1330 assert(Br && Br->getOpcode() == Instruction::Br &&
1331 "splitBasicBlock broken!");
1332 Br->setOperand(0, FirstNewBlock);
1335 // Now that the function is correct, make it a little bit nicer. In
1336 // particular, move the basic blocks inserted from the end of the function
1337 // into the space made by splitting the source basic block.
1338 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1339 FirstNewBlock, Caller->end());
1341 // Handle all of the return instructions that we just cloned in, and eliminate
1342 // any users of the original call/invoke instruction.
1343 Type *RTy = CalledFunc->getReturnType();
1345 PHINode *PHI = nullptr;
1346 if (Returns.size() > 1) {
1347 // The PHI node should go at the front of the new basic block to merge all
1348 // possible incoming values.
1349 if (!TheCall->use_empty()) {
1350 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1351 AfterCallBB->begin());
1352 // Anything that used the result of the function call should now use the
1353 // PHI node as their operand.
1354 TheCall->replaceAllUsesWith(PHI);
1355 }
1357 // Loop over all of the return instructions adding entries to the PHI node
1358 // as appropriate.
1359 if (PHI) {
1360 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1361 ReturnInst *RI = Returns[i];
1362 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1363 "Ret value not consistent in function!");
1364 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1365 }
1366 }
1369 // Add a branch to the merge points and remove return instructions.
1370 DebugLoc Loc;
1371 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1372 ReturnInst *RI = Returns[i];
1373 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1374 Loc = RI->getDebugLoc();
1375 BI->setDebugLoc(Loc);
1376 RI->eraseFromParent();
1377 }
1378 // We need to set the debug location to *somewhere* inside the
1379 // inlined function. The line number may be nonsensical, but the
1380 // instruction will at least be associated with the right
1381 // function.
1382 if (CreatedBranchToNormalDest)
1383 CreatedBranchToNormalDest->setDebugLoc(Loc);
1384 } else if (!Returns.empty()) {
1385 // Otherwise, if there is exactly one return value, just replace anything
1386 // using the return value of the call with the computed value.
1387 if (!TheCall->use_empty()) {
1388 if (TheCall == Returns[0]->getReturnValue())
1389 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1390 else
1391 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1392 }
1394 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1395 BasicBlock *ReturnBB = Returns[0]->getParent();
1396 ReturnBB->replaceAllUsesWith(AfterCallBB);
1398 // Splice the code from the return block into the block that it will return
1399 // to, which contains the code that was after the call.
1400 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1401 ReturnBB->getInstList());
1403 if (CreatedBranchToNormalDest)
1404 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1406 // Delete the return instruction now and empty ReturnBB now.
1407 Returns[0]->eraseFromParent();
1408 ReturnBB->eraseFromParent();
1409 } else if (!TheCall->use_empty()) {
1410 // No returns, but something is using the return value of the call. Just
1411 // nuke the result.
1412 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1413 }
1415 // Since we are now done with the Call/Invoke, we can delete it.
1416 TheCall->eraseFromParent();
1418 // If we inlined any musttail calls and the original return is now
1419 // unreachable, delete it. It can only contain a bitcast and ret.
1420 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1421 AfterCallBB->eraseFromParent();
1423 // We should always be able to fold the entry block of the function into the
1424 // single predecessor of the block...
1425 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1426 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1428 // Splice the code entry block into calling block, right before the
1429 // unconditional branch.
1430 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1431 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1433 // Remove the unconditional branch.
1434 OrigBB->getInstList().erase(Br);
1436 // Now we can remove the CalleeEntry block, which is now empty.
1437 Caller->getBasicBlockList().erase(CalleeEntry);
1439 // If we inserted a phi node, check to see if it has a single value (e.g. all
1440 // the entries are the same or undef). If so, remove the PHI so it doesn't
1441 // block other optimizations.
1442 if (PHI) {
1443 if (Value *V = SimplifyInstruction(PHI, IFI.DL, nullptr, nullptr,
1444 &IFI.ACT->getAssumptionCache(*Caller))) {
1445 PHI->replaceAllUsesWith(V);
1446 PHI->eraseFromParent();
1447 }
1448 }
1450 return true;
1451 }